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  * Copyright 2010 Red Hat, Inc. and/or its affilates.
  11  *
  12  * Authors:
  13  *   Yaniv Kamay  <yaniv@qumranet.com>
  14  *   Avi Kivity   <avi@qumranet.com>
  15  *
  16  * This work is licensed under the terms of the GNU GPL, version 2.  See
  17  * the COPYING file in the top-level directory.
  18  *
  19  */
  20
  21 #include "mmu.h"
  22 #include "x86.h"
  23 #include "kvm_cache_regs.h"
  24
  25 #include <linux/kvm_host.h>
  26 #include <linux/types.h>
  27 #include <linux/string.h>
  28 #include <linux/mm.h>
  29 #include <linux/highmem.h>
  30 #include <linux/module.h>
  31 #include <linux/swap.h>
  32 #include <linux/hugetlb.h>
  33 #include <linux/compiler.h>
  34 #include <linux/srcu.h>
  35 #include <linux/slab.h>
  36 #include <linux/uaccess.h>
  37
  38 #include <asm/page.h>
  39 #include <asm/cmpxchg.h>
  40 #include <asm/io.h>
  41 #include <asm/vmx.h>
  42
  43 /*
  44  * When setting this variable to true it enables Two-Dimensional-Paging
  45  * where the hardware walks 2 page tables:
  46  * 1. the guest-virtual to guest-physical
  47  * 2. while doing 1. it walks guest-physical to host-physical
  48  * If the hardware supports that we don't need to do shadow paging.
  49  */
  50 bool tdp_enabled = false;
  51
  52 #undef MMU_DEBUG
  53
  54 #undef AUDIT
  55
  56 #ifdef AUDIT
  57 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  58 #else
  59 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  60 #endif
  61
  62 #ifdef MMU_DEBUG
  63
  64 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  65 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  66
  67 #else
  68
  69 #define pgprintk(x...) do { } while (0)
  70 #define rmap_printk(x...) do { } while (0)
  71
  72 #endif
  73
  74 #if defined(MMU_DEBUG) || defined(AUDIT)
  75 static int dbg = 0;
  76 module_param(dbg, bool, 0644);
  77 #endif
  78
  79 static int oos_shadow = 1;
  80 module_param(oos_shadow, bool, 0644);
  81
  82 #ifndef MMU_DEBUG
  83 #define ASSERT(x) do { } while (0)
  84 #else
  85 #define ASSERT(x)                                                       \
  86         if (!(x)) {                                                     \
  87                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  88                        __FILE__, __LINE__, #x);                         \
  89         }
  90 #endif
  91
  92 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  93 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  94
  95 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  96
  97 #define PT64_LEVEL_BITS 9
  98
  99 #define PT64_LEVEL_SHIFT(level) \
 100                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
 101
 102 #define PT64_LEVEL_MASK(level) \
 103                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 104
 105 #define PT64_INDEX(address, level)\
 106         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 107
 108
 109 #define PT32_LEVEL_BITS 10
 110
 111 #define PT32_LEVEL_SHIFT(level) \
 112                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 113
 114 #define PT32_LEVEL_MASK(level) \
 115                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 116 #define PT32_LVL_OFFSET_MASK(level) \
 117         (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 118                                                 * PT32_LEVEL_BITS))) - 1))
 119
 120 #define PT32_INDEX(address, level)\
 121         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 122
 123
 124 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 125 #define PT64_DIR_BASE_ADDR_MASK \
 126         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 127 #define PT64_LVL_ADDR_MASK(level) \
 128         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 129                                                 * PT64_LEVEL_BITS))) - 1))
 130 #define PT64_LVL_OFFSET_MASK(level) \
 131         (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 132                                                 * PT64_LEVEL_BITS))) - 1))
 133
 134 #define PT32_BASE_ADDR_MASK PAGE_MASK
 135 #define PT32_DIR_BASE_ADDR_MASK \
 136         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 137 #define PT32_LVL_ADDR_MASK(level) \
 138         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 139                                             * PT32_LEVEL_BITS))) - 1))
 140
 141 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 142                         | PT64_NX_MASK)
 143
 144 #define RMAP_EXT 4
 145
 146 #define ACC_EXEC_MASK    1
 147 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 148 #define ACC_USER_MASK    PT_USER_MASK
 149 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 150
 151 #include <trace/events/kvm.h>
 152
 153 #define CREATE_TRACE_POINTS
 154 #include "mmutrace.h"
 155
 156 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 157
 158 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 159
 160 struct kvm_rmap_desc {
 161         u64 *sptes[RMAP_EXT];
 162         struct kvm_rmap_desc *more;
 163 };
 164
 165 struct kvm_shadow_walk_iterator {
 166         u64 addr;
 167         hpa_t shadow_addr;
 168         int level;
 169         u64 *sptep;
 170         unsigned index;
 171 };
 172
 173 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
 174         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
 175              shadow_walk_okay(&(_walker));                      \
 176              shadow_walk_next(&(_walker)))
 177
 178 typedef int (*mmu_parent_walk_fn) (struct kvm_mmu_page *sp);
 179
 180 static struct kmem_cache *pte_chain_cache;
 181 static struct kmem_cache *rmap_desc_cache;
 182 static struct kmem_cache *mmu_page_header_cache;
 183
 184 static u64 __read_mostly shadow_trap_nonpresent_pte;
 185 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 186 static u64 __read_mostly shadow_base_present_pte;
 187 static u64 __read_mostly shadow_nx_mask;
 188 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 189 static u64 __read_mostly shadow_user_mask;
 190 static u64 __read_mostly shadow_accessed_mask;
 191 static u64 __read_mostly shadow_dirty_mask;
 192
 193 static inline u64 rsvd_bits(int s, int e)
 194 {
 195         return ((1ULL << (e - s + 1)) - 1) << s;
 196 }
 197
 198 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 199 {
 200         shadow_trap_nonpresent_pte = trap_pte;
 201         shadow_notrap_nonpresent_pte = notrap_pte;
 202 }
 203 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 204
 205 void kvm_mmu_set_base_ptes(u64 base_pte)
 206 {
 207         shadow_base_present_pte = base_pte;
 208 }
 209 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 210
 211 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 212                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 213 {
 214         shadow_user_mask = user_mask;
 215         shadow_accessed_mask = accessed_mask;
 216         shadow_dirty_mask = dirty_mask;
 217         shadow_nx_mask = nx_mask;
 218         shadow_x_mask = x_mask;
 219 }
 220 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 221
 222 static bool is_write_protection(struct kvm_vcpu *vcpu)
 223 {
 224         return kvm_read_cr0_bits(vcpu, X86_CR0_WP);
 225 }
 226
 227 static int is_cpuid_PSE36(void)
 228 {
 229         return 1;
 230 }
 231
 232 static int is_nx(struct kvm_vcpu *vcpu)
 233 {
 234         return vcpu->arch.efer & EFER_NX;
 235 }
 236
 237 static int is_shadow_present_pte(u64 pte)
 238 {
 239         return pte != shadow_trap_nonpresent_pte
 240                 && pte != shadow_notrap_nonpresent_pte;
 241 }
 242
 243 static int is_large_pte(u64 pte)
 244 {
 245         return pte & PT_PAGE_SIZE_MASK;
 246 }
 247
 248 static int is_writable_pte(unsigned long pte)
 249 {
 250         return pte & PT_WRITABLE_MASK;
 251 }
 252
 253 static int is_dirty_gpte(unsigned long pte)
 254 {
 255         return pte & PT_DIRTY_MASK;
 256 }
 257
 258 static int is_rmap_spte(u64 pte)
 259 {
 260         return is_shadow_present_pte(pte);
 261 }
 262
 263 static int is_last_spte(u64 pte, int level)
 264 {
 265         if (level == PT_PAGE_TABLE_LEVEL)
 266                 return 1;
 267         if (is_large_pte(pte))
 268                 return 1;
 269         return 0;
 270 }
 271
 272 static pfn_t spte_to_pfn(u64 pte)
 273 {
 274         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 275 }
 276
 277 static gfn_t pse36_gfn_delta(u32 gpte)
 278 {
 279         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 280
 281         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 282 }
 283
 284 static void __set_spte(u64 *sptep, u64 spte)
 285 {
 286 #ifdef CONFIG_X86_64
 287         set_64bit((unsigned long *)sptep, spte);
 288 #else
 289         set_64bit((unsigned long long *)sptep, spte);
 290 #endif
 291 }
 292
 293 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 294                                   struct kmem_cache *base_cache, int min)
 295 {
 296         void *obj;
 297
 298         if (cache->nobjs >= min)
 299                 return 0;
 300         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 301                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 302                 if (!obj)
 303                         return -ENOMEM;
 304                 cache->objects[cache->nobjs++] = obj;
 305         }
 306         return 0;
 307 }
 308
 309 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
 310                                   struct kmem_cache *cache)
 311 {
 312         while (mc->nobjs)
 313                 kmem_cache_free(cache, mc->objects[--mc->nobjs]);
 314 }
 315
 316 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 317                                        int min)
 318 {
 319         struct page *page;
 320
 321         if (cache->nobjs >= min)
 322                 return 0;
 323         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 324                 page = alloc_page(GFP_KERNEL);
 325                 if (!page)
 326                         return -ENOMEM;
 327                 cache->objects[cache->nobjs++] = page_address(page);
 328         }
 329         return 0;
 330 }
 331
 332 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 333 {
 334         while (mc->nobjs)
 335                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 336 }
 337
 338 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 339 {
 340         int r;
 341
 342         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 343                                    pte_chain_cache, 4);
 344         if (r)
 345                 goto out;
 346         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 347                                    rmap_desc_cache, 4);
 348         if (r)
 349                 goto out;
 350         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 351         if (r)
 352                 goto out;
 353         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 354                                    mmu_page_header_cache, 4);
 355 out:
 356         return r;
 357 }
 358
 359 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 360 {
 361         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache, pte_chain_cache);
 362         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache, rmap_desc_cache);
 363         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 364         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
 365                                 mmu_page_header_cache);
 366 }
 367
 368 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 369                                     size_t size)
 370 {
 371         void *p;
 372
 373         BUG_ON(!mc->nobjs);
 374         p = mc->objects[--mc->nobjs];
 375         return p;
 376 }
 377
 378 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 379 {
 380         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 381                                       sizeof(struct kvm_pte_chain));
 382 }
 383
 384 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 385 {
 386         kmem_cache_free(pte_chain_cache, pc);
 387 }
 388
 389 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 390 {
 391         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 392                                       sizeof(struct kvm_rmap_desc));
 393 }
 394
 395 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 396 {
 397         kmem_cache_free(rmap_desc_cache, rd);
 398 }
 399
 400 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
 401 {
 402         if (!sp->role.direct)
 403                 return sp->gfns[index];
 404
 405         return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
 406 }
 407
 408 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
 409 {
 410         if (sp->role.direct)
 411                 BUG_ON(gfn != kvm_mmu_page_get_gfn(sp, index));
 412         else
 413                 sp->gfns[index] = gfn;
 414 }
 415
 416 /*
 417  * Return the pointer to the largepage write count for a given
 418  * gfn, handling slots that are not large page aligned.
 419  */
 420 static int *slot_largepage_idx(gfn_t gfn,
 421                                struct kvm_memory_slot *slot,
 422                                int level)
 423 {
 424         unsigned long idx;
 425
 426         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 427               (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 428         return &slot->lpage_info[level - 2][idx].write_count;
 429 }
 430
 431 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 432 {
 433         struct kvm_memory_slot *slot;
 434         int *write_count;
 435         int i;
 436
 437         gfn = unalias_gfn(kvm, gfn);
 438
 439         slot = gfn_to_memslot_unaliased(kvm, gfn);
 440         for (i = PT_DIRECTORY_LEVEL;
 441              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 442                 write_count   = slot_largepage_idx(gfn, slot, i);
 443                 *write_count += 1;
 444         }
 445 }
 446
 447 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 448 {
 449         struct kvm_memory_slot *slot;
 450         int *write_count;
 451         int i;
 452
 453         gfn = unalias_gfn(kvm, gfn);
 454         slot = gfn_to_memslot_unaliased(kvm, gfn);
 455         for (i = PT_DIRECTORY_LEVEL;
 456              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 457                 write_count   = slot_largepage_idx(gfn, slot, i);
 458                 *write_count -= 1;
 459                 WARN_ON(*write_count < 0);
 460         }
 461 }
 462
 463 static int has_wrprotected_page(struct kvm *kvm,
 464                                 gfn_t gfn,
 465                                 int level)
 466 {
 467         struct kvm_memory_slot *slot;
 468         int *largepage_idx;
 469
 470         gfn = unalias_gfn(kvm, gfn);
 471         slot = gfn_to_memslot_unaliased(kvm, gfn);
 472         if (slot) {
 473                 largepage_idx = slot_largepage_idx(gfn, slot, level);
 474                 return *largepage_idx;
 475         }
 476
 477         return 1;
 478 }
 479
 480 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
 481 {
 482         unsigned long page_size;
 483         int i, ret = 0;
 484
 485         page_size = kvm_host_page_size(kvm, gfn);
 486
 487         for (i = PT_PAGE_TABLE_LEVEL;
 488              i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
 489                 if (page_size >= KVM_HPAGE_SIZE(i))
 490                         ret = i;
 491                 else
 492                         break;
 493         }
 494
 495         return ret;
 496 }
 497
 498 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 499 {
 500         struct kvm_memory_slot *slot;
 501         int host_level, level, max_level;
 502
 503         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 504         if (slot && slot->dirty_bitmap)
 505                 return PT_PAGE_TABLE_LEVEL;
 506
 507         host_level = host_mapping_level(vcpu->kvm, large_gfn);
 508
 509         if (host_level == PT_PAGE_TABLE_LEVEL)
 510                 return host_level;
 511
 512         max_level = kvm_x86_ops->get_lpage_level() < host_level ?
 513                 kvm_x86_ops->get_lpage_level() : host_level;
 514
 515         for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
 516                 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
 517                         break;
 518
 519         return level - 1;
 520 }
 521
 522 /*
 523  * Take gfn and return the reverse mapping to it.
 524  * Note: gfn must be unaliased before this function get called
 525  */
 526
 527 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
 528 {
 529         struct kvm_memory_slot *slot;
 530         unsigned long idx;
 531
 532         slot = gfn_to_memslot(kvm, gfn);
 533         if (likely(level == PT_PAGE_TABLE_LEVEL))
 534                 return &slot->rmap[gfn - slot->base_gfn];
 535
 536         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 537                 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 538
 539         return &slot->lpage_info[level - 2][idx].rmap_pde;
 540 }
 541
 542 /*
 543  * Reverse mapping data structures:
 544  *
 545  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 546  * that points to page_address(page).
 547  *
 548  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 549  * containing more mappings.
 550  *
 551  * Returns the number of rmap entries before the spte was added or zero if
 552  * the spte was not added.
 553  *
 554  */
 555 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 556 {
 557         struct kvm_mmu_page *sp;
 558         struct kvm_rmap_desc *desc;
 559         unsigned long *rmapp;
 560         int i, count = 0;
 561
 562         if (!is_rmap_spte(*spte))
 563                 return count;
 564         gfn = unalias_gfn(vcpu->kvm, gfn);
 565         sp = page_header(__pa(spte));
 566         kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
 567         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 568         if (!*rmapp) {
 569                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 570                 *rmapp = (unsigned long)spte;
 571         } else if (!(*rmapp & 1)) {
 572                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 573                 desc = mmu_alloc_rmap_desc(vcpu);
 574                 desc->sptes[0] = (u64 *)*rmapp;
 575                 desc->sptes[1] = spte;
 576                 *rmapp = (unsigned long)desc | 1;
 577         } else {
 578                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 579                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 580                 while (desc->sptes[RMAP_EXT-1] && desc->more) {
 581                         desc = desc->more;
 582                         count += RMAP_EXT;
 583                 }
 584                 if (desc->sptes[RMAP_EXT-1]) {
 585                         desc->more = mmu_alloc_rmap_desc(vcpu);
 586                         desc = desc->more;
 587                 }
 588                 for (i = 0; desc->sptes[i]; ++i)
 589                         ;
 590                 desc->sptes[i] = spte;
 591         }
 592         return count;
 593 }
 594
 595 static void rmap_desc_remove_entry(unsigned long *rmapp,
 596                                    struct kvm_rmap_desc *desc,
 597                                    int i,
 598                                    struct kvm_rmap_desc *prev_desc)
 599 {
 600         int j;
 601
 602         for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
 603                 ;
 604         desc->sptes[i] = desc->sptes[j];
 605         desc->sptes[j] = NULL;
 606         if (j != 0)
 607                 return;
 608         if (!prev_desc && !desc->more)
 609                 *rmapp = (unsigned long)desc->sptes[0];
 610         else
 611                 if (prev_desc)
 612                         prev_desc->more = desc->more;
 613                 else
 614                         *rmapp = (unsigned long)desc->more | 1;
 615         mmu_free_rmap_desc(desc);
 616 }
 617
 618 static void rmap_remove(struct kvm *kvm, u64 *spte)
 619 {
 620         struct kvm_rmap_desc *desc;
 621         struct kvm_rmap_desc *prev_desc;
 622         struct kvm_mmu_page *sp;
 623         pfn_t pfn;
 624         gfn_t gfn;
 625         unsigned long *rmapp;
 626         int i;
 627
 628         if (!is_rmap_spte(*spte))
 629                 return;
 630         sp = page_header(__pa(spte));
 631         pfn = spte_to_pfn(*spte);
 632         if (*spte & shadow_accessed_mask)
 633                 kvm_set_pfn_accessed(pfn);
 634         if (is_writable_pte(*spte))
 635                 kvm_set_pfn_dirty(pfn);
 636         gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
 637         rmapp = gfn_to_rmap(kvm, gfn, sp->role.level);
 638         if (!*rmapp) {
 639                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 640                 BUG();
 641         } else if (!(*rmapp & 1)) {
 642                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 643                 if ((u64 *)*rmapp != spte) {
 644                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 645                                spte, *spte);
 646                         BUG();
 647                 }
 648                 *rmapp = 0;
 649         } else {
 650                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 651                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 652                 prev_desc = NULL;
 653                 while (desc) {
 654                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
 655                                 if (desc->sptes[i] == spte) {
 656                                         rmap_desc_remove_entry(rmapp,
 657                                                                desc, i,
 658                                                                prev_desc);
 659                                         return;
 660                                 }
 661                         prev_desc = desc;
 662                         desc = desc->more;
 663                 }
 664                 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
 665                 BUG();
 666         }
 667 }
 668
 669 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 670 {
 671         struct kvm_rmap_desc *desc;
 672         u64 *prev_spte;
 673         int i;
 674
 675         if (!*rmapp)
 676                 return NULL;
 677         else if (!(*rmapp & 1)) {
 678                 if (!spte)
 679                         return (u64 *)*rmapp;
 680                 return NULL;
 681         }
 682         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 683         prev_spte = NULL;
 684         while (desc) {
 685                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
 686                         if (prev_spte == spte)
 687                                 return desc->sptes[i];
 688                         prev_spte = desc->sptes[i];
 689                 }
 690                 desc = desc->more;
 691         }
 692         return NULL;
 693 }
 694
 695 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 696 {
 697         unsigned long *rmapp;
 698         u64 *spte;
 699         int i, write_protected = 0;
 700
 701         gfn = unalias_gfn(kvm, gfn);
 702         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
 703
 704         spte = rmap_next(kvm, rmapp, NULL);
 705         while (spte) {
 706                 BUG_ON(!spte);
 707                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 708                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 709                 if (is_writable_pte(*spte)) {
 710                         __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
 711                         write_protected = 1;
 712                 }
 713                 spte = rmap_next(kvm, rmapp, spte);
 714         }
 715         if (write_protected) {
 716                 pfn_t pfn;
 717
 718                 spte = rmap_next(kvm, rmapp, NULL);
 719                 pfn = spte_to_pfn(*spte);
 720                 kvm_set_pfn_dirty(pfn);
 721         }
 722
 723         /* check for huge page mappings */
 724         for (i = PT_DIRECTORY_LEVEL;
 725              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 726                 rmapp = gfn_to_rmap(kvm, gfn, i);
 727                 spte = rmap_next(kvm, rmapp, NULL);
 728                 while (spte) {
 729                         BUG_ON(!spte);
 730                         BUG_ON(!(*spte & PT_PRESENT_MASK));
 731                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 732                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 733                         if (is_writable_pte(*spte)) {
 734                                 rmap_remove(kvm, spte);
 735                                 --kvm->stat.lpages;
 736                                 __set_spte(spte, shadow_trap_nonpresent_pte);
 737                                 spte = NULL;
 738                                 write_protected = 1;
 739                         }
 740                         spte = rmap_next(kvm, rmapp, spte);
 741                 }
 742         }
 743
 744         return write_protected;
 745 }
 746
 747 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 748                            unsigned long data)
 749 {
 750         u64 *spte;
 751         int need_tlb_flush = 0;
 752
 753         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 754                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 755                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 756                 rmap_remove(kvm, spte);
 757                 __set_spte(spte, shadow_trap_nonpresent_pte);
 758                 need_tlb_flush = 1;
 759         }
 760         return need_tlb_flush;
 761 }
 762
 763 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
 764                              unsigned long data)
 765 {
 766         int need_flush = 0;
 767         u64 *spte, new_spte;
 768         pte_t *ptep = (pte_t *)data;
 769         pfn_t new_pfn;
 770
 771         WARN_ON(pte_huge(*ptep));
 772         new_pfn = pte_pfn(*ptep);
 773         spte = rmap_next(kvm, rmapp, NULL);
 774         while (spte) {
 775                 BUG_ON(!is_shadow_present_pte(*spte));
 776                 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
 777                 need_flush = 1;
 778                 if (pte_write(*ptep)) {
 779                         rmap_remove(kvm, spte);
 780                         __set_spte(spte, shadow_trap_nonpresent_pte);
 781                         spte = rmap_next(kvm, rmapp, NULL);
 782                 } else {
 783                         new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
 784                         new_spte |= (u64)new_pfn << PAGE_SHIFT;
 785
 786                         new_spte &= ~PT_WRITABLE_MASK;
 787                         new_spte &= ~SPTE_HOST_WRITEABLE;
 788                         if (is_writable_pte(*spte))
 789                                 kvm_set_pfn_dirty(spte_to_pfn(*spte));
 790                         __set_spte(spte, new_spte);
 791                         spte = rmap_next(kvm, rmapp, spte);
 792                 }
 793         }
 794         if (need_flush)
 795                 kvm_flush_remote_tlbs(kvm);
 796
 797         return 0;
 798 }
 799
 800 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 801                           unsigned long data,
 802                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 803                                          unsigned long data))
 804 {
 805         int i, j;
 806         int ret;
 807         int retval = 0;
 808         struct kvm_memslots *slots;
 809
 810         slots = kvm_memslots(kvm);
 811
 812         for (i = 0; i < slots->nmemslots; i++) {
 813                 struct kvm_memory_slot *memslot = &slots->memslots[i];
 814                 unsigned long start = memslot->userspace_addr;
 815                 unsigned long end;
 816
 817                 end = start + (memslot->npages << PAGE_SHIFT);
 818                 if (hva >= start && hva < end) {
 819                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 820
 821                         ret = handler(kvm, &memslot->rmap[gfn_offset], data);
 822
 823                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
 824                                 int idx = gfn_offset;
 825                                 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
 826                                 ret |= handler(kvm,
 827                                         &memslot->lpage_info[j][idx].rmap_pde,
 828                                         data);
 829                         }
 830                         trace_kvm_age_page(hva, memslot, ret);
 831                         retval |= ret;
 832                 }
 833         }
 834
 835         return retval;
 836 }
 837
 838 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 839 {
 840         return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
 841 }
 842
 843 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
 844 {
 845         kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
 846 }
 847
 848 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 849                          unsigned long data)
 850 {
 851         u64 *spte;
 852         int young = 0;
 853
 854         /*
 855          * Emulate the accessed bit for EPT, by checking if this page has
 856          * an EPT mapping, and clearing it if it does. On the next access,
 857          * a new EPT mapping will be established.
 858          * This has some overhead, but not as much as the cost of swapping
 859          * out actively used pages or breaking up actively used hugepages.
 860          */
 861         if (!shadow_accessed_mask)
 862                 return kvm_unmap_rmapp(kvm, rmapp, data);
 863
 864         spte = rmap_next(kvm, rmapp, NULL);
 865         while (spte) {
 866                 int _young;
 867                 u64 _spte = *spte;
 868                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 869                 _young = _spte & PT_ACCESSED_MASK;
 870                 if (_young) {
 871                         young = 1;
 872                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 873                 }
 874                 spte = rmap_next(kvm, rmapp, spte);
 875         }
 876         return young;
 877 }
 878
 879 #define RMAP_RECYCLE_THRESHOLD 1000
 880
 881 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 882 {
 883         unsigned long *rmapp;
 884         struct kvm_mmu_page *sp;
 885
 886         sp = page_header(__pa(spte));
 887
 888         gfn = unalias_gfn(vcpu->kvm, gfn);
 889         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 890
 891         kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
 892         kvm_flush_remote_tlbs(vcpu->kvm);
 893 }
 894
 895 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 896 {
 897         return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
 898 }
 899
 900 #ifdef MMU_DEBUG
 901 static int is_empty_shadow_page(u64 *spt)
 902 {
 903         u64 *pos;
 904         u64 *end;
 905
 906         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 907                 if (is_shadow_present_pte(*pos)) {
 908                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 909                                pos, *pos);
 910                         return 0;
 911                 }
 912         return 1;
 913 }
 914 #endif
 915
 916 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 917 {
 918         ASSERT(is_empty_shadow_page(sp->spt));
 919         hlist_del(&sp->hash_link);
 920         list_del(&sp->link);
 921         __free_page(virt_to_page(sp->spt));
 922         if (!sp->role.direct)
 923                 __free_page(virt_to_page(sp->gfns));
 924         kmem_cache_free(mmu_page_header_cache, sp);
 925         ++kvm->arch.n_free_mmu_pages;
 926 }
 927
 928 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 929 {
 930         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 931 }
 932
 933 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 934                                                u64 *parent_pte, int direct)
 935 {
 936         struct kvm_mmu_page *sp;
 937
 938         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 939         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 940         if (!direct)
 941                 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache,
 942                                                   PAGE_SIZE);
 943         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 944         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 945         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 946         sp->multimapped = 0;
 947         sp->parent_pte = parent_pte;
 948         --vcpu->kvm->arch.n_free_mmu_pages;
 949         return sp;
 950 }
 951
 952 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 953                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 954 {
 955         struct kvm_pte_chain *pte_chain;
 956         struct hlist_node *node;
 957         int i;
 958
 959         if (!parent_pte)
 960                 return;
 961         if (!sp->multimapped) {
 962                 u64 *old = sp->parent_pte;
 963
 964                 if (!old) {
 965                         sp->parent_pte = parent_pte;
 966                         return;
 967                 }
 968                 sp->multimapped = 1;
 969                 pte_chain = mmu_alloc_pte_chain(vcpu);
 970                 INIT_HLIST_HEAD(&sp->parent_ptes);
 971                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 972                 pte_chain->parent_ptes[0] = old;
 973         }
 974         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 975                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 976                         continue;
 977                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 978                         if (!pte_chain->parent_ptes[i]) {
 979                                 pte_chain->parent_ptes[i] = parent_pte;
 980                                 return;
 981                         }
 982         }
 983         pte_chain = mmu_alloc_pte_chain(vcpu);
 984         BUG_ON(!pte_chain);
 985         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 986         pte_chain->parent_ptes[0] = parent_pte;
 987 }
 988
 989 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 990                                        u64 *parent_pte)
 991 {
 992         struct kvm_pte_chain *pte_chain;
 993         struct hlist_node *node;
 994         int i;
 995
 996         if (!sp->multimapped) {
 997                 BUG_ON(sp->parent_pte != parent_pte);
 998                 sp->parent_pte = NULL;
 999                 return;
1000         }
1001         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1002                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1003                         if (!pte_chain->parent_ptes[i])
1004                                 break;
1005                         if (pte_chain->parent_ptes[i] != parent_pte)
1006                                 continue;
1007                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
1008                                 && pte_chain->parent_ptes[i + 1]) {
1009                                 pte_chain->parent_ptes[i]
1010                                         = pte_chain->parent_ptes[i + 1];
1011                                 ++i;
1012                         }
1013                         pte_chain->parent_ptes[i] = NULL;
1014                         if (i == 0) {
1015                                 hlist_del(&pte_chain->link);
1016                                 mmu_free_pte_chain(pte_chain);
1017                                 if (hlist_empty(&sp->parent_ptes)) {
1018                                         sp->multimapped = 0;
1019                                         sp->parent_pte = NULL;
1020                                 }
1021                         }
1022                         return;
1023                 }
1024         BUG();
1025 }
1026
1027
1028 static void mmu_parent_walk(struct kvm_mmu_page *sp, mmu_parent_walk_fn fn)
1029 {
1030         struct kvm_pte_chain *pte_chain;
1031         struct hlist_node *node;
1032         struct kvm_mmu_page *parent_sp;
1033         int i;
1034
1035         if (!sp->multimapped && sp->parent_pte) {
1036                 parent_sp = page_header(__pa(sp->parent_pte));
1037                 fn(parent_sp);
1038                 mmu_parent_walk(parent_sp, fn);
1039                 return;
1040         }
1041         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1042                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1043                         if (!pte_chain->parent_ptes[i])
1044                                 break;
1045                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
1046                         fn(parent_sp);
1047                         mmu_parent_walk(parent_sp, fn);
1048                 }
1049 }
1050
1051 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1052 {
1053         unsigned int index;
1054         struct kvm_mmu_page *sp = page_header(__pa(spte));
1055
1056         index = spte - sp->spt;
1057         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1058                 sp->unsync_children++;
1059         WARN_ON(!sp->unsync_children);
1060 }
1061
1062 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1063 {
1064         struct kvm_pte_chain *pte_chain;
1065         struct hlist_node *node;
1066         int i;
1067
1068         if (!sp->parent_pte)
1069                 return;
1070
1071         if (!sp->multimapped) {
1072                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1073                 return;
1074         }
1075
1076         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1077                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1078                         if (!pte_chain->parent_ptes[i])
1079                                 break;
1080                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1081                 }
1082 }
1083
1084 static int unsync_walk_fn(struct kvm_mmu_page *sp)
1085 {
1086         kvm_mmu_update_parents_unsync(sp);
1087         return 1;
1088 }
1089
1090 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1091 {
1092         mmu_parent_walk(sp, unsync_walk_fn);
1093         kvm_mmu_update_parents_unsync(sp);
1094 }
1095
1096 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1097                                     struct kvm_mmu_page *sp)
1098 {
1099         int i;
1100
1101         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1102                 sp->spt[i] = shadow_trap_nonpresent_pte;
1103 }
1104
1105 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1106                                struct kvm_mmu_page *sp)
1107 {
1108         return 1;
1109 }
1110
1111 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1112 {
1113 }
1114
1115 #define KVM_PAGE_ARRAY_NR 16
1116
1117 struct kvm_mmu_pages {
1118         struct mmu_page_and_offset {
1119                 struct kvm_mmu_page *sp;
1120                 unsigned int idx;
1121         } page[KVM_PAGE_ARRAY_NR];
1122         unsigned int nr;
1123 };
1124
1125 #define for_each_unsync_children(bitmap, idx)           \
1126         for (idx = find_first_bit(bitmap, 512);         \
1127              idx < 512;                                 \
1128              idx = find_next_bit(bitmap, 512, idx+1))
1129
1130 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1131                          int idx)
1132 {
1133         int i;
1134
1135         if (sp->unsync)
1136                 for (i=0; i < pvec->nr; i++)
1137                         if (pvec->page[i].sp == sp)
1138                                 return 0;
1139
1140         pvec->page[pvec->nr].sp = sp;
1141         pvec->page[pvec->nr].idx = idx;
1142         pvec->nr++;
1143         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1144 }
1145
1146 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1147                            struct kvm_mmu_pages *pvec)
1148 {
1149         int i, ret, nr_unsync_leaf = 0;
1150
1151         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1152                 u64 ent = sp->spt[i];
1153
1154                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1155                         struct kvm_mmu_page *child;
1156                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1157
1158                         if (child->unsync_children) {
1159                                 if (mmu_pages_add(pvec, child, i))
1160                                         return -ENOSPC;
1161
1162                                 ret = __mmu_unsync_walk(child, pvec);
1163                                 if (!ret)
1164                                         __clear_bit(i, sp->unsync_child_bitmap);
1165                                 else if (ret > 0)
1166                                         nr_unsync_leaf += ret;
1167                                 else
1168                                         return ret;
1169                         }
1170
1171                         if (child->unsync) {
1172                                 nr_unsync_leaf++;
1173                                 if (mmu_pages_add(pvec, child, i))
1174                                         return -ENOSPC;
1175                         }
1176                 }
1177         }
1178
1179         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1180                 sp->unsync_children = 0;
1181
1182         return nr_unsync_leaf;
1183 }
1184
1185 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1186                            struct kvm_mmu_pages *pvec)
1187 {
1188         if (!sp->unsync_children)
1189                 return 0;
1190
1191         mmu_pages_add(pvec, sp, 0);
1192         return __mmu_unsync_walk(sp, pvec);
1193 }
1194
1195 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1196 {
1197         WARN_ON(!sp->unsync);
1198         trace_kvm_mmu_sync_page(sp);
1199         sp->unsync = 0;
1200         --kvm->stat.mmu_unsync;
1201 }
1202
1203 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1204                                     struct list_head *invalid_list);
1205 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1206                                     struct list_head *invalid_list);
1207
1208 #define for_each_gfn_sp(kvm, sp, gfn, pos)                              \
1209   hlist_for_each_entry(sp, pos,                                         \
1210    &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link)   \
1211         if ((sp)->gfn != (gfn)) {} else
1212
1213 #define for_each_gfn_indirect_valid_sp(kvm, sp, gfn, pos)               \
1214   hlist_for_each_entry(sp, pos,                                         \
1215    &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link)   \
1216                 if ((sp)->gfn != (gfn) || (sp)->role.direct ||          \
1217                         (sp)->role.invalid) {} else
1218
1219 /* @sp->gfn should be write-protected at the call site */
1220 static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1221                            struct list_head *invalid_list, bool clear_unsync)
1222 {
1223         if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1224                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1225                 return 1;
1226         }
1227
1228         if (clear_unsync)
1229                 kvm_unlink_unsync_page(vcpu->kvm, sp);
1230
1231         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1232                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1233                 return 1;
1234         }
1235
1236         kvm_mmu_flush_tlb(vcpu);
1237         return 0;
1238 }
1239
1240 static void mmu_convert_notrap(struct kvm_mmu_page *sp);
1241 static int kvm_sync_page_transient(struct kvm_vcpu *vcpu,
1242                                    struct kvm_mmu_page *sp)
1243 {
1244         LIST_HEAD(invalid_list);
1245         int ret;
1246
1247         ret = __kvm_sync_page(vcpu, sp, &invalid_list, false);
1248         if (!ret)
1249                 mmu_convert_notrap(sp);
1250         else
1251                 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1252
1253         return ret;
1254 }
1255
1256 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1257                          struct list_head *invalid_list)
1258 {
1259         return __kvm_sync_page(vcpu, sp, invalid_list, true);
1260 }
1261
1262 /* @gfn should be write-protected at the call site */
1263 static void kvm_sync_pages(struct kvm_vcpu *vcpu,  gfn_t gfn)
1264 {
1265         struct kvm_mmu_page *s;
1266         struct hlist_node *node;
1267         LIST_HEAD(invalid_list);
1268         bool flush = false;
1269
1270         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1271                 if (!s->unsync)
1272                         continue;
1273
1274                 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1275                 if ((s->role.cr4_pae != !!is_pae(vcpu)) ||
1276                         (vcpu->arch.mmu.sync_page(vcpu, s))) {
1277                         kvm_mmu_prepare_zap_page(vcpu->kvm, s, &invalid_list);
1278                         continue;
1279                 }
1280                 kvm_unlink_unsync_page(vcpu->kvm, s);
1281                 flush = true;
1282         }
1283
1284         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1285         if (flush)
1286                 kvm_mmu_flush_tlb(vcpu);
1287 }
1288
1289 struct mmu_page_path {
1290         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1291         unsigned int idx[PT64_ROOT_LEVEL-1];
1292 };
1293
1294 #define for_each_sp(pvec, sp, parents, i)                       \
1295                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1296                         sp = pvec.page[i].sp;                   \
1297                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1298                         i = mmu_pages_next(&pvec, &parents, i))
1299
1300 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1301                           struct mmu_page_path *parents,
1302                           int i)
1303 {
1304         int n;
1305
1306         for (n = i+1; n < pvec->nr; n++) {
1307                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1308
1309                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1310                         parents->idx[0] = pvec->page[n].idx;
1311                         return n;
1312                 }
1313
1314                 parents->parent[sp->role.level-2] = sp;
1315                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1316         }
1317
1318         return n;
1319 }
1320
1321 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1322 {
1323         struct kvm_mmu_page *sp;
1324         unsigned int level = 0;
1325
1326         do {
1327                 unsigned int idx = parents->idx[level];
1328
1329                 sp = parents->parent[level];
1330                 if (!sp)
1331                         return;
1332
1333                 --sp->unsync_children;
1334                 WARN_ON((int)sp->unsync_children < 0);
1335                 __clear_bit(idx, sp->unsync_child_bitmap);
1336                 level++;
1337         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1338 }
1339
1340 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1341                                struct mmu_page_path *parents,
1342                                struct kvm_mmu_pages *pvec)
1343 {
1344         parents->parent[parent->role.level-1] = NULL;
1345         pvec->nr = 0;
1346 }
1347
1348 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1349                               struct kvm_mmu_page *parent)
1350 {
1351         int i;
1352         struct kvm_mmu_page *sp;
1353         struct mmu_page_path parents;
1354         struct kvm_mmu_pages pages;
1355         LIST_HEAD(invalid_list);
1356
1357         kvm_mmu_pages_init(parent, &parents, &pages);
1358         while (mmu_unsync_walk(parent, &pages)) {
1359                 int protected = 0;
1360
1361                 for_each_sp(pages, sp, parents, i)
1362                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1363
1364                 if (protected)
1365                         kvm_flush_remote_tlbs(vcpu->kvm);
1366
1367                 for_each_sp(pages, sp, parents, i) {
1368                         kvm_sync_page(vcpu, sp, &invalid_list);
1369                         mmu_pages_clear_parents(&parents);
1370                 }
1371                 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1372                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1373                 kvm_mmu_pages_init(parent, &parents, &pages);
1374         }
1375 }
1376
1377 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1378                                              gfn_t gfn,
1379                                              gva_t gaddr,
1380                                              unsigned level,
1381                                              int direct,
1382                                              unsigned access,
1383                                              u64 *parent_pte)
1384 {
1385         union kvm_mmu_page_role role;
1386         unsigned quadrant;
1387         struct kvm_mmu_page *sp;
1388         struct hlist_node *node;
1389         bool need_sync = false;
1390
1391         role = vcpu->arch.mmu.base_role;
1392         role.level = level;
1393         role.direct = direct;
1394         if (role.direct)
1395                 role.cr4_pae = 0;
1396         role.access = access;
1397         if (!tdp_enabled && vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1398                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1399                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1400                 role.quadrant = quadrant;
1401         }
1402         for_each_gfn_sp(vcpu->kvm, sp, gfn, node) {
1403                 if (!need_sync && sp->unsync)
1404                         need_sync = true;
1405
1406                 if (sp->role.word != role.word)
1407                         continue;
1408
1409                 if (sp->unsync && kvm_sync_page_transient(vcpu, sp))
1410                         break;
1411
1412                 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1413                 if (sp->unsync_children) {
1414                         set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1415                         kvm_mmu_mark_parents_unsync(sp);
1416                 } else if (sp->unsync)
1417                         kvm_mmu_mark_parents_unsync(sp);
1418
1419                 trace_kvm_mmu_get_page(sp, false);
1420                 return sp;
1421         }
1422         ++vcpu->kvm->stat.mmu_cache_miss;
1423         sp = kvm_mmu_alloc_page(vcpu, parent_pte, direct);
1424         if (!sp)
1425                 return sp;
1426         sp->gfn = gfn;
1427         sp->role = role;
1428         hlist_add_head(&sp->hash_link,
1429                 &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
1430         if (!direct) {
1431                 if (rmap_write_protect(vcpu->kvm, gfn))
1432                         kvm_flush_remote_tlbs(vcpu->kvm);
1433                 if (level > PT_PAGE_TABLE_LEVEL && need_sync)
1434                         kvm_sync_pages(vcpu, gfn);
1435
1436                 account_shadowed(vcpu->kvm, gfn);
1437         }
1438         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1439                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1440         else
1441                 nonpaging_prefetch_page(vcpu, sp);
1442         trace_kvm_mmu_get_page(sp, true);
1443         return sp;
1444 }
1445
1446 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1447                              struct kvm_vcpu *vcpu, u64 addr)
1448 {
1449         iterator->addr = addr;
1450         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1451         iterator->level = vcpu->arch.mmu.shadow_root_level;
1452         if (iterator->level == PT32E_ROOT_LEVEL) {
1453                 iterator->shadow_addr
1454                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1455                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1456                 --iterator->level;
1457                 if (!iterator->shadow_addr)
1458                         iterator->level = 0;
1459         }
1460 }
1461
1462 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1463 {
1464         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1465                 return false;
1466
1467         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1468                 if (is_large_pte(*iterator->sptep))
1469                         return false;
1470
1471         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1472         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1473         return true;
1474 }
1475
1476 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1477 {
1478         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1479         --iterator->level;
1480 }
1481
1482 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1483                                          struct kvm_mmu_page *sp)
1484 {
1485         unsigned i;
1486         u64 *pt;
1487         u64 ent;
1488
1489         pt = sp->spt;
1490
1491         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1492                 ent = pt[i];
1493
1494                 if (is_shadow_present_pte(ent)) {
1495                         if (!is_last_spte(ent, sp->role.level)) {
1496                                 ent &= PT64_BASE_ADDR_MASK;
1497                                 mmu_page_remove_parent_pte(page_header(ent),
1498                                                            &pt[i]);
1499                         } else {
1500                                 if (is_large_pte(ent))
1501                                         --kvm->stat.lpages;
1502                                 rmap_remove(kvm, &pt[i]);
1503                         }
1504                 }
1505                 pt[i] = shadow_trap_nonpresent_pte;
1506         }
1507 }
1508
1509 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1510 {
1511         mmu_page_remove_parent_pte(sp, parent_pte);
1512 }
1513
1514 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1515 {
1516         int i;
1517         struct kvm_vcpu *vcpu;
1518
1519         kvm_for_each_vcpu(i, vcpu, kvm)
1520                 vcpu->arch.last_pte_updated = NULL;
1521 }
1522
1523 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1524 {
1525         u64 *parent_pte;
1526
1527         while (sp->multimapped || sp->parent_pte) {
1528                 if (!sp->multimapped)
1529                         parent_pte = sp->parent_pte;
1530                 else {
1531                         struct kvm_pte_chain *chain;
1532
1533                         chain = container_of(sp->parent_ptes.first,
1534                                              struct kvm_pte_chain, link);
1535                         parent_pte = chain->parent_ptes[0];
1536                 }
1537                 BUG_ON(!parent_pte);
1538                 kvm_mmu_put_page(sp, parent_pte);
1539                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1540         }
1541 }
1542
1543 static int mmu_zap_unsync_children(struct kvm *kvm,
1544                                    struct kvm_mmu_page *parent,
1545                                    struct list_head *invalid_list)
1546 {
1547         int i, zapped = 0;
1548         struct mmu_page_path parents;
1549         struct kvm_mmu_pages pages;
1550
1551         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1552                 return 0;
1553
1554         kvm_mmu_pages_init(parent, &parents, &pages);
1555         while (mmu_unsync_walk(parent, &pages)) {
1556                 struct kvm_mmu_page *sp;
1557
1558                 for_each_sp(pages, sp, parents, i) {
1559                         kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
1560                         mmu_pages_clear_parents(&parents);
1561                         zapped++;
1562                 }
1563                 kvm_mmu_pages_init(parent, &parents, &pages);
1564         }
1565
1566         return zapped;
1567 }
1568
1569 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1570                                     struct list_head *invalid_list)
1571 {
1572         int ret;
1573
1574         trace_kvm_mmu_prepare_zap_page(sp);
1575         ++kvm->stat.mmu_shadow_zapped;
1576         ret = mmu_zap_unsync_children(kvm, sp, invalid_list);
1577         kvm_mmu_page_unlink_children(kvm, sp);
1578         kvm_mmu_unlink_parents(kvm, sp);
1579         if (!sp->role.invalid && !sp->role.direct)
1580                 unaccount_shadowed(kvm, sp->gfn);
1581         if (sp->unsync)
1582                 kvm_unlink_unsync_page(kvm, sp);
1583         if (!sp->root_count) {
1584                 /* Count self */
1585                 ret++;
1586                 list_move(&sp->link, invalid_list);
1587         } else {
1588                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1589                 kvm_reload_remote_mmus(kvm);
1590         }
1591
1592         sp->role.invalid = 1;
1593         kvm_mmu_reset_last_pte_updated(kvm);
1594         return ret;
1595 }
1596
1597 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1598                                     struct list_head *invalid_list)
1599 {
1600         struct kvm_mmu_page *sp;
1601
1602         if (list_empty(invalid_list))
1603                 return;
1604
1605         kvm_flush_remote_tlbs(kvm);
1606
1607         do {
1608                 sp = list_first_entry(invalid_list, struct kvm_mmu_page, link);
1609                 WARN_ON(!sp->role.invalid || sp->root_count);
1610                 kvm_mmu_free_page(kvm, sp);
1611         } while (!list_empty(invalid_list));
1612
1613 }
1614
1615 /*
1616  * Changing the number of mmu pages allocated to the vm
1617  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1618  */
1619 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1620 {
1621         int used_pages;
1622         LIST_HEAD(invalid_list);
1623
1624         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1625         used_pages = max(0, used_pages);
1626
1627         /*
1628          * If we set the number of mmu pages to be smaller be than the
1629          * number of actived pages , we must to free some mmu pages before we
1630          * change the value
1631          */
1632
1633         if (used_pages > kvm_nr_mmu_pages) {
1634                 while (used_pages > kvm_nr_mmu_pages &&
1635                         !list_empty(&kvm->arch.active_mmu_pages)) {
1636                         struct kvm_mmu_page *page;
1637
1638                         page = container_of(kvm->arch.active_mmu_pages.prev,
1639                                             struct kvm_mmu_page, link);
1640                         used_pages -= kvm_mmu_prepare_zap_page(kvm, page,
1641                                                                &invalid_list);
1642                 }
1643                 kvm_mmu_commit_zap_page(kvm, &invalid_list);
1644                 kvm_nr_mmu_pages = used_pages;
1645                 kvm->arch.n_free_mmu_pages = 0;
1646         }
1647         else
1648                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1649                                          - kvm->arch.n_alloc_mmu_pages;
1650
1651         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1652 }
1653
1654 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1655 {
1656         struct kvm_mmu_page *sp;
1657         struct hlist_node *node;
1658         LIST_HEAD(invalid_list);
1659         int r;
1660
1661         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1662         r = 0;
1663
1664         for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1665                 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1666                          sp->role.word);
1667                 r = 1;
1668                 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1669         }
1670         kvm_mmu_commit_zap_page(kvm, &invalid_list);
1671         return r;
1672 }
1673
1674 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1675 {
1676         struct kvm_mmu_page *sp;
1677         struct hlist_node *node;
1678         LIST_HEAD(invalid_list);
1679
1680         for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1681                 pgprintk("%s: zap %lx %x\n",
1682                          __func__, gfn, sp->role.word);
1683                 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1684         }
1685         kvm_mmu_commit_zap_page(kvm, &invalid_list);
1686 }
1687
1688 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1689 {
1690         int slot = memslot_id(kvm, gfn);
1691         struct kvm_mmu_page *sp = page_header(__pa(pte));
1692
1693         __set_bit(slot, sp->slot_bitmap);
1694 }
1695
1696 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1697 {
1698         int i;
1699         u64 *pt = sp->spt;
1700
1701         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1702                 return;
1703
1704         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1705                 if (pt[i] == shadow_notrap_nonpresent_pte)
1706                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1707         }
1708 }
1709
1710 /*
1711  * The function is based on mtrr_type_lookup() in
1712  * arch/x86/kernel/cpu/mtrr/generic.c
1713  */
1714 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1715                          u64 start, u64 end)
1716 {
1717         int i;
1718         u64 base, mask;
1719         u8 prev_match, curr_match;
1720         int num_var_ranges = KVM_NR_VAR_MTRR;
1721
1722         if (!mtrr_state->enabled)
1723                 return 0xFF;
1724
1725         /* Make end inclusive end, instead of exclusive */
1726         end--;
1727
1728         /* Look in fixed ranges. Just return the type as per start */
1729         if (mtrr_state->have_fixed && (start < 0x100000)) {
1730                 int idx;
1731
1732                 if (start < 0x80000) {
1733                         idx = 0;
1734                         idx += (start >> 16);
1735                         return mtrr_state->fixed_ranges[idx];
1736                 } else if (start < 0xC0000) {
1737                         idx = 1 * 8;
1738                         idx += ((start - 0x80000) >> 14);
1739                         return mtrr_state->fixed_ranges[idx];
1740                 } else if (start < 0x1000000) {
1741                         idx = 3 * 8;
1742                         idx += ((start - 0xC0000) >> 12);
1743                         return mtrr_state->fixed_ranges[idx];
1744                 }
1745         }
1746
1747         /*
1748          * Look in variable ranges
1749          * Look of multiple ranges matching this address and pick type
1750          * as per MTRR precedence
1751          */
1752         if (!(mtrr_state->enabled & 2))
1753                 return mtrr_state->def_type;
1754
1755         prev_match = 0xFF;
1756         for (i = 0; i < num_var_ranges; ++i) {
1757                 unsigned short start_state, end_state;
1758
1759                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1760                         continue;
1761
1762                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1763                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1764                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1765                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1766
1767                 start_state = ((start & mask) == (base & mask));
1768                 end_state = ((end & mask) == (base & mask));
1769                 if (start_state != end_state)
1770                         return 0xFE;
1771
1772                 if ((start & mask) != (base & mask))
1773                         continue;
1774
1775                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1776                 if (prev_match == 0xFF) {
1777                         prev_match = curr_match;
1778                         continue;
1779                 }
1780
1781                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1782                     curr_match == MTRR_TYPE_UNCACHABLE)
1783                         return MTRR_TYPE_UNCACHABLE;
1784
1785                 if ((prev_match == MTRR_TYPE_WRBACK &&
1786                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1787                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1788                      curr_match == MTRR_TYPE_WRBACK)) {
1789                         prev_match = MTRR_TYPE_WRTHROUGH;
1790                         curr_match = MTRR_TYPE_WRTHROUGH;
1791                 }
1792
1793                 if (prev_match != curr_match)
1794                         return MTRR_TYPE_UNCACHABLE;
1795         }
1796
1797         if (prev_match != 0xFF)
1798                 return prev_match;
1799
1800         return mtrr_state->def_type;
1801 }
1802
1803 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1804 {
1805         u8 mtrr;
1806
1807         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1808                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1809         if (mtrr == 0xfe || mtrr == 0xff)
1810                 mtrr = MTRR_TYPE_WRBACK;
1811         return mtrr;
1812 }
1813 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1814
1815 static void __kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1816 {
1817         trace_kvm_mmu_unsync_page(sp);
1818         ++vcpu->kvm->stat.mmu_unsync;
1819         sp->unsync = 1;
1820
1821         kvm_mmu_mark_parents_unsync(sp);
1822         mmu_convert_notrap(sp);
1823 }
1824
1825 static void kvm_unsync_pages(struct kvm_vcpu *vcpu,  gfn_t gfn)
1826 {
1827         struct kvm_mmu_page *s;
1828         struct hlist_node *node;
1829
1830         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1831                 if (s->unsync)
1832                         continue;
1833                 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1834                 __kvm_unsync_page(vcpu, s);
1835         }
1836 }
1837
1838 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1839                                   bool can_unsync)
1840 {
1841         struct kvm_mmu_page *s;
1842         struct hlist_node *node;
1843         bool need_unsync = false;
1844
1845         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1846                 if (s->role.level != PT_PAGE_TABLE_LEVEL)
1847                         return 1;
1848
1849                 if (!need_unsync && !s->unsync) {
1850                         if (!can_unsync || !oos_shadow)
1851                                 return 1;
1852                         need_unsync = true;
1853                 }
1854         }
1855         if (need_unsync)
1856                 kvm_unsync_pages(vcpu, gfn);
1857         return 0;
1858 }
1859
1860 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1861                     unsigned pte_access, int user_fault,
1862                     int write_fault, int dirty, int level,
1863                     gfn_t gfn, pfn_t pfn, bool speculative,
1864                     bool can_unsync, bool reset_host_protection)
1865 {
1866         u64 spte;
1867         int ret = 0;
1868
1869         /*
1870          * We don't set the accessed bit, since we sometimes want to see
1871          * whether the guest actually used the pte (in order to detect
1872          * demand paging).
1873          */
1874         spte = shadow_base_present_pte | shadow_dirty_mask;
1875         if (!speculative)
1876                 spte |= shadow_accessed_mask;
1877         if (!dirty)
1878                 pte_access &= ~ACC_WRITE_MASK;
1879         if (pte_access & ACC_EXEC_MASK)
1880                 spte |= shadow_x_mask;
1881         else
1882                 spte |= shadow_nx_mask;
1883         if (pte_access & ACC_USER_MASK)
1884                 spte |= shadow_user_mask;
1885         if (level > PT_PAGE_TABLE_LEVEL)
1886                 spte |= PT_PAGE_SIZE_MASK;
1887         if (tdp_enabled)
1888                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1889                         kvm_is_mmio_pfn(pfn));
1890
1891         if (reset_host_protection)
1892                 spte |= SPTE_HOST_WRITEABLE;
1893
1894         spte |= (u64)pfn << PAGE_SHIFT;
1895
1896         if ((pte_access & ACC_WRITE_MASK)
1897             || (!tdp_enabled && write_fault && !is_write_protection(vcpu)
1898                 && !user_fault)) {
1899
1900                 if (level > PT_PAGE_TABLE_LEVEL &&
1901                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1902                         ret = 1;
1903                         rmap_remove(vcpu->kvm, sptep);
1904                         spte = shadow_trap_nonpresent_pte;
1905                         goto set_pte;
1906                 }
1907
1908                 spte |= PT_WRITABLE_MASK;
1909
1910                 if (!tdp_enabled && !(pte_access & ACC_WRITE_MASK))
1911                         spte &= ~PT_USER_MASK;
1912
1913                 /*
1914                  * Optimization: for pte sync, if spte was writable the hash
1915                  * lookup is unnecessary (and expensive). Write protection
1916                  * is responsibility of mmu_get_page / kvm_sync_page.
1917                  * Same reasoning can be applied to dirty page accounting.
1918                  */
1919                 if (!can_unsync && is_writable_pte(*sptep))
1920                         goto set_pte;
1921
1922                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1923                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1924                                  __func__, gfn);
1925                         ret = 1;
1926                         pte_access &= ~ACC_WRITE_MASK;
1927                         if (is_writable_pte(spte))
1928                                 spte &= ~PT_WRITABLE_MASK;
1929                 }
1930         }
1931
1932         if (pte_access & ACC_WRITE_MASK)
1933                 mark_page_dirty(vcpu->kvm, gfn);
1934
1935 set_pte:
1936         __set_spte(sptep, spte);
1937         return ret;
1938 }
1939
1940 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1941                          unsigned pt_access, unsigned pte_access,
1942                          int user_fault, int write_fault, int dirty,
1943                          int *ptwrite, int level, gfn_t gfn,
1944                          pfn_t pfn, bool speculative,
1945                          bool reset_host_protection)
1946 {
1947         int was_rmapped = 0;
1948         int was_writable = is_writable_pte(*sptep);
1949         int rmap_count;
1950
1951         pgprintk("%s: spte %llx access %x write_fault %d"
1952                  " user_fault %d gfn %lx\n",
1953                  __func__, *sptep, pt_access,
1954                  write_fault, user_fault, gfn);
1955
1956         if (is_rmap_spte(*sptep)) {
1957                 /*
1958                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1959                  * the parent of the now unreachable PTE.
1960                  */
1961                 if (level > PT_PAGE_TABLE_LEVEL &&
1962                     !is_large_pte(*sptep)) {
1963                         struct kvm_mmu_page *child;
1964                         u64 pte = *sptep;
1965
1966                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1967                         mmu_page_remove_parent_pte(child, sptep);
1968                         __set_spte(sptep, shadow_trap_nonpresent_pte);
1969                         kvm_flush_remote_tlbs(vcpu->kvm);
1970                 } else if (pfn != spte_to_pfn(*sptep)) {
1971                         pgprintk("hfn old %lx new %lx\n",
1972                                  spte_to_pfn(*sptep), pfn);
1973                         rmap_remove(vcpu->kvm, sptep);
1974                         __set_spte(sptep, shadow_trap_nonpresent_pte);
1975                         kvm_flush_remote_tlbs(vcpu->kvm);
1976                 } else
1977                         was_rmapped = 1;
1978         }
1979
1980         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1981                       dirty, level, gfn, pfn, speculative, true,
1982                       reset_host_protection)) {
1983                 if (write_fault)
1984                         *ptwrite = 1;
1985                 kvm_mmu_flush_tlb(vcpu);
1986         }
1987
1988         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1989         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1990                  is_large_pte(*sptep)? "2MB" : "4kB",
1991                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1992                  *sptep, sptep);
1993         if (!was_rmapped && is_large_pte(*sptep))
1994                 ++vcpu->kvm->stat.lpages;
1995
1996         page_header_update_slot(vcpu->kvm, sptep, gfn);
1997         if (!was_rmapped) {
1998                 rmap_count = rmap_add(vcpu, sptep, gfn);
1999                 kvm_release_pfn_clean(pfn);
2000                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
2001                         rmap_recycle(vcpu, sptep, gfn);
2002         } else {
2003                 if (was_writable)
2004                         kvm_release_pfn_dirty(pfn);
2005                 else
2006                         kvm_release_pfn_clean(pfn);
2007         }
2008         if (speculative) {
2009                 vcpu->arch.last_pte_updated = sptep;
2010                 vcpu->arch.last_pte_gfn = gfn;
2011         }
2012 }
2013
2014 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
2015 {
2016 }
2017
2018 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
2019                         int level, gfn_t gfn, pfn_t pfn)
2020 {
2021         struct kvm_shadow_walk_iterator iterator;
2022         struct kvm_mmu_page *sp;
2023         int pt_write = 0;
2024         gfn_t pseudo_gfn;
2025
2026         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
2027                 if (iterator.level == level) {
2028                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
2029                                      0, write, 1, &pt_write,
2030                                      level, gfn, pfn, false, true);
2031                         ++vcpu->stat.pf_fixed;
2032                         break;
2033                 }
2034
2035                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
2036                         u64 base_addr = iterator.addr;
2037
2038                         base_addr &= PT64_LVL_ADDR_MASK(iterator.level);
2039                         pseudo_gfn = base_addr >> PAGE_SHIFT;
2040                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
2041                                               iterator.level - 1,
2042                                               1, ACC_ALL, iterator.sptep);
2043                         if (!sp) {
2044                                 pgprintk("nonpaging_map: ENOMEM\n");
2045                                 kvm_release_pfn_clean(pfn);
2046                                 return -ENOMEM;
2047                         }
2048
2049                         __set_spte(iterator.sptep,
2050                                    __pa(sp->spt)
2051                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
2052                                    | shadow_user_mask | shadow_x_mask);
2053                 }
2054         }
2055         return pt_write;
2056 }
2057
2058 static void kvm_send_hwpoison_signal(struct kvm *kvm, gfn_t gfn)
2059 {
2060         char buf[1];
2061         void __user *hva;
2062         int r;
2063
2064         /* Touch the page, so send SIGBUS */
2065         hva = (void __user *)gfn_to_hva(kvm, gfn);
2066         r = copy_from_user(buf, hva, 1);
2067 }
2068
2069 static int kvm_handle_bad_page(struct kvm *kvm, gfn_t gfn, pfn_t pfn)
2070 {
2071         kvm_release_pfn_clean(pfn);
2072         if (is_hwpoison_pfn(pfn)) {
2073                 kvm_send_hwpoison_signal(kvm, gfn);
2074                 return 0;
2075         }
2076         return 1;
2077 }
2078
2079 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
2080 {
2081         int r;
2082         int level;
2083         pfn_t pfn;
2084         unsigned long mmu_seq;
2085
2086         level = mapping_level(vcpu, gfn);
2087
2088         /*
2089          * This path builds a PAE pagetable - so we can map 2mb pages at
2090          * maximum. Therefore check if the level is larger than that.
2091          */
2092         if (level > PT_DIRECTORY_LEVEL)
2093                 level = PT_DIRECTORY_LEVEL;
2094
2095         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2096
2097         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2098         smp_rmb();
2099         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2100
2101         /* mmio */
2102         if (is_error_pfn(pfn))
2103                 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2104
2105         spin_lock(&vcpu->kvm->mmu_lock);
2106         if (mmu_notifier_retry(vcpu, mmu_seq))
2107                 goto out_unlock;
2108         kvm_mmu_free_some_pages(vcpu);
2109         r = __direct_map(vcpu, v, write, level, gfn, pfn);
2110         spin_unlock(&vcpu->kvm->mmu_lock);
2111
2112
2113         return r;
2114
2115 out_unlock:
2116         spin_unlock(&vcpu->kvm->mmu_lock);
2117         kvm_release_pfn_clean(pfn);
2118         return 0;
2119 }
2120
2121
2122 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2123 {
2124         int i;
2125         struct kvm_mmu_page *sp;
2126         LIST_HEAD(invalid_list);
2127
2128         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2129                 return;
2130         spin_lock(&vcpu->kvm->mmu_lock);
2131         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2132                 hpa_t root = vcpu->arch.mmu.root_hpa;
2133
2134                 sp = page_header(root);
2135                 --sp->root_count;
2136                 if (!sp->root_count && sp->role.invalid) {
2137                         kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
2138                         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2139                 }
2140                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2141                 spin_unlock(&vcpu->kvm->mmu_lock);
2142                 return;
2143         }
2144         for (i = 0; i < 4; ++i) {
2145                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2146
2147                 if (root) {
2148                         root &= PT64_BASE_ADDR_MASK;
2149                         sp = page_header(root);
2150                         --sp->root_count;
2151                         if (!sp->root_count && sp->role.invalid)
2152                                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2153                                                          &invalid_list);
2154                 }
2155                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2156         }
2157         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2158         spin_unlock(&vcpu->kvm->mmu_lock);
2159         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2160 }
2161
2162 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2163 {
2164         int ret = 0;
2165
2166         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2167                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2168                 ret = 1;
2169         }
2170
2171         return ret;
2172 }
2173
2174 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2175 {
2176         int i;
2177         gfn_t root_gfn;
2178         struct kvm_mmu_page *sp;
2179         int direct = 0;
2180         u64 pdptr;
2181
2182         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2183
2184         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2185                 hpa_t root = vcpu->arch.mmu.root_hpa;
2186
2187                 ASSERT(!VALID_PAGE(root));
2188                 if (mmu_check_root(vcpu, root_gfn))
2189                         return 1;
2190                 if (tdp_enabled) {
2191                         direct = 1;
2192                         root_gfn = 0;
2193                 }
2194                 spin_lock(&vcpu->kvm->mmu_lock);
2195                 kvm_mmu_free_some_pages(vcpu);
2196                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2197                                       PT64_ROOT_LEVEL, direct,
2198                                       ACC_ALL, NULL);
2199                 root = __pa(sp->spt);
2200                 ++sp->root_count;
2201                 spin_unlock(&vcpu->kvm->mmu_lock);
2202                 vcpu->arch.mmu.root_hpa = root;
2203                 return 0;
2204         }
2205         direct = !is_paging(vcpu);
2206         for (i = 0; i < 4; ++i) {
2207                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2208
2209                 ASSERT(!VALID_PAGE(root));
2210                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2211                         pdptr = kvm_pdptr_read(vcpu, i);
2212                         if (!is_present_gpte(pdptr)) {
2213                                 vcpu->arch.mmu.pae_root[i] = 0;
2214                                 continue;
2215                         }
2216                         root_gfn = pdptr >> PAGE_SHIFT;
2217                 } else if (vcpu->arch.mmu.root_level == 0)
2218                         root_gfn = 0;
2219                 if (mmu_check_root(vcpu, root_gfn))
2220                         return 1;
2221                 if (tdp_enabled) {
2222                         direct = 1;
2223                         root_gfn = i << 30;
2224                 }
2225                 spin_lock(&vcpu->kvm->mmu_lock);
2226                 kvm_mmu_free_some_pages(vcpu);
2227                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2228                                       PT32_ROOT_LEVEL, direct,
2229                                       ACC_ALL, NULL);
2230                 root = __pa(sp->spt);
2231                 ++sp->root_count;
2232                 spin_unlock(&vcpu->kvm->mmu_lock);
2233
2234                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2235         }
2236         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2237         return 0;
2238 }
2239
2240 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2241 {
2242         int i;
2243         struct kvm_mmu_page *sp;
2244
2245         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2246                 return;
2247         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2248                 hpa_t root = vcpu->arch.mmu.root_hpa;
2249                 sp = page_header(root);
2250                 mmu_sync_children(vcpu, sp);
2251                 return;
2252         }
2253         for (i = 0; i < 4; ++i) {
2254                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2255
2256                 if (root && VALID_PAGE(root)) {
2257                         root &= PT64_BASE_ADDR_MASK;
2258                         sp = page_header(root);
2259                         mmu_sync_children(vcpu, sp);
2260                 }
2261         }
2262 }
2263
2264 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2265 {
2266         spin_lock(&vcpu->kvm->mmu_lock);
2267         mmu_sync_roots(vcpu);
2268         spin_unlock(&vcpu->kvm->mmu_lock);
2269 }
2270
2271 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2272                                   u32 access, u32 *error)
2273 {
2274         if (error)
2275                 *error = 0;
2276         return vaddr;
2277 }
2278
2279 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2280                                 u32 error_code)
2281 {
2282         gfn_t gfn;
2283         int r;
2284
2285         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2286         r = mmu_topup_memory_caches(vcpu);
2287         if (r)
2288                 return r;
2289
2290         ASSERT(vcpu);
2291         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2292
2293         gfn = gva >> PAGE_SHIFT;
2294
2295         return nonpaging_map(vcpu, gva & PAGE_MASK,
2296                              error_code & PFERR_WRITE_MASK, gfn);
2297 }
2298
2299 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2300                                 u32 error_code)
2301 {
2302         pfn_t pfn;
2303         int r;
2304         int level;
2305         gfn_t gfn = gpa >> PAGE_SHIFT;
2306         unsigned long mmu_seq;
2307
2308         ASSERT(vcpu);
2309         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2310
2311         r = mmu_topup_memory_caches(vcpu);
2312         if (r)
2313                 return r;
2314
2315         level = mapping_level(vcpu, gfn);
2316
2317         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2318
2319         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2320         smp_rmb();
2321         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2322         if (is_error_pfn(pfn))
2323                 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2324         spin_lock(&vcpu->kvm->mmu_lock);
2325         if (mmu_notifier_retry(vcpu, mmu_seq))
2326                 goto out_unlock;
2327         kvm_mmu_free_some_pages(vcpu);
2328         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2329                          level, gfn, pfn);
2330         spin_unlock(&vcpu->kvm->mmu_lock);
2331
2332         return r;
2333
2334 out_unlock:
2335         spin_unlock(&vcpu->kvm->mmu_lock);
2336         kvm_release_pfn_clean(pfn);
2337         return 0;
2338 }
2339
2340 static void nonpaging_free(struct kvm_vcpu *vcpu)
2341 {
2342         mmu_free_roots(vcpu);
2343 }
2344
2345 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2346 {
2347         struct kvm_mmu *context = &vcpu->arch.mmu;
2348
2349         context->new_cr3 = nonpaging_new_cr3;
2350         context->page_fault = nonpaging_page_fault;
2351         context->gva_to_gpa = nonpaging_gva_to_gpa;
2352         context->free = nonpaging_free;
2353         context->prefetch_page = nonpaging_prefetch_page;
2354         context->sync_page = nonpaging_sync_page;
2355         context->invlpg = nonpaging_invlpg;
2356         context->root_level = 0;
2357         context->shadow_root_level = PT32E_ROOT_LEVEL;
2358         context->root_hpa = INVALID_PAGE;
2359         return 0;
2360 }
2361
2362 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2363 {
2364         ++vcpu->stat.tlb_flush;
2365         set_bit(KVM_REQ_TLB_FLUSH, &vcpu->requests);
2366 }
2367
2368 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2369 {
2370         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2371         mmu_free_roots(vcpu);
2372 }
2373
2374 static void inject_page_fault(struct kvm_vcpu *vcpu,
2375                               u64 addr,
2376                               u32 err_code)
2377 {
2378         kvm_inject_page_fault(vcpu, addr, err_code);
2379 }
2380
2381 static void paging_free(struct kvm_vcpu *vcpu)
2382 {
2383         nonpaging_free(vcpu);
2384 }
2385
2386 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2387 {
2388         int bit7;
2389
2390         bit7 = (gpte >> 7) & 1;
2391         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2392 }
2393
2394 #define PTTYPE 64
2395 #include "paging_tmpl.h"
2396 #undef PTTYPE
2397
2398 #define PTTYPE 32
2399 #include "paging_tmpl.h"
2400 #undef PTTYPE
2401
2402 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2403 {
2404         struct kvm_mmu *context = &vcpu->arch.mmu;
2405         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2406         u64 exb_bit_rsvd = 0;
2407
2408         if (!is_nx(vcpu))
2409                 exb_bit_rsvd = rsvd_bits(63, 63);
2410         switch (level) {
2411         case PT32_ROOT_LEVEL:
2412                 /* no rsvd bits for 2 level 4K page table entries */
2413                 context->rsvd_bits_mask[0][1] = 0;
2414                 context->rsvd_bits_mask[0][0] = 0;
2415                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2416
2417                 if (!is_pse(vcpu)) {
2418                         context->rsvd_bits_mask[1][1] = 0;
2419                         break;
2420                 }
2421
2422                 if (is_cpuid_PSE36())
2423                         /* 36bits PSE 4MB page */
2424                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2425                 else
2426                         /* 32 bits PSE 4MB page */
2427                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2428                 break;
2429         case PT32E_ROOT_LEVEL:
2430                 context->rsvd_bits_mask[0][2] =
2431                         rsvd_bits(maxphyaddr, 63) |
2432                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2433                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2434                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2435                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2436                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2437                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2438                         rsvd_bits(maxphyaddr, 62) |
2439                         rsvd_bits(13, 20);              /* large page */
2440                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2441                 break;
2442         case PT64_ROOT_LEVEL:
2443                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2444                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2445                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2446                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2447                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2448                         rsvd_bits(maxphyaddr, 51);
2449                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2450                         rsvd_bits(maxphyaddr, 51);
2451                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2452                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2453                         rsvd_bits(maxphyaddr, 51) |
2454                         rsvd_bits(13, 29);
2455                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2456                         rsvd_bits(maxphyaddr, 51) |
2457                         rsvd_bits(13, 20);              /* large page */
2458                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2459                 break;
2460         }
2461 }
2462
2463 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2464 {
2465         struct kvm_mmu *context = &vcpu->arch.mmu;
2466
2467         ASSERT(is_pae(vcpu));
2468         context->new_cr3 = paging_new_cr3;
2469         context->page_fault = paging64_page_fault;
2470         context->gva_to_gpa = paging64_gva_to_gpa;
2471         context->prefetch_page = paging64_prefetch_page;
2472         context->sync_page = paging64_sync_page;
2473         context->invlpg = paging64_invlpg;
2474         context->free = paging_free;
2475         context->root_level = level;
2476         context->shadow_root_level = level;
2477         context->root_hpa = INVALID_PAGE;
2478         return 0;
2479 }
2480
2481 static int paging64_init_context(struct kvm_vcpu *vcpu)
2482 {
2483         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2484         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2485 }
2486
2487 static int paging32_init_context(struct kvm_vcpu *vcpu)
2488 {
2489         struct kvm_mmu *context = &vcpu->arch.mmu;
2490
2491         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2492         context->new_cr3 = paging_new_cr3;
2493         context->page_fault = paging32_page_fault;
2494         context->gva_to_gpa = paging32_gva_to_gpa;
2495         context->free = paging_free;
2496         context->prefetch_page = paging32_prefetch_page;
2497         context->sync_page = paging32_sync_page;
2498         context->invlpg = paging32_invlpg;
2499         context->root_level = PT32_ROOT_LEVEL;
2500         context->shadow_root_level = PT32E_ROOT_LEVEL;
2501         context->root_hpa = INVALID_PAGE;
2502         return 0;
2503 }
2504
2505 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2506 {
2507         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2508         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2509 }
2510
2511 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2512 {
2513         struct kvm_mmu *context = &vcpu->arch.mmu;
2514
2515         context->new_cr3 = nonpaging_new_cr3;
2516         context->page_fault = tdp_page_fault;
2517         context->free = nonpaging_free;
2518         context->prefetch_page = nonpaging_prefetch_page;
2519         context->sync_page = nonpaging_sync_page;
2520         context->invlpg = nonpaging_invlpg;
2521         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2522         context->root_hpa = INVALID_PAGE;
2523
2524         if (!is_paging(vcpu)) {
2525                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2526                 context->root_level = 0;
2527         } else if (is_long_mode(vcpu)) {
2528                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2529                 context->gva_to_gpa = paging64_gva_to_gpa;
2530                 context->root_level = PT64_ROOT_LEVEL;
2531         } else if (is_pae(vcpu)) {
2532                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2533                 context->gva_to_gpa = paging64_gva_to_gpa;
2534                 context->root_level = PT32E_ROOT_LEVEL;
2535         } else {
2536                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2537                 context->gva_to_gpa = paging32_gva_to_gpa;
2538                 context->root_level = PT32_ROOT_LEVEL;
2539         }
2540
2541         return 0;
2542 }
2543
2544 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2545 {
2546         int r;
2547
2548         ASSERT(vcpu);
2549         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2550
2551         if (!is_paging(vcpu))
2552                 r = nonpaging_init_context(vcpu);
2553         else if (is_long_mode(vcpu))
2554                 r = paging64_init_context(vcpu);
2555         else if (is_pae(vcpu))
2556                 r = paging32E_init_context(vcpu);
2557         else
2558                 r = paging32_init_context(vcpu);
2559
2560         vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
2561         vcpu->arch.mmu.base_role.cr0_wp = is_write_protection(vcpu);
2562
2563         return r;
2564 }
2565
2566 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2567 {
2568         vcpu->arch.update_pte.pfn = bad_pfn;
2569
2570         if (tdp_enabled)
2571                 return init_kvm_tdp_mmu(vcpu);
2572         else
2573                 return init_kvm_softmmu(vcpu);
2574 }
2575
2576 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2577 {
2578         ASSERT(vcpu);
2579         if (VALID_PAGE(vcpu->arch.mmu.root_hpa))
2580                 /* mmu.free() should set root_hpa = INVALID_PAGE */
2581                 vcpu->arch.mmu.free(vcpu);
2582 }
2583
2584 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2585 {
2586         destroy_kvm_mmu(vcpu);
2587         return init_kvm_mmu(vcpu);
2588 }
2589 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2590
2591 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2592 {
2593         int r;
2594
2595         r = mmu_topup_memory_caches(vcpu);
2596         if (r)
2597                 goto out;
2598         r = mmu_alloc_roots(vcpu);
2599         spin_lock(&vcpu->kvm->mmu_lock);
2600         mmu_sync_roots(vcpu);
2601         spin_unlock(&vcpu->kvm->mmu_lock);
2602         if (r)
2603                 goto out;
2604         /* set_cr3() should ensure TLB has been flushed */
2605         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2606 out:
2607         return r;
2608 }
2609 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2610
2611 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2612 {
2613         mmu_free_roots(vcpu);
2614 }
2615
2616 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2617                                   struct kvm_mmu_page *sp,
2618                                   u64 *spte)
2619 {
2620         u64 pte;
2621         struct kvm_mmu_page *child;
2622
2623         pte = *spte;
2624         if (is_shadow_present_pte(pte)) {
2625                 if (is_last_spte(pte, sp->role.level))
2626                         rmap_remove(vcpu->kvm, spte);
2627                 else {
2628                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2629                         mmu_page_remove_parent_pte(child, spte);
2630                 }
2631         }
2632         __set_spte(spte, shadow_trap_nonpresent_pte);
2633         if (is_large_pte(pte))
2634                 --vcpu->kvm->stat.lpages;
2635 }
2636
2637 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2638                                   struct kvm_mmu_page *sp,
2639                                   u64 *spte,
2640                                   const void *new)
2641 {
2642         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2643                 ++vcpu->kvm->stat.mmu_pde_zapped;
2644                 return;
2645         }
2646
2647         ++vcpu->kvm->stat.mmu_pte_updated;
2648         if (!sp->role.cr4_pae)
2649                 paging32_update_pte(vcpu, sp, spte, new);
2650         else
2651                 paging64_update_pte(vcpu, sp, spte, new);
2652 }
2653
2654 static bool need_remote_flush(u64 old, u64 new)
2655 {
2656         if (!is_shadow_present_pte(old))
2657                 return false;
2658         if (!is_shadow_present_pte(new))
2659                 return true;
2660         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2661                 return true;
2662         old ^= PT64_NX_MASK;
2663         new ^= PT64_NX_MASK;
2664         return (old & ~new & PT64_PERM_MASK) != 0;
2665 }
2666
2667 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, bool zap_page,
2668                                     bool remote_flush, bool local_flush)
2669 {
2670         if (zap_page)
2671                 return;
2672
2673         if (remote_flush)
2674                 kvm_flush_remote_tlbs(vcpu->kvm);
2675         else if (local_flush)
2676                 kvm_mmu_flush_tlb(vcpu);
2677 }
2678
2679 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2680 {
2681         u64 *spte = vcpu->arch.last_pte_updated;
2682
2683         return !!(spte && (*spte & shadow_accessed_mask));
2684 }
2685
2686 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2687                                           u64 gpte)
2688 {
2689         gfn_t gfn;
2690         pfn_t pfn;
2691
2692         if (!is_present_gpte(gpte))
2693                 return;
2694         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2695
2696         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2697         smp_rmb();
2698         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2699
2700         if (is_error_pfn(pfn)) {
2701                 kvm_release_pfn_clean(pfn);
2702                 return;
2703         }
2704         vcpu->arch.update_pte.gfn = gfn;
2705         vcpu->arch.update_pte.pfn = pfn;
2706 }
2707
2708 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2709 {
2710         u64 *spte = vcpu->arch.last_pte_updated;
2711
2712         if (spte
2713             && vcpu->arch.last_pte_gfn == gfn
2714             && shadow_accessed_mask
2715             && !(*spte & shadow_accessed_mask)
2716             && is_shadow_present_pte(*spte))
2717                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2718 }
2719
2720 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2721                        const u8 *new, int bytes,
2722                        bool guest_initiated)
2723 {
2724         gfn_t gfn = gpa >> PAGE_SHIFT;
2725         struct kvm_mmu_page *sp;
2726         struct hlist_node *node;
2727         LIST_HEAD(invalid_list);
2728         u64 entry, gentry;
2729         u64 *spte;
2730         unsigned offset = offset_in_page(gpa);
2731         unsigned pte_size;
2732         unsigned page_offset;
2733         unsigned misaligned;
2734         unsigned quadrant;
2735         int level;
2736         int flooded = 0;
2737         int npte;
2738         int r;
2739         int invlpg_counter;
2740         bool remote_flush, local_flush, zap_page;
2741
2742         zap_page = remote_flush = local_flush = false;
2743
2744         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2745
2746         invlpg_counter = atomic_read(&vcpu->kvm->arch.invlpg_counter);
2747
2748         /*
2749          * Assume that the pte write on a page table of the same type
2750          * as the current vcpu paging mode.  This is nearly always true
2751          * (might be false while changing modes).  Note it is verified later
2752          * by update_pte().
2753          */
2754         if ((is_pae(vcpu) && bytes == 4) || !new) {
2755                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2756                 if (is_pae(vcpu)) {
2757                         gpa &= ~(gpa_t)7;
2758                         bytes = 8;
2759                 }
2760                 r = kvm_read_guest(vcpu->kvm, gpa, &gentry, min(bytes, 8));
2761                 if (r)
2762                         gentry = 0;
2763                 new = (const u8 *)&gentry;
2764         }
2765
2766         switch (bytes) {
2767         case 4:
2768                 gentry = *(const u32 *)new;
2769                 break;
2770         case 8:
2771                 gentry = *(const u64 *)new;
2772                 break;
2773         default:
2774                 gentry = 0;
2775                 break;
2776         }
2777
2778         mmu_guess_page_from_pte_write(vcpu, gpa, gentry);
2779         spin_lock(&vcpu->kvm->mmu_lock);
2780         if (atomic_read(&vcpu->kvm->arch.invlpg_counter) != invlpg_counter)
2781                 gentry = 0;
2782         kvm_mmu_access_page(vcpu, gfn);
2783         kvm_mmu_free_some_pages(vcpu);
2784         ++vcpu->kvm->stat.mmu_pte_write;
2785         kvm_mmu_audit(vcpu, "pre pte write");
2786         if (guest_initiated) {
2787                 if (gfn == vcpu->arch.last_pt_write_gfn
2788                     && !last_updated_pte_accessed(vcpu)) {
2789                         ++vcpu->arch.last_pt_write_count;
2790                         if (vcpu->arch.last_pt_write_count >= 3)
2791                                 flooded = 1;
2792                 } else {
2793                         vcpu->arch.last_pt_write_gfn = gfn;
2794                         vcpu->arch.last_pt_write_count = 1;
2795                         vcpu->arch.last_pte_updated = NULL;
2796                 }
2797         }
2798
2799         for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn, node) {
2800                 pte_size = sp->role.cr4_pae ? 8 : 4;
2801                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2802                 misaligned |= bytes < 4;
2803                 if (misaligned || flooded) {
2804                         /*
2805                          * Misaligned accesses are too much trouble to fix
2806                          * up; also, they usually indicate a page is not used
2807                          * as a page table.
2808                          *
2809                          * If we're seeing too many writes to a page,
2810                          * it may no longer be a page table, or we may be
2811                          * forking, in which case it is better to unmap the
2812                          * page.
2813                          */
2814                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2815                                  gpa, bytes, sp->role.word);
2816                         zap_page |= !!kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2817                                                      &invalid_list);
2818                         ++vcpu->kvm->stat.mmu_flooded;
2819                         continue;
2820                 }
2821                 page_offset = offset;
2822                 level = sp->role.level;
2823                 npte = 1;
2824                 if (!sp->role.cr4_pae) {
2825                         page_offset <<= 1;      /* 32->64 */
2826                         /*
2827                          * A 32-bit pde maps 4MB while the shadow pdes map
2828                          * only 2MB.  So we need to double the offset again
2829                          * and zap two pdes instead of one.
2830                          */
2831                         if (level == PT32_ROOT_LEVEL) {
2832                                 page_offset &= ~7; /* kill rounding error */
2833                                 page_offset <<= 1;
2834                                 npte = 2;
2835                         }
2836                         quadrant = page_offset >> PAGE_SHIFT;
2837                         page_offset &= ~PAGE_MASK;
2838                         if (quadrant != sp->role.quadrant)
2839                                 continue;
2840                 }
2841                 local_flush = true;
2842                 spte = &sp->spt[page_offset / sizeof(*spte)];
2843                 while (npte--) {
2844                         entry = *spte;
2845                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2846                         if (gentry)
2847                                 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
2848                         if (!remote_flush && need_remote_flush(entry, *spte))
2849                                 remote_flush = true;
2850                         ++spte;
2851                 }
2852         }
2853         mmu_pte_write_flush_tlb(vcpu, zap_page, remote_flush, local_flush);
2854         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2855         kvm_mmu_audit(vcpu, "post pte write");
2856         spin_unlock(&vcpu->kvm->mmu_lock);
2857         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2858                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2859                 vcpu->arch.update_pte.pfn = bad_pfn;
2860         }
2861 }
2862
2863 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2864 {
2865         gpa_t gpa;
2866         int r;
2867
2868         if (tdp_enabled)
2869                 return 0;
2870
2871         gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2872
2873         spin_lock(&vcpu->kvm->mmu_lock);
2874         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2875         spin_unlock(&vcpu->kvm->mmu_lock);
2876         return r;
2877 }
2878 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2879
2880 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2881 {
2882         int free_pages;
2883         LIST_HEAD(invalid_list);
2884
2885         free_pages = vcpu->kvm->arch.n_free_mmu_pages;
2886         while (free_pages < KVM_REFILL_PAGES &&
2887                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2888                 struct kvm_mmu_page *sp;
2889
2890                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2891                                   struct kvm_mmu_page, link);
2892                 free_pages += kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2893                                                        &invalid_list);
2894                 ++vcpu->kvm->stat.mmu_recycled;
2895         }
2896         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2897 }
2898
2899 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2900 {
2901         int r;
2902         enum emulation_result er;
2903
2904         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2905         if (r < 0)
2906                 goto out;
2907
2908         if (!r) {
2909                 r = 1;
2910                 goto out;
2911         }
2912
2913         r = mmu_topup_memory_caches(vcpu);
2914         if (r)
2915                 goto out;
2916
2917         er = emulate_instruction(vcpu, cr2, error_code, 0);
2918
2919         switch (er) {
2920         case EMULATE_DONE:
2921                 return 1;
2922         case EMULATE_DO_MMIO:
2923                 ++vcpu->stat.mmio_exits;
2924                 /* fall through */
2925         case EMULATE_FAIL:
2926                 return 0;
2927         default:
2928                 BUG();
2929         }
2930 out:
2931         return r;
2932 }
2933 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2934
2935 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2936 {
2937         vcpu->arch.mmu.invlpg(vcpu, gva);
2938         kvm_mmu_flush_tlb(vcpu);
2939         ++vcpu->stat.invlpg;
2940 }
2941 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2942
2943 void kvm_enable_tdp(void)
2944 {
2945         tdp_enabled = true;
2946 }
2947 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2948
2949 void kvm_disable_tdp(void)
2950 {
2951         tdp_enabled = false;
2952 }
2953 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2954
2955 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2956 {
2957         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2958 }
2959
2960 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2961 {
2962         struct page *page;
2963         int i;
2964
2965         ASSERT(vcpu);
2966
2967         /*
2968          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2969          * Therefore we need to allocate shadow page tables in the first
2970          * 4GB of memory, which happens to fit the DMA32 zone.
2971          */
2972         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2973         if (!page)
2974                 return -ENOMEM;
2975
2976         vcpu->arch.mmu.pae_root = page_address(page);
2977         for (i = 0; i < 4; ++i)
2978                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2979
2980         return 0;
2981 }
2982
2983 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2984 {
2985         ASSERT(vcpu);
2986         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2987
2988         return alloc_mmu_pages(vcpu);
2989 }
2990
2991 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2992 {
2993         ASSERT(vcpu);
2994         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2995
2996         return init_kvm_mmu(vcpu);
2997 }
2998
2999 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
3000 {
3001         ASSERT(vcpu);
3002
3003         destroy_kvm_mmu(vcpu);
3004         free_mmu_pages(vcpu);
3005         mmu_free_memory_caches(vcpu);
3006 }
3007
3008 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
3009 {
3010         struct kvm_mmu_page *sp;
3011
3012         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
3013                 int i;
3014                 u64 *pt;
3015
3016                 if (!test_bit(slot, sp->slot_bitmap))
3017                         continue;
3018
3019                 pt = sp->spt;
3020                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
3021                         /* avoid RMW */
3022                         if (is_writable_pte(pt[i]))
3023                                 pt[i] &= ~PT_WRITABLE_MASK;
3024         }
3025         kvm_flush_remote_tlbs(kvm);
3026 }
3027
3028 void kvm_mmu_zap_all(struct kvm *kvm)
3029 {
3030         struct kvm_mmu_page *sp, *node;
3031         LIST_HEAD(invalid_list);
3032
3033         spin_lock(&kvm->mmu_lock);
3034 restart:
3035         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
3036                 if (kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list))
3037                         goto restart;
3038
3039         kvm_mmu_commit_zap_page(kvm, &invalid_list);
3040         spin_unlock(&kvm->mmu_lock);
3041 }
3042
3043 static int kvm_mmu_remove_some_alloc_mmu_pages(struct kvm *kvm,
3044                                                struct list_head *invalid_list)
3045 {
3046         struct kvm_mmu_page *page;
3047
3048         page = container_of(kvm->arch.active_mmu_pages.prev,
3049                             struct kvm_mmu_page, link);
3050         return kvm_mmu_prepare_zap_page(kvm, page, invalid_list);
3051 }
3052
3053 static int mmu_shrink(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
3054 {
3055         struct kvm *kvm;
3056         struct kvm *kvm_freed = NULL;
3057         int cache_count = 0;
3058
3059         spin_lock(&kvm_lock);
3060
3061         list_for_each_entry(kvm, &vm_list, vm_list) {
3062                 int npages, idx, freed_pages;
3063                 LIST_HEAD(invalid_list);
3064
3065                 idx = srcu_read_lock(&kvm->srcu);
3066                 spin_lock(&kvm->mmu_lock);
3067                 npages = kvm->arch.n_alloc_mmu_pages -
3068                          kvm->arch.n_free_mmu_pages;
3069                 cache_count += npages;
3070                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
3071                         freed_pages = kvm_mmu_remove_some_alloc_mmu_pages(kvm,
3072                                                           &invalid_list);
3073                         cache_count -= freed_pages;
3074                         kvm_freed = kvm;
3075                 }
3076                 nr_to_scan--;
3077
3078                 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3079                 spin_unlock(&kvm->mmu_lock);
3080                 srcu_read_unlock(&kvm->srcu, idx);
3081         }
3082         if (kvm_freed)
3083                 list_move_tail(&kvm_freed->vm_list, &vm_list);
3084
3085         spin_unlock(&kvm_lock);
3086
3087         return cache_count;
3088 }
3089
3090 static struct shrinker mmu_shrinker = {
3091         .shrink = mmu_shrink,
3092         .seeks = DEFAULT_SEEKS * 10,
3093 };
3094
3095 static void mmu_destroy_caches(void)
3096 {
3097         if (pte_chain_cache)
3098                 kmem_cache_destroy(pte_chain_cache);
3099         if (rmap_desc_cache)
3100                 kmem_cache_destroy(rmap_desc_cache);
3101         if (mmu_page_header_cache)
3102                 kmem_cache_destroy(mmu_page_header_cache);
3103 }
3104
3105 void kvm_mmu_module_exit(void)
3106 {
3107         mmu_destroy_caches();
3108         unregister_shrinker(&mmu_shrinker);
3109 }
3110
3111 int kvm_mmu_module_init(void)
3112 {
3113         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
3114                                             sizeof(struct kvm_pte_chain),
3115                                             0, 0, NULL);
3116         if (!pte_chain_cache)
3117                 goto nomem;
3118         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
3119                                             sizeof(struct kvm_rmap_desc),
3120                                             0, 0, NULL);
3121         if (!rmap_desc_cache)
3122                 goto nomem;
3123
3124         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
3125                                                   sizeof(struct kvm_mmu_page),
3126                                                   0, 0, NULL);
3127         if (!mmu_page_header_cache)
3128                 goto nomem;
3129
3130         register_shrinker(&mmu_shrinker);
3131
3132         return 0;
3133
3134 nomem:
3135         mmu_destroy_caches();
3136         return -ENOMEM;
3137 }
3138
3139 /*
3140  * Caculate mmu pages needed for kvm.
3141  */
3142 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3143 {
3144         int i;
3145         unsigned int nr_mmu_pages;
3146         unsigned int  nr_pages = 0;
3147         struct kvm_memslots *slots;
3148
3149         slots = kvm_memslots(kvm);
3150
3151         for (i = 0; i < slots->nmemslots; i++)
3152                 nr_pages += slots->memslots[i].npages;
3153
3154         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3155         nr_mmu_pages = max(nr_mmu_pages,
3156                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3157
3158         return nr_mmu_pages;
3159 }
3160
3161 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3162                                 unsigned len)
3163 {
3164         if (len > buffer->len)
3165                 return NULL;
3166         return buffer->ptr;
3167 }
3168
3169 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3170                                 unsigned len)
3171 {
3172         void *ret;
3173
3174         ret = pv_mmu_peek_buffer(buffer, len);
3175         if (!ret)
3176                 return ret;
3177         buffer->ptr += len;
3178         buffer->len -= len;
3179         buffer->processed += len;
3180         return ret;
3181 }
3182
3183 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3184                              gpa_t addr, gpa_t value)
3185 {
3186         int bytes = 8;
3187         int r;
3188
3189         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3190                 bytes = 4;
3191
3192         r = mmu_topup_memory_caches(vcpu);
3193         if (r)
3194                 return r;
3195
3196         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3197                 return -EFAULT;
3198
3199         return 1;
3200 }
3201
3202 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3203 {
3204         (void)kvm_set_cr3(vcpu, vcpu->arch.cr3);
3205         return 1;
3206 }
3207
3208 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3209 {
3210         spin_lock(&vcpu->kvm->mmu_lock);
3211         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3212         spin_unlock(&vcpu->kvm->mmu_lock);
3213         return 1;
3214 }
3215
3216 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3217                              struct kvm_pv_mmu_op_buffer *buffer)
3218 {
3219         struct kvm_mmu_op_header *header;
3220
3221         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3222         if (!header)
3223                 return 0;
3224         switch (header->op) {
3225         case KVM_MMU_OP_WRITE_PTE: {
3226                 struct kvm_mmu_op_write_pte *wpte;
3227
3228                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3229                 if (!wpte)
3230                         return 0;
3231                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3232                                         wpte->pte_val);
3233         }
3234         case KVM_MMU_OP_FLUSH_TLB: {
3235                 struct kvm_mmu_op_flush_tlb *ftlb;
3236
3237                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3238                 if (!ftlb)
3239                         return 0;
3240                 return kvm_pv_mmu_flush_tlb(vcpu);
3241         }
3242         case KVM_MMU_OP_RELEASE_PT: {
3243                 struct kvm_mmu_op_release_pt *rpt;
3244
3245                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3246                 if (!rpt)
3247                         return 0;
3248                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3249         }
3250         default: return 0;
3251         }
3252 }
3253
3254 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3255                   gpa_t addr, unsigned long *ret)
3256 {
3257         int r;
3258         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3259
3260         buffer->ptr = buffer->buf;
3261         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3262         buffer->processed = 0;
3263
3264         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3265         if (r)
3266                 goto out;
3267
3268         while (buffer->len) {
3269                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3270                 if (r < 0)
3271                         goto out;
3272                 if (r == 0)
3273                         break;
3274         }
3275
3276         r = 1;
3277 out:
3278         *ret = buffer->processed;
3279         return r;
3280 }
3281
3282 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3283 {
3284         struct kvm_shadow_walk_iterator iterator;
3285         int nr_sptes = 0;
3286
3287         spin_lock(&vcpu->kvm->mmu_lock);
3288         for_each_shadow_entry(vcpu, addr, iterator) {
3289                 sptes[iterator.level-1] = *iterator.sptep;
3290                 nr_sptes++;
3291                 if (!is_shadow_present_pte(*iterator.sptep))
3292                         break;
3293         }
3294         spin_unlock(&vcpu->kvm->mmu_lock);
3295
3296         return nr_sptes;
3297 }
3298 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3299
3300 #ifdef AUDIT
3301
3302 static const char *audit_msg;
3303
3304 static gva_t canonicalize(gva_t gva)
3305 {
3306 #ifdef CONFIG_X86_64
3307         gva = (long long)(gva << 16) >> 16;
3308 #endif
3309         return gva;
3310 }
3311
3312
3313 typedef void (*inspect_spte_fn) (struct kvm *kvm, u64 *sptep);
3314
3315 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3316                             inspect_spte_fn fn)
3317 {
3318         int i;
3319
3320         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3321                 u64 ent = sp->spt[i];
3322
3323                 if (is_shadow_present_pte(ent)) {
3324                         if (!is_last_spte(ent, sp->role.level)) {
3325                                 struct kvm_mmu_page *child;
3326                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3327                                 __mmu_spte_walk(kvm, child, fn);
3328                         } else
3329                                 fn(kvm, &sp->spt[i]);
3330                 }
3331         }
3332 }
3333
3334 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3335 {
3336         int i;
3337         struct kvm_mmu_page *sp;
3338
3339         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3340                 return;
3341         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3342                 hpa_t root = vcpu->arch.mmu.root_hpa;
3343                 sp = page_header(root);
3344                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3345                 return;
3346         }
3347         for (i = 0; i < 4; ++i) {
3348                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3349
3350                 if (root && VALID_PAGE(root)) {
3351                         root &= PT64_BASE_ADDR_MASK;
3352                         sp = page_header(root);
3353                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3354                 }
3355         }
3356         return;
3357 }
3358
3359 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3360                                 gva_t va, int level)
3361 {
3362         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3363         int i;
3364         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3365
3366         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3367                 u64 ent = pt[i];
3368
3369                 if (ent == shadow_trap_nonpresent_pte)
3370                         continue;
3371
3372                 va = canonicalize(va);
3373                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3374                         audit_mappings_page(vcpu, ent, va, level - 1);
3375                 else {
3376                         gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, va, NULL);
3377                         gfn_t gfn = gpa >> PAGE_SHIFT;
3378                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3379                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3380
3381                         if (is_error_pfn(pfn)) {
3382                                 kvm_release_pfn_clean(pfn);
3383                                 continue;
3384                         }
3385
3386                         if (is_shadow_present_pte(ent)
3387                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3388                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3389                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3390                                        audit_msg, vcpu->arch.mmu.root_level,
3391                                        va, gpa, hpa, ent,
3392                                        is_shadow_present_pte(ent));
3393                         else if (ent == shadow_notrap_nonpresent_pte
3394                                  && !is_error_hpa(hpa))
3395                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3396                                        " valid guest gva %lx\n", audit_msg, va);
3397                         kvm_release_pfn_clean(pfn);
3398
3399                 }
3400         }
3401 }
3402
3403 static void audit_mappings(struct kvm_vcpu *vcpu)
3404 {
3405         unsigned i;
3406
3407         if (vcpu->arch.mmu.root_level == 4)
3408                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3409         else
3410                 for (i = 0; i < 4; ++i)
3411                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3412                                 audit_mappings_page(vcpu,
3413                                                     vcpu->arch.mmu.pae_root[i],
3414                                                     i << 30,
3415                                                     2);
3416 }
3417
3418 static int count_rmaps(struct kvm_vcpu *vcpu)
3419 {
3420         struct kvm *kvm = vcpu->kvm;
3421         struct kvm_memslots *slots;
3422         int nmaps = 0;
3423         int i, j, k, idx;
3424
3425         idx = srcu_read_lock(&kvm->srcu);
3426         slots = kvm_memslots(kvm);
3427         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3428                 struct kvm_memory_slot *m = &slots->memslots[i];
3429                 struct kvm_rmap_desc *d;
3430
3431                 for (j = 0; j < m->npages; ++j) {
3432                         unsigned long *rmapp = &m->rmap[j];
3433
3434                         if (!*rmapp)
3435                                 continue;
3436                         if (!(*rmapp & 1)) {
3437                                 ++nmaps;
3438                                 continue;
3439                         }
3440                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3441                         while (d) {
3442                                 for (k = 0; k < RMAP_EXT; ++k)
3443                                         if (d->sptes[k])
3444                                                 ++nmaps;
3445                                         else
3446                                                 break;
3447                                 d = d->more;
3448                         }
3449                 }
3450         }
3451         srcu_read_unlock(&kvm->srcu, idx);
3452         return nmaps;
3453 }
3454
3455 void inspect_spte_has_rmap(struct kvm *kvm, u64 *sptep)
3456 {
3457         unsigned long *rmapp;
3458         struct kvm_mmu_page *rev_sp;
3459         gfn_t gfn;
3460
3461         if (is_writable_pte(*sptep)) {
3462                 rev_sp = page_header(__pa(sptep));
3463                 gfn = kvm_mmu_page_get_gfn(rev_sp, sptep - rev_sp->spt);
3464
3465                 if (!gfn_to_memslot(kvm, gfn)) {
3466                         if (!printk_ratelimit())
3467                                 return;
3468                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3469                                          audit_msg, gfn);
3470                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3471                                audit_msg, (long int)(sptep - rev_sp->spt),
3472                                         rev_sp->gfn);
3473                         dump_stack();
3474                         return;
3475                 }
3476
3477                 rmapp = gfn_to_rmap(kvm, gfn, rev_sp->role.level);
3478                 if (!*rmapp) {
3479                         if (!printk_ratelimit())
3480                                 return;
3481                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3482                                          audit_msg, *sptep);
3483                         dump_stack();
3484                 }
3485         }
3486
3487 }
3488
3489 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3490 {
3491         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3492 }
3493
3494 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3495 {
3496         struct kvm_mmu_page *sp;
3497         int i;
3498
3499         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3500                 u64 *pt = sp->spt;
3501
3502                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3503                         continue;
3504
3505                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3506                         u64 ent = pt[i];
3507
3508                         if (!(ent & PT_PRESENT_MASK))
3509                                 continue;
3510                         if (!is_writable_pte(ent))
3511                                 continue;
3512                         inspect_spte_has_rmap(vcpu->kvm, &pt[i]);
3513                 }
3514         }
3515         return;
3516 }
3517
3518 static void audit_rmap(struct kvm_vcpu *vcpu)
3519 {
3520         check_writable_mappings_rmap(vcpu);
3521         count_rmaps(vcpu);
3522 }
3523
3524 static void audit_write_protection(struct kvm_vcpu *vcpu)
3525 {
3526         struct kvm_mmu_page *sp;
3527         struct kvm_memory_slot *slot;
3528         unsigned long *rmapp;
3529         u64 *spte;
3530         gfn_t gfn;
3531
3532         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3533                 if (sp->role.direct)
3534                         continue;
3535                 if (sp->unsync)
3536                         continue;
3537
3538                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3539                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3540                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3541
3542                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3543                 while (spte) {
3544                         if (is_writable_pte(*spte))
3545                                 printk(KERN_ERR "%s: (%s) shadow page has "
3546                                 "writable mappings: gfn %lx role %x\n",
3547                                __func__, audit_msg, sp->gfn,
3548                                sp->role.word);
3549                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3550                 }
3551         }
3552 }
3553
3554 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3555 {
3556         int olddbg = dbg;
3557
3558         dbg = 0;
3559         audit_msg = msg;
3560         audit_rmap(vcpu);
3561         audit_write_protection(vcpu);
3562         if (strcmp("pre pte write", audit_msg) != 0)
3563                 audit_mappings(vcpu);
3564         audit_writable_sptes_have_rmaps(vcpu);
3565         dbg = olddbg;
3566 }
3567
3568 #endif