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

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "mmu.h"
  21 #include "x86.h"
  22 #include "kvm_cache_regs.h"
  23
  24 #include <linux/kvm_host.h>
  25 #include <linux/types.h>
  26 #include <linux/string.h>
  27 #include <linux/mm.h>
  28 #include <linux/highmem.h>
  29 #include <linux/module.h>
  30 #include <linux/swap.h>
  31 #include <linux/hugetlb.h>
  32 #include <linux/compiler.h>
  33 #include <linux/srcu.h>
  34 #include <linux/slab.h>
  35
  36 #include <asm/page.h>
  37 #include <asm/cmpxchg.h>
  38 #include <asm/io.h>
  39 #include <asm/vmx.h>
  40
  41 /*
  42  * When setting this variable to true it enables Two-Dimensional-Paging
  43  * where the hardware walks 2 page tables:
  44  * 1. the guest-virtual to guest-physical
  45  * 2. while doing 1. it walks guest-physical to host-physical
  46  * If the hardware supports that we don't need to do shadow paging.
  47  */
  48 bool tdp_enabled = false;
  49
  50 #undef MMU_DEBUG
  51
  52 #undef AUDIT
  53
  54 #ifdef AUDIT
  55 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  56 #else
  57 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  58 #endif
  59
  60 #ifdef MMU_DEBUG
  61
  62 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  63 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  64
  65 #else
  66
  67 #define pgprintk(x...) do { } while (0)
  68 #define rmap_printk(x...) do { } while (0)
  69
  70 #endif
  71
  72 #if defined(MMU_DEBUG) || defined(AUDIT)
  73 static int dbg = 0;
  74 module_param(dbg, bool, 0644);
  75 #endif
  76
  77 static int oos_shadow = 1;
  78 module_param(oos_shadow, bool, 0644);
  79
  80 #ifndef MMU_DEBUG
  81 #define ASSERT(x) do { } while (0)
  82 #else
  83 #define ASSERT(x)                                                       \
  84         if (!(x)) {                                                     \
  85                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  86                        __FILE__, __LINE__, #x);                         \
  87         }
  88 #endif
  89
  90 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  91 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  92
  93 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  94
  95 #define PT64_LEVEL_BITS 9
  96
  97 #define PT64_LEVEL_SHIFT(level) \
  98                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  99
 100 #define PT64_LEVEL_MASK(level) \
 101                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 102
 103 #define PT64_INDEX(address, level)\
 104         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 105
 106
 107 #define PT32_LEVEL_BITS 10
 108
 109 #define PT32_LEVEL_SHIFT(level) \
 110                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 111
 112 #define PT32_LEVEL_MASK(level) \
 113                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 114 #define PT32_LVL_OFFSET_MASK(level) \
 115         (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 116                                                 * PT32_LEVEL_BITS))) - 1))
 117
 118 #define PT32_INDEX(address, level)\
 119         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 120
 121
 122 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 123 #define PT64_DIR_BASE_ADDR_MASK \
 124         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 125 #define PT64_LVL_ADDR_MASK(level) \
 126         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 127                                                 * PT64_LEVEL_BITS))) - 1))
 128 #define PT64_LVL_OFFSET_MASK(level) \
 129         (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 130                                                 * PT64_LEVEL_BITS))) - 1))
 131
 132 #define PT32_BASE_ADDR_MASK PAGE_MASK
 133 #define PT32_DIR_BASE_ADDR_MASK \
 134         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 135 #define PT32_LVL_ADDR_MASK(level) \
 136         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 137                                             * PT32_LEVEL_BITS))) - 1))
 138
 139 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 140                         | PT64_NX_MASK)
 141
 142 #define RMAP_EXT 4
 143
 144 #define ACC_EXEC_MASK    1
 145 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 146 #define ACC_USER_MASK    PT_USER_MASK
 147 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 148
 149 #include <trace/events/kvm.h>
 150
 151 #define CREATE_TRACE_POINTS
 152 #include "mmutrace.h"
 153
 154 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 155
 156 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 157
 158 struct kvm_rmap_desc {
 159         u64 *sptes[RMAP_EXT];
 160         struct kvm_rmap_desc *more;
 161 };
 162
 163 struct kvm_shadow_walk_iterator {
 164         u64 addr;
 165         hpa_t shadow_addr;
 166         int level;
 167         u64 *sptep;
 168         unsigned index;
 169 };
 170
 171 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
 172         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
 173              shadow_walk_okay(&(_walker));                      \
 174              shadow_walk_next(&(_walker)))
 175
 176
 177 struct kvm_unsync_walk {
 178         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 179 };
 180
 181 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 182
 183 static struct kmem_cache *pte_chain_cache;
 184 static struct kmem_cache *rmap_desc_cache;
 185 static struct kmem_cache *mmu_page_header_cache;
 186
 187 static u64 __read_mostly shadow_trap_nonpresent_pte;
 188 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 189 static u64 __read_mostly shadow_base_present_pte;
 190 static u64 __read_mostly shadow_nx_mask;
 191 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 192 static u64 __read_mostly shadow_user_mask;
 193 static u64 __read_mostly shadow_accessed_mask;
 194 static u64 __read_mostly shadow_dirty_mask;
 195
 196 static inline u64 rsvd_bits(int s, int e)
 197 {
 198         return ((1ULL << (e - s + 1)) - 1) << s;
 199 }
 200
 201 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 202 {
 203         shadow_trap_nonpresent_pte = trap_pte;
 204         shadow_notrap_nonpresent_pte = notrap_pte;
 205 }
 206 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 207
 208 void kvm_mmu_set_base_ptes(u64 base_pte)
 209 {
 210         shadow_base_present_pte = base_pte;
 211 }
 212 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 213
 214 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 215                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 216 {
 217         shadow_user_mask = user_mask;
 218         shadow_accessed_mask = accessed_mask;
 219         shadow_dirty_mask = dirty_mask;
 220         shadow_nx_mask = nx_mask;
 221         shadow_x_mask = x_mask;
 222 }
 223 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 224
 225 static int is_write_protection(struct kvm_vcpu *vcpu)
 226 {
 227         return kvm_read_cr0_bits(vcpu, X86_CR0_WP);
 228 }
 229
 230 static int is_cpuid_PSE36(void)
 231 {
 232         return 1;
 233 }
 234
 235 static int is_nx(struct kvm_vcpu *vcpu)
 236 {
 237         return vcpu->arch.efer & EFER_NX;
 238 }
 239
 240 static int is_shadow_present_pte(u64 pte)
 241 {
 242         return pte != shadow_trap_nonpresent_pte
 243                 && pte != shadow_notrap_nonpresent_pte;
 244 }
 245
 246 static int is_large_pte(u64 pte)
 247 {
 248         return pte & PT_PAGE_SIZE_MASK;
 249 }
 250
 251 static int is_writable_pte(unsigned long pte)
 252 {
 253         return pte & PT_WRITABLE_MASK;
 254 }
 255
 256 static int is_dirty_gpte(unsigned long pte)
 257 {
 258         return pte & PT_DIRTY_MASK;
 259 }
 260
 261 static int is_rmap_spte(u64 pte)
 262 {
 263         return is_shadow_present_pte(pte);
 264 }
 265
 266 static int is_last_spte(u64 pte, int level)
 267 {
 268         if (level == PT_PAGE_TABLE_LEVEL)
 269                 return 1;
 270         if (is_large_pte(pte))
 271                 return 1;
 272         return 0;
 273 }
 274
 275 static pfn_t spte_to_pfn(u64 pte)
 276 {
 277         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 278 }
 279
 280 static gfn_t pse36_gfn_delta(u32 gpte)
 281 {
 282         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 283
 284         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 285 }
 286
 287 static void __set_spte(u64 *sptep, u64 spte)
 288 {
 289 #ifdef CONFIG_X86_64
 290         set_64bit((unsigned long *)sptep, spte);
 291 #else
 292         set_64bit((unsigned long long *)sptep, spte);
 293 #endif
 294 }
 295
 296 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 297                                   struct kmem_cache *base_cache, int min)
 298 {
 299         void *obj;
 300
 301         if (cache->nobjs >= min)
 302                 return 0;
 303         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 304                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 305                 if (!obj)
 306                         return -ENOMEM;
 307                 cache->objects[cache->nobjs++] = obj;
 308         }
 309         return 0;
 310 }
 311
 312 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 313 {
 314         while (mc->nobjs)
 315                 kfree(mc->objects[--mc->nobjs]);
 316 }
 317
 318 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 319                                        int min)
 320 {
 321         struct page *page;
 322
 323         if (cache->nobjs >= min)
 324                 return 0;
 325         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 326                 page = alloc_page(GFP_KERNEL);
 327                 if (!page)
 328                         return -ENOMEM;
 329                 cache->objects[cache->nobjs++] = page_address(page);
 330         }
 331         return 0;
 332 }
 333
 334 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 335 {
 336         while (mc->nobjs)
 337                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 338 }
 339
 340 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 341 {
 342         int r;
 343
 344         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 345                                    pte_chain_cache, 4);
 346         if (r)
 347                 goto out;
 348         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 349                                    rmap_desc_cache, 4);
 350         if (r)
 351                 goto out;
 352         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 353         if (r)
 354                 goto out;
 355         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 356                                    mmu_page_header_cache, 4);
 357 out:
 358         return r;
 359 }
 360
 361 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 362 {
 363         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 364         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 365         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 366         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 367 }
 368
 369 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 370                                     size_t size)
 371 {
 372         void *p;
 373
 374         BUG_ON(!mc->nobjs);
 375         p = mc->objects[--mc->nobjs];
 376         return p;
 377 }
 378
 379 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 380 {
 381         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 382                                       sizeof(struct kvm_pte_chain));
 383 }
 384
 385 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 386 {
 387         kfree(pc);
 388 }
 389
 390 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 391 {
 392         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 393                                       sizeof(struct kvm_rmap_desc));
 394 }
 395
 396 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 397 {
 398         kfree(rd);
 399 }
 400
 401 /*
 402  * Return the pointer to the largepage write count for a given
 403  * gfn, handling slots that are not large page aligned.
 404  */
 405 static int *slot_largepage_idx(gfn_t gfn,
 406                                struct kvm_memory_slot *slot,
 407                                int level)
 408 {
 409         unsigned long idx;
 410
 411         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 412               (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 413         return &slot->lpage_info[level - 2][idx].write_count;
 414 }
 415
 416 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 417 {
 418         struct kvm_memory_slot *slot;
 419         int *write_count;
 420         int i;
 421
 422         gfn = unalias_gfn(kvm, gfn);
 423
 424         slot = gfn_to_memslot_unaliased(kvm, gfn);
 425         for (i = PT_DIRECTORY_LEVEL;
 426              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 427                 write_count   = slot_largepage_idx(gfn, slot, i);
 428                 *write_count += 1;
 429         }
 430 }
 431
 432 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 433 {
 434         struct kvm_memory_slot *slot;
 435         int *write_count;
 436         int i;
 437
 438         gfn = unalias_gfn(kvm, gfn);
 439         for (i = PT_DIRECTORY_LEVEL;
 440              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 441                 slot          = gfn_to_memslot_unaliased(kvm, gfn);
 442                 write_count   = slot_largepage_idx(gfn, slot, i);
 443                 *write_count -= 1;
 444                 WARN_ON(*write_count < 0);
 445         }
 446 }
 447
 448 static int has_wrprotected_page(struct kvm *kvm,
 449                                 gfn_t gfn,
 450                                 int level)
 451 {
 452         struct kvm_memory_slot *slot;
 453         int *largepage_idx;
 454
 455         gfn = unalias_gfn(kvm, gfn);
 456         slot = gfn_to_memslot_unaliased(kvm, gfn);
 457         if (slot) {
 458                 largepage_idx = slot_largepage_idx(gfn, slot, level);
 459                 return *largepage_idx;
 460         }
 461
 462         return 1;
 463 }
 464
 465 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
 466 {
 467         unsigned long page_size;
 468         int i, ret = 0;
 469
 470         page_size = kvm_host_page_size(kvm, gfn);
 471
 472         for (i = PT_PAGE_TABLE_LEVEL;
 473              i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
 474                 if (page_size >= KVM_HPAGE_SIZE(i))
 475                         ret = i;
 476                 else
 477                         break;
 478         }
 479
 480         return ret;
 481 }
 482
 483 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 484 {
 485         struct kvm_memory_slot *slot;
 486         int host_level, level, max_level;
 487
 488         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 489         if (slot && slot->dirty_bitmap)
 490                 return PT_PAGE_TABLE_LEVEL;
 491
 492         host_level = host_mapping_level(vcpu->kvm, large_gfn);
 493
 494         if (host_level == PT_PAGE_TABLE_LEVEL)
 495                 return host_level;
 496
 497         max_level = kvm_x86_ops->get_lpage_level() < host_level ?
 498                 kvm_x86_ops->get_lpage_level() : host_level;
 499
 500         for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
 501                 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
 502                         break;
 503
 504         return level - 1;
 505 }
 506
 507 /*
 508  * Take gfn and return the reverse mapping to it.
 509  * Note: gfn must be unaliased before this function get called
 510  */
 511
 512 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
 513 {
 514         struct kvm_memory_slot *slot;
 515         unsigned long idx;
 516
 517         slot = gfn_to_memslot(kvm, gfn);
 518         if (likely(level == PT_PAGE_TABLE_LEVEL))
 519                 return &slot->rmap[gfn - slot->base_gfn];
 520
 521         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 522                 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 523
 524         return &slot->lpage_info[level - 2][idx].rmap_pde;
 525 }
 526
 527 /*
 528  * Reverse mapping data structures:
 529  *
 530  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 531  * that points to page_address(page).
 532  *
 533  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 534  * containing more mappings.
 535  *
 536  * Returns the number of rmap entries before the spte was added or zero if
 537  * the spte was not added.
 538  *
 539  */
 540 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 541 {
 542         struct kvm_mmu_page *sp;
 543         struct kvm_rmap_desc *desc;
 544         unsigned long *rmapp;
 545         int i, count = 0;
 546
 547         if (!is_rmap_spte(*spte))
 548                 return count;
 549         gfn = unalias_gfn(vcpu->kvm, gfn);
 550         sp = page_header(__pa(spte));
 551         sp->gfns[spte - sp->spt] = gfn;
 552         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 553         if (!*rmapp) {
 554                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 555                 *rmapp = (unsigned long)spte;
 556         } else if (!(*rmapp & 1)) {
 557                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 558                 desc = mmu_alloc_rmap_desc(vcpu);
 559                 desc->sptes[0] = (u64 *)*rmapp;
 560                 desc->sptes[1] = spte;
 561                 *rmapp = (unsigned long)desc | 1;
 562         } else {
 563                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 564                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 565                 while (desc->sptes[RMAP_EXT-1] && desc->more) {
 566                         desc = desc->more;
 567                         count += RMAP_EXT;
 568                 }
 569                 if (desc->sptes[RMAP_EXT-1]) {
 570                         desc->more = mmu_alloc_rmap_desc(vcpu);
 571                         desc = desc->more;
 572                 }
 573                 for (i = 0; desc->sptes[i]; ++i)
 574                         ;
 575                 desc->sptes[i] = spte;
 576         }
 577         return count;
 578 }
 579
 580 static void rmap_desc_remove_entry(unsigned long *rmapp,
 581                                    struct kvm_rmap_desc *desc,
 582                                    int i,
 583                                    struct kvm_rmap_desc *prev_desc)
 584 {
 585         int j;
 586
 587         for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
 588                 ;
 589         desc->sptes[i] = desc->sptes[j];
 590         desc->sptes[j] = NULL;
 591         if (j != 0)
 592                 return;
 593         if (!prev_desc && !desc->more)
 594                 *rmapp = (unsigned long)desc->sptes[0];
 595         else
 596                 if (prev_desc)
 597                         prev_desc->more = desc->more;
 598                 else
 599                         *rmapp = (unsigned long)desc->more | 1;
 600         mmu_free_rmap_desc(desc);
 601 }
 602
 603 static void rmap_remove(struct kvm *kvm, u64 *spte)
 604 {
 605         struct kvm_rmap_desc *desc;
 606         struct kvm_rmap_desc *prev_desc;
 607         struct kvm_mmu_page *sp;
 608         pfn_t pfn;
 609         unsigned long *rmapp;
 610         int i;
 611
 612         if (!is_rmap_spte(*spte))
 613                 return;
 614         sp = page_header(__pa(spte));
 615         pfn = spte_to_pfn(*spte);
 616         if (*spte & shadow_accessed_mask)
 617                 kvm_set_pfn_accessed(pfn);
 618         if (is_writable_pte(*spte))
 619                 kvm_set_pfn_dirty(pfn);
 620         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], sp->role.level);
 621         if (!*rmapp) {
 622                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 623                 BUG();
 624         } else if (!(*rmapp & 1)) {
 625                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 626                 if ((u64 *)*rmapp != spte) {
 627                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 628                                spte, *spte);
 629                         BUG();
 630                 }
 631                 *rmapp = 0;
 632         } else {
 633                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 634                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 635                 prev_desc = NULL;
 636                 while (desc) {
 637                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
 638                                 if (desc->sptes[i] == spte) {
 639                                         rmap_desc_remove_entry(rmapp,
 640                                                                desc, i,
 641                                                                prev_desc);
 642                                         return;
 643                                 }
 644                         prev_desc = desc;
 645                         desc = desc->more;
 646                 }
 647                 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
 648                 BUG();
 649         }
 650 }
 651
 652 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 653 {
 654         struct kvm_rmap_desc *desc;
 655         struct kvm_rmap_desc *prev_desc;
 656         u64 *prev_spte;
 657         int i;
 658
 659         if (!*rmapp)
 660                 return NULL;
 661         else if (!(*rmapp & 1)) {
 662                 if (!spte)
 663                         return (u64 *)*rmapp;
 664                 return NULL;
 665         }
 666         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 667         prev_desc = NULL;
 668         prev_spte = NULL;
 669         while (desc) {
 670                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
 671                         if (prev_spte == spte)
 672                                 return desc->sptes[i];
 673                         prev_spte = desc->sptes[i];
 674                 }
 675                 desc = desc->more;
 676         }
 677         return NULL;
 678 }
 679
 680 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 681 {
 682         unsigned long *rmapp;
 683         u64 *spte;
 684         int i, write_protected = 0;
 685
 686         gfn = unalias_gfn(kvm, gfn);
 687         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
 688
 689         spte = rmap_next(kvm, rmapp, NULL);
 690         while (spte) {
 691                 BUG_ON(!spte);
 692                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 693                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 694                 if (is_writable_pte(*spte)) {
 695                         __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
 696                         write_protected = 1;
 697                 }
 698                 spte = rmap_next(kvm, rmapp, spte);
 699         }
 700         if (write_protected) {
 701                 pfn_t pfn;
 702
 703                 spte = rmap_next(kvm, rmapp, NULL);
 704                 pfn = spte_to_pfn(*spte);
 705                 kvm_set_pfn_dirty(pfn);
 706         }
 707
 708         /* check for huge page mappings */
 709         for (i = PT_DIRECTORY_LEVEL;
 710              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 711                 rmapp = gfn_to_rmap(kvm, gfn, i);
 712                 spte = rmap_next(kvm, rmapp, NULL);
 713                 while (spte) {
 714                         BUG_ON(!spte);
 715                         BUG_ON(!(*spte & PT_PRESENT_MASK));
 716                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 717                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 718                         if (is_writable_pte(*spte)) {
 719                                 rmap_remove(kvm, spte);
 720                                 --kvm->stat.lpages;
 721                                 __set_spte(spte, shadow_trap_nonpresent_pte);
 722                                 spte = NULL;
 723                                 write_protected = 1;
 724                         }
 725                         spte = rmap_next(kvm, rmapp, spte);
 726                 }
 727         }
 728
 729         return write_protected;
 730 }
 731
 732 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 733                            unsigned long data)
 734 {
 735         u64 *spte;
 736         int need_tlb_flush = 0;
 737
 738         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 739                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 740                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 741                 rmap_remove(kvm, spte);
 742                 __set_spte(spte, shadow_trap_nonpresent_pte);
 743                 need_tlb_flush = 1;
 744         }
 745         return need_tlb_flush;
 746 }
 747
 748 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
 749                              unsigned long data)
 750 {
 751         int need_flush = 0;
 752         u64 *spte, new_spte;
 753         pte_t *ptep = (pte_t *)data;
 754         pfn_t new_pfn;
 755
 756         WARN_ON(pte_huge(*ptep));
 757         new_pfn = pte_pfn(*ptep);
 758         spte = rmap_next(kvm, rmapp, NULL);
 759         while (spte) {
 760                 BUG_ON(!is_shadow_present_pte(*spte));
 761                 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
 762                 need_flush = 1;
 763                 if (pte_write(*ptep)) {
 764                         rmap_remove(kvm, spte);
 765                         __set_spte(spte, shadow_trap_nonpresent_pte);
 766                         spte = rmap_next(kvm, rmapp, NULL);
 767                 } else {
 768                         new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
 769                         new_spte |= (u64)new_pfn << PAGE_SHIFT;
 770
 771                         new_spte &= ~PT_WRITABLE_MASK;
 772                         new_spte &= ~SPTE_HOST_WRITEABLE;
 773                         if (is_writable_pte(*spte))
 774                                 kvm_set_pfn_dirty(spte_to_pfn(*spte));
 775                         __set_spte(spte, new_spte);
 776                         spte = rmap_next(kvm, rmapp, spte);
 777                 }
 778         }
 779         if (need_flush)
 780                 kvm_flush_remote_tlbs(kvm);
 781
 782         return 0;
 783 }
 784
 785 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 786                           unsigned long data,
 787                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 788                                          unsigned long data))
 789 {
 790         int i, j;
 791         int ret;
 792         int retval = 0;
 793         struct kvm_memslots *slots;
 794
 795         slots = rcu_dereference(kvm->memslots);
 796
 797         for (i = 0; i < slots->nmemslots; i++) {
 798                 struct kvm_memory_slot *memslot = &slots->memslots[i];
 799                 unsigned long start = memslot->userspace_addr;
 800                 unsigned long end;
 801
 802                 end = start + (memslot->npages << PAGE_SHIFT);
 803                 if (hva >= start && hva < end) {
 804                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 805
 806                         ret = handler(kvm, &memslot->rmap[gfn_offset], data);
 807
 808                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
 809                                 int idx = gfn_offset;
 810                                 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
 811                                 ret |= handler(kvm,
 812                                         &memslot->lpage_info[j][idx].rmap_pde,
 813                                         data);
 814                         }
 815                         trace_kvm_age_page(hva, memslot, ret);
 816                         retval |= ret;
 817                 }
 818         }
 819
 820         return retval;
 821 }
 822
 823 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 824 {
 825         return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
 826 }
 827
 828 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
 829 {
 830         kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
 831 }
 832
 833 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 834                          unsigned long data)
 835 {
 836         u64 *spte;
 837         int young = 0;
 838
 839         /*
 840          * Emulate the accessed bit for EPT, by checking if this page has
 841          * an EPT mapping, and clearing it if it does. On the next access,
 842          * a new EPT mapping will be established.
 843          * This has some overhead, but not as much as the cost of swapping
 844          * out actively used pages or breaking up actively used hugepages.
 845          */
 846         if (!shadow_accessed_mask)
 847                 return kvm_unmap_rmapp(kvm, rmapp, data);
 848
 849         spte = rmap_next(kvm, rmapp, NULL);
 850         while (spte) {
 851                 int _young;
 852                 u64 _spte = *spte;
 853                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 854                 _young = _spte & PT_ACCESSED_MASK;
 855                 if (_young) {
 856                         young = 1;
 857                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 858                 }
 859                 spte = rmap_next(kvm, rmapp, spte);
 860         }
 861         return young;
 862 }
 863
 864 #define RMAP_RECYCLE_THRESHOLD 1000
 865
 866 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 867 {
 868         unsigned long *rmapp;
 869         struct kvm_mmu_page *sp;
 870
 871         sp = page_header(__pa(spte));
 872
 873         gfn = unalias_gfn(vcpu->kvm, gfn);
 874         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 875
 876         kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
 877         kvm_flush_remote_tlbs(vcpu->kvm);
 878 }
 879
 880 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 881 {
 882         return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
 883 }
 884
 885 #ifdef MMU_DEBUG
 886 static int is_empty_shadow_page(u64 *spt)
 887 {
 888         u64 *pos;
 889         u64 *end;
 890
 891         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 892                 if (is_shadow_present_pte(*pos)) {
 893                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 894                                pos, *pos);
 895                         return 0;
 896                 }
 897         return 1;
 898 }
 899 #endif
 900
 901 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 902 {
 903         ASSERT(is_empty_shadow_page(sp->spt));
 904         list_del(&sp->link);
 905         __free_page(virt_to_page(sp->spt));
 906         __free_page(virt_to_page(sp->gfns));
 907         kfree(sp);
 908         ++kvm->arch.n_free_mmu_pages;
 909 }
 910
 911 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 912 {
 913         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 914 }
 915
 916 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 917                                                u64 *parent_pte)
 918 {
 919         struct kvm_mmu_page *sp;
 920
 921         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 922         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 923         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 924         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 925         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 926         INIT_LIST_HEAD(&sp->oos_link);
 927         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 928         sp->multimapped = 0;
 929         sp->parent_pte = parent_pte;
 930         --vcpu->kvm->arch.n_free_mmu_pages;
 931         return sp;
 932 }
 933
 934 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 935                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 936 {
 937         struct kvm_pte_chain *pte_chain;
 938         struct hlist_node *node;
 939         int i;
 940
 941         if (!parent_pte)
 942                 return;
 943         if (!sp->multimapped) {
 944                 u64 *old = sp->parent_pte;
 945
 946                 if (!old) {
 947                         sp->parent_pte = parent_pte;
 948                         return;
 949                 }
 950                 sp->multimapped = 1;
 951                 pte_chain = mmu_alloc_pte_chain(vcpu);
 952                 INIT_HLIST_HEAD(&sp->parent_ptes);
 953                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 954                 pte_chain->parent_ptes[0] = old;
 955         }
 956         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 957                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 958                         continue;
 959                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 960                         if (!pte_chain->parent_ptes[i]) {
 961                                 pte_chain->parent_ptes[i] = parent_pte;
 962                                 return;
 963                         }
 964         }
 965         pte_chain = mmu_alloc_pte_chain(vcpu);
 966         BUG_ON(!pte_chain);
 967         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 968         pte_chain->parent_ptes[0] = parent_pte;
 969 }
 970
 971 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 972                                        u64 *parent_pte)
 973 {
 974         struct kvm_pte_chain *pte_chain;
 975         struct hlist_node *node;
 976         int i;
 977
 978         if (!sp->multimapped) {
 979                 BUG_ON(sp->parent_pte != parent_pte);
 980                 sp->parent_pte = NULL;
 981                 return;
 982         }
 983         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 984                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 985                         if (!pte_chain->parent_ptes[i])
 986                                 break;
 987                         if (pte_chain->parent_ptes[i] != parent_pte)
 988                                 continue;
 989                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 990                                 && pte_chain->parent_ptes[i + 1]) {
 991                                 pte_chain->parent_ptes[i]
 992                                         = pte_chain->parent_ptes[i + 1];
 993                                 ++i;
 994                         }
 995                         pte_chain->parent_ptes[i] = NULL;
 996                         if (i == 0) {
 997                                 hlist_del(&pte_chain->link);
 998                                 mmu_free_pte_chain(pte_chain);
 999                                 if (hlist_empty(&sp->parent_ptes)) {
1000                                         sp->multimapped = 0;
1001                                         sp->parent_pte = NULL;
1002                                 }
1003                         }
1004                         return;
1005                 }
1006         BUG();
1007 }
1008
1009
1010 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1011                             mmu_parent_walk_fn fn)
1012 {
1013         struct kvm_pte_chain *pte_chain;
1014         struct hlist_node *node;
1015         struct kvm_mmu_page *parent_sp;
1016         int i;
1017
1018         if (!sp->multimapped && sp->parent_pte) {
1019                 parent_sp = page_header(__pa(sp->parent_pte));
1020                 fn(vcpu, parent_sp);
1021                 mmu_parent_walk(vcpu, parent_sp, fn);
1022                 return;
1023         }
1024         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1025                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1026                         if (!pte_chain->parent_ptes[i])
1027                                 break;
1028                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
1029                         fn(vcpu, parent_sp);
1030                         mmu_parent_walk(vcpu, parent_sp, fn);
1031                 }
1032 }
1033
1034 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1035 {
1036         unsigned int index;
1037         struct kvm_mmu_page *sp = page_header(__pa(spte));
1038
1039         index = spte - sp->spt;
1040         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1041                 sp->unsync_children++;
1042         WARN_ON(!sp->unsync_children);
1043 }
1044
1045 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1046 {
1047         struct kvm_pte_chain *pte_chain;
1048         struct hlist_node *node;
1049         int i;
1050
1051         if (!sp->parent_pte)
1052                 return;
1053
1054         if (!sp->multimapped) {
1055                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1056                 return;
1057         }
1058
1059         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1060                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1061                         if (!pte_chain->parent_ptes[i])
1062                                 break;
1063                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1064                 }
1065 }
1066
1067 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1068 {
1069         kvm_mmu_update_parents_unsync(sp);
1070         return 1;
1071 }
1072
1073 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
1074                                         struct kvm_mmu_page *sp)
1075 {
1076         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
1077         kvm_mmu_update_parents_unsync(sp);
1078 }
1079
1080 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1081                                     struct kvm_mmu_page *sp)
1082 {
1083         int i;
1084
1085         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1086                 sp->spt[i] = shadow_trap_nonpresent_pte;
1087 }
1088
1089 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1090                                struct kvm_mmu_page *sp)
1091 {
1092         return 1;
1093 }
1094
1095 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1096 {
1097 }
1098
1099 #define KVM_PAGE_ARRAY_NR 16
1100
1101 struct kvm_mmu_pages {
1102         struct mmu_page_and_offset {
1103                 struct kvm_mmu_page *sp;
1104                 unsigned int idx;
1105         } page[KVM_PAGE_ARRAY_NR];
1106         unsigned int nr;
1107 };
1108
1109 #define for_each_unsync_children(bitmap, idx)           \
1110         for (idx = find_first_bit(bitmap, 512);         \
1111              idx < 512;                                 \
1112              idx = find_next_bit(bitmap, 512, idx+1))
1113
1114 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1115                          int idx)
1116 {
1117         int i;
1118
1119         if (sp->unsync)
1120                 for (i=0; i < pvec->nr; i++)
1121                         if (pvec->page[i].sp == sp)
1122                                 return 0;
1123
1124         pvec->page[pvec->nr].sp = sp;
1125         pvec->page[pvec->nr].idx = idx;
1126         pvec->nr++;
1127         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1128 }
1129
1130 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1131                            struct kvm_mmu_pages *pvec)
1132 {
1133         int i, ret, nr_unsync_leaf = 0;
1134
1135         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1136                 u64 ent = sp->spt[i];
1137
1138                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1139                         struct kvm_mmu_page *child;
1140                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1141
1142                         if (child->unsync_children) {
1143                                 if (mmu_pages_add(pvec, child, i))
1144                                         return -ENOSPC;
1145
1146                                 ret = __mmu_unsync_walk(child, pvec);
1147                                 if (!ret)
1148                                         __clear_bit(i, sp->unsync_child_bitmap);
1149                                 else if (ret > 0)
1150                                         nr_unsync_leaf += ret;
1151                                 else
1152                                         return ret;
1153                         }
1154
1155                         if (child->unsync) {
1156                                 nr_unsync_leaf++;
1157                                 if (mmu_pages_add(pvec, child, i))
1158                                         return -ENOSPC;
1159                         }
1160                 }
1161         }
1162
1163         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1164                 sp->unsync_children = 0;
1165
1166         return nr_unsync_leaf;
1167 }
1168
1169 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1170                            struct kvm_mmu_pages *pvec)
1171 {
1172         if (!sp->unsync_children)
1173                 return 0;
1174
1175         mmu_pages_add(pvec, sp, 0);
1176         return __mmu_unsync_walk(sp, pvec);
1177 }
1178
1179 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1180 {
1181         unsigned index;
1182         struct hlist_head *bucket;
1183         struct kvm_mmu_page *sp;
1184         struct hlist_node *node;
1185
1186         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1187         index = kvm_page_table_hashfn(gfn);
1188         bucket = &kvm->arch.mmu_page_hash[index];
1189         hlist_for_each_entry(sp, node, bucket, hash_link)
1190                 if (sp->gfn == gfn && !sp->role.direct
1191                     && !sp->role.invalid) {
1192                         pgprintk("%s: found role %x\n",
1193                                  __func__, sp->role.word);
1194                         return sp;
1195                 }
1196         return NULL;
1197 }
1198
1199 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1200 {
1201         WARN_ON(!sp->unsync);
1202         sp->unsync = 0;
1203         --kvm->stat.mmu_unsync;
1204 }
1205
1206 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1207
1208 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1209 {
1210         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1211                 kvm_mmu_zap_page(vcpu->kvm, sp);
1212                 return 1;
1213         }
1214
1215         trace_kvm_mmu_sync_page(sp);
1216         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1217                 kvm_flush_remote_tlbs(vcpu->kvm);
1218         kvm_unlink_unsync_page(vcpu->kvm, sp);
1219         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1220                 kvm_mmu_zap_page(vcpu->kvm, sp);
1221                 return 1;
1222         }
1223
1224         kvm_mmu_flush_tlb(vcpu);
1225         return 0;
1226 }
1227
1228 struct mmu_page_path {
1229         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1230         unsigned int idx[PT64_ROOT_LEVEL-1];
1231 };
1232
1233 #define for_each_sp(pvec, sp, parents, i)                       \
1234                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1235                         sp = pvec.page[i].sp;                   \
1236                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1237                         i = mmu_pages_next(&pvec, &parents, i))
1238
1239 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1240                           struct mmu_page_path *parents,
1241                           int i)
1242 {
1243         int n;
1244
1245         for (n = i+1; n < pvec->nr; n++) {
1246                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1247
1248                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1249                         parents->idx[0] = pvec->page[n].idx;
1250                         return n;
1251                 }
1252
1253                 parents->parent[sp->role.level-2] = sp;
1254                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1255         }
1256
1257         return n;
1258 }
1259
1260 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1261 {
1262         struct kvm_mmu_page *sp;
1263         unsigned int level = 0;
1264
1265         do {
1266                 unsigned int idx = parents->idx[level];
1267
1268                 sp = parents->parent[level];
1269                 if (!sp)
1270                         return;
1271
1272                 --sp->unsync_children;
1273                 WARN_ON((int)sp->unsync_children < 0);
1274                 __clear_bit(idx, sp->unsync_child_bitmap);
1275                 level++;
1276         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1277 }
1278
1279 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1280                                struct mmu_page_path *parents,
1281                                struct kvm_mmu_pages *pvec)
1282 {
1283         parents->parent[parent->role.level-1] = NULL;
1284         pvec->nr = 0;
1285 }
1286
1287 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1288                               struct kvm_mmu_page *parent)
1289 {
1290         int i;
1291         struct kvm_mmu_page *sp;
1292         struct mmu_page_path parents;
1293         struct kvm_mmu_pages pages;
1294
1295         kvm_mmu_pages_init(parent, &parents, &pages);
1296         while (mmu_unsync_walk(parent, &pages)) {
1297                 int protected = 0;
1298
1299                 for_each_sp(pages, sp, parents, i)
1300                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1301
1302                 if (protected)
1303                         kvm_flush_remote_tlbs(vcpu->kvm);
1304
1305                 for_each_sp(pages, sp, parents, i) {
1306                         kvm_sync_page(vcpu, sp);
1307                         mmu_pages_clear_parents(&parents);
1308                 }
1309                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1310                 kvm_mmu_pages_init(parent, &parents, &pages);
1311         }
1312 }
1313
1314 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1315                                              gfn_t gfn,
1316                                              gva_t gaddr,
1317                                              unsigned level,
1318                                              int direct,
1319                                              unsigned access,
1320                                              u64 *parent_pte)
1321 {
1322         union kvm_mmu_page_role role;
1323         unsigned index;
1324         unsigned quadrant;
1325         struct hlist_head *bucket;
1326         struct kvm_mmu_page *sp;
1327         struct hlist_node *node, *tmp;
1328
1329         role = vcpu->arch.mmu.base_role;
1330         role.level = level;
1331         role.direct = direct;
1332         if (role.direct)
1333                 role.glevels = 0;
1334         role.access = access;
1335         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1336                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1337                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1338                 role.quadrant = quadrant;
1339         }
1340         index = kvm_page_table_hashfn(gfn);
1341         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1342         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1343                 if (sp->gfn == gfn) {
1344                         if (sp->unsync)
1345                                 if (kvm_sync_page(vcpu, sp))
1346                                         continue;
1347
1348                         if (sp->role.word != role.word)
1349                                 continue;
1350
1351                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1352                         if (sp->unsync_children) {
1353                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1354                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1355                         }
1356                         trace_kvm_mmu_get_page(sp, false);
1357                         return sp;
1358                 }
1359         ++vcpu->kvm->stat.mmu_cache_miss;
1360         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1361         if (!sp)
1362                 return sp;
1363         sp->gfn = gfn;
1364         sp->role = role;
1365         hlist_add_head(&sp->hash_link, bucket);
1366         if (!direct) {
1367                 if (rmap_write_protect(vcpu->kvm, gfn))
1368                         kvm_flush_remote_tlbs(vcpu->kvm);
1369                 account_shadowed(vcpu->kvm, gfn);
1370         }
1371         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1372                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1373         else
1374                 nonpaging_prefetch_page(vcpu, sp);
1375         trace_kvm_mmu_get_page(sp, true);
1376         return sp;
1377 }
1378
1379 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1380                              struct kvm_vcpu *vcpu, u64 addr)
1381 {
1382         iterator->addr = addr;
1383         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1384         iterator->level = vcpu->arch.mmu.shadow_root_level;
1385         if (iterator->level == PT32E_ROOT_LEVEL) {
1386                 iterator->shadow_addr
1387                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1388                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1389                 --iterator->level;
1390                 if (!iterator->shadow_addr)
1391                         iterator->level = 0;
1392         }
1393 }
1394
1395 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1396 {
1397         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1398                 return false;
1399
1400         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1401                 if (is_large_pte(*iterator->sptep))
1402                         return false;
1403
1404         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1405         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1406         return true;
1407 }
1408
1409 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1410 {
1411         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1412         --iterator->level;
1413 }
1414
1415 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1416                                          struct kvm_mmu_page *sp)
1417 {
1418         unsigned i;
1419         u64 *pt;
1420         u64 ent;
1421
1422         pt = sp->spt;
1423
1424         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1425                 ent = pt[i];
1426
1427                 if (is_shadow_present_pte(ent)) {
1428                         if (!is_last_spte(ent, sp->role.level)) {
1429                                 ent &= PT64_BASE_ADDR_MASK;
1430                                 mmu_page_remove_parent_pte(page_header(ent),
1431                                                            &pt[i]);
1432                         } else {
1433                                 if (is_large_pte(ent))
1434                                         --kvm->stat.lpages;
1435                                 rmap_remove(kvm, &pt[i]);
1436                         }
1437                 }
1438                 pt[i] = shadow_trap_nonpresent_pte;
1439         }
1440 }
1441
1442 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1443 {
1444         mmu_page_remove_parent_pte(sp, parent_pte);
1445 }
1446
1447 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1448 {
1449         int i;
1450         struct kvm_vcpu *vcpu;
1451
1452         kvm_for_each_vcpu(i, vcpu, kvm)
1453                 vcpu->arch.last_pte_updated = NULL;
1454 }
1455
1456 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1457 {
1458         u64 *parent_pte;
1459
1460         while (sp->multimapped || sp->parent_pte) {
1461                 if (!sp->multimapped)
1462                         parent_pte = sp->parent_pte;
1463                 else {
1464                         struct kvm_pte_chain *chain;
1465
1466                         chain = container_of(sp->parent_ptes.first,
1467                                              struct kvm_pte_chain, link);
1468                         parent_pte = chain->parent_ptes[0];
1469                 }
1470                 BUG_ON(!parent_pte);
1471                 kvm_mmu_put_page(sp, parent_pte);
1472                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1473         }
1474 }
1475
1476 static int mmu_zap_unsync_children(struct kvm *kvm,
1477                                    struct kvm_mmu_page *parent)
1478 {
1479         int i, zapped = 0;
1480         struct mmu_page_path parents;
1481         struct kvm_mmu_pages pages;
1482
1483         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1484                 return 0;
1485
1486         kvm_mmu_pages_init(parent, &parents, &pages);
1487         while (mmu_unsync_walk(parent, &pages)) {
1488                 struct kvm_mmu_page *sp;
1489
1490                 for_each_sp(pages, sp, parents, i) {
1491                         kvm_mmu_zap_page(kvm, sp);
1492                         mmu_pages_clear_parents(&parents);
1493                         zapped++;
1494                 }
1495                 kvm_mmu_pages_init(parent, &parents, &pages);
1496         }
1497
1498         return zapped;
1499 }
1500
1501 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1502 {
1503         int ret;
1504
1505         trace_kvm_mmu_zap_page(sp);
1506         ++kvm->stat.mmu_shadow_zapped;
1507         ret = mmu_zap_unsync_children(kvm, sp);
1508         kvm_mmu_page_unlink_children(kvm, sp);
1509         kvm_mmu_unlink_parents(kvm, sp);
1510         kvm_flush_remote_tlbs(kvm);
1511         if (!sp->role.invalid && !sp->role.direct)
1512                 unaccount_shadowed(kvm, sp->gfn);
1513         if (sp->unsync)
1514                 kvm_unlink_unsync_page(kvm, sp);
1515         if (!sp->root_count) {
1516                 hlist_del(&sp->hash_link);
1517                 kvm_mmu_free_page(kvm, sp);
1518         } else {
1519                 sp->role.invalid = 1;
1520                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1521                 kvm_reload_remote_mmus(kvm);
1522         }
1523         kvm_mmu_reset_last_pte_updated(kvm);
1524         return ret;
1525 }
1526
1527 /*
1528  * Changing the number of mmu pages allocated to the vm
1529  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1530  */
1531 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1532 {
1533         int used_pages;
1534
1535         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1536         used_pages = max(0, used_pages);
1537
1538         /*
1539          * If we set the number of mmu pages to be smaller be than the
1540          * number of actived pages , we must to free some mmu pages before we
1541          * change the value
1542          */
1543
1544         if (used_pages > kvm_nr_mmu_pages) {
1545                 while (used_pages > kvm_nr_mmu_pages &&
1546                         !list_empty(&kvm->arch.active_mmu_pages)) {
1547                         struct kvm_mmu_page *page;
1548
1549                         page = container_of(kvm->arch.active_mmu_pages.prev,
1550                                             struct kvm_mmu_page, link);
1551                         used_pages -= kvm_mmu_zap_page(kvm, page);
1552                         used_pages--;
1553                 }
1554                 kvm_nr_mmu_pages = used_pages;
1555                 kvm->arch.n_free_mmu_pages = 0;
1556         }
1557         else
1558                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1559                                          - kvm->arch.n_alloc_mmu_pages;
1560
1561         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1562 }
1563
1564 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1565 {
1566         unsigned index;
1567         struct hlist_head *bucket;
1568         struct kvm_mmu_page *sp;
1569         struct hlist_node *node, *n;
1570         int r;
1571
1572         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1573         r = 0;
1574         index = kvm_page_table_hashfn(gfn);
1575         bucket = &kvm->arch.mmu_page_hash[index];
1576         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1577                 if (sp->gfn == gfn && !sp->role.direct) {
1578                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1579                                  sp->role.word);
1580                         r = 1;
1581                         if (kvm_mmu_zap_page(kvm, sp))
1582                                 n = bucket->first;
1583                 }
1584         return r;
1585 }
1586
1587 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1588 {
1589         unsigned index;
1590         struct hlist_head *bucket;
1591         struct kvm_mmu_page *sp;
1592         struct hlist_node *node, *nn;
1593
1594         index = kvm_page_table_hashfn(gfn);
1595         bucket = &kvm->arch.mmu_page_hash[index];
1596         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1597                 if (sp->gfn == gfn && !sp->role.direct
1598                     && !sp->role.invalid) {
1599                         pgprintk("%s: zap %lx %x\n",
1600                                  __func__, gfn, sp->role.word);
1601                         if (kvm_mmu_zap_page(kvm, sp))
1602                                 nn = bucket->first;
1603                 }
1604         }
1605 }
1606
1607 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1608 {
1609         int slot = memslot_id(kvm, gfn);
1610         struct kvm_mmu_page *sp = page_header(__pa(pte));
1611
1612         __set_bit(slot, sp->slot_bitmap);
1613 }
1614
1615 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1616 {
1617         int i;
1618         u64 *pt = sp->spt;
1619
1620         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1621                 return;
1622
1623         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1624                 if (pt[i] == shadow_notrap_nonpresent_pte)
1625                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1626         }
1627 }
1628
1629 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1630 {
1631         struct page *page;
1632
1633         gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
1634
1635         if (gpa == UNMAPPED_GVA)
1636                 return NULL;
1637
1638         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1639
1640         return page;
1641 }
1642
1643 /*
1644  * The function is based on mtrr_type_lookup() in
1645  * arch/x86/kernel/cpu/mtrr/generic.c
1646  */
1647 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1648                          u64 start, u64 end)
1649 {
1650         int i;
1651         u64 base, mask;
1652         u8 prev_match, curr_match;
1653         int num_var_ranges = KVM_NR_VAR_MTRR;
1654
1655         if (!mtrr_state->enabled)
1656                 return 0xFF;
1657
1658         /* Make end inclusive end, instead of exclusive */
1659         end--;
1660
1661         /* Look in fixed ranges. Just return the type as per start */
1662         if (mtrr_state->have_fixed && (start < 0x100000)) {
1663                 int idx;
1664
1665                 if (start < 0x80000) {
1666                         idx = 0;
1667                         idx += (start >> 16);
1668                         return mtrr_state->fixed_ranges[idx];
1669                 } else if (start < 0xC0000) {
1670                         idx = 1 * 8;
1671                         idx += ((start - 0x80000) >> 14);
1672                         return mtrr_state->fixed_ranges[idx];
1673                 } else if (start < 0x1000000) {
1674                         idx = 3 * 8;
1675                         idx += ((start - 0xC0000) >> 12);
1676                         return mtrr_state->fixed_ranges[idx];
1677                 }
1678         }
1679
1680         /*
1681          * Look in variable ranges
1682          * Look of multiple ranges matching this address and pick type
1683          * as per MTRR precedence
1684          */
1685         if (!(mtrr_state->enabled & 2))
1686                 return mtrr_state->def_type;
1687
1688         prev_match = 0xFF;
1689         for (i = 0; i < num_var_ranges; ++i) {
1690                 unsigned short start_state, end_state;
1691
1692                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1693                         continue;
1694
1695                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1696                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1697                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1698                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1699
1700                 start_state = ((start & mask) == (base & mask));
1701                 end_state = ((end & mask) == (base & mask));
1702                 if (start_state != end_state)
1703                         return 0xFE;
1704
1705                 if ((start & mask) != (base & mask))
1706                         continue;
1707
1708                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1709                 if (prev_match == 0xFF) {
1710                         prev_match = curr_match;
1711                         continue;
1712                 }
1713
1714                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1715                     curr_match == MTRR_TYPE_UNCACHABLE)
1716                         return MTRR_TYPE_UNCACHABLE;
1717
1718                 if ((prev_match == MTRR_TYPE_WRBACK &&
1719                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1720                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1721                      curr_match == MTRR_TYPE_WRBACK)) {
1722                         prev_match = MTRR_TYPE_WRTHROUGH;
1723                         curr_match = MTRR_TYPE_WRTHROUGH;
1724                 }
1725
1726                 if (prev_match != curr_match)
1727                         return MTRR_TYPE_UNCACHABLE;
1728         }
1729
1730         if (prev_match != 0xFF)
1731                 return prev_match;
1732
1733         return mtrr_state->def_type;
1734 }
1735
1736 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1737 {
1738         u8 mtrr;
1739
1740         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1741                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1742         if (mtrr == 0xfe || mtrr == 0xff)
1743                 mtrr = MTRR_TYPE_WRBACK;
1744         return mtrr;
1745 }
1746 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1747
1748 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1749 {
1750         unsigned index;
1751         struct hlist_head *bucket;
1752         struct kvm_mmu_page *s;
1753         struct hlist_node *node, *n;
1754
1755         trace_kvm_mmu_unsync_page(sp);
1756         index = kvm_page_table_hashfn(sp->gfn);
1757         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1758         /* don't unsync if pagetable is shadowed with multiple roles */
1759         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1760                 if (s->gfn != sp->gfn || s->role.direct)
1761                         continue;
1762                 if (s->role.word != sp->role.word)
1763                         return 1;
1764         }
1765         ++vcpu->kvm->stat.mmu_unsync;
1766         sp->unsync = 1;
1767
1768         kvm_mmu_mark_parents_unsync(vcpu, sp);
1769
1770         mmu_convert_notrap(sp);
1771         return 0;
1772 }
1773
1774 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1775                                   bool can_unsync)
1776 {
1777         struct kvm_mmu_page *shadow;
1778
1779         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1780         if (shadow) {
1781                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1782                         return 1;
1783                 if (shadow->unsync)
1784                         return 0;
1785                 if (can_unsync && oos_shadow)
1786                         return kvm_unsync_page(vcpu, shadow);
1787                 return 1;
1788         }
1789         return 0;
1790 }
1791
1792 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1793                     unsigned pte_access, int user_fault,
1794                     int write_fault, int dirty, int level,
1795                     gfn_t gfn, pfn_t pfn, bool speculative,
1796                     bool can_unsync, bool reset_host_protection)
1797 {
1798         u64 spte;
1799         int ret = 0;
1800
1801         /*
1802          * We don't set the accessed bit, since we sometimes want to see
1803          * whether the guest actually used the pte (in order to detect
1804          * demand paging).
1805          */
1806         spte = shadow_base_present_pte | shadow_dirty_mask;
1807         if (!speculative)
1808                 spte |= shadow_accessed_mask;
1809         if (!dirty)
1810                 pte_access &= ~ACC_WRITE_MASK;
1811         if (pte_access & ACC_EXEC_MASK)
1812                 spte |= shadow_x_mask;
1813         else
1814                 spte |= shadow_nx_mask;
1815         if (pte_access & ACC_USER_MASK)
1816                 spte |= shadow_user_mask;
1817         if (level > PT_PAGE_TABLE_LEVEL)
1818                 spte |= PT_PAGE_SIZE_MASK;
1819         if (tdp_enabled)
1820                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1821                         kvm_is_mmio_pfn(pfn));
1822
1823         if (reset_host_protection)
1824                 spte |= SPTE_HOST_WRITEABLE;
1825
1826         spte |= (u64)pfn << PAGE_SHIFT;
1827
1828         if ((pte_access & ACC_WRITE_MASK)
1829             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1830
1831                 if (level > PT_PAGE_TABLE_LEVEL &&
1832                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1833                         ret = 1;
1834                         spte = shadow_trap_nonpresent_pte;
1835                         goto set_pte;
1836                 }
1837
1838                 spte |= PT_WRITABLE_MASK;
1839
1840                 /*
1841                  * Optimization: for pte sync, if spte was writable the hash
1842                  * lookup is unnecessary (and expensive). Write protection
1843                  * is responsibility of mmu_get_page / kvm_sync_page.
1844                  * Same reasoning can be applied to dirty page accounting.
1845                  */
1846                 if (!can_unsync && is_writable_pte(*sptep))
1847                         goto set_pte;
1848
1849                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1850                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1851                                  __func__, gfn);
1852                         ret = 1;
1853                         pte_access &= ~ACC_WRITE_MASK;
1854                         if (is_writable_pte(spte))
1855                                 spte &= ~PT_WRITABLE_MASK;
1856                 }
1857         }
1858
1859         if (pte_access & ACC_WRITE_MASK)
1860                 mark_page_dirty(vcpu->kvm, gfn);
1861
1862 set_pte:
1863         __set_spte(sptep, spte);
1864         return ret;
1865 }
1866
1867 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1868                          unsigned pt_access, unsigned pte_access,
1869                          int user_fault, int write_fault, int dirty,
1870                          int *ptwrite, int level, gfn_t gfn,
1871                          pfn_t pfn, bool speculative,
1872                          bool reset_host_protection)
1873 {
1874         int was_rmapped = 0;
1875         int was_writable = is_writable_pte(*sptep);
1876         int rmap_count;
1877
1878         pgprintk("%s: spte %llx access %x write_fault %d"
1879                  " user_fault %d gfn %lx\n",
1880                  __func__, *sptep, pt_access,
1881                  write_fault, user_fault, gfn);
1882
1883         if (is_rmap_spte(*sptep)) {
1884                 /*
1885                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1886                  * the parent of the now unreachable PTE.
1887                  */
1888                 if (level > PT_PAGE_TABLE_LEVEL &&
1889                     !is_large_pte(*sptep)) {
1890                         struct kvm_mmu_page *child;
1891                         u64 pte = *sptep;
1892
1893                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1894                         mmu_page_remove_parent_pte(child, sptep);
1895                 } else if (pfn != spte_to_pfn(*sptep)) {
1896                         pgprintk("hfn old %lx new %lx\n",
1897                                  spte_to_pfn(*sptep), pfn);
1898                         rmap_remove(vcpu->kvm, sptep);
1899                 } else
1900                         was_rmapped = 1;
1901         }
1902
1903         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1904                       dirty, level, gfn, pfn, speculative, true,
1905                       reset_host_protection)) {
1906                 if (write_fault)
1907                         *ptwrite = 1;
1908                 kvm_x86_ops->tlb_flush(vcpu);
1909         }
1910
1911         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1912         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1913                  is_large_pte(*sptep)? "2MB" : "4kB",
1914                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1915                  *sptep, sptep);
1916         if (!was_rmapped && is_large_pte(*sptep))
1917                 ++vcpu->kvm->stat.lpages;
1918
1919         page_header_update_slot(vcpu->kvm, sptep, gfn);
1920         if (!was_rmapped) {
1921                 rmap_count = rmap_add(vcpu, sptep, gfn);
1922                 kvm_release_pfn_clean(pfn);
1923                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1924                         rmap_recycle(vcpu, sptep, gfn);
1925         } else {
1926                 if (was_writable)
1927                         kvm_release_pfn_dirty(pfn);
1928                 else
1929                         kvm_release_pfn_clean(pfn);
1930         }
1931         if (speculative) {
1932                 vcpu->arch.last_pte_updated = sptep;
1933                 vcpu->arch.last_pte_gfn = gfn;
1934         }
1935 }
1936
1937 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1938 {
1939 }
1940
1941 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1942                         int level, gfn_t gfn, pfn_t pfn)
1943 {
1944         struct kvm_shadow_walk_iterator iterator;
1945         struct kvm_mmu_page *sp;
1946         int pt_write = 0;
1947         gfn_t pseudo_gfn;
1948
1949         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1950                 if (iterator.level == level) {
1951                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1952                                      0, write, 1, &pt_write,
1953                                      level, gfn, pfn, false, true);
1954                         ++vcpu->stat.pf_fixed;
1955                         break;
1956                 }
1957
1958                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1959                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1960                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1961                                               iterator.level - 1,
1962                                               1, ACC_ALL, iterator.sptep);
1963                         if (!sp) {
1964                                 pgprintk("nonpaging_map: ENOMEM\n");
1965                                 kvm_release_pfn_clean(pfn);
1966                                 return -ENOMEM;
1967                         }
1968
1969                         __set_spte(iterator.sptep,
1970                                    __pa(sp->spt)
1971                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1972                                    | shadow_user_mask | shadow_x_mask);
1973                 }
1974         }
1975         return pt_write;
1976 }
1977
1978 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1979 {
1980         int r;
1981         int level;
1982         pfn_t pfn;
1983         unsigned long mmu_seq;
1984
1985         level = mapping_level(vcpu, gfn);
1986
1987         /*
1988          * This path builds a PAE pagetable - so we can map 2mb pages at
1989          * maximum. Therefore check if the level is larger than that.
1990          */
1991         if (level > PT_DIRECTORY_LEVEL)
1992                 level = PT_DIRECTORY_LEVEL;
1993
1994         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
1995
1996         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1997         smp_rmb();
1998         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1999
2000         /* mmio */
2001         if (is_error_pfn(pfn)) {
2002                 kvm_release_pfn_clean(pfn);
2003                 return 1;
2004         }
2005
2006         spin_lock(&vcpu->kvm->mmu_lock);
2007         if (mmu_notifier_retry(vcpu, mmu_seq))
2008                 goto out_unlock;
2009         kvm_mmu_free_some_pages(vcpu);
2010         r = __direct_map(vcpu, v, write, level, gfn, pfn);
2011         spin_unlock(&vcpu->kvm->mmu_lock);
2012
2013
2014         return r;
2015
2016 out_unlock:
2017         spin_unlock(&vcpu->kvm->mmu_lock);
2018         kvm_release_pfn_clean(pfn);
2019         return 0;
2020 }
2021
2022
2023 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2024 {
2025         int i;
2026         struct kvm_mmu_page *sp;
2027
2028         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2029                 return;
2030         spin_lock(&vcpu->kvm->mmu_lock);
2031         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2032                 hpa_t root = vcpu->arch.mmu.root_hpa;
2033
2034                 sp = page_header(root);
2035                 --sp->root_count;
2036                 if (!sp->root_count && sp->role.invalid)
2037                         kvm_mmu_zap_page(vcpu->kvm, sp);
2038                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2039                 spin_unlock(&vcpu->kvm->mmu_lock);
2040                 return;
2041         }
2042         for (i = 0; i < 4; ++i) {
2043                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2044
2045                 if (root) {
2046                         root &= PT64_BASE_ADDR_MASK;
2047                         sp = page_header(root);
2048                         --sp->root_count;
2049                         if (!sp->root_count && sp->role.invalid)
2050                                 kvm_mmu_zap_page(vcpu->kvm, sp);
2051                 }
2052                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2053         }
2054         spin_unlock(&vcpu->kvm->mmu_lock);
2055         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2056 }
2057
2058 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2059 {
2060         int ret = 0;
2061
2062         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2063                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2064                 ret = 1;
2065         }
2066
2067         return ret;
2068 }
2069
2070 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2071 {
2072         int i;
2073         gfn_t root_gfn;
2074         struct kvm_mmu_page *sp;
2075         int direct = 0;
2076         u64 pdptr;
2077
2078         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2079
2080         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2081                 hpa_t root = vcpu->arch.mmu.root_hpa;
2082
2083                 ASSERT(!VALID_PAGE(root));
2084                 if (tdp_enabled)
2085                         direct = 1;
2086                 if (mmu_check_root(vcpu, root_gfn))
2087                         return 1;
2088                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2089                                       PT64_ROOT_LEVEL, direct,
2090                                       ACC_ALL, NULL);
2091                 root = __pa(sp->spt);
2092                 ++sp->root_count;
2093                 vcpu->arch.mmu.root_hpa = root;
2094                 return 0;
2095         }
2096         direct = !is_paging(vcpu);
2097         if (tdp_enabled)
2098                 direct = 1;
2099         for (i = 0; i < 4; ++i) {
2100                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2101
2102                 ASSERT(!VALID_PAGE(root));
2103                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2104                         pdptr = kvm_pdptr_read(vcpu, i);
2105                         if (!is_present_gpte(pdptr)) {
2106                                 vcpu->arch.mmu.pae_root[i] = 0;
2107                                 continue;
2108                         }
2109                         root_gfn = pdptr >> PAGE_SHIFT;
2110                 } else if (vcpu->arch.mmu.root_level == 0)
2111                         root_gfn = 0;
2112                 if (mmu_check_root(vcpu, root_gfn))
2113                         return 1;
2114                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2115                                       PT32_ROOT_LEVEL, direct,
2116                                       ACC_ALL, NULL);
2117                 root = __pa(sp->spt);
2118                 ++sp->root_count;
2119                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2120         }
2121         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2122         return 0;
2123 }
2124
2125 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2126 {
2127         int i;
2128         struct kvm_mmu_page *sp;
2129
2130         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2131                 return;
2132         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2133                 hpa_t root = vcpu->arch.mmu.root_hpa;
2134                 sp = page_header(root);
2135                 mmu_sync_children(vcpu, sp);
2136                 return;
2137         }
2138         for (i = 0; i < 4; ++i) {
2139                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2140
2141                 if (root && VALID_PAGE(root)) {
2142                         root &= PT64_BASE_ADDR_MASK;
2143                         sp = page_header(root);
2144                         mmu_sync_children(vcpu, sp);
2145                 }
2146         }
2147 }
2148
2149 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2150 {
2151         spin_lock(&vcpu->kvm->mmu_lock);
2152         mmu_sync_roots(vcpu);
2153         spin_unlock(&vcpu->kvm->mmu_lock);
2154 }
2155
2156 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2157                                   u32 access, u32 *error)
2158 {
2159         if (error)
2160                 *error = 0;
2161         return vaddr;
2162 }
2163
2164 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2165                                 u32 error_code)
2166 {
2167         gfn_t gfn;
2168         int r;
2169
2170         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2171         r = mmu_topup_memory_caches(vcpu);
2172         if (r)
2173                 return r;
2174
2175         ASSERT(vcpu);
2176         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2177
2178         gfn = gva >> PAGE_SHIFT;
2179
2180         return nonpaging_map(vcpu, gva & PAGE_MASK,
2181                              error_code & PFERR_WRITE_MASK, gfn);
2182 }
2183
2184 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2185                                 u32 error_code)
2186 {
2187         pfn_t pfn;
2188         int r;
2189         int level;
2190         gfn_t gfn = gpa >> PAGE_SHIFT;
2191         unsigned long mmu_seq;
2192
2193         ASSERT(vcpu);
2194         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2195
2196         r = mmu_topup_memory_caches(vcpu);
2197         if (r)
2198                 return r;
2199
2200         level = mapping_level(vcpu, gfn);
2201
2202         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2203
2204         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2205         smp_rmb();
2206         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2207         if (is_error_pfn(pfn)) {
2208                 kvm_release_pfn_clean(pfn);
2209                 return 1;
2210         }
2211         spin_lock(&vcpu->kvm->mmu_lock);
2212         if (mmu_notifier_retry(vcpu, mmu_seq))
2213                 goto out_unlock;
2214         kvm_mmu_free_some_pages(vcpu);
2215         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2216                          level, gfn, pfn);
2217         spin_unlock(&vcpu->kvm->mmu_lock);
2218
2219         return r;
2220
2221 out_unlock:
2222         spin_unlock(&vcpu->kvm->mmu_lock);
2223         kvm_release_pfn_clean(pfn);
2224         return 0;
2225 }
2226
2227 static void nonpaging_free(struct kvm_vcpu *vcpu)
2228 {
2229         mmu_free_roots(vcpu);
2230 }
2231
2232 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2233 {
2234         struct kvm_mmu *context = &vcpu->arch.mmu;
2235
2236         context->new_cr3 = nonpaging_new_cr3;
2237         context->page_fault = nonpaging_page_fault;
2238         context->gva_to_gpa = nonpaging_gva_to_gpa;
2239         context->free = nonpaging_free;
2240         context->prefetch_page = nonpaging_prefetch_page;
2241         context->sync_page = nonpaging_sync_page;
2242         context->invlpg = nonpaging_invlpg;
2243         context->root_level = 0;
2244         context->shadow_root_level = PT32E_ROOT_LEVEL;
2245         context->root_hpa = INVALID_PAGE;
2246         return 0;
2247 }
2248
2249 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2250 {
2251         ++vcpu->stat.tlb_flush;
2252         kvm_x86_ops->tlb_flush(vcpu);
2253 }
2254
2255 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2256 {
2257         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2258         mmu_free_roots(vcpu);
2259 }
2260
2261 static void inject_page_fault(struct kvm_vcpu *vcpu,
2262                               u64 addr,
2263                               u32 err_code)
2264 {
2265         kvm_inject_page_fault(vcpu, addr, err_code);
2266 }
2267
2268 static void paging_free(struct kvm_vcpu *vcpu)
2269 {
2270         nonpaging_free(vcpu);
2271 }
2272
2273 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2274 {
2275         int bit7;
2276
2277         bit7 = (gpte >> 7) & 1;
2278         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2279 }
2280
2281 #define PTTYPE 64
2282 #include "paging_tmpl.h"
2283 #undef PTTYPE
2284
2285 #define PTTYPE 32
2286 #include "paging_tmpl.h"
2287 #undef PTTYPE
2288
2289 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2290 {
2291         struct kvm_mmu *context = &vcpu->arch.mmu;
2292         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2293         u64 exb_bit_rsvd = 0;
2294
2295         if (!is_nx(vcpu))
2296                 exb_bit_rsvd = rsvd_bits(63, 63);
2297         switch (level) {
2298         case PT32_ROOT_LEVEL:
2299                 /* no rsvd bits for 2 level 4K page table entries */
2300                 context->rsvd_bits_mask[0][1] = 0;
2301                 context->rsvd_bits_mask[0][0] = 0;
2302                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2303
2304                 if (!is_pse(vcpu)) {
2305                         context->rsvd_bits_mask[1][1] = 0;
2306                         break;
2307                 }
2308
2309                 if (is_cpuid_PSE36())
2310                         /* 36bits PSE 4MB page */
2311                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2312                 else
2313                         /* 32 bits PSE 4MB page */
2314                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2315                 break;
2316         case PT32E_ROOT_LEVEL:
2317                 context->rsvd_bits_mask[0][2] =
2318                         rsvd_bits(maxphyaddr, 63) |
2319                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2320                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2321                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2322                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2323                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2324                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2325                         rsvd_bits(maxphyaddr, 62) |
2326                         rsvd_bits(13, 20);              /* large page */
2327                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2328                 break;
2329         case PT64_ROOT_LEVEL:
2330                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2331                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2332                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2333                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2334                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2335                         rsvd_bits(maxphyaddr, 51);
2336                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2337                         rsvd_bits(maxphyaddr, 51);
2338                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2339                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2340                         rsvd_bits(maxphyaddr, 51) |
2341                         rsvd_bits(13, 29);
2342                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2343                         rsvd_bits(maxphyaddr, 51) |
2344                         rsvd_bits(13, 20);              /* large page */
2345                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2346                 break;
2347         }
2348 }
2349
2350 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2351 {
2352         struct kvm_mmu *context = &vcpu->arch.mmu;
2353
2354         ASSERT(is_pae(vcpu));
2355         context->new_cr3 = paging_new_cr3;
2356         context->page_fault = paging64_page_fault;
2357         context->gva_to_gpa = paging64_gva_to_gpa;
2358         context->prefetch_page = paging64_prefetch_page;
2359         context->sync_page = paging64_sync_page;
2360         context->invlpg = paging64_invlpg;
2361         context->free = paging_free;
2362         context->root_level = level;
2363         context->shadow_root_level = level;
2364         context->root_hpa = INVALID_PAGE;
2365         return 0;
2366 }
2367
2368 static int paging64_init_context(struct kvm_vcpu *vcpu)
2369 {
2370         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2371         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2372 }
2373
2374 static int paging32_init_context(struct kvm_vcpu *vcpu)
2375 {
2376         struct kvm_mmu *context = &vcpu->arch.mmu;
2377
2378         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2379         context->new_cr3 = paging_new_cr3;
2380         context->page_fault = paging32_page_fault;
2381         context->gva_to_gpa = paging32_gva_to_gpa;
2382         context->free = paging_free;
2383         context->prefetch_page = paging32_prefetch_page;
2384         context->sync_page = paging32_sync_page;
2385         context->invlpg = paging32_invlpg;
2386         context->root_level = PT32_ROOT_LEVEL;
2387         context->shadow_root_level = PT32E_ROOT_LEVEL;
2388         context->root_hpa = INVALID_PAGE;
2389         return 0;
2390 }
2391
2392 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2393 {
2394         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2395         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2396 }
2397
2398 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2399 {
2400         struct kvm_mmu *context = &vcpu->arch.mmu;
2401
2402         context->new_cr3 = nonpaging_new_cr3;
2403         context->page_fault = tdp_page_fault;
2404         context->free = nonpaging_free;
2405         context->prefetch_page = nonpaging_prefetch_page;
2406         context->sync_page = nonpaging_sync_page;
2407         context->invlpg = nonpaging_invlpg;
2408         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2409         context->root_hpa = INVALID_PAGE;
2410
2411         if (!is_paging(vcpu)) {
2412                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2413                 context->root_level = 0;
2414         } else if (is_long_mode(vcpu)) {
2415                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2416                 context->gva_to_gpa = paging64_gva_to_gpa;
2417                 context->root_level = PT64_ROOT_LEVEL;
2418         } else if (is_pae(vcpu)) {
2419                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2420                 context->gva_to_gpa = paging64_gva_to_gpa;
2421                 context->root_level = PT32E_ROOT_LEVEL;
2422         } else {
2423                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2424                 context->gva_to_gpa = paging32_gva_to_gpa;
2425                 context->root_level = PT32_ROOT_LEVEL;
2426         }
2427
2428         return 0;
2429 }
2430
2431 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2432 {
2433         int r;
2434
2435         ASSERT(vcpu);
2436         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2437
2438         if (!is_paging(vcpu))
2439                 r = nonpaging_init_context(vcpu);
2440         else if (is_long_mode(vcpu))
2441                 r = paging64_init_context(vcpu);
2442         else if (is_pae(vcpu))
2443                 r = paging32E_init_context(vcpu);
2444         else
2445                 r = paging32_init_context(vcpu);
2446
2447         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2448
2449         return r;
2450 }
2451
2452 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2453 {
2454         vcpu->arch.update_pte.pfn = bad_pfn;
2455
2456         if (tdp_enabled)
2457                 return init_kvm_tdp_mmu(vcpu);
2458         else
2459                 return init_kvm_softmmu(vcpu);
2460 }
2461
2462 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2463 {
2464         ASSERT(vcpu);
2465         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2466                 vcpu->arch.mmu.free(vcpu);
2467                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2468         }
2469 }
2470
2471 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2472 {
2473         destroy_kvm_mmu(vcpu);
2474         return init_kvm_mmu(vcpu);
2475 }
2476 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2477
2478 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2479 {
2480         int r;
2481
2482         r = mmu_topup_memory_caches(vcpu);
2483         if (r)
2484                 goto out;
2485         spin_lock(&vcpu->kvm->mmu_lock);
2486         kvm_mmu_free_some_pages(vcpu);
2487         r = mmu_alloc_roots(vcpu);
2488         mmu_sync_roots(vcpu);
2489         spin_unlock(&vcpu->kvm->mmu_lock);
2490         if (r)
2491                 goto out;
2492         /* set_cr3() should ensure TLB has been flushed */
2493         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2494 out:
2495         return r;
2496 }
2497 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2498
2499 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2500 {
2501         mmu_free_roots(vcpu);
2502 }
2503
2504 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2505                                   struct kvm_mmu_page *sp,
2506                                   u64 *spte)
2507 {
2508         u64 pte;
2509         struct kvm_mmu_page *child;
2510
2511         pte = *spte;
2512         if (is_shadow_present_pte(pte)) {
2513                 if (is_last_spte(pte, sp->role.level))
2514                         rmap_remove(vcpu->kvm, spte);
2515                 else {
2516                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2517                         mmu_page_remove_parent_pte(child, spte);
2518                 }
2519         }
2520         __set_spte(spte, shadow_trap_nonpresent_pte);
2521         if (is_large_pte(pte))
2522                 --vcpu->kvm->stat.lpages;
2523 }
2524
2525 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2526                                   struct kvm_mmu_page *sp,
2527                                   u64 *spte,
2528                                   const void *new)
2529 {
2530         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2531                 ++vcpu->kvm->stat.mmu_pde_zapped;
2532                 return;
2533         }
2534
2535         ++vcpu->kvm->stat.mmu_pte_updated;
2536         if (sp->role.glevels == PT32_ROOT_LEVEL)
2537                 paging32_update_pte(vcpu, sp, spte, new);
2538         else
2539                 paging64_update_pte(vcpu, sp, spte, new);
2540 }
2541
2542 static bool need_remote_flush(u64 old, u64 new)
2543 {
2544         if (!is_shadow_present_pte(old))
2545                 return false;
2546         if (!is_shadow_present_pte(new))
2547                 return true;
2548         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2549                 return true;
2550         old ^= PT64_NX_MASK;
2551         new ^= PT64_NX_MASK;
2552         return (old & ~new & PT64_PERM_MASK) != 0;
2553 }
2554
2555 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2556 {
2557         if (need_remote_flush(old, new))
2558                 kvm_flush_remote_tlbs(vcpu->kvm);
2559         else
2560                 kvm_mmu_flush_tlb(vcpu);
2561 }
2562
2563 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2564 {
2565         u64 *spte = vcpu->arch.last_pte_updated;
2566
2567         return !!(spte && (*spte & shadow_accessed_mask));
2568 }
2569
2570 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2571                                           u64 gpte)
2572 {
2573         gfn_t gfn;
2574         pfn_t pfn;
2575
2576         if (!is_present_gpte(gpte))
2577                 return;
2578         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2579
2580         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2581         smp_rmb();
2582         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2583
2584         if (is_error_pfn(pfn)) {
2585                 kvm_release_pfn_clean(pfn);
2586                 return;
2587         }
2588         vcpu->arch.update_pte.gfn = gfn;
2589         vcpu->arch.update_pte.pfn = pfn;
2590 }
2591
2592 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2593 {
2594         u64 *spte = vcpu->arch.last_pte_updated;
2595
2596         if (spte
2597             && vcpu->arch.last_pte_gfn == gfn
2598             && shadow_accessed_mask
2599             && !(*spte & shadow_accessed_mask)
2600             && is_shadow_present_pte(*spte))
2601                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2602 }
2603
2604 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2605                        const u8 *new, int bytes,
2606                        bool guest_initiated)
2607 {
2608         gfn_t gfn = gpa >> PAGE_SHIFT;
2609         struct kvm_mmu_page *sp;
2610         struct hlist_node *node, *n;
2611         struct hlist_head *bucket;
2612         unsigned index;
2613         u64 entry, gentry;
2614         u64 *spte;
2615         unsigned offset = offset_in_page(gpa);
2616         unsigned pte_size;
2617         unsigned page_offset;
2618         unsigned misaligned;
2619         unsigned quadrant;
2620         int level;
2621         int flooded = 0;
2622         int npte;
2623         int r;
2624         int invlpg_counter;
2625
2626         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2627
2628         invlpg_counter = atomic_read(&vcpu->kvm->arch.invlpg_counter);
2629
2630         /*
2631          * Assume that the pte write on a page table of the same type
2632          * as the current vcpu paging mode.  This is nearly always true
2633          * (might be false while changing modes).  Note it is verified later
2634          * by update_pte().
2635          */
2636         if ((is_pae(vcpu) && bytes == 4) || !new) {
2637                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2638                 if (is_pae(vcpu)) {
2639                         gpa &= ~(gpa_t)7;
2640                         bytes = 8;
2641                 }
2642                 r = kvm_read_guest(vcpu->kvm, gpa, &gentry, min(bytes, 8));
2643                 if (r)
2644                         gentry = 0;
2645                 new = (const u8 *)&gentry;
2646         }
2647
2648         switch (bytes) {
2649         case 4:
2650                 gentry = *(const u32 *)new;
2651                 break;
2652         case 8:
2653                 gentry = *(const u64 *)new;
2654                 break;
2655         default:
2656                 gentry = 0;
2657                 break;
2658         }
2659
2660         mmu_guess_page_from_pte_write(vcpu, gpa, gentry);
2661         spin_lock(&vcpu->kvm->mmu_lock);
2662         if (atomic_read(&vcpu->kvm->arch.invlpg_counter) != invlpg_counter)
2663                 gentry = 0;
2664         kvm_mmu_access_page(vcpu, gfn);
2665         kvm_mmu_free_some_pages(vcpu);
2666         ++vcpu->kvm->stat.mmu_pte_write;
2667         kvm_mmu_audit(vcpu, "pre pte write");
2668         if (guest_initiated) {
2669                 if (gfn == vcpu->arch.last_pt_write_gfn
2670                     && !last_updated_pte_accessed(vcpu)) {
2671                         ++vcpu->arch.last_pt_write_count;
2672                         if (vcpu->arch.last_pt_write_count >= 3)
2673                                 flooded = 1;
2674                 } else {
2675                         vcpu->arch.last_pt_write_gfn = gfn;
2676                         vcpu->arch.last_pt_write_count = 1;
2677                         vcpu->arch.last_pte_updated = NULL;
2678                 }
2679         }
2680         index = kvm_page_table_hashfn(gfn);
2681         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2682         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2683                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2684                         continue;
2685                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2686                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2687                 misaligned |= bytes < 4;
2688                 if (misaligned || flooded) {
2689                         /*
2690                          * Misaligned accesses are too much trouble to fix
2691                          * up; also, they usually indicate a page is not used
2692                          * as a page table.
2693                          *
2694                          * If we're seeing too many writes to a page,
2695                          * it may no longer be a page table, or we may be
2696                          * forking, in which case it is better to unmap the
2697                          * page.
2698                          */
2699                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2700                                  gpa, bytes, sp->role.word);
2701                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2702                                 n = bucket->first;
2703                         ++vcpu->kvm->stat.mmu_flooded;
2704                         continue;
2705                 }
2706                 page_offset = offset;
2707                 level = sp->role.level;
2708                 npte = 1;
2709                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2710                         page_offset <<= 1;      /* 32->64 */
2711                         /*
2712                          * A 32-bit pde maps 4MB while the shadow pdes map
2713                          * only 2MB.  So we need to double the offset again
2714                          * and zap two pdes instead of one.
2715                          */
2716                         if (level == PT32_ROOT_LEVEL) {
2717                                 page_offset &= ~7; /* kill rounding error */
2718                                 page_offset <<= 1;
2719                                 npte = 2;
2720                         }
2721                         quadrant = page_offset >> PAGE_SHIFT;
2722                         page_offset &= ~PAGE_MASK;
2723                         if (quadrant != sp->role.quadrant)
2724                                 continue;
2725                 }
2726                 spte = &sp->spt[page_offset / sizeof(*spte)];
2727                 while (npte--) {
2728                         entry = *spte;
2729                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2730                         if (gentry)
2731                                 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
2732                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2733                         ++spte;
2734                 }
2735         }
2736         kvm_mmu_audit(vcpu, "post pte write");
2737         spin_unlock(&vcpu->kvm->mmu_lock);
2738         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2739                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2740                 vcpu->arch.update_pte.pfn = bad_pfn;
2741         }
2742 }
2743
2744 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2745 {
2746         gpa_t gpa;
2747         int r;
2748
2749         if (tdp_enabled)
2750                 return 0;
2751
2752         gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2753
2754         spin_lock(&vcpu->kvm->mmu_lock);
2755         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2756         spin_unlock(&vcpu->kvm->mmu_lock);
2757         return r;
2758 }
2759 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2760
2761 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2762 {
2763         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2764                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2765                 struct kvm_mmu_page *sp;
2766
2767                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2768                                   struct kvm_mmu_page, link);
2769                 kvm_mmu_zap_page(vcpu->kvm, sp);
2770                 ++vcpu->kvm->stat.mmu_recycled;
2771         }
2772 }
2773
2774 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2775 {
2776         int r;
2777         enum emulation_result er;
2778
2779         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2780         if (r < 0)
2781                 goto out;
2782
2783         if (!r) {
2784                 r = 1;
2785                 goto out;
2786         }
2787
2788         r = mmu_topup_memory_caches(vcpu);
2789         if (r)
2790                 goto out;
2791
2792         er = emulate_instruction(vcpu, cr2, error_code, 0);
2793
2794         switch (er) {
2795         case EMULATE_DONE:
2796                 return 1;
2797         case EMULATE_DO_MMIO:
2798                 ++vcpu->stat.mmio_exits;
2799                 return 0;
2800         case EMULATE_FAIL:
2801                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2802                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2803                 vcpu->run->internal.ndata = 0;
2804                 return 0;
2805         default:
2806                 BUG();
2807         }
2808 out:
2809         return r;
2810 }
2811 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2812
2813 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2814 {
2815         vcpu->arch.mmu.invlpg(vcpu, gva);
2816         kvm_mmu_flush_tlb(vcpu);
2817         ++vcpu->stat.invlpg;
2818 }
2819 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2820
2821 void kvm_enable_tdp(void)
2822 {
2823         tdp_enabled = true;
2824 }
2825 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2826
2827 void kvm_disable_tdp(void)
2828 {
2829         tdp_enabled = false;
2830 }
2831 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2832
2833 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2834 {
2835         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2836 }
2837
2838 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2839 {
2840         struct page *page;
2841         int i;
2842
2843         ASSERT(vcpu);
2844
2845         /*
2846          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2847          * Therefore we need to allocate shadow page tables in the first
2848          * 4GB of memory, which happens to fit the DMA32 zone.
2849          */
2850         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2851         if (!page)
2852                 return -ENOMEM;
2853
2854         vcpu->arch.mmu.pae_root = page_address(page);
2855         for (i = 0; i < 4; ++i)
2856                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2857
2858         return 0;
2859 }
2860
2861 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2862 {
2863         ASSERT(vcpu);
2864         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2865
2866         return alloc_mmu_pages(vcpu);
2867 }
2868
2869 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2870 {
2871         ASSERT(vcpu);
2872         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2873
2874         return init_kvm_mmu(vcpu);
2875 }
2876
2877 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2878 {
2879         ASSERT(vcpu);
2880
2881         destroy_kvm_mmu(vcpu);
2882         free_mmu_pages(vcpu);
2883         mmu_free_memory_caches(vcpu);
2884 }
2885
2886 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2887 {
2888         struct kvm_mmu_page *sp;
2889
2890         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2891                 int i;
2892                 u64 *pt;
2893
2894                 if (!test_bit(slot, sp->slot_bitmap))
2895                         continue;
2896
2897                 pt = sp->spt;
2898                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2899                         /* avoid RMW */
2900                         if (pt[i] & PT_WRITABLE_MASK)
2901                                 pt[i] &= ~PT_WRITABLE_MASK;
2902         }
2903         kvm_flush_remote_tlbs(kvm);
2904 }
2905
2906 void kvm_mmu_zap_all(struct kvm *kvm)
2907 {
2908         struct kvm_mmu_page *sp, *node;
2909
2910         spin_lock(&kvm->mmu_lock);
2911         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2912                 if (kvm_mmu_zap_page(kvm, sp))
2913                         node = container_of(kvm->arch.active_mmu_pages.next,
2914                                             struct kvm_mmu_page, link);
2915         spin_unlock(&kvm->mmu_lock);
2916
2917         kvm_flush_remote_tlbs(kvm);
2918 }
2919
2920 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2921 {
2922         struct kvm_mmu_page *page;
2923
2924         page = container_of(kvm->arch.active_mmu_pages.prev,
2925                             struct kvm_mmu_page, link);
2926         kvm_mmu_zap_page(kvm, page);
2927 }
2928
2929 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2930 {
2931         struct kvm *kvm;
2932         struct kvm *kvm_freed = NULL;
2933         int cache_count = 0;
2934
2935         spin_lock(&kvm_lock);
2936
2937         list_for_each_entry(kvm, &vm_list, vm_list) {
2938                 int npages, idx;
2939
2940                 idx = srcu_read_lock(&kvm->srcu);
2941                 spin_lock(&kvm->mmu_lock);
2942                 npages = kvm->arch.n_alloc_mmu_pages -
2943                          kvm->arch.n_free_mmu_pages;
2944                 cache_count += npages;
2945                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2946                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2947                         cache_count--;
2948                         kvm_freed = kvm;
2949                 }
2950                 nr_to_scan--;
2951
2952                 spin_unlock(&kvm->mmu_lock);
2953                 srcu_read_unlock(&kvm->srcu, idx);
2954         }
2955         if (kvm_freed)
2956                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2957
2958         spin_unlock(&kvm_lock);
2959
2960         return cache_count;
2961 }
2962
2963 static struct shrinker mmu_shrinker = {
2964         .shrink = mmu_shrink,
2965         .seeks = DEFAULT_SEEKS * 10,
2966 };
2967
2968 static void mmu_destroy_caches(void)
2969 {
2970         if (pte_chain_cache)
2971                 kmem_cache_destroy(pte_chain_cache);
2972         if (rmap_desc_cache)
2973                 kmem_cache_destroy(rmap_desc_cache);
2974         if (mmu_page_header_cache)
2975                 kmem_cache_destroy(mmu_page_header_cache);
2976 }
2977
2978 void kvm_mmu_module_exit(void)
2979 {
2980         mmu_destroy_caches();
2981         unregister_shrinker(&mmu_shrinker);
2982 }
2983
2984 int kvm_mmu_module_init(void)
2985 {
2986         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2987                                             sizeof(struct kvm_pte_chain),
2988                                             0, 0, NULL);
2989         if (!pte_chain_cache)
2990                 goto nomem;
2991         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2992                                             sizeof(struct kvm_rmap_desc),
2993                                             0, 0, NULL);
2994         if (!rmap_desc_cache)
2995                 goto nomem;
2996
2997         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2998                                                   sizeof(struct kvm_mmu_page),
2999                                                   0, 0, NULL);
3000         if (!mmu_page_header_cache)
3001                 goto nomem;
3002
3003         register_shrinker(&mmu_shrinker);
3004
3005         return 0;
3006
3007 nomem:
3008         mmu_destroy_caches();
3009         return -ENOMEM;
3010 }
3011
3012 /*
3013  * Caculate mmu pages needed for kvm.
3014  */
3015 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3016 {
3017         int i;
3018         unsigned int nr_mmu_pages;
3019         unsigned int  nr_pages = 0;
3020         struct kvm_memslots *slots;
3021
3022         slots = rcu_dereference(kvm->memslots);
3023         for (i = 0; i < slots->nmemslots; i++)
3024                 nr_pages += slots->memslots[i].npages;
3025
3026         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3027         nr_mmu_pages = max(nr_mmu_pages,
3028                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3029
3030         return nr_mmu_pages;
3031 }
3032
3033 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3034                                 unsigned len)
3035 {
3036         if (len > buffer->len)
3037                 return NULL;
3038         return buffer->ptr;
3039 }
3040
3041 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3042                                 unsigned len)
3043 {
3044         void *ret;
3045
3046         ret = pv_mmu_peek_buffer(buffer, len);
3047         if (!ret)
3048                 return ret;
3049         buffer->ptr += len;
3050         buffer->len -= len;
3051         buffer->processed += len;
3052         return ret;
3053 }
3054
3055 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3056                              gpa_t addr, gpa_t value)
3057 {
3058         int bytes = 8;
3059         int r;
3060
3061         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3062                 bytes = 4;
3063
3064         r = mmu_topup_memory_caches(vcpu);
3065         if (r)
3066                 return r;
3067
3068         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3069                 return -EFAULT;
3070
3071         return 1;
3072 }
3073
3074 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3075 {
3076         kvm_set_cr3(vcpu, vcpu->arch.cr3);
3077         return 1;
3078 }
3079
3080 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3081 {
3082         spin_lock(&vcpu->kvm->mmu_lock);
3083         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3084         spin_unlock(&vcpu->kvm->mmu_lock);
3085         return 1;
3086 }
3087
3088 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3089                              struct kvm_pv_mmu_op_buffer *buffer)
3090 {
3091         struct kvm_mmu_op_header *header;
3092
3093         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3094         if (!header)
3095                 return 0;
3096         switch (header->op) {
3097         case KVM_MMU_OP_WRITE_PTE: {
3098                 struct kvm_mmu_op_write_pte *wpte;
3099
3100                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3101                 if (!wpte)
3102                         return 0;
3103                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3104                                         wpte->pte_val);
3105         }
3106         case KVM_MMU_OP_FLUSH_TLB: {
3107                 struct kvm_mmu_op_flush_tlb *ftlb;
3108
3109                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3110                 if (!ftlb)
3111                         return 0;
3112                 return kvm_pv_mmu_flush_tlb(vcpu);
3113         }
3114         case KVM_MMU_OP_RELEASE_PT: {
3115                 struct kvm_mmu_op_release_pt *rpt;
3116
3117                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3118                 if (!rpt)
3119                         return 0;
3120                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3121         }
3122         default: return 0;
3123         }
3124 }
3125
3126 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3127                   gpa_t addr, unsigned long *ret)
3128 {
3129         int r;
3130         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3131
3132         buffer->ptr = buffer->buf;
3133         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3134         buffer->processed = 0;
3135
3136         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3137         if (r)
3138                 goto out;
3139
3140         while (buffer->len) {
3141                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3142                 if (r < 0)
3143                         goto out;
3144                 if (r == 0)
3145                         break;
3146         }
3147
3148         r = 1;
3149 out:
3150         *ret = buffer->processed;
3151         return r;
3152 }
3153
3154 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3155 {
3156         struct kvm_shadow_walk_iterator iterator;
3157         int nr_sptes = 0;
3158
3159         spin_lock(&vcpu->kvm->mmu_lock);
3160         for_each_shadow_entry(vcpu, addr, iterator) {
3161                 sptes[iterator.level-1] = *iterator.sptep;
3162                 nr_sptes++;
3163                 if (!is_shadow_present_pte(*iterator.sptep))
3164                         break;
3165         }
3166         spin_unlock(&vcpu->kvm->mmu_lock);
3167
3168         return nr_sptes;
3169 }
3170 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3171
3172 #ifdef AUDIT
3173
3174 static const char *audit_msg;
3175
3176 static gva_t canonicalize(gva_t gva)
3177 {
3178 #ifdef CONFIG_X86_64
3179         gva = (long long)(gva << 16) >> 16;
3180 #endif
3181         return gva;
3182 }
3183
3184
3185 typedef void (*inspect_spte_fn) (struct kvm *kvm, u64 *sptep);
3186
3187 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3188                             inspect_spte_fn fn)
3189 {
3190         int i;
3191
3192         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3193                 u64 ent = sp->spt[i];
3194
3195                 if (is_shadow_present_pte(ent)) {
3196                         if (!is_last_spte(ent, sp->role.level)) {
3197                                 struct kvm_mmu_page *child;
3198                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3199                                 __mmu_spte_walk(kvm, child, fn);
3200                         } else
3201                                 fn(kvm, &sp->spt[i]);
3202                 }
3203         }
3204 }
3205
3206 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3207 {
3208         int i;
3209         struct kvm_mmu_page *sp;
3210
3211         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3212                 return;
3213         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3214                 hpa_t root = vcpu->arch.mmu.root_hpa;
3215                 sp = page_header(root);
3216                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3217                 return;
3218         }
3219         for (i = 0; i < 4; ++i) {
3220                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3221
3222                 if (root && VALID_PAGE(root)) {
3223                         root &= PT64_BASE_ADDR_MASK;
3224                         sp = page_header(root);
3225                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3226                 }
3227         }
3228         return;
3229 }
3230
3231 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3232                                 gva_t va, int level)
3233 {
3234         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3235         int i;
3236         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3237
3238         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3239                 u64 ent = pt[i];
3240
3241                 if (ent == shadow_trap_nonpresent_pte)
3242                         continue;
3243
3244                 va = canonicalize(va);
3245                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3246                         audit_mappings_page(vcpu, ent, va, level - 1);
3247                 else {
3248                         gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, va, NULL);
3249                         gfn_t gfn = gpa >> PAGE_SHIFT;
3250                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3251                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3252
3253                         if (is_error_pfn(pfn)) {
3254                                 kvm_release_pfn_clean(pfn);
3255                                 continue;
3256                         }
3257
3258                         if (is_shadow_present_pte(ent)
3259                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3260                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3261                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3262                                        audit_msg, vcpu->arch.mmu.root_level,
3263                                        va, gpa, hpa, ent,
3264                                        is_shadow_present_pte(ent));
3265                         else if (ent == shadow_notrap_nonpresent_pte
3266                                  && !is_error_hpa(hpa))
3267                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3268                                        " valid guest gva %lx\n", audit_msg, va);
3269                         kvm_release_pfn_clean(pfn);
3270
3271                 }
3272         }
3273 }
3274
3275 static void audit_mappings(struct kvm_vcpu *vcpu)
3276 {
3277         unsigned i;
3278
3279         if (vcpu->arch.mmu.root_level == 4)
3280                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3281         else
3282                 for (i = 0; i < 4; ++i)
3283                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3284                                 audit_mappings_page(vcpu,
3285                                                     vcpu->arch.mmu.pae_root[i],
3286                                                     i << 30,
3287                                                     2);
3288 }
3289
3290 static int count_rmaps(struct kvm_vcpu *vcpu)
3291 {
3292         struct kvm *kvm = vcpu->kvm;
3293         struct kvm_memslots *slots;
3294         int nmaps = 0;
3295         int i, j, k, idx;
3296
3297         idx = srcu_read_lock(&kvm->srcu);
3298         slots = rcu_dereference(kvm->memslots);
3299         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3300                 struct kvm_memory_slot *m = &slots->memslots[i];
3301                 struct kvm_rmap_desc *d;
3302
3303                 for (j = 0; j < m->npages; ++j) {
3304                         unsigned long *rmapp = &m->rmap[j];
3305
3306                         if (!*rmapp)
3307                                 continue;
3308                         if (!(*rmapp & 1)) {
3309                                 ++nmaps;
3310                                 continue;
3311                         }
3312                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3313                         while (d) {
3314                                 for (k = 0; k < RMAP_EXT; ++k)
3315                                         if (d->sptes[k])
3316                                                 ++nmaps;
3317                                         else
3318                                                 break;
3319                                 d = d->more;
3320                         }
3321                 }
3322         }
3323         srcu_read_unlock(&kvm->srcu, idx);
3324         return nmaps;
3325 }
3326
3327 void inspect_spte_has_rmap(struct kvm *kvm, u64 *sptep)
3328 {
3329         unsigned long *rmapp;
3330         struct kvm_mmu_page *rev_sp;
3331         gfn_t gfn;
3332
3333         if (*sptep & PT_WRITABLE_MASK) {
3334                 rev_sp = page_header(__pa(sptep));
3335                 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3336
3337                 if (!gfn_to_memslot(kvm, gfn)) {
3338                         if (!printk_ratelimit())
3339                                 return;
3340                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3341                                          audit_msg, gfn);
3342                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3343                                audit_msg, (long int)(sptep - rev_sp->spt),
3344                                         rev_sp->gfn);
3345                         dump_stack();
3346                         return;
3347                 }
3348
3349                 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3350                                     rev_sp->role.level);
3351                 if (!*rmapp) {
3352                         if (!printk_ratelimit())
3353                                 return;
3354                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3355                                          audit_msg, *sptep);
3356                         dump_stack();
3357                 }
3358         }
3359
3360 }
3361
3362 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3363 {
3364         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3365 }
3366
3367 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3368 {
3369         struct kvm_mmu_page *sp;
3370         int i;
3371
3372         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3373                 u64 *pt = sp->spt;
3374
3375                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3376                         continue;
3377
3378                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3379                         u64 ent = pt[i];
3380
3381                         if (!(ent & PT_PRESENT_MASK))
3382                                 continue;
3383                         if (!(ent & PT_WRITABLE_MASK))
3384                                 continue;
3385                         inspect_spte_has_rmap(vcpu->kvm, &pt[i]);
3386                 }
3387         }
3388         return;
3389 }
3390
3391 static void audit_rmap(struct kvm_vcpu *vcpu)
3392 {
3393         check_writable_mappings_rmap(vcpu);
3394         count_rmaps(vcpu);
3395 }
3396
3397 static void audit_write_protection(struct kvm_vcpu *vcpu)
3398 {
3399         struct kvm_mmu_page *sp;
3400         struct kvm_memory_slot *slot;
3401         unsigned long *rmapp;
3402         u64 *spte;
3403         gfn_t gfn;
3404
3405         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3406                 if (sp->role.direct)
3407                         continue;
3408                 if (sp->unsync)
3409                         continue;
3410
3411                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3412                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3413                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3414
3415                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3416                 while (spte) {
3417                         if (*spte & PT_WRITABLE_MASK)
3418                                 printk(KERN_ERR "%s: (%s) shadow page has "
3419                                 "writable mappings: gfn %lx role %x\n",
3420                                __func__, audit_msg, sp->gfn,
3421                                sp->role.word);
3422                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3423                 }
3424         }
3425 }
3426
3427 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3428 {
3429         int olddbg = dbg;
3430
3431         dbg = 0;
3432         audit_msg = msg;
3433         audit_rmap(vcpu);
3434         audit_write_protection(vcpu);
3435         if (strcmp("pre pte write", audit_msg) != 0)
3436                 audit_mappings(vcpu);
3437         audit_writable_sptes_have_rmaps(vcpu);
3438         dbg = olddbg;
3439 }
3440
3441 #endif