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         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 927         sp->multimapped = 0;
 928         sp->parent_pte = parent_pte;
 929         --vcpu->kvm->arch.n_free_mmu_pages;
 930         return sp;
 931 }
 932
 933 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 934                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 935 {
 936         struct kvm_pte_chain *pte_chain;
 937         struct hlist_node *node;
 938         int i;
 939
 940         if (!parent_pte)
 941                 return;
 942         if (!sp->multimapped) {
 943                 u64 *old = sp->parent_pte;
 944
 945                 if (!old) {
 946                         sp->parent_pte = parent_pte;
 947                         return;
 948                 }
 949                 sp->multimapped = 1;
 950                 pte_chain = mmu_alloc_pte_chain(vcpu);
 951                 INIT_HLIST_HEAD(&sp->parent_ptes);
 952                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 953                 pte_chain->parent_ptes[0] = old;
 954         }
 955         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 956                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 957                         continue;
 958                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 959                         if (!pte_chain->parent_ptes[i]) {
 960                                 pte_chain->parent_ptes[i] = parent_pte;
 961                                 return;
 962                         }
 963         }
 964         pte_chain = mmu_alloc_pte_chain(vcpu);
 965         BUG_ON(!pte_chain);
 966         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 967         pte_chain->parent_ptes[0] = parent_pte;
 968 }
 969
 970 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 971                                        u64 *parent_pte)
 972 {
 973         struct kvm_pte_chain *pte_chain;
 974         struct hlist_node *node;
 975         int i;
 976
 977         if (!sp->multimapped) {
 978                 BUG_ON(sp->parent_pte != parent_pte);
 979                 sp->parent_pte = NULL;
 980                 return;
 981         }
 982         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 983                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 984                         if (!pte_chain->parent_ptes[i])
 985                                 break;
 986                         if (pte_chain->parent_ptes[i] != parent_pte)
 987                                 continue;
 988                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 989                                 && pte_chain->parent_ptes[i + 1]) {
 990                                 pte_chain->parent_ptes[i]
 991                                         = pte_chain->parent_ptes[i + 1];
 992                                 ++i;
 993                         }
 994                         pte_chain->parent_ptes[i] = NULL;
 995                         if (i == 0) {
 996                                 hlist_del(&pte_chain->link);
 997                                 mmu_free_pte_chain(pte_chain);
 998                                 if (hlist_empty(&sp->parent_ptes)) {
 999                                         sp->multimapped = 0;
1000                                         sp->parent_pte = NULL;
1001                                 }
1002                         }
1003                         return;
1004                 }
1005         BUG();
1006 }
1007
1008
1009 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1010                             mmu_parent_walk_fn fn)
1011 {
1012         struct kvm_pte_chain *pte_chain;
1013         struct hlist_node *node;
1014         struct kvm_mmu_page *parent_sp;
1015         int i;
1016
1017         if (!sp->multimapped && sp->parent_pte) {
1018                 parent_sp = page_header(__pa(sp->parent_pte));
1019                 fn(vcpu, parent_sp);
1020                 mmu_parent_walk(vcpu, parent_sp, fn);
1021                 return;
1022         }
1023         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1024                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1025                         if (!pte_chain->parent_ptes[i])
1026                                 break;
1027                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
1028                         fn(vcpu, parent_sp);
1029                         mmu_parent_walk(vcpu, parent_sp, fn);
1030                 }
1031 }
1032
1033 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1034 {
1035         unsigned int index;
1036         struct kvm_mmu_page *sp = page_header(__pa(spte));
1037
1038         index = spte - sp->spt;
1039         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1040                 sp->unsync_children++;
1041         WARN_ON(!sp->unsync_children);
1042 }
1043
1044 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1045 {
1046         struct kvm_pte_chain *pte_chain;
1047         struct hlist_node *node;
1048         int i;
1049
1050         if (!sp->parent_pte)
1051                 return;
1052
1053         if (!sp->multimapped) {
1054                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1055                 return;
1056         }
1057
1058         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1059                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1060                         if (!pte_chain->parent_ptes[i])
1061                                 break;
1062                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1063                 }
1064 }
1065
1066 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1067 {
1068         kvm_mmu_update_parents_unsync(sp);
1069         return 1;
1070 }
1071
1072 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
1073                                         struct kvm_mmu_page *sp)
1074 {
1075         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
1076         kvm_mmu_update_parents_unsync(sp);
1077 }
1078
1079 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1080                                     struct kvm_mmu_page *sp)
1081 {
1082         int i;
1083
1084         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1085                 sp->spt[i] = shadow_trap_nonpresent_pte;
1086 }
1087
1088 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1089                                struct kvm_mmu_page *sp)
1090 {
1091         return 1;
1092 }
1093
1094 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1095 {
1096 }
1097
1098 #define KVM_PAGE_ARRAY_NR 16
1099
1100 struct kvm_mmu_pages {
1101         struct mmu_page_and_offset {
1102                 struct kvm_mmu_page *sp;
1103                 unsigned int idx;
1104         } page[KVM_PAGE_ARRAY_NR];
1105         unsigned int nr;
1106 };
1107
1108 #define for_each_unsync_children(bitmap, idx)           \
1109         for (idx = find_first_bit(bitmap, 512);         \
1110              idx < 512;                                 \
1111              idx = find_next_bit(bitmap, 512, idx+1))
1112
1113 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1114                          int idx)
1115 {
1116         int i;
1117
1118         if (sp->unsync)
1119                 for (i=0; i < pvec->nr; i++)
1120                         if (pvec->page[i].sp == sp)
1121                                 return 0;
1122
1123         pvec->page[pvec->nr].sp = sp;
1124         pvec->page[pvec->nr].idx = idx;
1125         pvec->nr++;
1126         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1127 }
1128
1129 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1130                            struct kvm_mmu_pages *pvec)
1131 {
1132         int i, ret, nr_unsync_leaf = 0;
1133
1134         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1135                 u64 ent = sp->spt[i];
1136
1137                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1138                         struct kvm_mmu_page *child;
1139                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1140
1141                         if (child->unsync_children) {
1142                                 if (mmu_pages_add(pvec, child, i))
1143                                         return -ENOSPC;
1144
1145                                 ret = __mmu_unsync_walk(child, pvec);
1146                                 if (!ret)
1147                                         __clear_bit(i, sp->unsync_child_bitmap);
1148                                 else if (ret > 0)
1149                                         nr_unsync_leaf += ret;
1150                                 else
1151                                         return ret;
1152                         }
1153
1154                         if (child->unsync) {
1155                                 nr_unsync_leaf++;
1156                                 if (mmu_pages_add(pvec, child, i))
1157                                         return -ENOSPC;
1158                         }
1159                 }
1160         }
1161
1162         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1163                 sp->unsync_children = 0;
1164
1165         return nr_unsync_leaf;
1166 }
1167
1168 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1169                            struct kvm_mmu_pages *pvec)
1170 {
1171         if (!sp->unsync_children)
1172                 return 0;
1173
1174         mmu_pages_add(pvec, sp, 0);
1175         return __mmu_unsync_walk(sp, pvec);
1176 }
1177
1178 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1179 {
1180         unsigned index;
1181         struct hlist_head *bucket;
1182         struct kvm_mmu_page *sp;
1183         struct hlist_node *node;
1184
1185         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1186         index = kvm_page_table_hashfn(gfn);
1187         bucket = &kvm->arch.mmu_page_hash[index];
1188         hlist_for_each_entry(sp, node, bucket, hash_link)
1189                 if (sp->gfn == gfn && !sp->role.direct
1190                     && !sp->role.invalid) {
1191                         pgprintk("%s: found role %x\n",
1192                                  __func__, sp->role.word);
1193                         return sp;
1194                 }
1195         return NULL;
1196 }
1197
1198 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1199 {
1200         WARN_ON(!sp->unsync);
1201         sp->unsync = 0;
1202         --kvm->stat.mmu_unsync;
1203 }
1204
1205 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1206
1207 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1208 {
1209         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1210                 kvm_mmu_zap_page(vcpu->kvm, sp);
1211                 return 1;
1212         }
1213
1214         trace_kvm_mmu_sync_page(sp);
1215         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1216                 kvm_flush_remote_tlbs(vcpu->kvm);
1217         kvm_unlink_unsync_page(vcpu->kvm, sp);
1218         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1219                 kvm_mmu_zap_page(vcpu->kvm, sp);
1220                 return 1;
1221         }
1222
1223         kvm_mmu_flush_tlb(vcpu);
1224         return 0;
1225 }
1226
1227 struct mmu_page_path {
1228         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1229         unsigned int idx[PT64_ROOT_LEVEL-1];
1230 };
1231
1232 #define for_each_sp(pvec, sp, parents, i)                       \
1233                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1234                         sp = pvec.page[i].sp;                   \
1235                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1236                         i = mmu_pages_next(&pvec, &parents, i))
1237
1238 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1239                           struct mmu_page_path *parents,
1240                           int i)
1241 {
1242         int n;
1243
1244         for (n = i+1; n < pvec->nr; n++) {
1245                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1246
1247                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1248                         parents->idx[0] = pvec->page[n].idx;
1249                         return n;
1250                 }
1251
1252                 parents->parent[sp->role.level-2] = sp;
1253                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1254         }
1255
1256         return n;
1257 }
1258
1259 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1260 {
1261         struct kvm_mmu_page *sp;
1262         unsigned int level = 0;
1263
1264         do {
1265                 unsigned int idx = parents->idx[level];
1266
1267                 sp = parents->parent[level];
1268                 if (!sp)
1269                         return;
1270
1271                 --sp->unsync_children;
1272                 WARN_ON((int)sp->unsync_children < 0);
1273                 __clear_bit(idx, sp->unsync_child_bitmap);
1274                 level++;
1275         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1276 }
1277
1278 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1279                                struct mmu_page_path *parents,
1280                                struct kvm_mmu_pages *pvec)
1281 {
1282         parents->parent[parent->role.level-1] = NULL;
1283         pvec->nr = 0;
1284 }
1285
1286 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1287                               struct kvm_mmu_page *parent)
1288 {
1289         int i;
1290         struct kvm_mmu_page *sp;
1291         struct mmu_page_path parents;
1292         struct kvm_mmu_pages pages;
1293
1294         kvm_mmu_pages_init(parent, &parents, &pages);
1295         while (mmu_unsync_walk(parent, &pages)) {
1296                 int protected = 0;
1297
1298                 for_each_sp(pages, sp, parents, i)
1299                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1300
1301                 if (protected)
1302                         kvm_flush_remote_tlbs(vcpu->kvm);
1303
1304                 for_each_sp(pages, sp, parents, i) {
1305                         kvm_sync_page(vcpu, sp);
1306                         mmu_pages_clear_parents(&parents);
1307                 }
1308                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1309                 kvm_mmu_pages_init(parent, &parents, &pages);
1310         }
1311 }
1312
1313 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1314                                              gfn_t gfn,
1315                                              gva_t gaddr,
1316                                              unsigned level,
1317                                              int direct,
1318                                              unsigned access,
1319                                              u64 *parent_pte)
1320 {
1321         union kvm_mmu_page_role role;
1322         unsigned index;
1323         unsigned quadrant;
1324         struct hlist_head *bucket;
1325         struct kvm_mmu_page *sp;
1326         struct hlist_node *node, *tmp;
1327
1328         role = vcpu->arch.mmu.base_role;
1329         role.level = level;
1330         role.direct = direct;
1331         if (role.direct)
1332                 role.glevels = 0;
1333         role.access = access;
1334         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1335                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1336                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1337                 role.quadrant = quadrant;
1338         }
1339         index = kvm_page_table_hashfn(gfn);
1340         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1341         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1342                 if (sp->gfn == gfn) {
1343                         if (sp->unsync)
1344                                 if (kvm_sync_page(vcpu, sp))
1345                                         continue;
1346
1347                         if (sp->role.word != role.word)
1348                                 continue;
1349
1350                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1351                         if (sp->unsync_children) {
1352                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1353                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1354                         }
1355                         trace_kvm_mmu_get_page(sp, false);
1356                         return sp;
1357                 }
1358         ++vcpu->kvm->stat.mmu_cache_miss;
1359         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1360         if (!sp)
1361                 return sp;
1362         sp->gfn = gfn;
1363         sp->role = role;
1364         hlist_add_head(&sp->hash_link, bucket);
1365         if (!direct) {
1366                 if (rmap_write_protect(vcpu->kvm, gfn))
1367                         kvm_flush_remote_tlbs(vcpu->kvm);
1368                 account_shadowed(vcpu->kvm, gfn);
1369         }
1370         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1371                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1372         else
1373                 nonpaging_prefetch_page(vcpu, sp);
1374         trace_kvm_mmu_get_page(sp, true);
1375         return sp;
1376 }
1377
1378 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1379                              struct kvm_vcpu *vcpu, u64 addr)
1380 {
1381         iterator->addr = addr;
1382         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1383         iterator->level = vcpu->arch.mmu.shadow_root_level;
1384         if (iterator->level == PT32E_ROOT_LEVEL) {
1385                 iterator->shadow_addr
1386                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1387                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1388                 --iterator->level;
1389                 if (!iterator->shadow_addr)
1390                         iterator->level = 0;
1391         }
1392 }
1393
1394 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1395 {
1396         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1397                 return false;
1398
1399         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1400                 if (is_large_pte(*iterator->sptep))
1401                         return false;
1402
1403         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1404         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1405         return true;
1406 }
1407
1408 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1409 {
1410         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1411         --iterator->level;
1412 }
1413
1414 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1415                                          struct kvm_mmu_page *sp)
1416 {
1417         unsigned i;
1418         u64 *pt;
1419         u64 ent;
1420
1421         pt = sp->spt;
1422
1423         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1424                 ent = pt[i];
1425
1426                 if (is_shadow_present_pte(ent)) {
1427                         if (!is_last_spte(ent, sp->role.level)) {
1428                                 ent &= PT64_BASE_ADDR_MASK;
1429                                 mmu_page_remove_parent_pte(page_header(ent),
1430                                                            &pt[i]);
1431                         } else {
1432                                 if (is_large_pte(ent))
1433                                         --kvm->stat.lpages;
1434                                 rmap_remove(kvm, &pt[i]);
1435                         }
1436                 }
1437                 pt[i] = shadow_trap_nonpresent_pte;
1438         }
1439 }
1440
1441 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1442 {
1443         mmu_page_remove_parent_pte(sp, parent_pte);
1444 }
1445
1446 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1447 {
1448         int i;
1449         struct kvm_vcpu *vcpu;
1450
1451         kvm_for_each_vcpu(i, vcpu, kvm)
1452                 vcpu->arch.last_pte_updated = NULL;
1453 }
1454
1455 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1456 {
1457         u64 *parent_pte;
1458
1459         while (sp->multimapped || sp->parent_pte) {
1460                 if (!sp->multimapped)
1461                         parent_pte = sp->parent_pte;
1462                 else {
1463                         struct kvm_pte_chain *chain;
1464
1465                         chain = container_of(sp->parent_ptes.first,
1466                                              struct kvm_pte_chain, link);
1467                         parent_pte = chain->parent_ptes[0];
1468                 }
1469                 BUG_ON(!parent_pte);
1470                 kvm_mmu_put_page(sp, parent_pte);
1471                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1472         }
1473 }
1474
1475 static int mmu_zap_unsync_children(struct kvm *kvm,
1476                                    struct kvm_mmu_page *parent)
1477 {
1478         int i, zapped = 0;
1479         struct mmu_page_path parents;
1480         struct kvm_mmu_pages pages;
1481
1482         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1483                 return 0;
1484
1485         kvm_mmu_pages_init(parent, &parents, &pages);
1486         while (mmu_unsync_walk(parent, &pages)) {
1487                 struct kvm_mmu_page *sp;
1488
1489                 for_each_sp(pages, sp, parents, i) {
1490                         kvm_mmu_zap_page(kvm, sp);
1491                         mmu_pages_clear_parents(&parents);
1492                         zapped++;
1493                 }
1494                 kvm_mmu_pages_init(parent, &parents, &pages);
1495         }
1496
1497         return zapped;
1498 }
1499
1500 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1501 {
1502         int ret;
1503
1504         trace_kvm_mmu_zap_page(sp);
1505         ++kvm->stat.mmu_shadow_zapped;
1506         ret = mmu_zap_unsync_children(kvm, sp);
1507         kvm_mmu_page_unlink_children(kvm, sp);
1508         kvm_mmu_unlink_parents(kvm, sp);
1509         kvm_flush_remote_tlbs(kvm);
1510         if (!sp->role.invalid && !sp->role.direct)
1511                 unaccount_shadowed(kvm, sp->gfn);
1512         if (sp->unsync)
1513                 kvm_unlink_unsync_page(kvm, sp);
1514         if (!sp->root_count) {
1515                 hlist_del(&sp->hash_link);
1516                 kvm_mmu_free_page(kvm, sp);
1517         } else {
1518                 sp->role.invalid = 1;
1519                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1520                 kvm_reload_remote_mmus(kvm);
1521         }
1522         kvm_mmu_reset_last_pte_updated(kvm);
1523         return ret;
1524 }
1525
1526 /*
1527  * Changing the number of mmu pages allocated to the vm
1528  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1529  */
1530 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1531 {
1532         int used_pages;
1533
1534         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1535         used_pages = max(0, used_pages);
1536
1537         /*
1538          * If we set the number of mmu pages to be smaller be than the
1539          * number of actived pages , we must to free some mmu pages before we
1540          * change the value
1541          */
1542
1543         if (used_pages > kvm_nr_mmu_pages) {
1544                 while (used_pages > kvm_nr_mmu_pages &&
1545                         !list_empty(&kvm->arch.active_mmu_pages)) {
1546                         struct kvm_mmu_page *page;
1547
1548                         page = container_of(kvm->arch.active_mmu_pages.prev,
1549                                             struct kvm_mmu_page, link);
1550                         used_pages -= kvm_mmu_zap_page(kvm, page);
1551                         used_pages--;
1552                 }
1553                 kvm_nr_mmu_pages = used_pages;
1554                 kvm->arch.n_free_mmu_pages = 0;
1555         }
1556         else
1557                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1558                                          - kvm->arch.n_alloc_mmu_pages;
1559
1560         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1561 }
1562
1563 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1564 {
1565         unsigned index;
1566         struct hlist_head *bucket;
1567         struct kvm_mmu_page *sp;
1568         struct hlist_node *node, *n;
1569         int r;
1570
1571         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1572         r = 0;
1573         index = kvm_page_table_hashfn(gfn);
1574         bucket = &kvm->arch.mmu_page_hash[index];
1575         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1576                 if (sp->gfn == gfn && !sp->role.direct) {
1577                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1578                                  sp->role.word);
1579                         r = 1;
1580                         if (kvm_mmu_zap_page(kvm, sp))
1581                                 n = bucket->first;
1582                 }
1583         return r;
1584 }
1585
1586 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1587 {
1588         unsigned index;
1589         struct hlist_head *bucket;
1590         struct kvm_mmu_page *sp;
1591         struct hlist_node *node, *nn;
1592
1593         index = kvm_page_table_hashfn(gfn);
1594         bucket = &kvm->arch.mmu_page_hash[index];
1595         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1596                 if (sp->gfn == gfn && !sp->role.direct
1597                     && !sp->role.invalid) {
1598                         pgprintk("%s: zap %lx %x\n",
1599                                  __func__, gfn, sp->role.word);
1600                         if (kvm_mmu_zap_page(kvm, sp))
1601                                 nn = bucket->first;
1602                 }
1603         }
1604 }
1605
1606 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1607 {
1608         int slot = memslot_id(kvm, gfn);
1609         struct kvm_mmu_page *sp = page_header(__pa(pte));
1610
1611         __set_bit(slot, sp->slot_bitmap);
1612 }
1613
1614 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1615 {
1616         int i;
1617         u64 *pt = sp->spt;
1618
1619         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1620                 return;
1621
1622         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1623                 if (pt[i] == shadow_notrap_nonpresent_pte)
1624                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1625         }
1626 }
1627
1628 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1629 {
1630         struct page *page;
1631
1632         gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
1633
1634         if (gpa == UNMAPPED_GVA)
1635                 return NULL;
1636
1637         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1638
1639         return page;
1640 }
1641
1642 /*
1643  * The function is based on mtrr_type_lookup() in
1644  * arch/x86/kernel/cpu/mtrr/generic.c
1645  */
1646 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1647                          u64 start, u64 end)
1648 {
1649         int i;
1650         u64 base, mask;
1651         u8 prev_match, curr_match;
1652         int num_var_ranges = KVM_NR_VAR_MTRR;
1653
1654         if (!mtrr_state->enabled)
1655                 return 0xFF;
1656
1657         /* Make end inclusive end, instead of exclusive */
1658         end--;
1659
1660         /* Look in fixed ranges. Just return the type as per start */
1661         if (mtrr_state->have_fixed && (start < 0x100000)) {
1662                 int idx;
1663
1664                 if (start < 0x80000) {
1665                         idx = 0;
1666                         idx += (start >> 16);
1667                         return mtrr_state->fixed_ranges[idx];
1668                 } else if (start < 0xC0000) {
1669                         idx = 1 * 8;
1670                         idx += ((start - 0x80000) >> 14);
1671                         return mtrr_state->fixed_ranges[idx];
1672                 } else if (start < 0x1000000) {
1673                         idx = 3 * 8;
1674                         idx += ((start - 0xC0000) >> 12);
1675                         return mtrr_state->fixed_ranges[idx];
1676                 }
1677         }
1678
1679         /*
1680          * Look in variable ranges
1681          * Look of multiple ranges matching this address and pick type
1682          * as per MTRR precedence
1683          */
1684         if (!(mtrr_state->enabled & 2))
1685                 return mtrr_state->def_type;
1686
1687         prev_match = 0xFF;
1688         for (i = 0; i < num_var_ranges; ++i) {
1689                 unsigned short start_state, end_state;
1690
1691                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1692                         continue;
1693
1694                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1695                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1696                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1697                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1698
1699                 start_state = ((start & mask) == (base & mask));
1700                 end_state = ((end & mask) == (base & mask));
1701                 if (start_state != end_state)
1702                         return 0xFE;
1703
1704                 if ((start & mask) != (base & mask))
1705                         continue;
1706
1707                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1708                 if (prev_match == 0xFF) {
1709                         prev_match = curr_match;
1710                         continue;
1711                 }
1712
1713                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1714                     curr_match == MTRR_TYPE_UNCACHABLE)
1715                         return MTRR_TYPE_UNCACHABLE;
1716
1717                 if ((prev_match == MTRR_TYPE_WRBACK &&
1718                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1719                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1720                      curr_match == MTRR_TYPE_WRBACK)) {
1721                         prev_match = MTRR_TYPE_WRTHROUGH;
1722                         curr_match = MTRR_TYPE_WRTHROUGH;
1723                 }
1724
1725                 if (prev_match != curr_match)
1726                         return MTRR_TYPE_UNCACHABLE;
1727         }
1728
1729         if (prev_match != 0xFF)
1730                 return prev_match;
1731
1732         return mtrr_state->def_type;
1733 }
1734
1735 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1736 {
1737         u8 mtrr;
1738
1739         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1740                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1741         if (mtrr == 0xfe || mtrr == 0xff)
1742                 mtrr = MTRR_TYPE_WRBACK;
1743         return mtrr;
1744 }
1745 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1746
1747 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1748 {
1749         unsigned index;
1750         struct hlist_head *bucket;
1751         struct kvm_mmu_page *s;
1752         struct hlist_node *node, *n;
1753
1754         trace_kvm_mmu_unsync_page(sp);
1755         index = kvm_page_table_hashfn(sp->gfn);
1756         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1757         /* don't unsync if pagetable is shadowed with multiple roles */
1758         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1759                 if (s->gfn != sp->gfn || s->role.direct)
1760                         continue;
1761                 if (s->role.word != sp->role.word)
1762                         return 1;
1763         }
1764         ++vcpu->kvm->stat.mmu_unsync;
1765         sp->unsync = 1;
1766
1767         kvm_mmu_mark_parents_unsync(vcpu, sp);
1768
1769         mmu_convert_notrap(sp);
1770         return 0;
1771 }
1772
1773 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1774                                   bool can_unsync)
1775 {
1776         struct kvm_mmu_page *shadow;
1777
1778         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1779         if (shadow) {
1780                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1781                         return 1;
1782                 if (shadow->unsync)
1783                         return 0;
1784                 if (can_unsync && oos_shadow)
1785                         return kvm_unsync_page(vcpu, shadow);
1786                 return 1;
1787         }
1788         return 0;
1789 }
1790
1791 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1792                     unsigned pte_access, int user_fault,
1793                     int write_fault, int dirty, int level,
1794                     gfn_t gfn, pfn_t pfn, bool speculative,
1795                     bool can_unsync, bool reset_host_protection)
1796 {
1797         u64 spte;
1798         int ret = 0;
1799
1800         /*
1801          * We don't set the accessed bit, since we sometimes want to see
1802          * whether the guest actually used the pte (in order to detect
1803          * demand paging).
1804          */
1805         spte = shadow_base_present_pte | shadow_dirty_mask;
1806         if (!speculative)
1807                 spte |= shadow_accessed_mask;
1808         if (!dirty)
1809                 pte_access &= ~ACC_WRITE_MASK;
1810         if (pte_access & ACC_EXEC_MASK)
1811                 spte |= shadow_x_mask;
1812         else
1813                 spte |= shadow_nx_mask;
1814         if (pte_access & ACC_USER_MASK)
1815                 spte |= shadow_user_mask;
1816         if (level > PT_PAGE_TABLE_LEVEL)
1817                 spte |= PT_PAGE_SIZE_MASK;
1818         if (tdp_enabled)
1819                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1820                         kvm_is_mmio_pfn(pfn));
1821
1822         if (reset_host_protection)
1823                 spte |= SPTE_HOST_WRITEABLE;
1824
1825         spte |= (u64)pfn << PAGE_SHIFT;
1826
1827         if ((pte_access & ACC_WRITE_MASK)
1828             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1829
1830                 if (level > PT_PAGE_TABLE_LEVEL &&
1831                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1832                         ret = 1;
1833                         spte = shadow_trap_nonpresent_pte;
1834                         goto set_pte;
1835                 }
1836
1837                 spte |= PT_WRITABLE_MASK;
1838
1839                 /*
1840                  * Optimization: for pte sync, if spte was writable the hash
1841                  * lookup is unnecessary (and expensive). Write protection
1842                  * is responsibility of mmu_get_page / kvm_sync_page.
1843                  * Same reasoning can be applied to dirty page accounting.
1844                  */
1845                 if (!can_unsync && is_writable_pte(*sptep))
1846                         goto set_pte;
1847
1848                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1849                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1850                                  __func__, gfn);
1851                         ret = 1;
1852                         pte_access &= ~ACC_WRITE_MASK;
1853                         if (is_writable_pte(spte))
1854                                 spte &= ~PT_WRITABLE_MASK;
1855                 }
1856         }
1857
1858         if (pte_access & ACC_WRITE_MASK)
1859                 mark_page_dirty(vcpu->kvm, gfn);
1860
1861 set_pte:
1862         __set_spte(sptep, spte);
1863         return ret;
1864 }
1865
1866 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1867                          unsigned pt_access, unsigned pte_access,
1868                          int user_fault, int write_fault, int dirty,
1869                          int *ptwrite, int level, gfn_t gfn,
1870                          pfn_t pfn, bool speculative,
1871                          bool reset_host_protection)
1872 {
1873         int was_rmapped = 0;
1874         int was_writable = is_writable_pte(*sptep);
1875         int rmap_count;
1876
1877         pgprintk("%s: spte %llx access %x write_fault %d"
1878                  " user_fault %d gfn %lx\n",
1879                  __func__, *sptep, pt_access,
1880                  write_fault, user_fault, gfn);
1881
1882         if (is_rmap_spte(*sptep)) {
1883                 /*
1884                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1885                  * the parent of the now unreachable PTE.
1886                  */
1887                 if (level > PT_PAGE_TABLE_LEVEL &&
1888                     !is_large_pte(*sptep)) {
1889                         struct kvm_mmu_page *child;
1890                         u64 pte = *sptep;
1891
1892                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1893                         mmu_page_remove_parent_pte(child, sptep);
1894                 } else if (pfn != spte_to_pfn(*sptep)) {
1895                         pgprintk("hfn old %lx new %lx\n",
1896                                  spte_to_pfn(*sptep), pfn);
1897                         rmap_remove(vcpu->kvm, sptep);
1898                 } else
1899                         was_rmapped = 1;
1900         }
1901
1902         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1903                       dirty, level, gfn, pfn, speculative, true,
1904                       reset_host_protection)) {
1905                 if (write_fault)
1906                         *ptwrite = 1;
1907                 kvm_x86_ops->tlb_flush(vcpu);
1908         }
1909
1910         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1911         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1912                  is_large_pte(*sptep)? "2MB" : "4kB",
1913                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1914                  *sptep, sptep);
1915         if (!was_rmapped && is_large_pte(*sptep))
1916                 ++vcpu->kvm->stat.lpages;
1917
1918         page_header_update_slot(vcpu->kvm, sptep, gfn);
1919         if (!was_rmapped) {
1920                 rmap_count = rmap_add(vcpu, sptep, gfn);
1921                 kvm_release_pfn_clean(pfn);
1922                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1923                         rmap_recycle(vcpu, sptep, gfn);
1924         } else {
1925                 if (was_writable)
1926                         kvm_release_pfn_dirty(pfn);
1927                 else
1928                         kvm_release_pfn_clean(pfn);
1929         }
1930         if (speculative) {
1931                 vcpu->arch.last_pte_updated = sptep;
1932                 vcpu->arch.last_pte_gfn = gfn;
1933         }
1934 }
1935
1936 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1937 {
1938 }
1939
1940 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1941                         int level, gfn_t gfn, pfn_t pfn)
1942 {
1943         struct kvm_shadow_walk_iterator iterator;
1944         struct kvm_mmu_page *sp;
1945         int pt_write = 0;
1946         gfn_t pseudo_gfn;
1947
1948         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1949                 if (iterator.level == level) {
1950                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1951                                      0, write, 1, &pt_write,
1952                                      level, gfn, pfn, false, true);
1953                         ++vcpu->stat.pf_fixed;
1954                         break;
1955                 }
1956
1957                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1958                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1959                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1960                                               iterator.level - 1,
1961                                               1, ACC_ALL, iterator.sptep);
1962                         if (!sp) {
1963                                 pgprintk("nonpaging_map: ENOMEM\n");
1964                                 kvm_release_pfn_clean(pfn);
1965                                 return -ENOMEM;
1966                         }
1967
1968                         __set_spte(iterator.sptep,
1969                                    __pa(sp->spt)
1970                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1971                                    | shadow_user_mask | shadow_x_mask);
1972                 }
1973         }
1974         return pt_write;
1975 }
1976
1977 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1978 {
1979         int r;
1980         int level;
1981         pfn_t pfn;
1982         unsigned long mmu_seq;
1983
1984         level = mapping_level(vcpu, gfn);
1985
1986         /*
1987          * This path builds a PAE pagetable - so we can map 2mb pages at
1988          * maximum. Therefore check if the level is larger than that.
1989          */
1990         if (level > PT_DIRECTORY_LEVEL)
1991                 level = PT_DIRECTORY_LEVEL;
1992
1993         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
1994
1995         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1996         smp_rmb();
1997         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1998
1999         /* mmio */
2000         if (is_error_pfn(pfn)) {
2001                 kvm_release_pfn_clean(pfn);
2002                 return 1;
2003         }
2004
2005         spin_lock(&vcpu->kvm->mmu_lock);
2006         if (mmu_notifier_retry(vcpu, mmu_seq))
2007                 goto out_unlock;
2008         kvm_mmu_free_some_pages(vcpu);
2009         r = __direct_map(vcpu, v, write, level, gfn, pfn);
2010         spin_unlock(&vcpu->kvm->mmu_lock);
2011
2012
2013         return r;
2014
2015 out_unlock:
2016         spin_unlock(&vcpu->kvm->mmu_lock);
2017         kvm_release_pfn_clean(pfn);
2018         return 0;
2019 }
2020
2021
2022 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2023 {
2024         int i;
2025         struct kvm_mmu_page *sp;
2026
2027         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2028                 return;
2029         spin_lock(&vcpu->kvm->mmu_lock);
2030         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2031                 hpa_t root = vcpu->arch.mmu.root_hpa;
2032
2033                 sp = page_header(root);
2034                 --sp->root_count;
2035                 if (!sp->root_count && sp->role.invalid)
2036                         kvm_mmu_zap_page(vcpu->kvm, sp);
2037                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2038                 spin_unlock(&vcpu->kvm->mmu_lock);
2039                 return;
2040         }
2041         for (i = 0; i < 4; ++i) {
2042                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2043
2044                 if (root) {
2045                         root &= PT64_BASE_ADDR_MASK;
2046                         sp = page_header(root);
2047                         --sp->root_count;
2048                         if (!sp->root_count && sp->role.invalid)
2049                                 kvm_mmu_zap_page(vcpu->kvm, sp);
2050                 }
2051                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2052         }
2053         spin_unlock(&vcpu->kvm->mmu_lock);
2054         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2055 }
2056
2057 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2058 {
2059         int ret = 0;
2060
2061         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2062                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2063                 ret = 1;
2064         }
2065
2066         return ret;
2067 }
2068
2069 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2070 {
2071         int i;
2072         gfn_t root_gfn;
2073         struct kvm_mmu_page *sp;
2074         int direct = 0;
2075         u64 pdptr;
2076
2077         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2078
2079         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2080                 hpa_t root = vcpu->arch.mmu.root_hpa;
2081
2082                 ASSERT(!VALID_PAGE(root));
2083                 if (tdp_enabled)
2084                         direct = 1;
2085                 if (mmu_check_root(vcpu, root_gfn))
2086                         return 1;
2087                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2088                                       PT64_ROOT_LEVEL, direct,
2089                                       ACC_ALL, NULL);
2090                 root = __pa(sp->spt);
2091                 ++sp->root_count;
2092                 vcpu->arch.mmu.root_hpa = root;
2093                 return 0;
2094         }
2095         direct = !is_paging(vcpu);
2096         if (tdp_enabled)
2097                 direct = 1;
2098         for (i = 0; i < 4; ++i) {
2099                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2100
2101                 ASSERT(!VALID_PAGE(root));
2102                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2103                         pdptr = kvm_pdptr_read(vcpu, i);
2104                         if (!is_present_gpte(pdptr)) {
2105                                 vcpu->arch.mmu.pae_root[i] = 0;
2106                                 continue;
2107                         }
2108                         root_gfn = pdptr >> PAGE_SHIFT;
2109                 } else if (vcpu->arch.mmu.root_level == 0)
2110                         root_gfn = 0;
2111                 if (mmu_check_root(vcpu, root_gfn))
2112                         return 1;
2113                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2114                                       PT32_ROOT_LEVEL, direct,
2115                                       ACC_ALL, NULL);
2116                 root = __pa(sp->spt);
2117                 ++sp->root_count;
2118                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2119         }
2120         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2121         return 0;
2122 }
2123
2124 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2125 {
2126         int i;
2127         struct kvm_mmu_page *sp;
2128
2129         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2130                 return;
2131         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2132                 hpa_t root = vcpu->arch.mmu.root_hpa;
2133                 sp = page_header(root);
2134                 mmu_sync_children(vcpu, sp);
2135                 return;
2136         }
2137         for (i = 0; i < 4; ++i) {
2138                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2139
2140                 if (root && VALID_PAGE(root)) {
2141                         root &= PT64_BASE_ADDR_MASK;
2142                         sp = page_header(root);
2143                         mmu_sync_children(vcpu, sp);
2144                 }
2145         }
2146 }
2147
2148 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2149 {
2150         spin_lock(&vcpu->kvm->mmu_lock);
2151         mmu_sync_roots(vcpu);
2152         spin_unlock(&vcpu->kvm->mmu_lock);
2153 }
2154
2155 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2156                                   u32 access, u32 *error)
2157 {
2158         if (error)
2159                 *error = 0;
2160         return vaddr;
2161 }
2162
2163 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2164                                 u32 error_code)
2165 {
2166         gfn_t gfn;
2167         int r;
2168
2169         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2170         r = mmu_topup_memory_caches(vcpu);
2171         if (r)
2172                 return r;
2173
2174         ASSERT(vcpu);
2175         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2176
2177         gfn = gva >> PAGE_SHIFT;
2178
2179         return nonpaging_map(vcpu, gva & PAGE_MASK,
2180                              error_code & PFERR_WRITE_MASK, gfn);
2181 }
2182
2183 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2184                                 u32 error_code)
2185 {
2186         pfn_t pfn;
2187         int r;
2188         int level;
2189         gfn_t gfn = gpa >> PAGE_SHIFT;
2190         unsigned long mmu_seq;
2191
2192         ASSERT(vcpu);
2193         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2194
2195         r = mmu_topup_memory_caches(vcpu);
2196         if (r)
2197                 return r;
2198
2199         level = mapping_level(vcpu, gfn);
2200
2201         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2202
2203         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2204         smp_rmb();
2205         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2206         if (is_error_pfn(pfn)) {
2207                 kvm_release_pfn_clean(pfn);
2208                 return 1;
2209         }
2210         spin_lock(&vcpu->kvm->mmu_lock);
2211         if (mmu_notifier_retry(vcpu, mmu_seq))
2212                 goto out_unlock;
2213         kvm_mmu_free_some_pages(vcpu);
2214         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2215                          level, gfn, pfn);
2216         spin_unlock(&vcpu->kvm->mmu_lock);
2217
2218         return r;
2219
2220 out_unlock:
2221         spin_unlock(&vcpu->kvm->mmu_lock);
2222         kvm_release_pfn_clean(pfn);
2223         return 0;
2224 }
2225
2226 static void nonpaging_free(struct kvm_vcpu *vcpu)
2227 {
2228         mmu_free_roots(vcpu);
2229 }
2230
2231 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2232 {
2233         struct kvm_mmu *context = &vcpu->arch.mmu;
2234
2235         context->new_cr3 = nonpaging_new_cr3;
2236         context->page_fault = nonpaging_page_fault;
2237         context->gva_to_gpa = nonpaging_gva_to_gpa;
2238         context->free = nonpaging_free;
2239         context->prefetch_page = nonpaging_prefetch_page;
2240         context->sync_page = nonpaging_sync_page;
2241         context->invlpg = nonpaging_invlpg;
2242         context->root_level = 0;
2243         context->shadow_root_level = PT32E_ROOT_LEVEL;
2244         context->root_hpa = INVALID_PAGE;
2245         return 0;
2246 }
2247
2248 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2249 {
2250         ++vcpu->stat.tlb_flush;
2251         kvm_x86_ops->tlb_flush(vcpu);
2252 }
2253
2254 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2255 {
2256         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2257         mmu_free_roots(vcpu);
2258 }
2259
2260 static void inject_page_fault(struct kvm_vcpu *vcpu,
2261                               u64 addr,
2262                               u32 err_code)
2263 {
2264         kvm_inject_page_fault(vcpu, addr, err_code);
2265 }
2266
2267 static void paging_free(struct kvm_vcpu *vcpu)
2268 {
2269         nonpaging_free(vcpu);
2270 }
2271
2272 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2273 {
2274         int bit7;
2275
2276         bit7 = (gpte >> 7) & 1;
2277         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2278 }
2279
2280 #define PTTYPE 64
2281 #include "paging_tmpl.h"
2282 #undef PTTYPE
2283
2284 #define PTTYPE 32
2285 #include "paging_tmpl.h"
2286 #undef PTTYPE
2287
2288 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2289 {
2290         struct kvm_mmu *context = &vcpu->arch.mmu;
2291         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2292         u64 exb_bit_rsvd = 0;
2293
2294         if (!is_nx(vcpu))
2295                 exb_bit_rsvd = rsvd_bits(63, 63);
2296         switch (level) {
2297         case PT32_ROOT_LEVEL:
2298                 /* no rsvd bits for 2 level 4K page table entries */
2299                 context->rsvd_bits_mask[0][1] = 0;
2300                 context->rsvd_bits_mask[0][0] = 0;
2301                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2302
2303                 if (!is_pse(vcpu)) {
2304                         context->rsvd_bits_mask[1][1] = 0;
2305                         break;
2306                 }
2307
2308                 if (is_cpuid_PSE36())
2309                         /* 36bits PSE 4MB page */
2310                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2311                 else
2312                         /* 32 bits PSE 4MB page */
2313                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2314                 break;
2315         case PT32E_ROOT_LEVEL:
2316                 context->rsvd_bits_mask[0][2] =
2317                         rsvd_bits(maxphyaddr, 63) |
2318                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2319                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2320                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2321                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2322                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2323                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2324                         rsvd_bits(maxphyaddr, 62) |
2325                         rsvd_bits(13, 20);              /* large page */
2326                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2327                 break;
2328         case PT64_ROOT_LEVEL:
2329                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2330                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2331                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2332                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2333                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2334                         rsvd_bits(maxphyaddr, 51);
2335                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2336                         rsvd_bits(maxphyaddr, 51);
2337                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2338                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2339                         rsvd_bits(maxphyaddr, 51) |
2340                         rsvd_bits(13, 29);
2341                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2342                         rsvd_bits(maxphyaddr, 51) |
2343                         rsvd_bits(13, 20);              /* large page */
2344                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2345                 break;
2346         }
2347 }
2348
2349 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2350 {
2351         struct kvm_mmu *context = &vcpu->arch.mmu;
2352
2353         ASSERT(is_pae(vcpu));
2354         context->new_cr3 = paging_new_cr3;
2355         context->page_fault = paging64_page_fault;
2356         context->gva_to_gpa = paging64_gva_to_gpa;
2357         context->prefetch_page = paging64_prefetch_page;
2358         context->sync_page = paging64_sync_page;
2359         context->invlpg = paging64_invlpg;
2360         context->free = paging_free;
2361         context->root_level = level;
2362         context->shadow_root_level = level;
2363         context->root_hpa = INVALID_PAGE;
2364         return 0;
2365 }
2366
2367 static int paging64_init_context(struct kvm_vcpu *vcpu)
2368 {
2369         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2370         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2371 }
2372
2373 static int paging32_init_context(struct kvm_vcpu *vcpu)
2374 {
2375         struct kvm_mmu *context = &vcpu->arch.mmu;
2376
2377         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2378         context->new_cr3 = paging_new_cr3;
2379         context->page_fault = paging32_page_fault;
2380         context->gva_to_gpa = paging32_gva_to_gpa;
2381         context->free = paging_free;
2382         context->prefetch_page = paging32_prefetch_page;
2383         context->sync_page = paging32_sync_page;
2384         context->invlpg = paging32_invlpg;
2385         context->root_level = PT32_ROOT_LEVEL;
2386         context->shadow_root_level = PT32E_ROOT_LEVEL;
2387         context->root_hpa = INVALID_PAGE;
2388         return 0;
2389 }
2390
2391 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2392 {
2393         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2394         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2395 }
2396
2397 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2398 {
2399         struct kvm_mmu *context = &vcpu->arch.mmu;
2400
2401         context->new_cr3 = nonpaging_new_cr3;
2402         context->page_fault = tdp_page_fault;
2403         context->free = nonpaging_free;
2404         context->prefetch_page = nonpaging_prefetch_page;
2405         context->sync_page = nonpaging_sync_page;
2406         context->invlpg = nonpaging_invlpg;
2407         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2408         context->root_hpa = INVALID_PAGE;
2409
2410         if (!is_paging(vcpu)) {
2411                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2412                 context->root_level = 0;
2413         } else if (is_long_mode(vcpu)) {
2414                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2415                 context->gva_to_gpa = paging64_gva_to_gpa;
2416                 context->root_level = PT64_ROOT_LEVEL;
2417         } else if (is_pae(vcpu)) {
2418                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2419                 context->gva_to_gpa = paging64_gva_to_gpa;
2420                 context->root_level = PT32E_ROOT_LEVEL;
2421         } else {
2422                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2423                 context->gva_to_gpa = paging32_gva_to_gpa;
2424                 context->root_level = PT32_ROOT_LEVEL;
2425         }
2426
2427         return 0;
2428 }
2429
2430 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2431 {
2432         int r;
2433
2434         ASSERT(vcpu);
2435         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2436
2437         if (!is_paging(vcpu))
2438                 r = nonpaging_init_context(vcpu);
2439         else if (is_long_mode(vcpu))
2440                 r = paging64_init_context(vcpu);
2441         else if (is_pae(vcpu))
2442                 r = paging32E_init_context(vcpu);
2443         else
2444                 r = paging32_init_context(vcpu);
2445
2446         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2447
2448         return r;
2449 }
2450
2451 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2452 {
2453         vcpu->arch.update_pte.pfn = bad_pfn;
2454
2455         if (tdp_enabled)
2456                 return init_kvm_tdp_mmu(vcpu);
2457         else
2458                 return init_kvm_softmmu(vcpu);
2459 }
2460
2461 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2462 {
2463         ASSERT(vcpu);
2464         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2465                 vcpu->arch.mmu.free(vcpu);
2466                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2467         }
2468 }
2469
2470 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2471 {
2472         destroy_kvm_mmu(vcpu);
2473         return init_kvm_mmu(vcpu);
2474 }
2475 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2476
2477 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2478 {
2479         int r;
2480
2481         r = mmu_topup_memory_caches(vcpu);
2482         if (r)
2483                 goto out;
2484         spin_lock(&vcpu->kvm->mmu_lock);
2485         kvm_mmu_free_some_pages(vcpu);
2486         r = mmu_alloc_roots(vcpu);
2487         mmu_sync_roots(vcpu);
2488         spin_unlock(&vcpu->kvm->mmu_lock);
2489         if (r)
2490                 goto out;
2491         /* set_cr3() should ensure TLB has been flushed */
2492         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2493 out:
2494         return r;
2495 }
2496 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2497
2498 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2499 {
2500         mmu_free_roots(vcpu);
2501 }
2502
2503 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2504                                   struct kvm_mmu_page *sp,
2505                                   u64 *spte)
2506 {
2507         u64 pte;
2508         struct kvm_mmu_page *child;
2509
2510         pte = *spte;
2511         if (is_shadow_present_pte(pte)) {
2512                 if (is_last_spte(pte, sp->role.level))
2513                         rmap_remove(vcpu->kvm, spte);
2514                 else {
2515                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2516                         mmu_page_remove_parent_pte(child, spte);
2517                 }
2518         }
2519         __set_spte(spte, shadow_trap_nonpresent_pte);
2520         if (is_large_pte(pte))
2521                 --vcpu->kvm->stat.lpages;
2522 }
2523
2524 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2525                                   struct kvm_mmu_page *sp,
2526                                   u64 *spte,
2527                                   const void *new)
2528 {
2529         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2530                 ++vcpu->kvm->stat.mmu_pde_zapped;
2531                 return;
2532         }
2533
2534         ++vcpu->kvm->stat.mmu_pte_updated;
2535         if (sp->role.glevels == PT32_ROOT_LEVEL)
2536                 paging32_update_pte(vcpu, sp, spte, new);
2537         else
2538                 paging64_update_pte(vcpu, sp, spte, new);
2539 }
2540
2541 static bool need_remote_flush(u64 old, u64 new)
2542 {
2543         if (!is_shadow_present_pte(old))
2544                 return false;
2545         if (!is_shadow_present_pte(new))
2546                 return true;
2547         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2548                 return true;
2549         old ^= PT64_NX_MASK;
2550         new ^= PT64_NX_MASK;
2551         return (old & ~new & PT64_PERM_MASK) != 0;
2552 }
2553
2554 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2555 {
2556         if (need_remote_flush(old, new))
2557                 kvm_flush_remote_tlbs(vcpu->kvm);
2558         else
2559                 kvm_mmu_flush_tlb(vcpu);
2560 }
2561
2562 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2563 {
2564         u64 *spte = vcpu->arch.last_pte_updated;
2565
2566         return !!(spte && (*spte & shadow_accessed_mask));
2567 }
2568
2569 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2570                                           u64 gpte)
2571 {
2572         gfn_t gfn;
2573         pfn_t pfn;
2574
2575         if (!is_present_gpte(gpte))
2576                 return;
2577         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2578
2579         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2580         smp_rmb();
2581         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2582
2583         if (is_error_pfn(pfn)) {
2584                 kvm_release_pfn_clean(pfn);
2585                 return;
2586         }
2587         vcpu->arch.update_pte.gfn = gfn;
2588         vcpu->arch.update_pte.pfn = pfn;
2589 }
2590
2591 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2592 {
2593         u64 *spte = vcpu->arch.last_pte_updated;
2594
2595         if (spte
2596             && vcpu->arch.last_pte_gfn == gfn
2597             && shadow_accessed_mask
2598             && !(*spte & shadow_accessed_mask)
2599             && is_shadow_present_pte(*spte))
2600                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2601 }
2602
2603 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2604                        const u8 *new, int bytes,
2605                        bool guest_initiated)
2606 {
2607         gfn_t gfn = gpa >> PAGE_SHIFT;
2608         struct kvm_mmu_page *sp;
2609         struct hlist_node *node, *n;
2610         struct hlist_head *bucket;
2611         unsigned index;
2612         u64 entry, gentry;
2613         u64 *spte;
2614         unsigned offset = offset_in_page(gpa);
2615         unsigned pte_size;
2616         unsigned page_offset;
2617         unsigned misaligned;
2618         unsigned quadrant;
2619         int level;
2620         int flooded = 0;
2621         int npte;
2622         int r;
2623         int invlpg_counter;
2624
2625         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2626
2627         invlpg_counter = atomic_read(&vcpu->kvm->arch.invlpg_counter);
2628
2629         /*
2630          * Assume that the pte write on a page table of the same type
2631          * as the current vcpu paging mode.  This is nearly always true
2632          * (might be false while changing modes).  Note it is verified later
2633          * by update_pte().
2634          */
2635         if ((is_pae(vcpu) && bytes == 4) || !new) {
2636                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2637                 if (is_pae(vcpu)) {
2638                         gpa &= ~(gpa_t)7;
2639                         bytes = 8;
2640                 }
2641                 r = kvm_read_guest(vcpu->kvm, gpa, &gentry, min(bytes, 8));
2642                 if (r)
2643                         gentry = 0;
2644                 new = (const u8 *)&gentry;
2645         }
2646
2647         switch (bytes) {
2648         case 4:
2649                 gentry = *(const u32 *)new;
2650                 break;
2651         case 8:
2652                 gentry = *(const u64 *)new;
2653                 break;
2654         default:
2655                 gentry = 0;
2656                 break;
2657         }
2658
2659         mmu_guess_page_from_pte_write(vcpu, gpa, gentry);
2660         spin_lock(&vcpu->kvm->mmu_lock);
2661         if (atomic_read(&vcpu->kvm->arch.invlpg_counter) != invlpg_counter)
2662                 gentry = 0;
2663         kvm_mmu_access_page(vcpu, gfn);
2664         kvm_mmu_free_some_pages(vcpu);
2665         ++vcpu->kvm->stat.mmu_pte_write;
2666         kvm_mmu_audit(vcpu, "pre pte write");
2667         if (guest_initiated) {
2668                 if (gfn == vcpu->arch.last_pt_write_gfn
2669                     && !last_updated_pte_accessed(vcpu)) {
2670                         ++vcpu->arch.last_pt_write_count;
2671                         if (vcpu->arch.last_pt_write_count >= 3)
2672                                 flooded = 1;
2673                 } else {
2674                         vcpu->arch.last_pt_write_gfn = gfn;
2675                         vcpu->arch.last_pt_write_count = 1;
2676                         vcpu->arch.last_pte_updated = NULL;
2677                 }
2678         }
2679         index = kvm_page_table_hashfn(gfn);
2680         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2681         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2682                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2683                         continue;
2684                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2685                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2686                 misaligned |= bytes < 4;
2687                 if (misaligned || flooded) {
2688                         /*
2689                          * Misaligned accesses are too much trouble to fix
2690                          * up; also, they usually indicate a page is not used
2691                          * as a page table.
2692                          *
2693                          * If we're seeing too many writes to a page,
2694                          * it may no longer be a page table, or we may be
2695                          * forking, in which case it is better to unmap the
2696                          * page.
2697                          */
2698                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2699                                  gpa, bytes, sp->role.word);
2700                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2701                                 n = bucket->first;
2702                         ++vcpu->kvm->stat.mmu_flooded;
2703                         continue;
2704                 }
2705                 page_offset = offset;
2706                 level = sp->role.level;
2707                 npte = 1;
2708                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2709                         page_offset <<= 1;      /* 32->64 */
2710                         /*
2711                          * A 32-bit pde maps 4MB while the shadow pdes map
2712                          * only 2MB.  So we need to double the offset again
2713                          * and zap two pdes instead of one.
2714                          */
2715                         if (level == PT32_ROOT_LEVEL) {
2716                                 page_offset &= ~7; /* kill rounding error */
2717                                 page_offset <<= 1;
2718                                 npte = 2;
2719                         }
2720                         quadrant = page_offset >> PAGE_SHIFT;
2721                         page_offset &= ~PAGE_MASK;
2722                         if (quadrant != sp->role.quadrant)
2723                                 continue;
2724                 }
2725                 spte = &sp->spt[page_offset / sizeof(*spte)];
2726                 while (npte--) {
2727                         entry = *spte;
2728                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2729                         if (gentry)
2730                                 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
2731                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2732                         ++spte;
2733                 }
2734         }
2735         kvm_mmu_audit(vcpu, "post pte write");
2736         spin_unlock(&vcpu->kvm->mmu_lock);
2737         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2738                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2739                 vcpu->arch.update_pte.pfn = bad_pfn;
2740         }
2741 }
2742
2743 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2744 {
2745         gpa_t gpa;
2746         int r;
2747
2748         if (tdp_enabled)
2749                 return 0;
2750
2751         gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2752
2753         spin_lock(&vcpu->kvm->mmu_lock);
2754         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2755         spin_unlock(&vcpu->kvm->mmu_lock);
2756         return r;
2757 }
2758 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2759
2760 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2761 {
2762         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2763                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2764                 struct kvm_mmu_page *sp;
2765
2766                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2767                                   struct kvm_mmu_page, link);
2768                 kvm_mmu_zap_page(vcpu->kvm, sp);
2769                 ++vcpu->kvm->stat.mmu_recycled;
2770         }
2771 }
2772
2773 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2774 {
2775         int r;
2776         enum emulation_result er;
2777
2778         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2779         if (r < 0)
2780                 goto out;
2781
2782         if (!r) {
2783                 r = 1;
2784                 goto out;
2785         }
2786
2787         r = mmu_topup_memory_caches(vcpu);
2788         if (r)
2789                 goto out;
2790
2791         er = emulate_instruction(vcpu, cr2, error_code, 0);
2792
2793         switch (er) {
2794         case EMULATE_DONE:
2795                 return 1;
2796         case EMULATE_DO_MMIO:
2797                 ++vcpu->stat.mmio_exits;
2798                 return 0;
2799         case EMULATE_FAIL:
2800                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2801                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2802                 vcpu->run->internal.ndata = 0;
2803                 return 0;
2804         default:
2805                 BUG();
2806         }
2807 out:
2808         return r;
2809 }
2810 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2811
2812 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2813 {
2814         vcpu->arch.mmu.invlpg(vcpu, gva);
2815         kvm_mmu_flush_tlb(vcpu);
2816         ++vcpu->stat.invlpg;
2817 }
2818 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2819
2820 void kvm_enable_tdp(void)
2821 {
2822         tdp_enabled = true;
2823 }
2824 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2825
2826 void kvm_disable_tdp(void)
2827 {
2828         tdp_enabled = false;
2829 }
2830 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2831
2832 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2833 {
2834         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2835 }
2836
2837 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2838 {
2839         struct page *page;
2840         int i;
2841
2842         ASSERT(vcpu);
2843
2844         /*
2845          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2846          * Therefore we need to allocate shadow page tables in the first
2847          * 4GB of memory, which happens to fit the DMA32 zone.
2848          */
2849         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2850         if (!page)
2851                 return -ENOMEM;
2852
2853         vcpu->arch.mmu.pae_root = page_address(page);
2854         for (i = 0; i < 4; ++i)
2855                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2856
2857         return 0;
2858 }
2859
2860 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2861 {
2862         ASSERT(vcpu);
2863         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2864
2865         return alloc_mmu_pages(vcpu);
2866 }
2867
2868 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2869 {
2870         ASSERT(vcpu);
2871         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2872
2873         return init_kvm_mmu(vcpu);
2874 }
2875
2876 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2877 {
2878         ASSERT(vcpu);
2879
2880         destroy_kvm_mmu(vcpu);
2881         free_mmu_pages(vcpu);
2882         mmu_free_memory_caches(vcpu);
2883 }
2884
2885 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2886 {
2887         struct kvm_mmu_page *sp;
2888
2889         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2890                 int i;
2891                 u64 *pt;
2892
2893                 if (!test_bit(slot, sp->slot_bitmap))
2894                         continue;
2895
2896                 pt = sp->spt;
2897                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2898                         /* avoid RMW */
2899                         if (pt[i] & PT_WRITABLE_MASK)
2900                                 pt[i] &= ~PT_WRITABLE_MASK;
2901         }
2902         kvm_flush_remote_tlbs(kvm);
2903 }
2904
2905 void kvm_mmu_zap_all(struct kvm *kvm)
2906 {
2907         struct kvm_mmu_page *sp, *node;
2908
2909         spin_lock(&kvm->mmu_lock);
2910         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2911                 if (kvm_mmu_zap_page(kvm, sp))
2912                         node = container_of(kvm->arch.active_mmu_pages.next,
2913                                             struct kvm_mmu_page, link);
2914         spin_unlock(&kvm->mmu_lock);
2915
2916         kvm_flush_remote_tlbs(kvm);
2917 }
2918
2919 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2920 {
2921         struct kvm_mmu_page *page;
2922
2923         page = container_of(kvm->arch.active_mmu_pages.prev,
2924                             struct kvm_mmu_page, link);
2925         kvm_mmu_zap_page(kvm, page);
2926 }
2927
2928 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2929 {
2930         struct kvm *kvm;
2931         struct kvm *kvm_freed = NULL;
2932         int cache_count = 0;
2933
2934         spin_lock(&kvm_lock);
2935
2936         list_for_each_entry(kvm, &vm_list, vm_list) {
2937                 int npages, idx;
2938
2939                 idx = srcu_read_lock(&kvm->srcu);
2940                 spin_lock(&kvm->mmu_lock);
2941                 npages = kvm->arch.n_alloc_mmu_pages -
2942                          kvm->arch.n_free_mmu_pages;
2943                 cache_count += npages;
2944                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2945                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2946                         cache_count--;
2947                         kvm_freed = kvm;
2948                 }
2949                 nr_to_scan--;
2950
2951                 spin_unlock(&kvm->mmu_lock);
2952                 srcu_read_unlock(&kvm->srcu, idx);
2953         }
2954         if (kvm_freed)
2955                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2956
2957         spin_unlock(&kvm_lock);
2958
2959         return cache_count;
2960 }
2961
2962 static struct shrinker mmu_shrinker = {
2963         .shrink = mmu_shrink,
2964         .seeks = DEFAULT_SEEKS * 10,
2965 };
2966
2967 static void mmu_destroy_caches(void)
2968 {
2969         if (pte_chain_cache)
2970                 kmem_cache_destroy(pte_chain_cache);
2971         if (rmap_desc_cache)
2972                 kmem_cache_destroy(rmap_desc_cache);
2973         if (mmu_page_header_cache)
2974                 kmem_cache_destroy(mmu_page_header_cache);
2975 }
2976
2977 void kvm_mmu_module_exit(void)
2978 {
2979         mmu_destroy_caches();
2980         unregister_shrinker(&mmu_shrinker);
2981 }
2982
2983 int kvm_mmu_module_init(void)
2984 {
2985         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2986                                             sizeof(struct kvm_pte_chain),
2987                                             0, 0, NULL);
2988         if (!pte_chain_cache)
2989                 goto nomem;
2990         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2991                                             sizeof(struct kvm_rmap_desc),
2992                                             0, 0, NULL);
2993         if (!rmap_desc_cache)
2994                 goto nomem;
2995
2996         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2997                                                   sizeof(struct kvm_mmu_page),
2998                                                   0, 0, NULL);
2999         if (!mmu_page_header_cache)
3000                 goto nomem;
3001
3002         register_shrinker(&mmu_shrinker);
3003
3004         return 0;
3005
3006 nomem:
3007         mmu_destroy_caches();
3008         return -ENOMEM;
3009 }
3010
3011 /*
3012  * Caculate mmu pages needed for kvm.
3013  */
3014 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3015 {
3016         int i;
3017         unsigned int nr_mmu_pages;
3018         unsigned int  nr_pages = 0;
3019         struct kvm_memslots *slots;
3020
3021         slots = rcu_dereference(kvm->memslots);
3022         for (i = 0; i < slots->nmemslots; i++)
3023                 nr_pages += slots->memslots[i].npages;
3024
3025         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3026         nr_mmu_pages = max(nr_mmu_pages,
3027                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3028
3029         return nr_mmu_pages;
3030 }
3031
3032 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3033                                 unsigned len)
3034 {
3035         if (len > buffer->len)
3036                 return NULL;
3037         return buffer->ptr;
3038 }
3039
3040 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3041                                 unsigned len)
3042 {
3043         void *ret;
3044
3045         ret = pv_mmu_peek_buffer(buffer, len);
3046         if (!ret)
3047                 return ret;
3048         buffer->ptr += len;
3049         buffer->len -= len;
3050         buffer->processed += len;
3051         return ret;
3052 }
3053
3054 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3055                              gpa_t addr, gpa_t value)
3056 {
3057         int bytes = 8;
3058         int r;
3059
3060         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3061                 bytes = 4;
3062
3063         r = mmu_topup_memory_caches(vcpu);
3064         if (r)
3065                 return r;
3066
3067         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3068                 return -EFAULT;
3069
3070         return 1;
3071 }
3072
3073 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3074 {
3075         kvm_set_cr3(vcpu, vcpu->arch.cr3);
3076         return 1;
3077 }
3078
3079 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3080 {
3081         spin_lock(&vcpu->kvm->mmu_lock);
3082         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3083         spin_unlock(&vcpu->kvm->mmu_lock);
3084         return 1;
3085 }
3086
3087 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3088                              struct kvm_pv_mmu_op_buffer *buffer)
3089 {
3090         struct kvm_mmu_op_header *header;
3091
3092         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3093         if (!header)
3094                 return 0;
3095         switch (header->op) {
3096         case KVM_MMU_OP_WRITE_PTE: {
3097                 struct kvm_mmu_op_write_pte *wpte;
3098
3099                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3100                 if (!wpte)
3101                         return 0;
3102                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3103                                         wpte->pte_val);
3104         }
3105         case KVM_MMU_OP_FLUSH_TLB: {
3106                 struct kvm_mmu_op_flush_tlb *ftlb;
3107
3108                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3109                 if (!ftlb)
3110                         return 0;
3111                 return kvm_pv_mmu_flush_tlb(vcpu);
3112         }
3113         case KVM_MMU_OP_RELEASE_PT: {
3114                 struct kvm_mmu_op_release_pt *rpt;
3115
3116                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3117                 if (!rpt)
3118                         return 0;
3119                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3120         }
3121         default: return 0;
3122         }
3123 }
3124
3125 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3126                   gpa_t addr, unsigned long *ret)
3127 {
3128         int r;
3129         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3130
3131         buffer->ptr = buffer->buf;
3132         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3133         buffer->processed = 0;
3134
3135         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3136         if (r)
3137                 goto out;
3138
3139         while (buffer->len) {
3140                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3141                 if (r < 0)
3142                         goto out;
3143                 if (r == 0)
3144                         break;
3145         }
3146
3147         r = 1;
3148 out:
3149         *ret = buffer->processed;
3150         return r;
3151 }
3152
3153 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3154 {
3155         struct kvm_shadow_walk_iterator iterator;
3156         int nr_sptes = 0;
3157
3158         spin_lock(&vcpu->kvm->mmu_lock);
3159         for_each_shadow_entry(vcpu, addr, iterator) {
3160                 sptes[iterator.level-1] = *iterator.sptep;
3161                 nr_sptes++;
3162                 if (!is_shadow_present_pte(*iterator.sptep))
3163                         break;
3164         }
3165         spin_unlock(&vcpu->kvm->mmu_lock);
3166
3167         return nr_sptes;
3168 }
3169 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3170
3171 #ifdef AUDIT
3172
3173 static const char *audit_msg;
3174
3175 static gva_t canonicalize(gva_t gva)
3176 {
3177 #ifdef CONFIG_X86_64
3178         gva = (long long)(gva << 16) >> 16;
3179 #endif
3180         return gva;
3181 }
3182
3183
3184 typedef void (*inspect_spte_fn) (struct kvm *kvm, u64 *sptep);
3185
3186 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3187                             inspect_spte_fn fn)
3188 {
3189         int i;
3190
3191         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3192                 u64 ent = sp->spt[i];
3193
3194                 if (is_shadow_present_pte(ent)) {
3195                         if (!is_last_spte(ent, sp->role.level)) {
3196                                 struct kvm_mmu_page *child;
3197                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3198                                 __mmu_spte_walk(kvm, child, fn);
3199                         } else
3200                                 fn(kvm, &sp->spt[i]);
3201                 }
3202         }
3203 }
3204
3205 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3206 {
3207         int i;
3208         struct kvm_mmu_page *sp;
3209
3210         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3211                 return;
3212         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3213                 hpa_t root = vcpu->arch.mmu.root_hpa;
3214                 sp = page_header(root);
3215                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3216                 return;
3217         }
3218         for (i = 0; i < 4; ++i) {
3219                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3220
3221                 if (root && VALID_PAGE(root)) {
3222                         root &= PT64_BASE_ADDR_MASK;
3223                         sp = page_header(root);
3224                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3225                 }
3226         }
3227         return;
3228 }
3229
3230 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3231                                 gva_t va, int level)
3232 {
3233         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3234         int i;
3235         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3236
3237         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3238                 u64 ent = pt[i];
3239
3240                 if (ent == shadow_trap_nonpresent_pte)
3241                         continue;
3242
3243                 va = canonicalize(va);
3244                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3245                         audit_mappings_page(vcpu, ent, va, level - 1);
3246                 else {
3247                         gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, va, NULL);
3248                         gfn_t gfn = gpa >> PAGE_SHIFT;
3249                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3250                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3251
3252                         if (is_error_pfn(pfn)) {
3253                                 kvm_release_pfn_clean(pfn);
3254                                 continue;
3255                         }
3256
3257                         if (is_shadow_present_pte(ent)
3258                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3259                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3260                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3261                                        audit_msg, vcpu->arch.mmu.root_level,
3262                                        va, gpa, hpa, ent,
3263                                        is_shadow_present_pte(ent));
3264                         else if (ent == shadow_notrap_nonpresent_pte
3265                                  && !is_error_hpa(hpa))
3266                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3267                                        " valid guest gva %lx\n", audit_msg, va);
3268                         kvm_release_pfn_clean(pfn);
3269
3270                 }
3271         }
3272 }
3273
3274 static void audit_mappings(struct kvm_vcpu *vcpu)
3275 {
3276         unsigned i;
3277
3278         if (vcpu->arch.mmu.root_level == 4)
3279                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3280         else
3281                 for (i = 0; i < 4; ++i)
3282                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3283                                 audit_mappings_page(vcpu,
3284                                                     vcpu->arch.mmu.pae_root[i],
3285                                                     i << 30,
3286                                                     2);
3287 }
3288
3289 static int count_rmaps(struct kvm_vcpu *vcpu)
3290 {
3291         struct kvm *kvm = vcpu->kvm;
3292         struct kvm_memslots *slots;
3293         int nmaps = 0;
3294         int i, j, k, idx;
3295
3296         idx = srcu_read_lock(&kvm->srcu);
3297         slots = rcu_dereference(kvm->memslots);
3298         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3299                 struct kvm_memory_slot *m = &slots->memslots[i];
3300                 struct kvm_rmap_desc *d;
3301
3302                 for (j = 0; j < m->npages; ++j) {
3303                         unsigned long *rmapp = &m->rmap[j];
3304
3305                         if (!*rmapp)
3306                                 continue;
3307                         if (!(*rmapp & 1)) {
3308                                 ++nmaps;
3309                                 continue;
3310                         }
3311                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3312                         while (d) {
3313                                 for (k = 0; k < RMAP_EXT; ++k)
3314                                         if (d->sptes[k])
3315                                                 ++nmaps;
3316                                         else
3317                                                 break;
3318                                 d = d->more;
3319                         }
3320                 }
3321         }
3322         srcu_read_unlock(&kvm->srcu, idx);
3323         return nmaps;
3324 }
3325
3326 void inspect_spte_has_rmap(struct kvm *kvm, u64 *sptep)
3327 {
3328         unsigned long *rmapp;
3329         struct kvm_mmu_page *rev_sp;
3330         gfn_t gfn;
3331
3332         if (*sptep & PT_WRITABLE_MASK) {
3333                 rev_sp = page_header(__pa(sptep));
3334                 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3335
3336                 if (!gfn_to_memslot(kvm, gfn)) {
3337                         if (!printk_ratelimit())
3338                                 return;
3339                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3340                                          audit_msg, gfn);
3341                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3342                                audit_msg, (long int)(sptep - rev_sp->spt),
3343                                         rev_sp->gfn);
3344                         dump_stack();
3345                         return;
3346                 }
3347
3348                 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3349                                     rev_sp->role.level);
3350                 if (!*rmapp) {
3351                         if (!printk_ratelimit())
3352                                 return;
3353                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3354                                          audit_msg, *sptep);
3355                         dump_stack();
3356                 }
3357         }
3358
3359 }
3360
3361 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3362 {
3363         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3364 }
3365
3366 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3367 {
3368         struct kvm_mmu_page *sp;
3369         int i;
3370
3371         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3372                 u64 *pt = sp->spt;
3373
3374                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3375                         continue;
3376
3377                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3378                         u64 ent = pt[i];
3379
3380                         if (!(ent & PT_PRESENT_MASK))
3381                                 continue;
3382                         if (!(ent & PT_WRITABLE_MASK))
3383                                 continue;
3384                         inspect_spte_has_rmap(vcpu->kvm, &pt[i]);
3385                 }
3386         }
3387         return;
3388 }
3389
3390 static void audit_rmap(struct kvm_vcpu *vcpu)
3391 {
3392         check_writable_mappings_rmap(vcpu);
3393         count_rmaps(vcpu);
3394 }
3395
3396 static void audit_write_protection(struct kvm_vcpu *vcpu)
3397 {
3398         struct kvm_mmu_page *sp;
3399         struct kvm_memory_slot *slot;
3400         unsigned long *rmapp;
3401         u64 *spte;
3402         gfn_t gfn;
3403
3404         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3405                 if (sp->role.direct)
3406                         continue;
3407                 if (sp->unsync)
3408                         continue;
3409
3410                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3411                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3412                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3413
3414                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3415                 while (spte) {
3416                         if (*spte & PT_WRITABLE_MASK)
3417                                 printk(KERN_ERR "%s: (%s) shadow page has "
3418                                 "writable mappings: gfn %lx role %x\n",
3419                                __func__, audit_msg, sp->gfn,
3420                                sp->role.word);
3421                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3422                 }
3423         }
3424 }
3425
3426 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3427 {
3428         int olddbg = dbg;
3429
3430         dbg = 0;
3431         audit_msg = msg;
3432         audit_rmap(vcpu);
3433         audit_write_protection(vcpu);
3434         if (strcmp("pre pte write", audit_msg) != 0)
3435                 audit_mappings(vcpu);
3436         audit_writable_sptes_have_rmaps(vcpu);
3437         dbg = olddbg;
3438 }
3439
3440 #endif