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