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         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1569                 if (sp->gfn == gfn && !sp->role.direct) {
1570                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1571                                  sp->role.word);
1572                         r = 1;
1573                         if (kvm_mmu_zap_page(kvm, sp))
1574                                 n = bucket->first;
1575                 }
1576         return r;
1577 }
1578
1579 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1580 {
1581         unsigned index;
1582         struct hlist_head *bucket;
1583         struct kvm_mmu_page *sp;
1584         struct hlist_node *node, *nn;
1585
1586         index = kvm_page_table_hashfn(gfn);
1587         bucket = &kvm->arch.mmu_page_hash[index];
1588         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1589                 if (sp->gfn == gfn && !sp->role.direct
1590                     && !sp->role.invalid) {
1591                         pgprintk("%s: zap %lx %x\n",
1592                                  __func__, gfn, sp->role.word);
1593                         if (kvm_mmu_zap_page(kvm, sp))
1594                                 nn = bucket->first;
1595                 }
1596         }
1597 }
1598
1599 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1600 {
1601         int slot = memslot_id(kvm, gfn);
1602         struct kvm_mmu_page *sp = page_header(__pa(pte));
1603
1604         __set_bit(slot, sp->slot_bitmap);
1605 }
1606
1607 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1608 {
1609         int i;
1610         u64 *pt = sp->spt;
1611
1612         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1613                 return;
1614
1615         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1616                 if (pt[i] == shadow_notrap_nonpresent_pte)
1617                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1618         }
1619 }
1620
1621 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1622 {
1623         struct page *page;
1624
1625         gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
1626
1627         if (gpa == UNMAPPED_GVA)
1628                 return NULL;
1629
1630         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1631
1632         return page;
1633 }
1634
1635 /*
1636  * The function is based on mtrr_type_lookup() in
1637  * arch/x86/kernel/cpu/mtrr/generic.c
1638  */
1639 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1640                          u64 start, u64 end)
1641 {
1642         int i;
1643         u64 base, mask;
1644         u8 prev_match, curr_match;
1645         int num_var_ranges = KVM_NR_VAR_MTRR;
1646
1647         if (!mtrr_state->enabled)
1648                 return 0xFF;
1649
1650         /* Make end inclusive end, instead of exclusive */
1651         end--;
1652
1653         /* Look in fixed ranges. Just return the type as per start */
1654         if (mtrr_state->have_fixed && (start < 0x100000)) {
1655                 int idx;
1656
1657                 if (start < 0x80000) {
1658                         idx = 0;
1659                         idx += (start >> 16);
1660                         return mtrr_state->fixed_ranges[idx];
1661                 } else if (start < 0xC0000) {
1662                         idx = 1 * 8;
1663                         idx += ((start - 0x80000) >> 14);
1664                         return mtrr_state->fixed_ranges[idx];
1665                 } else if (start < 0x1000000) {
1666                         idx = 3 * 8;
1667                         idx += ((start - 0xC0000) >> 12);
1668                         return mtrr_state->fixed_ranges[idx];
1669                 }
1670         }
1671
1672         /*
1673          * Look in variable ranges
1674          * Look of multiple ranges matching this address and pick type
1675          * as per MTRR precedence
1676          */
1677         if (!(mtrr_state->enabled & 2))
1678                 return mtrr_state->def_type;
1679
1680         prev_match = 0xFF;
1681         for (i = 0; i < num_var_ranges; ++i) {
1682                 unsigned short start_state, end_state;
1683
1684                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1685                         continue;
1686
1687                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1688                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1689                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1690                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1691
1692                 start_state = ((start & mask) == (base & mask));
1693                 end_state = ((end & mask) == (base & mask));
1694                 if (start_state != end_state)
1695                         return 0xFE;
1696
1697                 if ((start & mask) != (base & mask))
1698                         continue;
1699
1700                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1701                 if (prev_match == 0xFF) {
1702                         prev_match = curr_match;
1703                         continue;
1704                 }
1705
1706                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1707                     curr_match == MTRR_TYPE_UNCACHABLE)
1708                         return MTRR_TYPE_UNCACHABLE;
1709
1710                 if ((prev_match == MTRR_TYPE_WRBACK &&
1711                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1712                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1713                      curr_match == MTRR_TYPE_WRBACK)) {
1714                         prev_match = MTRR_TYPE_WRTHROUGH;
1715                         curr_match = MTRR_TYPE_WRTHROUGH;
1716                 }
1717
1718                 if (prev_match != curr_match)
1719                         return MTRR_TYPE_UNCACHABLE;
1720         }
1721
1722         if (prev_match != 0xFF)
1723                 return prev_match;
1724
1725         return mtrr_state->def_type;
1726 }
1727
1728 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1729 {
1730         u8 mtrr;
1731
1732         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1733                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1734         if (mtrr == 0xfe || mtrr == 0xff)
1735                 mtrr = MTRR_TYPE_WRBACK;
1736         return mtrr;
1737 }
1738 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1739
1740 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1741 {
1742         unsigned index;
1743         struct hlist_head *bucket;
1744         struct kvm_mmu_page *s;
1745         struct hlist_node *node, *n;
1746
1747         trace_kvm_mmu_unsync_page(sp);
1748         index = kvm_page_table_hashfn(sp->gfn);
1749         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1750         /* don't unsync if pagetable is shadowed with multiple roles */
1751         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1752                 if (s->gfn != sp->gfn || s->role.direct)
1753                         continue;
1754                 if (s->role.word != sp->role.word)
1755                         return 1;
1756         }
1757         ++vcpu->kvm->stat.mmu_unsync;
1758         sp->unsync = 1;
1759
1760         kvm_mmu_mark_parents_unsync(sp);
1761
1762         mmu_convert_notrap(sp);
1763         return 0;
1764 }
1765
1766 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1767                                   bool can_unsync)
1768 {
1769         struct kvm_mmu_page *shadow;
1770
1771         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1772         if (shadow) {
1773                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1774                         return 1;
1775                 if (shadow->unsync)
1776                         return 0;
1777                 if (can_unsync && oos_shadow)
1778                         return kvm_unsync_page(vcpu, shadow);
1779                 return 1;
1780         }
1781         return 0;
1782 }
1783
1784 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1785                     unsigned pte_access, int user_fault,
1786                     int write_fault, int dirty, int level,
1787                     gfn_t gfn, pfn_t pfn, bool speculative,
1788                     bool can_unsync, bool reset_host_protection)
1789 {
1790         u64 spte;
1791         int ret = 0;
1792
1793         /*
1794          * We don't set the accessed bit, since we sometimes want to see
1795          * whether the guest actually used the pte (in order to detect
1796          * demand paging).
1797          */
1798         spte = shadow_base_present_pte | shadow_dirty_mask;
1799         if (!speculative)
1800                 spte |= shadow_accessed_mask;
1801         if (!dirty)
1802                 pte_access &= ~ACC_WRITE_MASK;
1803         if (pte_access & ACC_EXEC_MASK)
1804                 spte |= shadow_x_mask;
1805         else
1806                 spte |= shadow_nx_mask;
1807         if (pte_access & ACC_USER_MASK)
1808                 spte |= shadow_user_mask;
1809         if (level > PT_PAGE_TABLE_LEVEL)
1810                 spte |= PT_PAGE_SIZE_MASK;
1811         if (tdp_enabled)
1812                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1813                         kvm_is_mmio_pfn(pfn));
1814
1815         if (reset_host_protection)
1816                 spte |= SPTE_HOST_WRITEABLE;
1817
1818         spte |= (u64)pfn << PAGE_SHIFT;
1819
1820         if ((pte_access & ACC_WRITE_MASK)
1821             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1822
1823                 if (level > PT_PAGE_TABLE_LEVEL &&
1824                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1825                         ret = 1;
1826                         spte = shadow_trap_nonpresent_pte;
1827                         goto set_pte;
1828                 }
1829
1830                 spte |= PT_WRITABLE_MASK;
1831
1832                 /*
1833                  * Optimization: for pte sync, if spte was writable the hash
1834                  * lookup is unnecessary (and expensive). Write protection
1835                  * is responsibility of mmu_get_page / kvm_sync_page.
1836                  * Same reasoning can be applied to dirty page accounting.
1837                  */
1838                 if (!can_unsync && is_writable_pte(*sptep))
1839                         goto set_pte;
1840
1841                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1842                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1843                                  __func__, gfn);
1844                         ret = 1;
1845                         pte_access &= ~ACC_WRITE_MASK;
1846                         if (is_writable_pte(spte))
1847                                 spte &= ~PT_WRITABLE_MASK;
1848                 }
1849         }
1850
1851         if (pte_access & ACC_WRITE_MASK)
1852                 mark_page_dirty(vcpu->kvm, gfn);
1853
1854 set_pte:
1855         __set_spte(sptep, spte);
1856         return ret;
1857 }
1858
1859 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1860                          unsigned pt_access, unsigned pte_access,
1861                          int user_fault, int write_fault, int dirty,
1862                          int *ptwrite, int level, gfn_t gfn,
1863                          pfn_t pfn, bool speculative,
1864                          bool reset_host_protection)
1865 {
1866         int was_rmapped = 0;
1867         int was_writable = is_writable_pte(*sptep);
1868         int rmap_count;
1869
1870         pgprintk("%s: spte %llx access %x write_fault %d"
1871                  " user_fault %d gfn %lx\n",
1872                  __func__, *sptep, pt_access,
1873                  write_fault, user_fault, gfn);
1874
1875         if (is_rmap_spte(*sptep)) {
1876                 /*
1877                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1878                  * the parent of the now unreachable PTE.
1879                  */
1880                 if (level > PT_PAGE_TABLE_LEVEL &&
1881                     !is_large_pte(*sptep)) {
1882                         struct kvm_mmu_page *child;
1883                         u64 pte = *sptep;
1884
1885                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1886                         mmu_page_remove_parent_pte(child, sptep);
1887                 } else if (pfn != spte_to_pfn(*sptep)) {
1888                         pgprintk("hfn old %lx new %lx\n",
1889                                  spte_to_pfn(*sptep), pfn);
1890                         rmap_remove(vcpu->kvm, sptep);
1891                 } else
1892                         was_rmapped = 1;
1893         }
1894
1895         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1896                       dirty, level, gfn, pfn, speculative, true,
1897                       reset_host_protection)) {
1898                 if (write_fault)
1899                         *ptwrite = 1;
1900                 kvm_x86_ops->tlb_flush(vcpu);
1901         }
1902
1903         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1904         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1905                  is_large_pte(*sptep)? "2MB" : "4kB",
1906                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1907                  *sptep, sptep);
1908         if (!was_rmapped && is_large_pte(*sptep))
1909                 ++vcpu->kvm->stat.lpages;
1910
1911         page_header_update_slot(vcpu->kvm, sptep, gfn);
1912         if (!was_rmapped) {
1913                 rmap_count = rmap_add(vcpu, sptep, gfn);
1914                 kvm_release_pfn_clean(pfn);
1915                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1916                         rmap_recycle(vcpu, sptep, gfn);
1917         } else {
1918                 if (was_writable)
1919                         kvm_release_pfn_dirty(pfn);
1920                 else
1921                         kvm_release_pfn_clean(pfn);
1922         }
1923         if (speculative) {
1924                 vcpu->arch.last_pte_updated = sptep;
1925                 vcpu->arch.last_pte_gfn = gfn;
1926         }
1927 }
1928
1929 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1930 {
1931 }
1932
1933 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1934                         int level, gfn_t gfn, pfn_t pfn)
1935 {
1936         struct kvm_shadow_walk_iterator iterator;
1937         struct kvm_mmu_page *sp;
1938         int pt_write = 0;
1939         gfn_t pseudo_gfn;
1940
1941         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1942                 if (iterator.level == level) {
1943                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1944                                      0, write, 1, &pt_write,
1945                                      level, gfn, pfn, false, true);
1946                         ++vcpu->stat.pf_fixed;
1947                         break;
1948                 }
1949
1950                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1951                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1952                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1953                                               iterator.level - 1,
1954                                               1, ACC_ALL, iterator.sptep);
1955                         if (!sp) {
1956                                 pgprintk("nonpaging_map: ENOMEM\n");
1957                                 kvm_release_pfn_clean(pfn);
1958                                 return -ENOMEM;
1959                         }
1960
1961                         __set_spte(iterator.sptep,
1962                                    __pa(sp->spt)
1963                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1964                                    | shadow_user_mask | shadow_x_mask);
1965                 }
1966         }
1967         return pt_write;
1968 }
1969
1970 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1971 {
1972         int r;
1973         int level;
1974         pfn_t pfn;
1975         unsigned long mmu_seq;
1976
1977         level = mapping_level(vcpu, gfn);
1978
1979         /*
1980          * This path builds a PAE pagetable - so we can map 2mb pages at
1981          * maximum. Therefore check if the level is larger than that.
1982          */
1983         if (level > PT_DIRECTORY_LEVEL)
1984                 level = PT_DIRECTORY_LEVEL;
1985
1986         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
1987
1988         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1989         smp_rmb();
1990         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1991
1992         /* mmio */
1993         if (is_error_pfn(pfn)) {
1994                 kvm_release_pfn_clean(pfn);
1995                 return 1;
1996         }
1997
1998         spin_lock(&vcpu->kvm->mmu_lock);
1999         if (mmu_notifier_retry(vcpu, mmu_seq))
2000                 goto out_unlock;
2001         kvm_mmu_free_some_pages(vcpu);
2002         r = __direct_map(vcpu, v, write, level, gfn, pfn);
2003         spin_unlock(&vcpu->kvm->mmu_lock);
2004
2005
2006         return r;
2007
2008 out_unlock:
2009         spin_unlock(&vcpu->kvm->mmu_lock);
2010         kvm_release_pfn_clean(pfn);
2011         return 0;
2012 }
2013
2014
2015 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2016 {
2017         int i;
2018         struct kvm_mmu_page *sp;
2019
2020         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2021                 return;
2022         spin_lock(&vcpu->kvm->mmu_lock);
2023         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2024                 hpa_t root = vcpu->arch.mmu.root_hpa;
2025
2026                 sp = page_header(root);
2027                 --sp->root_count;
2028                 if (!sp->root_count && sp->role.invalid)
2029                         kvm_mmu_zap_page(vcpu->kvm, sp);
2030                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2031                 spin_unlock(&vcpu->kvm->mmu_lock);
2032                 return;
2033         }
2034         for (i = 0; i < 4; ++i) {
2035                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2036
2037                 if (root) {
2038                         root &= PT64_BASE_ADDR_MASK;
2039                         sp = page_header(root);
2040                         --sp->root_count;
2041                         if (!sp->root_count && sp->role.invalid)
2042                                 kvm_mmu_zap_page(vcpu->kvm, sp);
2043                 }
2044                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2045         }
2046         spin_unlock(&vcpu->kvm->mmu_lock);
2047         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2048 }
2049
2050 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2051 {
2052         int ret = 0;
2053
2054         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2055                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2056                 ret = 1;
2057         }
2058
2059         return ret;
2060 }
2061
2062 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2063 {
2064         int i;
2065         gfn_t root_gfn;
2066         struct kvm_mmu_page *sp;
2067         int direct = 0;
2068         u64 pdptr;
2069
2070         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2071
2072         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2073                 hpa_t root = vcpu->arch.mmu.root_hpa;
2074
2075                 ASSERT(!VALID_PAGE(root));
2076                 if (tdp_enabled)
2077                         direct = 1;
2078                 if (mmu_check_root(vcpu, root_gfn))
2079                         return 1;
2080                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2081                                       PT64_ROOT_LEVEL, direct,
2082                                       ACC_ALL, NULL);
2083                 root = __pa(sp->spt);
2084                 ++sp->root_count;
2085                 vcpu->arch.mmu.root_hpa = root;
2086                 return 0;
2087         }
2088         direct = !is_paging(vcpu);
2089         if (tdp_enabled)
2090                 direct = 1;
2091         for (i = 0; i < 4; ++i) {
2092                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2093
2094                 ASSERT(!VALID_PAGE(root));
2095                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2096                         pdptr = kvm_pdptr_read(vcpu, i);
2097                         if (!is_present_gpte(pdptr)) {
2098                                 vcpu->arch.mmu.pae_root[i] = 0;
2099                                 continue;
2100                         }
2101                         root_gfn = pdptr >> PAGE_SHIFT;
2102                 } else if (vcpu->arch.mmu.root_level == 0)
2103                         root_gfn = 0;
2104                 if (mmu_check_root(vcpu, root_gfn))
2105                         return 1;
2106                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2107                                       PT32_ROOT_LEVEL, direct,
2108                                       ACC_ALL, NULL);
2109                 root = __pa(sp->spt);
2110                 ++sp->root_count;
2111                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2112         }
2113         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2114         return 0;
2115 }
2116
2117 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2118 {
2119         int i;
2120         struct kvm_mmu_page *sp;
2121
2122         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2123                 return;
2124         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2125                 hpa_t root = vcpu->arch.mmu.root_hpa;
2126                 sp = page_header(root);
2127                 mmu_sync_children(vcpu, sp);
2128                 return;
2129         }
2130         for (i = 0; i < 4; ++i) {
2131                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2132
2133                 if (root && VALID_PAGE(root)) {
2134                         root &= PT64_BASE_ADDR_MASK;
2135                         sp = page_header(root);
2136                         mmu_sync_children(vcpu, sp);
2137                 }
2138         }
2139 }
2140
2141 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2142 {
2143         spin_lock(&vcpu->kvm->mmu_lock);
2144         mmu_sync_roots(vcpu);
2145         spin_unlock(&vcpu->kvm->mmu_lock);
2146 }
2147
2148 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2149                                   u32 access, u32 *error)
2150 {
2151         if (error)
2152                 *error = 0;
2153         return vaddr;
2154 }
2155
2156 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2157                                 u32 error_code)
2158 {
2159         gfn_t gfn;
2160         int r;
2161
2162         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2163         r = mmu_topup_memory_caches(vcpu);
2164         if (r)
2165                 return r;
2166
2167         ASSERT(vcpu);
2168         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2169
2170         gfn = gva >> PAGE_SHIFT;
2171
2172         return nonpaging_map(vcpu, gva & PAGE_MASK,
2173                              error_code & PFERR_WRITE_MASK, gfn);
2174 }
2175
2176 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2177                                 u32 error_code)
2178 {
2179         pfn_t pfn;
2180         int r;
2181         int level;
2182         gfn_t gfn = gpa >> PAGE_SHIFT;
2183         unsigned long mmu_seq;
2184
2185         ASSERT(vcpu);
2186         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2187
2188         r = mmu_topup_memory_caches(vcpu);
2189         if (r)
2190                 return r;
2191
2192         level = mapping_level(vcpu, gfn);
2193
2194         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2195
2196         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2197         smp_rmb();
2198         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2199         if (is_error_pfn(pfn)) {
2200                 kvm_release_pfn_clean(pfn);
2201                 return 1;
2202         }
2203         spin_lock(&vcpu->kvm->mmu_lock);
2204         if (mmu_notifier_retry(vcpu, mmu_seq))
2205                 goto out_unlock;
2206         kvm_mmu_free_some_pages(vcpu);
2207         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2208                          level, gfn, pfn);
2209         spin_unlock(&vcpu->kvm->mmu_lock);
2210
2211         return r;
2212
2213 out_unlock:
2214         spin_unlock(&vcpu->kvm->mmu_lock);
2215         kvm_release_pfn_clean(pfn);
2216         return 0;
2217 }
2218
2219 static void nonpaging_free(struct kvm_vcpu *vcpu)
2220 {
2221         mmu_free_roots(vcpu);
2222 }
2223
2224 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2225 {
2226         struct kvm_mmu *context = &vcpu->arch.mmu;
2227
2228         context->new_cr3 = nonpaging_new_cr3;
2229         context->page_fault = nonpaging_page_fault;
2230         context->gva_to_gpa = nonpaging_gva_to_gpa;
2231         context->free = nonpaging_free;
2232         context->prefetch_page = nonpaging_prefetch_page;
2233         context->sync_page = nonpaging_sync_page;
2234         context->invlpg = nonpaging_invlpg;
2235         context->root_level = 0;
2236         context->shadow_root_level = PT32E_ROOT_LEVEL;
2237         context->root_hpa = INVALID_PAGE;
2238         return 0;
2239 }
2240
2241 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2242 {
2243         ++vcpu->stat.tlb_flush;
2244         kvm_x86_ops->tlb_flush(vcpu);
2245 }
2246
2247 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2248 {
2249         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2250         mmu_free_roots(vcpu);
2251 }
2252
2253 static void inject_page_fault(struct kvm_vcpu *vcpu,
2254                               u64 addr,
2255                               u32 err_code)
2256 {
2257         kvm_inject_page_fault(vcpu, addr, err_code);
2258 }
2259
2260 static void paging_free(struct kvm_vcpu *vcpu)
2261 {
2262         nonpaging_free(vcpu);
2263 }
2264
2265 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2266 {
2267         int bit7;
2268
2269         bit7 = (gpte >> 7) & 1;
2270         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2271 }
2272
2273 #define PTTYPE 64
2274 #include "paging_tmpl.h"
2275 #undef PTTYPE
2276
2277 #define PTTYPE 32
2278 #include "paging_tmpl.h"
2279 #undef PTTYPE
2280
2281 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2282 {
2283         struct kvm_mmu *context = &vcpu->arch.mmu;
2284         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2285         u64 exb_bit_rsvd = 0;
2286
2287         if (!is_nx(vcpu))
2288                 exb_bit_rsvd = rsvd_bits(63, 63);
2289         switch (level) {
2290         case PT32_ROOT_LEVEL:
2291                 /* no rsvd bits for 2 level 4K page table entries */
2292                 context->rsvd_bits_mask[0][1] = 0;
2293                 context->rsvd_bits_mask[0][0] = 0;
2294                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2295
2296                 if (!is_pse(vcpu)) {
2297                         context->rsvd_bits_mask[1][1] = 0;
2298                         break;
2299                 }
2300
2301                 if (is_cpuid_PSE36())
2302                         /* 36bits PSE 4MB page */
2303                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2304                 else
2305                         /* 32 bits PSE 4MB page */
2306                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2307                 break;
2308         case PT32E_ROOT_LEVEL:
2309                 context->rsvd_bits_mask[0][2] =
2310                         rsvd_bits(maxphyaddr, 63) |
2311                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2312                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2313                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2314                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2315                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2316                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2317                         rsvd_bits(maxphyaddr, 62) |
2318                         rsvd_bits(13, 20);              /* large page */
2319                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2320                 break;
2321         case PT64_ROOT_LEVEL:
2322                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2323                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2324                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2325                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2326                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2327                         rsvd_bits(maxphyaddr, 51);
2328                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2329                         rsvd_bits(maxphyaddr, 51);
2330                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2331                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2332                         rsvd_bits(maxphyaddr, 51) |
2333                         rsvd_bits(13, 29);
2334                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2335                         rsvd_bits(maxphyaddr, 51) |
2336                         rsvd_bits(13, 20);              /* large page */
2337                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2338                 break;
2339         }
2340 }
2341
2342 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2343 {
2344         struct kvm_mmu *context = &vcpu->arch.mmu;
2345
2346         ASSERT(is_pae(vcpu));
2347         context->new_cr3 = paging_new_cr3;
2348         context->page_fault = paging64_page_fault;
2349         context->gva_to_gpa = paging64_gva_to_gpa;
2350         context->prefetch_page = paging64_prefetch_page;
2351         context->sync_page = paging64_sync_page;
2352         context->invlpg = paging64_invlpg;
2353         context->free = paging_free;
2354         context->root_level = level;
2355         context->shadow_root_level = level;
2356         context->root_hpa = INVALID_PAGE;
2357         return 0;
2358 }
2359
2360 static int paging64_init_context(struct kvm_vcpu *vcpu)
2361 {
2362         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2363         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2364 }
2365
2366 static int paging32_init_context(struct kvm_vcpu *vcpu)
2367 {
2368         struct kvm_mmu *context = &vcpu->arch.mmu;
2369
2370         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2371         context->new_cr3 = paging_new_cr3;
2372         context->page_fault = paging32_page_fault;
2373         context->gva_to_gpa = paging32_gva_to_gpa;
2374         context->free = paging_free;
2375         context->prefetch_page = paging32_prefetch_page;
2376         context->sync_page = paging32_sync_page;
2377         context->invlpg = paging32_invlpg;
2378         context->root_level = PT32_ROOT_LEVEL;
2379         context->shadow_root_level = PT32E_ROOT_LEVEL;
2380         context->root_hpa = INVALID_PAGE;
2381         return 0;
2382 }
2383
2384 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2385 {
2386         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2387         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2388 }
2389
2390 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2391 {
2392         struct kvm_mmu *context = &vcpu->arch.mmu;
2393
2394         context->new_cr3 = nonpaging_new_cr3;
2395         context->page_fault = tdp_page_fault;
2396         context->free = nonpaging_free;
2397         context->prefetch_page = nonpaging_prefetch_page;
2398         context->sync_page = nonpaging_sync_page;
2399         context->invlpg = nonpaging_invlpg;
2400         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2401         context->root_hpa = INVALID_PAGE;
2402
2403         if (!is_paging(vcpu)) {
2404                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2405                 context->root_level = 0;
2406         } else if (is_long_mode(vcpu)) {
2407                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2408                 context->gva_to_gpa = paging64_gva_to_gpa;
2409                 context->root_level = PT64_ROOT_LEVEL;
2410         } else if (is_pae(vcpu)) {
2411                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2412                 context->gva_to_gpa = paging64_gva_to_gpa;
2413                 context->root_level = PT32E_ROOT_LEVEL;
2414         } else {
2415                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2416                 context->gva_to_gpa = paging32_gva_to_gpa;
2417                 context->root_level = PT32_ROOT_LEVEL;
2418         }
2419
2420         return 0;
2421 }
2422
2423 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2424 {
2425         int r;
2426
2427         ASSERT(vcpu);
2428         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2429
2430         if (!is_paging(vcpu))
2431                 r = nonpaging_init_context(vcpu);
2432         else if (is_long_mode(vcpu))
2433                 r = paging64_init_context(vcpu);
2434         else if (is_pae(vcpu))
2435                 r = paging32E_init_context(vcpu);
2436         else
2437                 r = paging32_init_context(vcpu);
2438
2439         vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
2440
2441         return r;
2442 }
2443
2444 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2445 {
2446         vcpu->arch.update_pte.pfn = bad_pfn;
2447
2448         if (tdp_enabled)
2449                 return init_kvm_tdp_mmu(vcpu);
2450         else
2451                 return init_kvm_softmmu(vcpu);
2452 }
2453
2454 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2455 {
2456         ASSERT(vcpu);
2457         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2458                 vcpu->arch.mmu.free(vcpu);
2459                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2460         }
2461 }
2462
2463 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2464 {
2465         destroy_kvm_mmu(vcpu);
2466         return init_kvm_mmu(vcpu);
2467 }
2468 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2469
2470 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2471 {
2472         int r;
2473
2474         r = mmu_topup_memory_caches(vcpu);
2475         if (r)
2476                 goto out;
2477         spin_lock(&vcpu->kvm->mmu_lock);
2478         kvm_mmu_free_some_pages(vcpu);
2479         r = mmu_alloc_roots(vcpu);
2480         mmu_sync_roots(vcpu);
2481         spin_unlock(&vcpu->kvm->mmu_lock);
2482         if (r)
2483                 goto out;
2484         /* set_cr3() should ensure TLB has been flushed */
2485         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2486 out:
2487         return r;
2488 }
2489 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2490
2491 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2492 {
2493         mmu_free_roots(vcpu);
2494 }
2495
2496 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2497                                   struct kvm_mmu_page *sp,
2498                                   u64 *spte)
2499 {
2500         u64 pte;
2501         struct kvm_mmu_page *child;
2502
2503         pte = *spte;
2504         if (is_shadow_present_pte(pte)) {
2505                 if (is_last_spte(pte, sp->role.level))
2506                         rmap_remove(vcpu->kvm, spte);
2507                 else {
2508                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2509                         mmu_page_remove_parent_pte(child, spte);
2510                 }
2511         }
2512         __set_spte(spte, shadow_trap_nonpresent_pte);
2513         if (is_large_pte(pte))
2514                 --vcpu->kvm->stat.lpages;
2515 }
2516
2517 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2518                                   struct kvm_mmu_page *sp,
2519                                   u64 *spte,
2520                                   const void *new)
2521 {
2522         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2523                 ++vcpu->kvm->stat.mmu_pde_zapped;
2524                 return;
2525         }
2526
2527         ++vcpu->kvm->stat.mmu_pte_updated;
2528         if (!sp->role.cr4_pae)
2529                 paging32_update_pte(vcpu, sp, spte, new);
2530         else
2531                 paging64_update_pte(vcpu, sp, spte, new);
2532 }
2533
2534 static bool need_remote_flush(u64 old, u64 new)
2535 {
2536         if (!is_shadow_present_pte(old))
2537                 return false;
2538         if (!is_shadow_present_pte(new))
2539                 return true;
2540         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2541                 return true;
2542         old ^= PT64_NX_MASK;
2543         new ^= PT64_NX_MASK;
2544         return (old & ~new & PT64_PERM_MASK) != 0;
2545 }
2546
2547 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2548 {
2549         if (need_remote_flush(old, new))
2550                 kvm_flush_remote_tlbs(vcpu->kvm);
2551         else
2552                 kvm_mmu_flush_tlb(vcpu);
2553 }
2554
2555 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2556 {
2557         u64 *spte = vcpu->arch.last_pte_updated;
2558
2559         return !!(spte && (*spte & shadow_accessed_mask));
2560 }
2561
2562 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2563                                           u64 gpte)
2564 {
2565         gfn_t gfn;
2566         pfn_t pfn;
2567
2568         if (!is_present_gpte(gpte))
2569                 return;
2570         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2571
2572         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2573         smp_rmb();
2574         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2575
2576         if (is_error_pfn(pfn)) {
2577                 kvm_release_pfn_clean(pfn);
2578                 return;
2579         }
2580         vcpu->arch.update_pte.gfn = gfn;
2581         vcpu->arch.update_pte.pfn = pfn;
2582 }
2583
2584 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2585 {
2586         u64 *spte = vcpu->arch.last_pte_updated;
2587
2588         if (spte
2589             && vcpu->arch.last_pte_gfn == gfn
2590             && shadow_accessed_mask
2591             && !(*spte & shadow_accessed_mask)
2592             && is_shadow_present_pte(*spte))
2593                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2594 }
2595
2596 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2597                        const u8 *new, int bytes,
2598                        bool guest_initiated)
2599 {
2600         gfn_t gfn = gpa >> PAGE_SHIFT;
2601         struct kvm_mmu_page *sp;
2602         struct hlist_node *node, *n;
2603         struct hlist_head *bucket;
2604         unsigned index;
2605         u64 entry, gentry;
2606         u64 *spte;
2607         unsigned offset = offset_in_page(gpa);
2608         unsigned pte_size;
2609         unsigned page_offset;
2610         unsigned misaligned;
2611         unsigned quadrant;
2612         int level;
2613         int flooded = 0;
2614         int npte;
2615         int r;
2616         int invlpg_counter;
2617
2618         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2619
2620         invlpg_counter = atomic_read(&vcpu->kvm->arch.invlpg_counter);
2621
2622         /*
2623          * Assume that the pte write on a page table of the same type
2624          * as the current vcpu paging mode.  This is nearly always true
2625          * (might be false while changing modes).  Note it is verified later
2626          * by update_pte().
2627          */
2628         if ((is_pae(vcpu) && bytes == 4) || !new) {
2629                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2630                 if (is_pae(vcpu)) {
2631                         gpa &= ~(gpa_t)7;
2632                         bytes = 8;
2633                 }
2634                 r = kvm_read_guest(vcpu->kvm, gpa, &gentry, min(bytes, 8));
2635                 if (r)
2636                         gentry = 0;
2637                 new = (const u8 *)&gentry;
2638         }
2639
2640         switch (bytes) {
2641         case 4:
2642                 gentry = *(const u32 *)new;
2643                 break;
2644         case 8:
2645                 gentry = *(const u64 *)new;
2646                 break;
2647         default:
2648                 gentry = 0;
2649                 break;
2650         }
2651
2652         mmu_guess_page_from_pte_write(vcpu, gpa, gentry);
2653         spin_lock(&vcpu->kvm->mmu_lock);
2654         if (atomic_read(&vcpu->kvm->arch.invlpg_counter) != invlpg_counter)
2655                 gentry = 0;
2656         kvm_mmu_access_page(vcpu, gfn);
2657         kvm_mmu_free_some_pages(vcpu);
2658         ++vcpu->kvm->stat.mmu_pte_write;
2659         kvm_mmu_audit(vcpu, "pre pte write");
2660         if (guest_initiated) {
2661                 if (gfn == vcpu->arch.last_pt_write_gfn
2662                     && !last_updated_pte_accessed(vcpu)) {
2663                         ++vcpu->arch.last_pt_write_count;
2664                         if (vcpu->arch.last_pt_write_count >= 3)
2665                                 flooded = 1;
2666                 } else {
2667                         vcpu->arch.last_pt_write_gfn = gfn;
2668                         vcpu->arch.last_pt_write_count = 1;
2669                         vcpu->arch.last_pte_updated = NULL;
2670                 }
2671         }
2672         index = kvm_page_table_hashfn(gfn);
2673         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2674         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2675                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2676                         continue;
2677                 pte_size = sp->role.cr4_pae ? 8 : 4;
2678                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2679                 misaligned |= bytes < 4;
2680                 if (misaligned || flooded) {
2681                         /*
2682                          * Misaligned accesses are too much trouble to fix
2683                          * up; also, they usually indicate a page is not used
2684                          * as a page table.
2685                          *
2686                          * If we're seeing too many writes to a page,
2687                          * it may no longer be a page table, or we may be
2688                          * forking, in which case it is better to unmap the
2689                          * page.
2690                          */
2691                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2692                                  gpa, bytes, sp->role.word);
2693                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2694                                 n = bucket->first;
2695                         ++vcpu->kvm->stat.mmu_flooded;
2696                         continue;
2697                 }
2698                 page_offset = offset;
2699                 level = sp->role.level;
2700                 npte = 1;
2701                 if (!sp->role.cr4_pae) {
2702                         page_offset <<= 1;      /* 32->64 */
2703                         /*
2704                          * A 32-bit pde maps 4MB while the shadow pdes map
2705                          * only 2MB.  So we need to double the offset again
2706                          * and zap two pdes instead of one.
2707                          */
2708                         if (level == PT32_ROOT_LEVEL) {
2709                                 page_offset &= ~7; /* kill rounding error */
2710                                 page_offset <<= 1;
2711                                 npte = 2;
2712                         }
2713                         quadrant = page_offset >> PAGE_SHIFT;
2714                         page_offset &= ~PAGE_MASK;
2715                         if (quadrant != sp->role.quadrant)
2716                                 continue;
2717                 }
2718                 spte = &sp->spt[page_offset / sizeof(*spte)];
2719                 while (npte--) {
2720                         entry = *spte;
2721                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2722                         if (gentry)
2723                                 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
2724                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2725                         ++spte;
2726                 }
2727         }
2728         kvm_mmu_audit(vcpu, "post pte write");
2729         spin_unlock(&vcpu->kvm->mmu_lock);
2730         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2731                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2732                 vcpu->arch.update_pte.pfn = bad_pfn;
2733         }
2734 }
2735
2736 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2737 {
2738         gpa_t gpa;
2739         int r;
2740
2741         if (tdp_enabled)
2742                 return 0;
2743
2744         gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2745
2746         spin_lock(&vcpu->kvm->mmu_lock);
2747         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2748         spin_unlock(&vcpu->kvm->mmu_lock);
2749         return r;
2750 }
2751 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2752
2753 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2754 {
2755         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2756                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2757                 struct kvm_mmu_page *sp;
2758
2759                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2760                                   struct kvm_mmu_page, link);
2761                 kvm_mmu_zap_page(vcpu->kvm, sp);
2762                 ++vcpu->kvm->stat.mmu_recycled;
2763         }
2764 }
2765
2766 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2767 {
2768         int r;
2769         enum emulation_result er;
2770
2771         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2772         if (r < 0)
2773                 goto out;
2774
2775         if (!r) {
2776                 r = 1;
2777                 goto out;
2778         }
2779
2780         r = mmu_topup_memory_caches(vcpu);
2781         if (r)
2782                 goto out;
2783
2784         er = emulate_instruction(vcpu, cr2, error_code, 0);
2785
2786         switch (er) {
2787         case EMULATE_DONE:
2788                 return 1;
2789         case EMULATE_DO_MMIO:
2790                 ++vcpu->stat.mmio_exits;
2791                 return 0;
2792         case EMULATE_FAIL:
2793                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2794                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2795                 vcpu->run->internal.ndata = 0;
2796                 return 0;
2797         default:
2798                 BUG();
2799         }
2800 out:
2801         return r;
2802 }
2803 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2804
2805 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2806 {
2807         vcpu->arch.mmu.invlpg(vcpu, gva);
2808         kvm_mmu_flush_tlb(vcpu);
2809         ++vcpu->stat.invlpg;
2810 }
2811 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2812
2813 void kvm_enable_tdp(void)
2814 {
2815         tdp_enabled = true;
2816 }
2817 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2818
2819 void kvm_disable_tdp(void)
2820 {
2821         tdp_enabled = false;
2822 }
2823 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2824
2825 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2826 {
2827         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2828 }
2829
2830 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2831 {
2832         struct page *page;
2833         int i;
2834
2835         ASSERT(vcpu);
2836
2837         /*
2838          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2839          * Therefore we need to allocate shadow page tables in the first
2840          * 4GB of memory, which happens to fit the DMA32 zone.
2841          */
2842         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2843         if (!page)
2844                 return -ENOMEM;
2845
2846         vcpu->arch.mmu.pae_root = page_address(page);
2847         for (i = 0; i < 4; ++i)
2848                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2849
2850         return 0;
2851 }
2852
2853 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2854 {
2855         ASSERT(vcpu);
2856         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2857
2858         return alloc_mmu_pages(vcpu);
2859 }
2860
2861 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2862 {
2863         ASSERT(vcpu);
2864         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2865
2866         return init_kvm_mmu(vcpu);
2867 }
2868
2869 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2870 {
2871         ASSERT(vcpu);
2872
2873         destroy_kvm_mmu(vcpu);
2874         free_mmu_pages(vcpu);
2875         mmu_free_memory_caches(vcpu);
2876 }
2877
2878 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2879 {
2880         struct kvm_mmu_page *sp;
2881
2882         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2883                 int i;
2884                 u64 *pt;
2885
2886                 if (!test_bit(slot, sp->slot_bitmap))
2887                         continue;
2888
2889                 pt = sp->spt;
2890                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2891                         /* avoid RMW */
2892                         if (pt[i] & PT_WRITABLE_MASK)
2893                                 pt[i] &= ~PT_WRITABLE_MASK;
2894         }
2895         kvm_flush_remote_tlbs(kvm);
2896 }
2897
2898 void kvm_mmu_zap_all(struct kvm *kvm)
2899 {
2900         struct kvm_mmu_page *sp, *node;
2901
2902         spin_lock(&kvm->mmu_lock);
2903         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2904                 if (kvm_mmu_zap_page(kvm, sp))
2905                         node = container_of(kvm->arch.active_mmu_pages.next,
2906                                             struct kvm_mmu_page, link);
2907         spin_unlock(&kvm->mmu_lock);
2908
2909         kvm_flush_remote_tlbs(kvm);
2910 }
2911
2912 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2913 {
2914         struct kvm_mmu_page *page;
2915
2916         page = container_of(kvm->arch.active_mmu_pages.prev,
2917                             struct kvm_mmu_page, link);
2918         kvm_mmu_zap_page(kvm, page);
2919 }
2920
2921 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2922 {
2923         struct kvm *kvm;
2924         struct kvm *kvm_freed = NULL;
2925         int cache_count = 0;
2926
2927         spin_lock(&kvm_lock);
2928
2929         list_for_each_entry(kvm, &vm_list, vm_list) {
2930                 int npages, idx;
2931
2932                 idx = srcu_read_lock(&kvm->srcu);
2933                 spin_lock(&kvm->mmu_lock);
2934                 npages = kvm->arch.n_alloc_mmu_pages -
2935                          kvm->arch.n_free_mmu_pages;
2936                 cache_count += npages;
2937                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2938                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2939                         cache_count--;
2940                         kvm_freed = kvm;
2941                 }
2942                 nr_to_scan--;
2943
2944                 spin_unlock(&kvm->mmu_lock);
2945                 srcu_read_unlock(&kvm->srcu, idx);
2946         }
2947         if (kvm_freed)
2948                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2949
2950         spin_unlock(&kvm_lock);
2951
2952         return cache_count;
2953 }
2954
2955 static struct shrinker mmu_shrinker = {
2956         .shrink = mmu_shrink,
2957         .seeks = DEFAULT_SEEKS * 10,
2958 };
2959
2960 static void mmu_destroy_caches(void)
2961 {
2962         if (pte_chain_cache)
2963                 kmem_cache_destroy(pte_chain_cache);
2964         if (rmap_desc_cache)
2965                 kmem_cache_destroy(rmap_desc_cache);
2966         if (mmu_page_header_cache)
2967                 kmem_cache_destroy(mmu_page_header_cache);
2968 }
2969
2970 void kvm_mmu_module_exit(void)
2971 {
2972         mmu_destroy_caches();
2973         unregister_shrinker(&mmu_shrinker);
2974 }
2975
2976 int kvm_mmu_module_init(void)
2977 {
2978         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2979                                             sizeof(struct kvm_pte_chain),
2980                                             0, 0, NULL);
2981         if (!pte_chain_cache)
2982                 goto nomem;
2983         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2984                                             sizeof(struct kvm_rmap_desc),
2985                                             0, 0, NULL);
2986         if (!rmap_desc_cache)
2987                 goto nomem;
2988
2989         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2990                                                   sizeof(struct kvm_mmu_page),
2991                                                   0, 0, NULL);
2992         if (!mmu_page_header_cache)
2993                 goto nomem;
2994
2995         register_shrinker(&mmu_shrinker);
2996
2997         return 0;
2998
2999 nomem:
3000         mmu_destroy_caches();
3001         return -ENOMEM;
3002 }
3003
3004 /*
3005  * Caculate mmu pages needed for kvm.
3006  */
3007 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3008 {
3009         int i;
3010         unsigned int nr_mmu_pages;
3011         unsigned int  nr_pages = 0;
3012         struct kvm_memslots *slots;
3013
3014         slots = rcu_dereference(kvm->memslots);
3015         for (i = 0; i < slots->nmemslots; i++)
3016                 nr_pages += slots->memslots[i].npages;
3017
3018         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3019         nr_mmu_pages = max(nr_mmu_pages,
3020                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3021
3022         return nr_mmu_pages;
3023 }
3024
3025 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3026                                 unsigned len)
3027 {
3028         if (len > buffer->len)
3029                 return NULL;
3030         return buffer->ptr;
3031 }
3032
3033 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3034                                 unsigned len)
3035 {
3036         void *ret;
3037
3038         ret = pv_mmu_peek_buffer(buffer, len);
3039         if (!ret)
3040                 return ret;
3041         buffer->ptr += len;
3042         buffer->len -= len;
3043         buffer->processed += len;
3044         return ret;
3045 }
3046
3047 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3048                              gpa_t addr, gpa_t value)
3049 {
3050         int bytes = 8;
3051         int r;
3052
3053         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3054                 bytes = 4;
3055
3056         r = mmu_topup_memory_caches(vcpu);
3057         if (r)
3058                 return r;
3059
3060         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3061                 return -EFAULT;
3062
3063         return 1;
3064 }
3065
3066 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3067 {
3068         kvm_set_cr3(vcpu, vcpu->arch.cr3);
3069         return 1;
3070 }
3071
3072 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3073 {
3074         spin_lock(&vcpu->kvm->mmu_lock);
3075         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3076         spin_unlock(&vcpu->kvm->mmu_lock);
3077         return 1;
3078 }
3079
3080 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3081                              struct kvm_pv_mmu_op_buffer *buffer)
3082 {
3083         struct kvm_mmu_op_header *header;
3084
3085         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3086         if (!header)
3087                 return 0;
3088         switch (header->op) {
3089         case KVM_MMU_OP_WRITE_PTE: {
3090                 struct kvm_mmu_op_write_pte *wpte;
3091
3092                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3093                 if (!wpte)
3094                         return 0;
3095                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3096                                         wpte->pte_val);
3097         }
3098         case KVM_MMU_OP_FLUSH_TLB: {
3099                 struct kvm_mmu_op_flush_tlb *ftlb;
3100
3101                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3102                 if (!ftlb)
3103                         return 0;
3104                 return kvm_pv_mmu_flush_tlb(vcpu);
3105         }
3106         case KVM_MMU_OP_RELEASE_PT: {
3107                 struct kvm_mmu_op_release_pt *rpt;
3108
3109                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3110                 if (!rpt)
3111                         return 0;
3112                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3113         }
3114         default: return 0;
3115         }
3116 }
3117
3118 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3119                   gpa_t addr, unsigned long *ret)
3120 {
3121         int r;
3122         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3123
3124         buffer->ptr = buffer->buf;
3125         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3126         buffer->processed = 0;
3127
3128         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3129         if (r)
3130                 goto out;
3131
3132         while (buffer->len) {
3133                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3134                 if (r < 0)
3135                         goto out;
3136                 if (r == 0)
3137                         break;
3138         }
3139
3140         r = 1;
3141 out:
3142         *ret = buffer->processed;
3143         return r;
3144 }
3145
3146 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3147 {
3148         struct kvm_shadow_walk_iterator iterator;
3149         int nr_sptes = 0;
3150
3151         spin_lock(&vcpu->kvm->mmu_lock);
3152         for_each_shadow_entry(vcpu, addr, iterator) {
3153                 sptes[iterator.level-1] = *iterator.sptep;
3154                 nr_sptes++;
3155                 if (!is_shadow_present_pte(*iterator.sptep))
3156                         break;
3157         }
3158         spin_unlock(&vcpu->kvm->mmu_lock);
3159
3160         return nr_sptes;
3161 }
3162 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3163
3164 #ifdef AUDIT
3165
3166 static const char *audit_msg;
3167
3168 static gva_t canonicalize(gva_t gva)
3169 {
3170 #ifdef CONFIG_X86_64
3171         gva = (long long)(gva << 16) >> 16;
3172 #endif
3173         return gva;
3174 }
3175
3176
3177 typedef void (*inspect_spte_fn) (struct kvm *kvm, u64 *sptep);
3178
3179 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3180                             inspect_spte_fn fn)
3181 {
3182         int i;
3183
3184         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3185                 u64 ent = sp->spt[i];
3186
3187                 if (is_shadow_present_pte(ent)) {
3188                         if (!is_last_spte(ent, sp->role.level)) {
3189                                 struct kvm_mmu_page *child;
3190                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3191                                 __mmu_spte_walk(kvm, child, fn);
3192                         } else
3193                                 fn(kvm, &sp->spt[i]);
3194                 }
3195         }
3196 }
3197
3198 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3199 {
3200         int i;
3201         struct kvm_mmu_page *sp;
3202
3203         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3204                 return;
3205         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3206                 hpa_t root = vcpu->arch.mmu.root_hpa;
3207                 sp = page_header(root);
3208                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3209                 return;
3210         }
3211         for (i = 0; i < 4; ++i) {
3212                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3213
3214                 if (root && VALID_PAGE(root)) {
3215                         root &= PT64_BASE_ADDR_MASK;
3216                         sp = page_header(root);
3217                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3218                 }
3219         }
3220         return;
3221 }
3222
3223 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3224                                 gva_t va, int level)
3225 {
3226         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3227         int i;
3228         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3229
3230         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3231                 u64 ent = pt[i];
3232
3233                 if (ent == shadow_trap_nonpresent_pte)
3234                         continue;
3235
3236                 va = canonicalize(va);
3237                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3238                         audit_mappings_page(vcpu, ent, va, level - 1);
3239                 else {
3240                         gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, va, NULL);
3241                         gfn_t gfn = gpa >> PAGE_SHIFT;
3242                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3243                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3244
3245                         if (is_error_pfn(pfn)) {
3246                                 kvm_release_pfn_clean(pfn);
3247                                 continue;
3248                         }
3249
3250                         if (is_shadow_present_pte(ent)
3251                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3252                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3253                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3254                                        audit_msg, vcpu->arch.mmu.root_level,
3255                                        va, gpa, hpa, ent,
3256                                        is_shadow_present_pte(ent));
3257                         else if (ent == shadow_notrap_nonpresent_pte
3258                                  && !is_error_hpa(hpa))
3259                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3260                                        " valid guest gva %lx\n", audit_msg, va);
3261                         kvm_release_pfn_clean(pfn);
3262
3263                 }
3264         }
3265 }
3266
3267 static void audit_mappings(struct kvm_vcpu *vcpu)
3268 {
3269         unsigned i;
3270
3271         if (vcpu->arch.mmu.root_level == 4)
3272                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3273         else
3274                 for (i = 0; i < 4; ++i)
3275                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3276                                 audit_mappings_page(vcpu,
3277                                                     vcpu->arch.mmu.pae_root[i],
3278                                                     i << 30,
3279                                                     2);
3280 }
3281
3282 static int count_rmaps(struct kvm_vcpu *vcpu)
3283 {
3284         struct kvm *kvm = vcpu->kvm;
3285         struct kvm_memslots *slots;
3286         int nmaps = 0;
3287         int i, j, k, idx;
3288
3289         idx = srcu_read_lock(&kvm->srcu);
3290         slots = rcu_dereference(kvm->memslots);
3291         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3292                 struct kvm_memory_slot *m = &slots->memslots[i];
3293                 struct kvm_rmap_desc *d;
3294
3295                 for (j = 0; j < m->npages; ++j) {
3296                         unsigned long *rmapp = &m->rmap[j];
3297
3298                         if (!*rmapp)
3299                                 continue;
3300                         if (!(*rmapp & 1)) {
3301                                 ++nmaps;
3302                                 continue;
3303                         }
3304                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3305                         while (d) {
3306                                 for (k = 0; k < RMAP_EXT; ++k)
3307                                         if (d->sptes[k])
3308                                                 ++nmaps;
3309                                         else
3310                                                 break;
3311                                 d = d->more;
3312                         }
3313                 }
3314         }
3315         srcu_read_unlock(&kvm->srcu, idx);
3316         return nmaps;
3317 }
3318
3319 void inspect_spte_has_rmap(struct kvm *kvm, u64 *sptep)
3320 {
3321         unsigned long *rmapp;
3322         struct kvm_mmu_page *rev_sp;
3323         gfn_t gfn;
3324
3325         if (*sptep & PT_WRITABLE_MASK) {
3326                 rev_sp = page_header(__pa(sptep));
3327                 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3328
3329                 if (!gfn_to_memslot(kvm, gfn)) {
3330                         if (!printk_ratelimit())
3331                                 return;
3332                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3333                                          audit_msg, gfn);
3334                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3335                                audit_msg, (long int)(sptep - rev_sp->spt),
3336                                         rev_sp->gfn);
3337                         dump_stack();
3338                         return;
3339                 }
3340
3341                 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3342                                     rev_sp->role.level);
3343                 if (!*rmapp) {
3344                         if (!printk_ratelimit())
3345                                 return;
3346                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3347                                          audit_msg, *sptep);
3348                         dump_stack();
3349                 }
3350         }
3351
3352 }
3353
3354 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3355 {
3356         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3357 }
3358
3359 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3360 {
3361         struct kvm_mmu_page *sp;
3362         int i;
3363
3364         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3365                 u64 *pt = sp->spt;
3366
3367                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3368                         continue;
3369
3370                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3371                         u64 ent = pt[i];
3372
3373                         if (!(ent & PT_PRESENT_MASK))
3374                                 continue;
3375                         if (!(ent & PT_WRITABLE_MASK))
3376                                 continue;
3377                         inspect_spte_has_rmap(vcpu->kvm, &pt[i]);
3378                 }
3379         }
3380         return;
3381 }
3382
3383 static void audit_rmap(struct kvm_vcpu *vcpu)
3384 {
3385         check_writable_mappings_rmap(vcpu);
3386         count_rmaps(vcpu);
3387 }
3388
3389 static void audit_write_protection(struct kvm_vcpu *vcpu)
3390 {
3391         struct kvm_mmu_page *sp;
3392         struct kvm_memory_slot *slot;
3393         unsigned long *rmapp;
3394         u64 *spte;
3395         gfn_t gfn;
3396
3397         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3398                 if (sp->role.direct)
3399                         continue;
3400                 if (sp->unsync)
3401                         continue;
3402
3403                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3404                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3405                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3406
3407                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3408                 while (spte) {
3409                         if (*spte & PT_WRITABLE_MASK)
3410                                 printk(KERN_ERR "%s: (%s) shadow page has "
3411                                 "writable mappings: gfn %lx role %x\n",
3412                                __func__, audit_msg, sp->gfn,
3413                                sp->role.word);
3414                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3415                 }
3416         }
3417 }
3418
3419 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3420 {
3421         int olddbg = dbg;
3422
3423         dbg = 0;
3424         audit_msg = msg;
3425         audit_rmap(vcpu);
3426         audit_write_protection(vcpu);
3427         if (strcmp("pre pte write", audit_msg) != 0)
3428                 audit_mappings(vcpu);
3429         audit_writable_sptes_have_rmaps(vcpu);
3430         dbg = olddbg;
3431 }
3432
3433 #endif