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