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