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 "kvm_cache_regs.h"
  22
  23 #include <linux/kvm_host.h>
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29 #include <linux/swap.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/compiler.h>
  32
  33 #include <asm/page.h>
  34 #include <asm/cmpxchg.h>
  35 #include <asm/io.h>
  36 #include <asm/vmx.h>
  37
  38 /*
  39  * When setting this variable to true it enables Two-Dimensional-Paging
  40  * where the hardware walks 2 page tables:
  41  * 1. the guest-virtual to guest-physical
  42  * 2. while doing 1. it walks guest-physical to host-physical
  43  * If the hardware supports that we don't need to do shadow paging.
  44  */
  45 bool tdp_enabled = false;
  46
  47 #undef MMU_DEBUG
  48
  49 #undef AUDIT
  50
  51 #ifdef AUDIT
  52 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  53 #else
  54 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  55 #endif
  56
  57 #ifdef MMU_DEBUG
  58
  59 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  60 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  61
  62 #else
  63
  64 #define pgprintk(x...) do { } while (0)
  65 #define rmap_printk(x...) do { } while (0)
  66
  67 #endif
  68
  69 #if defined(MMU_DEBUG) || defined(AUDIT)
  70 static int dbg = 0;
  71 module_param(dbg, bool, 0644);
  72 #endif
  73
  74 static int oos_shadow = 1;
  75 module_param(oos_shadow, bool, 0644);
  76
  77 #ifndef MMU_DEBUG
  78 #define ASSERT(x) do { } while (0)
  79 #else
  80 #define ASSERT(x)                                                       \
  81         if (!(x)) {                                                     \
  82                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  83                        __FILE__, __LINE__, #x);                         \
  84         }
  85 #endif
  86
  87 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  88 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  89
  90 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  91
  92 #define PT64_LEVEL_BITS 9
  93
  94 #define PT64_LEVEL_SHIFT(level) \
  95                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  96
  97 #define PT64_LEVEL_MASK(level) \
  98                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  99
 100 #define PT64_INDEX(address, level)\
 101         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 102
 103
 104 #define PT32_LEVEL_BITS 10
 105
 106 #define PT32_LEVEL_SHIFT(level) \
 107                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 108
 109 #define PT32_LEVEL_MASK(level) \
 110                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 111 #define PT32_LVL_OFFSET_MASK(level) \
 112         (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 113                                                 * PT32_LEVEL_BITS))) - 1))
 114
 115 #define PT32_INDEX(address, level)\
 116         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 117
 118
 119 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 120 #define PT64_DIR_BASE_ADDR_MASK \
 121         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 122 #define PT64_LVL_ADDR_MASK(level) \
 123         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 124                                                 * PT64_LEVEL_BITS))) - 1))
 125 #define PT64_LVL_OFFSET_MASK(level) \
 126         (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 127                                                 * PT64_LEVEL_BITS))) - 1))
 128
 129 #define PT32_BASE_ADDR_MASK PAGE_MASK
 130 #define PT32_DIR_BASE_ADDR_MASK \
 131         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 132 #define PT32_LVL_ADDR_MASK(level) \
 133         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 134                                             * PT32_LEVEL_BITS))) - 1))
 135
 136 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 137                         | PT64_NX_MASK)
 138
 139 #define PFERR_PRESENT_MASK (1U << 0)
 140 #define PFERR_WRITE_MASK (1U << 1)
 141 #define PFERR_USER_MASK (1U << 2)
 142 #define PFERR_RSVD_MASK (1U << 3)
 143 #define PFERR_FETCH_MASK (1U << 4)
 144
 145 #define PT_PDPE_LEVEL 3
 146 #define PT_DIRECTORY_LEVEL 2
 147 #define PT_PAGE_TABLE_LEVEL 1
 148
 149 #define RMAP_EXT 4
 150
 151 #define ACC_EXEC_MASK    1
 152 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 153 #define ACC_USER_MASK    PT_USER_MASK
 154 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 155
 156 #define CREATE_TRACE_POINTS
 157 #include "mmutrace.h"
 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 vcpu->arch.cr0 & 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.shadow_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_writeble_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 = PAGE_SIZE;
 472         struct vm_area_struct *vma;
 473         unsigned long addr;
 474         int i, ret = 0;
 475
 476         addr = gfn_to_hva(kvm, gfn);
 477         if (kvm_is_error_hva(addr))
 478                 return page_size;
 479
 480         down_read(&current->mm->mmap_sem);
 481         vma = find_vma(current->mm, addr);
 482         if (!vma)
 483                 goto out;
 484
 485         page_size = vma_kernel_pagesize(vma);
 486
 487 out:
 488         up_read(&current->mm->mmap_sem);
 489
 490         for (i = PT_PAGE_TABLE_LEVEL;
 491              i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
 492                 if (page_size >= KVM_HPAGE_SIZE(i))
 493                         ret = i;
 494                 else
 495                         break;
 496         }
 497
 498         return ret;
 499 }
 500
 501 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 502 {
 503         struct kvm_memory_slot *slot;
 504         int host_level;
 505         int level = PT_PAGE_TABLE_LEVEL;
 506
 507         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 508         if (slot && slot->dirty_bitmap)
 509                 return PT_PAGE_TABLE_LEVEL;
 510
 511         host_level = host_mapping_level(vcpu->kvm, large_gfn);
 512
 513         if (host_level == PT_PAGE_TABLE_LEVEL)
 514                 return host_level;
 515
 516         for (level = PT_DIRECTORY_LEVEL; level <= host_level; ++level) {
 517
 518                 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
 519                         break;
 520         }
 521
 522         return level - 1;
 523 }
 524
 525 /*
 526  * Take gfn and return the reverse mapping to it.
 527  * Note: gfn must be unaliased before this function get called
 528  */
 529
 530 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
 531 {
 532         struct kvm_memory_slot *slot;
 533         unsigned long idx;
 534
 535         slot = gfn_to_memslot(kvm, gfn);
 536         if (likely(level == PT_PAGE_TABLE_LEVEL))
 537                 return &slot->rmap[gfn - slot->base_gfn];
 538
 539         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 540                 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 541
 542         return &slot->lpage_info[level - 2][idx].rmap_pde;
 543 }
 544
 545 /*
 546  * Reverse mapping data structures:
 547  *
 548  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 549  * that points to page_address(page).
 550  *
 551  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 552  * containing more mappings.
 553  *
 554  * Returns the number of rmap entries before the spte was added or zero if
 555  * the spte was not added.
 556  *
 557  */
 558 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 559 {
 560         struct kvm_mmu_page *sp;
 561         struct kvm_rmap_desc *desc;
 562         unsigned long *rmapp;
 563         int i, count = 0;
 564
 565         if (!is_rmap_spte(*spte))
 566                 return count;
 567         gfn = unalias_gfn(vcpu->kvm, gfn);
 568         sp = page_header(__pa(spte));
 569         sp->gfns[spte - sp->spt] = gfn;
 570         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 571         if (!*rmapp) {
 572                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 573                 *rmapp = (unsigned long)spte;
 574         } else if (!(*rmapp & 1)) {
 575                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 576                 desc = mmu_alloc_rmap_desc(vcpu);
 577                 desc->sptes[0] = (u64 *)*rmapp;
 578                 desc->sptes[1] = spte;
 579                 *rmapp = (unsigned long)desc | 1;
 580         } else {
 581                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 582                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 583                 while (desc->sptes[RMAP_EXT-1] && desc->more) {
 584                         desc = desc->more;
 585                         count += RMAP_EXT;
 586                 }
 587                 if (desc->sptes[RMAP_EXT-1]) {
 588                         desc->more = mmu_alloc_rmap_desc(vcpu);
 589                         desc = desc->more;
 590                 }
 591                 for (i = 0; desc->sptes[i]; ++i)
 592                         ;
 593                 desc->sptes[i] = spte;
 594         }
 595         return count;
 596 }
 597
 598 static void rmap_desc_remove_entry(unsigned long *rmapp,
 599                                    struct kvm_rmap_desc *desc,
 600                                    int i,
 601                                    struct kvm_rmap_desc *prev_desc)
 602 {
 603         int j;
 604
 605         for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
 606                 ;
 607         desc->sptes[i] = desc->sptes[j];
 608         desc->sptes[j] = NULL;
 609         if (j != 0)
 610                 return;
 611         if (!prev_desc && !desc->more)
 612                 *rmapp = (unsigned long)desc->sptes[0];
 613         else
 614                 if (prev_desc)
 615                         prev_desc->more = desc->more;
 616                 else
 617                         *rmapp = (unsigned long)desc->more | 1;
 618         mmu_free_rmap_desc(desc);
 619 }
 620
 621 static void rmap_remove(struct kvm *kvm, u64 *spte)
 622 {
 623         struct kvm_rmap_desc *desc;
 624         struct kvm_rmap_desc *prev_desc;
 625         struct kvm_mmu_page *sp;
 626         pfn_t pfn;
 627         unsigned long *rmapp;
 628         int i;
 629
 630         if (!is_rmap_spte(*spte))
 631                 return;
 632         sp = page_header(__pa(spte));
 633         pfn = spte_to_pfn(*spte);
 634         if (*spte & shadow_accessed_mask)
 635                 kvm_set_pfn_accessed(pfn);
 636         if (is_writeble_pte(*spte))
 637                 kvm_set_pfn_dirty(pfn);
 638         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], sp->role.level);
 639         if (!*rmapp) {
 640                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 641                 BUG();
 642         } else if (!(*rmapp & 1)) {
 643                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 644                 if ((u64 *)*rmapp != spte) {
 645                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 646                                spte, *spte);
 647                         BUG();
 648                 }
 649                 *rmapp = 0;
 650         } else {
 651                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 652                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 653                 prev_desc = NULL;
 654                 while (desc) {
 655                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
 656                                 if (desc->sptes[i] == spte) {
 657                                         rmap_desc_remove_entry(rmapp,
 658                                                                desc, i,
 659                                                                prev_desc);
 660                                         return;
 661                                 }
 662                         prev_desc = desc;
 663                         desc = desc->more;
 664                 }
 665                 BUG();
 666         }
 667 }
 668
 669 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 670 {
 671         struct kvm_rmap_desc *desc;
 672         struct kvm_rmap_desc *prev_desc;
 673         u64 *prev_spte;
 674         int i;
 675
 676         if (!*rmapp)
 677                 return NULL;
 678         else if (!(*rmapp & 1)) {
 679                 if (!spte)
 680                         return (u64 *)*rmapp;
 681                 return NULL;
 682         }
 683         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 684         prev_desc = NULL;
 685         prev_spte = NULL;
 686         while (desc) {
 687                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
 688                         if (prev_spte == spte)
 689                                 return desc->sptes[i];
 690                         prev_spte = desc->sptes[i];
 691                 }
 692                 desc = desc->more;
 693         }
 694         return NULL;
 695 }
 696
 697 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 698 {
 699         unsigned long *rmapp;
 700         u64 *spte;
 701         int i, write_protected = 0;
 702
 703         gfn = unalias_gfn(kvm, gfn);
 704         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
 705
 706         spte = rmap_next(kvm, rmapp, NULL);
 707         while (spte) {
 708                 BUG_ON(!spte);
 709                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 710                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 711                 if (is_writeble_pte(*spte)) {
 712                         __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
 713                         write_protected = 1;
 714                 }
 715                 spte = rmap_next(kvm, rmapp, spte);
 716         }
 717         if (write_protected) {
 718                 pfn_t pfn;
 719
 720                 spte = rmap_next(kvm, rmapp, NULL);
 721                 pfn = spte_to_pfn(*spte);
 722                 kvm_set_pfn_dirty(pfn);
 723         }
 724
 725         /* check for huge page mappings */
 726         for (i = PT_DIRECTORY_LEVEL;
 727              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 728                 rmapp = gfn_to_rmap(kvm, gfn, i);
 729                 spte = rmap_next(kvm, rmapp, NULL);
 730                 while (spte) {
 731                         BUG_ON(!spte);
 732                         BUG_ON(!(*spte & PT_PRESENT_MASK));
 733                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 734                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 735                         if (is_writeble_pte(*spte)) {
 736                                 rmap_remove(kvm, spte);
 737                                 --kvm->stat.lpages;
 738                                 __set_spte(spte, shadow_trap_nonpresent_pte);
 739                                 spte = NULL;
 740                                 write_protected = 1;
 741                         }
 742                         spte = rmap_next(kvm, rmapp, spte);
 743                 }
 744         }
 745
 746         return write_protected;
 747 }
 748
 749 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 750 {
 751         u64 *spte;
 752         int need_tlb_flush = 0;
 753
 754         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 755                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 756                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 757                 rmap_remove(kvm, spte);
 758                 __set_spte(spte, shadow_trap_nonpresent_pte);
 759                 need_tlb_flush = 1;
 760         }
 761         return need_tlb_flush;
 762 }
 763
 764 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 765                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 766 {
 767         int i, j;
 768         int retval = 0;
 769
 770         /*
 771          * If mmap_sem isn't taken, we can look the memslots with only
 772          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 773          */
 774         for (i = 0; i < kvm->nmemslots; i++) {
 775                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 776                 unsigned long start = memslot->userspace_addr;
 777                 unsigned long end;
 778
 779                 /* mmu_lock protects userspace_addr */
 780                 if (!start)
 781                         continue;
 782
 783                 end = start + (memslot->npages << PAGE_SHIFT);
 784                 if (hva >= start && hva < end) {
 785                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 786
 787                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 788
 789                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
 790                                 int idx = gfn_offset;
 791                                 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
 792                                 retval |= handler(kvm,
 793                                         &memslot->lpage_info[j][idx].rmap_pde);
 794                         }
 795                 }
 796         }
 797
 798         return retval;
 799 }
 800
 801 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 802 {
 803         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 804 }
 805
 806 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 807 {
 808         u64 *spte;
 809         int young = 0;
 810
 811         /* always return old for EPT */
 812         if (!shadow_accessed_mask)
 813                 return 0;
 814
 815         spte = rmap_next(kvm, rmapp, NULL);
 816         while (spte) {
 817                 int _young;
 818                 u64 _spte = *spte;
 819                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 820                 _young = _spte & PT_ACCESSED_MASK;
 821                 if (_young) {
 822                         young = 1;
 823                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 824                 }
 825                 spte = rmap_next(kvm, rmapp, spte);
 826         }
 827         return young;
 828 }
 829
 830 #define RMAP_RECYCLE_THRESHOLD 1000
 831
 832 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 833 {
 834         unsigned long *rmapp;
 835         struct kvm_mmu_page *sp;
 836
 837         sp = page_header(__pa(spte));
 838
 839         gfn = unalias_gfn(vcpu->kvm, gfn);
 840         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 841
 842         kvm_unmap_rmapp(vcpu->kvm, rmapp);
 843         kvm_flush_remote_tlbs(vcpu->kvm);
 844 }
 845
 846 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 847 {
 848         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 849 }
 850
 851 #ifdef MMU_DEBUG
 852 static int is_empty_shadow_page(u64 *spt)
 853 {
 854         u64 *pos;
 855         u64 *end;
 856
 857         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 858                 if (is_shadow_present_pte(*pos)) {
 859                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 860                                pos, *pos);
 861                         return 0;
 862                 }
 863         return 1;
 864 }
 865 #endif
 866
 867 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 868 {
 869         ASSERT(is_empty_shadow_page(sp->spt));
 870         list_del(&sp->link);
 871         __free_page(virt_to_page(sp->spt));
 872         __free_page(virt_to_page(sp->gfns));
 873         kfree(sp);
 874         ++kvm->arch.n_free_mmu_pages;
 875 }
 876
 877 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 878 {
 879         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 880 }
 881
 882 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 883                                                u64 *parent_pte)
 884 {
 885         struct kvm_mmu_page *sp;
 886
 887         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 888         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 889         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 890         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 891         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 892         INIT_LIST_HEAD(&sp->oos_link);
 893         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 894         sp->multimapped = 0;
 895         sp->parent_pte = parent_pte;
 896         --vcpu->kvm->arch.n_free_mmu_pages;
 897         return sp;
 898 }
 899
 900 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 901                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 902 {
 903         struct kvm_pte_chain *pte_chain;
 904         struct hlist_node *node;
 905         int i;
 906
 907         if (!parent_pte)
 908                 return;
 909         if (!sp->multimapped) {
 910                 u64 *old = sp->parent_pte;
 911
 912                 if (!old) {
 913                         sp->parent_pte = parent_pte;
 914                         return;
 915                 }
 916                 sp->multimapped = 1;
 917                 pte_chain = mmu_alloc_pte_chain(vcpu);
 918                 INIT_HLIST_HEAD(&sp->parent_ptes);
 919                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 920                 pte_chain->parent_ptes[0] = old;
 921         }
 922         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 923                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 924                         continue;
 925                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 926                         if (!pte_chain->parent_ptes[i]) {
 927                                 pte_chain->parent_ptes[i] = parent_pte;
 928                                 return;
 929                         }
 930         }
 931         pte_chain = mmu_alloc_pte_chain(vcpu);
 932         BUG_ON(!pte_chain);
 933         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 934         pte_chain->parent_ptes[0] = parent_pte;
 935 }
 936
 937 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 938                                        u64 *parent_pte)
 939 {
 940         struct kvm_pte_chain *pte_chain;
 941         struct hlist_node *node;
 942         int i;
 943
 944         if (!sp->multimapped) {
 945                 BUG_ON(sp->parent_pte != parent_pte);
 946                 sp->parent_pte = NULL;
 947                 return;
 948         }
 949         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 950                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 951                         if (!pte_chain->parent_ptes[i])
 952                                 break;
 953                         if (pte_chain->parent_ptes[i] != parent_pte)
 954                                 continue;
 955                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 956                                 && pte_chain->parent_ptes[i + 1]) {
 957                                 pte_chain->parent_ptes[i]
 958                                         = pte_chain->parent_ptes[i + 1];
 959                                 ++i;
 960                         }
 961                         pte_chain->parent_ptes[i] = NULL;
 962                         if (i == 0) {
 963                                 hlist_del(&pte_chain->link);
 964                                 mmu_free_pte_chain(pte_chain);
 965                                 if (hlist_empty(&sp->parent_ptes)) {
 966                                         sp->multimapped = 0;
 967                                         sp->parent_pte = NULL;
 968                                 }
 969                         }
 970                         return;
 971                 }
 972         BUG();
 973 }
 974
 975
 976 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 977                             mmu_parent_walk_fn fn)
 978 {
 979         struct kvm_pte_chain *pte_chain;
 980         struct hlist_node *node;
 981         struct kvm_mmu_page *parent_sp;
 982         int i;
 983
 984         if (!sp->multimapped && sp->parent_pte) {
 985                 parent_sp = page_header(__pa(sp->parent_pte));
 986                 fn(vcpu, parent_sp);
 987                 mmu_parent_walk(vcpu, parent_sp, fn);
 988                 return;
 989         }
 990         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 991                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 992                         if (!pte_chain->parent_ptes[i])
 993                                 break;
 994                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 995                         fn(vcpu, parent_sp);
 996                         mmu_parent_walk(vcpu, parent_sp, fn);
 997                 }
 998 }
 999
1000 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1001 {
1002         unsigned int index;
1003         struct kvm_mmu_page *sp = page_header(__pa(spte));
1004
1005         index = spte - sp->spt;
1006         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1007                 sp->unsync_children++;
1008         WARN_ON(!sp->unsync_children);
1009 }
1010
1011 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1012 {
1013         struct kvm_pte_chain *pte_chain;
1014         struct hlist_node *node;
1015         int i;
1016
1017         if (!sp->parent_pte)
1018                 return;
1019
1020         if (!sp->multimapped) {
1021                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1022                 return;
1023         }
1024
1025         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1026                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1027                         if (!pte_chain->parent_ptes[i])
1028                                 break;
1029                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1030                 }
1031 }
1032
1033 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1034 {
1035         kvm_mmu_update_parents_unsync(sp);
1036         return 1;
1037 }
1038
1039 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
1040                                         struct kvm_mmu_page *sp)
1041 {
1042         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
1043         kvm_mmu_update_parents_unsync(sp);
1044 }
1045
1046 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1047                                     struct kvm_mmu_page *sp)
1048 {
1049         int i;
1050
1051         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1052                 sp->spt[i] = shadow_trap_nonpresent_pte;
1053 }
1054
1055 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1056                                struct kvm_mmu_page *sp)
1057 {
1058         return 1;
1059 }
1060
1061 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1062 {
1063 }
1064
1065 #define KVM_PAGE_ARRAY_NR 16
1066
1067 struct kvm_mmu_pages {
1068         struct mmu_page_and_offset {
1069                 struct kvm_mmu_page *sp;
1070                 unsigned int idx;
1071         } page[KVM_PAGE_ARRAY_NR];
1072         unsigned int nr;
1073 };
1074
1075 #define for_each_unsync_children(bitmap, idx)           \
1076         for (idx = find_first_bit(bitmap, 512);         \
1077              idx < 512;                                 \
1078              idx = find_next_bit(bitmap, 512, idx+1))
1079
1080 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1081                          int idx)
1082 {
1083         int i;
1084
1085         if (sp->unsync)
1086                 for (i=0; i < pvec->nr; i++)
1087                         if (pvec->page[i].sp == sp)
1088                                 return 0;
1089
1090         pvec->page[pvec->nr].sp = sp;
1091         pvec->page[pvec->nr].idx = idx;
1092         pvec->nr++;
1093         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1094 }
1095
1096 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1097                            struct kvm_mmu_pages *pvec)
1098 {
1099         int i, ret, nr_unsync_leaf = 0;
1100
1101         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1102                 u64 ent = sp->spt[i];
1103
1104                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1105                         struct kvm_mmu_page *child;
1106                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1107
1108                         if (child->unsync_children) {
1109                                 if (mmu_pages_add(pvec, child, i))
1110                                         return -ENOSPC;
1111
1112                                 ret = __mmu_unsync_walk(child, pvec);
1113                                 if (!ret)
1114                                         __clear_bit(i, sp->unsync_child_bitmap);
1115                                 else if (ret > 0)
1116                                         nr_unsync_leaf += ret;
1117                                 else
1118                                         return ret;
1119                         }
1120
1121                         if (child->unsync) {
1122                                 nr_unsync_leaf++;
1123                                 if (mmu_pages_add(pvec, child, i))
1124                                         return -ENOSPC;
1125                         }
1126                 }
1127         }
1128
1129         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1130                 sp->unsync_children = 0;
1131
1132         return nr_unsync_leaf;
1133 }
1134
1135 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1136                            struct kvm_mmu_pages *pvec)
1137 {
1138         if (!sp->unsync_children)
1139                 return 0;
1140
1141         mmu_pages_add(pvec, sp, 0);
1142         return __mmu_unsync_walk(sp, pvec);
1143 }
1144
1145 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1146 {
1147         unsigned index;
1148         struct hlist_head *bucket;
1149         struct kvm_mmu_page *sp;
1150         struct hlist_node *node;
1151
1152         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1153         index = kvm_page_table_hashfn(gfn);
1154         bucket = &kvm->arch.mmu_page_hash[index];
1155         hlist_for_each_entry(sp, node, bucket, hash_link)
1156                 if (sp->gfn == gfn && !sp->role.direct
1157                     && !sp->role.invalid) {
1158                         pgprintk("%s: found role %x\n",
1159                                  __func__, sp->role.word);
1160                         return sp;
1161                 }
1162         return NULL;
1163 }
1164
1165 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1166 {
1167         WARN_ON(!sp->unsync);
1168         sp->unsync = 0;
1169         --kvm->stat.mmu_unsync;
1170 }
1171
1172 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1173
1174 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1175 {
1176         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1177                 kvm_mmu_zap_page(vcpu->kvm, sp);
1178                 return 1;
1179         }
1180
1181         trace_kvm_mmu_sync_page(sp);
1182         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1183                 kvm_flush_remote_tlbs(vcpu->kvm);
1184         kvm_unlink_unsync_page(vcpu->kvm, sp);
1185         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1186                 kvm_mmu_zap_page(vcpu->kvm, sp);
1187                 return 1;
1188         }
1189
1190         kvm_mmu_flush_tlb(vcpu);
1191         return 0;
1192 }
1193
1194 struct mmu_page_path {
1195         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1196         unsigned int idx[PT64_ROOT_LEVEL-1];
1197 };
1198
1199 #define for_each_sp(pvec, sp, parents, i)                       \
1200                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1201                         sp = pvec.page[i].sp;                   \
1202                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1203                         i = mmu_pages_next(&pvec, &parents, i))
1204
1205 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1206                           struct mmu_page_path *parents,
1207                           int i)
1208 {
1209         int n;
1210
1211         for (n = i+1; n < pvec->nr; n++) {
1212                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1213
1214                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1215                         parents->idx[0] = pvec->page[n].idx;
1216                         return n;
1217                 }
1218
1219                 parents->parent[sp->role.level-2] = sp;
1220                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1221         }
1222
1223         return n;
1224 }
1225
1226 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1227 {
1228         struct kvm_mmu_page *sp;
1229         unsigned int level = 0;
1230
1231         do {
1232                 unsigned int idx = parents->idx[level];
1233
1234                 sp = parents->parent[level];
1235                 if (!sp)
1236                         return;
1237
1238                 --sp->unsync_children;
1239                 WARN_ON((int)sp->unsync_children < 0);
1240                 __clear_bit(idx, sp->unsync_child_bitmap);
1241                 level++;
1242         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1243 }
1244
1245 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1246                                struct mmu_page_path *parents,
1247                                struct kvm_mmu_pages *pvec)
1248 {
1249         parents->parent[parent->role.level-1] = NULL;
1250         pvec->nr = 0;
1251 }
1252
1253 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1254                               struct kvm_mmu_page *parent)
1255 {
1256         int i;
1257         struct kvm_mmu_page *sp;
1258         struct mmu_page_path parents;
1259         struct kvm_mmu_pages pages;
1260
1261         kvm_mmu_pages_init(parent, &parents, &pages);
1262         while (mmu_unsync_walk(parent, &pages)) {
1263                 int protected = 0;
1264
1265                 for_each_sp(pages, sp, parents, i)
1266                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1267
1268                 if (protected)
1269                         kvm_flush_remote_tlbs(vcpu->kvm);
1270
1271                 for_each_sp(pages, sp, parents, i) {
1272                         kvm_sync_page(vcpu, sp);
1273                         mmu_pages_clear_parents(&parents);
1274                 }
1275                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1276                 kvm_mmu_pages_init(parent, &parents, &pages);
1277         }
1278 }
1279
1280 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1281                                              gfn_t gfn,
1282                                              gva_t gaddr,
1283                                              unsigned level,
1284                                              int direct,
1285                                              unsigned access,
1286                                              u64 *parent_pte)
1287 {
1288         union kvm_mmu_page_role role;
1289         unsigned index;
1290         unsigned quadrant;
1291         struct hlist_head *bucket;
1292         struct kvm_mmu_page *sp;
1293         struct hlist_node *node, *tmp;
1294
1295         role = vcpu->arch.mmu.base_role;
1296         role.level = level;
1297         role.direct = direct;
1298         role.access = access;
1299         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1300                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1301                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1302                 role.quadrant = quadrant;
1303         }
1304         index = kvm_page_table_hashfn(gfn);
1305         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1306         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1307                 if (sp->gfn == gfn) {
1308                         if (sp->unsync)
1309                                 if (kvm_sync_page(vcpu, sp))
1310                                         continue;
1311
1312                         if (sp->role.word != role.word)
1313                                 continue;
1314
1315                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1316                         if (sp->unsync_children) {
1317                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1318                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1319                         }
1320                         trace_kvm_mmu_get_page(sp, false);
1321                         return sp;
1322                 }
1323         ++vcpu->kvm->stat.mmu_cache_miss;
1324         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1325         if (!sp)
1326                 return sp;
1327         sp->gfn = gfn;
1328         sp->role = role;
1329         hlist_add_head(&sp->hash_link, bucket);
1330         if (!direct) {
1331                 if (rmap_write_protect(vcpu->kvm, gfn))
1332                         kvm_flush_remote_tlbs(vcpu->kvm);
1333                 account_shadowed(vcpu->kvm, gfn);
1334         }
1335         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1336                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1337         else
1338                 nonpaging_prefetch_page(vcpu, sp);
1339         trace_kvm_mmu_get_page(sp, true);
1340         return sp;
1341 }
1342
1343 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1344                              struct kvm_vcpu *vcpu, u64 addr)
1345 {
1346         iterator->addr = addr;
1347         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1348         iterator->level = vcpu->arch.mmu.shadow_root_level;
1349         if (iterator->level == PT32E_ROOT_LEVEL) {
1350                 iterator->shadow_addr
1351                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1352                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1353                 --iterator->level;
1354                 if (!iterator->shadow_addr)
1355                         iterator->level = 0;
1356         }
1357 }
1358
1359 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1360 {
1361         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1362                 return false;
1363
1364         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1365                 if (is_large_pte(*iterator->sptep))
1366                         return false;
1367
1368         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1369         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1370         return true;
1371 }
1372
1373 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1374 {
1375         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1376         --iterator->level;
1377 }
1378
1379 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1380                                          struct kvm_mmu_page *sp)
1381 {
1382         unsigned i;
1383         u64 *pt;
1384         u64 ent;
1385
1386         pt = sp->spt;
1387
1388         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1389                 ent = pt[i];
1390
1391                 if (is_shadow_present_pte(ent)) {
1392                         if (!is_last_spte(ent, sp->role.level)) {
1393                                 ent &= PT64_BASE_ADDR_MASK;
1394                                 mmu_page_remove_parent_pte(page_header(ent),
1395                                                            &pt[i]);
1396                         } else {
1397                                 if (is_large_pte(ent))
1398                                         --kvm->stat.lpages;
1399                                 rmap_remove(kvm, &pt[i]);
1400                         }
1401                 }
1402                 pt[i] = shadow_trap_nonpresent_pte;
1403         }
1404 }
1405
1406 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1407 {
1408         mmu_page_remove_parent_pte(sp, parent_pte);
1409 }
1410
1411 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1412 {
1413         int i;
1414         struct kvm_vcpu *vcpu;
1415
1416         kvm_for_each_vcpu(i, vcpu, kvm)
1417                 vcpu->arch.last_pte_updated = NULL;
1418 }
1419
1420 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1421 {
1422         u64 *parent_pte;
1423
1424         while (sp->multimapped || sp->parent_pte) {
1425                 if (!sp->multimapped)
1426                         parent_pte = sp->parent_pte;
1427                 else {
1428                         struct kvm_pte_chain *chain;
1429
1430                         chain = container_of(sp->parent_ptes.first,
1431                                              struct kvm_pte_chain, link);
1432                         parent_pte = chain->parent_ptes[0];
1433                 }
1434                 BUG_ON(!parent_pte);
1435                 kvm_mmu_put_page(sp, parent_pte);
1436                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1437         }
1438 }
1439
1440 static int mmu_zap_unsync_children(struct kvm *kvm,
1441                                    struct kvm_mmu_page *parent)
1442 {
1443         int i, zapped = 0;
1444         struct mmu_page_path parents;
1445         struct kvm_mmu_pages pages;
1446
1447         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1448                 return 0;
1449
1450         kvm_mmu_pages_init(parent, &parents, &pages);
1451         while (mmu_unsync_walk(parent, &pages)) {
1452                 struct kvm_mmu_page *sp;
1453
1454                 for_each_sp(pages, sp, parents, i) {
1455                         kvm_mmu_zap_page(kvm, sp);
1456                         mmu_pages_clear_parents(&parents);
1457                 }
1458                 zapped += pages.nr;
1459                 kvm_mmu_pages_init(parent, &parents, &pages);
1460         }
1461
1462         return zapped;
1463 }
1464
1465 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1466 {
1467         int ret;
1468
1469         trace_kvm_mmu_zap_page(sp);
1470         ++kvm->stat.mmu_shadow_zapped;
1471         ret = mmu_zap_unsync_children(kvm, sp);
1472         kvm_mmu_page_unlink_children(kvm, sp);
1473         kvm_mmu_unlink_parents(kvm, sp);
1474         kvm_flush_remote_tlbs(kvm);
1475         if (!sp->role.invalid && !sp->role.direct)
1476                 unaccount_shadowed(kvm, sp->gfn);
1477         if (sp->unsync)
1478                 kvm_unlink_unsync_page(kvm, sp);
1479         if (!sp->root_count) {
1480                 hlist_del(&sp->hash_link);
1481                 kvm_mmu_free_page(kvm, sp);
1482         } else {
1483                 sp->role.invalid = 1;
1484                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1485                 kvm_reload_remote_mmus(kvm);
1486         }
1487         kvm_mmu_reset_last_pte_updated(kvm);
1488         return ret;
1489 }
1490
1491 /*
1492  * Changing the number of mmu pages allocated to the vm
1493  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1494  */
1495 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1496 {
1497         int used_pages;
1498
1499         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1500         used_pages = max(0, used_pages);
1501
1502         /*
1503          * If we set the number of mmu pages to be smaller be than the
1504          * number of actived pages , we must to free some mmu pages before we
1505          * change the value
1506          */
1507
1508         if (used_pages > kvm_nr_mmu_pages) {
1509                 while (used_pages > kvm_nr_mmu_pages) {
1510                         struct kvm_mmu_page *page;
1511
1512                         page = container_of(kvm->arch.active_mmu_pages.prev,
1513                                             struct kvm_mmu_page, link);
1514                         kvm_mmu_zap_page(kvm, page);
1515                         used_pages--;
1516                 }
1517                 kvm->arch.n_free_mmu_pages = 0;
1518         }
1519         else
1520                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1521                                          - kvm->arch.n_alloc_mmu_pages;
1522
1523         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1524 }
1525
1526 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1527 {
1528         unsigned index;
1529         struct hlist_head *bucket;
1530         struct kvm_mmu_page *sp;
1531         struct hlist_node *node, *n;
1532         int r;
1533
1534         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1535         r = 0;
1536         index = kvm_page_table_hashfn(gfn);
1537         bucket = &kvm->arch.mmu_page_hash[index];
1538         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1539                 if (sp->gfn == gfn && !sp->role.direct) {
1540                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1541                                  sp->role.word);
1542                         r = 1;
1543                         if (kvm_mmu_zap_page(kvm, sp))
1544                                 n = bucket->first;
1545                 }
1546         return r;
1547 }
1548
1549 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1550 {
1551         unsigned index;
1552         struct hlist_head *bucket;
1553         struct kvm_mmu_page *sp;
1554         struct hlist_node *node, *nn;
1555
1556         index = kvm_page_table_hashfn(gfn);
1557         bucket = &kvm->arch.mmu_page_hash[index];
1558         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1559                 if (sp->gfn == gfn && !sp->role.direct
1560                     && !sp->role.invalid) {
1561                         pgprintk("%s: zap %lx %x\n",
1562                                  __func__, gfn, sp->role.word);
1563                         kvm_mmu_zap_page(kvm, sp);
1564                 }
1565         }
1566 }
1567
1568 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1569 {
1570         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1571         struct kvm_mmu_page *sp = page_header(__pa(pte));
1572
1573         __set_bit(slot, sp->slot_bitmap);
1574 }
1575
1576 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1577 {
1578         int i;
1579         u64 *pt = sp->spt;
1580
1581         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1582                 return;
1583
1584         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1585                 if (pt[i] == shadow_notrap_nonpresent_pte)
1586                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1587         }
1588 }
1589
1590 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1591 {
1592         struct page *page;
1593
1594         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1595
1596         if (gpa == UNMAPPED_GVA)
1597                 return NULL;
1598
1599         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1600
1601         return page;
1602 }
1603
1604 /*
1605  * The function is based on mtrr_type_lookup() in
1606  * arch/x86/kernel/cpu/mtrr/generic.c
1607  */
1608 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1609                          u64 start, u64 end)
1610 {
1611         int i;
1612         u64 base, mask;
1613         u8 prev_match, curr_match;
1614         int num_var_ranges = KVM_NR_VAR_MTRR;
1615
1616         if (!mtrr_state->enabled)
1617                 return 0xFF;
1618
1619         /* Make end inclusive end, instead of exclusive */
1620         end--;
1621
1622         /* Look in fixed ranges. Just return the type as per start */
1623         if (mtrr_state->have_fixed && (start < 0x100000)) {
1624                 int idx;
1625
1626                 if (start < 0x80000) {
1627                         idx = 0;
1628                         idx += (start >> 16);
1629                         return mtrr_state->fixed_ranges[idx];
1630                 } else if (start < 0xC0000) {
1631                         idx = 1 * 8;
1632                         idx += ((start - 0x80000) >> 14);
1633                         return mtrr_state->fixed_ranges[idx];
1634                 } else if (start < 0x1000000) {
1635                         idx = 3 * 8;
1636                         idx += ((start - 0xC0000) >> 12);
1637                         return mtrr_state->fixed_ranges[idx];
1638                 }
1639         }
1640
1641         /*
1642          * Look in variable ranges
1643          * Look of multiple ranges matching this address and pick type
1644          * as per MTRR precedence
1645          */
1646         if (!(mtrr_state->enabled & 2))
1647                 return mtrr_state->def_type;
1648
1649         prev_match = 0xFF;
1650         for (i = 0; i < num_var_ranges; ++i) {
1651                 unsigned short start_state, end_state;
1652
1653                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1654                         continue;
1655
1656                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1657                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1658                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1659                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1660
1661                 start_state = ((start & mask) == (base & mask));
1662                 end_state = ((end & mask) == (base & mask));
1663                 if (start_state != end_state)
1664                         return 0xFE;
1665
1666                 if ((start & mask) != (base & mask))
1667                         continue;
1668
1669                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1670                 if (prev_match == 0xFF) {
1671                         prev_match = curr_match;
1672                         continue;
1673                 }
1674
1675                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1676                     curr_match == MTRR_TYPE_UNCACHABLE)
1677                         return MTRR_TYPE_UNCACHABLE;
1678
1679                 if ((prev_match == MTRR_TYPE_WRBACK &&
1680                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1681                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1682                      curr_match == MTRR_TYPE_WRBACK)) {
1683                         prev_match = MTRR_TYPE_WRTHROUGH;
1684                         curr_match = MTRR_TYPE_WRTHROUGH;
1685                 }
1686
1687                 if (prev_match != curr_match)
1688                         return MTRR_TYPE_UNCACHABLE;
1689         }
1690
1691         if (prev_match != 0xFF)
1692                 return prev_match;
1693
1694         return mtrr_state->def_type;
1695 }
1696
1697 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1698 {
1699         u8 mtrr;
1700
1701         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1702                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1703         if (mtrr == 0xfe || mtrr == 0xff)
1704                 mtrr = MTRR_TYPE_WRBACK;
1705         return mtrr;
1706 }
1707 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1708
1709 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1710 {
1711         unsigned index;
1712         struct hlist_head *bucket;
1713         struct kvm_mmu_page *s;
1714         struct hlist_node *node, *n;
1715
1716         trace_kvm_mmu_unsync_page(sp);
1717         index = kvm_page_table_hashfn(sp->gfn);
1718         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1719         /* don't unsync if pagetable is shadowed with multiple roles */
1720         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1721                 if (s->gfn != sp->gfn || s->role.direct)
1722                         continue;
1723                 if (s->role.word != sp->role.word)
1724                         return 1;
1725         }
1726         ++vcpu->kvm->stat.mmu_unsync;
1727         sp->unsync = 1;
1728
1729         kvm_mmu_mark_parents_unsync(vcpu, sp);
1730
1731         mmu_convert_notrap(sp);
1732         return 0;
1733 }
1734
1735 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1736                                   bool can_unsync)
1737 {
1738         struct kvm_mmu_page *shadow;
1739
1740         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1741         if (shadow) {
1742                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1743                         return 1;
1744                 if (shadow->unsync)
1745                         return 0;
1746                 if (can_unsync && oos_shadow)
1747                         return kvm_unsync_page(vcpu, shadow);
1748                 return 1;
1749         }
1750         return 0;
1751 }
1752
1753 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1754                     unsigned pte_access, int user_fault,
1755                     int write_fault, int dirty, int level,
1756                     gfn_t gfn, pfn_t pfn, bool speculative,
1757                     bool can_unsync)
1758 {
1759         u64 spte;
1760         int ret = 0;
1761
1762         /*
1763          * We don't set the accessed bit, since we sometimes want to see
1764          * whether the guest actually used the pte (in order to detect
1765          * demand paging).
1766          */
1767         spte = shadow_base_present_pte | shadow_dirty_mask;
1768         if (!speculative)
1769                 spte |= shadow_accessed_mask;
1770         if (!dirty)
1771                 pte_access &= ~ACC_WRITE_MASK;
1772         if (pte_access & ACC_EXEC_MASK)
1773                 spte |= shadow_x_mask;
1774         else
1775                 spte |= shadow_nx_mask;
1776         if (pte_access & ACC_USER_MASK)
1777                 spte |= shadow_user_mask;
1778         if (level > PT_PAGE_TABLE_LEVEL)
1779                 spte |= PT_PAGE_SIZE_MASK;
1780         if (tdp_enabled)
1781                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1782                         kvm_is_mmio_pfn(pfn));
1783
1784         spte |= (u64)pfn << PAGE_SHIFT;
1785
1786         if ((pte_access & ACC_WRITE_MASK)
1787             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1788
1789                 if (level > PT_PAGE_TABLE_LEVEL &&
1790                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1791                         ret = 1;
1792                         spte = shadow_trap_nonpresent_pte;
1793                         goto set_pte;
1794                 }
1795
1796                 spte |= PT_WRITABLE_MASK;
1797
1798                 /*
1799                  * Optimization: for pte sync, if spte was writable the hash
1800                  * lookup is unnecessary (and expensive). Write protection
1801                  * is responsibility of mmu_get_page / kvm_sync_page.
1802                  * Same reasoning can be applied to dirty page accounting.
1803                  */
1804                 if (!can_unsync && is_writeble_pte(*sptep))
1805                         goto set_pte;
1806
1807                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1808                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1809                                  __func__, gfn);
1810                         ret = 1;
1811                         pte_access &= ~ACC_WRITE_MASK;
1812                         if (is_writeble_pte(spte))
1813                                 spte &= ~PT_WRITABLE_MASK;
1814                 }
1815         }
1816
1817         if (pte_access & ACC_WRITE_MASK)
1818                 mark_page_dirty(vcpu->kvm, gfn);
1819
1820 set_pte:
1821         __set_spte(sptep, spte);
1822         return ret;
1823 }
1824
1825 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1826                          unsigned pt_access, unsigned pte_access,
1827                          int user_fault, int write_fault, int dirty,
1828                          int *ptwrite, int level, gfn_t gfn,
1829                          pfn_t pfn, bool speculative)
1830 {
1831         int was_rmapped = 0;
1832         int was_writeble = is_writeble_pte(*sptep);
1833         int rmap_count;
1834
1835         pgprintk("%s: spte %llx access %x write_fault %d"
1836                  " user_fault %d gfn %lx\n",
1837                  __func__, *sptep, pt_access,
1838                  write_fault, user_fault, gfn);
1839
1840         if (is_rmap_spte(*sptep)) {
1841                 /*
1842                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1843                  * the parent of the now unreachable PTE.
1844                  */
1845                 if (level > PT_PAGE_TABLE_LEVEL &&
1846                     !is_large_pte(*sptep)) {
1847                         struct kvm_mmu_page *child;
1848                         u64 pte = *sptep;
1849
1850                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1851                         mmu_page_remove_parent_pte(child, sptep);
1852                 } else if (pfn != spte_to_pfn(*sptep)) {
1853                         pgprintk("hfn old %lx new %lx\n",
1854                                  spte_to_pfn(*sptep), pfn);
1855                         rmap_remove(vcpu->kvm, sptep);
1856                 } else
1857                         was_rmapped = 1;
1858         }
1859
1860         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1861                       dirty, level, gfn, pfn, speculative, true)) {
1862                 if (write_fault)
1863                         *ptwrite = 1;
1864                 kvm_x86_ops->tlb_flush(vcpu);
1865         }
1866
1867         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1868         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1869                  is_large_pte(*sptep)? "2MB" : "4kB",
1870                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1871                  *sptep, sptep);
1872         if (!was_rmapped && is_large_pte(*sptep))
1873                 ++vcpu->kvm->stat.lpages;
1874
1875         page_header_update_slot(vcpu->kvm, sptep, gfn);
1876         if (!was_rmapped) {
1877                 rmap_count = rmap_add(vcpu, sptep, gfn);
1878                 kvm_release_pfn_clean(pfn);
1879                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1880                         rmap_recycle(vcpu, sptep, gfn);
1881         } else {
1882                 if (was_writeble)
1883                         kvm_release_pfn_dirty(pfn);
1884                 else
1885                         kvm_release_pfn_clean(pfn);
1886         }
1887         if (speculative) {
1888                 vcpu->arch.last_pte_updated = sptep;
1889                 vcpu->arch.last_pte_gfn = gfn;
1890         }
1891 }
1892
1893 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1894 {
1895 }
1896
1897 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1898                         int level, gfn_t gfn, pfn_t pfn)
1899 {
1900         struct kvm_shadow_walk_iterator iterator;
1901         struct kvm_mmu_page *sp;
1902         int pt_write = 0;
1903         gfn_t pseudo_gfn;
1904
1905         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1906                 if (iterator.level == level) {
1907                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1908                                      0, write, 1, &pt_write,
1909                                      level, gfn, pfn, false);
1910                         ++vcpu->stat.pf_fixed;
1911                         break;
1912                 }
1913
1914                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1915                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1916                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1917                                               iterator.level - 1,
1918                                               1, ACC_ALL, iterator.sptep);
1919                         if (!sp) {
1920                                 pgprintk("nonpaging_map: ENOMEM\n");
1921                                 kvm_release_pfn_clean(pfn);
1922                                 return -ENOMEM;
1923                         }
1924
1925                         __set_spte(iterator.sptep,
1926                                    __pa(sp->spt)
1927                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1928                                    | shadow_user_mask | shadow_x_mask);
1929                 }
1930         }
1931         return pt_write;
1932 }
1933
1934 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1935 {
1936         int r;
1937         int level;
1938         pfn_t pfn;
1939         unsigned long mmu_seq;
1940
1941         level = mapping_level(vcpu, gfn);
1942
1943         /*
1944          * This path builds a PAE pagetable - so we can map 2mb pages at
1945          * maximum. Therefore check if the level is larger than that.
1946          */
1947         if (level > PT_DIRECTORY_LEVEL)
1948                 level = PT_DIRECTORY_LEVEL;
1949
1950         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
1951
1952         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1953         smp_rmb();
1954         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1955
1956         /* mmio */
1957         if (is_error_pfn(pfn)) {
1958                 kvm_release_pfn_clean(pfn);
1959                 return 1;
1960         }
1961
1962         spin_lock(&vcpu->kvm->mmu_lock);
1963         if (mmu_notifier_retry(vcpu, mmu_seq))
1964                 goto out_unlock;
1965         kvm_mmu_free_some_pages(vcpu);
1966         r = __direct_map(vcpu, v, write, level, gfn, pfn);
1967         spin_unlock(&vcpu->kvm->mmu_lock);
1968
1969
1970         return r;
1971
1972 out_unlock:
1973         spin_unlock(&vcpu->kvm->mmu_lock);
1974         kvm_release_pfn_clean(pfn);
1975         return 0;
1976 }
1977
1978
1979 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1980 {
1981         int i;
1982         struct kvm_mmu_page *sp;
1983
1984         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1985                 return;
1986         spin_lock(&vcpu->kvm->mmu_lock);
1987         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1988                 hpa_t root = vcpu->arch.mmu.root_hpa;
1989
1990                 sp = page_header(root);
1991                 --sp->root_count;
1992                 if (!sp->root_count && sp->role.invalid)
1993                         kvm_mmu_zap_page(vcpu->kvm, sp);
1994                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1995                 spin_unlock(&vcpu->kvm->mmu_lock);
1996                 return;
1997         }
1998         for (i = 0; i < 4; ++i) {
1999                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2000
2001                 if (root) {
2002                         root &= PT64_BASE_ADDR_MASK;
2003                         sp = page_header(root);
2004                         --sp->root_count;
2005                         if (!sp->root_count && sp->role.invalid)
2006                                 kvm_mmu_zap_page(vcpu->kvm, sp);
2007                 }
2008                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2009         }
2010         spin_unlock(&vcpu->kvm->mmu_lock);
2011         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2012 }
2013
2014 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2015 {
2016         int ret = 0;
2017
2018         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2019                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2020                 ret = 1;
2021         }
2022
2023         return ret;
2024 }
2025
2026 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2027 {
2028         int i;
2029         gfn_t root_gfn;
2030         struct kvm_mmu_page *sp;
2031         int direct = 0;
2032         u64 pdptr;
2033
2034         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2035
2036         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2037                 hpa_t root = vcpu->arch.mmu.root_hpa;
2038
2039                 ASSERT(!VALID_PAGE(root));
2040                 if (tdp_enabled)
2041                         direct = 1;
2042                 if (mmu_check_root(vcpu, root_gfn))
2043                         return 1;
2044                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2045                                       PT64_ROOT_LEVEL, direct,
2046                                       ACC_ALL, NULL);
2047                 root = __pa(sp->spt);
2048                 ++sp->root_count;
2049                 vcpu->arch.mmu.root_hpa = root;
2050                 return 0;
2051         }
2052         direct = !is_paging(vcpu);
2053         if (tdp_enabled)
2054                 direct = 1;
2055         for (i = 0; i < 4; ++i) {
2056                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2057
2058                 ASSERT(!VALID_PAGE(root));
2059                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2060                         pdptr = kvm_pdptr_read(vcpu, i);
2061                         if (!is_present_gpte(pdptr)) {
2062                                 vcpu->arch.mmu.pae_root[i] = 0;
2063                                 continue;
2064                         }
2065                         root_gfn = pdptr >> PAGE_SHIFT;
2066                 } else if (vcpu->arch.mmu.root_level == 0)
2067                         root_gfn = 0;
2068                 if (mmu_check_root(vcpu, root_gfn))
2069                         return 1;
2070                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2071                                       PT32_ROOT_LEVEL, direct,
2072                                       ACC_ALL, NULL);
2073                 root = __pa(sp->spt);
2074                 ++sp->root_count;
2075                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2076         }
2077         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2078         return 0;
2079 }
2080
2081 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2082 {
2083         int i;
2084         struct kvm_mmu_page *sp;
2085
2086         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2087                 return;
2088         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2089                 hpa_t root = vcpu->arch.mmu.root_hpa;
2090                 sp = page_header(root);
2091                 mmu_sync_children(vcpu, sp);
2092                 return;
2093         }
2094         for (i = 0; i < 4; ++i) {
2095                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2096
2097                 if (root && VALID_PAGE(root)) {
2098                         root &= PT64_BASE_ADDR_MASK;
2099                         sp = page_header(root);
2100                         mmu_sync_children(vcpu, sp);
2101                 }
2102         }
2103 }
2104
2105 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2106 {
2107         spin_lock(&vcpu->kvm->mmu_lock);
2108         mmu_sync_roots(vcpu);
2109         spin_unlock(&vcpu->kvm->mmu_lock);
2110 }
2111
2112 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2113 {
2114         return vaddr;
2115 }
2116
2117 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2118                                 u32 error_code)
2119 {
2120         gfn_t gfn;
2121         int r;
2122
2123         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2124         r = mmu_topup_memory_caches(vcpu);
2125         if (r)
2126                 return r;
2127
2128         ASSERT(vcpu);
2129         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2130
2131         gfn = gva >> PAGE_SHIFT;
2132
2133         return nonpaging_map(vcpu, gva & PAGE_MASK,
2134                              error_code & PFERR_WRITE_MASK, gfn);
2135 }
2136
2137 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2138                                 u32 error_code)
2139 {
2140         pfn_t pfn;
2141         int r;
2142         int level;
2143         gfn_t gfn = gpa >> PAGE_SHIFT;
2144         unsigned long mmu_seq;
2145
2146         ASSERT(vcpu);
2147         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2148
2149         r = mmu_topup_memory_caches(vcpu);
2150         if (r)
2151                 return r;
2152
2153         level = mapping_level(vcpu, gfn);
2154
2155         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2156
2157         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2158         smp_rmb();
2159         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2160         if (is_error_pfn(pfn)) {
2161                 kvm_release_pfn_clean(pfn);
2162                 return 1;
2163         }
2164         spin_lock(&vcpu->kvm->mmu_lock);
2165         if (mmu_notifier_retry(vcpu, mmu_seq))
2166                 goto out_unlock;
2167         kvm_mmu_free_some_pages(vcpu);
2168         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2169                          level, gfn, pfn);
2170         spin_unlock(&vcpu->kvm->mmu_lock);
2171
2172         return r;
2173
2174 out_unlock:
2175         spin_unlock(&vcpu->kvm->mmu_lock);
2176         kvm_release_pfn_clean(pfn);
2177         return 0;
2178 }
2179
2180 static void nonpaging_free(struct kvm_vcpu *vcpu)
2181 {
2182         mmu_free_roots(vcpu);
2183 }
2184
2185 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2186 {
2187         struct kvm_mmu *context = &vcpu->arch.mmu;
2188
2189         context->new_cr3 = nonpaging_new_cr3;
2190         context->page_fault = nonpaging_page_fault;
2191         context->gva_to_gpa = nonpaging_gva_to_gpa;
2192         context->free = nonpaging_free;
2193         context->prefetch_page = nonpaging_prefetch_page;
2194         context->sync_page = nonpaging_sync_page;
2195         context->invlpg = nonpaging_invlpg;
2196         context->root_level = 0;
2197         context->shadow_root_level = PT32E_ROOT_LEVEL;
2198         context->root_hpa = INVALID_PAGE;
2199         return 0;
2200 }
2201
2202 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2203 {
2204         ++vcpu->stat.tlb_flush;
2205         kvm_x86_ops->tlb_flush(vcpu);
2206 }
2207
2208 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2209 {
2210         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2211         mmu_free_roots(vcpu);
2212 }
2213
2214 static void inject_page_fault(struct kvm_vcpu *vcpu,
2215                               u64 addr,
2216                               u32 err_code)
2217 {
2218         kvm_inject_page_fault(vcpu, addr, err_code);
2219 }
2220
2221 static void paging_free(struct kvm_vcpu *vcpu)
2222 {
2223         nonpaging_free(vcpu);
2224 }
2225
2226 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2227 {
2228         int bit7;
2229
2230         bit7 = (gpte >> 7) & 1;
2231         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2232 }
2233
2234 #define PTTYPE 64
2235 #include "paging_tmpl.h"
2236 #undef PTTYPE
2237
2238 #define PTTYPE 32
2239 #include "paging_tmpl.h"
2240 #undef PTTYPE
2241
2242 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2243 {
2244         struct kvm_mmu *context = &vcpu->arch.mmu;
2245         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2246         u64 exb_bit_rsvd = 0;
2247
2248         if (!is_nx(vcpu))
2249                 exb_bit_rsvd = rsvd_bits(63, 63);
2250         switch (level) {
2251         case PT32_ROOT_LEVEL:
2252                 /* no rsvd bits for 2 level 4K page table entries */
2253                 context->rsvd_bits_mask[0][1] = 0;
2254                 context->rsvd_bits_mask[0][0] = 0;
2255                 if (is_cpuid_PSE36())
2256                         /* 36bits PSE 4MB page */
2257                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2258                 else
2259                         /* 32 bits PSE 4MB page */
2260                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2261                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2262                 break;
2263         case PT32E_ROOT_LEVEL:
2264                 context->rsvd_bits_mask[0][2] =
2265                         rsvd_bits(maxphyaddr, 63) |
2266                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2267                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2268                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2269                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2270                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2271                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2272                         rsvd_bits(maxphyaddr, 62) |
2273                         rsvd_bits(13, 20);              /* large page */
2274                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2275                 break;
2276         case PT64_ROOT_LEVEL:
2277                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2278                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2279                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2280                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2281                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2282                         rsvd_bits(maxphyaddr, 51);
2283                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2284                         rsvd_bits(maxphyaddr, 51);
2285                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2286                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2287                         rsvd_bits(maxphyaddr, 51) |
2288                         rsvd_bits(13, 29);
2289                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2290                         rsvd_bits(maxphyaddr, 51) |
2291                         rsvd_bits(13, 20);              /* large page */
2292                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2293                 break;
2294         }
2295 }
2296
2297 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2298 {
2299         struct kvm_mmu *context = &vcpu->arch.mmu;
2300
2301         ASSERT(is_pae(vcpu));
2302         context->new_cr3 = paging_new_cr3;
2303         context->page_fault = paging64_page_fault;
2304         context->gva_to_gpa = paging64_gva_to_gpa;
2305         context->prefetch_page = paging64_prefetch_page;
2306         context->sync_page = paging64_sync_page;
2307         context->invlpg = paging64_invlpg;
2308         context->free = paging_free;
2309         context->root_level = level;
2310         context->shadow_root_level = level;
2311         context->root_hpa = INVALID_PAGE;
2312         return 0;
2313 }
2314
2315 static int paging64_init_context(struct kvm_vcpu *vcpu)
2316 {
2317         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2318         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2319 }
2320
2321 static int paging32_init_context(struct kvm_vcpu *vcpu)
2322 {
2323         struct kvm_mmu *context = &vcpu->arch.mmu;
2324
2325         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2326         context->new_cr3 = paging_new_cr3;
2327         context->page_fault = paging32_page_fault;
2328         context->gva_to_gpa = paging32_gva_to_gpa;
2329         context->free = paging_free;
2330         context->prefetch_page = paging32_prefetch_page;
2331         context->sync_page = paging32_sync_page;
2332         context->invlpg = paging32_invlpg;
2333         context->root_level = PT32_ROOT_LEVEL;
2334         context->shadow_root_level = PT32E_ROOT_LEVEL;
2335         context->root_hpa = INVALID_PAGE;
2336         return 0;
2337 }
2338
2339 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2340 {
2341         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2342         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2343 }
2344
2345 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2346 {
2347         struct kvm_mmu *context = &vcpu->arch.mmu;
2348
2349         context->new_cr3 = nonpaging_new_cr3;
2350         context->page_fault = tdp_page_fault;
2351         context->free = nonpaging_free;
2352         context->prefetch_page = nonpaging_prefetch_page;
2353         context->sync_page = nonpaging_sync_page;
2354         context->invlpg = nonpaging_invlpg;
2355         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2356         context->root_hpa = INVALID_PAGE;
2357
2358         if (!is_paging(vcpu)) {
2359                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2360                 context->root_level = 0;
2361         } else if (is_long_mode(vcpu)) {
2362                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2363                 context->gva_to_gpa = paging64_gva_to_gpa;
2364                 context->root_level = PT64_ROOT_LEVEL;
2365         } else if (is_pae(vcpu)) {
2366                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2367                 context->gva_to_gpa = paging64_gva_to_gpa;
2368                 context->root_level = PT32E_ROOT_LEVEL;
2369         } else {
2370                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2371                 context->gva_to_gpa = paging32_gva_to_gpa;
2372                 context->root_level = PT32_ROOT_LEVEL;
2373         }
2374
2375         return 0;
2376 }
2377
2378 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2379 {
2380         int r;
2381
2382         ASSERT(vcpu);
2383         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2384
2385         if (!is_paging(vcpu))
2386                 r = nonpaging_init_context(vcpu);
2387         else if (is_long_mode(vcpu))
2388                 r = paging64_init_context(vcpu);
2389         else if (is_pae(vcpu))
2390                 r = paging32E_init_context(vcpu);
2391         else
2392                 r = paging32_init_context(vcpu);
2393
2394         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2395
2396         return r;
2397 }
2398
2399 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2400 {
2401         vcpu->arch.update_pte.pfn = bad_pfn;
2402
2403         if (tdp_enabled)
2404                 return init_kvm_tdp_mmu(vcpu);
2405         else
2406                 return init_kvm_softmmu(vcpu);
2407 }
2408
2409 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2410 {
2411         ASSERT(vcpu);
2412         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2413                 vcpu->arch.mmu.free(vcpu);
2414                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2415         }
2416 }
2417
2418 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2419 {
2420         destroy_kvm_mmu(vcpu);
2421         return init_kvm_mmu(vcpu);
2422 }
2423 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2424
2425 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2426 {
2427         int r;
2428
2429         r = mmu_topup_memory_caches(vcpu);
2430         if (r)
2431                 goto out;
2432         spin_lock(&vcpu->kvm->mmu_lock);
2433         kvm_mmu_free_some_pages(vcpu);
2434         r = mmu_alloc_roots(vcpu);
2435         mmu_sync_roots(vcpu);
2436         spin_unlock(&vcpu->kvm->mmu_lock);
2437         if (r)
2438                 goto out;
2439         /* set_cr3() should ensure TLB has been flushed */
2440         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2441 out:
2442         return r;
2443 }
2444 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2445
2446 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2447 {
2448         mmu_free_roots(vcpu);
2449 }
2450
2451 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2452                                   struct kvm_mmu_page *sp,
2453                                   u64 *spte)
2454 {
2455         u64 pte;
2456         struct kvm_mmu_page *child;
2457
2458         pte = *spte;
2459         if (is_shadow_present_pte(pte)) {
2460                 if (is_last_spte(pte, sp->role.level))
2461                         rmap_remove(vcpu->kvm, spte);
2462                 else {
2463                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2464                         mmu_page_remove_parent_pte(child, spte);
2465                 }
2466         }
2467         __set_spte(spte, shadow_trap_nonpresent_pte);
2468         if (is_large_pte(pte))
2469                 --vcpu->kvm->stat.lpages;
2470 }
2471
2472 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2473                                   struct kvm_mmu_page *sp,
2474                                   u64 *spte,
2475                                   const void *new)
2476 {
2477         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2478                 ++vcpu->kvm->stat.mmu_pde_zapped;
2479                 return;
2480         }
2481
2482         ++vcpu->kvm->stat.mmu_pte_updated;
2483         if (sp->role.glevels == PT32_ROOT_LEVEL)
2484                 paging32_update_pte(vcpu, sp, spte, new);
2485         else
2486                 paging64_update_pte(vcpu, sp, spte, new);
2487 }
2488
2489 static bool need_remote_flush(u64 old, u64 new)
2490 {
2491         if (!is_shadow_present_pte(old))
2492                 return false;
2493         if (!is_shadow_present_pte(new))
2494                 return true;
2495         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2496                 return true;
2497         old ^= PT64_NX_MASK;
2498         new ^= PT64_NX_MASK;
2499         return (old & ~new & PT64_PERM_MASK) != 0;
2500 }
2501
2502 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2503 {
2504         if (need_remote_flush(old, new))
2505                 kvm_flush_remote_tlbs(vcpu->kvm);
2506         else
2507                 kvm_mmu_flush_tlb(vcpu);
2508 }
2509
2510 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2511 {
2512         u64 *spte = vcpu->arch.last_pte_updated;
2513
2514         return !!(spte && (*spte & shadow_accessed_mask));
2515 }
2516
2517 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2518                                           const u8 *new, int bytes)
2519 {
2520         gfn_t gfn;
2521         int r;
2522         u64 gpte = 0;
2523         pfn_t pfn;
2524
2525         if (bytes != 4 && bytes != 8)
2526                 return;
2527
2528         /*
2529          * Assume that the pte write on a page table of the same type
2530          * as the current vcpu paging mode.  This is nearly always true
2531          * (might be false while changing modes).  Note it is verified later
2532          * by update_pte().
2533          */
2534         if (is_pae(vcpu)) {
2535                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2536                 if ((bytes == 4) && (gpa % 4 == 0)) {
2537                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2538                         if (r)
2539                                 return;
2540                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2541                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2542                         memcpy((void *)&gpte, new, 8);
2543                 }
2544         } else {
2545                 if ((bytes == 4) && (gpa % 4 == 0))
2546                         memcpy((void *)&gpte, new, 4);
2547         }
2548         if (!is_present_gpte(gpte))
2549                 return;
2550         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2551
2552         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2553         smp_rmb();
2554         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2555
2556         if (is_error_pfn(pfn)) {
2557                 kvm_release_pfn_clean(pfn);
2558                 return;
2559         }
2560         vcpu->arch.update_pte.gfn = gfn;
2561         vcpu->arch.update_pte.pfn = pfn;
2562 }
2563
2564 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2565 {
2566         u64 *spte = vcpu->arch.last_pte_updated;
2567
2568         if (spte
2569             && vcpu->arch.last_pte_gfn == gfn
2570             && shadow_accessed_mask
2571             && !(*spte & shadow_accessed_mask)
2572             && is_shadow_present_pte(*spte))
2573                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2574 }
2575
2576 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2577                        const u8 *new, int bytes,
2578                        bool guest_initiated)
2579 {
2580         gfn_t gfn = gpa >> PAGE_SHIFT;
2581         struct kvm_mmu_page *sp;
2582         struct hlist_node *node, *n;
2583         struct hlist_head *bucket;
2584         unsigned index;
2585         u64 entry, gentry;
2586         u64 *spte;
2587         unsigned offset = offset_in_page(gpa);
2588         unsigned pte_size;
2589         unsigned page_offset;
2590         unsigned misaligned;
2591         unsigned quadrant;
2592         int level;
2593         int flooded = 0;
2594         int npte;
2595         int r;
2596
2597         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2598         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2599         spin_lock(&vcpu->kvm->mmu_lock);
2600         kvm_mmu_access_page(vcpu, gfn);
2601         kvm_mmu_free_some_pages(vcpu);
2602         ++vcpu->kvm->stat.mmu_pte_write;
2603         kvm_mmu_audit(vcpu, "pre pte write");
2604         if (guest_initiated) {
2605                 if (gfn == vcpu->arch.last_pt_write_gfn
2606                     && !last_updated_pte_accessed(vcpu)) {
2607                         ++vcpu->arch.last_pt_write_count;
2608                         if (vcpu->arch.last_pt_write_count >= 3)
2609                                 flooded = 1;
2610                 } else {
2611                         vcpu->arch.last_pt_write_gfn = gfn;
2612                         vcpu->arch.last_pt_write_count = 1;
2613                         vcpu->arch.last_pte_updated = NULL;
2614                 }
2615         }
2616         index = kvm_page_table_hashfn(gfn);
2617         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2618         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2619                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2620                         continue;
2621                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2622                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2623                 misaligned |= bytes < 4;
2624                 if (misaligned || flooded) {
2625                         /*
2626                          * Misaligned accesses are too much trouble to fix
2627                          * up; also, they usually indicate a page is not used
2628                          * as a page table.
2629                          *
2630                          * If we're seeing too many writes to a page,
2631                          * it may no longer be a page table, or we may be
2632                          * forking, in which case it is better to unmap the
2633                          * page.
2634                          */
2635                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2636                                  gpa, bytes, sp->role.word);
2637                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2638                                 n = bucket->first;
2639                         ++vcpu->kvm->stat.mmu_flooded;
2640                         continue;
2641                 }
2642                 page_offset = offset;
2643                 level = sp->role.level;
2644                 npte = 1;
2645                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2646                         page_offset <<= 1;      /* 32->64 */
2647                         /*
2648                          * A 32-bit pde maps 4MB while the shadow pdes map
2649                          * only 2MB.  So we need to double the offset again
2650                          * and zap two pdes instead of one.
2651                          */
2652                         if (level == PT32_ROOT_LEVEL) {
2653                                 page_offset &= ~7; /* kill rounding error */
2654                                 page_offset <<= 1;
2655                                 npte = 2;
2656                         }
2657                         quadrant = page_offset >> PAGE_SHIFT;
2658                         page_offset &= ~PAGE_MASK;
2659                         if (quadrant != sp->role.quadrant)
2660                                 continue;
2661                 }
2662                 spte = &sp->spt[page_offset / sizeof(*spte)];
2663                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2664                         gentry = 0;
2665                         r = kvm_read_guest_atomic(vcpu->kvm,
2666                                                   gpa & ~(u64)(pte_size - 1),
2667                                                   &gentry, pte_size);
2668                         new = (const void *)&gentry;
2669                         if (r < 0)
2670                                 new = NULL;
2671                 }
2672                 while (npte--) {
2673                         entry = *spte;
2674                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2675                         if (new)
2676                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2677                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2678                         ++spte;
2679                 }
2680         }
2681         kvm_mmu_audit(vcpu, "post pte write");
2682         spin_unlock(&vcpu->kvm->mmu_lock);
2683         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2684                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2685                 vcpu->arch.update_pte.pfn = bad_pfn;
2686         }
2687 }
2688
2689 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2690 {
2691         gpa_t gpa;
2692         int r;
2693
2694         if (tdp_enabled)
2695                 return 0;
2696
2697         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2698
2699         spin_lock(&vcpu->kvm->mmu_lock);
2700         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2701         spin_unlock(&vcpu->kvm->mmu_lock);
2702         return r;
2703 }
2704 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2705
2706 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2707 {
2708         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2709                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2710                 struct kvm_mmu_page *sp;
2711
2712                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2713                                   struct kvm_mmu_page, link);
2714                 kvm_mmu_zap_page(vcpu->kvm, sp);
2715                 ++vcpu->kvm->stat.mmu_recycled;
2716         }
2717 }
2718
2719 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2720 {
2721         int r;
2722         enum emulation_result er;
2723
2724         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2725         if (r < 0)
2726                 goto out;
2727
2728         if (!r) {
2729                 r = 1;
2730                 goto out;
2731         }
2732
2733         r = mmu_topup_memory_caches(vcpu);
2734         if (r)
2735                 goto out;
2736
2737         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2738
2739         switch (er) {
2740         case EMULATE_DONE:
2741                 return 1;
2742         case EMULATE_DO_MMIO:
2743                 ++vcpu->stat.mmio_exits;
2744                 return 0;
2745         case EMULATE_FAIL:
2746                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2747                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2748                 return 0;
2749         default:
2750                 BUG();
2751         }
2752 out:
2753         return r;
2754 }
2755 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2756
2757 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2758 {
2759         vcpu->arch.mmu.invlpg(vcpu, gva);
2760         kvm_mmu_flush_tlb(vcpu);
2761         ++vcpu->stat.invlpg;
2762 }
2763 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2764
2765 void kvm_enable_tdp(void)
2766 {
2767         tdp_enabled = true;
2768 }
2769 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2770
2771 void kvm_disable_tdp(void)
2772 {
2773         tdp_enabled = false;
2774 }
2775 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2776
2777 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2778 {
2779         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2780 }
2781
2782 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2783 {
2784         struct page *page;
2785         int i;
2786
2787         ASSERT(vcpu);
2788
2789         /*
2790          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2791          * Therefore we need to allocate shadow page tables in the first
2792          * 4GB of memory, which happens to fit the DMA32 zone.
2793          */
2794         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2795         if (!page)
2796                 goto error_1;
2797         vcpu->arch.mmu.pae_root = page_address(page);
2798         for (i = 0; i < 4; ++i)
2799                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2800
2801         return 0;
2802
2803 error_1:
2804         free_mmu_pages(vcpu);
2805         return -ENOMEM;
2806 }
2807
2808 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2809 {
2810         ASSERT(vcpu);
2811         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2812
2813         return alloc_mmu_pages(vcpu);
2814 }
2815
2816 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2817 {
2818         ASSERT(vcpu);
2819         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2820
2821         return init_kvm_mmu(vcpu);
2822 }
2823
2824 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2825 {
2826         ASSERT(vcpu);
2827
2828         destroy_kvm_mmu(vcpu);
2829         free_mmu_pages(vcpu);
2830         mmu_free_memory_caches(vcpu);
2831 }
2832
2833 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2834 {
2835         struct kvm_mmu_page *sp;
2836
2837         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2838                 int i;
2839                 u64 *pt;
2840
2841                 if (!test_bit(slot, sp->slot_bitmap))
2842                         continue;
2843
2844                 pt = sp->spt;
2845                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2846                         /* avoid RMW */
2847                         if (pt[i] & PT_WRITABLE_MASK)
2848                                 pt[i] &= ~PT_WRITABLE_MASK;
2849         }
2850         kvm_flush_remote_tlbs(kvm);
2851 }
2852
2853 void kvm_mmu_zap_all(struct kvm *kvm)
2854 {
2855         struct kvm_mmu_page *sp, *node;
2856
2857         spin_lock(&kvm->mmu_lock);
2858         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2859                 if (kvm_mmu_zap_page(kvm, sp))
2860                         node = container_of(kvm->arch.active_mmu_pages.next,
2861                                             struct kvm_mmu_page, link);
2862         spin_unlock(&kvm->mmu_lock);
2863
2864         kvm_flush_remote_tlbs(kvm);
2865 }
2866
2867 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2868 {
2869         struct kvm_mmu_page *page;
2870
2871         page = container_of(kvm->arch.active_mmu_pages.prev,
2872                             struct kvm_mmu_page, link);
2873         kvm_mmu_zap_page(kvm, page);
2874 }
2875
2876 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2877 {
2878         struct kvm *kvm;
2879         struct kvm *kvm_freed = NULL;
2880         int cache_count = 0;
2881
2882         spin_lock(&kvm_lock);
2883
2884         list_for_each_entry(kvm, &vm_list, vm_list) {
2885                 int npages;
2886
2887                 if (!down_read_trylock(&kvm->slots_lock))
2888                         continue;
2889                 spin_lock(&kvm->mmu_lock);
2890                 npages = kvm->arch.n_alloc_mmu_pages -
2891                          kvm->arch.n_free_mmu_pages;
2892                 cache_count += npages;
2893                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2894                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2895                         cache_count--;
2896                         kvm_freed = kvm;
2897                 }
2898                 nr_to_scan--;
2899
2900                 spin_unlock(&kvm->mmu_lock);
2901                 up_read(&kvm->slots_lock);
2902         }
2903         if (kvm_freed)
2904                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2905
2906         spin_unlock(&kvm_lock);
2907
2908         return cache_count;
2909 }
2910
2911 static struct shrinker mmu_shrinker = {
2912         .shrink = mmu_shrink,
2913         .seeks = DEFAULT_SEEKS * 10,
2914 };
2915
2916 static void mmu_destroy_caches(void)
2917 {
2918         if (pte_chain_cache)
2919                 kmem_cache_destroy(pte_chain_cache);
2920         if (rmap_desc_cache)
2921                 kmem_cache_destroy(rmap_desc_cache);
2922         if (mmu_page_header_cache)
2923                 kmem_cache_destroy(mmu_page_header_cache);
2924 }
2925
2926 void kvm_mmu_module_exit(void)
2927 {
2928         mmu_destroy_caches();
2929         unregister_shrinker(&mmu_shrinker);
2930 }
2931
2932 int kvm_mmu_module_init(void)
2933 {
2934         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2935                                             sizeof(struct kvm_pte_chain),
2936                                             0, 0, NULL);
2937         if (!pte_chain_cache)
2938                 goto nomem;
2939         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2940                                             sizeof(struct kvm_rmap_desc),
2941                                             0, 0, NULL);
2942         if (!rmap_desc_cache)
2943                 goto nomem;
2944
2945         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2946                                                   sizeof(struct kvm_mmu_page),
2947                                                   0, 0, NULL);
2948         if (!mmu_page_header_cache)
2949                 goto nomem;
2950
2951         register_shrinker(&mmu_shrinker);
2952
2953         return 0;
2954
2955 nomem:
2956         mmu_destroy_caches();
2957         return -ENOMEM;
2958 }
2959
2960 /*
2961  * Caculate mmu pages needed for kvm.
2962  */
2963 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2964 {
2965         int i;
2966         unsigned int nr_mmu_pages;
2967         unsigned int  nr_pages = 0;
2968
2969         for (i = 0; i < kvm->nmemslots; i++)
2970                 nr_pages += kvm->memslots[i].npages;
2971
2972         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2973         nr_mmu_pages = max(nr_mmu_pages,
2974                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2975
2976         return nr_mmu_pages;
2977 }
2978
2979 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2980                                 unsigned len)
2981 {
2982         if (len > buffer->len)
2983                 return NULL;
2984         return buffer->ptr;
2985 }
2986
2987 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2988                                 unsigned len)
2989 {
2990         void *ret;
2991
2992         ret = pv_mmu_peek_buffer(buffer, len);
2993         if (!ret)
2994                 return ret;
2995         buffer->ptr += len;
2996         buffer->len -= len;
2997         buffer->processed += len;
2998         return ret;
2999 }
3000
3001 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3002                              gpa_t addr, gpa_t value)
3003 {
3004         int bytes = 8;
3005         int r;
3006
3007         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3008                 bytes = 4;
3009
3010         r = mmu_topup_memory_caches(vcpu);
3011         if (r)
3012                 return r;
3013
3014         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3015                 return -EFAULT;
3016
3017         return 1;
3018 }
3019
3020 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3021 {
3022         kvm_set_cr3(vcpu, vcpu->arch.cr3);
3023         return 1;
3024 }
3025
3026 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3027 {
3028         spin_lock(&vcpu->kvm->mmu_lock);
3029         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3030         spin_unlock(&vcpu->kvm->mmu_lock);
3031         return 1;
3032 }
3033
3034 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3035                              struct kvm_pv_mmu_op_buffer *buffer)
3036 {
3037         struct kvm_mmu_op_header *header;
3038
3039         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3040         if (!header)
3041                 return 0;
3042         switch (header->op) {
3043         case KVM_MMU_OP_WRITE_PTE: {
3044                 struct kvm_mmu_op_write_pte *wpte;
3045
3046                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3047                 if (!wpte)
3048                         return 0;
3049                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3050                                         wpte->pte_val);
3051         }
3052         case KVM_MMU_OP_FLUSH_TLB: {
3053                 struct kvm_mmu_op_flush_tlb *ftlb;
3054
3055                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3056                 if (!ftlb)
3057                         return 0;
3058                 return kvm_pv_mmu_flush_tlb(vcpu);
3059         }
3060         case KVM_MMU_OP_RELEASE_PT: {
3061                 struct kvm_mmu_op_release_pt *rpt;
3062
3063                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3064                 if (!rpt)
3065                         return 0;
3066                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3067         }
3068         default: return 0;
3069         }
3070 }
3071
3072 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3073                   gpa_t addr, unsigned long *ret)
3074 {
3075         int r;
3076         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3077
3078         buffer->ptr = buffer->buf;
3079         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3080         buffer->processed = 0;
3081
3082         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3083         if (r)
3084                 goto out;
3085
3086         while (buffer->len) {
3087                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3088                 if (r < 0)
3089                         goto out;
3090                 if (r == 0)
3091                         break;
3092         }
3093
3094         r = 1;
3095 out:
3096         *ret = buffer->processed;
3097         return r;
3098 }
3099
3100 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3101 {
3102         struct kvm_shadow_walk_iterator iterator;
3103         int nr_sptes = 0;
3104
3105         spin_lock(&vcpu->kvm->mmu_lock);
3106         for_each_shadow_entry(vcpu, addr, iterator) {
3107                 sptes[iterator.level-1] = *iterator.sptep;
3108                 nr_sptes++;
3109                 if (!is_shadow_present_pte(*iterator.sptep))
3110                         break;
3111         }
3112         spin_unlock(&vcpu->kvm->mmu_lock);
3113
3114         return nr_sptes;
3115 }
3116 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3117
3118 #ifdef AUDIT
3119
3120 static const char *audit_msg;
3121
3122 static gva_t canonicalize(gva_t gva)
3123 {
3124 #ifdef CONFIG_X86_64
3125         gva = (long long)(gva << 16) >> 16;
3126 #endif
3127         return gva;
3128 }
3129
3130
3131 typedef void (*inspect_spte_fn) (struct kvm *kvm, struct kvm_mmu_page *sp,
3132                                  u64 *sptep);
3133
3134 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3135                             inspect_spte_fn fn)
3136 {
3137         int i;
3138
3139         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3140                 u64 ent = sp->spt[i];
3141
3142                 if (is_shadow_present_pte(ent)) {
3143                         if (!is_last_spte(ent, sp->role.level)) {
3144                                 struct kvm_mmu_page *child;
3145                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3146                                 __mmu_spte_walk(kvm, child, fn);
3147                         } else
3148                                 fn(kvm, sp, &sp->spt[i]);
3149                 }
3150         }
3151 }
3152
3153 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3154 {
3155         int i;
3156         struct kvm_mmu_page *sp;
3157
3158         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3159                 return;
3160         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3161                 hpa_t root = vcpu->arch.mmu.root_hpa;
3162                 sp = page_header(root);
3163                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3164                 return;
3165         }
3166         for (i = 0; i < 4; ++i) {
3167                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3168
3169                 if (root && VALID_PAGE(root)) {
3170                         root &= PT64_BASE_ADDR_MASK;
3171                         sp = page_header(root);
3172                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3173                 }
3174         }
3175         return;
3176 }
3177
3178 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3179                                 gva_t va, int level)
3180 {
3181         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3182         int i;
3183         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3184
3185         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3186                 u64 ent = pt[i];
3187
3188                 if (ent == shadow_trap_nonpresent_pte)
3189                         continue;
3190
3191                 va = canonicalize(va);
3192                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3193                         audit_mappings_page(vcpu, ent, va, level - 1);
3194                 else {
3195                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3196                         gfn_t gfn = gpa >> PAGE_SHIFT;
3197                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3198                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3199
3200                         if (is_error_pfn(pfn)) {
3201                                 kvm_release_pfn_clean(pfn);
3202                                 continue;
3203                         }
3204
3205                         if (is_shadow_present_pte(ent)
3206                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3207                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3208                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3209                                        audit_msg, vcpu->arch.mmu.root_level,
3210                                        va, gpa, hpa, ent,
3211                                        is_shadow_present_pte(ent));
3212                         else if (ent == shadow_notrap_nonpresent_pte
3213                                  && !is_error_hpa(hpa))
3214                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3215                                        " valid guest gva %lx\n", audit_msg, va);
3216                         kvm_release_pfn_clean(pfn);
3217
3218                 }
3219         }
3220 }
3221
3222 static void audit_mappings(struct kvm_vcpu *vcpu)
3223 {
3224         unsigned i;
3225
3226         if (vcpu->arch.mmu.root_level == 4)
3227                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3228         else
3229                 for (i = 0; i < 4; ++i)
3230                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3231                                 audit_mappings_page(vcpu,
3232                                                     vcpu->arch.mmu.pae_root[i],
3233                                                     i << 30,
3234                                                     2);
3235 }
3236
3237 static int count_rmaps(struct kvm_vcpu *vcpu)
3238 {
3239         int nmaps = 0;
3240         int i, j, k;
3241
3242         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3243                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3244                 struct kvm_rmap_desc *d;
3245
3246                 for (j = 0; j < m->npages; ++j) {
3247                         unsigned long *rmapp = &m->rmap[j];
3248
3249                         if (!*rmapp)
3250                                 continue;
3251                         if (!(*rmapp & 1)) {
3252                                 ++nmaps;
3253                                 continue;
3254                         }
3255                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3256                         while (d) {
3257                                 for (k = 0; k < RMAP_EXT; ++k)
3258                                         if (d->sptes[k])
3259                                                 ++nmaps;
3260                                         else
3261                                                 break;
3262                                 d = d->more;
3263                         }
3264                 }
3265         }
3266         return nmaps;
3267 }
3268
3269 void inspect_spte_has_rmap(struct kvm *kvm, struct kvm_mmu_page *sp, u64 *sptep)
3270 {
3271         unsigned long *rmapp;
3272         struct kvm_mmu_page *rev_sp;
3273         gfn_t gfn;
3274
3275         if (*sptep & PT_WRITABLE_MASK) {
3276                 rev_sp = page_header(__pa(sptep));
3277                 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3278
3279                 if (!gfn_to_memslot(kvm, gfn)) {
3280                         if (!printk_ratelimit())
3281                                 return;
3282                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3283                                          audit_msg, gfn);
3284                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3285                                         audit_msg, sptep - rev_sp->spt,
3286                                         rev_sp->gfn);
3287                         dump_stack();
3288                         return;
3289                 }
3290
3291                 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3292                                     is_large_pte(*sptep));
3293                 if (!*rmapp) {
3294                         if (!printk_ratelimit())
3295                                 return;
3296                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3297                                          audit_msg, *sptep);
3298                         dump_stack();
3299                 }
3300         }
3301
3302 }
3303
3304 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3305 {
3306         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3307 }
3308
3309 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3310 {
3311         struct kvm_mmu_page *sp;
3312         int i;
3313
3314         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3315                 u64 *pt = sp->spt;
3316
3317                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3318                         continue;
3319
3320                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3321                         u64 ent = pt[i];
3322
3323                         if (!(ent & PT_PRESENT_MASK))
3324                                 continue;
3325                         if (!(ent & PT_WRITABLE_MASK))
3326                                 continue;
3327                         inspect_spte_has_rmap(vcpu->kvm, sp, &pt[i]);
3328                 }
3329         }
3330         return;
3331 }
3332
3333 static void audit_rmap(struct kvm_vcpu *vcpu)
3334 {
3335         check_writable_mappings_rmap(vcpu);
3336         count_rmaps(vcpu);
3337 }
3338
3339 static void audit_write_protection(struct kvm_vcpu *vcpu)
3340 {
3341         struct kvm_mmu_page *sp;
3342         struct kvm_memory_slot *slot;
3343         unsigned long *rmapp;
3344         u64 *spte;
3345         gfn_t gfn;
3346
3347         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3348                 if (sp->role.direct)
3349                         continue;
3350                 if (sp->unsync)
3351                         continue;
3352
3353                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3354                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3355                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3356
3357                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3358                 while (spte) {
3359                         if (*spte & PT_WRITABLE_MASK)
3360                                 printk(KERN_ERR "%s: (%s) shadow page has "
3361                                 "writable mappings: gfn %lx role %x\n",
3362                                __func__, audit_msg, sp->gfn,
3363                                sp->role.word);
3364                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3365                 }
3366         }
3367 }
3368
3369 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3370 {
3371         int olddbg = dbg;
3372
3373         dbg = 0;
3374         audit_msg = msg;
3375         audit_rmap(vcpu);
3376         audit_write_protection(vcpu);
3377         if (strcmp("pre pte write", audit_msg) != 0)
3378                 audit_mappings(vcpu);
3379         audit_writable_sptes_have_rmaps(vcpu);
3380         dbg = olddbg;
3381 }
3382
3383 #endif