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_release_pfn_dirty(pfn);
 638         else
 639                 kvm_release_pfn_clean(pfn);
 640         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], sp->role.level);
 641         if (!*rmapp) {
 642                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 643                 BUG();
 644         } else if (!(*rmapp & 1)) {
 645                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 646                 if ((u64 *)*rmapp != spte) {
 647                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 648                                spte, *spte);
 649                         BUG();
 650                 }
 651                 *rmapp = 0;
 652         } else {
 653                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 654                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 655                 prev_desc = NULL;
 656                 while (desc) {
 657                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
 658                                 if (desc->sptes[i] == spte) {
 659                                         rmap_desc_remove_entry(rmapp,
 660                                                                desc, i,
 661                                                                prev_desc);
 662                                         return;
 663                                 }
 664                         prev_desc = desc;
 665                         desc = desc->more;
 666                 }
 667                 BUG();
 668         }
 669 }
 670
 671 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 672 {
 673         struct kvm_rmap_desc *desc;
 674         struct kvm_rmap_desc *prev_desc;
 675         u64 *prev_spte;
 676         int i;
 677
 678         if (!*rmapp)
 679                 return NULL;
 680         else if (!(*rmapp & 1)) {
 681                 if (!spte)
 682                         return (u64 *)*rmapp;
 683                 return NULL;
 684         }
 685         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 686         prev_desc = NULL;
 687         prev_spte = NULL;
 688         while (desc) {
 689                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
 690                         if (prev_spte == spte)
 691                                 return desc->sptes[i];
 692                         prev_spte = desc->sptes[i];
 693                 }
 694                 desc = desc->more;
 695         }
 696         return NULL;
 697 }
 698
 699 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 700 {
 701         unsigned long *rmapp;
 702         u64 *spte;
 703         int i, write_protected = 0;
 704
 705         gfn = unalias_gfn(kvm, gfn);
 706         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
 707
 708         spte = rmap_next(kvm, rmapp, NULL);
 709         while (spte) {
 710                 BUG_ON(!spte);
 711                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 712                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 713                 if (is_writeble_pte(*spte)) {
 714                         __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
 715                         write_protected = 1;
 716                 }
 717                 spte = rmap_next(kvm, rmapp, spte);
 718         }
 719         if (write_protected) {
 720                 pfn_t pfn;
 721
 722                 spte = rmap_next(kvm, rmapp, NULL);
 723                 pfn = spte_to_pfn(*spte);
 724                 kvm_set_pfn_dirty(pfn);
 725         }
 726
 727         /* check for huge page mappings */
 728         for (i = PT_DIRECTORY_LEVEL;
 729              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 730                 rmapp = gfn_to_rmap(kvm, gfn, i);
 731                 spte = rmap_next(kvm, rmapp, NULL);
 732                 while (spte) {
 733                         BUG_ON(!spte);
 734                         BUG_ON(!(*spte & PT_PRESENT_MASK));
 735                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 736                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 737                         if (is_writeble_pte(*spte)) {
 738                                 rmap_remove(kvm, spte);
 739                                 --kvm->stat.lpages;
 740                                 __set_spte(spte, shadow_trap_nonpresent_pte);
 741                                 spte = NULL;
 742                                 write_protected = 1;
 743                         }
 744                         spte = rmap_next(kvm, rmapp, spte);
 745                 }
 746         }
 747
 748         return write_protected;
 749 }
 750
 751 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 752 {
 753         u64 *spte;
 754         int need_tlb_flush = 0;
 755
 756         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 757                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 758                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 759                 rmap_remove(kvm, spte);
 760                 __set_spte(spte, shadow_trap_nonpresent_pte);
 761                 need_tlb_flush = 1;
 762         }
 763         return need_tlb_flush;
 764 }
 765
 766 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 767                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 768 {
 769         int i, j;
 770         int retval = 0;
 771
 772         /*
 773          * If mmap_sem isn't taken, we can look the memslots with only
 774          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 775          */
 776         for (i = 0; i < kvm->nmemslots; i++) {
 777                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 778                 unsigned long start = memslot->userspace_addr;
 779                 unsigned long end;
 780
 781                 /* mmu_lock protects userspace_addr */
 782                 if (!start)
 783                         continue;
 784
 785                 end = start + (memslot->npages << PAGE_SHIFT);
 786                 if (hva >= start && hva < end) {
 787                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 788
 789                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 790
 791                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
 792                                 int idx = gfn_offset;
 793                                 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
 794                                 retval |= handler(kvm,
 795                                         &memslot->lpage_info[j][idx].rmap_pde);
 796                         }
 797                 }
 798         }
 799
 800         return retval;
 801 }
 802
 803 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 804 {
 805         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 806 }
 807
 808 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 809 {
 810         u64 *spte;
 811         int young = 0;
 812
 813         /* always return old for EPT */
 814         if (!shadow_accessed_mask)
 815                 return 0;
 816
 817         spte = rmap_next(kvm, rmapp, NULL);
 818         while (spte) {
 819                 int _young;
 820                 u64 _spte = *spte;
 821                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 822                 _young = _spte & PT_ACCESSED_MASK;
 823                 if (_young) {
 824                         young = 1;
 825                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 826                 }
 827                 spte = rmap_next(kvm, rmapp, spte);
 828         }
 829         return young;
 830 }
 831
 832 #define RMAP_RECYCLE_THRESHOLD 1000
 833
 834 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 835 {
 836         unsigned long *rmapp;
 837         struct kvm_mmu_page *sp;
 838
 839         sp = page_header(__pa(spte));
 840
 841         gfn = unalias_gfn(vcpu->kvm, gfn);
 842         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 843
 844         kvm_unmap_rmapp(vcpu->kvm, rmapp);
 845         kvm_flush_remote_tlbs(vcpu->kvm);
 846 }
 847
 848 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 849 {
 850         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 851 }
 852
 853 #ifdef MMU_DEBUG
 854 static int is_empty_shadow_page(u64 *spt)
 855 {
 856         u64 *pos;
 857         u64 *end;
 858
 859         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 860                 if (is_shadow_present_pte(*pos)) {
 861                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 862                                pos, *pos);
 863                         return 0;
 864                 }
 865         return 1;
 866 }
 867 #endif
 868
 869 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 870 {
 871         ASSERT(is_empty_shadow_page(sp->spt));
 872         list_del(&sp->link);
 873         __free_page(virt_to_page(sp->spt));
 874         __free_page(virt_to_page(sp->gfns));
 875         kfree(sp);
 876         ++kvm->arch.n_free_mmu_pages;
 877 }
 878
 879 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 880 {
 881         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 882 }
 883
 884 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 885                                                u64 *parent_pte)
 886 {
 887         struct kvm_mmu_page *sp;
 888
 889         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 890         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 891         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 892         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 893         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 894         INIT_LIST_HEAD(&sp->oos_link);
 895         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 896         sp->multimapped = 0;
 897         sp->parent_pte = parent_pte;
 898         --vcpu->kvm->arch.n_free_mmu_pages;
 899         return sp;
 900 }
 901
 902 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 903                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 904 {
 905         struct kvm_pte_chain *pte_chain;
 906         struct hlist_node *node;
 907         int i;
 908
 909         if (!parent_pte)
 910                 return;
 911         if (!sp->multimapped) {
 912                 u64 *old = sp->parent_pte;
 913
 914                 if (!old) {
 915                         sp->parent_pte = parent_pte;
 916                         return;
 917                 }
 918                 sp->multimapped = 1;
 919                 pte_chain = mmu_alloc_pte_chain(vcpu);
 920                 INIT_HLIST_HEAD(&sp->parent_ptes);
 921                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 922                 pte_chain->parent_ptes[0] = old;
 923         }
 924         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 925                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 926                         continue;
 927                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 928                         if (!pte_chain->parent_ptes[i]) {
 929                                 pte_chain->parent_ptes[i] = parent_pte;
 930                                 return;
 931                         }
 932         }
 933         pte_chain = mmu_alloc_pte_chain(vcpu);
 934         BUG_ON(!pte_chain);
 935         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 936         pte_chain->parent_ptes[0] = parent_pte;
 937 }
 938
 939 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 940                                        u64 *parent_pte)
 941 {
 942         struct kvm_pte_chain *pte_chain;
 943         struct hlist_node *node;
 944         int i;
 945
 946         if (!sp->multimapped) {
 947                 BUG_ON(sp->parent_pte != parent_pte);
 948                 sp->parent_pte = NULL;
 949                 return;
 950         }
 951         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 952                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 953                         if (!pte_chain->parent_ptes[i])
 954                                 break;
 955                         if (pte_chain->parent_ptes[i] != parent_pte)
 956                                 continue;
 957                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 958                                 && pte_chain->parent_ptes[i + 1]) {
 959                                 pte_chain->parent_ptes[i]
 960                                         = pte_chain->parent_ptes[i + 1];
 961                                 ++i;
 962                         }
 963                         pte_chain->parent_ptes[i] = NULL;
 964                         if (i == 0) {
 965                                 hlist_del(&pte_chain->link);
 966                                 mmu_free_pte_chain(pte_chain);
 967                                 if (hlist_empty(&sp->parent_ptes)) {
 968                                         sp->multimapped = 0;
 969                                         sp->parent_pte = NULL;
 970                                 }
 971                         }
 972                         return;
 973                 }
 974         BUG();
 975 }
 976
 977
 978 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 979                             mmu_parent_walk_fn fn)
 980 {
 981         struct kvm_pte_chain *pte_chain;
 982         struct hlist_node *node;
 983         struct kvm_mmu_page *parent_sp;
 984         int i;
 985
 986         if (!sp->multimapped && sp->parent_pte) {
 987                 parent_sp = page_header(__pa(sp->parent_pte));
 988                 fn(vcpu, parent_sp);
 989                 mmu_parent_walk(vcpu, parent_sp, fn);
 990                 return;
 991         }
 992         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 993                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 994                         if (!pte_chain->parent_ptes[i])
 995                                 break;
 996                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 997                         fn(vcpu, parent_sp);
 998                         mmu_parent_walk(vcpu, parent_sp, fn);
 999                 }
1000 }
1001
1002 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1003 {
1004         unsigned int index;
1005         struct kvm_mmu_page *sp = page_header(__pa(spte));
1006
1007         index = spte - sp->spt;
1008         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1009                 sp->unsync_children++;
1010         WARN_ON(!sp->unsync_children);
1011 }
1012
1013 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1014 {
1015         struct kvm_pte_chain *pte_chain;
1016         struct hlist_node *node;
1017         int i;
1018
1019         if (!sp->parent_pte)
1020                 return;
1021
1022         if (!sp->multimapped) {
1023                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1024                 return;
1025         }
1026
1027         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1028                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1029                         if (!pte_chain->parent_ptes[i])
1030                                 break;
1031                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1032                 }
1033 }
1034
1035 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1036 {
1037         kvm_mmu_update_parents_unsync(sp);
1038         return 1;
1039 }
1040
1041 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
1042                                         struct kvm_mmu_page *sp)
1043 {
1044         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
1045         kvm_mmu_update_parents_unsync(sp);
1046 }
1047
1048 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1049                                     struct kvm_mmu_page *sp)
1050 {
1051         int i;
1052
1053         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1054                 sp->spt[i] = shadow_trap_nonpresent_pte;
1055 }
1056
1057 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1058                                struct kvm_mmu_page *sp)
1059 {
1060         return 1;
1061 }
1062
1063 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1064 {
1065 }
1066
1067 #define KVM_PAGE_ARRAY_NR 16
1068
1069 struct kvm_mmu_pages {
1070         struct mmu_page_and_offset {
1071                 struct kvm_mmu_page *sp;
1072                 unsigned int idx;
1073         } page[KVM_PAGE_ARRAY_NR];
1074         unsigned int nr;
1075 };
1076
1077 #define for_each_unsync_children(bitmap, idx)           \
1078         for (idx = find_first_bit(bitmap, 512);         \
1079              idx < 512;                                 \
1080              idx = find_next_bit(bitmap, 512, idx+1))
1081
1082 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1083                          int idx)
1084 {
1085         int i;
1086
1087         if (sp->unsync)
1088                 for (i=0; i < pvec->nr; i++)
1089                         if (pvec->page[i].sp == sp)
1090                                 return 0;
1091
1092         pvec->page[pvec->nr].sp = sp;
1093         pvec->page[pvec->nr].idx = idx;
1094         pvec->nr++;
1095         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1096 }
1097
1098 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1099                            struct kvm_mmu_pages *pvec)
1100 {
1101         int i, ret, nr_unsync_leaf = 0;
1102
1103         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1104                 u64 ent = sp->spt[i];
1105
1106                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1107                         struct kvm_mmu_page *child;
1108                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1109
1110                         if (child->unsync_children) {
1111                                 if (mmu_pages_add(pvec, child, i))
1112                                         return -ENOSPC;
1113
1114                                 ret = __mmu_unsync_walk(child, pvec);
1115                                 if (!ret)
1116                                         __clear_bit(i, sp->unsync_child_bitmap);
1117                                 else if (ret > 0)
1118                                         nr_unsync_leaf += ret;
1119                                 else
1120                                         return ret;
1121                         }
1122
1123                         if (child->unsync) {
1124                                 nr_unsync_leaf++;
1125                                 if (mmu_pages_add(pvec, child, i))
1126                                         return -ENOSPC;
1127                         }
1128                 }
1129         }
1130
1131         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1132                 sp->unsync_children = 0;
1133
1134         return nr_unsync_leaf;
1135 }
1136
1137 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1138                            struct kvm_mmu_pages *pvec)
1139 {
1140         if (!sp->unsync_children)
1141                 return 0;
1142
1143         mmu_pages_add(pvec, sp, 0);
1144         return __mmu_unsync_walk(sp, pvec);
1145 }
1146
1147 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1148 {
1149         unsigned index;
1150         struct hlist_head *bucket;
1151         struct kvm_mmu_page *sp;
1152         struct hlist_node *node;
1153
1154         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1155         index = kvm_page_table_hashfn(gfn);
1156         bucket = &kvm->arch.mmu_page_hash[index];
1157         hlist_for_each_entry(sp, node, bucket, hash_link)
1158                 if (sp->gfn == gfn && !sp->role.direct
1159                     && !sp->role.invalid) {
1160                         pgprintk("%s: found role %x\n",
1161                                  __func__, sp->role.word);
1162                         return sp;
1163                 }
1164         return NULL;
1165 }
1166
1167 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1168 {
1169         WARN_ON(!sp->unsync);
1170         sp->unsync = 0;
1171         --kvm->stat.mmu_unsync;
1172 }
1173
1174 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1175
1176 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1177 {
1178         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1179                 kvm_mmu_zap_page(vcpu->kvm, sp);
1180                 return 1;
1181         }
1182
1183         trace_kvm_mmu_sync_page(sp);
1184         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1185                 kvm_flush_remote_tlbs(vcpu->kvm);
1186         kvm_unlink_unsync_page(vcpu->kvm, sp);
1187         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1188                 kvm_mmu_zap_page(vcpu->kvm, sp);
1189                 return 1;
1190         }
1191
1192         kvm_mmu_flush_tlb(vcpu);
1193         return 0;
1194 }
1195
1196 struct mmu_page_path {
1197         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1198         unsigned int idx[PT64_ROOT_LEVEL-1];
1199 };
1200
1201 #define for_each_sp(pvec, sp, parents, i)                       \
1202                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1203                         sp = pvec.page[i].sp;                   \
1204                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1205                         i = mmu_pages_next(&pvec, &parents, i))
1206
1207 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1208                           struct mmu_page_path *parents,
1209                           int i)
1210 {
1211         int n;
1212
1213         for (n = i+1; n < pvec->nr; n++) {
1214                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1215
1216                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1217                         parents->idx[0] = pvec->page[n].idx;
1218                         return n;
1219                 }
1220
1221                 parents->parent[sp->role.level-2] = sp;
1222                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1223         }
1224
1225         return n;
1226 }
1227
1228 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1229 {
1230         struct kvm_mmu_page *sp;
1231         unsigned int level = 0;
1232
1233         do {
1234                 unsigned int idx = parents->idx[level];
1235
1236                 sp = parents->parent[level];
1237                 if (!sp)
1238                         return;
1239
1240                 --sp->unsync_children;
1241                 WARN_ON((int)sp->unsync_children < 0);
1242                 __clear_bit(idx, sp->unsync_child_bitmap);
1243                 level++;
1244         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1245 }
1246
1247 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1248                                struct mmu_page_path *parents,
1249                                struct kvm_mmu_pages *pvec)
1250 {
1251         parents->parent[parent->role.level-1] = NULL;
1252         pvec->nr = 0;
1253 }
1254
1255 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1256                               struct kvm_mmu_page *parent)
1257 {
1258         int i;
1259         struct kvm_mmu_page *sp;
1260         struct mmu_page_path parents;
1261         struct kvm_mmu_pages pages;
1262
1263         kvm_mmu_pages_init(parent, &parents, &pages);
1264         while (mmu_unsync_walk(parent, &pages)) {
1265                 int protected = 0;
1266
1267                 for_each_sp(pages, sp, parents, i)
1268                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1269
1270                 if (protected)
1271                         kvm_flush_remote_tlbs(vcpu->kvm);
1272
1273                 for_each_sp(pages, sp, parents, i) {
1274                         kvm_sync_page(vcpu, sp);
1275                         mmu_pages_clear_parents(&parents);
1276                 }
1277                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1278                 kvm_mmu_pages_init(parent, &parents, &pages);
1279         }
1280 }
1281
1282 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1283                                              gfn_t gfn,
1284                                              gva_t gaddr,
1285                                              unsigned level,
1286                                              int direct,
1287                                              unsigned access,
1288                                              u64 *parent_pte)
1289 {
1290         union kvm_mmu_page_role role;
1291         unsigned index;
1292         unsigned quadrant;
1293         struct hlist_head *bucket;
1294         struct kvm_mmu_page *sp;
1295         struct hlist_node *node, *tmp;
1296
1297         role = vcpu->arch.mmu.base_role;
1298         role.level = level;
1299         role.direct = direct;
1300         role.access = access;
1301         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1302                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1303                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1304                 role.quadrant = quadrant;
1305         }
1306         index = kvm_page_table_hashfn(gfn);
1307         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1308         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1309                 if (sp->gfn == gfn) {
1310                         if (sp->unsync)
1311                                 if (kvm_sync_page(vcpu, sp))
1312                                         continue;
1313
1314                         if (sp->role.word != role.word)
1315                                 continue;
1316
1317                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1318                         if (sp->unsync_children) {
1319                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1320                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1321                         }
1322                         trace_kvm_mmu_get_page(sp, false);
1323                         return sp;
1324                 }
1325         ++vcpu->kvm->stat.mmu_cache_miss;
1326         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1327         if (!sp)
1328                 return sp;
1329         sp->gfn = gfn;
1330         sp->role = role;
1331         hlist_add_head(&sp->hash_link, bucket);
1332         if (!direct) {
1333                 if (rmap_write_protect(vcpu->kvm, gfn))
1334                         kvm_flush_remote_tlbs(vcpu->kvm);
1335                 account_shadowed(vcpu->kvm, gfn);
1336         }
1337         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1338                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1339         else
1340                 nonpaging_prefetch_page(vcpu, sp);
1341         trace_kvm_mmu_get_page(sp, true);
1342         return sp;
1343 }
1344
1345 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1346                              struct kvm_vcpu *vcpu, u64 addr)
1347 {
1348         iterator->addr = addr;
1349         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1350         iterator->level = vcpu->arch.mmu.shadow_root_level;
1351         if (iterator->level == PT32E_ROOT_LEVEL) {
1352                 iterator->shadow_addr
1353                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1354                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1355                 --iterator->level;
1356                 if (!iterator->shadow_addr)
1357                         iterator->level = 0;
1358         }
1359 }
1360
1361 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1362 {
1363         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1364                 return false;
1365
1366         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1367                 if (is_large_pte(*iterator->sptep))
1368                         return false;
1369
1370         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1371         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1372         return true;
1373 }
1374
1375 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1376 {
1377         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1378         --iterator->level;
1379 }
1380
1381 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1382                                          struct kvm_mmu_page *sp)
1383 {
1384         unsigned i;
1385         u64 *pt;
1386         u64 ent;
1387
1388         pt = sp->spt;
1389
1390         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1391                 ent = pt[i];
1392
1393                 if (is_shadow_present_pte(ent)) {
1394                         if (!is_last_spte(ent, sp->role.level)) {
1395                                 ent &= PT64_BASE_ADDR_MASK;
1396                                 mmu_page_remove_parent_pte(page_header(ent),
1397                                                            &pt[i]);
1398                         } else {
1399                                 if (is_large_pte(ent))
1400                                         --kvm->stat.lpages;
1401                                 rmap_remove(kvm, &pt[i]);
1402                         }
1403                 }
1404                 pt[i] = shadow_trap_nonpresent_pte;
1405         }
1406 }
1407
1408 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1409 {
1410         mmu_page_remove_parent_pte(sp, parent_pte);
1411 }
1412
1413 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1414 {
1415         int i;
1416         struct kvm_vcpu *vcpu;
1417
1418         kvm_for_each_vcpu(i, vcpu, kvm)
1419                 vcpu->arch.last_pte_updated = NULL;
1420 }
1421
1422 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1423 {
1424         u64 *parent_pte;
1425
1426         while (sp->multimapped || sp->parent_pte) {
1427                 if (!sp->multimapped)
1428                         parent_pte = sp->parent_pte;
1429                 else {
1430                         struct kvm_pte_chain *chain;
1431
1432                         chain = container_of(sp->parent_ptes.first,
1433                                              struct kvm_pte_chain, link);
1434                         parent_pte = chain->parent_ptes[0];
1435                 }
1436                 BUG_ON(!parent_pte);
1437                 kvm_mmu_put_page(sp, parent_pte);
1438                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1439         }
1440 }
1441
1442 static int mmu_zap_unsync_children(struct kvm *kvm,
1443                                    struct kvm_mmu_page *parent)
1444 {
1445         int i, zapped = 0;
1446         struct mmu_page_path parents;
1447         struct kvm_mmu_pages pages;
1448
1449         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1450                 return 0;
1451
1452         kvm_mmu_pages_init(parent, &parents, &pages);
1453         while (mmu_unsync_walk(parent, &pages)) {
1454                 struct kvm_mmu_page *sp;
1455
1456                 for_each_sp(pages, sp, parents, i) {
1457                         kvm_mmu_zap_page(kvm, sp);
1458                         mmu_pages_clear_parents(&parents);
1459                 }
1460                 zapped += pages.nr;
1461                 kvm_mmu_pages_init(parent, &parents, &pages);
1462         }
1463
1464         return zapped;
1465 }
1466
1467 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1468 {
1469         int ret;
1470
1471         trace_kvm_mmu_zap_page(sp);
1472         ++kvm->stat.mmu_shadow_zapped;
1473         ret = mmu_zap_unsync_children(kvm, sp);
1474         kvm_mmu_page_unlink_children(kvm, sp);
1475         kvm_mmu_unlink_parents(kvm, sp);
1476         kvm_flush_remote_tlbs(kvm);
1477         if (!sp->role.invalid && !sp->role.direct)
1478                 unaccount_shadowed(kvm, sp->gfn);
1479         if (sp->unsync)
1480                 kvm_unlink_unsync_page(kvm, sp);
1481         if (!sp->root_count) {
1482                 hlist_del(&sp->hash_link);
1483                 kvm_mmu_free_page(kvm, sp);
1484         } else {
1485                 sp->role.invalid = 1;
1486                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1487                 kvm_reload_remote_mmus(kvm);
1488         }
1489         kvm_mmu_reset_last_pte_updated(kvm);
1490         return ret;
1491 }
1492
1493 /*
1494  * Changing the number of mmu pages allocated to the vm
1495  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1496  */
1497 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1498 {
1499         int used_pages;
1500
1501         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1502         used_pages = max(0, used_pages);
1503
1504         /*
1505          * If we set the number of mmu pages to be smaller be than the
1506          * number of actived pages , we must to free some mmu pages before we
1507          * change the value
1508          */
1509
1510         if (used_pages > kvm_nr_mmu_pages) {
1511                 while (used_pages > kvm_nr_mmu_pages) {
1512                         struct kvm_mmu_page *page;
1513
1514                         page = container_of(kvm->arch.active_mmu_pages.prev,
1515                                             struct kvm_mmu_page, link);
1516                         kvm_mmu_zap_page(kvm, page);
1517                         used_pages--;
1518                 }
1519                 kvm->arch.n_free_mmu_pages = 0;
1520         }
1521         else
1522                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1523                                          - kvm->arch.n_alloc_mmu_pages;
1524
1525         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1526 }
1527
1528 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1529 {
1530         unsigned index;
1531         struct hlist_head *bucket;
1532         struct kvm_mmu_page *sp;
1533         struct hlist_node *node, *n;
1534         int r;
1535
1536         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1537         r = 0;
1538         index = kvm_page_table_hashfn(gfn);
1539         bucket = &kvm->arch.mmu_page_hash[index];
1540         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1541                 if (sp->gfn == gfn && !sp->role.direct) {
1542                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1543                                  sp->role.word);
1544                         r = 1;
1545                         if (kvm_mmu_zap_page(kvm, sp))
1546                                 n = bucket->first;
1547                 }
1548         return r;
1549 }
1550
1551 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1552 {
1553         unsigned index;
1554         struct hlist_head *bucket;
1555         struct kvm_mmu_page *sp;
1556         struct hlist_node *node, *nn;
1557
1558         index = kvm_page_table_hashfn(gfn);
1559         bucket = &kvm->arch.mmu_page_hash[index];
1560         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1561                 if (sp->gfn == gfn && !sp->role.direct
1562                     && !sp->role.invalid) {
1563                         pgprintk("%s: zap %lx %x\n",
1564                                  __func__, gfn, sp->role.word);
1565                         kvm_mmu_zap_page(kvm, sp);
1566                 }
1567         }
1568 }
1569
1570 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1571 {
1572         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1573         struct kvm_mmu_page *sp = page_header(__pa(pte));
1574
1575         __set_bit(slot, sp->slot_bitmap);
1576 }
1577
1578 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1579 {
1580         int i;
1581         u64 *pt = sp->spt;
1582
1583         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1584                 return;
1585
1586         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1587                 if (pt[i] == shadow_notrap_nonpresent_pte)
1588                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1589         }
1590 }
1591
1592 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1593 {
1594         struct page *page;
1595
1596         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1597
1598         if (gpa == UNMAPPED_GVA)
1599                 return NULL;
1600
1601         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1602
1603         return page;
1604 }
1605
1606 /*
1607  * The function is based on mtrr_type_lookup() in
1608  * arch/x86/kernel/cpu/mtrr/generic.c
1609  */
1610 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1611                          u64 start, u64 end)
1612 {
1613         int i;
1614         u64 base, mask;
1615         u8 prev_match, curr_match;
1616         int num_var_ranges = KVM_NR_VAR_MTRR;
1617
1618         if (!mtrr_state->enabled)
1619                 return 0xFF;
1620
1621         /* Make end inclusive end, instead of exclusive */
1622         end--;
1623
1624         /* Look in fixed ranges. Just return the type as per start */
1625         if (mtrr_state->have_fixed && (start < 0x100000)) {
1626                 int idx;
1627
1628                 if (start < 0x80000) {
1629                         idx = 0;
1630                         idx += (start >> 16);
1631                         return mtrr_state->fixed_ranges[idx];
1632                 } else if (start < 0xC0000) {
1633                         idx = 1 * 8;
1634                         idx += ((start - 0x80000) >> 14);
1635                         return mtrr_state->fixed_ranges[idx];
1636                 } else if (start < 0x1000000) {
1637                         idx = 3 * 8;
1638                         idx += ((start - 0xC0000) >> 12);
1639                         return mtrr_state->fixed_ranges[idx];
1640                 }
1641         }
1642
1643         /*
1644          * Look in variable ranges
1645          * Look of multiple ranges matching this address and pick type
1646          * as per MTRR precedence
1647          */
1648         if (!(mtrr_state->enabled & 2))
1649                 return mtrr_state->def_type;
1650
1651         prev_match = 0xFF;
1652         for (i = 0; i < num_var_ranges; ++i) {
1653                 unsigned short start_state, end_state;
1654
1655                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1656                         continue;
1657
1658                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1659                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1660                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1661                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1662
1663                 start_state = ((start & mask) == (base & mask));
1664                 end_state = ((end & mask) == (base & mask));
1665                 if (start_state != end_state)
1666                         return 0xFE;
1667
1668                 if ((start & mask) != (base & mask))
1669                         continue;
1670
1671                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1672                 if (prev_match == 0xFF) {
1673                         prev_match = curr_match;
1674                         continue;
1675                 }
1676
1677                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1678                     curr_match == MTRR_TYPE_UNCACHABLE)
1679                         return MTRR_TYPE_UNCACHABLE;
1680
1681                 if ((prev_match == MTRR_TYPE_WRBACK &&
1682                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1683                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1684                      curr_match == MTRR_TYPE_WRBACK)) {
1685                         prev_match = MTRR_TYPE_WRTHROUGH;
1686                         curr_match = MTRR_TYPE_WRTHROUGH;
1687                 }
1688
1689                 if (prev_match != curr_match)
1690                         return MTRR_TYPE_UNCACHABLE;
1691         }
1692
1693         if (prev_match != 0xFF)
1694                 return prev_match;
1695
1696         return mtrr_state->def_type;
1697 }
1698
1699 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1700 {
1701         u8 mtrr;
1702
1703         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1704                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1705         if (mtrr == 0xfe || mtrr == 0xff)
1706                 mtrr = MTRR_TYPE_WRBACK;
1707         return mtrr;
1708 }
1709 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1710
1711 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1712 {
1713         unsigned index;
1714         struct hlist_head *bucket;
1715         struct kvm_mmu_page *s;
1716         struct hlist_node *node, *n;
1717
1718         trace_kvm_mmu_unsync_page(sp);
1719         index = kvm_page_table_hashfn(sp->gfn);
1720         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1721         /* don't unsync if pagetable is shadowed with multiple roles */
1722         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1723                 if (s->gfn != sp->gfn || s->role.direct)
1724                         continue;
1725                 if (s->role.word != sp->role.word)
1726                         return 1;
1727         }
1728         ++vcpu->kvm->stat.mmu_unsync;
1729         sp->unsync = 1;
1730
1731         kvm_mmu_mark_parents_unsync(vcpu, sp);
1732
1733         mmu_convert_notrap(sp);
1734         return 0;
1735 }
1736
1737 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1738                                   bool can_unsync)
1739 {
1740         struct kvm_mmu_page *shadow;
1741
1742         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1743         if (shadow) {
1744                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1745                         return 1;
1746                 if (shadow->unsync)
1747                         return 0;
1748                 if (can_unsync && oos_shadow)
1749                         return kvm_unsync_page(vcpu, shadow);
1750                 return 1;
1751         }
1752         return 0;
1753 }
1754
1755 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1756                     unsigned pte_access, int user_fault,
1757                     int write_fault, int dirty, int level,
1758                     gfn_t gfn, pfn_t pfn, bool speculative,
1759                     bool can_unsync)
1760 {
1761         u64 spte;
1762         int ret = 0;
1763
1764         /*
1765          * We don't set the accessed bit, since we sometimes want to see
1766          * whether the guest actually used the pte (in order to detect
1767          * demand paging).
1768          */
1769         spte = shadow_base_present_pte | shadow_dirty_mask;
1770         if (!speculative)
1771                 spte |= shadow_accessed_mask;
1772         if (!dirty)
1773                 pte_access &= ~ACC_WRITE_MASK;
1774         if (pte_access & ACC_EXEC_MASK)
1775                 spte |= shadow_x_mask;
1776         else
1777                 spte |= shadow_nx_mask;
1778         if (pte_access & ACC_USER_MASK)
1779                 spte |= shadow_user_mask;
1780         if (level > PT_PAGE_TABLE_LEVEL)
1781                 spte |= PT_PAGE_SIZE_MASK;
1782         if (tdp_enabled)
1783                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1784                         kvm_is_mmio_pfn(pfn));
1785
1786         spte |= (u64)pfn << PAGE_SHIFT;
1787
1788         if ((pte_access & ACC_WRITE_MASK)
1789             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1790
1791                 if (level > PT_PAGE_TABLE_LEVEL &&
1792                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1793                         ret = 1;
1794                         spte = shadow_trap_nonpresent_pte;
1795                         goto set_pte;
1796                 }
1797
1798                 spte |= PT_WRITABLE_MASK;
1799
1800                 /*
1801                  * Optimization: for pte sync, if spte was writable the hash
1802                  * lookup is unnecessary (and expensive). Write protection
1803                  * is responsibility of mmu_get_page / kvm_sync_page.
1804                  * Same reasoning can be applied to dirty page accounting.
1805                  */
1806                 if (!can_unsync && is_writeble_pte(*sptep))
1807                         goto set_pte;
1808
1809                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1810                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1811                                  __func__, gfn);
1812                         ret = 1;
1813                         pte_access &= ~ACC_WRITE_MASK;
1814                         if (is_writeble_pte(spte))
1815                                 spte &= ~PT_WRITABLE_MASK;
1816                 }
1817         }
1818
1819         if (pte_access & ACC_WRITE_MASK)
1820                 mark_page_dirty(vcpu->kvm, gfn);
1821
1822 set_pte:
1823         __set_spte(sptep, spte);
1824         return ret;
1825 }
1826
1827 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1828                          unsigned pt_access, unsigned pte_access,
1829                          int user_fault, int write_fault, int dirty,
1830                          int *ptwrite, int level, gfn_t gfn,
1831                          pfn_t pfn, bool speculative)
1832 {
1833         int was_rmapped = 0;
1834         int was_writeble = is_writeble_pte(*sptep);
1835         int rmap_count;
1836
1837         pgprintk("%s: spte %llx access %x write_fault %d"
1838                  " user_fault %d gfn %lx\n",
1839                  __func__, *sptep, pt_access,
1840                  write_fault, user_fault, gfn);
1841
1842         if (is_rmap_spte(*sptep)) {
1843                 /*
1844                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1845                  * the parent of the now unreachable PTE.
1846                  */
1847                 if (level > PT_PAGE_TABLE_LEVEL &&
1848                     !is_large_pte(*sptep)) {
1849                         struct kvm_mmu_page *child;
1850                         u64 pte = *sptep;
1851
1852                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1853                         mmu_page_remove_parent_pte(child, sptep);
1854                 } else if (pfn != spte_to_pfn(*sptep)) {
1855                         pgprintk("hfn old %lx new %lx\n",
1856                                  spte_to_pfn(*sptep), pfn);
1857                         rmap_remove(vcpu->kvm, sptep);
1858                 } else
1859                         was_rmapped = 1;
1860         }
1861
1862         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1863                       dirty, level, gfn, pfn, speculative, true)) {
1864                 if (write_fault)
1865                         *ptwrite = 1;
1866                 kvm_x86_ops->tlb_flush(vcpu);
1867         }
1868
1869         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1870         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1871                  is_large_pte(*sptep)? "2MB" : "4kB",
1872                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1873                  *sptep, sptep);
1874         if (!was_rmapped && is_large_pte(*sptep))
1875                 ++vcpu->kvm->stat.lpages;
1876
1877         page_header_update_slot(vcpu->kvm, sptep, gfn);
1878         if (!was_rmapped) {
1879                 rmap_count = rmap_add(vcpu, sptep, gfn);
1880                 if (!is_rmap_spte(*sptep))
1881                         kvm_release_pfn_clean(pfn);
1882                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1883                         rmap_recycle(vcpu, sptep, gfn);
1884         } else {
1885                 if (was_writeble)
1886                         kvm_release_pfn_dirty(pfn);
1887                 else
1888                         kvm_release_pfn_clean(pfn);
1889         }
1890         if (speculative) {
1891                 vcpu->arch.last_pte_updated = sptep;
1892                 vcpu->arch.last_pte_gfn = gfn;
1893         }
1894 }
1895
1896 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1897 {
1898 }
1899
1900 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1901                         int level, gfn_t gfn, pfn_t pfn)
1902 {
1903         struct kvm_shadow_walk_iterator iterator;
1904         struct kvm_mmu_page *sp;
1905         int pt_write = 0;
1906         gfn_t pseudo_gfn;
1907
1908         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1909                 if (iterator.level == level) {
1910                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1911                                      0, write, 1, &pt_write,
1912                                      level, gfn, pfn, false);
1913                         ++vcpu->stat.pf_fixed;
1914                         break;
1915                 }
1916
1917                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1918                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1919                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1920                                               iterator.level - 1,
1921                                               1, ACC_ALL, iterator.sptep);
1922                         if (!sp) {
1923                                 pgprintk("nonpaging_map: ENOMEM\n");
1924                                 kvm_release_pfn_clean(pfn);
1925                                 return -ENOMEM;
1926                         }
1927
1928                         __set_spte(iterator.sptep,
1929                                    __pa(sp->spt)
1930                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1931                                    | shadow_user_mask | shadow_x_mask);
1932                 }
1933         }
1934         return pt_write;
1935 }
1936
1937 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1938 {
1939         int r;
1940         int level;
1941         pfn_t pfn;
1942         unsigned long mmu_seq;
1943
1944         level = mapping_level(vcpu, gfn);
1945
1946         /*
1947          * This path builds a PAE pagetable - so we can map 2mb pages at
1948          * maximum. Therefore check if the level is larger than that.
1949          */
1950         if (level > PT_DIRECTORY_LEVEL)
1951                 level = PT_DIRECTORY_LEVEL;
1952
1953         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
1954
1955         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1956         smp_rmb();
1957         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1958
1959         /* mmio */
1960         if (is_error_pfn(pfn)) {
1961                 kvm_release_pfn_clean(pfn);
1962                 return 1;
1963         }
1964
1965         spin_lock(&vcpu->kvm->mmu_lock);
1966         if (mmu_notifier_retry(vcpu, mmu_seq))
1967                 goto out_unlock;
1968         kvm_mmu_free_some_pages(vcpu);
1969         r = __direct_map(vcpu, v, write, level, gfn, pfn);
1970         spin_unlock(&vcpu->kvm->mmu_lock);
1971
1972
1973         return r;
1974
1975 out_unlock:
1976         spin_unlock(&vcpu->kvm->mmu_lock);
1977         kvm_release_pfn_clean(pfn);
1978         return 0;
1979 }
1980
1981
1982 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1983 {
1984         int i;
1985         struct kvm_mmu_page *sp;
1986
1987         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1988                 return;
1989         spin_lock(&vcpu->kvm->mmu_lock);
1990         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1991                 hpa_t root = vcpu->arch.mmu.root_hpa;
1992
1993                 sp = page_header(root);
1994                 --sp->root_count;
1995                 if (!sp->root_count && sp->role.invalid)
1996                         kvm_mmu_zap_page(vcpu->kvm, sp);
1997                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1998                 spin_unlock(&vcpu->kvm->mmu_lock);
1999                 return;
2000         }
2001         for (i = 0; i < 4; ++i) {
2002                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2003
2004                 if (root) {
2005                         root &= PT64_BASE_ADDR_MASK;
2006                         sp = page_header(root);
2007                         --sp->root_count;
2008                         if (!sp->root_count && sp->role.invalid)
2009                                 kvm_mmu_zap_page(vcpu->kvm, sp);
2010                 }
2011                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2012         }
2013         spin_unlock(&vcpu->kvm->mmu_lock);
2014         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2015 }
2016
2017 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2018 {
2019         int ret = 0;
2020
2021         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2022                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2023                 ret = 1;
2024         }
2025
2026         return ret;
2027 }
2028
2029 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2030 {
2031         int i;
2032         gfn_t root_gfn;
2033         struct kvm_mmu_page *sp;
2034         int direct = 0;
2035         u64 pdptr;
2036
2037         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2038
2039         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2040                 hpa_t root = vcpu->arch.mmu.root_hpa;
2041
2042                 ASSERT(!VALID_PAGE(root));
2043                 if (tdp_enabled)
2044                         direct = 1;
2045                 if (mmu_check_root(vcpu, root_gfn))
2046                         return 1;
2047                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2048                                       PT64_ROOT_LEVEL, direct,
2049                                       ACC_ALL, NULL);
2050                 root = __pa(sp->spt);
2051                 ++sp->root_count;
2052                 vcpu->arch.mmu.root_hpa = root;
2053                 return 0;
2054         }
2055         direct = !is_paging(vcpu);
2056         if (tdp_enabled)
2057                 direct = 1;
2058         for (i = 0; i < 4; ++i) {
2059                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2060
2061                 ASSERT(!VALID_PAGE(root));
2062                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2063                         pdptr = kvm_pdptr_read(vcpu, i);
2064                         if (!is_present_gpte(pdptr)) {
2065                                 vcpu->arch.mmu.pae_root[i] = 0;
2066                                 continue;
2067                         }
2068                         root_gfn = pdptr >> PAGE_SHIFT;
2069                 } else if (vcpu->arch.mmu.root_level == 0)
2070                         root_gfn = 0;
2071                 if (mmu_check_root(vcpu, root_gfn))
2072                         return 1;
2073                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2074                                       PT32_ROOT_LEVEL, direct,
2075                                       ACC_ALL, NULL);
2076                 root = __pa(sp->spt);
2077                 ++sp->root_count;
2078                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2079         }
2080         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2081         return 0;
2082 }
2083
2084 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2085 {
2086         int i;
2087         struct kvm_mmu_page *sp;
2088
2089         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2090                 return;
2091         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2092                 hpa_t root = vcpu->arch.mmu.root_hpa;
2093                 sp = page_header(root);
2094                 mmu_sync_children(vcpu, sp);
2095                 return;
2096         }
2097         for (i = 0; i < 4; ++i) {
2098                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2099
2100                 if (root && VALID_PAGE(root)) {
2101                         root &= PT64_BASE_ADDR_MASK;
2102                         sp = page_header(root);
2103                         mmu_sync_children(vcpu, sp);
2104                 }
2105         }
2106 }
2107
2108 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2109 {
2110         spin_lock(&vcpu->kvm->mmu_lock);
2111         mmu_sync_roots(vcpu);
2112         spin_unlock(&vcpu->kvm->mmu_lock);
2113 }
2114
2115 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2116 {
2117         return vaddr;
2118 }
2119
2120 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2121                                 u32 error_code)
2122 {
2123         gfn_t gfn;
2124         int r;
2125
2126         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2127         r = mmu_topup_memory_caches(vcpu);
2128         if (r)
2129                 return r;
2130
2131         ASSERT(vcpu);
2132         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2133
2134         gfn = gva >> PAGE_SHIFT;
2135
2136         return nonpaging_map(vcpu, gva & PAGE_MASK,
2137                              error_code & PFERR_WRITE_MASK, gfn);
2138 }
2139
2140 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2141                                 u32 error_code)
2142 {
2143         pfn_t pfn;
2144         int r;
2145         int level;
2146         gfn_t gfn = gpa >> PAGE_SHIFT;
2147         unsigned long mmu_seq;
2148
2149         ASSERT(vcpu);
2150         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2151
2152         r = mmu_topup_memory_caches(vcpu);
2153         if (r)
2154                 return r;
2155
2156         level = mapping_level(vcpu, gfn);
2157
2158         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2159
2160         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2161         smp_rmb();
2162         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2163         if (is_error_pfn(pfn)) {
2164                 kvm_release_pfn_clean(pfn);
2165                 return 1;
2166         }
2167         spin_lock(&vcpu->kvm->mmu_lock);
2168         if (mmu_notifier_retry(vcpu, mmu_seq))
2169                 goto out_unlock;
2170         kvm_mmu_free_some_pages(vcpu);
2171         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2172                          level, gfn, pfn);
2173         spin_unlock(&vcpu->kvm->mmu_lock);
2174
2175         return r;
2176
2177 out_unlock:
2178         spin_unlock(&vcpu->kvm->mmu_lock);
2179         kvm_release_pfn_clean(pfn);
2180         return 0;
2181 }
2182
2183 static void nonpaging_free(struct kvm_vcpu *vcpu)
2184 {
2185         mmu_free_roots(vcpu);
2186 }
2187
2188 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2189 {
2190         struct kvm_mmu *context = &vcpu->arch.mmu;
2191
2192         context->new_cr3 = nonpaging_new_cr3;
2193         context->page_fault = nonpaging_page_fault;
2194         context->gva_to_gpa = nonpaging_gva_to_gpa;
2195         context->free = nonpaging_free;
2196         context->prefetch_page = nonpaging_prefetch_page;
2197         context->sync_page = nonpaging_sync_page;
2198         context->invlpg = nonpaging_invlpg;
2199         context->root_level = 0;
2200         context->shadow_root_level = PT32E_ROOT_LEVEL;
2201         context->root_hpa = INVALID_PAGE;
2202         return 0;
2203 }
2204
2205 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2206 {
2207         ++vcpu->stat.tlb_flush;
2208         kvm_x86_ops->tlb_flush(vcpu);
2209 }
2210
2211 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2212 {
2213         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2214         mmu_free_roots(vcpu);
2215 }
2216
2217 static void inject_page_fault(struct kvm_vcpu *vcpu,
2218                               u64 addr,
2219                               u32 err_code)
2220 {
2221         kvm_inject_page_fault(vcpu, addr, err_code);
2222 }
2223
2224 static void paging_free(struct kvm_vcpu *vcpu)
2225 {
2226         nonpaging_free(vcpu);
2227 }
2228
2229 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2230 {
2231         int bit7;
2232
2233         bit7 = (gpte >> 7) & 1;
2234         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2235 }
2236
2237 #define PTTYPE 64
2238 #include "paging_tmpl.h"
2239 #undef PTTYPE
2240
2241 #define PTTYPE 32
2242 #include "paging_tmpl.h"
2243 #undef PTTYPE
2244
2245 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2246 {
2247         struct kvm_mmu *context = &vcpu->arch.mmu;
2248         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2249         u64 exb_bit_rsvd = 0;
2250
2251         if (!is_nx(vcpu))
2252                 exb_bit_rsvd = rsvd_bits(63, 63);
2253         switch (level) {
2254         case PT32_ROOT_LEVEL:
2255                 /* no rsvd bits for 2 level 4K page table entries */
2256                 context->rsvd_bits_mask[0][1] = 0;
2257                 context->rsvd_bits_mask[0][0] = 0;
2258                 if (is_cpuid_PSE36())
2259                         /* 36bits PSE 4MB page */
2260                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2261                 else
2262                         /* 32 bits PSE 4MB page */
2263                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2264                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2265                 break;
2266         case PT32E_ROOT_LEVEL:
2267                 context->rsvd_bits_mask[0][2] =
2268                         rsvd_bits(maxphyaddr, 63) |
2269                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2270                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2271                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2272                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2273                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2274                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2275                         rsvd_bits(maxphyaddr, 62) |
2276                         rsvd_bits(13, 20);              /* large page */
2277                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2278                 break;
2279         case PT64_ROOT_LEVEL:
2280                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2281                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2282                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2283                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2284                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2285                         rsvd_bits(maxphyaddr, 51);
2286                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2287                         rsvd_bits(maxphyaddr, 51);
2288                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2289                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2290                         rsvd_bits(maxphyaddr, 51) |
2291                         rsvd_bits(13, 29);
2292                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2293                         rsvd_bits(maxphyaddr, 51) |
2294                         rsvd_bits(13, 20);              /* large page */
2295                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2296                 break;
2297         }
2298 }
2299
2300 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2301 {
2302         struct kvm_mmu *context = &vcpu->arch.mmu;
2303
2304         ASSERT(is_pae(vcpu));
2305         context->new_cr3 = paging_new_cr3;
2306         context->page_fault = paging64_page_fault;
2307         context->gva_to_gpa = paging64_gva_to_gpa;
2308         context->prefetch_page = paging64_prefetch_page;
2309         context->sync_page = paging64_sync_page;
2310         context->invlpg = paging64_invlpg;
2311         context->free = paging_free;
2312         context->root_level = level;
2313         context->shadow_root_level = level;
2314         context->root_hpa = INVALID_PAGE;
2315         return 0;
2316 }
2317
2318 static int paging64_init_context(struct kvm_vcpu *vcpu)
2319 {
2320         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2321         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2322 }
2323
2324 static int paging32_init_context(struct kvm_vcpu *vcpu)
2325 {
2326         struct kvm_mmu *context = &vcpu->arch.mmu;
2327
2328         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2329         context->new_cr3 = paging_new_cr3;
2330         context->page_fault = paging32_page_fault;
2331         context->gva_to_gpa = paging32_gva_to_gpa;
2332         context->free = paging_free;
2333         context->prefetch_page = paging32_prefetch_page;
2334         context->sync_page = paging32_sync_page;
2335         context->invlpg = paging32_invlpg;
2336         context->root_level = PT32_ROOT_LEVEL;
2337         context->shadow_root_level = PT32E_ROOT_LEVEL;
2338         context->root_hpa = INVALID_PAGE;
2339         return 0;
2340 }
2341
2342 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2343 {
2344         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2345         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2346 }
2347
2348 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2349 {
2350         struct kvm_mmu *context = &vcpu->arch.mmu;
2351
2352         context->new_cr3 = nonpaging_new_cr3;
2353         context->page_fault = tdp_page_fault;
2354         context->free = nonpaging_free;
2355         context->prefetch_page = nonpaging_prefetch_page;
2356         context->sync_page = nonpaging_sync_page;
2357         context->invlpg = nonpaging_invlpg;
2358         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2359         context->root_hpa = INVALID_PAGE;
2360
2361         if (!is_paging(vcpu)) {
2362                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2363                 context->root_level = 0;
2364         } else if (is_long_mode(vcpu)) {
2365                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2366                 context->gva_to_gpa = paging64_gva_to_gpa;
2367                 context->root_level = PT64_ROOT_LEVEL;
2368         } else if (is_pae(vcpu)) {
2369                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2370                 context->gva_to_gpa = paging64_gva_to_gpa;
2371                 context->root_level = PT32E_ROOT_LEVEL;
2372         } else {
2373                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2374                 context->gva_to_gpa = paging32_gva_to_gpa;
2375                 context->root_level = PT32_ROOT_LEVEL;
2376         }
2377
2378         return 0;
2379 }
2380
2381 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2382 {
2383         int r;
2384
2385         ASSERT(vcpu);
2386         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2387
2388         if (!is_paging(vcpu))
2389                 r = nonpaging_init_context(vcpu);
2390         else if (is_long_mode(vcpu))
2391                 r = paging64_init_context(vcpu);
2392         else if (is_pae(vcpu))
2393                 r = paging32E_init_context(vcpu);
2394         else
2395                 r = paging32_init_context(vcpu);
2396
2397         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2398
2399         return r;
2400 }
2401
2402 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2403 {
2404         vcpu->arch.update_pte.pfn = bad_pfn;
2405
2406         if (tdp_enabled)
2407                 return init_kvm_tdp_mmu(vcpu);
2408         else
2409                 return init_kvm_softmmu(vcpu);
2410 }
2411
2412 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2413 {
2414         ASSERT(vcpu);
2415         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2416                 vcpu->arch.mmu.free(vcpu);
2417                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2418         }
2419 }
2420
2421 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2422 {
2423         destroy_kvm_mmu(vcpu);
2424         return init_kvm_mmu(vcpu);
2425 }
2426 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2427
2428 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2429 {
2430         int r;
2431
2432         r = mmu_topup_memory_caches(vcpu);
2433         if (r)
2434                 goto out;
2435         spin_lock(&vcpu->kvm->mmu_lock);
2436         kvm_mmu_free_some_pages(vcpu);
2437         r = mmu_alloc_roots(vcpu);
2438         mmu_sync_roots(vcpu);
2439         spin_unlock(&vcpu->kvm->mmu_lock);
2440         if (r)
2441                 goto out;
2442         /* set_cr3() should ensure TLB has been flushed */
2443         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2444 out:
2445         return r;
2446 }
2447 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2448
2449 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2450 {
2451         mmu_free_roots(vcpu);
2452 }
2453
2454 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2455                                   struct kvm_mmu_page *sp,
2456                                   u64 *spte)
2457 {
2458         u64 pte;
2459         struct kvm_mmu_page *child;
2460
2461         pte = *spte;
2462         if (is_shadow_present_pte(pte)) {
2463                 if (is_last_spte(pte, sp->role.level))
2464                         rmap_remove(vcpu->kvm, spte);
2465                 else {
2466                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2467                         mmu_page_remove_parent_pte(child, spte);
2468                 }
2469         }
2470         __set_spte(spte, shadow_trap_nonpresent_pte);
2471         if (is_large_pte(pte))
2472                 --vcpu->kvm->stat.lpages;
2473 }
2474
2475 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2476                                   struct kvm_mmu_page *sp,
2477                                   u64 *spte,
2478                                   const void *new)
2479 {
2480         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2481                 ++vcpu->kvm->stat.mmu_pde_zapped;
2482                 return;
2483         }
2484
2485         ++vcpu->kvm->stat.mmu_pte_updated;
2486         if (sp->role.glevels == PT32_ROOT_LEVEL)
2487                 paging32_update_pte(vcpu, sp, spte, new);
2488         else
2489                 paging64_update_pte(vcpu, sp, spte, new);
2490 }
2491
2492 static bool need_remote_flush(u64 old, u64 new)
2493 {
2494         if (!is_shadow_present_pte(old))
2495                 return false;
2496         if (!is_shadow_present_pte(new))
2497                 return true;
2498         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2499                 return true;
2500         old ^= PT64_NX_MASK;
2501         new ^= PT64_NX_MASK;
2502         return (old & ~new & PT64_PERM_MASK) != 0;
2503 }
2504
2505 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2506 {
2507         if (need_remote_flush(old, new))
2508                 kvm_flush_remote_tlbs(vcpu->kvm);
2509         else
2510                 kvm_mmu_flush_tlb(vcpu);
2511 }
2512
2513 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2514 {
2515         u64 *spte = vcpu->arch.last_pte_updated;
2516
2517         return !!(spte && (*spte & shadow_accessed_mask));
2518 }
2519
2520 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2521                                           const u8 *new, int bytes)
2522 {
2523         gfn_t gfn;
2524         int r;
2525         u64 gpte = 0;
2526         pfn_t pfn;
2527
2528         if (bytes != 4 && bytes != 8)
2529                 return;
2530
2531         /*
2532          * Assume that the pte write on a page table of the same type
2533          * as the current vcpu paging mode.  This is nearly always true
2534          * (might be false while changing modes).  Note it is verified later
2535          * by update_pte().
2536          */
2537         if (is_pae(vcpu)) {
2538                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2539                 if ((bytes == 4) && (gpa % 4 == 0)) {
2540                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2541                         if (r)
2542                                 return;
2543                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2544                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2545                         memcpy((void *)&gpte, new, 8);
2546                 }
2547         } else {
2548                 if ((bytes == 4) && (gpa % 4 == 0))
2549                         memcpy((void *)&gpte, new, 4);
2550         }
2551         if (!is_present_gpte(gpte))
2552                 return;
2553         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2554
2555         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2556         smp_rmb();
2557         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2558
2559         if (is_error_pfn(pfn)) {
2560                 kvm_release_pfn_clean(pfn);
2561                 return;
2562         }
2563         vcpu->arch.update_pte.gfn = gfn;
2564         vcpu->arch.update_pte.pfn = pfn;
2565 }
2566
2567 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2568 {
2569         u64 *spte = vcpu->arch.last_pte_updated;
2570
2571         if (spte
2572             && vcpu->arch.last_pte_gfn == gfn
2573             && shadow_accessed_mask
2574             && !(*spte & shadow_accessed_mask)
2575             && is_shadow_present_pte(*spte))
2576                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2577 }
2578
2579 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2580                        const u8 *new, int bytes,
2581                        bool guest_initiated)
2582 {
2583         gfn_t gfn = gpa >> PAGE_SHIFT;
2584         struct kvm_mmu_page *sp;
2585         struct hlist_node *node, *n;
2586         struct hlist_head *bucket;
2587         unsigned index;
2588         u64 entry, gentry;
2589         u64 *spte;
2590         unsigned offset = offset_in_page(gpa);
2591         unsigned pte_size;
2592         unsigned page_offset;
2593         unsigned misaligned;
2594         unsigned quadrant;
2595         int level;
2596         int flooded = 0;
2597         int npte;
2598         int r;
2599
2600         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2601         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2602         spin_lock(&vcpu->kvm->mmu_lock);
2603         kvm_mmu_access_page(vcpu, gfn);
2604         kvm_mmu_free_some_pages(vcpu);
2605         ++vcpu->kvm->stat.mmu_pte_write;
2606         kvm_mmu_audit(vcpu, "pre pte write");
2607         if (guest_initiated) {
2608                 if (gfn == vcpu->arch.last_pt_write_gfn
2609                     && !last_updated_pte_accessed(vcpu)) {
2610                         ++vcpu->arch.last_pt_write_count;
2611                         if (vcpu->arch.last_pt_write_count >= 3)
2612                                 flooded = 1;
2613                 } else {
2614                         vcpu->arch.last_pt_write_gfn = gfn;
2615                         vcpu->arch.last_pt_write_count = 1;
2616                         vcpu->arch.last_pte_updated = NULL;
2617                 }
2618         }
2619         index = kvm_page_table_hashfn(gfn);
2620         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2621         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2622                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2623                         continue;
2624                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2625                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2626                 misaligned |= bytes < 4;
2627                 if (misaligned || flooded) {
2628                         /*
2629                          * Misaligned accesses are too much trouble to fix
2630                          * up; also, they usually indicate a page is not used
2631                          * as a page table.
2632                          *
2633                          * If we're seeing too many writes to a page,
2634                          * it may no longer be a page table, or we may be
2635                          * forking, in which case it is better to unmap the
2636                          * page.
2637                          */
2638                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2639                                  gpa, bytes, sp->role.word);
2640                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2641                                 n = bucket->first;
2642                         ++vcpu->kvm->stat.mmu_flooded;
2643                         continue;
2644                 }
2645                 page_offset = offset;
2646                 level = sp->role.level;
2647                 npte = 1;
2648                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2649                         page_offset <<= 1;      /* 32->64 */
2650                         /*
2651                          * A 32-bit pde maps 4MB while the shadow pdes map
2652                          * only 2MB.  So we need to double the offset again
2653                          * and zap two pdes instead of one.
2654                          */
2655                         if (level == PT32_ROOT_LEVEL) {
2656                                 page_offset &= ~7; /* kill rounding error */
2657                                 page_offset <<= 1;
2658                                 npte = 2;
2659                         }
2660                         quadrant = page_offset >> PAGE_SHIFT;
2661                         page_offset &= ~PAGE_MASK;
2662                         if (quadrant != sp->role.quadrant)
2663                                 continue;
2664                 }
2665                 spte = &sp->spt[page_offset / sizeof(*spte)];
2666                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2667                         gentry = 0;
2668                         r = kvm_read_guest_atomic(vcpu->kvm,
2669                                                   gpa & ~(u64)(pte_size - 1),
2670                                                   &gentry, pte_size);
2671                         new = (const void *)&gentry;
2672                         if (r < 0)
2673                                 new = NULL;
2674                 }
2675                 while (npte--) {
2676                         entry = *spte;
2677                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2678                         if (new)
2679                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2680                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2681                         ++spte;
2682                 }
2683         }
2684         kvm_mmu_audit(vcpu, "post pte write");
2685         spin_unlock(&vcpu->kvm->mmu_lock);
2686         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2687                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2688                 vcpu->arch.update_pte.pfn = bad_pfn;
2689         }
2690 }
2691
2692 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2693 {
2694         gpa_t gpa;
2695         int r;
2696
2697         if (tdp_enabled)
2698                 return 0;
2699
2700         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2701
2702         spin_lock(&vcpu->kvm->mmu_lock);
2703         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2704         spin_unlock(&vcpu->kvm->mmu_lock);
2705         return r;
2706 }
2707 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2708
2709 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2710 {
2711         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2712                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2713                 struct kvm_mmu_page *sp;
2714
2715                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2716                                   struct kvm_mmu_page, link);
2717                 kvm_mmu_zap_page(vcpu->kvm, sp);
2718                 ++vcpu->kvm->stat.mmu_recycled;
2719         }
2720 }
2721
2722 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2723 {
2724         int r;
2725         enum emulation_result er;
2726
2727         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2728         if (r < 0)
2729                 goto out;
2730
2731         if (!r) {
2732                 r = 1;
2733                 goto out;
2734         }
2735
2736         r = mmu_topup_memory_caches(vcpu);
2737         if (r)
2738                 goto out;
2739
2740         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2741
2742         switch (er) {
2743         case EMULATE_DONE:
2744                 return 1;
2745         case EMULATE_DO_MMIO:
2746                 ++vcpu->stat.mmio_exits;
2747                 return 0;
2748         case EMULATE_FAIL:
2749                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2750                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2751                 return 0;
2752         default:
2753                 BUG();
2754         }
2755 out:
2756         return r;
2757 }
2758 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2759
2760 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2761 {
2762         vcpu->arch.mmu.invlpg(vcpu, gva);
2763         kvm_mmu_flush_tlb(vcpu);
2764         ++vcpu->stat.invlpg;
2765 }
2766 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2767
2768 void kvm_enable_tdp(void)
2769 {
2770         tdp_enabled = true;
2771 }
2772 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2773
2774 void kvm_disable_tdp(void)
2775 {
2776         tdp_enabled = false;
2777 }
2778 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2779
2780 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2781 {
2782         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2783 }
2784
2785 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2786 {
2787         struct page *page;
2788         int i;
2789
2790         ASSERT(vcpu);
2791
2792         /*
2793          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2794          * Therefore we need to allocate shadow page tables in the first
2795          * 4GB of memory, which happens to fit the DMA32 zone.
2796          */
2797         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2798         if (!page)
2799                 goto error_1;
2800         vcpu->arch.mmu.pae_root = page_address(page);
2801         for (i = 0; i < 4; ++i)
2802                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2803
2804         return 0;
2805
2806 error_1:
2807         free_mmu_pages(vcpu);
2808         return -ENOMEM;
2809 }
2810
2811 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2812 {
2813         ASSERT(vcpu);
2814         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2815
2816         return alloc_mmu_pages(vcpu);
2817 }
2818
2819 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2820 {
2821         ASSERT(vcpu);
2822         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2823
2824         return init_kvm_mmu(vcpu);
2825 }
2826
2827 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2828 {
2829         ASSERT(vcpu);
2830
2831         destroy_kvm_mmu(vcpu);
2832         free_mmu_pages(vcpu);
2833         mmu_free_memory_caches(vcpu);
2834 }
2835
2836 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2837 {
2838         struct kvm_mmu_page *sp;
2839
2840         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2841                 int i;
2842                 u64 *pt;
2843
2844                 if (!test_bit(slot, sp->slot_bitmap))
2845                         continue;
2846
2847                 pt = sp->spt;
2848                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2849                         /* avoid RMW */
2850                         if (pt[i] & PT_WRITABLE_MASK)
2851                                 pt[i] &= ~PT_WRITABLE_MASK;
2852         }
2853         kvm_flush_remote_tlbs(kvm);
2854 }
2855
2856 void kvm_mmu_zap_all(struct kvm *kvm)
2857 {
2858         struct kvm_mmu_page *sp, *node;
2859
2860         spin_lock(&kvm->mmu_lock);
2861         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2862                 if (kvm_mmu_zap_page(kvm, sp))
2863                         node = container_of(kvm->arch.active_mmu_pages.next,
2864                                             struct kvm_mmu_page, link);
2865         spin_unlock(&kvm->mmu_lock);
2866
2867         kvm_flush_remote_tlbs(kvm);
2868 }
2869
2870 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2871 {
2872         struct kvm_mmu_page *page;
2873
2874         page = container_of(kvm->arch.active_mmu_pages.prev,
2875                             struct kvm_mmu_page, link);
2876         kvm_mmu_zap_page(kvm, page);
2877 }
2878
2879 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2880 {
2881         struct kvm *kvm;
2882         struct kvm *kvm_freed = NULL;
2883         int cache_count = 0;
2884
2885         spin_lock(&kvm_lock);
2886
2887         list_for_each_entry(kvm, &vm_list, vm_list) {
2888                 int npages;
2889
2890                 if (!down_read_trylock(&kvm->slots_lock))
2891                         continue;
2892                 spin_lock(&kvm->mmu_lock);
2893                 npages = kvm->arch.n_alloc_mmu_pages -
2894                          kvm->arch.n_free_mmu_pages;
2895                 cache_count += npages;
2896                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2897                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2898                         cache_count--;
2899                         kvm_freed = kvm;
2900                 }
2901                 nr_to_scan--;
2902
2903                 spin_unlock(&kvm->mmu_lock);
2904                 up_read(&kvm->slots_lock);
2905         }
2906         if (kvm_freed)
2907                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2908
2909         spin_unlock(&kvm_lock);
2910
2911         return cache_count;
2912 }
2913
2914 static struct shrinker mmu_shrinker = {
2915         .shrink = mmu_shrink,
2916         .seeks = DEFAULT_SEEKS * 10,
2917 };
2918
2919 static void mmu_destroy_caches(void)
2920 {
2921         if (pte_chain_cache)
2922                 kmem_cache_destroy(pte_chain_cache);
2923         if (rmap_desc_cache)
2924                 kmem_cache_destroy(rmap_desc_cache);
2925         if (mmu_page_header_cache)
2926                 kmem_cache_destroy(mmu_page_header_cache);
2927 }
2928
2929 void kvm_mmu_module_exit(void)
2930 {
2931         mmu_destroy_caches();
2932         unregister_shrinker(&mmu_shrinker);
2933 }
2934
2935 int kvm_mmu_module_init(void)
2936 {
2937         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2938                                             sizeof(struct kvm_pte_chain),
2939                                             0, 0, NULL);
2940         if (!pte_chain_cache)
2941                 goto nomem;
2942         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2943                                             sizeof(struct kvm_rmap_desc),
2944                                             0, 0, NULL);
2945         if (!rmap_desc_cache)
2946                 goto nomem;
2947
2948         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2949                                                   sizeof(struct kvm_mmu_page),
2950                                                   0, 0, NULL);
2951         if (!mmu_page_header_cache)
2952                 goto nomem;
2953
2954         register_shrinker(&mmu_shrinker);
2955
2956         return 0;
2957
2958 nomem:
2959         mmu_destroy_caches();
2960         return -ENOMEM;
2961 }
2962
2963 /*
2964  * Caculate mmu pages needed for kvm.
2965  */
2966 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2967 {
2968         int i;
2969         unsigned int nr_mmu_pages;
2970         unsigned int  nr_pages = 0;
2971
2972         for (i = 0; i < kvm->nmemslots; i++)
2973                 nr_pages += kvm->memslots[i].npages;
2974
2975         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2976         nr_mmu_pages = max(nr_mmu_pages,
2977                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2978
2979         return nr_mmu_pages;
2980 }
2981
2982 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2983                                 unsigned len)
2984 {
2985         if (len > buffer->len)
2986                 return NULL;
2987         return buffer->ptr;
2988 }
2989
2990 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2991                                 unsigned len)
2992 {
2993         void *ret;
2994
2995         ret = pv_mmu_peek_buffer(buffer, len);
2996         if (!ret)
2997                 return ret;
2998         buffer->ptr += len;
2999         buffer->len -= len;
3000         buffer->processed += len;
3001         return ret;
3002 }
3003
3004 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3005                              gpa_t addr, gpa_t value)
3006 {
3007         int bytes = 8;
3008         int r;
3009
3010         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3011                 bytes = 4;
3012
3013         r = mmu_topup_memory_caches(vcpu);
3014         if (r)
3015                 return r;
3016
3017         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3018                 return -EFAULT;
3019
3020         return 1;
3021 }
3022
3023 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3024 {
3025         kvm_set_cr3(vcpu, vcpu->arch.cr3);
3026         return 1;
3027 }
3028
3029 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3030 {
3031         spin_lock(&vcpu->kvm->mmu_lock);
3032         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3033         spin_unlock(&vcpu->kvm->mmu_lock);
3034         return 1;
3035 }
3036
3037 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3038                              struct kvm_pv_mmu_op_buffer *buffer)
3039 {
3040         struct kvm_mmu_op_header *header;
3041
3042         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3043         if (!header)
3044                 return 0;
3045         switch (header->op) {
3046         case KVM_MMU_OP_WRITE_PTE: {
3047                 struct kvm_mmu_op_write_pte *wpte;
3048
3049                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3050                 if (!wpte)
3051                         return 0;
3052                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3053                                         wpte->pte_val);
3054         }
3055         case KVM_MMU_OP_FLUSH_TLB: {
3056                 struct kvm_mmu_op_flush_tlb *ftlb;
3057
3058                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3059                 if (!ftlb)
3060                         return 0;
3061                 return kvm_pv_mmu_flush_tlb(vcpu);
3062         }
3063         case KVM_MMU_OP_RELEASE_PT: {
3064                 struct kvm_mmu_op_release_pt *rpt;
3065
3066                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3067                 if (!rpt)
3068                         return 0;
3069                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3070         }
3071         default: return 0;
3072         }
3073 }
3074
3075 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3076                   gpa_t addr, unsigned long *ret)
3077 {
3078         int r;
3079         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3080
3081         buffer->ptr = buffer->buf;
3082         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3083         buffer->processed = 0;
3084
3085         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3086         if (r)
3087                 goto out;
3088
3089         while (buffer->len) {
3090                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3091                 if (r < 0)
3092                         goto out;
3093                 if (r == 0)
3094                         break;
3095         }
3096
3097         r = 1;
3098 out:
3099         *ret = buffer->processed;
3100         return r;
3101 }
3102
3103 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3104 {
3105         struct kvm_shadow_walk_iterator iterator;
3106         int nr_sptes = 0;
3107
3108         spin_lock(&vcpu->kvm->mmu_lock);
3109         for_each_shadow_entry(vcpu, addr, iterator) {
3110                 sptes[iterator.level-1] = *iterator.sptep;
3111                 nr_sptes++;
3112                 if (!is_shadow_present_pte(*iterator.sptep))
3113                         break;
3114         }
3115         spin_unlock(&vcpu->kvm->mmu_lock);
3116
3117         return nr_sptes;
3118 }
3119 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3120
3121 #ifdef AUDIT
3122
3123 static const char *audit_msg;
3124
3125 static gva_t canonicalize(gva_t gva)
3126 {
3127 #ifdef CONFIG_X86_64
3128         gva = (long long)(gva << 16) >> 16;
3129 #endif
3130         return gva;
3131 }
3132
3133
3134 typedef void (*inspect_spte_fn) (struct kvm *kvm, struct kvm_mmu_page *sp,
3135                                  u64 *sptep);
3136
3137 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3138                             inspect_spte_fn fn)
3139 {
3140         int i;
3141
3142         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3143                 u64 ent = sp->spt[i];
3144
3145                 if (is_shadow_present_pte(ent)) {
3146                         if (!is_last_spte(ent, sp->role.level)) {
3147                                 struct kvm_mmu_page *child;
3148                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3149                                 __mmu_spte_walk(kvm, child, fn);
3150                         } else
3151                                 fn(kvm, sp, &sp->spt[i]);
3152                 }
3153         }
3154 }
3155
3156 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3157 {
3158         int i;
3159         struct kvm_mmu_page *sp;
3160
3161         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3162                 return;
3163         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3164                 hpa_t root = vcpu->arch.mmu.root_hpa;
3165                 sp = page_header(root);
3166                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3167                 return;
3168         }
3169         for (i = 0; i < 4; ++i) {
3170                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3171
3172                 if (root && VALID_PAGE(root)) {
3173                         root &= PT64_BASE_ADDR_MASK;
3174                         sp = page_header(root);
3175                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3176                 }
3177         }
3178         return;
3179 }
3180
3181 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3182                                 gva_t va, int level)
3183 {
3184         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3185         int i;
3186         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3187
3188         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3189                 u64 ent = pt[i];
3190
3191                 if (ent == shadow_trap_nonpresent_pte)
3192                         continue;
3193
3194                 va = canonicalize(va);
3195                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3196                         audit_mappings_page(vcpu, ent, va, level - 1);
3197                 else {
3198                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3199                         gfn_t gfn = gpa >> PAGE_SHIFT;
3200                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3201                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3202
3203                         if (is_error_pfn(pfn)) {
3204                                 kvm_release_pfn_clean(pfn);
3205                                 continue;
3206                         }
3207
3208                         if (is_shadow_present_pte(ent)
3209                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3210                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3211                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3212                                        audit_msg, vcpu->arch.mmu.root_level,
3213                                        va, gpa, hpa, ent,
3214                                        is_shadow_present_pte(ent));
3215                         else if (ent == shadow_notrap_nonpresent_pte
3216                                  && !is_error_hpa(hpa))
3217                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3218                                        " valid guest gva %lx\n", audit_msg, va);
3219                         kvm_release_pfn_clean(pfn);
3220
3221                 }
3222         }
3223 }
3224
3225 static void audit_mappings(struct kvm_vcpu *vcpu)
3226 {
3227         unsigned i;
3228
3229         if (vcpu->arch.mmu.root_level == 4)
3230                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3231         else
3232                 for (i = 0; i < 4; ++i)
3233                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3234                                 audit_mappings_page(vcpu,
3235                                                     vcpu->arch.mmu.pae_root[i],
3236                                                     i << 30,
3237                                                     2);
3238 }
3239
3240 static int count_rmaps(struct kvm_vcpu *vcpu)
3241 {
3242         int nmaps = 0;
3243         int i, j, k;
3244
3245         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3246                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3247                 struct kvm_rmap_desc *d;
3248
3249                 for (j = 0; j < m->npages; ++j) {
3250                         unsigned long *rmapp = &m->rmap[j];
3251
3252                         if (!*rmapp)
3253                                 continue;
3254                         if (!(*rmapp & 1)) {
3255                                 ++nmaps;
3256                                 continue;
3257                         }
3258                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3259                         while (d) {
3260                                 for (k = 0; k < RMAP_EXT; ++k)
3261                                         if (d->sptes[k])
3262                                                 ++nmaps;
3263                                         else
3264                                                 break;
3265                                 d = d->more;
3266                         }
3267                 }
3268         }
3269         return nmaps;
3270 }
3271
3272 void inspect_spte_has_rmap(struct kvm *kvm, struct kvm_mmu_page *sp, u64 *sptep)
3273 {
3274         unsigned long *rmapp;
3275         struct kvm_mmu_page *rev_sp;
3276         gfn_t gfn;
3277
3278         if (*sptep & PT_WRITABLE_MASK) {
3279                 rev_sp = page_header(__pa(sptep));
3280                 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3281
3282                 if (!gfn_to_memslot(kvm, gfn)) {
3283                         if (!printk_ratelimit())
3284                                 return;
3285                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3286                                          audit_msg, gfn);
3287                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3288                                         audit_msg, sptep - rev_sp->spt,
3289                                         rev_sp->gfn);
3290                         dump_stack();
3291                         return;
3292                 }
3293
3294                 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3295                                     is_large_pte(*sptep));
3296                 if (!*rmapp) {
3297                         if (!printk_ratelimit())
3298                                 return;
3299                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3300                                          audit_msg, *sptep);
3301                         dump_stack();
3302                 }
3303         }
3304
3305 }
3306
3307 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3308 {
3309         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3310 }
3311
3312 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3313 {
3314         struct kvm_mmu_page *sp;
3315         int i;
3316
3317         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3318                 u64 *pt = sp->spt;
3319
3320                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3321                         continue;
3322
3323                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3324                         u64 ent = pt[i];
3325
3326                         if (!(ent & PT_PRESENT_MASK))
3327                                 continue;
3328                         if (!(ent & PT_WRITABLE_MASK))
3329                                 continue;
3330                         inspect_spte_has_rmap(vcpu->kvm, sp, &pt[i]);
3331                 }
3332         }
3333         return;
3334 }
3335
3336 static void audit_rmap(struct kvm_vcpu *vcpu)
3337 {
3338         check_writable_mappings_rmap(vcpu);
3339         count_rmaps(vcpu);
3340 }
3341
3342 static void audit_write_protection(struct kvm_vcpu *vcpu)
3343 {
3344         struct kvm_mmu_page *sp;
3345         struct kvm_memory_slot *slot;
3346         unsigned long *rmapp;
3347         u64 *spte;
3348         gfn_t gfn;
3349
3350         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3351                 if (sp->role.direct)
3352                         continue;
3353                 if (sp->unsync)
3354                         continue;
3355
3356                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3357                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3358                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3359
3360                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3361                 while (spte) {
3362                         if (*spte & PT_WRITABLE_MASK)
3363                                 printk(KERN_ERR "%s: (%s) shadow page has "
3364                                 "writable mappings: gfn %lx role %x\n",
3365                                __func__, audit_msg, sp->gfn,
3366                                sp->role.word);
3367                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3368                 }
3369         }
3370 }
3371
3372 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3373 {
3374         int olddbg = dbg;
3375
3376         dbg = 0;
3377         audit_msg = msg;
3378         audit_rmap(vcpu);
3379         audit_write_protection(vcpu);
3380         if (strcmp("pre pte write", audit_msg) != 0)
3381                 audit_mappings(vcpu);
3382         audit_writable_sptes_have_rmaps(vcpu);
3383         dbg = olddbg;
3384 }
3385
3386 #endif