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
  22 #include <linux/kvm_host.h>
  23 #include <linux/types.h>
  24 #include <linux/string.h>
  25 #include <linux/mm.h>
  26 #include <linux/highmem.h>
  27 #include <linux/module.h>
  28 #include <linux/swap.h>
  29 #include <linux/hugetlb.h>
  30 #include <linux/compiler.h>
  31
  32 #include <asm/page.h>
  33 #include <asm/cmpxchg.h>
  34 #include <asm/io.h>
  35 #include <asm/vmx.h>
  36
  37 /*
  38  * When setting this variable to true it enables Two-Dimensional-Paging
  39  * where the hardware walks 2 page tables:
  40  * 1. the guest-virtual to guest-physical
  41  * 2. while doing 1. it walks guest-physical to host-physical
  42  * If the hardware supports that we don't need to do shadow paging.
  43  */
  44 bool tdp_enabled = false;
  45
  46 #undef MMU_DEBUG
  47
  48 #undef AUDIT
  49
  50 #ifdef AUDIT
  51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  52 #else
  53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  54 #endif
  55
  56 #ifdef MMU_DEBUG
  57
  58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  60
  61 #else
  62
  63 #define pgprintk(x...) do { } while (0)
  64 #define rmap_printk(x...) do { } while (0)
  65
  66 #endif
  67
  68 #if defined(MMU_DEBUG) || defined(AUDIT)
  69 static int dbg = 0;
  70 module_param(dbg, bool, 0644);
  71 #endif
  72
  73 static int oos_shadow = 1;
  74 module_param(oos_shadow, bool, 0644);
  75
  76 #ifndef MMU_DEBUG
  77 #define ASSERT(x) do { } while (0)
  78 #else
  79 #define ASSERT(x)                                                       \
  80         if (!(x)) {                                                     \
  81                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  82                        __FILE__, __LINE__, #x);                         \
  83         }
  84 #endif
  85
  86 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  87 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  88
  89 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  90
  91 #define PT64_LEVEL_BITS 9
  92
  93 #define PT64_LEVEL_SHIFT(level) \
  94                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  95
  96 #define PT64_LEVEL_MASK(level) \
  97                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  98
  99 #define PT64_INDEX(address, level)\
 100         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 101
 102
 103 #define PT32_LEVEL_BITS 10
 104
 105 #define PT32_LEVEL_SHIFT(level) \
 106                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 107
 108 #define PT32_LEVEL_MASK(level) \
 109                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 110
 111 #define PT32_INDEX(address, level)\
 112         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 113
 114
 115 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 116 #define PT64_DIR_BASE_ADDR_MASK \
 117         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 118
 119 #define PT32_BASE_ADDR_MASK PAGE_MASK
 120 #define PT32_DIR_BASE_ADDR_MASK \
 121         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 122
 123 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 124                         | PT64_NX_MASK)
 125
 126 #define PFERR_PRESENT_MASK (1U << 0)
 127 #define PFERR_WRITE_MASK (1U << 1)
 128 #define PFERR_USER_MASK (1U << 2)
 129 #define PFERR_RSVD_MASK (1U << 3)
 130 #define PFERR_FETCH_MASK (1U << 4)
 131
 132 #define PT_DIRECTORY_LEVEL 2
 133 #define PT_PAGE_TABLE_LEVEL 1
 134
 135 #define RMAP_EXT 4
 136
 137 #define ACC_EXEC_MASK    1
 138 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 139 #define ACC_USER_MASK    PT_USER_MASK
 140 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 141
 142 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 143
 144 struct kvm_rmap_desc {
 145         u64 *shadow_ptes[RMAP_EXT];
 146         struct kvm_rmap_desc *more;
 147 };
 148
 149 struct kvm_shadow_walk_iterator {
 150         u64 addr;
 151         hpa_t shadow_addr;
 152         int level;
 153         u64 *sptep;
 154         unsigned index;
 155 };
 156
 157 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
 158         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
 159              shadow_walk_okay(&(_walker));                      \
 160              shadow_walk_next(&(_walker)))
 161
 162
 163 struct kvm_unsync_walk {
 164         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 165 };
 166
 167 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 168
 169 static struct kmem_cache *pte_chain_cache;
 170 static struct kmem_cache *rmap_desc_cache;
 171 static struct kmem_cache *mmu_page_header_cache;
 172
 173 static u64 __read_mostly shadow_trap_nonpresent_pte;
 174 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 175 static u64 __read_mostly shadow_base_present_pte;
 176 static u64 __read_mostly shadow_nx_mask;
 177 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 178 static u64 __read_mostly shadow_user_mask;
 179 static u64 __read_mostly shadow_accessed_mask;
 180 static u64 __read_mostly shadow_dirty_mask;
 181
 182 static inline u64 rsvd_bits(int s, int e)
 183 {
 184         return ((1ULL << (e - s + 1)) - 1) << s;
 185 }
 186
 187 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 188 {
 189         shadow_trap_nonpresent_pte = trap_pte;
 190         shadow_notrap_nonpresent_pte = notrap_pte;
 191 }
 192 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 193
 194 void kvm_mmu_set_base_ptes(u64 base_pte)
 195 {
 196         shadow_base_present_pte = base_pte;
 197 }
 198 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 199
 200 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 201                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 202 {
 203         shadow_user_mask = user_mask;
 204         shadow_accessed_mask = accessed_mask;
 205         shadow_dirty_mask = dirty_mask;
 206         shadow_nx_mask = nx_mask;
 207         shadow_x_mask = x_mask;
 208 }
 209 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 210
 211 static int is_write_protection(struct kvm_vcpu *vcpu)
 212 {
 213         return vcpu->arch.cr0 & X86_CR0_WP;
 214 }
 215
 216 static int is_cpuid_PSE36(void)
 217 {
 218         return 1;
 219 }
 220
 221 static int is_nx(struct kvm_vcpu *vcpu)
 222 {
 223         return vcpu->arch.shadow_efer & EFER_NX;
 224 }
 225
 226 static int is_shadow_present_pte(u64 pte)
 227 {
 228         return pte != shadow_trap_nonpresent_pte
 229                 && pte != shadow_notrap_nonpresent_pte;
 230 }
 231
 232 static int is_large_pte(u64 pte)
 233 {
 234         return pte & PT_PAGE_SIZE_MASK;
 235 }
 236
 237 static int is_writeble_pte(unsigned long pte)
 238 {
 239         return pte & PT_WRITABLE_MASK;
 240 }
 241
 242 static int is_dirty_pte(unsigned long pte)
 243 {
 244         return pte & shadow_dirty_mask;
 245 }
 246
 247 static int is_rmap_pte(u64 pte)
 248 {
 249         return is_shadow_present_pte(pte);
 250 }
 251
 252 static pfn_t spte_to_pfn(u64 pte)
 253 {
 254         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 255 }
 256
 257 static gfn_t pse36_gfn_delta(u32 gpte)
 258 {
 259         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 260
 261         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 262 }
 263
 264 static void set_shadow_pte(u64 *sptep, u64 spte)
 265 {
 266 #ifdef CONFIG_X86_64
 267         set_64bit((unsigned long *)sptep, spte);
 268 #else
 269         set_64bit((unsigned long long *)sptep, spte);
 270 #endif
 271 }
 272
 273 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 274                                   struct kmem_cache *base_cache, int min)
 275 {
 276         void *obj;
 277
 278         if (cache->nobjs >= min)
 279                 return 0;
 280         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 281                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 282                 if (!obj)
 283                         return -ENOMEM;
 284                 cache->objects[cache->nobjs++] = obj;
 285         }
 286         return 0;
 287 }
 288
 289 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 290 {
 291         while (mc->nobjs)
 292                 kfree(mc->objects[--mc->nobjs]);
 293 }
 294
 295 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 296                                        int min)
 297 {
 298         struct page *page;
 299
 300         if (cache->nobjs >= min)
 301                 return 0;
 302         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 303                 page = alloc_page(GFP_KERNEL);
 304                 if (!page)
 305                         return -ENOMEM;
 306                 set_page_private(page, 0);
 307                 cache->objects[cache->nobjs++] = page_address(page);
 308         }
 309         return 0;
 310 }
 311
 312 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 313 {
 314         while (mc->nobjs)
 315                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 316 }
 317
 318 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 319 {
 320         int r;
 321
 322         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 323                                    pte_chain_cache, 4);
 324         if (r)
 325                 goto out;
 326         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 327                                    rmap_desc_cache, 4);
 328         if (r)
 329                 goto out;
 330         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 331         if (r)
 332                 goto out;
 333         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 334                                    mmu_page_header_cache, 4);
 335 out:
 336         return r;
 337 }
 338
 339 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 340 {
 341         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 342         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 343         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 344         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 345 }
 346
 347 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 348                                     size_t size)
 349 {
 350         void *p;
 351
 352         BUG_ON(!mc->nobjs);
 353         p = mc->objects[--mc->nobjs];
 354         return p;
 355 }
 356
 357 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 358 {
 359         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 360                                       sizeof(struct kvm_pte_chain));
 361 }
 362
 363 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 364 {
 365         kfree(pc);
 366 }
 367
 368 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 369 {
 370         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 371                                       sizeof(struct kvm_rmap_desc));
 372 }
 373
 374 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 375 {
 376         kfree(rd);
 377 }
 378
 379 /*
 380  * Return the pointer to the largepage write count for a given
 381  * gfn, handling slots that are not large page aligned.
 382  */
 383 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 384 {
 385         unsigned long idx;
 386
 387         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 388               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 389         return &slot->lpage_info[idx].write_count;
 390 }
 391
 392 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 393 {
 394         int *write_count;
 395
 396         gfn = unalias_gfn(kvm, gfn);
 397         write_count = slot_largepage_idx(gfn,
 398                                          gfn_to_memslot_unaliased(kvm, gfn));
 399         *write_count += 1;
 400 }
 401
 402 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 403 {
 404         int *write_count;
 405
 406         gfn = unalias_gfn(kvm, gfn);
 407         write_count = slot_largepage_idx(gfn,
 408                                          gfn_to_memslot_unaliased(kvm, gfn));
 409         *write_count -= 1;
 410         WARN_ON(*write_count < 0);
 411 }
 412
 413 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 414 {
 415         struct kvm_memory_slot *slot;
 416         int *largepage_idx;
 417
 418         gfn = unalias_gfn(kvm, gfn);
 419         slot = gfn_to_memslot_unaliased(kvm, gfn);
 420         if (slot) {
 421                 largepage_idx = slot_largepage_idx(gfn, slot);
 422                 return *largepage_idx;
 423         }
 424
 425         return 1;
 426 }
 427
 428 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 429 {
 430         struct vm_area_struct *vma;
 431         unsigned long addr;
 432         int ret = 0;
 433
 434         addr = gfn_to_hva(kvm, gfn);
 435         if (kvm_is_error_hva(addr))
 436                 return ret;
 437
 438         down_read(&current->mm->mmap_sem);
 439         vma = find_vma(current->mm, addr);
 440         if (vma && is_vm_hugetlb_page(vma))
 441                 ret = 1;
 442         up_read(&current->mm->mmap_sem);
 443
 444         return ret;
 445 }
 446
 447 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 448 {
 449         struct kvm_memory_slot *slot;
 450
 451         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 452                 return 0;
 453
 454         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 455                 return 0;
 456
 457         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 458         if (slot && slot->dirty_bitmap)
 459                 return 0;
 460
 461         return 1;
 462 }
 463
 464 /*
 465  * Take gfn and return the reverse mapping to it.
 466  * Note: gfn must be unaliased before this function get called
 467  */
 468
 469 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 470 {
 471         struct kvm_memory_slot *slot;
 472         unsigned long idx;
 473
 474         slot = gfn_to_memslot(kvm, gfn);
 475         if (!lpage)
 476                 return &slot->rmap[gfn - slot->base_gfn];
 477
 478         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 479               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 480
 481         return &slot->lpage_info[idx].rmap_pde;
 482 }
 483
 484 /*
 485  * Reverse mapping data structures:
 486  *
 487  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 488  * that points to page_address(page).
 489  *
 490  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 491  * containing more mappings.
 492  *
 493  * Returns the number of rmap entries before the spte was added or zero if
 494  * the spte was not added.
 495  *
 496  */
 497 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 498 {
 499         struct kvm_mmu_page *sp;
 500         struct kvm_rmap_desc *desc;
 501         unsigned long *rmapp;
 502         int i, count = 0;
 503
 504         if (!is_rmap_pte(*spte))
 505                 return count;
 506         gfn = unalias_gfn(vcpu->kvm, gfn);
 507         sp = page_header(__pa(spte));
 508         sp->gfns[spte - sp->spt] = gfn;
 509         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 510         if (!*rmapp) {
 511                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 512                 *rmapp = (unsigned long)spte;
 513         } else if (!(*rmapp & 1)) {
 514                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 515                 desc = mmu_alloc_rmap_desc(vcpu);
 516                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 517                 desc->shadow_ptes[1] = spte;
 518                 *rmapp = (unsigned long)desc | 1;
 519         } else {
 520                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 521                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 522                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more) {
 523                         desc = desc->more;
 524                         count += RMAP_EXT;
 525                 }
 526                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 527                         desc->more = mmu_alloc_rmap_desc(vcpu);
 528                         desc = desc->more;
 529                 }
 530                 for (i = 0; desc->shadow_ptes[i]; ++i)
 531                         ;
 532                 desc->shadow_ptes[i] = spte;
 533         }
 534         return count;
 535 }
 536
 537 static void rmap_desc_remove_entry(unsigned long *rmapp,
 538                                    struct kvm_rmap_desc *desc,
 539                                    int i,
 540                                    struct kvm_rmap_desc *prev_desc)
 541 {
 542         int j;
 543
 544         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 545                 ;
 546         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 547         desc->shadow_ptes[j] = NULL;
 548         if (j != 0)
 549                 return;
 550         if (!prev_desc && !desc->more)
 551                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 552         else
 553                 if (prev_desc)
 554                         prev_desc->more = desc->more;
 555                 else
 556                         *rmapp = (unsigned long)desc->more | 1;
 557         mmu_free_rmap_desc(desc);
 558 }
 559
 560 static void rmap_remove(struct kvm *kvm, u64 *spte)
 561 {
 562         struct kvm_rmap_desc *desc;
 563         struct kvm_rmap_desc *prev_desc;
 564         struct kvm_mmu_page *sp;
 565         pfn_t pfn;
 566         unsigned long *rmapp;
 567         int i;
 568
 569         if (!is_rmap_pte(*spte))
 570                 return;
 571         sp = page_header(__pa(spte));
 572         pfn = spte_to_pfn(*spte);
 573         if (*spte & shadow_accessed_mask)
 574                 kvm_set_pfn_accessed(pfn);
 575         if (is_writeble_pte(*spte))
 576                 kvm_release_pfn_dirty(pfn);
 577         else
 578                 kvm_release_pfn_clean(pfn);
 579         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 580         if (!*rmapp) {
 581                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 582                 BUG();
 583         } else if (!(*rmapp & 1)) {
 584                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 585                 if ((u64 *)*rmapp != spte) {
 586                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 587                                spte, *spte);
 588                         BUG();
 589                 }
 590                 *rmapp = 0;
 591         } else {
 592                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 593                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 594                 prev_desc = NULL;
 595                 while (desc) {
 596                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 597                                 if (desc->shadow_ptes[i] == spte) {
 598                                         rmap_desc_remove_entry(rmapp,
 599                                                                desc, i,
 600                                                                prev_desc);
 601                                         return;
 602                                 }
 603                         prev_desc = desc;
 604                         desc = desc->more;
 605                 }
 606                 BUG();
 607         }
 608 }
 609
 610 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 611 {
 612         struct kvm_rmap_desc *desc;
 613         struct kvm_rmap_desc *prev_desc;
 614         u64 *prev_spte;
 615         int i;
 616
 617         if (!*rmapp)
 618                 return NULL;
 619         else if (!(*rmapp & 1)) {
 620                 if (!spte)
 621                         return (u64 *)*rmapp;
 622                 return NULL;
 623         }
 624         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 625         prev_desc = NULL;
 626         prev_spte = NULL;
 627         while (desc) {
 628                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 629                         if (prev_spte == spte)
 630                                 return desc->shadow_ptes[i];
 631                         prev_spte = desc->shadow_ptes[i];
 632                 }
 633                 desc = desc->more;
 634         }
 635         return NULL;
 636 }
 637
 638 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 639 {
 640         unsigned long *rmapp;
 641         u64 *spte;
 642         int write_protected = 0;
 643
 644         gfn = unalias_gfn(kvm, gfn);
 645         rmapp = gfn_to_rmap(kvm, gfn, 0);
 646
 647         spte = rmap_next(kvm, rmapp, NULL);
 648         while (spte) {
 649                 BUG_ON(!spte);
 650                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 651                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 652                 if (is_writeble_pte(*spte)) {
 653                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 654                         write_protected = 1;
 655                 }
 656                 spte = rmap_next(kvm, rmapp, spte);
 657         }
 658         if (write_protected) {
 659                 pfn_t pfn;
 660
 661                 spte = rmap_next(kvm, rmapp, NULL);
 662                 pfn = spte_to_pfn(*spte);
 663                 kvm_set_pfn_dirty(pfn);
 664         }
 665
 666         /* check for huge page mappings */
 667         rmapp = gfn_to_rmap(kvm, gfn, 1);
 668         spte = rmap_next(kvm, rmapp, NULL);
 669         while (spte) {
 670                 BUG_ON(!spte);
 671                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 672                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 673                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 674                 if (is_writeble_pte(*spte)) {
 675                         rmap_remove(kvm, spte);
 676                         --kvm->stat.lpages;
 677                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 678                         spte = NULL;
 679                         write_protected = 1;
 680                 }
 681                 spte = rmap_next(kvm, rmapp, spte);
 682         }
 683
 684         return write_protected;
 685 }
 686
 687 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 688 {
 689         u64 *spte;
 690         int need_tlb_flush = 0;
 691
 692         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 693                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 694                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 695                 rmap_remove(kvm, spte);
 696                 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 697                 need_tlb_flush = 1;
 698         }
 699         return need_tlb_flush;
 700 }
 701
 702 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 703                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 704 {
 705         int i;
 706         int retval = 0;
 707
 708         /*
 709          * If mmap_sem isn't taken, we can look the memslots with only
 710          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 711          */
 712         for (i = 0; i < kvm->nmemslots; i++) {
 713                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 714                 unsigned long start = memslot->userspace_addr;
 715                 unsigned long end;
 716
 717                 /* mmu_lock protects userspace_addr */
 718                 if (!start)
 719                         continue;
 720
 721                 end = start + (memslot->npages << PAGE_SHIFT);
 722                 if (hva >= start && hva < end) {
 723                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 724                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 725                         retval |= handler(kvm,
 726                                           &memslot->lpage_info[
 727                                                   gfn_offset /
 728                                                   KVM_PAGES_PER_HPAGE].rmap_pde);
 729                 }
 730         }
 731
 732         return retval;
 733 }
 734
 735 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 736 {
 737         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 738 }
 739
 740 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 741 {
 742         u64 *spte;
 743         int young = 0;
 744
 745         /* always return old for EPT */
 746         if (!shadow_accessed_mask)
 747                 return 0;
 748
 749         spte = rmap_next(kvm, rmapp, NULL);
 750         while (spte) {
 751                 int _young;
 752                 u64 _spte = *spte;
 753                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 754                 _young = _spte & PT_ACCESSED_MASK;
 755                 if (_young) {
 756                         young = 1;
 757                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 758                 }
 759                 spte = rmap_next(kvm, rmapp, spte);
 760         }
 761         return young;
 762 }
 763
 764 #define RMAP_RECYCLE_THRESHOLD 1000
 765
 766 static void rmap_recycle(struct kvm_vcpu *vcpu, gfn_t gfn, int lpage)
 767 {
 768         unsigned long *rmapp;
 769
 770         gfn = unalias_gfn(vcpu->kvm, gfn);
 771         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 772
 773         kvm_unmap_rmapp(vcpu->kvm, rmapp);
 774         kvm_flush_remote_tlbs(vcpu->kvm);
 775 }
 776
 777 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 778 {
 779         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 780 }
 781
 782 #ifdef MMU_DEBUG
 783 static int is_empty_shadow_page(u64 *spt)
 784 {
 785         u64 *pos;
 786         u64 *end;
 787
 788         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 789                 if (is_shadow_present_pte(*pos)) {
 790                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 791                                pos, *pos);
 792                         return 0;
 793                 }
 794         return 1;
 795 }
 796 #endif
 797
 798 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 799 {
 800         ASSERT(is_empty_shadow_page(sp->spt));
 801         list_del(&sp->link);
 802         __free_page(virt_to_page(sp->spt));
 803         __free_page(virt_to_page(sp->gfns));
 804         kfree(sp);
 805         ++kvm->arch.n_free_mmu_pages;
 806 }
 807
 808 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 809 {
 810         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 811 }
 812
 813 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 814                                                u64 *parent_pte)
 815 {
 816         struct kvm_mmu_page *sp;
 817
 818         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 819         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 820         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 821         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 822         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 823         INIT_LIST_HEAD(&sp->oos_link);
 824         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 825         sp->multimapped = 0;
 826         sp->parent_pte = parent_pte;
 827         --vcpu->kvm->arch.n_free_mmu_pages;
 828         return sp;
 829 }
 830
 831 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 832                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 833 {
 834         struct kvm_pte_chain *pte_chain;
 835         struct hlist_node *node;
 836         int i;
 837
 838         if (!parent_pte)
 839                 return;
 840         if (!sp->multimapped) {
 841                 u64 *old = sp->parent_pte;
 842
 843                 if (!old) {
 844                         sp->parent_pte = parent_pte;
 845                         return;
 846                 }
 847                 sp->multimapped = 1;
 848                 pte_chain = mmu_alloc_pte_chain(vcpu);
 849                 INIT_HLIST_HEAD(&sp->parent_ptes);
 850                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 851                 pte_chain->parent_ptes[0] = old;
 852         }
 853         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 854                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 855                         continue;
 856                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 857                         if (!pte_chain->parent_ptes[i]) {
 858                                 pte_chain->parent_ptes[i] = parent_pte;
 859                                 return;
 860                         }
 861         }
 862         pte_chain = mmu_alloc_pte_chain(vcpu);
 863         BUG_ON(!pte_chain);
 864         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 865         pte_chain->parent_ptes[0] = parent_pte;
 866 }
 867
 868 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 869                                        u64 *parent_pte)
 870 {
 871         struct kvm_pte_chain *pte_chain;
 872         struct hlist_node *node;
 873         int i;
 874
 875         if (!sp->multimapped) {
 876                 BUG_ON(sp->parent_pte != parent_pte);
 877                 sp->parent_pte = NULL;
 878                 return;
 879         }
 880         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 881                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 882                         if (!pte_chain->parent_ptes[i])
 883                                 break;
 884                         if (pte_chain->parent_ptes[i] != parent_pte)
 885                                 continue;
 886                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 887                                 && pte_chain->parent_ptes[i + 1]) {
 888                                 pte_chain->parent_ptes[i]
 889                                         = pte_chain->parent_ptes[i + 1];
 890                                 ++i;
 891                         }
 892                         pte_chain->parent_ptes[i] = NULL;
 893                         if (i == 0) {
 894                                 hlist_del(&pte_chain->link);
 895                                 mmu_free_pte_chain(pte_chain);
 896                                 if (hlist_empty(&sp->parent_ptes)) {
 897                                         sp->multimapped = 0;
 898                                         sp->parent_pte = NULL;
 899                                 }
 900                         }
 901                         return;
 902                 }
 903         BUG();
 904 }
 905
 906
 907 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 908                             mmu_parent_walk_fn fn)
 909 {
 910         struct kvm_pte_chain *pte_chain;
 911         struct hlist_node *node;
 912         struct kvm_mmu_page *parent_sp;
 913         int i;
 914
 915         if (!sp->multimapped && sp->parent_pte) {
 916                 parent_sp = page_header(__pa(sp->parent_pte));
 917                 fn(vcpu, parent_sp);
 918                 mmu_parent_walk(vcpu, parent_sp, fn);
 919                 return;
 920         }
 921         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 922                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 923                         if (!pte_chain->parent_ptes[i])
 924                                 break;
 925                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 926                         fn(vcpu, parent_sp);
 927                         mmu_parent_walk(vcpu, parent_sp, fn);
 928                 }
 929 }
 930
 931 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 932 {
 933         unsigned int index;
 934         struct kvm_mmu_page *sp = page_header(__pa(spte));
 935
 936         index = spte - sp->spt;
 937         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
 938                 sp->unsync_children++;
 939         WARN_ON(!sp->unsync_children);
 940 }
 941
 942 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 943 {
 944         struct kvm_pte_chain *pte_chain;
 945         struct hlist_node *node;
 946         int i;
 947
 948         if (!sp->parent_pte)
 949                 return;
 950
 951         if (!sp->multimapped) {
 952                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 953                 return;
 954         }
 955
 956         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 957                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 958                         if (!pte_chain->parent_ptes[i])
 959                                 break;
 960                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 961                 }
 962 }
 963
 964 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 965 {
 966         kvm_mmu_update_parents_unsync(sp);
 967         return 1;
 968 }
 969
 970 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 971                                         struct kvm_mmu_page *sp)
 972 {
 973         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 974         kvm_mmu_update_parents_unsync(sp);
 975 }
 976
 977 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 978                                     struct kvm_mmu_page *sp)
 979 {
 980         int i;
 981
 982         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 983                 sp->spt[i] = shadow_trap_nonpresent_pte;
 984 }
 985
 986 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
 987                                struct kvm_mmu_page *sp)
 988 {
 989         return 1;
 990 }
 991
 992 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
 993 {
 994 }
 995
 996 #define KVM_PAGE_ARRAY_NR 16
 997
 998 struct kvm_mmu_pages {
 999         struct mmu_page_and_offset {
1000                 struct kvm_mmu_page *sp;
1001                 unsigned int idx;
1002         } page[KVM_PAGE_ARRAY_NR];
1003         unsigned int nr;
1004 };
1005
1006 #define for_each_unsync_children(bitmap, idx)           \
1007         for (idx = find_first_bit(bitmap, 512);         \
1008              idx < 512;                                 \
1009              idx = find_next_bit(bitmap, 512, idx+1))
1010
1011 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1012                          int idx)
1013 {
1014         int i;
1015
1016         if (sp->unsync)
1017                 for (i=0; i < pvec->nr; i++)
1018                         if (pvec->page[i].sp == sp)
1019                                 return 0;
1020
1021         pvec->page[pvec->nr].sp = sp;
1022         pvec->page[pvec->nr].idx = idx;
1023         pvec->nr++;
1024         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1025 }
1026
1027 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1028                            struct kvm_mmu_pages *pvec)
1029 {
1030         int i, ret, nr_unsync_leaf = 0;
1031
1032         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1033                 u64 ent = sp->spt[i];
1034
1035                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1036                         struct kvm_mmu_page *child;
1037                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1038
1039                         if (child->unsync_children) {
1040                                 if (mmu_pages_add(pvec, child, i))
1041                                         return -ENOSPC;
1042
1043                                 ret = __mmu_unsync_walk(child, pvec);
1044                                 if (!ret)
1045                                         __clear_bit(i, sp->unsync_child_bitmap);
1046                                 else if (ret > 0)
1047                                         nr_unsync_leaf += ret;
1048                                 else
1049                                         return ret;
1050                         }
1051
1052                         if (child->unsync) {
1053                                 nr_unsync_leaf++;
1054                                 if (mmu_pages_add(pvec, child, i))
1055                                         return -ENOSPC;
1056                         }
1057                 }
1058         }
1059
1060         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1061                 sp->unsync_children = 0;
1062
1063         return nr_unsync_leaf;
1064 }
1065
1066 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1067                            struct kvm_mmu_pages *pvec)
1068 {
1069         if (!sp->unsync_children)
1070                 return 0;
1071
1072         mmu_pages_add(pvec, sp, 0);
1073         return __mmu_unsync_walk(sp, pvec);
1074 }
1075
1076 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1077 {
1078         unsigned index;
1079         struct hlist_head *bucket;
1080         struct kvm_mmu_page *sp;
1081         struct hlist_node *node;
1082
1083         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1084         index = kvm_page_table_hashfn(gfn);
1085         bucket = &kvm->arch.mmu_page_hash[index];
1086         hlist_for_each_entry(sp, node, bucket, hash_link)
1087                 if (sp->gfn == gfn && !sp->role.direct
1088                     && !sp->role.invalid) {
1089                         pgprintk("%s: found role %x\n",
1090                                  __func__, sp->role.word);
1091                         return sp;
1092                 }
1093         return NULL;
1094 }
1095
1096 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1097 {
1098         WARN_ON(!sp->unsync);
1099         sp->unsync = 0;
1100         --kvm->stat.mmu_unsync;
1101 }
1102
1103 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1104
1105 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1106 {
1107         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1108                 kvm_mmu_zap_page(vcpu->kvm, sp);
1109                 return 1;
1110         }
1111
1112         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1113                 kvm_flush_remote_tlbs(vcpu->kvm);
1114         kvm_unlink_unsync_page(vcpu->kvm, sp);
1115         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1116                 kvm_mmu_zap_page(vcpu->kvm, sp);
1117                 return 1;
1118         }
1119
1120         kvm_mmu_flush_tlb(vcpu);
1121         return 0;
1122 }
1123
1124 struct mmu_page_path {
1125         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1126         unsigned int idx[PT64_ROOT_LEVEL-1];
1127 };
1128
1129 #define for_each_sp(pvec, sp, parents, i)                       \
1130                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1131                         sp = pvec.page[i].sp;                   \
1132                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1133                         i = mmu_pages_next(&pvec, &parents, i))
1134
1135 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1136                           struct mmu_page_path *parents,
1137                           int i)
1138 {
1139         int n;
1140
1141         for (n = i+1; n < pvec->nr; n++) {
1142                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1143
1144                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1145                         parents->idx[0] = pvec->page[n].idx;
1146                         return n;
1147                 }
1148
1149                 parents->parent[sp->role.level-2] = sp;
1150                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1151         }
1152
1153         return n;
1154 }
1155
1156 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1157 {
1158         struct kvm_mmu_page *sp;
1159         unsigned int level = 0;
1160
1161         do {
1162                 unsigned int idx = parents->idx[level];
1163
1164                 sp = parents->parent[level];
1165                 if (!sp)
1166                         return;
1167
1168                 --sp->unsync_children;
1169                 WARN_ON((int)sp->unsync_children < 0);
1170                 __clear_bit(idx, sp->unsync_child_bitmap);
1171                 level++;
1172         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1173 }
1174
1175 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1176                                struct mmu_page_path *parents,
1177                                struct kvm_mmu_pages *pvec)
1178 {
1179         parents->parent[parent->role.level-1] = NULL;
1180         pvec->nr = 0;
1181 }
1182
1183 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1184                               struct kvm_mmu_page *parent)
1185 {
1186         int i;
1187         struct kvm_mmu_page *sp;
1188         struct mmu_page_path parents;
1189         struct kvm_mmu_pages pages;
1190
1191         kvm_mmu_pages_init(parent, &parents, &pages);
1192         while (mmu_unsync_walk(parent, &pages)) {
1193                 int protected = 0;
1194
1195                 for_each_sp(pages, sp, parents, i)
1196                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1197
1198                 if (protected)
1199                         kvm_flush_remote_tlbs(vcpu->kvm);
1200
1201                 for_each_sp(pages, sp, parents, i) {
1202                         kvm_sync_page(vcpu, sp);
1203                         mmu_pages_clear_parents(&parents);
1204                 }
1205                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1206                 kvm_mmu_pages_init(parent, &parents, &pages);
1207         }
1208 }
1209
1210 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1211                                              gfn_t gfn,
1212                                              gva_t gaddr,
1213                                              unsigned level,
1214                                              int direct,
1215                                              unsigned access,
1216                                              u64 *parent_pte)
1217 {
1218         union kvm_mmu_page_role role;
1219         unsigned index;
1220         unsigned quadrant;
1221         struct hlist_head *bucket;
1222         struct kvm_mmu_page *sp;
1223         struct hlist_node *node, *tmp;
1224
1225         role = vcpu->arch.mmu.base_role;
1226         role.level = level;
1227         role.direct = direct;
1228         role.access = access;
1229         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1230                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1231                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1232                 role.quadrant = quadrant;
1233         }
1234         pgprintk("%s: looking gfn %lx role %x\n", __func__,
1235                  gfn, role.word);
1236         index = kvm_page_table_hashfn(gfn);
1237         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1238         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1239                 if (sp->gfn == gfn) {
1240                         if (sp->unsync)
1241                                 if (kvm_sync_page(vcpu, sp))
1242                                         continue;
1243
1244                         if (sp->role.word != role.word)
1245                                 continue;
1246
1247                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1248                         if (sp->unsync_children) {
1249                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1250                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1251                         }
1252                         pgprintk("%s: found\n", __func__);
1253                         return sp;
1254                 }
1255         ++vcpu->kvm->stat.mmu_cache_miss;
1256         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1257         if (!sp)
1258                 return sp;
1259         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1260         sp->gfn = gfn;
1261         sp->role = role;
1262         hlist_add_head(&sp->hash_link, bucket);
1263         if (!direct) {
1264                 if (rmap_write_protect(vcpu->kvm, gfn))
1265                         kvm_flush_remote_tlbs(vcpu->kvm);
1266                 account_shadowed(vcpu->kvm, gfn);
1267         }
1268         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1269                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1270         else
1271                 nonpaging_prefetch_page(vcpu, sp);
1272         return sp;
1273 }
1274
1275 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1276                              struct kvm_vcpu *vcpu, u64 addr)
1277 {
1278         iterator->addr = addr;
1279         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1280         iterator->level = vcpu->arch.mmu.shadow_root_level;
1281         if (iterator->level == PT32E_ROOT_LEVEL) {
1282                 iterator->shadow_addr
1283                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1284                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1285                 --iterator->level;
1286                 if (!iterator->shadow_addr)
1287                         iterator->level = 0;
1288         }
1289 }
1290
1291 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1292 {
1293         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1294                 return false;
1295         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1296         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1297         return true;
1298 }
1299
1300 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1301 {
1302         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1303         --iterator->level;
1304 }
1305
1306 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1307                                          struct kvm_mmu_page *sp)
1308 {
1309         unsigned i;
1310         u64 *pt;
1311         u64 ent;
1312
1313         pt = sp->spt;
1314
1315         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1316                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1317                         if (is_shadow_present_pte(pt[i]))
1318                                 rmap_remove(kvm, &pt[i]);
1319                         pt[i] = shadow_trap_nonpresent_pte;
1320                 }
1321                 return;
1322         }
1323
1324         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1325                 ent = pt[i];
1326
1327                 if (is_shadow_present_pte(ent)) {
1328                         if (!is_large_pte(ent)) {
1329                                 ent &= PT64_BASE_ADDR_MASK;
1330                                 mmu_page_remove_parent_pte(page_header(ent),
1331                                                            &pt[i]);
1332                         } else {
1333                                 --kvm->stat.lpages;
1334                                 rmap_remove(kvm, &pt[i]);
1335                         }
1336                 }
1337                 pt[i] = shadow_trap_nonpresent_pte;
1338         }
1339 }
1340
1341 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1342 {
1343         mmu_page_remove_parent_pte(sp, parent_pte);
1344 }
1345
1346 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1347 {
1348         int i;
1349
1350         for (i = 0; i < KVM_MAX_VCPUS; ++i)
1351                 if (kvm->vcpus[i])
1352                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
1353 }
1354
1355 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1356 {
1357         u64 *parent_pte;
1358
1359         while (sp->multimapped || sp->parent_pte) {
1360                 if (!sp->multimapped)
1361                         parent_pte = sp->parent_pte;
1362                 else {
1363                         struct kvm_pte_chain *chain;
1364
1365                         chain = container_of(sp->parent_ptes.first,
1366                                              struct kvm_pte_chain, link);
1367                         parent_pte = chain->parent_ptes[0];
1368                 }
1369                 BUG_ON(!parent_pte);
1370                 kvm_mmu_put_page(sp, parent_pte);
1371                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1372         }
1373 }
1374
1375 static int mmu_zap_unsync_children(struct kvm *kvm,
1376                                    struct kvm_mmu_page *parent)
1377 {
1378         int i, zapped = 0;
1379         struct mmu_page_path parents;
1380         struct kvm_mmu_pages pages;
1381
1382         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1383                 return 0;
1384
1385         kvm_mmu_pages_init(parent, &parents, &pages);
1386         while (mmu_unsync_walk(parent, &pages)) {
1387                 struct kvm_mmu_page *sp;
1388
1389                 for_each_sp(pages, sp, parents, i) {
1390                         kvm_mmu_zap_page(kvm, sp);
1391                         mmu_pages_clear_parents(&parents);
1392                 }
1393                 zapped += pages.nr;
1394                 kvm_mmu_pages_init(parent, &parents, &pages);
1395         }
1396
1397         return zapped;
1398 }
1399
1400 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1401 {
1402         int ret;
1403         ++kvm->stat.mmu_shadow_zapped;
1404         ret = mmu_zap_unsync_children(kvm, sp);
1405         kvm_mmu_page_unlink_children(kvm, sp);
1406         kvm_mmu_unlink_parents(kvm, sp);
1407         kvm_flush_remote_tlbs(kvm);
1408         if (!sp->role.invalid && !sp->role.direct)
1409                 unaccount_shadowed(kvm, sp->gfn);
1410         if (sp->unsync)
1411                 kvm_unlink_unsync_page(kvm, sp);
1412         if (!sp->root_count) {
1413                 hlist_del(&sp->hash_link);
1414                 kvm_mmu_free_page(kvm, sp);
1415         } else {
1416                 sp->role.invalid = 1;
1417                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1418                 kvm_reload_remote_mmus(kvm);
1419         }
1420         kvm_mmu_reset_last_pte_updated(kvm);
1421         return ret;
1422 }
1423
1424 /*
1425  * Changing the number of mmu pages allocated to the vm
1426  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1427  */
1428 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1429 {
1430         int used_pages;
1431
1432         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1433         used_pages = max(0, used_pages);
1434
1435         /*
1436          * If we set the number of mmu pages to be smaller be than the
1437          * number of actived pages , we must to free some mmu pages before we
1438          * change the value
1439          */
1440
1441         if (used_pages > kvm_nr_mmu_pages) {
1442                 while (used_pages > kvm_nr_mmu_pages) {
1443                         struct kvm_mmu_page *page;
1444
1445                         page = container_of(kvm->arch.active_mmu_pages.prev,
1446                                             struct kvm_mmu_page, link);
1447                         kvm_mmu_zap_page(kvm, page);
1448                         used_pages--;
1449                 }
1450                 kvm->arch.n_free_mmu_pages = 0;
1451         }
1452         else
1453                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1454                                          - kvm->arch.n_alloc_mmu_pages;
1455
1456         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1457 }
1458
1459 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1460 {
1461         unsigned index;
1462         struct hlist_head *bucket;
1463         struct kvm_mmu_page *sp;
1464         struct hlist_node *node, *n;
1465         int r;
1466
1467         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1468         r = 0;
1469         index = kvm_page_table_hashfn(gfn);
1470         bucket = &kvm->arch.mmu_page_hash[index];
1471         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1472                 if (sp->gfn == gfn && !sp->role.direct) {
1473                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1474                                  sp->role.word);
1475                         r = 1;
1476                         if (kvm_mmu_zap_page(kvm, sp))
1477                                 n = bucket->first;
1478                 }
1479         return r;
1480 }
1481
1482 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1483 {
1484         unsigned index;
1485         struct hlist_head *bucket;
1486         struct kvm_mmu_page *sp;
1487         struct hlist_node *node, *nn;
1488
1489         index = kvm_page_table_hashfn(gfn);
1490         bucket = &kvm->arch.mmu_page_hash[index];
1491         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1492                 if (sp->gfn == gfn && !sp->role.direct
1493                     && !sp->role.invalid) {
1494                         pgprintk("%s: zap %lx %x\n",
1495                                  __func__, gfn, sp->role.word);
1496                         kvm_mmu_zap_page(kvm, sp);
1497                 }
1498         }
1499 }
1500
1501 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1502 {
1503         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1504         struct kvm_mmu_page *sp = page_header(__pa(pte));
1505
1506         __set_bit(slot, sp->slot_bitmap);
1507 }
1508
1509 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1510 {
1511         int i;
1512         u64 *pt = sp->spt;
1513
1514         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1515                 return;
1516
1517         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1518                 if (pt[i] == shadow_notrap_nonpresent_pte)
1519                         set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1520         }
1521 }
1522
1523 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1524 {
1525         struct page *page;
1526
1527         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1528
1529         if (gpa == UNMAPPED_GVA)
1530                 return NULL;
1531
1532         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1533
1534         return page;
1535 }
1536
1537 /*
1538  * The function is based on mtrr_type_lookup() in
1539  * arch/x86/kernel/cpu/mtrr/generic.c
1540  */
1541 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1542                          u64 start, u64 end)
1543 {
1544         int i;
1545         u64 base, mask;
1546         u8 prev_match, curr_match;
1547         int num_var_ranges = KVM_NR_VAR_MTRR;
1548
1549         if (!mtrr_state->enabled)
1550                 return 0xFF;
1551
1552         /* Make end inclusive end, instead of exclusive */
1553         end--;
1554
1555         /* Look in fixed ranges. Just return the type as per start */
1556         if (mtrr_state->have_fixed && (start < 0x100000)) {
1557                 int idx;
1558
1559                 if (start < 0x80000) {
1560                         idx = 0;
1561                         idx += (start >> 16);
1562                         return mtrr_state->fixed_ranges[idx];
1563                 } else if (start < 0xC0000) {
1564                         idx = 1 * 8;
1565                         idx += ((start - 0x80000) >> 14);
1566                         return mtrr_state->fixed_ranges[idx];
1567                 } else if (start < 0x1000000) {
1568                         idx = 3 * 8;
1569                         idx += ((start - 0xC0000) >> 12);
1570                         return mtrr_state->fixed_ranges[idx];
1571                 }
1572         }
1573
1574         /*
1575          * Look in variable ranges
1576          * Look of multiple ranges matching this address and pick type
1577          * as per MTRR precedence
1578          */
1579         if (!(mtrr_state->enabled & 2))
1580                 return mtrr_state->def_type;
1581
1582         prev_match = 0xFF;
1583         for (i = 0; i < num_var_ranges; ++i) {
1584                 unsigned short start_state, end_state;
1585
1586                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1587                         continue;
1588
1589                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1590                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1591                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1592                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1593
1594                 start_state = ((start & mask) == (base & mask));
1595                 end_state = ((end & mask) == (base & mask));
1596                 if (start_state != end_state)
1597                         return 0xFE;
1598
1599                 if ((start & mask) != (base & mask))
1600                         continue;
1601
1602                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1603                 if (prev_match == 0xFF) {
1604                         prev_match = curr_match;
1605                         continue;
1606                 }
1607
1608                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1609                     curr_match == MTRR_TYPE_UNCACHABLE)
1610                         return MTRR_TYPE_UNCACHABLE;
1611
1612                 if ((prev_match == MTRR_TYPE_WRBACK &&
1613                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1614                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1615                      curr_match == MTRR_TYPE_WRBACK)) {
1616                         prev_match = MTRR_TYPE_WRTHROUGH;
1617                         curr_match = MTRR_TYPE_WRTHROUGH;
1618                 }
1619
1620                 if (prev_match != curr_match)
1621                         return MTRR_TYPE_UNCACHABLE;
1622         }
1623
1624         if (prev_match != 0xFF)
1625                 return prev_match;
1626
1627         return mtrr_state->def_type;
1628 }
1629
1630 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1631 {
1632         u8 mtrr;
1633
1634         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1635                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1636         if (mtrr == 0xfe || mtrr == 0xff)
1637                 mtrr = MTRR_TYPE_WRBACK;
1638         return mtrr;
1639 }
1640 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1641
1642 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1643 {
1644         unsigned index;
1645         struct hlist_head *bucket;
1646         struct kvm_mmu_page *s;
1647         struct hlist_node *node, *n;
1648
1649         index = kvm_page_table_hashfn(sp->gfn);
1650         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1651         /* don't unsync if pagetable is shadowed with multiple roles */
1652         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1653                 if (s->gfn != sp->gfn || s->role.direct)
1654                         continue;
1655                 if (s->role.word != sp->role.word)
1656                         return 1;
1657         }
1658         ++vcpu->kvm->stat.mmu_unsync;
1659         sp->unsync = 1;
1660
1661         kvm_mmu_mark_parents_unsync(vcpu, sp);
1662
1663         mmu_convert_notrap(sp);
1664         return 0;
1665 }
1666
1667 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1668                                   bool can_unsync)
1669 {
1670         struct kvm_mmu_page *shadow;
1671
1672         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1673         if (shadow) {
1674                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1675                         return 1;
1676                 if (shadow->unsync)
1677                         return 0;
1678                 if (can_unsync && oos_shadow)
1679                         return kvm_unsync_page(vcpu, shadow);
1680                 return 1;
1681         }
1682         return 0;
1683 }
1684
1685 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1686                     unsigned pte_access, int user_fault,
1687                     int write_fault, int dirty, int largepage,
1688                     gfn_t gfn, pfn_t pfn, bool speculative,
1689                     bool can_unsync)
1690 {
1691         u64 spte;
1692         int ret = 0;
1693
1694         /*
1695          * We don't set the accessed bit, since we sometimes want to see
1696          * whether the guest actually used the pte (in order to detect
1697          * demand paging).
1698          */
1699         spte = shadow_base_present_pte | shadow_dirty_mask;
1700         if (!speculative)
1701                 spte |= shadow_accessed_mask;
1702         if (!dirty)
1703                 pte_access &= ~ACC_WRITE_MASK;
1704         if (pte_access & ACC_EXEC_MASK)
1705                 spte |= shadow_x_mask;
1706         else
1707                 spte |= shadow_nx_mask;
1708         if (pte_access & ACC_USER_MASK)
1709                 spte |= shadow_user_mask;
1710         if (largepage)
1711                 spte |= PT_PAGE_SIZE_MASK;
1712         if (tdp_enabled)
1713                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1714                         kvm_is_mmio_pfn(pfn));
1715
1716         spte |= (u64)pfn << PAGE_SHIFT;
1717
1718         if ((pte_access & ACC_WRITE_MASK)
1719             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1720
1721                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1722                         ret = 1;
1723                         spte = shadow_trap_nonpresent_pte;
1724                         goto set_pte;
1725                 }
1726
1727                 spte |= PT_WRITABLE_MASK;
1728
1729                 /*
1730                  * Optimization: for pte sync, if spte was writable the hash
1731                  * lookup is unnecessary (and expensive). Write protection
1732                  * is responsibility of mmu_get_page / kvm_sync_page.
1733                  * Same reasoning can be applied to dirty page accounting.
1734                  */
1735                 if (!can_unsync && is_writeble_pte(*shadow_pte))
1736                         goto set_pte;
1737
1738                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1739                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1740                                  __func__, gfn);
1741                         ret = 1;
1742                         pte_access &= ~ACC_WRITE_MASK;
1743                         if (is_writeble_pte(spte))
1744                                 spte &= ~PT_WRITABLE_MASK;
1745                 }
1746         }
1747
1748         if (pte_access & ACC_WRITE_MASK)
1749                 mark_page_dirty(vcpu->kvm, gfn);
1750
1751 set_pte:
1752         set_shadow_pte(shadow_pte, spte);
1753         return ret;
1754 }
1755
1756 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1757                          unsigned pt_access, unsigned pte_access,
1758                          int user_fault, int write_fault, int dirty,
1759                          int *ptwrite, int largepage, gfn_t gfn,
1760                          pfn_t pfn, bool speculative)
1761 {
1762         int was_rmapped = 0;
1763         int was_writeble = is_writeble_pte(*shadow_pte);
1764         int rmap_count;
1765
1766         pgprintk("%s: spte %llx access %x write_fault %d"
1767                  " user_fault %d gfn %lx\n",
1768                  __func__, *shadow_pte, pt_access,
1769                  write_fault, user_fault, gfn);
1770
1771         if (is_rmap_pte(*shadow_pte)) {
1772                 /*
1773                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1774                  * the parent of the now unreachable PTE.
1775                  */
1776                 if (largepage && !is_large_pte(*shadow_pte)) {
1777                         struct kvm_mmu_page *child;
1778                         u64 pte = *shadow_pte;
1779
1780                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1781                         mmu_page_remove_parent_pte(child, shadow_pte);
1782                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1783                         pgprintk("hfn old %lx new %lx\n",
1784                                  spte_to_pfn(*shadow_pte), pfn);
1785                         rmap_remove(vcpu->kvm, shadow_pte);
1786                 } else
1787                         was_rmapped = 1;
1788         }
1789         if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1790                       dirty, largepage, gfn, pfn, speculative, true)) {
1791                 if (write_fault)
1792                         *ptwrite = 1;
1793                 kvm_x86_ops->tlb_flush(vcpu);
1794         }
1795
1796         pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1797         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1798                  is_large_pte(*shadow_pte)? "2MB" : "4kB",
1799                  is_present_pte(*shadow_pte)?"RW":"R", gfn,
1800                  *shadow_pte, shadow_pte);
1801         if (!was_rmapped && is_large_pte(*shadow_pte))
1802                 ++vcpu->kvm->stat.lpages;
1803
1804         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1805         if (!was_rmapped) {
1806                 rmap_count = rmap_add(vcpu, shadow_pte, gfn, largepage);
1807                 if (!is_rmap_pte(*shadow_pte))
1808                         kvm_release_pfn_clean(pfn);
1809                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1810                         rmap_recycle(vcpu, gfn, largepage);
1811         } else {
1812                 if (was_writeble)
1813                         kvm_release_pfn_dirty(pfn);
1814                 else
1815                         kvm_release_pfn_clean(pfn);
1816         }
1817         if (speculative) {
1818                 vcpu->arch.last_pte_updated = shadow_pte;
1819                 vcpu->arch.last_pte_gfn = gfn;
1820         }
1821 }
1822
1823 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1824 {
1825 }
1826
1827 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1828                         int largepage, gfn_t gfn, pfn_t pfn)
1829 {
1830         struct kvm_shadow_walk_iterator iterator;
1831         struct kvm_mmu_page *sp;
1832         int pt_write = 0;
1833         gfn_t pseudo_gfn;
1834
1835         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1836                 if (iterator.level == PT_PAGE_TABLE_LEVEL
1837                     || (largepage && iterator.level == PT_DIRECTORY_LEVEL)) {
1838                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1839                                      0, write, 1, &pt_write,
1840                                      largepage, gfn, pfn, false);
1841                         ++vcpu->stat.pf_fixed;
1842                         break;
1843                 }
1844
1845                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1846                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1847                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1848                                               iterator.level - 1,
1849                                               1, ACC_ALL, iterator.sptep);
1850                         if (!sp) {
1851                                 pgprintk("nonpaging_map: ENOMEM\n");
1852                                 kvm_release_pfn_clean(pfn);
1853                                 return -ENOMEM;
1854                         }
1855
1856                         set_shadow_pte(iterator.sptep,
1857                                        __pa(sp->spt)
1858                                        | PT_PRESENT_MASK | PT_WRITABLE_MASK
1859                                        | shadow_user_mask | shadow_x_mask);
1860                 }
1861         }
1862         return pt_write;
1863 }
1864
1865 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1866 {
1867         int r;
1868         int largepage = 0;
1869         pfn_t pfn;
1870         unsigned long mmu_seq;
1871
1872         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1873                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1874                 largepage = 1;
1875         }
1876
1877         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1878         smp_rmb();
1879         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1880
1881         /* mmio */
1882         if (is_error_pfn(pfn)) {
1883                 kvm_release_pfn_clean(pfn);
1884                 return 1;
1885         }
1886
1887         spin_lock(&vcpu->kvm->mmu_lock);
1888         if (mmu_notifier_retry(vcpu, mmu_seq))
1889                 goto out_unlock;
1890         kvm_mmu_free_some_pages(vcpu);
1891         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1892         spin_unlock(&vcpu->kvm->mmu_lock);
1893
1894
1895         return r;
1896
1897 out_unlock:
1898         spin_unlock(&vcpu->kvm->mmu_lock);
1899         kvm_release_pfn_clean(pfn);
1900         return 0;
1901 }
1902
1903
1904 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1905 {
1906         int i;
1907         struct kvm_mmu_page *sp;
1908
1909         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1910                 return;
1911         spin_lock(&vcpu->kvm->mmu_lock);
1912         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1913                 hpa_t root = vcpu->arch.mmu.root_hpa;
1914
1915                 sp = page_header(root);
1916                 --sp->root_count;
1917                 if (!sp->root_count && sp->role.invalid)
1918                         kvm_mmu_zap_page(vcpu->kvm, sp);
1919                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1920                 spin_unlock(&vcpu->kvm->mmu_lock);
1921                 return;
1922         }
1923         for (i = 0; i < 4; ++i) {
1924                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1925
1926                 if (root) {
1927                         root &= PT64_BASE_ADDR_MASK;
1928                         sp = page_header(root);
1929                         --sp->root_count;
1930                         if (!sp->root_count && sp->role.invalid)
1931                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1932                 }
1933                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1934         }
1935         spin_unlock(&vcpu->kvm->mmu_lock);
1936         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1937 }
1938
1939 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
1940 {
1941         int ret = 0;
1942
1943         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
1944                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
1945                 ret = 1;
1946         }
1947
1948         return ret;
1949 }
1950
1951 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
1952 {
1953         int i;
1954         gfn_t root_gfn;
1955         struct kvm_mmu_page *sp;
1956         int direct = 0;
1957
1958         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1959
1960         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1961                 hpa_t root = vcpu->arch.mmu.root_hpa;
1962
1963                 ASSERT(!VALID_PAGE(root));
1964                 if (tdp_enabled)
1965                         direct = 1;
1966                 if (mmu_check_root(vcpu, root_gfn))
1967                         return 1;
1968                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1969                                       PT64_ROOT_LEVEL, direct,
1970                                       ACC_ALL, NULL);
1971                 root = __pa(sp->spt);
1972                 ++sp->root_count;
1973                 vcpu->arch.mmu.root_hpa = root;
1974                 return 0;
1975         }
1976         direct = !is_paging(vcpu);
1977         if (tdp_enabled)
1978                 direct = 1;
1979         for (i = 0; i < 4; ++i) {
1980                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1981
1982                 ASSERT(!VALID_PAGE(root));
1983                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1984                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1985                                 vcpu->arch.mmu.pae_root[i] = 0;
1986                                 continue;
1987                         }
1988                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1989                 } else if (vcpu->arch.mmu.root_level == 0)
1990                         root_gfn = 0;
1991                 if (mmu_check_root(vcpu, root_gfn))
1992                         return 1;
1993                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1994                                       PT32_ROOT_LEVEL, direct,
1995                                       ACC_ALL, NULL);
1996                 root = __pa(sp->spt);
1997                 ++sp->root_count;
1998                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1999         }
2000         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2001         return 0;
2002 }
2003
2004 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2005 {
2006         int i;
2007         struct kvm_mmu_page *sp;
2008
2009         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2010                 return;
2011         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2012                 hpa_t root = vcpu->arch.mmu.root_hpa;
2013                 sp = page_header(root);
2014                 mmu_sync_children(vcpu, sp);
2015                 return;
2016         }
2017         for (i = 0; i < 4; ++i) {
2018                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2019
2020                 if (root && VALID_PAGE(root)) {
2021                         root &= PT64_BASE_ADDR_MASK;
2022                         sp = page_header(root);
2023                         mmu_sync_children(vcpu, sp);
2024                 }
2025         }
2026 }
2027
2028 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2029 {
2030         spin_lock(&vcpu->kvm->mmu_lock);
2031         mmu_sync_roots(vcpu);
2032         spin_unlock(&vcpu->kvm->mmu_lock);
2033 }
2034
2035 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2036 {
2037         return vaddr;
2038 }
2039
2040 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2041                                 u32 error_code)
2042 {
2043         gfn_t gfn;
2044         int r;
2045
2046         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2047         r = mmu_topup_memory_caches(vcpu);
2048         if (r)
2049                 return r;
2050
2051         ASSERT(vcpu);
2052         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2053
2054         gfn = gva >> PAGE_SHIFT;
2055
2056         return nonpaging_map(vcpu, gva & PAGE_MASK,
2057                              error_code & PFERR_WRITE_MASK, gfn);
2058 }
2059
2060 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2061                                 u32 error_code)
2062 {
2063         pfn_t pfn;
2064         int r;
2065         int largepage = 0;
2066         gfn_t gfn = gpa >> PAGE_SHIFT;
2067         unsigned long mmu_seq;
2068
2069         ASSERT(vcpu);
2070         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2071
2072         r = mmu_topup_memory_caches(vcpu);
2073         if (r)
2074                 return r;
2075
2076         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
2077                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2078                 largepage = 1;
2079         }
2080         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2081         smp_rmb();
2082         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2083         if (is_error_pfn(pfn)) {
2084                 kvm_release_pfn_clean(pfn);
2085                 return 1;
2086         }
2087         spin_lock(&vcpu->kvm->mmu_lock);
2088         if (mmu_notifier_retry(vcpu, mmu_seq))
2089                 goto out_unlock;
2090         kvm_mmu_free_some_pages(vcpu);
2091         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2092                          largepage, gfn, pfn);
2093         spin_unlock(&vcpu->kvm->mmu_lock);
2094
2095         return r;
2096
2097 out_unlock:
2098         spin_unlock(&vcpu->kvm->mmu_lock);
2099         kvm_release_pfn_clean(pfn);
2100         return 0;
2101 }
2102
2103 static void nonpaging_free(struct kvm_vcpu *vcpu)
2104 {
2105         mmu_free_roots(vcpu);
2106 }
2107
2108 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2109 {
2110         struct kvm_mmu *context = &vcpu->arch.mmu;
2111
2112         context->new_cr3 = nonpaging_new_cr3;
2113         context->page_fault = nonpaging_page_fault;
2114         context->gva_to_gpa = nonpaging_gva_to_gpa;
2115         context->free = nonpaging_free;
2116         context->prefetch_page = nonpaging_prefetch_page;
2117         context->sync_page = nonpaging_sync_page;
2118         context->invlpg = nonpaging_invlpg;
2119         context->root_level = 0;
2120         context->shadow_root_level = PT32E_ROOT_LEVEL;
2121         context->root_hpa = INVALID_PAGE;
2122         return 0;
2123 }
2124
2125 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2126 {
2127         ++vcpu->stat.tlb_flush;
2128         kvm_x86_ops->tlb_flush(vcpu);
2129 }
2130
2131 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2132 {
2133         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2134         mmu_free_roots(vcpu);
2135 }
2136
2137 static void inject_page_fault(struct kvm_vcpu *vcpu,
2138                               u64 addr,
2139                               u32 err_code)
2140 {
2141         kvm_inject_page_fault(vcpu, addr, err_code);
2142 }
2143
2144 static void paging_free(struct kvm_vcpu *vcpu)
2145 {
2146         nonpaging_free(vcpu);
2147 }
2148
2149 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2150 {
2151         int bit7;
2152
2153         bit7 = (gpte >> 7) & 1;
2154         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2155 }
2156
2157 #define PTTYPE 64
2158 #include "paging_tmpl.h"
2159 #undef PTTYPE
2160
2161 #define PTTYPE 32
2162 #include "paging_tmpl.h"
2163 #undef PTTYPE
2164
2165 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2166 {
2167         struct kvm_mmu *context = &vcpu->arch.mmu;
2168         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2169         u64 exb_bit_rsvd = 0;
2170
2171         if (!is_nx(vcpu))
2172                 exb_bit_rsvd = rsvd_bits(63, 63);
2173         switch (level) {
2174         case PT32_ROOT_LEVEL:
2175                 /* no rsvd bits for 2 level 4K page table entries */
2176                 context->rsvd_bits_mask[0][1] = 0;
2177                 context->rsvd_bits_mask[0][0] = 0;
2178                 if (is_cpuid_PSE36())
2179                         /* 36bits PSE 4MB page */
2180                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2181                 else
2182                         /* 32 bits PSE 4MB page */
2183                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2184                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2185                 break;
2186         case PT32E_ROOT_LEVEL:
2187                 context->rsvd_bits_mask[0][2] =
2188                         rsvd_bits(maxphyaddr, 63) |
2189                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2190                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2191                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2192                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2193                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2194                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2195                         rsvd_bits(maxphyaddr, 62) |
2196                         rsvd_bits(13, 20);              /* large page */
2197                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2198                 break;
2199         case PT64_ROOT_LEVEL:
2200                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2201                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2202                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2203                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2204                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2205                         rsvd_bits(maxphyaddr, 51);
2206                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2207                         rsvd_bits(maxphyaddr, 51);
2208                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2209                 context->rsvd_bits_mask[1][2] = context->rsvd_bits_mask[0][2];
2210                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2211                         rsvd_bits(maxphyaddr, 51) |
2212                         rsvd_bits(13, 20);              /* large page */
2213                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2214                 break;
2215         }
2216 }
2217
2218 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2219 {
2220         struct kvm_mmu *context = &vcpu->arch.mmu;
2221
2222         ASSERT(is_pae(vcpu));
2223         context->new_cr3 = paging_new_cr3;
2224         context->page_fault = paging64_page_fault;
2225         context->gva_to_gpa = paging64_gva_to_gpa;
2226         context->prefetch_page = paging64_prefetch_page;
2227         context->sync_page = paging64_sync_page;
2228         context->invlpg = paging64_invlpg;
2229         context->free = paging_free;
2230         context->root_level = level;
2231         context->shadow_root_level = level;
2232         context->root_hpa = INVALID_PAGE;
2233         return 0;
2234 }
2235
2236 static int paging64_init_context(struct kvm_vcpu *vcpu)
2237 {
2238         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2239         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2240 }
2241
2242 static int paging32_init_context(struct kvm_vcpu *vcpu)
2243 {
2244         struct kvm_mmu *context = &vcpu->arch.mmu;
2245
2246         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2247         context->new_cr3 = paging_new_cr3;
2248         context->page_fault = paging32_page_fault;
2249         context->gva_to_gpa = paging32_gva_to_gpa;
2250         context->free = paging_free;
2251         context->prefetch_page = paging32_prefetch_page;
2252         context->sync_page = paging32_sync_page;
2253         context->invlpg = paging32_invlpg;
2254         context->root_level = PT32_ROOT_LEVEL;
2255         context->shadow_root_level = PT32E_ROOT_LEVEL;
2256         context->root_hpa = INVALID_PAGE;
2257         return 0;
2258 }
2259
2260 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2261 {
2262         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2263         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2264 }
2265
2266 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2267 {
2268         struct kvm_mmu *context = &vcpu->arch.mmu;
2269
2270         context->new_cr3 = nonpaging_new_cr3;
2271         context->page_fault = tdp_page_fault;
2272         context->free = nonpaging_free;
2273         context->prefetch_page = nonpaging_prefetch_page;
2274         context->sync_page = nonpaging_sync_page;
2275         context->invlpg = nonpaging_invlpg;
2276         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2277         context->root_hpa = INVALID_PAGE;
2278
2279         if (!is_paging(vcpu)) {
2280                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2281                 context->root_level = 0;
2282         } else if (is_long_mode(vcpu)) {
2283                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2284                 context->gva_to_gpa = paging64_gva_to_gpa;
2285                 context->root_level = PT64_ROOT_LEVEL;
2286         } else if (is_pae(vcpu)) {
2287                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2288                 context->gva_to_gpa = paging64_gva_to_gpa;
2289                 context->root_level = PT32E_ROOT_LEVEL;
2290         } else {
2291                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2292                 context->gva_to_gpa = paging32_gva_to_gpa;
2293                 context->root_level = PT32_ROOT_LEVEL;
2294         }
2295
2296         return 0;
2297 }
2298
2299 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2300 {
2301         int r;
2302
2303         ASSERT(vcpu);
2304         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2305
2306         if (!is_paging(vcpu))
2307                 r = nonpaging_init_context(vcpu);
2308         else if (is_long_mode(vcpu))
2309                 r = paging64_init_context(vcpu);
2310         else if (is_pae(vcpu))
2311                 r = paging32E_init_context(vcpu);
2312         else
2313                 r = paging32_init_context(vcpu);
2314
2315         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2316
2317         return r;
2318 }
2319
2320 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2321 {
2322         vcpu->arch.update_pte.pfn = bad_pfn;
2323
2324         if (tdp_enabled)
2325                 return init_kvm_tdp_mmu(vcpu);
2326         else
2327                 return init_kvm_softmmu(vcpu);
2328 }
2329
2330 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2331 {
2332         ASSERT(vcpu);
2333         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2334                 vcpu->arch.mmu.free(vcpu);
2335                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2336         }
2337 }
2338
2339 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2340 {
2341         destroy_kvm_mmu(vcpu);
2342         return init_kvm_mmu(vcpu);
2343 }
2344 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2345
2346 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2347 {
2348         int r;
2349
2350         r = mmu_topup_memory_caches(vcpu);
2351         if (r)
2352                 goto out;
2353         spin_lock(&vcpu->kvm->mmu_lock);
2354         kvm_mmu_free_some_pages(vcpu);
2355         r = mmu_alloc_roots(vcpu);
2356         mmu_sync_roots(vcpu);
2357         spin_unlock(&vcpu->kvm->mmu_lock);
2358         if (r)
2359                 goto out;
2360         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2361         kvm_mmu_flush_tlb(vcpu);
2362 out:
2363         return r;
2364 }
2365 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2366
2367 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2368 {
2369         mmu_free_roots(vcpu);
2370 }
2371
2372 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2373                                   struct kvm_mmu_page *sp,
2374                                   u64 *spte)
2375 {
2376         u64 pte;
2377         struct kvm_mmu_page *child;
2378
2379         pte = *spte;
2380         if (is_shadow_present_pte(pte)) {
2381                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2382                     is_large_pte(pte))
2383                         rmap_remove(vcpu->kvm, spte);
2384                 else {
2385                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2386                         mmu_page_remove_parent_pte(child, spte);
2387                 }
2388         }
2389         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2390         if (is_large_pte(pte))
2391                 --vcpu->kvm->stat.lpages;
2392 }
2393
2394 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2395                                   struct kvm_mmu_page *sp,
2396                                   u64 *spte,
2397                                   const void *new)
2398 {
2399         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2400                 if (!vcpu->arch.update_pte.largepage ||
2401                     sp->role.glevels == PT32_ROOT_LEVEL) {
2402                         ++vcpu->kvm->stat.mmu_pde_zapped;
2403                         return;
2404                 }
2405         }
2406
2407         ++vcpu->kvm->stat.mmu_pte_updated;
2408         if (sp->role.glevels == PT32_ROOT_LEVEL)
2409                 paging32_update_pte(vcpu, sp, spte, new);
2410         else
2411                 paging64_update_pte(vcpu, sp, spte, new);
2412 }
2413
2414 static bool need_remote_flush(u64 old, u64 new)
2415 {
2416         if (!is_shadow_present_pte(old))
2417                 return false;
2418         if (!is_shadow_present_pte(new))
2419                 return true;
2420         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2421                 return true;
2422         old ^= PT64_NX_MASK;
2423         new ^= PT64_NX_MASK;
2424         return (old & ~new & PT64_PERM_MASK) != 0;
2425 }
2426
2427 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2428 {
2429         if (need_remote_flush(old, new))
2430                 kvm_flush_remote_tlbs(vcpu->kvm);
2431         else
2432                 kvm_mmu_flush_tlb(vcpu);
2433 }
2434
2435 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2436 {
2437         u64 *spte = vcpu->arch.last_pte_updated;
2438
2439         return !!(spte && (*spte & shadow_accessed_mask));
2440 }
2441
2442 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2443                                           const u8 *new, int bytes)
2444 {
2445         gfn_t gfn;
2446         int r;
2447         u64 gpte = 0;
2448         pfn_t pfn;
2449
2450         vcpu->arch.update_pte.largepage = 0;
2451
2452         if (bytes != 4 && bytes != 8)
2453                 return;
2454
2455         /*
2456          * Assume that the pte write on a page table of the same type
2457          * as the current vcpu paging mode.  This is nearly always true
2458          * (might be false while changing modes).  Note it is verified later
2459          * by update_pte().
2460          */
2461         if (is_pae(vcpu)) {
2462                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2463                 if ((bytes == 4) && (gpa % 4 == 0)) {
2464                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2465                         if (r)
2466                                 return;
2467                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2468                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2469                         memcpy((void *)&gpte, new, 8);
2470                 }
2471         } else {
2472                 if ((bytes == 4) && (gpa % 4 == 0))
2473                         memcpy((void *)&gpte, new, 4);
2474         }
2475         if (!is_present_pte(gpte))
2476                 return;
2477         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2478
2479         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2480                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2481                 vcpu->arch.update_pte.largepage = 1;
2482         }
2483         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2484         smp_rmb();
2485         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2486
2487         if (is_error_pfn(pfn)) {
2488                 kvm_release_pfn_clean(pfn);
2489                 return;
2490         }
2491         vcpu->arch.update_pte.gfn = gfn;
2492         vcpu->arch.update_pte.pfn = pfn;
2493 }
2494
2495 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2496 {
2497         u64 *spte = vcpu->arch.last_pte_updated;
2498
2499         if (spte
2500             && vcpu->arch.last_pte_gfn == gfn
2501             && shadow_accessed_mask
2502             && !(*spte & shadow_accessed_mask)
2503             && is_shadow_present_pte(*spte))
2504                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2505 }
2506
2507 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2508                        const u8 *new, int bytes,
2509                        bool guest_initiated)
2510 {
2511         gfn_t gfn = gpa >> PAGE_SHIFT;
2512         struct kvm_mmu_page *sp;
2513         struct hlist_node *node, *n;
2514         struct hlist_head *bucket;
2515         unsigned index;
2516         u64 entry, gentry;
2517         u64 *spte;
2518         unsigned offset = offset_in_page(gpa);
2519         unsigned pte_size;
2520         unsigned page_offset;
2521         unsigned misaligned;
2522         unsigned quadrant;
2523         int level;
2524         int flooded = 0;
2525         int npte;
2526         int r;
2527
2528         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2529         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2530         spin_lock(&vcpu->kvm->mmu_lock);
2531         kvm_mmu_access_page(vcpu, gfn);
2532         kvm_mmu_free_some_pages(vcpu);
2533         ++vcpu->kvm->stat.mmu_pte_write;
2534         kvm_mmu_audit(vcpu, "pre pte write");
2535         if (guest_initiated) {
2536                 if (gfn == vcpu->arch.last_pt_write_gfn
2537                     && !last_updated_pte_accessed(vcpu)) {
2538                         ++vcpu->arch.last_pt_write_count;
2539                         if (vcpu->arch.last_pt_write_count >= 3)
2540                                 flooded = 1;
2541                 } else {
2542                         vcpu->arch.last_pt_write_gfn = gfn;
2543                         vcpu->arch.last_pt_write_count = 1;
2544                         vcpu->arch.last_pte_updated = NULL;
2545                 }
2546         }
2547         index = kvm_page_table_hashfn(gfn);
2548         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2549         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2550                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2551                         continue;
2552                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2553                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2554                 misaligned |= bytes < 4;
2555                 if (misaligned || flooded) {
2556                         /*
2557                          * Misaligned accesses are too much trouble to fix
2558                          * up; also, they usually indicate a page is not used
2559                          * as a page table.
2560                          *
2561                          * If we're seeing too many writes to a page,
2562                          * it may no longer be a page table, or we may be
2563                          * forking, in which case it is better to unmap the
2564                          * page.
2565                          */
2566                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2567                                  gpa, bytes, sp->role.word);
2568                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2569                                 n = bucket->first;
2570                         ++vcpu->kvm->stat.mmu_flooded;
2571                         continue;
2572                 }
2573                 page_offset = offset;
2574                 level = sp->role.level;
2575                 npte = 1;
2576                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2577                         page_offset <<= 1;      /* 32->64 */
2578                         /*
2579                          * A 32-bit pde maps 4MB while the shadow pdes map
2580                          * only 2MB.  So we need to double the offset again
2581                          * and zap two pdes instead of one.
2582                          */
2583                         if (level == PT32_ROOT_LEVEL) {
2584                                 page_offset &= ~7; /* kill rounding error */
2585                                 page_offset <<= 1;
2586                                 npte = 2;
2587                         }
2588                         quadrant = page_offset >> PAGE_SHIFT;
2589                         page_offset &= ~PAGE_MASK;
2590                         if (quadrant != sp->role.quadrant)
2591                                 continue;
2592                 }
2593                 spte = &sp->spt[page_offset / sizeof(*spte)];
2594                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2595                         gentry = 0;
2596                         r = kvm_read_guest_atomic(vcpu->kvm,
2597                                                   gpa & ~(u64)(pte_size - 1),
2598                                                   &gentry, pte_size);
2599                         new = (const void *)&gentry;
2600                         if (r < 0)
2601                                 new = NULL;
2602                 }
2603                 while (npte--) {
2604                         entry = *spte;
2605                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2606                         if (new)
2607                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2608                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2609                         ++spte;
2610                 }
2611         }
2612         kvm_mmu_audit(vcpu, "post pte write");
2613         spin_unlock(&vcpu->kvm->mmu_lock);
2614         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2615                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2616                 vcpu->arch.update_pte.pfn = bad_pfn;
2617         }
2618 }
2619
2620 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2621 {
2622         gpa_t gpa;
2623         int r;
2624
2625         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2626
2627         spin_lock(&vcpu->kvm->mmu_lock);
2628         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2629         spin_unlock(&vcpu->kvm->mmu_lock);
2630         return r;
2631 }
2632 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2633
2634 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2635 {
2636         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2637                 struct kvm_mmu_page *sp;
2638
2639                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2640                                   struct kvm_mmu_page, link);
2641                 kvm_mmu_zap_page(vcpu->kvm, sp);
2642                 ++vcpu->kvm->stat.mmu_recycled;
2643         }
2644 }
2645
2646 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2647 {
2648         int r;
2649         enum emulation_result er;
2650
2651         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2652         if (r < 0)
2653                 goto out;
2654
2655         if (!r) {
2656                 r = 1;
2657                 goto out;
2658         }
2659
2660         r = mmu_topup_memory_caches(vcpu);
2661         if (r)
2662                 goto out;
2663
2664         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2665
2666         switch (er) {
2667         case EMULATE_DONE:
2668                 return 1;
2669         case EMULATE_DO_MMIO:
2670                 ++vcpu->stat.mmio_exits;
2671                 return 0;
2672         case EMULATE_FAIL:
2673                 kvm_report_emulation_failure(vcpu, "pagetable");
2674                 return 1;
2675         default:
2676                 BUG();
2677         }
2678 out:
2679         return r;
2680 }
2681 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2682
2683 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2684 {
2685         vcpu->arch.mmu.invlpg(vcpu, gva);
2686         kvm_mmu_flush_tlb(vcpu);
2687         ++vcpu->stat.invlpg;
2688 }
2689 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2690
2691 void kvm_enable_tdp(void)
2692 {
2693         tdp_enabled = true;
2694 }
2695 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2696
2697 void kvm_disable_tdp(void)
2698 {
2699         tdp_enabled = false;
2700 }
2701 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2702
2703 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2704 {
2705         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2706 }
2707
2708 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2709 {
2710         struct page *page;
2711         int i;
2712
2713         ASSERT(vcpu);
2714
2715         if (vcpu->kvm->arch.n_requested_mmu_pages)
2716                 vcpu->kvm->arch.n_free_mmu_pages =
2717                                         vcpu->kvm->arch.n_requested_mmu_pages;
2718         else
2719                 vcpu->kvm->arch.n_free_mmu_pages =
2720                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2721         /*
2722          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2723          * Therefore we need to allocate shadow page tables in the first
2724          * 4GB of memory, which happens to fit the DMA32 zone.
2725          */
2726         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2727         if (!page)
2728                 goto error_1;
2729         vcpu->arch.mmu.pae_root = page_address(page);
2730         for (i = 0; i < 4; ++i)
2731                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2732
2733         return 0;
2734
2735 error_1:
2736         free_mmu_pages(vcpu);
2737         return -ENOMEM;
2738 }
2739
2740 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2741 {
2742         ASSERT(vcpu);
2743         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2744
2745         return alloc_mmu_pages(vcpu);
2746 }
2747
2748 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2749 {
2750         ASSERT(vcpu);
2751         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2752
2753         return init_kvm_mmu(vcpu);
2754 }
2755
2756 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2757 {
2758         ASSERT(vcpu);
2759
2760         destroy_kvm_mmu(vcpu);
2761         free_mmu_pages(vcpu);
2762         mmu_free_memory_caches(vcpu);
2763 }
2764
2765 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2766 {
2767         struct kvm_mmu_page *sp;
2768
2769         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2770                 int i;
2771                 u64 *pt;
2772
2773                 if (!test_bit(slot, sp->slot_bitmap))
2774                         continue;
2775
2776                 pt = sp->spt;
2777                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2778                         /* avoid RMW */
2779                         if (pt[i] & PT_WRITABLE_MASK)
2780                                 pt[i] &= ~PT_WRITABLE_MASK;
2781         }
2782         kvm_flush_remote_tlbs(kvm);
2783 }
2784
2785 void kvm_mmu_zap_all(struct kvm *kvm)
2786 {
2787         struct kvm_mmu_page *sp, *node;
2788
2789         spin_lock(&kvm->mmu_lock);
2790         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2791                 if (kvm_mmu_zap_page(kvm, sp))
2792                         node = container_of(kvm->arch.active_mmu_pages.next,
2793                                             struct kvm_mmu_page, link);
2794         spin_unlock(&kvm->mmu_lock);
2795
2796         kvm_flush_remote_tlbs(kvm);
2797 }
2798
2799 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2800 {
2801         struct kvm_mmu_page *page;
2802
2803         page = container_of(kvm->arch.active_mmu_pages.prev,
2804                             struct kvm_mmu_page, link);
2805         kvm_mmu_zap_page(kvm, page);
2806 }
2807
2808 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2809 {
2810         struct kvm *kvm;
2811         struct kvm *kvm_freed = NULL;
2812         int cache_count = 0;
2813
2814         spin_lock(&kvm_lock);
2815
2816         list_for_each_entry(kvm, &vm_list, vm_list) {
2817                 int npages;
2818
2819                 if (!down_read_trylock(&kvm->slots_lock))
2820                         continue;
2821                 spin_lock(&kvm->mmu_lock);
2822                 npages = kvm->arch.n_alloc_mmu_pages -
2823                          kvm->arch.n_free_mmu_pages;
2824                 cache_count += npages;
2825                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2826                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2827                         cache_count--;
2828                         kvm_freed = kvm;
2829                 }
2830                 nr_to_scan--;
2831
2832                 spin_unlock(&kvm->mmu_lock);
2833                 up_read(&kvm->slots_lock);
2834         }
2835         if (kvm_freed)
2836                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2837
2838         spin_unlock(&kvm_lock);
2839
2840         return cache_count;
2841 }
2842
2843 static struct shrinker mmu_shrinker = {
2844         .shrink = mmu_shrink,
2845         .seeks = DEFAULT_SEEKS * 10,
2846 };
2847
2848 static void mmu_destroy_caches(void)
2849 {
2850         if (pte_chain_cache)
2851                 kmem_cache_destroy(pte_chain_cache);
2852         if (rmap_desc_cache)
2853                 kmem_cache_destroy(rmap_desc_cache);
2854         if (mmu_page_header_cache)
2855                 kmem_cache_destroy(mmu_page_header_cache);
2856 }
2857
2858 void kvm_mmu_module_exit(void)
2859 {
2860         mmu_destroy_caches();
2861         unregister_shrinker(&mmu_shrinker);
2862 }
2863
2864 int kvm_mmu_module_init(void)
2865 {
2866         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2867                                             sizeof(struct kvm_pte_chain),
2868                                             0, 0, NULL);
2869         if (!pte_chain_cache)
2870                 goto nomem;
2871         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2872                                             sizeof(struct kvm_rmap_desc),
2873                                             0, 0, NULL);
2874         if (!rmap_desc_cache)
2875                 goto nomem;
2876
2877         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2878                                                   sizeof(struct kvm_mmu_page),
2879                                                   0, 0, NULL);
2880         if (!mmu_page_header_cache)
2881                 goto nomem;
2882
2883         register_shrinker(&mmu_shrinker);
2884
2885         return 0;
2886
2887 nomem:
2888         mmu_destroy_caches();
2889         return -ENOMEM;
2890 }
2891
2892 /*
2893  * Caculate mmu pages needed for kvm.
2894  */
2895 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2896 {
2897         int i;
2898         unsigned int nr_mmu_pages;
2899         unsigned int  nr_pages = 0;
2900
2901         for (i = 0; i < kvm->nmemslots; i++)
2902                 nr_pages += kvm->memslots[i].npages;
2903
2904         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2905         nr_mmu_pages = max(nr_mmu_pages,
2906                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2907
2908         return nr_mmu_pages;
2909 }
2910
2911 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2912                                 unsigned len)
2913 {
2914         if (len > buffer->len)
2915                 return NULL;
2916         return buffer->ptr;
2917 }
2918
2919 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2920                                 unsigned len)
2921 {
2922         void *ret;
2923
2924         ret = pv_mmu_peek_buffer(buffer, len);
2925         if (!ret)
2926                 return ret;
2927         buffer->ptr += len;
2928         buffer->len -= len;
2929         buffer->processed += len;
2930         return ret;
2931 }
2932
2933 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2934                              gpa_t addr, gpa_t value)
2935 {
2936         int bytes = 8;
2937         int r;
2938
2939         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2940                 bytes = 4;
2941
2942         r = mmu_topup_memory_caches(vcpu);
2943         if (r)
2944                 return r;
2945
2946         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2947                 return -EFAULT;
2948
2949         return 1;
2950 }
2951
2952 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2953 {
2954         kvm_set_cr3(vcpu, vcpu->arch.cr3);
2955         return 1;
2956 }
2957
2958 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2959 {
2960         spin_lock(&vcpu->kvm->mmu_lock);
2961         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2962         spin_unlock(&vcpu->kvm->mmu_lock);
2963         return 1;
2964 }
2965
2966 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2967                              struct kvm_pv_mmu_op_buffer *buffer)
2968 {
2969         struct kvm_mmu_op_header *header;
2970
2971         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2972         if (!header)
2973                 return 0;
2974         switch (header->op) {
2975         case KVM_MMU_OP_WRITE_PTE: {
2976                 struct kvm_mmu_op_write_pte *wpte;
2977
2978                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2979                 if (!wpte)
2980                         return 0;
2981                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2982                                         wpte->pte_val);
2983         }
2984         case KVM_MMU_OP_FLUSH_TLB: {
2985                 struct kvm_mmu_op_flush_tlb *ftlb;
2986
2987                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2988                 if (!ftlb)
2989                         return 0;
2990                 return kvm_pv_mmu_flush_tlb(vcpu);
2991         }
2992         case KVM_MMU_OP_RELEASE_PT: {
2993                 struct kvm_mmu_op_release_pt *rpt;
2994
2995                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2996                 if (!rpt)
2997                         return 0;
2998                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2999         }
3000         default: return 0;
3001         }
3002 }
3003
3004 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3005                   gpa_t addr, unsigned long *ret)
3006 {
3007         int r;
3008         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3009
3010         buffer->ptr = buffer->buf;
3011         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3012         buffer->processed = 0;
3013
3014         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3015         if (r)
3016                 goto out;
3017
3018         while (buffer->len) {
3019                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3020                 if (r < 0)
3021                         goto out;
3022                 if (r == 0)
3023                         break;
3024         }
3025
3026         r = 1;
3027 out:
3028         *ret = buffer->processed;
3029         return r;
3030 }
3031
3032 #ifdef AUDIT
3033
3034 static const char *audit_msg;
3035
3036 static gva_t canonicalize(gva_t gva)
3037 {
3038 #ifdef CONFIG_X86_64
3039         gva = (long long)(gva << 16) >> 16;
3040 #endif
3041         return gva;
3042 }
3043
3044 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3045                                 gva_t va, int level)
3046 {
3047         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3048         int i;
3049         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3050
3051         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3052                 u64 ent = pt[i];
3053
3054                 if (ent == shadow_trap_nonpresent_pte)
3055                         continue;
3056
3057                 va = canonicalize(va);
3058                 if (level > 1) {
3059                         if (ent == shadow_notrap_nonpresent_pte)
3060                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
3061                                        " in nonleaf level: levels %d gva %lx"
3062                                        " level %d pte %llx\n", audit_msg,
3063                                        vcpu->arch.mmu.root_level, va, level, ent);
3064                         else
3065                                 audit_mappings_page(vcpu, ent, va, level - 1);
3066                 } else {
3067                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3068                         gfn_t gfn = gpa >> PAGE_SHIFT;
3069                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3070                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3071
3072                         if (is_shadow_present_pte(ent)
3073                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3074                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3075                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3076                                        audit_msg, vcpu->arch.mmu.root_level,
3077                                        va, gpa, hpa, ent,
3078                                        is_shadow_present_pte(ent));
3079                         else if (ent == shadow_notrap_nonpresent_pte
3080                                  && !is_error_hpa(hpa))
3081                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3082                                        " valid guest gva %lx\n", audit_msg, va);
3083                         kvm_release_pfn_clean(pfn);
3084
3085                 }
3086         }
3087 }
3088
3089 static void audit_mappings(struct kvm_vcpu *vcpu)
3090 {
3091         unsigned i;
3092
3093         if (vcpu->arch.mmu.root_level == 4)
3094                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3095         else
3096                 for (i = 0; i < 4; ++i)
3097                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3098                                 audit_mappings_page(vcpu,
3099                                                     vcpu->arch.mmu.pae_root[i],
3100                                                     i << 30,
3101                                                     2);
3102 }
3103
3104 static int count_rmaps(struct kvm_vcpu *vcpu)
3105 {
3106         int nmaps = 0;
3107         int i, j, k;
3108
3109         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3110                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3111                 struct kvm_rmap_desc *d;
3112
3113                 for (j = 0; j < m->npages; ++j) {
3114                         unsigned long *rmapp = &m->rmap[j];
3115
3116                         if (!*rmapp)
3117                                 continue;
3118                         if (!(*rmapp & 1)) {
3119                                 ++nmaps;
3120                                 continue;
3121                         }
3122                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3123                         while (d) {
3124                                 for (k = 0; k < RMAP_EXT; ++k)
3125                                         if (d->shadow_ptes[k])
3126                                                 ++nmaps;
3127                                         else
3128                                                 break;
3129                                 d = d->more;
3130                         }
3131                 }
3132         }
3133         return nmaps;
3134 }
3135
3136 static int count_writable_mappings(struct kvm_vcpu *vcpu)
3137 {
3138         int nmaps = 0;
3139         struct kvm_mmu_page *sp;
3140         int i;
3141
3142         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3143                 u64 *pt = sp->spt;
3144
3145                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3146                         continue;
3147
3148                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3149                         u64 ent = pt[i];
3150
3151                         if (!(ent & PT_PRESENT_MASK))
3152                                 continue;
3153                         if (!(ent & PT_WRITABLE_MASK))
3154                                 continue;
3155                         ++nmaps;
3156                 }
3157         }
3158         return nmaps;
3159 }
3160
3161 static void audit_rmap(struct kvm_vcpu *vcpu)
3162 {
3163         int n_rmap = count_rmaps(vcpu);
3164         int n_actual = count_writable_mappings(vcpu);
3165
3166         if (n_rmap != n_actual)
3167                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
3168                        __func__, audit_msg, n_rmap, n_actual);
3169 }
3170
3171 static void audit_write_protection(struct kvm_vcpu *vcpu)
3172 {
3173         struct kvm_mmu_page *sp;
3174         struct kvm_memory_slot *slot;
3175         unsigned long *rmapp;
3176         gfn_t gfn;
3177
3178         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3179                 if (sp->role.direct)
3180                         continue;
3181
3182                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3183                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3184                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3185                 if (*rmapp)
3186                         printk(KERN_ERR "%s: (%s) shadow page has writable"
3187                                " mappings: gfn %lx role %x\n",
3188                                __func__, audit_msg, sp->gfn,
3189                                sp->role.word);
3190         }
3191 }
3192
3193 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3194 {
3195         int olddbg = dbg;
3196
3197         dbg = 0;
3198         audit_msg = msg;
3199         audit_rmap(vcpu);
3200         audit_write_protection(vcpu);
3201         audit_mappings(vcpu);
3202         dbg = olddbg;
3203 }
3204
3205 #endif