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