SAFE public projects git trees. - safe/jmp/linux-2.6/blob - arch/x86/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "mmu.h"
  21
  22 #include <linux/kvm_host.h>
  23 #include <linux/types.h>
  24 #include <linux/string.h>
  25 #include <linux/mm.h>
  26 #include <linux/highmem.h>
  27 #include <linux/module.h>
  28 #include <linux/swap.h>
  29 #include <linux/hugetlb.h>
  30 #include <linux/compiler.h>
  31
  32 #include <asm/page.h>
  33 #include <asm/cmpxchg.h>
  34 #include <asm/io.h>
  35 #include <asm/vmx.h>
  36
  37 /*
  38  * When setting this variable to true it enables Two-Dimensional-Paging
  39  * where the hardware walks 2 page tables:
  40  * 1. the guest-virtual to guest-physical
  41  * 2. while doing 1. it walks guest-physical to host-physical
  42  * If the hardware supports that we don't need to do shadow paging.
  43  */
  44 bool tdp_enabled = false;
  45
  46 #undef MMU_DEBUG
  47
  48 #undef AUDIT
  49
  50 #ifdef AUDIT
  51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  52 #else
  53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  54 #endif
  55
  56 #ifdef MMU_DEBUG
  57
  58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  60
  61 #else
  62
  63 #define pgprintk(x...) do { } while (0)
  64 #define rmap_printk(x...) do { } while (0)
  65
  66 #endif
  67
  68 #if defined(MMU_DEBUG) || defined(AUDIT)
  69 static int dbg = 0;
  70 module_param(dbg, bool, 0644);
  71 #endif
  72
  73 static int oos_shadow = 1;
  74 module_param(oos_shadow, bool, 0644);
  75
  76 #ifndef MMU_DEBUG
  77 #define ASSERT(x) do { } while (0)
  78 #else
  79 #define ASSERT(x)                                                       \
  80         if (!(x)) {                                                     \
  81                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  82                        __FILE__, __LINE__, #x);                         \
  83         }
  84 #endif
  85
  86 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  87 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  88
  89 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  90
  91 #define PT64_LEVEL_BITS 9
  92
  93 #define PT64_LEVEL_SHIFT(level) \
  94                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  95
  96 #define PT64_LEVEL_MASK(level) \
  97                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  98
  99 #define PT64_INDEX(address, level)\
 100         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 101
 102
 103 #define PT32_LEVEL_BITS 10
 104
 105 #define PT32_LEVEL_SHIFT(level) \
 106                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 107
 108 #define PT32_LEVEL_MASK(level) \
 109                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 110
 111 #define PT32_INDEX(address, level)\
 112         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 113
 114
 115 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 116 #define PT64_DIR_BASE_ADDR_MASK \
 117         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 118
 119 #define PT32_BASE_ADDR_MASK PAGE_MASK
 120 #define PT32_DIR_BASE_ADDR_MASK \
 121         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 122
 123 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 124                         | PT64_NX_MASK)
 125
 126 #define PFERR_PRESENT_MASK (1U << 0)
 127 #define PFERR_WRITE_MASK (1U << 1)
 128 #define PFERR_USER_MASK (1U << 2)
 129 #define PFERR_FETCH_MASK (1U << 4)
 130
 131 #define PT_DIRECTORY_LEVEL 2
 132 #define PT_PAGE_TABLE_LEVEL 1
 133
 134 #define RMAP_EXT 4
 135
 136 #define ACC_EXEC_MASK    1
 137 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 138 #define ACC_USER_MASK    PT_USER_MASK
 139 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 140
 141 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 142
 143 struct kvm_rmap_desc {
 144         u64 *shadow_ptes[RMAP_EXT];
 145         struct kvm_rmap_desc *more;
 146 };
 147
 148 struct kvm_shadow_walk {
 149         int (*entry)(struct kvm_shadow_walk *walk, struct kvm_vcpu *vcpu,
 150                      u64 addr, u64 *spte, int level);
 151 };
 152
 153 struct kvm_unsync_walk {
 154         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 155 };
 156
 157 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 158
 159 static struct kmem_cache *pte_chain_cache;
 160 static struct kmem_cache *rmap_desc_cache;
 161 static struct kmem_cache *mmu_page_header_cache;
 162
 163 static u64 __read_mostly shadow_trap_nonpresent_pte;
 164 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 165 static u64 __read_mostly shadow_base_present_pte;
 166 static u64 __read_mostly shadow_nx_mask;
 167 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 168 static u64 __read_mostly shadow_user_mask;
 169 static u64 __read_mostly shadow_accessed_mask;
 170 static u64 __read_mostly shadow_dirty_mask;
 171 static u64 __read_mostly shadow_mt_mask;
 172
 173 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 174 {
 175         shadow_trap_nonpresent_pte = trap_pte;
 176         shadow_notrap_nonpresent_pte = notrap_pte;
 177 }
 178 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 179
 180 void kvm_mmu_set_base_ptes(u64 base_pte)
 181 {
 182         shadow_base_present_pte = base_pte;
 183 }
 184 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 185
 186 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 187                 u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 mt_mask)
 188 {
 189         shadow_user_mask = user_mask;
 190         shadow_accessed_mask = accessed_mask;
 191         shadow_dirty_mask = dirty_mask;
 192         shadow_nx_mask = nx_mask;
 193         shadow_x_mask = x_mask;
 194         shadow_mt_mask = mt_mask;
 195 }
 196 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 197
 198 static int is_write_protection(struct kvm_vcpu *vcpu)
 199 {
 200         return vcpu->arch.cr0 & X86_CR0_WP;
 201 }
 202
 203 static int is_cpuid_PSE36(void)
 204 {
 205         return 1;
 206 }
 207
 208 static int is_nx(struct kvm_vcpu *vcpu)
 209 {
 210         return vcpu->arch.shadow_efer & EFER_NX;
 211 }
 212
 213 static int is_present_pte(unsigned long pte)
 214 {
 215         return pte & PT_PRESENT_MASK;
 216 }
 217
 218 static int is_shadow_present_pte(u64 pte)
 219 {
 220         return pte != shadow_trap_nonpresent_pte
 221                 && pte != shadow_notrap_nonpresent_pte;
 222 }
 223
 224 static int is_large_pte(u64 pte)
 225 {
 226         return pte & PT_PAGE_SIZE_MASK;
 227 }
 228
 229 static int is_writeble_pte(unsigned long pte)
 230 {
 231         return pte & PT_WRITABLE_MASK;
 232 }
 233
 234 static int is_dirty_pte(unsigned long pte)
 235 {
 236         return pte & shadow_dirty_mask;
 237 }
 238
 239 static int is_rmap_pte(u64 pte)
 240 {
 241         return is_shadow_present_pte(pte);
 242 }
 243
 244 static pfn_t spte_to_pfn(u64 pte)
 245 {
 246         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 247 }
 248
 249 static gfn_t pse36_gfn_delta(u32 gpte)
 250 {
 251         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 252
 253         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 254 }
 255
 256 static void set_shadow_pte(u64 *sptep, u64 spte)
 257 {
 258 #ifdef CONFIG_X86_64
 259         set_64bit((unsigned long *)sptep, spte);
 260 #else
 261         set_64bit((unsigned long long *)sptep, spte);
 262 #endif
 263 }
 264
 265 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 266                                   struct kmem_cache *base_cache, int min)
 267 {
 268         void *obj;
 269
 270         if (cache->nobjs >= min)
 271                 return 0;
 272         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 273                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 274                 if (!obj)
 275                         return -ENOMEM;
 276                 cache->objects[cache->nobjs++] = obj;
 277         }
 278         return 0;
 279 }
 280
 281 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 282 {
 283         while (mc->nobjs)
 284                 kfree(mc->objects[--mc->nobjs]);
 285 }
 286
 287 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 288                                        int min)
 289 {
 290         struct page *page;
 291
 292         if (cache->nobjs >= min)
 293                 return 0;
 294         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 295                 page = alloc_page(GFP_KERNEL);
 296                 if (!page)
 297                         return -ENOMEM;
 298                 set_page_private(page, 0);
 299                 cache->objects[cache->nobjs++] = page_address(page);
 300         }
 301         return 0;
 302 }
 303
 304 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 305 {
 306         while (mc->nobjs)
 307                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 308 }
 309
 310 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 311 {
 312         int r;
 313
 314         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 315                                    pte_chain_cache, 4);
 316         if (r)
 317                 goto out;
 318         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 319                                    rmap_desc_cache, 4);
 320         if (r)
 321                 goto out;
 322         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 323         if (r)
 324                 goto out;
 325         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 326                                    mmu_page_header_cache, 4);
 327 out:
 328         return r;
 329 }
 330
 331 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 332 {
 333         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 334         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 335         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 336         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 337 }
 338
 339 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 340                                     size_t size)
 341 {
 342         void *p;
 343
 344         BUG_ON(!mc->nobjs);
 345         p = mc->objects[--mc->nobjs];
 346         memset(p, 0, size);
 347         return p;
 348 }
 349
 350 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 351 {
 352         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 353                                       sizeof(struct kvm_pte_chain));
 354 }
 355
 356 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 357 {
 358         kfree(pc);
 359 }
 360
 361 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 362 {
 363         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 364                                       sizeof(struct kvm_rmap_desc));
 365 }
 366
 367 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 368 {
 369         kfree(rd);
 370 }
 371
 372 /*
 373  * Return the pointer to the largepage write count for a given
 374  * gfn, handling slots that are not large page aligned.
 375  */
 376 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 377 {
 378         unsigned long idx;
 379
 380         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 381               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 382         return &slot->lpage_info[idx].write_count;
 383 }
 384
 385 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 386 {
 387         int *write_count;
 388
 389         gfn = unalias_gfn(kvm, gfn);
 390         write_count = slot_largepage_idx(gfn,
 391                                          gfn_to_memslot_unaliased(kvm, gfn));
 392         *write_count += 1;
 393 }
 394
 395 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 396 {
 397         int *write_count;
 398
 399         gfn = unalias_gfn(kvm, gfn);
 400         write_count = slot_largepage_idx(gfn,
 401                                          gfn_to_memslot_unaliased(kvm, gfn));
 402         *write_count -= 1;
 403         WARN_ON(*write_count < 0);
 404 }
 405
 406 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 407 {
 408         struct kvm_memory_slot *slot;
 409         int *largepage_idx;
 410
 411         gfn = unalias_gfn(kvm, gfn);
 412         slot = gfn_to_memslot_unaliased(kvm, gfn);
 413         if (slot) {
 414                 largepage_idx = slot_largepage_idx(gfn, slot);
 415                 return *largepage_idx;
 416         }
 417
 418         return 1;
 419 }
 420
 421 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 422 {
 423         struct vm_area_struct *vma;
 424         unsigned long addr;
 425         int ret = 0;
 426
 427         addr = gfn_to_hva(kvm, gfn);
 428         if (kvm_is_error_hva(addr))
 429                 return ret;
 430
 431         down_read(&current->mm->mmap_sem);
 432         vma = find_vma(current->mm, addr);
 433         if (vma && is_vm_hugetlb_page(vma))
 434                 ret = 1;
 435         up_read(&current->mm->mmap_sem);
 436
 437         return ret;
 438 }
 439
 440 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 441 {
 442         struct kvm_memory_slot *slot;
 443
 444         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 445                 return 0;
 446
 447         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 448                 return 0;
 449
 450         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 451         if (slot && slot->dirty_bitmap)
 452                 return 0;
 453
 454         return 1;
 455 }
 456
 457 /*
 458  * Take gfn and return the reverse mapping to it.
 459  * Note: gfn must be unaliased before this function get called
 460  */
 461
 462 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 463 {
 464         struct kvm_memory_slot *slot;
 465         unsigned long idx;
 466
 467         slot = gfn_to_memslot(kvm, gfn);
 468         if (!lpage)
 469                 return &slot->rmap[gfn - slot->base_gfn];
 470
 471         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 472               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 473
 474         return &slot->lpage_info[idx].rmap_pde;
 475 }
 476
 477 /*
 478  * Reverse mapping data structures:
 479  *
 480  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 481  * that points to page_address(page).
 482  *
 483  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 484  * containing more mappings.
 485  */
 486 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 487 {
 488         struct kvm_mmu_page *sp;
 489         struct kvm_rmap_desc *desc;
 490         unsigned long *rmapp;
 491         int i;
 492
 493         if (!is_rmap_pte(*spte))
 494                 return;
 495         gfn = unalias_gfn(vcpu->kvm, gfn);
 496         sp = page_header(__pa(spte));
 497         sp->gfns[spte - sp->spt] = gfn;
 498         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 499         if (!*rmapp) {
 500                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 501                 *rmapp = (unsigned long)spte;
 502         } else if (!(*rmapp & 1)) {
 503                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 504                 desc = mmu_alloc_rmap_desc(vcpu);
 505                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 506                 desc->shadow_ptes[1] = spte;
 507                 *rmapp = (unsigned long)desc | 1;
 508         } else {
 509                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 510                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 511                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 512                         desc = desc->more;
 513                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 514                         desc->more = mmu_alloc_rmap_desc(vcpu);
 515                         desc = desc->more;
 516                 }
 517                 for (i = 0; desc->shadow_ptes[i]; ++i)
 518                         ;
 519                 desc->shadow_ptes[i] = spte;
 520         }
 521 }
 522
 523 static void rmap_desc_remove_entry(unsigned long *rmapp,
 524                                    struct kvm_rmap_desc *desc,
 525                                    int i,
 526                                    struct kvm_rmap_desc *prev_desc)
 527 {
 528         int j;
 529
 530         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 531                 ;
 532         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 533         desc->shadow_ptes[j] = NULL;
 534         if (j != 0)
 535                 return;
 536         if (!prev_desc && !desc->more)
 537                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 538         else
 539                 if (prev_desc)
 540                         prev_desc->more = desc->more;
 541                 else
 542                         *rmapp = (unsigned long)desc->more | 1;
 543         mmu_free_rmap_desc(desc);
 544 }
 545
 546 static void rmap_remove(struct kvm *kvm, u64 *spte)
 547 {
 548         struct kvm_rmap_desc *desc;
 549         struct kvm_rmap_desc *prev_desc;
 550         struct kvm_mmu_page *sp;
 551         pfn_t pfn;
 552         unsigned long *rmapp;
 553         int i;
 554
 555         if (!is_rmap_pte(*spte))
 556                 return;
 557         sp = page_header(__pa(spte));
 558         pfn = spte_to_pfn(*spte);
 559         if (*spte & shadow_accessed_mask)
 560                 kvm_set_pfn_accessed(pfn);
 561         if (is_writeble_pte(*spte))
 562                 kvm_release_pfn_dirty(pfn);
 563         else
 564                 kvm_release_pfn_clean(pfn);
 565         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 566         if (!*rmapp) {
 567                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 568                 BUG();
 569         } else if (!(*rmapp & 1)) {
 570                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 571                 if ((u64 *)*rmapp != spte) {
 572                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 573                                spte, *spte);
 574                         BUG();
 575                 }
 576                 *rmapp = 0;
 577         } else {
 578                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 579                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 580                 prev_desc = NULL;
 581                 while (desc) {
 582                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 583                                 if (desc->shadow_ptes[i] == spte) {
 584                                         rmap_desc_remove_entry(rmapp,
 585                                                                desc, i,
 586                                                                prev_desc);
 587                                         return;
 588                                 }
 589                         prev_desc = desc;
 590                         desc = desc->more;
 591                 }
 592                 BUG();
 593         }
 594 }
 595
 596 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 597 {
 598         struct kvm_rmap_desc *desc;
 599         struct kvm_rmap_desc *prev_desc;
 600         u64 *prev_spte;
 601         int i;
 602
 603         if (!*rmapp)
 604                 return NULL;
 605         else if (!(*rmapp & 1)) {
 606                 if (!spte)
 607                         return (u64 *)*rmapp;
 608                 return NULL;
 609         }
 610         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 611         prev_desc = NULL;
 612         prev_spte = NULL;
 613         while (desc) {
 614                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 615                         if (prev_spte == spte)
 616                                 return desc->shadow_ptes[i];
 617                         prev_spte = desc->shadow_ptes[i];
 618                 }
 619                 desc = desc->more;
 620         }
 621         return NULL;
 622 }
 623
 624 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 625 {
 626         unsigned long *rmapp;
 627         u64 *spte;
 628         int write_protected = 0;
 629
 630         gfn = unalias_gfn(kvm, gfn);
 631         rmapp = gfn_to_rmap(kvm, gfn, 0);
 632
 633         spte = rmap_next(kvm, rmapp, NULL);
 634         while (spte) {
 635                 BUG_ON(!spte);
 636                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 637                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 638                 if (is_writeble_pte(*spte)) {
 639                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 640                         write_protected = 1;
 641                 }
 642                 spte = rmap_next(kvm, rmapp, spte);
 643         }
 644         if (write_protected) {
 645                 pfn_t pfn;
 646
 647                 spte = rmap_next(kvm, rmapp, NULL);
 648                 pfn = spte_to_pfn(*spte);
 649                 kvm_set_pfn_dirty(pfn);
 650         }
 651
 652         /* check for huge page mappings */
 653         rmapp = gfn_to_rmap(kvm, gfn, 1);
 654         spte = rmap_next(kvm, rmapp, NULL);
 655         while (spte) {
 656                 BUG_ON(!spte);
 657                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 658                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 659                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 660                 if (is_writeble_pte(*spte)) {
 661                         rmap_remove(kvm, spte);
 662                         --kvm->stat.lpages;
 663                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 664                         spte = NULL;
 665                         write_protected = 1;
 666                 }
 667                 spte = rmap_next(kvm, rmapp, spte);
 668         }
 669
 670         return write_protected;
 671 }
 672
 673 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 674 {
 675         u64 *spte;
 676         int need_tlb_flush = 0;
 677
 678         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 679                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 680                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 681                 rmap_remove(kvm, spte);
 682                 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 683                 need_tlb_flush = 1;
 684         }
 685         return need_tlb_flush;
 686 }
 687
 688 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 689                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 690 {
 691         int i;
 692         int retval = 0;
 693
 694         /*
 695          * If mmap_sem isn't taken, we can look the memslots with only
 696          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 697          */
 698         for (i = 0; i < kvm->nmemslots; i++) {
 699                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 700                 unsigned long start = memslot->userspace_addr;
 701                 unsigned long end;
 702
 703                 /* mmu_lock protects userspace_addr */
 704                 if (!start)
 705                         continue;
 706
 707                 end = start + (memslot->npages << PAGE_SHIFT);
 708                 if (hva >= start && hva < end) {
 709                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 710                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 711                         retval |= handler(kvm,
 712                                           &memslot->lpage_info[
 713                                                   gfn_offset /
 714                                                   KVM_PAGES_PER_HPAGE].rmap_pde);
 715                 }
 716         }
 717
 718         return retval;
 719 }
 720
 721 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 722 {
 723         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 724 }
 725
 726 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 727 {
 728         u64 *spte;
 729         int young = 0;
 730
 731         /* always return old for EPT */
 732         if (!shadow_accessed_mask)
 733                 return 0;
 734
 735         spte = rmap_next(kvm, rmapp, NULL);
 736         while (spte) {
 737                 int _young;
 738                 u64 _spte = *spte;
 739                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 740                 _young = _spte & PT_ACCESSED_MASK;
 741                 if (_young) {
 742                         young = 1;
 743                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 744                 }
 745                 spte = rmap_next(kvm, rmapp, spte);
 746         }
 747         return young;
 748 }
 749
 750 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 751 {
 752         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 753 }
 754
 755 #ifdef MMU_DEBUG
 756 static int is_empty_shadow_page(u64 *spt)
 757 {
 758         u64 *pos;
 759         u64 *end;
 760
 761         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 762                 if (is_shadow_present_pte(*pos)) {
 763                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 764                                pos, *pos);
 765                         return 0;
 766                 }
 767         return 1;
 768 }
 769 #endif
 770
 771 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 772 {
 773         ASSERT(is_empty_shadow_page(sp->spt));
 774         list_del(&sp->link);
 775         __free_page(virt_to_page(sp->spt));
 776         __free_page(virt_to_page(sp->gfns));
 777         kfree(sp);
 778         ++kvm->arch.n_free_mmu_pages;
 779 }
 780
 781 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 782 {
 783         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 784 }
 785
 786 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 787                                                u64 *parent_pte)
 788 {
 789         struct kvm_mmu_page *sp;
 790
 791         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 792         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 793         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 794         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 795         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 796         INIT_LIST_HEAD(&sp->oos_link);
 797         ASSERT(is_empty_shadow_page(sp->spt));
 798         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 799         sp->multimapped = 0;
 800         sp->parent_pte = parent_pte;
 801         --vcpu->kvm->arch.n_free_mmu_pages;
 802         return sp;
 803 }
 804
 805 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 806                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 807 {
 808         struct kvm_pte_chain *pte_chain;
 809         struct hlist_node *node;
 810         int i;
 811
 812         if (!parent_pte)
 813                 return;
 814         if (!sp->multimapped) {
 815                 u64 *old = sp->parent_pte;
 816
 817                 if (!old) {
 818                         sp->parent_pte = parent_pte;
 819                         return;
 820                 }
 821                 sp->multimapped = 1;
 822                 pte_chain = mmu_alloc_pte_chain(vcpu);
 823                 INIT_HLIST_HEAD(&sp->parent_ptes);
 824                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 825                 pte_chain->parent_ptes[0] = old;
 826         }
 827         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 828                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 829                         continue;
 830                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 831                         if (!pte_chain->parent_ptes[i]) {
 832                                 pte_chain->parent_ptes[i] = parent_pte;
 833                                 return;
 834                         }
 835         }
 836         pte_chain = mmu_alloc_pte_chain(vcpu);
 837         BUG_ON(!pte_chain);
 838         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 839         pte_chain->parent_ptes[0] = parent_pte;
 840 }
 841
 842 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 843                                        u64 *parent_pte)
 844 {
 845         struct kvm_pte_chain *pte_chain;
 846         struct hlist_node *node;
 847         int i;
 848
 849         if (!sp->multimapped) {
 850                 BUG_ON(sp->parent_pte != parent_pte);
 851                 sp->parent_pte = NULL;
 852                 return;
 853         }
 854         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 855                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 856                         if (!pte_chain->parent_ptes[i])
 857                                 break;
 858                         if (pte_chain->parent_ptes[i] != parent_pte)
 859                                 continue;
 860                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 861                                 && pte_chain->parent_ptes[i + 1]) {
 862                                 pte_chain->parent_ptes[i]
 863                                         = pte_chain->parent_ptes[i + 1];
 864                                 ++i;
 865                         }
 866                         pte_chain->parent_ptes[i] = NULL;
 867                         if (i == 0) {
 868                                 hlist_del(&pte_chain->link);
 869                                 mmu_free_pte_chain(pte_chain);
 870                                 if (hlist_empty(&sp->parent_ptes)) {
 871                                         sp->multimapped = 0;
 872                                         sp->parent_pte = NULL;
 873                                 }
 874                         }
 875                         return;
 876                 }
 877         BUG();
 878 }
 879
 880
 881 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 882                             mmu_parent_walk_fn fn)
 883 {
 884         struct kvm_pte_chain *pte_chain;
 885         struct hlist_node *node;
 886         struct kvm_mmu_page *parent_sp;
 887         int i;
 888
 889         if (!sp->multimapped && sp->parent_pte) {
 890                 parent_sp = page_header(__pa(sp->parent_pte));
 891                 fn(vcpu, parent_sp);
 892                 mmu_parent_walk(vcpu, parent_sp, fn);
 893                 return;
 894         }
 895         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 896                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 897                         if (!pte_chain->parent_ptes[i])
 898                                 break;
 899                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 900                         fn(vcpu, parent_sp);
 901                         mmu_parent_walk(vcpu, parent_sp, fn);
 902                 }
 903 }
 904
 905 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 906 {
 907         unsigned int index;
 908         struct kvm_mmu_page *sp = page_header(__pa(spte));
 909
 910         index = spte - sp->spt;
 911         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
 912                 sp->unsync_children++;
 913         WARN_ON(!sp->unsync_children);
 914 }
 915
 916 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 917 {
 918         struct kvm_pte_chain *pte_chain;
 919         struct hlist_node *node;
 920         int i;
 921
 922         if (!sp->parent_pte)
 923                 return;
 924
 925         if (!sp->multimapped) {
 926                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 927                 return;
 928         }
 929
 930         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 931                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 932                         if (!pte_chain->parent_ptes[i])
 933                                 break;
 934                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 935                 }
 936 }
 937
 938 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 939 {
 940         kvm_mmu_update_parents_unsync(sp);
 941         return 1;
 942 }
 943
 944 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 945                                         struct kvm_mmu_page *sp)
 946 {
 947         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 948         kvm_mmu_update_parents_unsync(sp);
 949 }
 950
 951 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 952                                     struct kvm_mmu_page *sp)
 953 {
 954         int i;
 955
 956         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 957                 sp->spt[i] = shadow_trap_nonpresent_pte;
 958 }
 959
 960 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
 961                                struct kvm_mmu_page *sp)
 962 {
 963         return 1;
 964 }
 965
 966 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
 967 {
 968 }
 969
 970 #define KVM_PAGE_ARRAY_NR 16
 971
 972 struct kvm_mmu_pages {
 973         struct mmu_page_and_offset {
 974                 struct kvm_mmu_page *sp;
 975                 unsigned int idx;
 976         } page[KVM_PAGE_ARRAY_NR];
 977         unsigned int nr;
 978 };
 979
 980 #define for_each_unsync_children(bitmap, idx)           \
 981         for (idx = find_first_bit(bitmap, 512);         \
 982              idx < 512;                                 \
 983              idx = find_next_bit(bitmap, 512, idx+1))
 984
 985 int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
 986                    int idx)
 987 {
 988         int i;
 989
 990         if (sp->unsync)
 991                 for (i=0; i < pvec->nr; i++)
 992                         if (pvec->page[i].sp == sp)
 993                                 return 0;
 994
 995         pvec->page[pvec->nr].sp = sp;
 996         pvec->page[pvec->nr].idx = idx;
 997         pvec->nr++;
 998         return (pvec->nr == KVM_PAGE_ARRAY_NR);
 999 }
1000
1001 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1002                            struct kvm_mmu_pages *pvec)
1003 {
1004         int i, ret, nr_unsync_leaf = 0;
1005
1006         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1007                 u64 ent = sp->spt[i];
1008
1009                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1010                         struct kvm_mmu_page *child;
1011                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1012
1013                         if (child->unsync_children) {
1014                                 if (mmu_pages_add(pvec, child, i))
1015                                         return -ENOSPC;
1016
1017                                 ret = __mmu_unsync_walk(child, pvec);
1018                                 if (!ret)
1019                                         __clear_bit(i, sp->unsync_child_bitmap);
1020                                 else if (ret > 0)
1021                                         nr_unsync_leaf += ret;
1022                                 else
1023                                         return ret;
1024                         }
1025
1026                         if (child->unsync) {
1027                                 nr_unsync_leaf++;
1028                                 if (mmu_pages_add(pvec, child, i))
1029                                         return -ENOSPC;
1030                         }
1031                 }
1032         }
1033
1034         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1035                 sp->unsync_children = 0;
1036
1037         return nr_unsync_leaf;
1038 }
1039
1040 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1041                            struct kvm_mmu_pages *pvec)
1042 {
1043         if (!sp->unsync_children)
1044                 return 0;
1045
1046         mmu_pages_add(pvec, sp, 0);
1047         return __mmu_unsync_walk(sp, pvec);
1048 }
1049
1050 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1051 {
1052         unsigned index;
1053         struct hlist_head *bucket;
1054         struct kvm_mmu_page *sp;
1055         struct hlist_node *node;
1056
1057         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1058         index = kvm_page_table_hashfn(gfn);
1059         bucket = &kvm->arch.mmu_page_hash[index];
1060         hlist_for_each_entry(sp, node, bucket, hash_link)
1061                 if (sp->gfn == gfn && !sp->role.metaphysical
1062                     && !sp->role.invalid) {
1063                         pgprintk("%s: found role %x\n",
1064                                  __func__, sp->role.word);
1065                         return sp;
1066                 }
1067         return NULL;
1068 }
1069
1070 static void kvm_unlink_unsync_global(struct kvm *kvm, struct kvm_mmu_page *sp)
1071 {
1072         list_del(&sp->oos_link);
1073         --kvm->stat.mmu_unsync_global;
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         if (sp->global)
1081                 kvm_unlink_unsync_global(kvm, sp);
1082         --kvm->stat.mmu_unsync;
1083 }
1084
1085 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1086
1087 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1088 {
1089         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1090                 kvm_mmu_zap_page(vcpu->kvm, sp);
1091                 return 1;
1092         }
1093
1094         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1095                 kvm_flush_remote_tlbs(vcpu->kvm);
1096         kvm_unlink_unsync_page(vcpu->kvm, sp);
1097         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1098                 kvm_mmu_zap_page(vcpu->kvm, sp);
1099                 return 1;
1100         }
1101
1102         kvm_mmu_flush_tlb(vcpu);
1103         return 0;
1104 }
1105
1106 struct mmu_page_path {
1107         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1108         unsigned int idx[PT64_ROOT_LEVEL-1];
1109 };
1110
1111 #define for_each_sp(pvec, sp, parents, i)                       \
1112                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1113                         sp = pvec.page[i].sp;                   \
1114                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1115                         i = mmu_pages_next(&pvec, &parents, i))
1116
1117 int mmu_pages_next(struct kvm_mmu_pages *pvec, struct mmu_page_path *parents,
1118                    int i)
1119 {
1120         int n;
1121
1122         for (n = i+1; n < pvec->nr; n++) {
1123                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1124
1125                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1126                         parents->idx[0] = pvec->page[n].idx;
1127                         return n;
1128                 }
1129
1130                 parents->parent[sp->role.level-2] = sp;
1131                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1132         }
1133
1134         return n;
1135 }
1136
1137 void mmu_pages_clear_parents(struct mmu_page_path *parents)
1138 {
1139         struct kvm_mmu_page *sp;
1140         unsigned int level = 0;
1141
1142         do {
1143                 unsigned int idx = parents->idx[level];
1144
1145                 sp = parents->parent[level];
1146                 if (!sp)
1147                         return;
1148
1149                 --sp->unsync_children;
1150                 WARN_ON((int)sp->unsync_children < 0);
1151                 __clear_bit(idx, sp->unsync_child_bitmap);
1152                 level++;
1153         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1154 }
1155
1156 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1157                                struct mmu_page_path *parents,
1158                                struct kvm_mmu_pages *pvec)
1159 {
1160         parents->parent[parent->role.level-1] = NULL;
1161         pvec->nr = 0;
1162 }
1163
1164 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1165                               struct kvm_mmu_page *parent)
1166 {
1167         int i;
1168         struct kvm_mmu_page *sp;
1169         struct mmu_page_path parents;
1170         struct kvm_mmu_pages pages;
1171
1172         kvm_mmu_pages_init(parent, &parents, &pages);
1173         while (mmu_unsync_walk(parent, &pages)) {
1174                 int protected = 0;
1175
1176                 for_each_sp(pages, sp, parents, i)
1177                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1178
1179                 if (protected)
1180                         kvm_flush_remote_tlbs(vcpu->kvm);
1181
1182                 for_each_sp(pages, sp, parents, i) {
1183                         kvm_sync_page(vcpu, sp);
1184                         mmu_pages_clear_parents(&parents);
1185                 }
1186                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1187                 kvm_mmu_pages_init(parent, &parents, &pages);
1188         }
1189 }
1190
1191 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1192                                              gfn_t gfn,
1193                                              gva_t gaddr,
1194                                              unsigned level,
1195                                              int metaphysical,
1196                                              unsigned access,
1197                                              u64 *parent_pte)
1198 {
1199         union kvm_mmu_page_role role;
1200         unsigned index;
1201         unsigned quadrant;
1202         struct hlist_head *bucket;
1203         struct kvm_mmu_page *sp;
1204         struct hlist_node *node, *tmp;
1205
1206         role = vcpu->arch.mmu.base_role;
1207         role.level = level;
1208         role.metaphysical = metaphysical;
1209         role.access = access;
1210         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1211                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1212                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1213                 role.quadrant = quadrant;
1214         }
1215         pgprintk("%s: looking gfn %lx role %x\n", __func__,
1216                  gfn, role.word);
1217         index = kvm_page_table_hashfn(gfn);
1218         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1219         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1220                 if (sp->gfn == gfn) {
1221                         if (sp->unsync)
1222                                 if (kvm_sync_page(vcpu, sp))
1223                                         continue;
1224
1225                         if (sp->role.word != role.word)
1226                                 continue;
1227
1228                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1229                         if (sp->unsync_children) {
1230                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1231                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1232                         }
1233                         pgprintk("%s: found\n", __func__);
1234                         return sp;
1235                 }
1236         ++vcpu->kvm->stat.mmu_cache_miss;
1237         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1238         if (!sp)
1239                 return sp;
1240         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1241         sp->gfn = gfn;
1242         sp->role = role;
1243         sp->global = role.cr4_pge;
1244         hlist_add_head(&sp->hash_link, bucket);
1245         if (!metaphysical) {
1246                 if (rmap_write_protect(vcpu->kvm, gfn))
1247                         kvm_flush_remote_tlbs(vcpu->kvm);
1248                 account_shadowed(vcpu->kvm, gfn);
1249         }
1250         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1251                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1252         else
1253                 nonpaging_prefetch_page(vcpu, sp);
1254         return sp;
1255 }
1256
1257 static int walk_shadow(struct kvm_shadow_walk *walker,
1258                        struct kvm_vcpu *vcpu, u64 addr)
1259 {
1260         hpa_t shadow_addr;
1261         int level;
1262         int r;
1263         u64 *sptep;
1264         unsigned index;
1265
1266         shadow_addr = vcpu->arch.mmu.root_hpa;
1267         level = vcpu->arch.mmu.shadow_root_level;
1268         if (level == PT32E_ROOT_LEVEL) {
1269                 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1270                 shadow_addr &= PT64_BASE_ADDR_MASK;
1271                 if (!shadow_addr)
1272                         return 1;
1273                 --level;
1274         }
1275
1276         while (level >= PT_PAGE_TABLE_LEVEL) {
1277                 index = SHADOW_PT_INDEX(addr, level);
1278                 sptep = ((u64 *)__va(shadow_addr)) + index;
1279                 r = walker->entry(walker, vcpu, addr, sptep, level);
1280                 if (r)
1281                         return r;
1282                 shadow_addr = *sptep & PT64_BASE_ADDR_MASK;
1283                 --level;
1284         }
1285         return 0;
1286 }
1287
1288 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1289                                          struct kvm_mmu_page *sp)
1290 {
1291         unsigned i;
1292         u64 *pt;
1293         u64 ent;
1294
1295         pt = sp->spt;
1296
1297         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1298                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1299                         if (is_shadow_present_pte(pt[i]))
1300                                 rmap_remove(kvm, &pt[i]);
1301                         pt[i] = shadow_trap_nonpresent_pte;
1302                 }
1303                 return;
1304         }
1305
1306         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1307                 ent = pt[i];
1308
1309                 if (is_shadow_present_pte(ent)) {
1310                         if (!is_large_pte(ent)) {
1311                                 ent &= PT64_BASE_ADDR_MASK;
1312                                 mmu_page_remove_parent_pte(page_header(ent),
1313                                                            &pt[i]);
1314                         } else {
1315                                 --kvm->stat.lpages;
1316                                 rmap_remove(kvm, &pt[i]);
1317                         }
1318                 }
1319                 pt[i] = shadow_trap_nonpresent_pte;
1320         }
1321 }
1322
1323 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1324 {
1325         mmu_page_remove_parent_pte(sp, parent_pte);
1326 }
1327
1328 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1329 {
1330         int i;
1331
1332         for (i = 0; i < KVM_MAX_VCPUS; ++i)
1333                 if (kvm->vcpus[i])
1334                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
1335 }
1336
1337 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1338 {
1339         u64 *parent_pte;
1340
1341         while (sp->multimapped || sp->parent_pte) {
1342                 if (!sp->multimapped)
1343                         parent_pte = sp->parent_pte;
1344                 else {
1345                         struct kvm_pte_chain *chain;
1346
1347                         chain = container_of(sp->parent_ptes.first,
1348                                              struct kvm_pte_chain, link);
1349                         parent_pte = chain->parent_ptes[0];
1350                 }
1351                 BUG_ON(!parent_pte);
1352                 kvm_mmu_put_page(sp, parent_pte);
1353                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1354         }
1355 }
1356
1357 static int mmu_zap_unsync_children(struct kvm *kvm,
1358                                    struct kvm_mmu_page *parent)
1359 {
1360         int i, zapped = 0;
1361         struct mmu_page_path parents;
1362         struct kvm_mmu_pages pages;
1363
1364         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1365                 return 0;
1366
1367         kvm_mmu_pages_init(parent, &parents, &pages);
1368         while (mmu_unsync_walk(parent, &pages)) {
1369                 struct kvm_mmu_page *sp;
1370
1371                 for_each_sp(pages, sp, parents, i) {
1372                         kvm_mmu_zap_page(kvm, sp);
1373                         mmu_pages_clear_parents(&parents);
1374                 }
1375                 zapped += pages.nr;
1376                 kvm_mmu_pages_init(parent, &parents, &pages);
1377         }
1378
1379         return zapped;
1380 }
1381
1382 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1383 {
1384         int ret;
1385         ++kvm->stat.mmu_shadow_zapped;
1386         ret = mmu_zap_unsync_children(kvm, sp);
1387         kvm_mmu_page_unlink_children(kvm, sp);
1388         kvm_mmu_unlink_parents(kvm, sp);
1389         kvm_flush_remote_tlbs(kvm);
1390         if (!sp->role.invalid && !sp->role.metaphysical)
1391                 unaccount_shadowed(kvm, sp->gfn);
1392         if (sp->unsync)
1393                 kvm_unlink_unsync_page(kvm, sp);
1394         if (!sp->root_count) {
1395                 hlist_del(&sp->hash_link);
1396                 kvm_mmu_free_page(kvm, sp);
1397         } else {
1398                 sp->role.invalid = 1;
1399                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1400                 kvm_reload_remote_mmus(kvm);
1401         }
1402         kvm_mmu_reset_last_pte_updated(kvm);
1403         return ret;
1404 }
1405
1406 /*
1407  * Changing the number of mmu pages allocated to the vm
1408  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1409  */
1410 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1411 {
1412         /*
1413          * If we set the number of mmu pages to be smaller be than the
1414          * number of actived pages , we must to free some mmu pages before we
1415          * change the value
1416          */
1417
1418         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1419             kvm_nr_mmu_pages) {
1420                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1421                                        - kvm->arch.n_free_mmu_pages;
1422
1423                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1424                         struct kvm_mmu_page *page;
1425
1426                         page = container_of(kvm->arch.active_mmu_pages.prev,
1427                                             struct kvm_mmu_page, link);
1428                         kvm_mmu_zap_page(kvm, page);
1429                         n_used_mmu_pages--;
1430                 }
1431                 kvm->arch.n_free_mmu_pages = 0;
1432         }
1433         else
1434                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1435                                          - kvm->arch.n_alloc_mmu_pages;
1436
1437         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1438 }
1439
1440 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1441 {
1442         unsigned index;
1443         struct hlist_head *bucket;
1444         struct kvm_mmu_page *sp;
1445         struct hlist_node *node, *n;
1446         int r;
1447
1448         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1449         r = 0;
1450         index = kvm_page_table_hashfn(gfn);
1451         bucket = &kvm->arch.mmu_page_hash[index];
1452         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1453                 if (sp->gfn == gfn && !sp->role.metaphysical) {
1454                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1455                                  sp->role.word);
1456                         r = 1;
1457                         if (kvm_mmu_zap_page(kvm, sp))
1458                                 n = bucket->first;
1459                 }
1460         return r;
1461 }
1462
1463 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1464 {
1465         struct kvm_mmu_page *sp;
1466
1467         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1468                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1469                 kvm_mmu_zap_page(kvm, sp);
1470         }
1471 }
1472
1473 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1474 {
1475         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1476         struct kvm_mmu_page *sp = page_header(__pa(pte));
1477
1478         __set_bit(slot, sp->slot_bitmap);
1479 }
1480
1481 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1482 {
1483         int i;
1484         u64 *pt = sp->spt;
1485
1486         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1487                 return;
1488
1489         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1490                 if (pt[i] == shadow_notrap_nonpresent_pte)
1491                         set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1492         }
1493 }
1494
1495 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1496 {
1497         struct page *page;
1498
1499         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1500
1501         if (gpa == UNMAPPED_GVA)
1502                 return NULL;
1503
1504         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1505
1506         return page;
1507 }
1508
1509 /*
1510  * The function is based on mtrr_type_lookup() in
1511  * arch/x86/kernel/cpu/mtrr/generic.c
1512  */
1513 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1514                          u64 start, u64 end)
1515 {
1516         int i;
1517         u64 base, mask;
1518         u8 prev_match, curr_match;
1519         int num_var_ranges = KVM_NR_VAR_MTRR;
1520
1521         if (!mtrr_state->enabled)
1522                 return 0xFF;
1523
1524         /* Make end inclusive end, instead of exclusive */
1525         end--;
1526
1527         /* Look in fixed ranges. Just return the type as per start */
1528         if (mtrr_state->have_fixed && (start < 0x100000)) {
1529                 int idx;
1530
1531                 if (start < 0x80000) {
1532                         idx = 0;
1533                         idx += (start >> 16);
1534                         return mtrr_state->fixed_ranges[idx];
1535                 } else if (start < 0xC0000) {
1536                         idx = 1 * 8;
1537                         idx += ((start - 0x80000) >> 14);
1538                         return mtrr_state->fixed_ranges[idx];
1539                 } else if (start < 0x1000000) {
1540                         idx = 3 * 8;
1541                         idx += ((start - 0xC0000) >> 12);
1542                         return mtrr_state->fixed_ranges[idx];
1543                 }
1544         }
1545
1546         /*
1547          * Look in variable ranges
1548          * Look of multiple ranges matching this address and pick type
1549          * as per MTRR precedence
1550          */
1551         if (!(mtrr_state->enabled & 2))
1552                 return mtrr_state->def_type;
1553
1554         prev_match = 0xFF;
1555         for (i = 0; i < num_var_ranges; ++i) {
1556                 unsigned short start_state, end_state;
1557
1558                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1559                         continue;
1560
1561                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1562                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1563                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1564                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1565
1566                 start_state = ((start & mask) == (base & mask));
1567                 end_state = ((end & mask) == (base & mask));
1568                 if (start_state != end_state)
1569                         return 0xFE;
1570
1571                 if ((start & mask) != (base & mask))
1572                         continue;
1573
1574                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1575                 if (prev_match == 0xFF) {
1576                         prev_match = curr_match;
1577                         continue;
1578                 }
1579
1580                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1581                     curr_match == MTRR_TYPE_UNCACHABLE)
1582                         return MTRR_TYPE_UNCACHABLE;
1583
1584                 if ((prev_match == MTRR_TYPE_WRBACK &&
1585                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1586                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1587                      curr_match == MTRR_TYPE_WRBACK)) {
1588                         prev_match = MTRR_TYPE_WRTHROUGH;
1589                         curr_match = MTRR_TYPE_WRTHROUGH;
1590                 }
1591
1592                 if (prev_match != curr_match)
1593                         return MTRR_TYPE_UNCACHABLE;
1594         }
1595
1596         if (prev_match != 0xFF)
1597                 return prev_match;
1598
1599         return mtrr_state->def_type;
1600 }
1601
1602 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1603 {
1604         u8 mtrr;
1605
1606         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1607                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1608         if (mtrr == 0xfe || mtrr == 0xff)
1609                 mtrr = MTRR_TYPE_WRBACK;
1610         return mtrr;
1611 }
1612
1613 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1614 {
1615         unsigned index;
1616         struct hlist_head *bucket;
1617         struct kvm_mmu_page *s;
1618         struct hlist_node *node, *n;
1619
1620         index = kvm_page_table_hashfn(sp->gfn);
1621         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1622         /* don't unsync if pagetable is shadowed with multiple roles */
1623         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1624                 if (s->gfn != sp->gfn || s->role.metaphysical)
1625                         continue;
1626                 if (s->role.word != sp->role.word)
1627                         return 1;
1628         }
1629         ++vcpu->kvm->stat.mmu_unsync;
1630         sp->unsync = 1;
1631
1632         if (sp->global) {
1633                 list_add(&sp->oos_link, &vcpu->kvm->arch.oos_global_pages);
1634                 ++vcpu->kvm->stat.mmu_unsync_global;
1635         } else
1636                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1637
1638         mmu_convert_notrap(sp);
1639         return 0;
1640 }
1641
1642 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1643                                   bool can_unsync)
1644 {
1645         struct kvm_mmu_page *shadow;
1646
1647         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1648         if (shadow) {
1649                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1650                         return 1;
1651                 if (shadow->unsync)
1652                         return 0;
1653                 if (can_unsync && oos_shadow)
1654                         return kvm_unsync_page(vcpu, shadow);
1655                 return 1;
1656         }
1657         return 0;
1658 }
1659
1660 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1661                     unsigned pte_access, int user_fault,
1662                     int write_fault, int dirty, int largepage,
1663                     int global, gfn_t gfn, pfn_t pfn, bool speculative,
1664                     bool can_unsync)
1665 {
1666         u64 spte;
1667         int ret = 0;
1668         u64 mt_mask = shadow_mt_mask;
1669         struct kvm_mmu_page *sp = page_header(__pa(shadow_pte));
1670
1671         if (!global && sp->global) {
1672                 sp->global = 0;
1673                 if (sp->unsync) {
1674                         kvm_unlink_unsync_global(vcpu->kvm, sp);
1675                         kvm_mmu_mark_parents_unsync(vcpu, sp);
1676                 }
1677         }
1678
1679         /*
1680          * We don't set the accessed bit, since we sometimes want to see
1681          * whether the guest actually used the pte (in order to detect
1682          * demand paging).
1683          */
1684         spte = shadow_base_present_pte | shadow_dirty_mask;
1685         if (!speculative)
1686                 spte |= shadow_accessed_mask;
1687         if (!dirty)
1688                 pte_access &= ~ACC_WRITE_MASK;
1689         if (pte_access & ACC_EXEC_MASK)
1690                 spte |= shadow_x_mask;
1691         else
1692                 spte |= shadow_nx_mask;
1693         if (pte_access & ACC_USER_MASK)
1694                 spte |= shadow_user_mask;
1695         if (largepage)
1696                 spte |= PT_PAGE_SIZE_MASK;
1697         if (mt_mask) {
1698                 if (!kvm_is_mmio_pfn(pfn)) {
1699                         mt_mask = get_memory_type(vcpu, gfn) <<
1700                                 kvm_x86_ops->get_mt_mask_shift();
1701                         mt_mask |= VMX_EPT_IGMT_BIT;
1702                 } else
1703                         mt_mask = MTRR_TYPE_UNCACHABLE <<
1704                                 kvm_x86_ops->get_mt_mask_shift();
1705                 spte |= mt_mask;
1706         }
1707
1708         spte |= (u64)pfn << PAGE_SHIFT;
1709
1710         if ((pte_access & ACC_WRITE_MASK)
1711             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1712
1713                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1714                         ret = 1;
1715                         spte = shadow_trap_nonpresent_pte;
1716                         goto set_pte;
1717                 }
1718
1719                 spte |= PT_WRITABLE_MASK;
1720
1721                 /*
1722                  * Optimization: for pte sync, if spte was writable the hash
1723                  * lookup is unnecessary (and expensive). Write protection
1724                  * is responsibility of mmu_get_page / kvm_sync_page.
1725                  * Same reasoning can be applied to dirty page accounting.
1726                  */
1727                 if (!can_unsync && is_writeble_pte(*shadow_pte))
1728                         goto set_pte;
1729
1730                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1731                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1732                                  __func__, gfn);
1733                         ret = 1;
1734                         pte_access &= ~ACC_WRITE_MASK;
1735                         if (is_writeble_pte(spte))
1736                                 spte &= ~PT_WRITABLE_MASK;
1737                 }
1738         }
1739
1740         if (pte_access & ACC_WRITE_MASK)
1741                 mark_page_dirty(vcpu->kvm, gfn);
1742
1743 set_pte:
1744         set_shadow_pte(shadow_pte, spte);
1745         return ret;
1746 }
1747
1748 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1749                          unsigned pt_access, unsigned pte_access,
1750                          int user_fault, int write_fault, int dirty,
1751                          int *ptwrite, int largepage, int global,
1752                          gfn_t gfn, pfn_t pfn, bool speculative)
1753 {
1754         int was_rmapped = 0;
1755         int was_writeble = is_writeble_pte(*shadow_pte);
1756
1757         pgprintk("%s: spte %llx access %x write_fault %d"
1758                  " user_fault %d gfn %lx\n",
1759                  __func__, *shadow_pte, pt_access,
1760                  write_fault, user_fault, gfn);
1761
1762         if (is_rmap_pte(*shadow_pte)) {
1763                 /*
1764                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1765                  * the parent of the now unreachable PTE.
1766                  */
1767                 if (largepage && !is_large_pte(*shadow_pte)) {
1768                         struct kvm_mmu_page *child;
1769                         u64 pte = *shadow_pte;
1770
1771                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1772                         mmu_page_remove_parent_pte(child, shadow_pte);
1773                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1774                         pgprintk("hfn old %lx new %lx\n",
1775                                  spte_to_pfn(*shadow_pte), pfn);
1776                         rmap_remove(vcpu->kvm, shadow_pte);
1777                 } else {
1778                         if (largepage)
1779                                 was_rmapped = is_large_pte(*shadow_pte);
1780                         else
1781                                 was_rmapped = 1;
1782                 }
1783         }
1784         if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1785                       dirty, largepage, global, gfn, pfn, speculative, true)) {
1786                 if (write_fault)
1787                         *ptwrite = 1;
1788                 kvm_x86_ops->tlb_flush(vcpu);
1789         }
1790
1791         pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1792         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1793                  is_large_pte(*shadow_pte)? "2MB" : "4kB",
1794                  is_present_pte(*shadow_pte)?"RW":"R", gfn,
1795                  *shadow_pte, shadow_pte);
1796         if (!was_rmapped && is_large_pte(*shadow_pte))
1797                 ++vcpu->kvm->stat.lpages;
1798
1799         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1800         if (!was_rmapped) {
1801                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1802                 if (!is_rmap_pte(*shadow_pte))
1803                         kvm_release_pfn_clean(pfn);
1804         } else {
1805                 if (was_writeble)
1806                         kvm_release_pfn_dirty(pfn);
1807                 else
1808                         kvm_release_pfn_clean(pfn);
1809         }
1810         if (speculative) {
1811                 vcpu->arch.last_pte_updated = shadow_pte;
1812                 vcpu->arch.last_pte_gfn = gfn;
1813         }
1814 }
1815
1816 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1817 {
1818 }
1819
1820 struct direct_shadow_walk {
1821         struct kvm_shadow_walk walker;
1822         pfn_t pfn;
1823         int write;
1824         int largepage;
1825         int pt_write;
1826 };
1827
1828 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1829                             struct kvm_vcpu *vcpu,
1830                             u64 addr, u64 *sptep, int level)
1831 {
1832         struct direct_shadow_walk *walk =
1833                 container_of(_walk, struct direct_shadow_walk, walker);
1834         struct kvm_mmu_page *sp;
1835         gfn_t pseudo_gfn;
1836         gfn_t gfn = addr >> PAGE_SHIFT;
1837
1838         if (level == PT_PAGE_TABLE_LEVEL
1839             || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1840                 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1841                              0, walk->write, 1, &walk->pt_write,
1842                              walk->largepage, 0, gfn, walk->pfn, false);
1843                 ++vcpu->stat.pf_fixed;
1844                 return 1;
1845         }
1846
1847         if (*sptep == shadow_trap_nonpresent_pte) {
1848                 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1849                 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1850                                       1, ACC_ALL, sptep);
1851                 if (!sp) {
1852                         pgprintk("nonpaging_map: ENOMEM\n");
1853                         kvm_release_pfn_clean(walk->pfn);
1854                         return -ENOMEM;
1855                 }
1856
1857                 set_shadow_pte(sptep,
1858                                __pa(sp->spt)
1859                                | PT_PRESENT_MASK | PT_WRITABLE_MASK
1860                                | shadow_user_mask | shadow_x_mask);
1861         }
1862         return 0;
1863 }
1864
1865 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1866                         int largepage, gfn_t gfn, pfn_t pfn)
1867 {
1868         int r;
1869         struct direct_shadow_walk walker = {
1870                 .walker = { .entry = direct_map_entry, },
1871                 .pfn = pfn,
1872                 .largepage = largepage,
1873                 .write = write,
1874                 .pt_write = 0,
1875         };
1876
1877         r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1878         if (r < 0)
1879                 return r;
1880         return walker.pt_write;
1881 }
1882
1883 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1884 {
1885         int r;
1886         int largepage = 0;
1887         pfn_t pfn;
1888         unsigned long mmu_seq;
1889
1890         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1891                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1892                 largepage = 1;
1893         }
1894
1895         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1896         smp_rmb();
1897         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1898
1899         /* mmio */
1900         if (is_error_pfn(pfn)) {
1901                 kvm_release_pfn_clean(pfn);
1902                 return 1;
1903         }
1904
1905         spin_lock(&vcpu->kvm->mmu_lock);
1906         if (mmu_notifier_retry(vcpu, mmu_seq))
1907                 goto out_unlock;
1908         kvm_mmu_free_some_pages(vcpu);
1909         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1910         spin_unlock(&vcpu->kvm->mmu_lock);
1911
1912
1913         return r;
1914
1915 out_unlock:
1916         spin_unlock(&vcpu->kvm->mmu_lock);
1917         kvm_release_pfn_clean(pfn);
1918         return 0;
1919 }
1920
1921
1922 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1923 {
1924         int i;
1925         struct kvm_mmu_page *sp;
1926
1927         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1928                 return;
1929         spin_lock(&vcpu->kvm->mmu_lock);
1930         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1931                 hpa_t root = vcpu->arch.mmu.root_hpa;
1932
1933                 sp = page_header(root);
1934                 --sp->root_count;
1935                 if (!sp->root_count && sp->role.invalid)
1936                         kvm_mmu_zap_page(vcpu->kvm, sp);
1937                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1938                 spin_unlock(&vcpu->kvm->mmu_lock);
1939                 return;
1940         }
1941         for (i = 0; i < 4; ++i) {
1942                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1943
1944                 if (root) {
1945                         root &= PT64_BASE_ADDR_MASK;
1946                         sp = page_header(root);
1947                         --sp->root_count;
1948                         if (!sp->root_count && sp->role.invalid)
1949                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1950                 }
1951                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1952         }
1953         spin_unlock(&vcpu->kvm->mmu_lock);
1954         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1955 }
1956
1957 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1958 {
1959         int i;
1960         gfn_t root_gfn;
1961         struct kvm_mmu_page *sp;
1962         int metaphysical = 0;
1963
1964         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1965
1966         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1967                 hpa_t root = vcpu->arch.mmu.root_hpa;
1968
1969                 ASSERT(!VALID_PAGE(root));
1970                 if (tdp_enabled)
1971                         metaphysical = 1;
1972                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1973                                       PT64_ROOT_LEVEL, metaphysical,
1974                                       ACC_ALL, NULL);
1975                 root = __pa(sp->spt);
1976                 ++sp->root_count;
1977                 vcpu->arch.mmu.root_hpa = root;
1978                 return;
1979         }
1980         metaphysical = !is_paging(vcpu);
1981         if (tdp_enabled)
1982                 metaphysical = 1;
1983         for (i = 0; i < 4; ++i) {
1984                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1985
1986                 ASSERT(!VALID_PAGE(root));
1987                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1988                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1989                                 vcpu->arch.mmu.pae_root[i] = 0;
1990                                 continue;
1991                         }
1992                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1993                 } else if (vcpu->arch.mmu.root_level == 0)
1994                         root_gfn = 0;
1995                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1996                                       PT32_ROOT_LEVEL, metaphysical,
1997                                       ACC_ALL, NULL);
1998                 root = __pa(sp->spt);
1999                 ++sp->root_count;
2000                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2001         }
2002         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2003 }
2004
2005 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2006 {
2007         int i;
2008         struct kvm_mmu_page *sp;
2009
2010         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2011                 return;
2012         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2013                 hpa_t root = vcpu->arch.mmu.root_hpa;
2014                 sp = page_header(root);
2015                 mmu_sync_children(vcpu, sp);
2016                 return;
2017         }
2018         for (i = 0; i < 4; ++i) {
2019                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2020
2021                 if (root) {
2022                         root &= PT64_BASE_ADDR_MASK;
2023                         sp = page_header(root);
2024                         mmu_sync_children(vcpu, sp);
2025                 }
2026         }
2027 }
2028
2029 static void mmu_sync_global(struct kvm_vcpu *vcpu)
2030 {
2031         struct kvm *kvm = vcpu->kvm;
2032         struct kvm_mmu_page *sp, *n;
2033
2034         list_for_each_entry_safe(sp, n, &kvm->arch.oos_global_pages, oos_link)
2035                 kvm_sync_page(vcpu, sp);
2036 }
2037
2038 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2039 {
2040         spin_lock(&vcpu->kvm->mmu_lock);
2041         mmu_sync_roots(vcpu);
2042         spin_unlock(&vcpu->kvm->mmu_lock);
2043 }
2044
2045 void kvm_mmu_sync_global(struct kvm_vcpu *vcpu)
2046 {
2047         spin_lock(&vcpu->kvm->mmu_lock);
2048         mmu_sync_global(vcpu);
2049         spin_unlock(&vcpu->kvm->mmu_lock);
2050 }
2051
2052 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2053 {
2054         return vaddr;
2055 }
2056
2057 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2058                                 u32 error_code)
2059 {
2060         gfn_t gfn;
2061         int r;
2062
2063         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2064         r = mmu_topup_memory_caches(vcpu);
2065         if (r)
2066                 return r;
2067
2068         ASSERT(vcpu);
2069         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2070
2071         gfn = gva >> PAGE_SHIFT;
2072
2073         return nonpaging_map(vcpu, gva & PAGE_MASK,
2074                              error_code & PFERR_WRITE_MASK, gfn);
2075 }
2076
2077 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2078                                 u32 error_code)
2079 {
2080         pfn_t pfn;
2081         int r;
2082         int largepage = 0;
2083         gfn_t gfn = gpa >> PAGE_SHIFT;
2084         unsigned long mmu_seq;
2085
2086         ASSERT(vcpu);
2087         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2088
2089         r = mmu_topup_memory_caches(vcpu);
2090         if (r)
2091                 return r;
2092
2093         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
2094                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2095                 largepage = 1;
2096         }
2097         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2098         smp_rmb();
2099         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2100         if (is_error_pfn(pfn)) {
2101                 kvm_release_pfn_clean(pfn);
2102                 return 1;
2103         }
2104         spin_lock(&vcpu->kvm->mmu_lock);
2105         if (mmu_notifier_retry(vcpu, mmu_seq))
2106                 goto out_unlock;
2107         kvm_mmu_free_some_pages(vcpu);
2108         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2109                          largepage, gfn, pfn);
2110         spin_unlock(&vcpu->kvm->mmu_lock);
2111
2112         return r;
2113
2114 out_unlock:
2115         spin_unlock(&vcpu->kvm->mmu_lock);
2116         kvm_release_pfn_clean(pfn);
2117         return 0;
2118 }
2119
2120 static void nonpaging_free(struct kvm_vcpu *vcpu)
2121 {
2122         mmu_free_roots(vcpu);
2123 }
2124
2125 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2126 {
2127         struct kvm_mmu *context = &vcpu->arch.mmu;
2128
2129         context->new_cr3 = nonpaging_new_cr3;
2130         context->page_fault = nonpaging_page_fault;
2131         context->gva_to_gpa = nonpaging_gva_to_gpa;
2132         context->free = nonpaging_free;
2133         context->prefetch_page = nonpaging_prefetch_page;
2134         context->sync_page = nonpaging_sync_page;
2135         context->invlpg = nonpaging_invlpg;
2136         context->root_level = 0;
2137         context->shadow_root_level = PT32E_ROOT_LEVEL;
2138         context->root_hpa = INVALID_PAGE;
2139         return 0;
2140 }
2141
2142 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2143 {
2144         ++vcpu->stat.tlb_flush;
2145         kvm_x86_ops->tlb_flush(vcpu);
2146 }
2147
2148 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2149 {
2150         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2151         mmu_free_roots(vcpu);
2152 }
2153
2154 static void inject_page_fault(struct kvm_vcpu *vcpu,
2155                               u64 addr,
2156                               u32 err_code)
2157 {
2158         kvm_inject_page_fault(vcpu, addr, err_code);
2159 }
2160
2161 static void paging_free(struct kvm_vcpu *vcpu)
2162 {
2163         nonpaging_free(vcpu);
2164 }
2165
2166 #define PTTYPE 64
2167 #include "paging_tmpl.h"
2168 #undef PTTYPE
2169
2170 #define PTTYPE 32
2171 #include "paging_tmpl.h"
2172 #undef PTTYPE
2173
2174 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2175 {
2176         struct kvm_mmu *context = &vcpu->arch.mmu;
2177
2178         ASSERT(is_pae(vcpu));
2179         context->new_cr3 = paging_new_cr3;
2180         context->page_fault = paging64_page_fault;
2181         context->gva_to_gpa = paging64_gva_to_gpa;
2182         context->prefetch_page = paging64_prefetch_page;
2183         context->sync_page = paging64_sync_page;
2184         context->invlpg = paging64_invlpg;
2185         context->free = paging_free;
2186         context->root_level = level;
2187         context->shadow_root_level = level;
2188         context->root_hpa = INVALID_PAGE;
2189         return 0;
2190 }
2191
2192 static int paging64_init_context(struct kvm_vcpu *vcpu)
2193 {
2194         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2195 }
2196
2197 static int paging32_init_context(struct kvm_vcpu *vcpu)
2198 {
2199         struct kvm_mmu *context = &vcpu->arch.mmu;
2200
2201         context->new_cr3 = paging_new_cr3;
2202         context->page_fault = paging32_page_fault;
2203         context->gva_to_gpa = paging32_gva_to_gpa;
2204         context->free = paging_free;
2205         context->prefetch_page = paging32_prefetch_page;
2206         context->sync_page = paging32_sync_page;
2207         context->invlpg = paging32_invlpg;
2208         context->root_level = PT32_ROOT_LEVEL;
2209         context->shadow_root_level = PT32E_ROOT_LEVEL;
2210         context->root_hpa = INVALID_PAGE;
2211         return 0;
2212 }
2213
2214 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2215 {
2216         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2217 }
2218
2219 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2220 {
2221         struct kvm_mmu *context = &vcpu->arch.mmu;
2222
2223         context->new_cr3 = nonpaging_new_cr3;
2224         context->page_fault = tdp_page_fault;
2225         context->free = nonpaging_free;
2226         context->prefetch_page = nonpaging_prefetch_page;
2227         context->sync_page = nonpaging_sync_page;
2228         context->invlpg = nonpaging_invlpg;
2229         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2230         context->root_hpa = INVALID_PAGE;
2231
2232         if (!is_paging(vcpu)) {
2233                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2234                 context->root_level = 0;
2235         } else if (is_long_mode(vcpu)) {
2236                 context->gva_to_gpa = paging64_gva_to_gpa;
2237                 context->root_level = PT64_ROOT_LEVEL;
2238         } else if (is_pae(vcpu)) {
2239                 context->gva_to_gpa = paging64_gva_to_gpa;
2240                 context->root_level = PT32E_ROOT_LEVEL;
2241         } else {
2242                 context->gva_to_gpa = paging32_gva_to_gpa;
2243                 context->root_level = PT32_ROOT_LEVEL;
2244         }
2245
2246         return 0;
2247 }
2248
2249 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2250 {
2251         int r;
2252
2253         ASSERT(vcpu);
2254         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2255
2256         if (!is_paging(vcpu))
2257                 r = nonpaging_init_context(vcpu);
2258         else if (is_long_mode(vcpu))
2259                 r = paging64_init_context(vcpu);
2260         else if (is_pae(vcpu))
2261                 r = paging32E_init_context(vcpu);
2262         else
2263                 r = paging32_init_context(vcpu);
2264
2265         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2266
2267         return r;
2268 }
2269
2270 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2271 {
2272         vcpu->arch.update_pte.pfn = bad_pfn;
2273
2274         if (tdp_enabled)
2275                 return init_kvm_tdp_mmu(vcpu);
2276         else
2277                 return init_kvm_softmmu(vcpu);
2278 }
2279
2280 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2281 {
2282         ASSERT(vcpu);
2283         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2284                 vcpu->arch.mmu.free(vcpu);
2285                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2286         }
2287 }
2288
2289 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2290 {
2291         destroy_kvm_mmu(vcpu);
2292         return init_kvm_mmu(vcpu);
2293 }
2294 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2295
2296 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2297 {
2298         int r;
2299
2300         r = mmu_topup_memory_caches(vcpu);
2301         if (r)
2302                 goto out;
2303         spin_lock(&vcpu->kvm->mmu_lock);
2304         kvm_mmu_free_some_pages(vcpu);
2305         mmu_alloc_roots(vcpu);
2306         mmu_sync_roots(vcpu);
2307         spin_unlock(&vcpu->kvm->mmu_lock);
2308         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2309         kvm_mmu_flush_tlb(vcpu);
2310 out:
2311         return r;
2312 }
2313 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2314
2315 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2316 {
2317         mmu_free_roots(vcpu);
2318 }
2319
2320 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2321                                   struct kvm_mmu_page *sp,
2322                                   u64 *spte)
2323 {
2324         u64 pte;
2325         struct kvm_mmu_page *child;
2326
2327         pte = *spte;
2328         if (is_shadow_present_pte(pte)) {
2329                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2330                     is_large_pte(pte))
2331                         rmap_remove(vcpu->kvm, spte);
2332                 else {
2333                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2334                         mmu_page_remove_parent_pte(child, spte);
2335                 }
2336         }
2337         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2338         if (is_large_pte(pte))
2339                 --vcpu->kvm->stat.lpages;
2340 }
2341
2342 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2343                                   struct kvm_mmu_page *sp,
2344                                   u64 *spte,
2345                                   const void *new)
2346 {
2347         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2348                 if (!vcpu->arch.update_pte.largepage ||
2349                     sp->role.glevels == PT32_ROOT_LEVEL) {
2350                         ++vcpu->kvm->stat.mmu_pde_zapped;
2351                         return;
2352                 }
2353         }
2354
2355         ++vcpu->kvm->stat.mmu_pte_updated;
2356         if (sp->role.glevels == PT32_ROOT_LEVEL)
2357                 paging32_update_pte(vcpu, sp, spte, new);
2358         else
2359                 paging64_update_pte(vcpu, sp, spte, new);
2360 }
2361
2362 static bool need_remote_flush(u64 old, u64 new)
2363 {
2364         if (!is_shadow_present_pte(old))
2365                 return false;
2366         if (!is_shadow_present_pte(new))
2367                 return true;
2368         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2369                 return true;
2370         old ^= PT64_NX_MASK;
2371         new ^= PT64_NX_MASK;
2372         return (old & ~new & PT64_PERM_MASK) != 0;
2373 }
2374
2375 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2376 {
2377         if (need_remote_flush(old, new))
2378                 kvm_flush_remote_tlbs(vcpu->kvm);
2379         else
2380                 kvm_mmu_flush_tlb(vcpu);
2381 }
2382
2383 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2384 {
2385         u64 *spte = vcpu->arch.last_pte_updated;
2386
2387         return !!(spte && (*spte & shadow_accessed_mask));
2388 }
2389
2390 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2391                                           const u8 *new, int bytes)
2392 {
2393         gfn_t gfn;
2394         int r;
2395         u64 gpte = 0;
2396         pfn_t pfn;
2397
2398         vcpu->arch.update_pte.largepage = 0;
2399
2400         if (bytes != 4 && bytes != 8)
2401                 return;
2402
2403         /*
2404          * Assume that the pte write on a page table of the same type
2405          * as the current vcpu paging mode.  This is nearly always true
2406          * (might be false while changing modes).  Note it is verified later
2407          * by update_pte().
2408          */
2409         if (is_pae(vcpu)) {
2410                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2411                 if ((bytes == 4) && (gpa % 4 == 0)) {
2412                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2413                         if (r)
2414                                 return;
2415                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2416                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2417                         memcpy((void *)&gpte, new, 8);
2418                 }
2419         } else {
2420                 if ((bytes == 4) && (gpa % 4 == 0))
2421                         memcpy((void *)&gpte, new, 4);
2422         }
2423         if (!is_present_pte(gpte))
2424                 return;
2425         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2426
2427         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2428                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2429                 vcpu->arch.update_pte.largepage = 1;
2430         }
2431         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2432         smp_rmb();
2433         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2434
2435         if (is_error_pfn(pfn)) {
2436                 kvm_release_pfn_clean(pfn);
2437                 return;
2438         }
2439         vcpu->arch.update_pte.gfn = gfn;
2440         vcpu->arch.update_pte.pfn = pfn;
2441 }
2442
2443 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2444 {
2445         u64 *spte = vcpu->arch.last_pte_updated;
2446
2447         if (spte
2448             && vcpu->arch.last_pte_gfn == gfn
2449             && shadow_accessed_mask
2450             && !(*spte & shadow_accessed_mask)
2451             && is_shadow_present_pte(*spte))
2452                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2453 }
2454
2455 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2456                        const u8 *new, int bytes,
2457                        bool guest_initiated)
2458 {
2459         gfn_t gfn = gpa >> PAGE_SHIFT;
2460         struct kvm_mmu_page *sp;
2461         struct hlist_node *node, *n;
2462         struct hlist_head *bucket;
2463         unsigned index;
2464         u64 entry, gentry;
2465         u64 *spte;
2466         unsigned offset = offset_in_page(gpa);
2467         unsigned pte_size;
2468         unsigned page_offset;
2469         unsigned misaligned;
2470         unsigned quadrant;
2471         int level;
2472         int flooded = 0;
2473         int npte;
2474         int r;
2475
2476         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2477         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2478         spin_lock(&vcpu->kvm->mmu_lock);
2479         kvm_mmu_access_page(vcpu, gfn);
2480         kvm_mmu_free_some_pages(vcpu);
2481         ++vcpu->kvm->stat.mmu_pte_write;
2482         kvm_mmu_audit(vcpu, "pre pte write");
2483         if (guest_initiated) {
2484                 if (gfn == vcpu->arch.last_pt_write_gfn
2485                     && !last_updated_pte_accessed(vcpu)) {
2486                         ++vcpu->arch.last_pt_write_count;
2487                         if (vcpu->arch.last_pt_write_count >= 3)
2488                                 flooded = 1;
2489                 } else {
2490                         vcpu->arch.last_pt_write_gfn = gfn;
2491                         vcpu->arch.last_pt_write_count = 1;
2492                         vcpu->arch.last_pte_updated = NULL;
2493                 }
2494         }
2495         index = kvm_page_table_hashfn(gfn);
2496         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2497         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2498                 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
2499                         continue;
2500                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2501                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2502                 misaligned |= bytes < 4;
2503                 if (misaligned || flooded) {
2504                         /*
2505                          * Misaligned accesses are too much trouble to fix
2506                          * up; also, they usually indicate a page is not used
2507                          * as a page table.
2508                          *
2509                          * If we're seeing too many writes to a page,
2510                          * it may no longer be a page table, or we may be
2511                          * forking, in which case it is better to unmap the
2512                          * page.
2513                          */
2514                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2515                                  gpa, bytes, sp->role.word);
2516                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2517                                 n = bucket->first;
2518                         ++vcpu->kvm->stat.mmu_flooded;
2519                         continue;
2520                 }
2521                 page_offset = offset;
2522                 level = sp->role.level;
2523                 npte = 1;
2524                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2525                         page_offset <<= 1;      /* 32->64 */
2526                         /*
2527                          * A 32-bit pde maps 4MB while the shadow pdes map
2528                          * only 2MB.  So we need to double the offset again
2529                          * and zap two pdes instead of one.
2530                          */
2531                         if (level == PT32_ROOT_LEVEL) {
2532                                 page_offset &= ~7; /* kill rounding error */
2533                                 page_offset <<= 1;
2534                                 npte = 2;
2535                         }
2536                         quadrant = page_offset >> PAGE_SHIFT;
2537                         page_offset &= ~PAGE_MASK;
2538                         if (quadrant != sp->role.quadrant)
2539                                 continue;
2540                 }
2541                 spte = &sp->spt[page_offset / sizeof(*spte)];
2542                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2543                         gentry = 0;
2544                         r = kvm_read_guest_atomic(vcpu->kvm,
2545                                                   gpa & ~(u64)(pte_size - 1),
2546                                                   &gentry, pte_size);
2547                         new = (const void *)&gentry;
2548                         if (r < 0)
2549                                 new = NULL;
2550                 }
2551                 while (npte--) {
2552                         entry = *spte;
2553                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2554                         if (new)
2555                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2556                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2557                         ++spte;
2558                 }
2559         }
2560         kvm_mmu_audit(vcpu, "post pte write");
2561         spin_unlock(&vcpu->kvm->mmu_lock);
2562         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2563                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2564                 vcpu->arch.update_pte.pfn = bad_pfn;
2565         }
2566 }
2567
2568 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2569 {
2570         gpa_t gpa;
2571         int r;
2572
2573         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2574
2575         spin_lock(&vcpu->kvm->mmu_lock);
2576         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2577         spin_unlock(&vcpu->kvm->mmu_lock);
2578         return r;
2579 }
2580 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2581
2582 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2583 {
2584         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2585                 struct kvm_mmu_page *sp;
2586
2587                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2588                                   struct kvm_mmu_page, link);
2589                 kvm_mmu_zap_page(vcpu->kvm, sp);
2590                 ++vcpu->kvm->stat.mmu_recycled;
2591         }
2592 }
2593
2594 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2595 {
2596         int r;
2597         enum emulation_result er;
2598
2599         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2600         if (r < 0)
2601                 goto out;
2602
2603         if (!r) {
2604                 r = 1;
2605                 goto out;
2606         }
2607
2608         r = mmu_topup_memory_caches(vcpu);
2609         if (r)
2610                 goto out;
2611
2612         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2613
2614         switch (er) {
2615         case EMULATE_DONE:
2616                 return 1;
2617         case EMULATE_DO_MMIO:
2618                 ++vcpu->stat.mmio_exits;
2619                 return 0;
2620         case EMULATE_FAIL:
2621                 kvm_report_emulation_failure(vcpu, "pagetable");
2622                 return 1;
2623         default:
2624                 BUG();
2625         }
2626 out:
2627         return r;
2628 }
2629 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2630
2631 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2632 {
2633         vcpu->arch.mmu.invlpg(vcpu, gva);
2634         kvm_mmu_flush_tlb(vcpu);
2635         ++vcpu->stat.invlpg;
2636 }
2637 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2638
2639 void kvm_enable_tdp(void)
2640 {
2641         tdp_enabled = true;
2642 }
2643 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2644
2645 void kvm_disable_tdp(void)
2646 {
2647         tdp_enabled = false;
2648 }
2649 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2650
2651 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2652 {
2653         struct kvm_mmu_page *sp;
2654
2655         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2656                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2657                                   struct kvm_mmu_page, link);
2658                 kvm_mmu_zap_page(vcpu->kvm, sp);
2659                 cond_resched();
2660         }
2661         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2662 }
2663
2664 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2665 {
2666         struct page *page;
2667         int i;
2668
2669         ASSERT(vcpu);
2670
2671         if (vcpu->kvm->arch.n_requested_mmu_pages)
2672                 vcpu->kvm->arch.n_free_mmu_pages =
2673                                         vcpu->kvm->arch.n_requested_mmu_pages;
2674         else
2675                 vcpu->kvm->arch.n_free_mmu_pages =
2676                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2677         /*
2678          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2679          * Therefore we need to allocate shadow page tables in the first
2680          * 4GB of memory, which happens to fit the DMA32 zone.
2681          */
2682         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2683         if (!page)
2684                 goto error_1;
2685         vcpu->arch.mmu.pae_root = page_address(page);
2686         for (i = 0; i < 4; ++i)
2687                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2688
2689         return 0;
2690
2691 error_1:
2692         free_mmu_pages(vcpu);
2693         return -ENOMEM;
2694 }
2695
2696 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2697 {
2698         ASSERT(vcpu);
2699         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2700
2701         return alloc_mmu_pages(vcpu);
2702 }
2703
2704 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2705 {
2706         ASSERT(vcpu);
2707         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2708
2709         return init_kvm_mmu(vcpu);
2710 }
2711
2712 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2713 {
2714         ASSERT(vcpu);
2715
2716         destroy_kvm_mmu(vcpu);
2717         free_mmu_pages(vcpu);
2718         mmu_free_memory_caches(vcpu);
2719 }
2720
2721 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2722 {
2723         struct kvm_mmu_page *sp;
2724
2725         spin_lock(&kvm->mmu_lock);
2726         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2727                 int i;
2728                 u64 *pt;
2729
2730                 if (!test_bit(slot, sp->slot_bitmap))
2731                         continue;
2732
2733                 pt = sp->spt;
2734                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2735                         /* avoid RMW */
2736                         if (pt[i] & PT_WRITABLE_MASK)
2737                                 pt[i] &= ~PT_WRITABLE_MASK;
2738         }
2739         kvm_flush_remote_tlbs(kvm);
2740         spin_unlock(&kvm->mmu_lock);
2741 }
2742
2743 void kvm_mmu_zap_all(struct kvm *kvm)
2744 {
2745         struct kvm_mmu_page *sp, *node;
2746
2747         spin_lock(&kvm->mmu_lock);
2748         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2749                 if (kvm_mmu_zap_page(kvm, sp))
2750                         node = container_of(kvm->arch.active_mmu_pages.next,
2751                                             struct kvm_mmu_page, link);
2752         spin_unlock(&kvm->mmu_lock);
2753
2754         kvm_flush_remote_tlbs(kvm);
2755 }
2756
2757 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2758 {
2759         struct kvm_mmu_page *page;
2760
2761         page = container_of(kvm->arch.active_mmu_pages.prev,
2762                             struct kvm_mmu_page, link);
2763         kvm_mmu_zap_page(kvm, page);
2764 }
2765
2766 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2767 {
2768         struct kvm *kvm;
2769         struct kvm *kvm_freed = NULL;
2770         int cache_count = 0;
2771
2772         spin_lock(&kvm_lock);
2773
2774         list_for_each_entry(kvm, &vm_list, vm_list) {
2775                 int npages;
2776
2777                 if (!down_read_trylock(&kvm->slots_lock))
2778                         continue;
2779                 spin_lock(&kvm->mmu_lock);
2780                 npages = kvm->arch.n_alloc_mmu_pages -
2781                          kvm->arch.n_free_mmu_pages;
2782                 cache_count += npages;
2783                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2784                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2785                         cache_count--;
2786                         kvm_freed = kvm;
2787                 }
2788                 nr_to_scan--;
2789
2790                 spin_unlock(&kvm->mmu_lock);
2791                 up_read(&kvm->slots_lock);
2792         }
2793         if (kvm_freed)
2794                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2795
2796         spin_unlock(&kvm_lock);
2797
2798         return cache_count;
2799 }
2800
2801 static struct shrinker mmu_shrinker = {
2802         .shrink = mmu_shrink,
2803         .seeks = DEFAULT_SEEKS * 10,
2804 };
2805
2806 static void mmu_destroy_caches(void)
2807 {
2808         if (pte_chain_cache)
2809                 kmem_cache_destroy(pte_chain_cache);
2810         if (rmap_desc_cache)
2811                 kmem_cache_destroy(rmap_desc_cache);
2812         if (mmu_page_header_cache)
2813                 kmem_cache_destroy(mmu_page_header_cache);
2814 }
2815
2816 void kvm_mmu_module_exit(void)
2817 {
2818         mmu_destroy_caches();
2819         unregister_shrinker(&mmu_shrinker);
2820 }
2821
2822 int kvm_mmu_module_init(void)
2823 {
2824         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2825                                             sizeof(struct kvm_pte_chain),
2826                                             0, 0, NULL);
2827         if (!pte_chain_cache)
2828                 goto nomem;
2829         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2830                                             sizeof(struct kvm_rmap_desc),
2831                                             0, 0, NULL);
2832         if (!rmap_desc_cache)
2833                 goto nomem;
2834
2835         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2836                                                   sizeof(struct kvm_mmu_page),
2837                                                   0, 0, NULL);
2838         if (!mmu_page_header_cache)
2839                 goto nomem;
2840
2841         register_shrinker(&mmu_shrinker);
2842
2843         return 0;
2844
2845 nomem:
2846         mmu_destroy_caches();
2847         return -ENOMEM;
2848 }
2849
2850 /*
2851  * Caculate mmu pages needed for kvm.
2852  */
2853 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2854 {
2855         int i;
2856         unsigned int nr_mmu_pages;
2857         unsigned int  nr_pages = 0;
2858
2859         for (i = 0; i < kvm->nmemslots; i++)
2860                 nr_pages += kvm->memslots[i].npages;
2861
2862         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2863         nr_mmu_pages = max(nr_mmu_pages,
2864                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2865
2866         return nr_mmu_pages;
2867 }
2868
2869 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2870                                 unsigned len)
2871 {
2872         if (len > buffer->len)
2873                 return NULL;
2874         return buffer->ptr;
2875 }
2876
2877 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2878                                 unsigned len)
2879 {
2880         void *ret;
2881
2882         ret = pv_mmu_peek_buffer(buffer, len);
2883         if (!ret)
2884                 return ret;
2885         buffer->ptr += len;
2886         buffer->len -= len;
2887         buffer->processed += len;
2888         return ret;
2889 }
2890
2891 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2892                              gpa_t addr, gpa_t value)
2893 {
2894         int bytes = 8;
2895         int r;
2896
2897         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2898                 bytes = 4;
2899
2900         r = mmu_topup_memory_caches(vcpu);
2901         if (r)
2902                 return r;
2903
2904         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2905                 return -EFAULT;
2906
2907         return 1;
2908 }
2909
2910 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2911 {
2912         kvm_x86_ops->tlb_flush(vcpu);
2913         set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
2914         return 1;
2915 }
2916
2917 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2918 {
2919         spin_lock(&vcpu->kvm->mmu_lock);
2920         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2921         spin_unlock(&vcpu->kvm->mmu_lock);
2922         return 1;
2923 }
2924
2925 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2926                              struct kvm_pv_mmu_op_buffer *buffer)
2927 {
2928         struct kvm_mmu_op_header *header;
2929
2930         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2931         if (!header)
2932                 return 0;
2933         switch (header->op) {
2934         case KVM_MMU_OP_WRITE_PTE: {
2935                 struct kvm_mmu_op_write_pte *wpte;
2936
2937                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2938                 if (!wpte)
2939                         return 0;
2940                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2941                                         wpte->pte_val);
2942         }
2943         case KVM_MMU_OP_FLUSH_TLB: {
2944                 struct kvm_mmu_op_flush_tlb *ftlb;
2945
2946                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2947                 if (!ftlb)
2948                         return 0;
2949                 return kvm_pv_mmu_flush_tlb(vcpu);
2950         }
2951         case KVM_MMU_OP_RELEASE_PT: {
2952                 struct kvm_mmu_op_release_pt *rpt;
2953
2954                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2955                 if (!rpt)
2956                         return 0;
2957                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2958         }
2959         default: return 0;
2960         }
2961 }
2962
2963 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2964                   gpa_t addr, unsigned long *ret)
2965 {
2966         int r;
2967         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2968
2969         buffer->ptr = buffer->buf;
2970         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2971         buffer->processed = 0;
2972
2973         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2974         if (r)
2975                 goto out;
2976
2977         while (buffer->len) {
2978                 r = kvm_pv_mmu_op_one(vcpu, buffer);
2979                 if (r < 0)
2980                         goto out;
2981                 if (r == 0)
2982                         break;
2983         }
2984
2985         r = 1;
2986 out:
2987         *ret = buffer->processed;
2988         return r;
2989 }
2990
2991 #ifdef AUDIT
2992
2993 static const char *audit_msg;
2994
2995 static gva_t canonicalize(gva_t gva)
2996 {
2997 #ifdef CONFIG_X86_64
2998         gva = (long long)(gva << 16) >> 16;
2999 #endif
3000         return gva;
3001 }
3002
3003 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3004                                 gva_t va, int level)
3005 {
3006         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3007         int i;
3008         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3009
3010         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3011                 u64 ent = pt[i];
3012
3013                 if (ent == shadow_trap_nonpresent_pte)
3014                         continue;
3015
3016                 va = canonicalize(va);
3017                 if (level > 1) {
3018                         if (ent == shadow_notrap_nonpresent_pte)
3019                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
3020                                        " in nonleaf level: levels %d gva %lx"
3021                                        " level %d pte %llx\n", audit_msg,
3022                                        vcpu->arch.mmu.root_level, va, level, ent);
3023
3024                         audit_mappings_page(vcpu, ent, va, level - 1);
3025                 } else {
3026                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3027                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
3028
3029                         if (is_shadow_present_pte(ent)
3030                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3031                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3032                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3033                                        audit_msg, vcpu->arch.mmu.root_level,
3034                                        va, gpa, hpa, ent,
3035                                        is_shadow_present_pte(ent));
3036                         else if (ent == shadow_notrap_nonpresent_pte
3037                                  && !is_error_hpa(hpa))
3038                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3039                                        " valid guest gva %lx\n", audit_msg, va);
3040                         kvm_release_pfn_clean(pfn);
3041
3042                 }
3043         }
3044 }
3045
3046 static void audit_mappings(struct kvm_vcpu *vcpu)
3047 {
3048         unsigned i;
3049
3050         if (vcpu->arch.mmu.root_level == 4)
3051                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3052         else
3053                 for (i = 0; i < 4; ++i)
3054                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3055                                 audit_mappings_page(vcpu,
3056                                                     vcpu->arch.mmu.pae_root[i],
3057                                                     i << 30,
3058                                                     2);
3059 }
3060
3061 static int count_rmaps(struct kvm_vcpu *vcpu)
3062 {
3063         int nmaps = 0;
3064         int i, j, k;
3065
3066         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3067                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3068                 struct kvm_rmap_desc *d;
3069
3070                 for (j = 0; j < m->npages; ++j) {
3071                         unsigned long *rmapp = &m->rmap[j];
3072
3073                         if (!*rmapp)
3074                                 continue;
3075                         if (!(*rmapp & 1)) {
3076                                 ++nmaps;
3077                                 continue;
3078                         }
3079                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3080                         while (d) {
3081                                 for (k = 0; k < RMAP_EXT; ++k)
3082                                         if (d->shadow_ptes[k])
3083                                                 ++nmaps;
3084                                         else
3085                                                 break;
3086                                 d = d->more;
3087                         }
3088                 }
3089         }
3090         return nmaps;
3091 }
3092
3093 static int count_writable_mappings(struct kvm_vcpu *vcpu)
3094 {
3095         int nmaps = 0;
3096         struct kvm_mmu_page *sp;
3097         int i;
3098
3099         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3100                 u64 *pt = sp->spt;
3101
3102                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3103                         continue;
3104
3105                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3106                         u64 ent = pt[i];
3107
3108                         if (!(ent & PT_PRESENT_MASK))
3109                                 continue;
3110                         if (!(ent & PT_WRITABLE_MASK))
3111                                 continue;
3112                         ++nmaps;
3113                 }
3114         }
3115         return nmaps;
3116 }
3117
3118 static void audit_rmap(struct kvm_vcpu *vcpu)
3119 {
3120         int n_rmap = count_rmaps(vcpu);
3121         int n_actual = count_writable_mappings(vcpu);
3122
3123         if (n_rmap != n_actual)
3124                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
3125                        __func__, audit_msg, n_rmap, n_actual);
3126 }
3127
3128 static void audit_write_protection(struct kvm_vcpu *vcpu)
3129 {
3130         struct kvm_mmu_page *sp;
3131         struct kvm_memory_slot *slot;
3132         unsigned long *rmapp;
3133         gfn_t gfn;
3134
3135         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3136                 if (sp->role.metaphysical)
3137                         continue;
3138
3139                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3140                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3141                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3142                 if (*rmapp)
3143                         printk(KERN_ERR "%s: (%s) shadow page has writable"
3144                                " mappings: gfn %lx role %x\n",
3145                                __func__, audit_msg, sp->gfn,
3146                                sp->role.word);
3147         }
3148 }
3149
3150 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3151 {
3152         int olddbg = dbg;
3153
3154         dbg = 0;
3155         audit_msg = msg;
3156         audit_rmap(vcpu);
3157         audit_write_protection(vcpu);
3158         audit_mappings(vcpu);
3159         dbg = olddbg;
3160 }
3161
3162 #endif