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