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

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "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 int 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         return 0;
1099 }
1100
1101 /*
1102  * Changing the number of mmu pages allocated to the vm
1103  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1104  */
1105 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1106 {
1107         /*
1108          * If we set the number of mmu pages to be smaller be than the
1109          * number of actived pages , we must to free some mmu pages before we
1110          * change the value
1111          */
1112
1113         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1114             kvm_nr_mmu_pages) {
1115                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1116                                        - kvm->arch.n_free_mmu_pages;
1117
1118                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1119                         struct kvm_mmu_page *page;
1120
1121                         page = container_of(kvm->arch.active_mmu_pages.prev,
1122                                             struct kvm_mmu_page, link);
1123                         kvm_mmu_zap_page(kvm, page);
1124                         n_used_mmu_pages--;
1125                 }
1126                 kvm->arch.n_free_mmu_pages = 0;
1127         }
1128         else
1129                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1130                                          - kvm->arch.n_alloc_mmu_pages;
1131
1132         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1133 }
1134
1135 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1136 {
1137         unsigned index;
1138         struct hlist_head *bucket;
1139         struct kvm_mmu_page *sp;
1140         struct hlist_node *node, *n;
1141         int r;
1142
1143         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1144         r = 0;
1145         index = kvm_page_table_hashfn(gfn);
1146         bucket = &kvm->arch.mmu_page_hash[index];
1147         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1148                 if (sp->gfn == gfn && !sp->role.metaphysical) {
1149                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1150                                  sp->role.word);
1151                         r = 1;
1152                         if (kvm_mmu_zap_page(kvm, sp))
1153                                 n = bucket->first;
1154                 }
1155         return r;
1156 }
1157
1158 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1159 {
1160         struct kvm_mmu_page *sp;
1161
1162         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1163                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1164                 kvm_mmu_zap_page(kvm, sp);
1165         }
1166 }
1167
1168 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1169 {
1170         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1171         struct kvm_mmu_page *sp = page_header(__pa(pte));
1172
1173         __set_bit(slot, &sp->slot_bitmap);
1174 }
1175
1176 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1177 {
1178         int i;
1179         u64 *pt = sp->spt;
1180
1181         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1182                 return;
1183
1184         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1185                 if (pt[i] == shadow_notrap_nonpresent_pte)
1186                         set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1187         }
1188 }
1189
1190 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1191 {
1192         struct page *page;
1193
1194         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1195
1196         if (gpa == UNMAPPED_GVA)
1197                 return NULL;
1198
1199         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1200
1201         return page;
1202 }
1203
1204 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1205                     unsigned pte_access, int user_fault,
1206                     int write_fault, int dirty, int largepage,
1207                     gfn_t gfn, pfn_t pfn, bool speculative)
1208 {
1209         u64 spte;
1210         int ret = 0;
1211         /*
1212          * We don't set the accessed bit, since we sometimes want to see
1213          * whether the guest actually used the pte (in order to detect
1214          * demand paging).
1215          */
1216         spte = shadow_base_present_pte | shadow_dirty_mask;
1217         if (!speculative)
1218                 spte |= shadow_accessed_mask;
1219         if (!dirty)
1220                 pte_access &= ~ACC_WRITE_MASK;
1221         if (pte_access & ACC_EXEC_MASK)
1222                 spte |= shadow_x_mask;
1223         else
1224                 spte |= shadow_nx_mask;
1225         if (pte_access & ACC_USER_MASK)
1226                 spte |= shadow_user_mask;
1227         if (largepage)
1228                 spte |= PT_PAGE_SIZE_MASK;
1229
1230         spte |= (u64)pfn << PAGE_SHIFT;
1231
1232         if ((pte_access & ACC_WRITE_MASK)
1233             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1234                 struct kvm_mmu_page *shadow;
1235
1236                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1237                         ret = 1;
1238                         spte = shadow_trap_nonpresent_pte;
1239                         goto set_pte;
1240                 }
1241
1242                 spte |= PT_WRITABLE_MASK;
1243
1244                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1245                 if (shadow) {
1246                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1247                                  __func__, gfn);
1248                         ret = 1;
1249                         pte_access &= ~ACC_WRITE_MASK;
1250                         if (is_writeble_pte(spte))
1251                                 spte &= ~PT_WRITABLE_MASK;
1252                 }
1253         }
1254
1255         if (pte_access & ACC_WRITE_MASK)
1256                 mark_page_dirty(vcpu->kvm, gfn);
1257
1258 set_pte:
1259         set_shadow_pte(shadow_pte, spte);
1260         return ret;
1261 }
1262
1263
1264 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1265                          unsigned pt_access, unsigned pte_access,
1266                          int user_fault, int write_fault, int dirty,
1267                          int *ptwrite, int largepage, gfn_t gfn,
1268                          pfn_t pfn, bool speculative)
1269 {
1270         int was_rmapped = 0;
1271         int was_writeble = is_writeble_pte(*shadow_pte);
1272
1273         pgprintk("%s: spte %llx access %x write_fault %d"
1274                  " user_fault %d gfn %lx\n",
1275                  __func__, *shadow_pte, pt_access,
1276                  write_fault, user_fault, gfn);
1277
1278         if (is_rmap_pte(*shadow_pte)) {
1279                 /*
1280                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1281                  * the parent of the now unreachable PTE.
1282                  */
1283                 if (largepage && !is_large_pte(*shadow_pte)) {
1284                         struct kvm_mmu_page *child;
1285                         u64 pte = *shadow_pte;
1286
1287                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1288                         mmu_page_remove_parent_pte(child, shadow_pte);
1289                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1290                         pgprintk("hfn old %lx new %lx\n",
1291                                  spte_to_pfn(*shadow_pte), pfn);
1292                         rmap_remove(vcpu->kvm, shadow_pte);
1293                 } else {
1294                         if (largepage)
1295                                 was_rmapped = is_large_pte(*shadow_pte);
1296                         else
1297                                 was_rmapped = 1;
1298                 }
1299         }
1300         if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1301                       dirty, largepage, gfn, pfn, speculative)) {
1302                 if (write_fault)
1303                         *ptwrite = 1;
1304                 kvm_x86_ops->tlb_flush(vcpu);
1305         }
1306
1307         pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1308         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1309                  is_large_pte(*shadow_pte)? "2MB" : "4kB",
1310                  is_present_pte(*shadow_pte)?"RW":"R", gfn,
1311                  *shadow_pte, shadow_pte);
1312         if (!was_rmapped && is_large_pte(*shadow_pte))
1313                 ++vcpu->kvm->stat.lpages;
1314
1315         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1316         if (!was_rmapped) {
1317                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1318                 if (!is_rmap_pte(*shadow_pte))
1319                         kvm_release_pfn_clean(pfn);
1320         } else {
1321                 if (was_writeble)
1322                         kvm_release_pfn_dirty(pfn);
1323                 else
1324                         kvm_release_pfn_clean(pfn);
1325         }
1326         if (speculative) {
1327                 vcpu->arch.last_pte_updated = shadow_pte;
1328                 vcpu->arch.last_pte_gfn = gfn;
1329         }
1330 }
1331
1332 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1333 {
1334 }
1335
1336 struct direct_shadow_walk {
1337         struct kvm_shadow_walk walker;
1338         pfn_t pfn;
1339         int write;
1340         int largepage;
1341         int pt_write;
1342 };
1343
1344 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1345                             struct kvm_vcpu *vcpu,
1346                             u64 addr, u64 *sptep, int level)
1347 {
1348         struct direct_shadow_walk *walk =
1349                 container_of(_walk, struct direct_shadow_walk, walker);
1350         struct kvm_mmu_page *sp;
1351         gfn_t pseudo_gfn;
1352         gfn_t gfn = addr >> PAGE_SHIFT;
1353
1354         if (level == PT_PAGE_TABLE_LEVEL
1355             || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1356                 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1357                              0, walk->write, 1, &walk->pt_write,
1358                              walk->largepage, gfn, walk->pfn, false);
1359                 ++vcpu->stat.pf_fixed;
1360                 return 1;
1361         }
1362
1363         if (*sptep == shadow_trap_nonpresent_pte) {
1364                 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1365                 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1366                                       1, ACC_ALL, sptep);
1367                 if (!sp) {
1368                         pgprintk("nonpaging_map: ENOMEM\n");
1369                         kvm_release_pfn_clean(walk->pfn);
1370                         return -ENOMEM;
1371                 }
1372
1373                 set_shadow_pte(sptep,
1374                                __pa(sp->spt)
1375                                | PT_PRESENT_MASK | PT_WRITABLE_MASK
1376                                | shadow_user_mask | shadow_x_mask);
1377         }
1378         return 0;
1379 }
1380
1381 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1382                         int largepage, gfn_t gfn, pfn_t pfn)
1383 {
1384         int r;
1385         struct direct_shadow_walk walker = {
1386                 .walker = { .entry = direct_map_entry, },
1387                 .pfn = pfn,
1388                 .largepage = largepage,
1389                 .write = write,
1390                 .pt_write = 0,
1391         };
1392
1393         r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1394         if (r < 0)
1395                 return r;
1396         return walker.pt_write;
1397 }
1398
1399 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1400 {
1401         int r;
1402         int largepage = 0;
1403         pfn_t pfn;
1404         unsigned long mmu_seq;
1405
1406         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1407                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1408                 largepage = 1;
1409         }
1410
1411         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1412         smp_rmb();
1413         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1414
1415         /* mmio */
1416         if (is_error_pfn(pfn)) {
1417                 kvm_release_pfn_clean(pfn);
1418                 return 1;
1419         }
1420
1421         spin_lock(&vcpu->kvm->mmu_lock);
1422         if (mmu_notifier_retry(vcpu, mmu_seq))
1423                 goto out_unlock;
1424         kvm_mmu_free_some_pages(vcpu);
1425         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1426         spin_unlock(&vcpu->kvm->mmu_lock);
1427
1428
1429         return r;
1430
1431 out_unlock:
1432         spin_unlock(&vcpu->kvm->mmu_lock);
1433         kvm_release_pfn_clean(pfn);
1434         return 0;
1435 }
1436
1437
1438 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1439 {
1440         int i;
1441         struct kvm_mmu_page *sp;
1442
1443         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1444                 return;
1445         spin_lock(&vcpu->kvm->mmu_lock);
1446         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1447                 hpa_t root = vcpu->arch.mmu.root_hpa;
1448
1449                 sp = page_header(root);
1450                 --sp->root_count;
1451                 if (!sp->root_count && sp->role.invalid)
1452                         kvm_mmu_zap_page(vcpu->kvm, sp);
1453                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1454                 spin_unlock(&vcpu->kvm->mmu_lock);
1455                 return;
1456         }
1457         for (i = 0; i < 4; ++i) {
1458                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1459
1460                 if (root) {
1461                         root &= PT64_BASE_ADDR_MASK;
1462                         sp = page_header(root);
1463                         --sp->root_count;
1464                         if (!sp->root_count && sp->role.invalid)
1465                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1466                 }
1467                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1468         }
1469         spin_unlock(&vcpu->kvm->mmu_lock);
1470         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1471 }
1472
1473 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1474 {
1475         int i;
1476         gfn_t root_gfn;
1477         struct kvm_mmu_page *sp;
1478         int metaphysical = 0;
1479
1480         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1481
1482         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1483                 hpa_t root = vcpu->arch.mmu.root_hpa;
1484
1485                 ASSERT(!VALID_PAGE(root));
1486                 if (tdp_enabled)
1487                         metaphysical = 1;
1488                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1489                                       PT64_ROOT_LEVEL, metaphysical,
1490                                       ACC_ALL, NULL);
1491                 root = __pa(sp->spt);
1492                 ++sp->root_count;
1493                 vcpu->arch.mmu.root_hpa = root;
1494                 return;
1495         }
1496         metaphysical = !is_paging(vcpu);
1497         if (tdp_enabled)
1498                 metaphysical = 1;
1499         for (i = 0; i < 4; ++i) {
1500                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1501
1502                 ASSERT(!VALID_PAGE(root));
1503                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1504                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1505                                 vcpu->arch.mmu.pae_root[i] = 0;
1506                                 continue;
1507                         }
1508                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1509                 } else if (vcpu->arch.mmu.root_level == 0)
1510                         root_gfn = 0;
1511                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1512                                       PT32_ROOT_LEVEL, metaphysical,
1513                                       ACC_ALL, NULL);
1514                 root = __pa(sp->spt);
1515                 ++sp->root_count;
1516                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1517         }
1518         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1519 }
1520
1521 static void mmu_sync_children(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1522 {
1523 }
1524
1525 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
1526 {
1527         int i;
1528         struct kvm_mmu_page *sp;
1529
1530         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1531                 return;
1532         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1533                 hpa_t root = vcpu->arch.mmu.root_hpa;
1534                 sp = page_header(root);
1535                 mmu_sync_children(vcpu, sp);
1536                 return;
1537         }
1538         for (i = 0; i < 4; ++i) {
1539                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1540
1541                 if (root) {
1542                         root &= PT64_BASE_ADDR_MASK;
1543                         sp = page_header(root);
1544                         mmu_sync_children(vcpu, sp);
1545                 }
1546         }
1547 }
1548
1549 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
1550 {
1551         spin_lock(&vcpu->kvm->mmu_lock);
1552         mmu_sync_roots(vcpu);
1553         spin_unlock(&vcpu->kvm->mmu_lock);
1554 }
1555
1556 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1557 {
1558         return vaddr;
1559 }
1560
1561 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1562                                 u32 error_code)
1563 {
1564         gfn_t gfn;
1565         int r;
1566
1567         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1568         r = mmu_topup_memory_caches(vcpu);
1569         if (r)
1570                 return r;
1571
1572         ASSERT(vcpu);
1573         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1574
1575         gfn = gva >> PAGE_SHIFT;
1576
1577         return nonpaging_map(vcpu, gva & PAGE_MASK,
1578                              error_code & PFERR_WRITE_MASK, gfn);
1579 }
1580
1581 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1582                                 u32 error_code)
1583 {
1584         pfn_t pfn;
1585         int r;
1586         int largepage = 0;
1587         gfn_t gfn = gpa >> PAGE_SHIFT;
1588         unsigned long mmu_seq;
1589
1590         ASSERT(vcpu);
1591         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1592
1593         r = mmu_topup_memory_caches(vcpu);
1594         if (r)
1595                 return r;
1596
1597         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1598                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1599                 largepage = 1;
1600         }
1601         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1602         smp_rmb();
1603         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1604         if (is_error_pfn(pfn)) {
1605                 kvm_release_pfn_clean(pfn);
1606                 return 1;
1607         }
1608         spin_lock(&vcpu->kvm->mmu_lock);
1609         if (mmu_notifier_retry(vcpu, mmu_seq))
1610                 goto out_unlock;
1611         kvm_mmu_free_some_pages(vcpu);
1612         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1613                          largepage, gfn, pfn);
1614         spin_unlock(&vcpu->kvm->mmu_lock);
1615
1616         return r;
1617
1618 out_unlock:
1619         spin_unlock(&vcpu->kvm->mmu_lock);
1620         kvm_release_pfn_clean(pfn);
1621         return 0;
1622 }
1623
1624 static void nonpaging_free(struct kvm_vcpu *vcpu)
1625 {
1626         mmu_free_roots(vcpu);
1627 }
1628
1629 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1630 {
1631         struct kvm_mmu *context = &vcpu->arch.mmu;
1632
1633         context->new_cr3 = nonpaging_new_cr3;
1634         context->page_fault = nonpaging_page_fault;
1635         context->gva_to_gpa = nonpaging_gva_to_gpa;
1636         context->free = nonpaging_free;
1637         context->prefetch_page = nonpaging_prefetch_page;
1638         context->sync_page = nonpaging_sync_page;
1639         context->invlpg = nonpaging_invlpg;
1640         context->root_level = 0;
1641         context->shadow_root_level = PT32E_ROOT_LEVEL;
1642         context->root_hpa = INVALID_PAGE;
1643         return 0;
1644 }
1645
1646 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1647 {
1648         ++vcpu->stat.tlb_flush;
1649         kvm_x86_ops->tlb_flush(vcpu);
1650 }
1651
1652 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1653 {
1654         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1655         mmu_free_roots(vcpu);
1656 }
1657
1658 static void inject_page_fault(struct kvm_vcpu *vcpu,
1659                               u64 addr,
1660                               u32 err_code)
1661 {
1662         kvm_inject_page_fault(vcpu, addr, err_code);
1663 }
1664
1665 static void paging_free(struct kvm_vcpu *vcpu)
1666 {
1667         nonpaging_free(vcpu);
1668 }
1669
1670 #define PTTYPE 64
1671 #include "paging_tmpl.h"
1672 #undef PTTYPE
1673
1674 #define PTTYPE 32
1675 #include "paging_tmpl.h"
1676 #undef PTTYPE
1677
1678 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1679 {
1680         struct kvm_mmu *context = &vcpu->arch.mmu;
1681
1682         ASSERT(is_pae(vcpu));
1683         context->new_cr3 = paging_new_cr3;
1684         context->page_fault = paging64_page_fault;
1685         context->gva_to_gpa = paging64_gva_to_gpa;
1686         context->prefetch_page = paging64_prefetch_page;
1687         context->sync_page = paging64_sync_page;
1688         context->invlpg = paging64_invlpg;
1689         context->free = paging_free;
1690         context->root_level = level;
1691         context->shadow_root_level = level;
1692         context->root_hpa = INVALID_PAGE;
1693         return 0;
1694 }
1695
1696 static int paging64_init_context(struct kvm_vcpu *vcpu)
1697 {
1698         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1699 }
1700
1701 static int paging32_init_context(struct kvm_vcpu *vcpu)
1702 {
1703         struct kvm_mmu *context = &vcpu->arch.mmu;
1704
1705         context->new_cr3 = paging_new_cr3;
1706         context->page_fault = paging32_page_fault;
1707         context->gva_to_gpa = paging32_gva_to_gpa;
1708         context->free = paging_free;
1709         context->prefetch_page = paging32_prefetch_page;
1710         context->sync_page = paging32_sync_page;
1711         context->invlpg = paging32_invlpg;
1712         context->root_level = PT32_ROOT_LEVEL;
1713         context->shadow_root_level = PT32E_ROOT_LEVEL;
1714         context->root_hpa = INVALID_PAGE;
1715         return 0;
1716 }
1717
1718 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1719 {
1720         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1721 }
1722
1723 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1724 {
1725         struct kvm_mmu *context = &vcpu->arch.mmu;
1726
1727         context->new_cr3 = nonpaging_new_cr3;
1728         context->page_fault = tdp_page_fault;
1729         context->free = nonpaging_free;
1730         context->prefetch_page = nonpaging_prefetch_page;
1731         context->sync_page = nonpaging_sync_page;
1732         context->invlpg = nonpaging_invlpg;
1733         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
1734         context->root_hpa = INVALID_PAGE;
1735
1736         if (!is_paging(vcpu)) {
1737                 context->gva_to_gpa = nonpaging_gva_to_gpa;
1738                 context->root_level = 0;
1739         } else if (is_long_mode(vcpu)) {
1740                 context->gva_to_gpa = paging64_gva_to_gpa;
1741                 context->root_level = PT64_ROOT_LEVEL;
1742         } else if (is_pae(vcpu)) {
1743                 context->gva_to_gpa = paging64_gva_to_gpa;
1744                 context->root_level = PT32E_ROOT_LEVEL;
1745         } else {
1746                 context->gva_to_gpa = paging32_gva_to_gpa;
1747                 context->root_level = PT32_ROOT_LEVEL;
1748         }
1749
1750         return 0;
1751 }
1752
1753 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1754 {
1755         ASSERT(vcpu);
1756         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1757
1758         if (!is_paging(vcpu))
1759                 return nonpaging_init_context(vcpu);
1760         else if (is_long_mode(vcpu))
1761                 return paging64_init_context(vcpu);
1762         else if (is_pae(vcpu))
1763                 return paging32E_init_context(vcpu);
1764         else
1765                 return paging32_init_context(vcpu);
1766 }
1767
1768 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1769 {
1770         vcpu->arch.update_pte.pfn = bad_pfn;
1771
1772         if (tdp_enabled)
1773                 return init_kvm_tdp_mmu(vcpu);
1774         else
1775                 return init_kvm_softmmu(vcpu);
1776 }
1777
1778 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1779 {
1780         ASSERT(vcpu);
1781         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1782                 vcpu->arch.mmu.free(vcpu);
1783                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1784         }
1785 }
1786
1787 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1788 {
1789         destroy_kvm_mmu(vcpu);
1790         return init_kvm_mmu(vcpu);
1791 }
1792 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1793
1794 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1795 {
1796         int r;
1797
1798         r = mmu_topup_memory_caches(vcpu);
1799         if (r)
1800                 goto out;
1801         spin_lock(&vcpu->kvm->mmu_lock);
1802         kvm_mmu_free_some_pages(vcpu);
1803         mmu_alloc_roots(vcpu);
1804         mmu_sync_roots(vcpu);
1805         spin_unlock(&vcpu->kvm->mmu_lock);
1806         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1807         kvm_mmu_flush_tlb(vcpu);
1808 out:
1809         return r;
1810 }
1811 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1812
1813 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1814 {
1815         mmu_free_roots(vcpu);
1816 }
1817
1818 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1819                                   struct kvm_mmu_page *sp,
1820                                   u64 *spte)
1821 {
1822         u64 pte;
1823         struct kvm_mmu_page *child;
1824
1825         pte = *spte;
1826         if (is_shadow_present_pte(pte)) {
1827                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1828                     is_large_pte(pte))
1829                         rmap_remove(vcpu->kvm, spte);
1830                 else {
1831                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1832                         mmu_page_remove_parent_pte(child, spte);
1833                 }
1834         }
1835         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1836         if (is_large_pte(pte))
1837                 --vcpu->kvm->stat.lpages;
1838 }
1839
1840 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1841                                   struct kvm_mmu_page *sp,
1842                                   u64 *spte,
1843                                   const void *new)
1844 {
1845         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1846                 if (!vcpu->arch.update_pte.largepage ||
1847                     sp->role.glevels == PT32_ROOT_LEVEL) {
1848                         ++vcpu->kvm->stat.mmu_pde_zapped;
1849                         return;
1850                 }
1851         }
1852
1853         ++vcpu->kvm->stat.mmu_pte_updated;
1854         if (sp->role.glevels == PT32_ROOT_LEVEL)
1855                 paging32_update_pte(vcpu, sp, spte, new);
1856         else
1857                 paging64_update_pte(vcpu, sp, spte, new);
1858 }
1859
1860 static bool need_remote_flush(u64 old, u64 new)
1861 {
1862         if (!is_shadow_present_pte(old))
1863                 return false;
1864         if (!is_shadow_present_pte(new))
1865                 return true;
1866         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1867                 return true;
1868         old ^= PT64_NX_MASK;
1869         new ^= PT64_NX_MASK;
1870         return (old & ~new & PT64_PERM_MASK) != 0;
1871 }
1872
1873 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1874 {
1875         if (need_remote_flush(old, new))
1876                 kvm_flush_remote_tlbs(vcpu->kvm);
1877         else
1878                 kvm_mmu_flush_tlb(vcpu);
1879 }
1880
1881 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1882 {
1883         u64 *spte = vcpu->arch.last_pte_updated;
1884
1885         return !!(spte && (*spte & shadow_accessed_mask));
1886 }
1887
1888 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1889                                           const u8 *new, int bytes)
1890 {
1891         gfn_t gfn;
1892         int r;
1893         u64 gpte = 0;
1894         pfn_t pfn;
1895
1896         vcpu->arch.update_pte.largepage = 0;
1897
1898         if (bytes != 4 && bytes != 8)
1899                 return;
1900
1901         /*
1902          * Assume that the pte write on a page table of the same type
1903          * as the current vcpu paging mode.  This is nearly always true
1904          * (might be false while changing modes).  Note it is verified later
1905          * by update_pte().
1906          */
1907         if (is_pae(vcpu)) {
1908                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1909                 if ((bytes == 4) && (gpa % 4 == 0)) {
1910                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1911                         if (r)
1912                                 return;
1913                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1914                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1915                         memcpy((void *)&gpte, new, 8);
1916                 }
1917         } else {
1918                 if ((bytes == 4) && (gpa % 4 == 0))
1919                         memcpy((void *)&gpte, new, 4);
1920         }
1921         if (!is_present_pte(gpte))
1922                 return;
1923         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1924
1925         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1926                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1927                 vcpu->arch.update_pte.largepage = 1;
1928         }
1929         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
1930         smp_rmb();
1931         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1932
1933         if (is_error_pfn(pfn)) {
1934                 kvm_release_pfn_clean(pfn);
1935                 return;
1936         }
1937         vcpu->arch.update_pte.gfn = gfn;
1938         vcpu->arch.update_pte.pfn = pfn;
1939 }
1940
1941 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1942 {
1943         u64 *spte = vcpu->arch.last_pte_updated;
1944
1945         if (spte
1946             && vcpu->arch.last_pte_gfn == gfn
1947             && shadow_accessed_mask
1948             && !(*spte & shadow_accessed_mask)
1949             && is_shadow_present_pte(*spte))
1950                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
1951 }
1952
1953 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1954                        const u8 *new, int bytes)
1955 {
1956         gfn_t gfn = gpa >> PAGE_SHIFT;
1957         struct kvm_mmu_page *sp;
1958         struct hlist_node *node, *n;
1959         struct hlist_head *bucket;
1960         unsigned index;
1961         u64 entry, gentry;
1962         u64 *spte;
1963         unsigned offset = offset_in_page(gpa);
1964         unsigned pte_size;
1965         unsigned page_offset;
1966         unsigned misaligned;
1967         unsigned quadrant;
1968         int level;
1969         int flooded = 0;
1970         int npte;
1971         int r;
1972
1973         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1974         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1975         spin_lock(&vcpu->kvm->mmu_lock);
1976         kvm_mmu_access_page(vcpu, gfn);
1977         kvm_mmu_free_some_pages(vcpu);
1978         ++vcpu->kvm->stat.mmu_pte_write;
1979         kvm_mmu_audit(vcpu, "pre pte write");
1980         if (gfn == vcpu->arch.last_pt_write_gfn
1981             && !last_updated_pte_accessed(vcpu)) {
1982                 ++vcpu->arch.last_pt_write_count;
1983                 if (vcpu->arch.last_pt_write_count >= 3)
1984                         flooded = 1;
1985         } else {
1986                 vcpu->arch.last_pt_write_gfn = gfn;
1987                 vcpu->arch.last_pt_write_count = 1;
1988                 vcpu->arch.last_pte_updated = NULL;
1989         }
1990         index = kvm_page_table_hashfn(gfn);
1991         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1992         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1993                 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
1994                         continue;
1995                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1996                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1997                 misaligned |= bytes < 4;
1998                 if (misaligned || flooded) {
1999                         /*
2000                          * Misaligned accesses are too much trouble to fix
2001                          * up; also, they usually indicate a page is not used
2002                          * as a page table.
2003                          *
2004                          * If we're seeing too many writes to a page,
2005                          * it may no longer be a page table, or we may be
2006                          * forking, in which case it is better to unmap the
2007                          * page.
2008                          */
2009                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2010                                  gpa, bytes, sp->role.word);
2011                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2012                                 n = bucket->first;
2013                         ++vcpu->kvm->stat.mmu_flooded;
2014                         continue;
2015                 }
2016                 page_offset = offset;
2017                 level = sp->role.level;
2018                 npte = 1;
2019                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2020                         page_offset <<= 1;      /* 32->64 */
2021                         /*
2022                          * A 32-bit pde maps 4MB while the shadow pdes map
2023                          * only 2MB.  So we need to double the offset again
2024                          * and zap two pdes instead of one.
2025                          */
2026                         if (level == PT32_ROOT_LEVEL) {
2027                                 page_offset &= ~7; /* kill rounding error */
2028                                 page_offset <<= 1;
2029                                 npte = 2;
2030                         }
2031                         quadrant = page_offset >> PAGE_SHIFT;
2032                         page_offset &= ~PAGE_MASK;
2033                         if (quadrant != sp->role.quadrant)
2034                                 continue;
2035                 }
2036                 spte = &sp->spt[page_offset / sizeof(*spte)];
2037                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2038                         gentry = 0;
2039                         r = kvm_read_guest_atomic(vcpu->kvm,
2040                                                   gpa & ~(u64)(pte_size - 1),
2041                                                   &gentry, pte_size);
2042                         new = (const void *)&gentry;
2043                         if (r < 0)
2044                                 new = NULL;
2045                 }
2046                 while (npte--) {
2047                         entry = *spte;
2048                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2049                         if (new)
2050                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2051                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2052                         ++spte;
2053                 }
2054         }
2055         kvm_mmu_audit(vcpu, "post pte write");
2056         spin_unlock(&vcpu->kvm->mmu_lock);
2057         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2058                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2059                 vcpu->arch.update_pte.pfn = bad_pfn;
2060         }
2061 }
2062
2063 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2064 {
2065         gpa_t gpa;
2066         int r;
2067
2068         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2069
2070         spin_lock(&vcpu->kvm->mmu_lock);
2071         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2072         spin_unlock(&vcpu->kvm->mmu_lock);
2073         return r;
2074 }
2075 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2076
2077 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2078 {
2079         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2080                 struct kvm_mmu_page *sp;
2081
2082                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2083                                   struct kvm_mmu_page, link);
2084                 kvm_mmu_zap_page(vcpu->kvm, sp);
2085                 ++vcpu->kvm->stat.mmu_recycled;
2086         }
2087 }
2088
2089 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2090 {
2091         int r;
2092         enum emulation_result er;
2093
2094         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2095         if (r < 0)
2096                 goto out;
2097
2098         if (!r) {
2099                 r = 1;
2100                 goto out;
2101         }
2102
2103         r = mmu_topup_memory_caches(vcpu);
2104         if (r)
2105                 goto out;
2106
2107         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2108
2109         switch (er) {
2110         case EMULATE_DONE:
2111                 return 1;
2112         case EMULATE_DO_MMIO:
2113                 ++vcpu->stat.mmio_exits;
2114                 return 0;
2115         case EMULATE_FAIL:
2116                 kvm_report_emulation_failure(vcpu, "pagetable");
2117                 return 1;
2118         default:
2119                 BUG();
2120         }
2121 out:
2122         return r;
2123 }
2124 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2125
2126 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2127 {
2128         spin_lock(&vcpu->kvm->mmu_lock);
2129         vcpu->arch.mmu.invlpg(vcpu, gva);
2130         spin_unlock(&vcpu->kvm->mmu_lock);
2131         kvm_mmu_flush_tlb(vcpu);
2132         ++vcpu->stat.invlpg;
2133 }
2134 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2135
2136 void kvm_enable_tdp(void)
2137 {
2138         tdp_enabled = true;
2139 }
2140 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2141
2142 void kvm_disable_tdp(void)
2143 {
2144         tdp_enabled = false;
2145 }
2146 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2147
2148 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2149 {
2150         struct kvm_mmu_page *sp;
2151
2152         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2153                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2154                                   struct kvm_mmu_page, link);
2155                 kvm_mmu_zap_page(vcpu->kvm, sp);
2156                 cond_resched();
2157         }
2158         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2159 }
2160
2161 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2162 {
2163         struct page *page;
2164         int i;
2165
2166         ASSERT(vcpu);
2167
2168         if (vcpu->kvm->arch.n_requested_mmu_pages)
2169                 vcpu->kvm->arch.n_free_mmu_pages =
2170                                         vcpu->kvm->arch.n_requested_mmu_pages;
2171         else
2172                 vcpu->kvm->arch.n_free_mmu_pages =
2173                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2174         /*
2175          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2176          * Therefore we need to allocate shadow page tables in the first
2177          * 4GB of memory, which happens to fit the DMA32 zone.
2178          */
2179         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2180         if (!page)
2181                 goto error_1;
2182         vcpu->arch.mmu.pae_root = page_address(page);
2183         for (i = 0; i < 4; ++i)
2184                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2185
2186         return 0;
2187
2188 error_1:
2189         free_mmu_pages(vcpu);
2190         return -ENOMEM;
2191 }
2192
2193 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2194 {
2195         ASSERT(vcpu);
2196         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2197
2198         return alloc_mmu_pages(vcpu);
2199 }
2200
2201 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2202 {
2203         ASSERT(vcpu);
2204         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2205
2206         return init_kvm_mmu(vcpu);
2207 }
2208
2209 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2210 {
2211         ASSERT(vcpu);
2212
2213         destroy_kvm_mmu(vcpu);
2214         free_mmu_pages(vcpu);
2215         mmu_free_memory_caches(vcpu);
2216 }
2217
2218 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2219 {
2220         struct kvm_mmu_page *sp;
2221
2222         spin_lock(&kvm->mmu_lock);
2223         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2224                 int i;
2225                 u64 *pt;
2226
2227                 if (!test_bit(slot, &sp->slot_bitmap))
2228                         continue;
2229
2230                 pt = sp->spt;
2231                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2232                         /* avoid RMW */
2233                         if (pt[i] & PT_WRITABLE_MASK)
2234                                 pt[i] &= ~PT_WRITABLE_MASK;
2235         }
2236         kvm_flush_remote_tlbs(kvm);
2237         spin_unlock(&kvm->mmu_lock);
2238 }
2239
2240 void kvm_mmu_zap_all(struct kvm *kvm)
2241 {
2242         struct kvm_mmu_page *sp, *node;
2243
2244         spin_lock(&kvm->mmu_lock);
2245         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2246                 if (kvm_mmu_zap_page(kvm, sp))
2247                         node = container_of(kvm->arch.active_mmu_pages.next,
2248                                             struct kvm_mmu_page, link);
2249         spin_unlock(&kvm->mmu_lock);
2250
2251         kvm_flush_remote_tlbs(kvm);
2252 }
2253
2254 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2255 {
2256         struct kvm_mmu_page *page;
2257
2258         page = container_of(kvm->arch.active_mmu_pages.prev,
2259                             struct kvm_mmu_page, link);
2260         kvm_mmu_zap_page(kvm, page);
2261 }
2262
2263 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2264 {
2265         struct kvm *kvm;
2266         struct kvm *kvm_freed = NULL;
2267         int cache_count = 0;
2268
2269         spin_lock(&kvm_lock);
2270
2271         list_for_each_entry(kvm, &vm_list, vm_list) {
2272                 int npages;
2273
2274                 if (!down_read_trylock(&kvm->slots_lock))
2275                         continue;
2276                 spin_lock(&kvm->mmu_lock);
2277                 npages = kvm->arch.n_alloc_mmu_pages -
2278                          kvm->arch.n_free_mmu_pages;
2279                 cache_count += npages;
2280                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2281                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2282                         cache_count--;
2283                         kvm_freed = kvm;
2284                 }
2285                 nr_to_scan--;
2286
2287                 spin_unlock(&kvm->mmu_lock);
2288                 up_read(&kvm->slots_lock);
2289         }
2290         if (kvm_freed)
2291                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2292
2293         spin_unlock(&kvm_lock);
2294
2295         return cache_count;
2296 }
2297
2298 static struct shrinker mmu_shrinker = {
2299         .shrink = mmu_shrink,
2300         .seeks = DEFAULT_SEEKS * 10,
2301 };
2302
2303 static void mmu_destroy_caches(void)
2304 {
2305         if (pte_chain_cache)
2306                 kmem_cache_destroy(pte_chain_cache);
2307         if (rmap_desc_cache)
2308                 kmem_cache_destroy(rmap_desc_cache);
2309         if (mmu_page_header_cache)
2310                 kmem_cache_destroy(mmu_page_header_cache);
2311 }
2312
2313 void kvm_mmu_module_exit(void)
2314 {
2315         mmu_destroy_caches();
2316         unregister_shrinker(&mmu_shrinker);
2317 }
2318
2319 int kvm_mmu_module_init(void)
2320 {
2321         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2322                                             sizeof(struct kvm_pte_chain),
2323                                             0, 0, NULL);
2324         if (!pte_chain_cache)
2325                 goto nomem;
2326         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2327                                             sizeof(struct kvm_rmap_desc),
2328                                             0, 0, NULL);
2329         if (!rmap_desc_cache)
2330                 goto nomem;
2331
2332         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2333                                                   sizeof(struct kvm_mmu_page),
2334                                                   0, 0, NULL);
2335         if (!mmu_page_header_cache)
2336                 goto nomem;
2337
2338         register_shrinker(&mmu_shrinker);
2339
2340         return 0;
2341
2342 nomem:
2343         mmu_destroy_caches();
2344         return -ENOMEM;
2345 }
2346
2347 /*
2348  * Caculate mmu pages needed for kvm.
2349  */
2350 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2351 {
2352         int i;
2353         unsigned int nr_mmu_pages;
2354         unsigned int  nr_pages = 0;
2355
2356         for (i = 0; i < kvm->nmemslots; i++)
2357                 nr_pages += kvm->memslots[i].npages;
2358
2359         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2360         nr_mmu_pages = max(nr_mmu_pages,
2361                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2362
2363         return nr_mmu_pages;
2364 }
2365
2366 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2367                                 unsigned len)
2368 {
2369         if (len > buffer->len)
2370                 return NULL;
2371         return buffer->ptr;
2372 }
2373
2374 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2375                                 unsigned len)
2376 {
2377         void *ret;
2378
2379         ret = pv_mmu_peek_buffer(buffer, len);
2380         if (!ret)
2381                 return ret;
2382         buffer->ptr += len;
2383         buffer->len -= len;
2384         buffer->processed += len;
2385         return ret;
2386 }
2387
2388 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2389                              gpa_t addr, gpa_t value)
2390 {
2391         int bytes = 8;
2392         int r;
2393
2394         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2395                 bytes = 4;
2396
2397         r = mmu_topup_memory_caches(vcpu);
2398         if (r)
2399                 return r;
2400
2401         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2402                 return -EFAULT;
2403
2404         return 1;
2405 }
2406
2407 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2408 {
2409         kvm_x86_ops->tlb_flush(vcpu);
2410         return 1;
2411 }
2412
2413 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2414 {
2415         spin_lock(&vcpu->kvm->mmu_lock);
2416         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2417         spin_unlock(&vcpu->kvm->mmu_lock);
2418         return 1;
2419 }
2420
2421 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2422                              struct kvm_pv_mmu_op_buffer *buffer)
2423 {
2424         struct kvm_mmu_op_header *header;
2425
2426         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2427         if (!header)
2428                 return 0;
2429         switch (header->op) {
2430         case KVM_MMU_OP_WRITE_PTE: {
2431                 struct kvm_mmu_op_write_pte *wpte;
2432
2433                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2434                 if (!wpte)
2435                         return 0;
2436                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2437                                         wpte->pte_val);
2438         }
2439         case KVM_MMU_OP_FLUSH_TLB: {
2440                 struct kvm_mmu_op_flush_tlb *ftlb;
2441
2442                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2443                 if (!ftlb)
2444                         return 0;
2445                 return kvm_pv_mmu_flush_tlb(vcpu);
2446         }
2447         case KVM_MMU_OP_RELEASE_PT: {
2448                 struct kvm_mmu_op_release_pt *rpt;
2449
2450                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2451                 if (!rpt)
2452                         return 0;
2453                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2454         }
2455         default: return 0;
2456         }
2457 }
2458
2459 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2460                   gpa_t addr, unsigned long *ret)
2461 {
2462         int r;
2463         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2464
2465         buffer->ptr = buffer->buf;
2466         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2467         buffer->processed = 0;
2468
2469         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2470         if (r)
2471                 goto out;
2472
2473         while (buffer->len) {
2474                 r = kvm_pv_mmu_op_one(vcpu, buffer);
2475                 if (r < 0)
2476                         goto out;
2477                 if (r == 0)
2478                         break;
2479         }
2480
2481         r = 1;
2482 out:
2483         *ret = buffer->processed;
2484         return r;
2485 }
2486
2487 #ifdef AUDIT
2488
2489 static const char *audit_msg;
2490
2491 static gva_t canonicalize(gva_t gva)
2492 {
2493 #ifdef CONFIG_X86_64
2494         gva = (long long)(gva << 16) >> 16;
2495 #endif
2496         return gva;
2497 }
2498
2499 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2500                                 gva_t va, int level)
2501 {
2502         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2503         int i;
2504         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2505
2506         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2507                 u64 ent = pt[i];
2508
2509                 if (ent == shadow_trap_nonpresent_pte)
2510                         continue;
2511
2512                 va = canonicalize(va);
2513                 if (level > 1) {
2514                         if (ent == shadow_notrap_nonpresent_pte)
2515                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2516                                        " in nonleaf level: levels %d gva %lx"
2517                                        " level %d pte %llx\n", audit_msg,
2518                                        vcpu->arch.mmu.root_level, va, level, ent);
2519
2520                         audit_mappings_page(vcpu, ent, va, level - 1);
2521                 } else {
2522                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2523                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2524
2525                         if (is_shadow_present_pte(ent)
2526                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2527                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2528                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2529                                        audit_msg, vcpu->arch.mmu.root_level,
2530                                        va, gpa, hpa, ent,
2531                                        is_shadow_present_pte(ent));
2532                         else if (ent == shadow_notrap_nonpresent_pte
2533                                  && !is_error_hpa(hpa))
2534                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2535                                        " valid guest gva %lx\n", audit_msg, va);
2536                         kvm_release_pfn_clean(pfn);
2537
2538                 }
2539         }
2540 }
2541
2542 static void audit_mappings(struct kvm_vcpu *vcpu)
2543 {
2544         unsigned i;
2545
2546         if (vcpu->arch.mmu.root_level == 4)
2547                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2548         else
2549                 for (i = 0; i < 4; ++i)
2550                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2551                                 audit_mappings_page(vcpu,
2552                                                     vcpu->arch.mmu.pae_root[i],
2553                                                     i << 30,
2554                                                     2);
2555 }
2556
2557 static int count_rmaps(struct kvm_vcpu *vcpu)
2558 {
2559         int nmaps = 0;
2560         int i, j, k;
2561
2562         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2563                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2564                 struct kvm_rmap_desc *d;
2565
2566                 for (j = 0; j < m->npages; ++j) {
2567                         unsigned long *rmapp = &m->rmap[j];
2568
2569                         if (!*rmapp)
2570                                 continue;
2571                         if (!(*rmapp & 1)) {
2572                                 ++nmaps;
2573                                 continue;
2574                         }
2575                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2576                         while (d) {
2577                                 for (k = 0; k < RMAP_EXT; ++k)
2578                                         if (d->shadow_ptes[k])
2579                                                 ++nmaps;
2580                                         else
2581                                                 break;
2582                                 d = d->more;
2583                         }
2584                 }
2585         }
2586         return nmaps;
2587 }
2588
2589 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2590 {
2591         int nmaps = 0;
2592         struct kvm_mmu_page *sp;
2593         int i;
2594
2595         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2596                 u64 *pt = sp->spt;
2597
2598                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2599                         continue;
2600
2601                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2602                         u64 ent = pt[i];
2603
2604                         if (!(ent & PT_PRESENT_MASK))
2605                                 continue;
2606                         if (!(ent & PT_WRITABLE_MASK))
2607                                 continue;
2608                         ++nmaps;
2609                 }
2610         }
2611         return nmaps;
2612 }
2613
2614 static void audit_rmap(struct kvm_vcpu *vcpu)
2615 {
2616         int n_rmap = count_rmaps(vcpu);
2617         int n_actual = count_writable_mappings(vcpu);
2618
2619         if (n_rmap != n_actual)
2620                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2621                        __func__, audit_msg, n_rmap, n_actual);
2622 }
2623
2624 static void audit_write_protection(struct kvm_vcpu *vcpu)
2625 {
2626         struct kvm_mmu_page *sp;
2627         struct kvm_memory_slot *slot;
2628         unsigned long *rmapp;
2629         gfn_t gfn;
2630
2631         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2632                 if (sp->role.metaphysical)
2633                         continue;
2634
2635                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2636                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2637                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2638                 if (*rmapp)
2639                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2640                                " mappings: gfn %lx role %x\n",
2641                                __func__, audit_msg, sp->gfn,
2642                                sp->role.word);
2643         }
2644 }
2645
2646 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2647 {
2648         int olddbg = dbg;
2649
2650         dbg = 0;
2651         audit_msg = msg;
2652         audit_rmap(vcpu);
2653         audit_write_protection(vcpu);
2654         audit_mappings(vcpu);
2655         dbg = olddbg;
2656 }
2657
2658 #endif