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 = 1;
  70 #endif
  71
  72 #ifndef MMU_DEBUG
  73 #define ASSERT(x) do { } while (0)
  74 #else
  75 #define ASSERT(x)                                                       \
  76         if (!(x)) {                                                     \
  77                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  78                        __FILE__, __LINE__, #x);                         \
  79         }
  80 #endif
  81
  82 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  83 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  84
  85 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  86
  87 #define PT64_LEVEL_BITS 9
  88
  89 #define PT64_LEVEL_SHIFT(level) \
  90                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  91
  92 #define PT64_LEVEL_MASK(level) \
  93                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  94
  95 #define PT64_INDEX(address, level)\
  96         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
  97
  98
  99 #define PT32_LEVEL_BITS 10
 100
 101 #define PT32_LEVEL_SHIFT(level) \
 102                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 103
 104 #define PT32_LEVEL_MASK(level) \
 105                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 106
 107 #define PT32_INDEX(address, level)\
 108         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 109
 110
 111 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 112 #define PT64_DIR_BASE_ADDR_MASK \
 113         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 114
 115 #define PT32_BASE_ADDR_MASK PAGE_MASK
 116 #define PT32_DIR_BASE_ADDR_MASK \
 117         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 118
 119 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 120                         | PT64_NX_MASK)
 121
 122 #define PFERR_PRESENT_MASK (1U << 0)
 123 #define PFERR_WRITE_MASK (1U << 1)
 124 #define PFERR_USER_MASK (1U << 2)
 125 #define PFERR_FETCH_MASK (1U << 4)
 126
 127 #define PT_DIRECTORY_LEVEL 2
 128 #define PT_PAGE_TABLE_LEVEL 1
 129
 130 #define RMAP_EXT 4
 131
 132 #define ACC_EXEC_MASK    1
 133 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 134 #define ACC_USER_MASK    PT_USER_MASK
 135 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 136
 137 struct kvm_pv_mmu_op_buffer {
 138         void *ptr;
 139         unsigned len;
 140         unsigned processed;
 141         char buf[512] __aligned(sizeof(long));
 142 };
 143
 144 struct kvm_rmap_desc {
 145         u64 *shadow_ptes[RMAP_EXT];
 146         struct kvm_rmap_desc *more;
 147 };
 148
 149 static struct kmem_cache *pte_chain_cache;
 150 static struct kmem_cache *rmap_desc_cache;
 151 static struct kmem_cache *mmu_page_header_cache;
 152
 153 static u64 __read_mostly shadow_trap_nonpresent_pte;
 154 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 155
 156 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 157 {
 158         shadow_trap_nonpresent_pte = trap_pte;
 159         shadow_notrap_nonpresent_pte = notrap_pte;
 160 }
 161 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 162
 163 static int is_write_protection(struct kvm_vcpu *vcpu)
 164 {
 165         return vcpu->arch.cr0 & X86_CR0_WP;
 166 }
 167
 168 static int is_cpuid_PSE36(void)
 169 {
 170         return 1;
 171 }
 172
 173 static int is_nx(struct kvm_vcpu *vcpu)
 174 {
 175         return vcpu->arch.shadow_efer & EFER_NX;
 176 }
 177
 178 static int is_present_pte(unsigned long pte)
 179 {
 180         return pte & PT_PRESENT_MASK;
 181 }
 182
 183 static int is_shadow_present_pte(u64 pte)
 184 {
 185         return pte != shadow_trap_nonpresent_pte
 186                 && pte != shadow_notrap_nonpresent_pte;
 187 }
 188
 189 static int is_large_pte(u64 pte)
 190 {
 191         return pte & PT_PAGE_SIZE_MASK;
 192 }
 193
 194 static int is_writeble_pte(unsigned long pte)
 195 {
 196         return pte & PT_WRITABLE_MASK;
 197 }
 198
 199 static int is_dirty_pte(unsigned long pte)
 200 {
 201         return pte & PT_DIRTY_MASK;
 202 }
 203
 204 static int is_rmap_pte(u64 pte)
 205 {
 206         return is_shadow_present_pte(pte);
 207 }
 208
 209 static pfn_t spte_to_pfn(u64 pte)
 210 {
 211         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 212 }
 213
 214 static gfn_t pse36_gfn_delta(u32 gpte)
 215 {
 216         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 217
 218         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 219 }
 220
 221 static void set_shadow_pte(u64 *sptep, u64 spte)
 222 {
 223 #ifdef CONFIG_X86_64
 224         set_64bit((unsigned long *)sptep, spte);
 225 #else
 226         set_64bit((unsigned long long *)sptep, spte);
 227 #endif
 228 }
 229
 230 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 231                                   struct kmem_cache *base_cache, int min)
 232 {
 233         void *obj;
 234
 235         if (cache->nobjs >= min)
 236                 return 0;
 237         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 238                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 239                 if (!obj)
 240                         return -ENOMEM;
 241                 cache->objects[cache->nobjs++] = obj;
 242         }
 243         return 0;
 244 }
 245
 246 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 247 {
 248         while (mc->nobjs)
 249                 kfree(mc->objects[--mc->nobjs]);
 250 }
 251
 252 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 253                                        int min)
 254 {
 255         struct page *page;
 256
 257         if (cache->nobjs >= min)
 258                 return 0;
 259         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 260                 page = alloc_page(GFP_KERNEL);
 261                 if (!page)
 262                         return -ENOMEM;
 263                 set_page_private(page, 0);
 264                 cache->objects[cache->nobjs++] = page_address(page);
 265         }
 266         return 0;
 267 }
 268
 269 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 270 {
 271         while (mc->nobjs)
 272                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 273 }
 274
 275 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 276 {
 277         int r;
 278
 279         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 280                                    pte_chain_cache, 4);
 281         if (r)
 282                 goto out;
 283         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 284                                    rmap_desc_cache, 1);
 285         if (r)
 286                 goto out;
 287         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 288         if (r)
 289                 goto out;
 290         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 291                                    mmu_page_header_cache, 4);
 292 out:
 293         return r;
 294 }
 295
 296 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 297 {
 298         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 299         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 300         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 301         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 302 }
 303
 304 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 305                                     size_t size)
 306 {
 307         void *p;
 308
 309         BUG_ON(!mc->nobjs);
 310         p = mc->objects[--mc->nobjs];
 311         memset(p, 0, size);
 312         return p;
 313 }
 314
 315 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 316 {
 317         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 318                                       sizeof(struct kvm_pte_chain));
 319 }
 320
 321 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 322 {
 323         kfree(pc);
 324 }
 325
 326 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 327 {
 328         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 329                                       sizeof(struct kvm_rmap_desc));
 330 }
 331
 332 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 333 {
 334         kfree(rd);
 335 }
 336
 337 /*
 338  * Return the pointer to the largepage write count for a given
 339  * gfn, handling slots that are not large page aligned.
 340  */
 341 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 342 {
 343         unsigned long idx;
 344
 345         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 346               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 347         return &slot->lpage_info[idx].write_count;
 348 }
 349
 350 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 351 {
 352         int *write_count;
 353
 354         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 355         *write_count += 1;
 356         WARN_ON(*write_count > KVM_PAGES_PER_HPAGE);
 357 }
 358
 359 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 360 {
 361         int *write_count;
 362
 363         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 364         *write_count -= 1;
 365         WARN_ON(*write_count < 0);
 366 }
 367
 368 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 369 {
 370         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
 371         int *largepage_idx;
 372
 373         if (slot) {
 374                 largepage_idx = slot_largepage_idx(gfn, slot);
 375                 return *largepage_idx;
 376         }
 377
 378         return 1;
 379 }
 380
 381 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 382 {
 383         struct vm_area_struct *vma;
 384         unsigned long addr;
 385
 386         addr = gfn_to_hva(kvm, gfn);
 387         if (kvm_is_error_hva(addr))
 388                 return 0;
 389
 390         vma = find_vma(current->mm, addr);
 391         if (vma && is_vm_hugetlb_page(vma))
 392                 return 1;
 393
 394         return 0;
 395 }
 396
 397 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 398 {
 399         struct kvm_memory_slot *slot;
 400
 401         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 402                 return 0;
 403
 404         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 405                 return 0;
 406
 407         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 408         if (slot && slot->dirty_bitmap)
 409                 return 0;
 410
 411         return 1;
 412 }
 413
 414 /*
 415  * Take gfn and return the reverse mapping to it.
 416  * Note: gfn must be unaliased before this function get called
 417  */
 418
 419 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 420 {
 421         struct kvm_memory_slot *slot;
 422         unsigned long idx;
 423
 424         slot = gfn_to_memslot(kvm, gfn);
 425         if (!lpage)
 426                 return &slot->rmap[gfn - slot->base_gfn];
 427
 428         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 429               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 430
 431         return &slot->lpage_info[idx].rmap_pde;
 432 }
 433
 434 /*
 435  * Reverse mapping data structures:
 436  *
 437  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 438  * that points to page_address(page).
 439  *
 440  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 441  * containing more mappings.
 442  */
 443 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 444 {
 445         struct kvm_mmu_page *sp;
 446         struct kvm_rmap_desc *desc;
 447         unsigned long *rmapp;
 448         int i;
 449
 450         if (!is_rmap_pte(*spte))
 451                 return;
 452         gfn = unalias_gfn(vcpu->kvm, gfn);
 453         sp = page_header(__pa(spte));
 454         sp->gfns[spte - sp->spt] = gfn;
 455         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 456         if (!*rmapp) {
 457                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 458                 *rmapp = (unsigned long)spte;
 459         } else if (!(*rmapp & 1)) {
 460                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 461                 desc = mmu_alloc_rmap_desc(vcpu);
 462                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 463                 desc->shadow_ptes[1] = spte;
 464                 *rmapp = (unsigned long)desc | 1;
 465         } else {
 466                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 467                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 468                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 469                         desc = desc->more;
 470                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 471                         desc->more = mmu_alloc_rmap_desc(vcpu);
 472                         desc = desc->more;
 473                 }
 474                 for (i = 0; desc->shadow_ptes[i]; ++i)
 475                         ;
 476                 desc->shadow_ptes[i] = spte;
 477         }
 478 }
 479
 480 static void rmap_desc_remove_entry(unsigned long *rmapp,
 481                                    struct kvm_rmap_desc *desc,
 482                                    int i,
 483                                    struct kvm_rmap_desc *prev_desc)
 484 {
 485         int j;
 486
 487         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 488                 ;
 489         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 490         desc->shadow_ptes[j] = NULL;
 491         if (j != 0)
 492                 return;
 493         if (!prev_desc && !desc->more)
 494                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 495         else
 496                 if (prev_desc)
 497                         prev_desc->more = desc->more;
 498                 else
 499                         *rmapp = (unsigned long)desc->more | 1;
 500         mmu_free_rmap_desc(desc);
 501 }
 502
 503 static void rmap_remove(struct kvm *kvm, u64 *spte)
 504 {
 505         struct kvm_rmap_desc *desc;
 506         struct kvm_rmap_desc *prev_desc;
 507         struct kvm_mmu_page *sp;
 508         pfn_t pfn;
 509         unsigned long *rmapp;
 510         int i;
 511
 512         if (!is_rmap_pte(*spte))
 513                 return;
 514         sp = page_header(__pa(spte));
 515         pfn = spte_to_pfn(*spte);
 516         if (*spte & PT_ACCESSED_MASK)
 517                 kvm_set_pfn_accessed(pfn);
 518         if (is_writeble_pte(*spte))
 519                 kvm_release_pfn_dirty(pfn);
 520         else
 521                 kvm_release_pfn_clean(pfn);
 522         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 523         if (!*rmapp) {
 524                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 525                 BUG();
 526         } else if (!(*rmapp & 1)) {
 527                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 528                 if ((u64 *)*rmapp != spte) {
 529                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 530                                spte, *spte);
 531                         BUG();
 532                 }
 533                 *rmapp = 0;
 534         } else {
 535                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 536                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 537                 prev_desc = NULL;
 538                 while (desc) {
 539                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 540                                 if (desc->shadow_ptes[i] == spte) {
 541                                         rmap_desc_remove_entry(rmapp,
 542                                                                desc, i,
 543                                                                prev_desc);
 544                                         return;
 545                                 }
 546                         prev_desc = desc;
 547                         desc = desc->more;
 548                 }
 549                 BUG();
 550         }
 551 }
 552
 553 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 554 {
 555         struct kvm_rmap_desc *desc;
 556         struct kvm_rmap_desc *prev_desc;
 557         u64 *prev_spte;
 558         int i;
 559
 560         if (!*rmapp)
 561                 return NULL;
 562         else if (!(*rmapp & 1)) {
 563                 if (!spte)
 564                         return (u64 *)*rmapp;
 565                 return NULL;
 566         }
 567         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 568         prev_desc = NULL;
 569         prev_spte = NULL;
 570         while (desc) {
 571                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 572                         if (prev_spte == spte)
 573                                 return desc->shadow_ptes[i];
 574                         prev_spte = desc->shadow_ptes[i];
 575                 }
 576                 desc = desc->more;
 577         }
 578         return NULL;
 579 }
 580
 581 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 582 {
 583         unsigned long *rmapp;
 584         u64 *spte;
 585         int write_protected = 0;
 586
 587         gfn = unalias_gfn(kvm, gfn);
 588         rmapp = gfn_to_rmap(kvm, gfn, 0);
 589
 590         spte = rmap_next(kvm, rmapp, NULL);
 591         while (spte) {
 592                 BUG_ON(!spte);
 593                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 594                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 595                 if (is_writeble_pte(*spte)) {
 596                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 597                         write_protected = 1;
 598                 }
 599                 spte = rmap_next(kvm, rmapp, spte);
 600         }
 601         if (write_protected) {
 602                 pfn_t pfn;
 603
 604                 spte = rmap_next(kvm, rmapp, NULL);
 605                 pfn = spte_to_pfn(*spte);
 606                 kvm_set_pfn_dirty(pfn);
 607         }
 608
 609         /* check for huge page mappings */
 610         rmapp = gfn_to_rmap(kvm, gfn, 1);
 611         spte = rmap_next(kvm, rmapp, NULL);
 612         while (spte) {
 613                 BUG_ON(!spte);
 614                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 615                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 616                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 617                 if (is_writeble_pte(*spte)) {
 618                         rmap_remove(kvm, spte);
 619                         --kvm->stat.lpages;
 620                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 621                         write_protected = 1;
 622                 }
 623                 spte = rmap_next(kvm, rmapp, spte);
 624         }
 625
 626         if (write_protected)
 627                 kvm_flush_remote_tlbs(kvm);
 628
 629         account_shadowed(kvm, gfn);
 630 }
 631
 632 #ifdef MMU_DEBUG
 633 static int is_empty_shadow_page(u64 *spt)
 634 {
 635         u64 *pos;
 636         u64 *end;
 637
 638         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 639                 if (*pos != shadow_trap_nonpresent_pte) {
 640                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 641                                pos, *pos);
 642                         return 0;
 643                 }
 644         return 1;
 645 }
 646 #endif
 647
 648 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 649 {
 650         ASSERT(is_empty_shadow_page(sp->spt));
 651         list_del(&sp->link);
 652         __free_page(virt_to_page(sp->spt));
 653         __free_page(virt_to_page(sp->gfns));
 654         kfree(sp);
 655         ++kvm->arch.n_free_mmu_pages;
 656 }
 657
 658 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 659 {
 660         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 661 }
 662
 663 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 664                                                u64 *parent_pte)
 665 {
 666         struct kvm_mmu_page *sp;
 667
 668         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 669         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 670         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 671         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 672         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 673         ASSERT(is_empty_shadow_page(sp->spt));
 674         sp->slot_bitmap = 0;
 675         sp->multimapped = 0;
 676         sp->parent_pte = parent_pte;
 677         --vcpu->kvm->arch.n_free_mmu_pages;
 678         return sp;
 679 }
 680
 681 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 682                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 683 {
 684         struct kvm_pte_chain *pte_chain;
 685         struct hlist_node *node;
 686         int i;
 687
 688         if (!parent_pte)
 689                 return;
 690         if (!sp->multimapped) {
 691                 u64 *old = sp->parent_pte;
 692
 693                 if (!old) {
 694                         sp->parent_pte = parent_pte;
 695                         return;
 696                 }
 697                 sp->multimapped = 1;
 698                 pte_chain = mmu_alloc_pte_chain(vcpu);
 699                 INIT_HLIST_HEAD(&sp->parent_ptes);
 700                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 701                 pte_chain->parent_ptes[0] = old;
 702         }
 703         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 704                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 705                         continue;
 706                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 707                         if (!pte_chain->parent_ptes[i]) {
 708                                 pte_chain->parent_ptes[i] = parent_pte;
 709                                 return;
 710                         }
 711         }
 712         pte_chain = mmu_alloc_pte_chain(vcpu);
 713         BUG_ON(!pte_chain);
 714         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 715         pte_chain->parent_ptes[0] = parent_pte;
 716 }
 717
 718 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 719                                        u64 *parent_pte)
 720 {
 721         struct kvm_pte_chain *pte_chain;
 722         struct hlist_node *node;
 723         int i;
 724
 725         if (!sp->multimapped) {
 726                 BUG_ON(sp->parent_pte != parent_pte);
 727                 sp->parent_pte = NULL;
 728                 return;
 729         }
 730         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 731                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 732                         if (!pte_chain->parent_ptes[i])
 733                                 break;
 734                         if (pte_chain->parent_ptes[i] != parent_pte)
 735                                 continue;
 736                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 737                                 && pte_chain->parent_ptes[i + 1]) {
 738                                 pte_chain->parent_ptes[i]
 739                                         = pte_chain->parent_ptes[i + 1];
 740                                 ++i;
 741                         }
 742                         pte_chain->parent_ptes[i] = NULL;
 743                         if (i == 0) {
 744                                 hlist_del(&pte_chain->link);
 745                                 mmu_free_pte_chain(pte_chain);
 746                                 if (hlist_empty(&sp->parent_ptes)) {
 747                                         sp->multimapped = 0;
 748                                         sp->parent_pte = NULL;
 749                                 }
 750                         }
 751                         return;
 752                 }
 753         BUG();
 754 }
 755
 756 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
 757 {
 758         unsigned index;
 759         struct hlist_head *bucket;
 760         struct kvm_mmu_page *sp;
 761         struct hlist_node *node;
 762
 763         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 764         index = kvm_page_table_hashfn(gfn);
 765         bucket = &kvm->arch.mmu_page_hash[index];
 766         hlist_for_each_entry(sp, node, bucket, hash_link)
 767                 if (sp->gfn == gfn && !sp->role.metaphysical
 768                     && !sp->role.invalid) {
 769                         pgprintk("%s: found role %x\n",
 770                                  __func__, sp->role.word);
 771                         return sp;
 772                 }
 773         return NULL;
 774 }
 775
 776 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 777                                              gfn_t gfn,
 778                                              gva_t gaddr,
 779                                              unsigned level,
 780                                              int metaphysical,
 781                                              unsigned access,
 782                                              u64 *parent_pte)
 783 {
 784         union kvm_mmu_page_role role;
 785         unsigned index;
 786         unsigned quadrant;
 787         struct hlist_head *bucket;
 788         struct kvm_mmu_page *sp;
 789         struct hlist_node *node;
 790
 791         role.word = 0;
 792         role.glevels = vcpu->arch.mmu.root_level;
 793         role.level = level;
 794         role.metaphysical = metaphysical;
 795         role.access = access;
 796         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
 797                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 798                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 799                 role.quadrant = quadrant;
 800         }
 801         pgprintk("%s: looking gfn %lx role %x\n", __func__,
 802                  gfn, role.word);
 803         index = kvm_page_table_hashfn(gfn);
 804         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
 805         hlist_for_each_entry(sp, node, bucket, hash_link)
 806                 if (sp->gfn == gfn && sp->role.word == role.word) {
 807                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
 808                         pgprintk("%s: found\n", __func__);
 809                         return sp;
 810                 }
 811         ++vcpu->kvm->stat.mmu_cache_miss;
 812         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
 813         if (!sp)
 814                 return sp;
 815         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
 816         sp->gfn = gfn;
 817         sp->role = role;
 818         hlist_add_head(&sp->hash_link, bucket);
 819         if (!metaphysical)
 820                 rmap_write_protect(vcpu->kvm, gfn);
 821         vcpu->arch.mmu.prefetch_page(vcpu, sp);
 822         return sp;
 823 }
 824
 825 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 826                                          struct kvm_mmu_page *sp)
 827 {
 828         unsigned i;
 829         u64 *pt;
 830         u64 ent;
 831
 832         pt = sp->spt;
 833
 834         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
 835                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 836                         if (is_shadow_present_pte(pt[i]))
 837                                 rmap_remove(kvm, &pt[i]);
 838                         pt[i] = shadow_trap_nonpresent_pte;
 839                 }
 840                 kvm_flush_remote_tlbs(kvm);
 841                 return;
 842         }
 843
 844         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 845                 ent = pt[i];
 846
 847                 if (is_shadow_present_pte(ent)) {
 848                         if (!is_large_pte(ent)) {
 849                                 ent &= PT64_BASE_ADDR_MASK;
 850                                 mmu_page_remove_parent_pte(page_header(ent),
 851                                                            &pt[i]);
 852                         } else {
 853                                 --kvm->stat.lpages;
 854                                 rmap_remove(kvm, &pt[i]);
 855                         }
 856                 }
 857                 pt[i] = shadow_trap_nonpresent_pte;
 858         }
 859         kvm_flush_remote_tlbs(kvm);
 860 }
 861
 862 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
 863 {
 864         mmu_page_remove_parent_pte(sp, parent_pte);
 865 }
 866
 867 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 868 {
 869         int i;
 870
 871         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 872                 if (kvm->vcpus[i])
 873                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
 874 }
 875
 876 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 877 {
 878         u64 *parent_pte;
 879
 880         ++kvm->stat.mmu_shadow_zapped;
 881         while (sp->multimapped || sp->parent_pte) {
 882                 if (!sp->multimapped)
 883                         parent_pte = sp->parent_pte;
 884                 else {
 885                         struct kvm_pte_chain *chain;
 886
 887                         chain = container_of(sp->parent_ptes.first,
 888                                              struct kvm_pte_chain, link);
 889                         parent_pte = chain->parent_ptes[0];
 890                 }
 891                 BUG_ON(!parent_pte);
 892                 kvm_mmu_put_page(sp, parent_pte);
 893                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 894         }
 895         kvm_mmu_page_unlink_children(kvm, sp);
 896         if (!sp->root_count) {
 897                 if (!sp->role.metaphysical)
 898                         unaccount_shadowed(kvm, sp->gfn);
 899                 hlist_del(&sp->hash_link);
 900                 kvm_mmu_free_page(kvm, sp);
 901         } else {
 902                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
 903                 sp->role.invalid = 1;
 904                 kvm_reload_remote_mmus(kvm);
 905         }
 906         kvm_mmu_reset_last_pte_updated(kvm);
 907 }
 908
 909 /*
 910  * Changing the number of mmu pages allocated to the vm
 911  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 912  */
 913 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 914 {
 915         /*
 916          * If we set the number of mmu pages to be smaller be than the
 917          * number of actived pages , we must to free some mmu pages before we
 918          * change the value
 919          */
 920
 921         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
 922             kvm_nr_mmu_pages) {
 923                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
 924                                        - kvm->arch.n_free_mmu_pages;
 925
 926                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 927                         struct kvm_mmu_page *page;
 928
 929                         page = container_of(kvm->arch.active_mmu_pages.prev,
 930                                             struct kvm_mmu_page, link);
 931                         kvm_mmu_zap_page(kvm, page);
 932                         n_used_mmu_pages--;
 933                 }
 934                 kvm->arch.n_free_mmu_pages = 0;
 935         }
 936         else
 937                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
 938                                          - kvm->arch.n_alloc_mmu_pages;
 939
 940         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
 941 }
 942
 943 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 944 {
 945         unsigned index;
 946         struct hlist_head *bucket;
 947         struct kvm_mmu_page *sp;
 948         struct hlist_node *node, *n;
 949         int r;
 950
 951         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 952         r = 0;
 953         index = kvm_page_table_hashfn(gfn);
 954         bucket = &kvm->arch.mmu_page_hash[index];
 955         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
 956                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 957                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
 958                                  sp->role.word);
 959                         kvm_mmu_zap_page(kvm, sp);
 960                         r = 1;
 961                 }
 962         return r;
 963 }
 964
 965 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
 966 {
 967         struct kvm_mmu_page *sp;
 968
 969         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
 970                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
 971                 kvm_mmu_zap_page(kvm, sp);
 972         }
 973 }
 974
 975 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
 976 {
 977         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
 978         struct kvm_mmu_page *sp = page_header(__pa(pte));
 979
 980         __set_bit(slot, &sp->slot_bitmap);
 981 }
 982
 983 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 984 {
 985         struct page *page;
 986
 987         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
 988
 989         if (gpa == UNMAPPED_GVA)
 990                 return NULL;
 991
 992         down_read(&current->mm->mmap_sem);
 993         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 994         up_read(&current->mm->mmap_sem);
 995
 996         return page;
 997 }
 998
 999 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1000                          unsigned pt_access, unsigned pte_access,
1001                          int user_fault, int write_fault, int dirty,
1002                          int *ptwrite, int largepage, gfn_t gfn,
1003                          pfn_t pfn, bool speculative)
1004 {
1005         u64 spte;
1006         int was_rmapped = 0;
1007         int was_writeble = is_writeble_pte(*shadow_pte);
1008
1009         pgprintk("%s: spte %llx access %x write_fault %d"
1010                  " user_fault %d gfn %lx\n",
1011                  __func__, *shadow_pte, pt_access,
1012                  write_fault, user_fault, gfn);
1013
1014         if (is_rmap_pte(*shadow_pte)) {
1015                 /*
1016                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1017                  * the parent of the now unreachable PTE.
1018                  */
1019                 if (largepage && !is_large_pte(*shadow_pte)) {
1020                         struct kvm_mmu_page *child;
1021                         u64 pte = *shadow_pte;
1022
1023                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1024                         mmu_page_remove_parent_pte(child, shadow_pte);
1025                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1026                         pgprintk("hfn old %lx new %lx\n",
1027                                  spte_to_pfn(*shadow_pte), pfn);
1028                         rmap_remove(vcpu->kvm, shadow_pte);
1029                 } else {
1030                         if (largepage)
1031                                 was_rmapped = is_large_pte(*shadow_pte);
1032                         else
1033                                 was_rmapped = 1;
1034                 }
1035         }
1036
1037         /*
1038          * We don't set the accessed bit, since we sometimes want to see
1039          * whether the guest actually used the pte (in order to detect
1040          * demand paging).
1041          */
1042         spte = PT_PRESENT_MASK | PT_DIRTY_MASK;
1043         if (!speculative)
1044                 pte_access |= PT_ACCESSED_MASK;
1045         if (!dirty)
1046                 pte_access &= ~ACC_WRITE_MASK;
1047         if (!(pte_access & ACC_EXEC_MASK))
1048                 spte |= PT64_NX_MASK;
1049
1050         spte |= PT_PRESENT_MASK;
1051         if (pte_access & ACC_USER_MASK)
1052                 spte |= PT_USER_MASK;
1053         if (largepage)
1054                 spte |= PT_PAGE_SIZE_MASK;
1055
1056         spte |= (u64)pfn << PAGE_SHIFT;
1057
1058         if ((pte_access & ACC_WRITE_MASK)
1059             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1060                 struct kvm_mmu_page *shadow;
1061
1062                 spte |= PT_WRITABLE_MASK;
1063                 if (user_fault) {
1064                         mmu_unshadow(vcpu->kvm, gfn);
1065                         goto unshadowed;
1066                 }
1067
1068                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1069                 if (shadow ||
1070                    (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
1071                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1072                                  __func__, gfn);
1073                         pte_access &= ~ACC_WRITE_MASK;
1074                         if (is_writeble_pte(spte)) {
1075                                 spte &= ~PT_WRITABLE_MASK;
1076                                 kvm_x86_ops->tlb_flush(vcpu);
1077                         }
1078                         if (write_fault)
1079                                 *ptwrite = 1;
1080                 }
1081         }
1082
1083 unshadowed:
1084
1085         if (pte_access & ACC_WRITE_MASK)
1086                 mark_page_dirty(vcpu->kvm, gfn);
1087
1088         pgprintk("%s: setting spte %llx\n", __func__, spte);
1089         pgprintk("instantiating %s PTE (%s) at %d (%llx) addr %llx\n",
1090                  (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
1091                  (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
1092         set_shadow_pte(shadow_pte, spte);
1093         if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
1094             && (spte & PT_PRESENT_MASK))
1095                 ++vcpu->kvm->stat.lpages;
1096
1097         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1098         if (!was_rmapped) {
1099                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1100                 if (!is_rmap_pte(*shadow_pte))
1101                         kvm_release_pfn_clean(pfn);
1102         } else {
1103                 if (was_writeble)
1104                         kvm_release_pfn_dirty(pfn);
1105                 else
1106                         kvm_release_pfn_clean(pfn);
1107         }
1108         if (!ptwrite || !*ptwrite)
1109                 vcpu->arch.last_pte_updated = shadow_pte;
1110 }
1111
1112 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1113 {
1114 }
1115
1116 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1117                            int largepage, gfn_t gfn, pfn_t pfn,
1118                            int level)
1119 {
1120         hpa_t table_addr = vcpu->arch.mmu.root_hpa;
1121         int pt_write = 0;
1122
1123         for (; ; level--) {
1124                 u32 index = PT64_INDEX(v, level);
1125                 u64 *table;
1126
1127                 ASSERT(VALID_PAGE(table_addr));
1128                 table = __va(table_addr);
1129
1130                 if (level == 1) {
1131                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1132                                      0, write, 1, &pt_write, 0, gfn, pfn, false);
1133                         return pt_write;
1134                 }
1135
1136                 if (largepage && level == 2) {
1137                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1138                                      0, write, 1, &pt_write, 1, gfn, pfn, false);
1139                         return pt_write;
1140                 }
1141
1142                 if (table[index] == shadow_trap_nonpresent_pte) {
1143                         struct kvm_mmu_page *new_table;
1144                         gfn_t pseudo_gfn;
1145
1146                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1147                                 >> PAGE_SHIFT;
1148                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1149                                                      v, level - 1,
1150                                                      1, ACC_ALL, &table[index]);
1151                         if (!new_table) {
1152                                 pgprintk("nonpaging_map: ENOMEM\n");
1153                                 kvm_release_pfn_clean(pfn);
1154                                 return -ENOMEM;
1155                         }
1156
1157                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
1158                                 | PT_WRITABLE_MASK | PT_USER_MASK;
1159                 }
1160                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1161         }
1162 }
1163
1164 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1165 {
1166         int r;
1167         int largepage = 0;
1168         pfn_t pfn;
1169
1170         down_read(&current->mm->mmap_sem);
1171         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1172                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1173                 largepage = 1;
1174         }
1175
1176         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1177         up_read(&current->mm->mmap_sem);
1178
1179         /* mmio */
1180         if (is_error_pfn(pfn)) {
1181                 kvm_release_pfn_clean(pfn);
1182                 return 1;
1183         }
1184
1185         spin_lock(&vcpu->kvm->mmu_lock);
1186         kvm_mmu_free_some_pages(vcpu);
1187         r = __direct_map(vcpu, v, write, largepage, gfn, pfn,
1188                          PT32E_ROOT_LEVEL);
1189         spin_unlock(&vcpu->kvm->mmu_lock);
1190
1191
1192         return r;
1193 }
1194
1195
1196 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1197                                     struct kvm_mmu_page *sp)
1198 {
1199         int i;
1200
1201         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1202                 sp->spt[i] = shadow_trap_nonpresent_pte;
1203 }
1204
1205 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1206 {
1207         int i;
1208         struct kvm_mmu_page *sp;
1209
1210         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1211                 return;
1212         spin_lock(&vcpu->kvm->mmu_lock);
1213 #ifdef CONFIG_X86_64
1214         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1215                 hpa_t root = vcpu->arch.mmu.root_hpa;
1216
1217                 sp = page_header(root);
1218                 --sp->root_count;
1219                 if (!sp->root_count && sp->role.invalid)
1220                         kvm_mmu_zap_page(vcpu->kvm, sp);
1221                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1222                 spin_unlock(&vcpu->kvm->mmu_lock);
1223                 return;
1224         }
1225 #endif
1226         for (i = 0; i < 4; ++i) {
1227                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1228
1229                 if (root) {
1230                         root &= PT64_BASE_ADDR_MASK;
1231                         sp = page_header(root);
1232                         --sp->root_count;
1233                         if (!sp->root_count && sp->role.invalid)
1234                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1235                 }
1236                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1237         }
1238         spin_unlock(&vcpu->kvm->mmu_lock);
1239         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1240 }
1241
1242 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1243 {
1244         int i;
1245         gfn_t root_gfn;
1246         struct kvm_mmu_page *sp;
1247         int metaphysical = 0;
1248
1249         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1250
1251 #ifdef CONFIG_X86_64
1252         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1253                 hpa_t root = vcpu->arch.mmu.root_hpa;
1254
1255                 ASSERT(!VALID_PAGE(root));
1256                 if (tdp_enabled)
1257                         metaphysical = 1;
1258                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1259                                       PT64_ROOT_LEVEL, metaphysical,
1260                                       ACC_ALL, NULL);
1261                 root = __pa(sp->spt);
1262                 ++sp->root_count;
1263                 vcpu->arch.mmu.root_hpa = root;
1264                 return;
1265         }
1266 #endif
1267         metaphysical = !is_paging(vcpu);
1268         if (tdp_enabled)
1269                 metaphysical = 1;
1270         for (i = 0; i < 4; ++i) {
1271                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1272
1273                 ASSERT(!VALID_PAGE(root));
1274                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1275                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1276                                 vcpu->arch.mmu.pae_root[i] = 0;
1277                                 continue;
1278                         }
1279                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1280                 } else if (vcpu->arch.mmu.root_level == 0)
1281                         root_gfn = 0;
1282                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1283                                       PT32_ROOT_LEVEL, metaphysical,
1284                                       ACC_ALL, NULL);
1285                 root = __pa(sp->spt);
1286                 ++sp->root_count;
1287                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1288         }
1289         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1290 }
1291
1292 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1293 {
1294         return vaddr;
1295 }
1296
1297 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1298                                 u32 error_code)
1299 {
1300         gfn_t gfn;
1301         int r;
1302
1303         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1304         r = mmu_topup_memory_caches(vcpu);
1305         if (r)
1306                 return r;
1307
1308         ASSERT(vcpu);
1309         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1310
1311         gfn = gva >> PAGE_SHIFT;
1312
1313         return nonpaging_map(vcpu, gva & PAGE_MASK,
1314                              error_code & PFERR_WRITE_MASK, gfn);
1315 }
1316
1317 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1318                                 u32 error_code)
1319 {
1320         pfn_t pfn;
1321         int r;
1322         int largepage = 0;
1323         gfn_t gfn = gpa >> PAGE_SHIFT;
1324
1325         ASSERT(vcpu);
1326         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1327
1328         r = mmu_topup_memory_caches(vcpu);
1329         if (r)
1330                 return r;
1331
1332         down_read(&current->mm->mmap_sem);
1333         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1334                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1335                 largepage = 1;
1336         }
1337         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1338         up_read(&current->mm->mmap_sem);
1339         if (is_error_pfn(pfn)) {
1340                 kvm_release_pfn_clean(pfn);
1341                 return 1;
1342         }
1343         spin_lock(&vcpu->kvm->mmu_lock);
1344         kvm_mmu_free_some_pages(vcpu);
1345         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1346                          largepage, gfn, pfn, TDP_ROOT_LEVEL);
1347         spin_unlock(&vcpu->kvm->mmu_lock);
1348
1349         return r;
1350 }
1351
1352 static void nonpaging_free(struct kvm_vcpu *vcpu)
1353 {
1354         mmu_free_roots(vcpu);
1355 }
1356
1357 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1358 {
1359         struct kvm_mmu *context = &vcpu->arch.mmu;
1360
1361         context->new_cr3 = nonpaging_new_cr3;
1362         context->page_fault = nonpaging_page_fault;
1363         context->gva_to_gpa = nonpaging_gva_to_gpa;
1364         context->free = nonpaging_free;
1365         context->prefetch_page = nonpaging_prefetch_page;
1366         context->root_level = 0;
1367         context->shadow_root_level = PT32E_ROOT_LEVEL;
1368         context->root_hpa = INVALID_PAGE;
1369         return 0;
1370 }
1371
1372 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1373 {
1374         ++vcpu->stat.tlb_flush;
1375         kvm_x86_ops->tlb_flush(vcpu);
1376 }
1377
1378 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1379 {
1380         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1381         mmu_free_roots(vcpu);
1382 }
1383
1384 static void inject_page_fault(struct kvm_vcpu *vcpu,
1385                               u64 addr,
1386                               u32 err_code)
1387 {
1388         kvm_inject_page_fault(vcpu, addr, err_code);
1389 }
1390
1391 static void paging_free(struct kvm_vcpu *vcpu)
1392 {
1393         nonpaging_free(vcpu);
1394 }
1395
1396 #define PTTYPE 64
1397 #include "paging_tmpl.h"
1398 #undef PTTYPE
1399
1400 #define PTTYPE 32
1401 #include "paging_tmpl.h"
1402 #undef PTTYPE
1403
1404 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1405 {
1406         struct kvm_mmu *context = &vcpu->arch.mmu;
1407
1408         ASSERT(is_pae(vcpu));
1409         context->new_cr3 = paging_new_cr3;
1410         context->page_fault = paging64_page_fault;
1411         context->gva_to_gpa = paging64_gva_to_gpa;
1412         context->prefetch_page = paging64_prefetch_page;
1413         context->free = paging_free;
1414         context->root_level = level;
1415         context->shadow_root_level = level;
1416         context->root_hpa = INVALID_PAGE;
1417         return 0;
1418 }
1419
1420 static int paging64_init_context(struct kvm_vcpu *vcpu)
1421 {
1422         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1423 }
1424
1425 static int paging32_init_context(struct kvm_vcpu *vcpu)
1426 {
1427         struct kvm_mmu *context = &vcpu->arch.mmu;
1428
1429         context->new_cr3 = paging_new_cr3;
1430         context->page_fault = paging32_page_fault;
1431         context->gva_to_gpa = paging32_gva_to_gpa;
1432         context->free = paging_free;
1433         context->prefetch_page = paging32_prefetch_page;
1434         context->root_level = PT32_ROOT_LEVEL;
1435         context->shadow_root_level = PT32E_ROOT_LEVEL;
1436         context->root_hpa = INVALID_PAGE;
1437         return 0;
1438 }
1439
1440 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1441 {
1442         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1443 }
1444
1445 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1446 {
1447         struct kvm_mmu *context = &vcpu->arch.mmu;
1448
1449         context->new_cr3 = nonpaging_new_cr3;
1450         context->page_fault = tdp_page_fault;
1451         context->free = nonpaging_free;
1452         context->prefetch_page = nonpaging_prefetch_page;
1453         context->shadow_root_level = TDP_ROOT_LEVEL;
1454         context->root_hpa = INVALID_PAGE;
1455
1456         if (!is_paging(vcpu)) {
1457                 context->gva_to_gpa = nonpaging_gva_to_gpa;
1458                 context->root_level = 0;
1459         } else if (is_long_mode(vcpu)) {
1460                 context->gva_to_gpa = paging64_gva_to_gpa;
1461                 context->root_level = PT64_ROOT_LEVEL;
1462         } else if (is_pae(vcpu)) {
1463                 context->gva_to_gpa = paging64_gva_to_gpa;
1464                 context->root_level = PT32E_ROOT_LEVEL;
1465         } else {
1466                 context->gva_to_gpa = paging32_gva_to_gpa;
1467                 context->root_level = PT32_ROOT_LEVEL;
1468         }
1469
1470         return 0;
1471 }
1472
1473 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1474 {
1475         ASSERT(vcpu);
1476         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1477
1478         if (!is_paging(vcpu))
1479                 return nonpaging_init_context(vcpu);
1480         else if (is_long_mode(vcpu))
1481                 return paging64_init_context(vcpu);
1482         else if (is_pae(vcpu))
1483                 return paging32E_init_context(vcpu);
1484         else
1485                 return paging32_init_context(vcpu);
1486 }
1487
1488 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1489 {
1490         vcpu->arch.update_pte.pfn = bad_pfn;
1491
1492         if (tdp_enabled)
1493                 return init_kvm_tdp_mmu(vcpu);
1494         else
1495                 return init_kvm_softmmu(vcpu);
1496 }
1497
1498 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1499 {
1500         ASSERT(vcpu);
1501         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1502                 vcpu->arch.mmu.free(vcpu);
1503                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1504         }
1505 }
1506
1507 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1508 {
1509         destroy_kvm_mmu(vcpu);
1510         return init_kvm_mmu(vcpu);
1511 }
1512 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1513
1514 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1515 {
1516         int r;
1517
1518         r = mmu_topup_memory_caches(vcpu);
1519         if (r)
1520                 goto out;
1521         spin_lock(&vcpu->kvm->mmu_lock);
1522         kvm_mmu_free_some_pages(vcpu);
1523         mmu_alloc_roots(vcpu);
1524         spin_unlock(&vcpu->kvm->mmu_lock);
1525         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1526         kvm_mmu_flush_tlb(vcpu);
1527 out:
1528         return r;
1529 }
1530 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1531
1532 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1533 {
1534         mmu_free_roots(vcpu);
1535 }
1536
1537 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1538                                   struct kvm_mmu_page *sp,
1539                                   u64 *spte)
1540 {
1541         u64 pte;
1542         struct kvm_mmu_page *child;
1543
1544         pte = *spte;
1545         if (is_shadow_present_pte(pte)) {
1546                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1547                     is_large_pte(pte))
1548                         rmap_remove(vcpu->kvm, spte);
1549                 else {
1550                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1551                         mmu_page_remove_parent_pte(child, spte);
1552                 }
1553         }
1554         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1555         if (is_large_pte(pte))
1556                 --vcpu->kvm->stat.lpages;
1557 }
1558
1559 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1560                                   struct kvm_mmu_page *sp,
1561                                   u64 *spte,
1562                                   const void *new)
1563 {
1564         if ((sp->role.level != PT_PAGE_TABLE_LEVEL)
1565             && !vcpu->arch.update_pte.largepage) {
1566                 ++vcpu->kvm->stat.mmu_pde_zapped;
1567                 return;
1568         }
1569
1570         ++vcpu->kvm->stat.mmu_pte_updated;
1571         if (sp->role.glevels == PT32_ROOT_LEVEL)
1572                 paging32_update_pte(vcpu, sp, spte, new);
1573         else
1574                 paging64_update_pte(vcpu, sp, spte, new);
1575 }
1576
1577 static bool need_remote_flush(u64 old, u64 new)
1578 {
1579         if (!is_shadow_present_pte(old))
1580                 return false;
1581         if (!is_shadow_present_pte(new))
1582                 return true;
1583         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1584                 return true;
1585         old ^= PT64_NX_MASK;
1586         new ^= PT64_NX_MASK;
1587         return (old & ~new & PT64_PERM_MASK) != 0;
1588 }
1589
1590 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1591 {
1592         if (need_remote_flush(old, new))
1593                 kvm_flush_remote_tlbs(vcpu->kvm);
1594         else
1595                 kvm_mmu_flush_tlb(vcpu);
1596 }
1597
1598 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1599 {
1600         u64 *spte = vcpu->arch.last_pte_updated;
1601
1602         return !!(spte && (*spte & PT_ACCESSED_MASK));
1603 }
1604
1605 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1606                                           const u8 *new, int bytes)
1607 {
1608         gfn_t gfn;
1609         int r;
1610         u64 gpte = 0;
1611         pfn_t pfn;
1612
1613         vcpu->arch.update_pte.largepage = 0;
1614
1615         if (bytes != 4 && bytes != 8)
1616                 return;
1617
1618         /*
1619          * Assume that the pte write on a page table of the same type
1620          * as the current vcpu paging mode.  This is nearly always true
1621          * (might be false while changing modes).  Note it is verified later
1622          * by update_pte().
1623          */
1624         if (is_pae(vcpu)) {
1625                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1626                 if ((bytes == 4) && (gpa % 4 == 0)) {
1627                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1628                         if (r)
1629                                 return;
1630                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1631                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1632                         memcpy((void *)&gpte, new, 8);
1633                 }
1634         } else {
1635                 if ((bytes == 4) && (gpa % 4 == 0))
1636                         memcpy((void *)&gpte, new, 4);
1637         }
1638         if (!is_present_pte(gpte))
1639                 return;
1640         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1641
1642         down_read(&current->mm->mmap_sem);
1643         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1644                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1645                 vcpu->arch.update_pte.largepage = 1;
1646         }
1647         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1648         up_read(&current->mm->mmap_sem);
1649
1650         if (is_error_pfn(pfn)) {
1651                 kvm_release_pfn_clean(pfn);
1652                 return;
1653         }
1654         vcpu->arch.update_pte.gfn = gfn;
1655         vcpu->arch.update_pte.pfn = pfn;
1656 }
1657
1658 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1659                        const u8 *new, int bytes)
1660 {
1661         gfn_t gfn = gpa >> PAGE_SHIFT;
1662         struct kvm_mmu_page *sp;
1663         struct hlist_node *node, *n;
1664         struct hlist_head *bucket;
1665         unsigned index;
1666         u64 entry, gentry;
1667         u64 *spte;
1668         unsigned offset = offset_in_page(gpa);
1669         unsigned pte_size;
1670         unsigned page_offset;
1671         unsigned misaligned;
1672         unsigned quadrant;
1673         int level;
1674         int flooded = 0;
1675         int npte;
1676         int r;
1677
1678         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1679         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1680         spin_lock(&vcpu->kvm->mmu_lock);
1681         kvm_mmu_free_some_pages(vcpu);
1682         ++vcpu->kvm->stat.mmu_pte_write;
1683         kvm_mmu_audit(vcpu, "pre pte write");
1684         if (gfn == vcpu->arch.last_pt_write_gfn
1685             && !last_updated_pte_accessed(vcpu)) {
1686                 ++vcpu->arch.last_pt_write_count;
1687                 if (vcpu->arch.last_pt_write_count >= 3)
1688                         flooded = 1;
1689         } else {
1690                 vcpu->arch.last_pt_write_gfn = gfn;
1691                 vcpu->arch.last_pt_write_count = 1;
1692                 vcpu->arch.last_pte_updated = NULL;
1693         }
1694         index = kvm_page_table_hashfn(gfn);
1695         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1696         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1697                 if (sp->gfn != gfn || sp->role.metaphysical)
1698                         continue;
1699                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1700                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1701                 misaligned |= bytes < 4;
1702                 if (misaligned || flooded) {
1703                         /*
1704                          * Misaligned accesses are too much trouble to fix
1705                          * up; also, they usually indicate a page is not used
1706                          * as a page table.
1707                          *
1708                          * If we're seeing too many writes to a page,
1709                          * it may no longer be a page table, or we may be
1710                          * forking, in which case it is better to unmap the
1711                          * page.
1712                          */
1713                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1714                                  gpa, bytes, sp->role.word);
1715                         kvm_mmu_zap_page(vcpu->kvm, sp);
1716                         ++vcpu->kvm->stat.mmu_flooded;
1717                         continue;
1718                 }
1719                 page_offset = offset;
1720                 level = sp->role.level;
1721                 npte = 1;
1722                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1723                         page_offset <<= 1;      /* 32->64 */
1724                         /*
1725                          * A 32-bit pde maps 4MB while the shadow pdes map
1726                          * only 2MB.  So we need to double the offset again
1727                          * and zap two pdes instead of one.
1728                          */
1729                         if (level == PT32_ROOT_LEVEL) {
1730                                 page_offset &= ~7; /* kill rounding error */
1731                                 page_offset <<= 1;
1732                                 npte = 2;
1733                         }
1734                         quadrant = page_offset >> PAGE_SHIFT;
1735                         page_offset &= ~PAGE_MASK;
1736                         if (quadrant != sp->role.quadrant)
1737                                 continue;
1738                 }
1739                 spte = &sp->spt[page_offset / sizeof(*spte)];
1740                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1741                         gentry = 0;
1742                         r = kvm_read_guest_atomic(vcpu->kvm,
1743                                                   gpa & ~(u64)(pte_size - 1),
1744                                                   &gentry, pte_size);
1745                         new = (const void *)&gentry;
1746                         if (r < 0)
1747                                 new = NULL;
1748                 }
1749                 while (npte--) {
1750                         entry = *spte;
1751                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1752                         if (new)
1753                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1754                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1755                         ++spte;
1756                 }
1757         }
1758         kvm_mmu_audit(vcpu, "post pte write");
1759         spin_unlock(&vcpu->kvm->mmu_lock);
1760         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
1761                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
1762                 vcpu->arch.update_pte.pfn = bad_pfn;
1763         }
1764 }
1765
1766 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1767 {
1768         gpa_t gpa;
1769         int r;
1770
1771         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1772
1773         spin_lock(&vcpu->kvm->mmu_lock);
1774         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1775         spin_unlock(&vcpu->kvm->mmu_lock);
1776         return r;
1777 }
1778
1779 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1780 {
1781         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1782                 struct kvm_mmu_page *sp;
1783
1784                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1785                                   struct kvm_mmu_page, link);
1786                 kvm_mmu_zap_page(vcpu->kvm, sp);
1787                 ++vcpu->kvm->stat.mmu_recycled;
1788         }
1789 }
1790
1791 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1792 {
1793         int r;
1794         enum emulation_result er;
1795
1796         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1797         if (r < 0)
1798                 goto out;
1799
1800         if (!r) {
1801                 r = 1;
1802                 goto out;
1803         }
1804
1805         r = mmu_topup_memory_caches(vcpu);
1806         if (r)
1807                 goto out;
1808
1809         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1810
1811         switch (er) {
1812         case EMULATE_DONE:
1813                 return 1;
1814         case EMULATE_DO_MMIO:
1815                 ++vcpu->stat.mmio_exits;
1816                 return 0;
1817         case EMULATE_FAIL:
1818                 kvm_report_emulation_failure(vcpu, "pagetable");
1819                 return 1;
1820         default:
1821                 BUG();
1822         }
1823 out:
1824         return r;
1825 }
1826 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1827
1828 void kvm_enable_tdp(void)
1829 {
1830         tdp_enabled = true;
1831 }
1832 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1833
1834 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1835 {
1836         struct kvm_mmu_page *sp;
1837
1838         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1839                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1840                                   struct kvm_mmu_page, link);
1841                 kvm_mmu_zap_page(vcpu->kvm, sp);
1842         }
1843         free_page((unsigned long)vcpu->arch.mmu.pae_root);
1844 }
1845
1846 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1847 {
1848         struct page *page;
1849         int i;
1850
1851         ASSERT(vcpu);
1852
1853         if (vcpu->kvm->arch.n_requested_mmu_pages)
1854                 vcpu->kvm->arch.n_free_mmu_pages =
1855                                         vcpu->kvm->arch.n_requested_mmu_pages;
1856         else
1857                 vcpu->kvm->arch.n_free_mmu_pages =
1858                                         vcpu->kvm->arch.n_alloc_mmu_pages;
1859         /*
1860          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1861          * Therefore we need to allocate shadow page tables in the first
1862          * 4GB of memory, which happens to fit the DMA32 zone.
1863          */
1864         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1865         if (!page)
1866                 goto error_1;
1867         vcpu->arch.mmu.pae_root = page_address(page);
1868         for (i = 0; i < 4; ++i)
1869                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1870
1871         return 0;
1872
1873 error_1:
1874         free_mmu_pages(vcpu);
1875         return -ENOMEM;
1876 }
1877
1878 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1879 {
1880         ASSERT(vcpu);
1881         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1882
1883         return alloc_mmu_pages(vcpu);
1884 }
1885
1886 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1887 {
1888         ASSERT(vcpu);
1889         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1890
1891         return init_kvm_mmu(vcpu);
1892 }
1893
1894 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1895 {
1896         ASSERT(vcpu);
1897
1898         destroy_kvm_mmu(vcpu);
1899         free_mmu_pages(vcpu);
1900         mmu_free_memory_caches(vcpu);
1901 }
1902
1903 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1904 {
1905         struct kvm_mmu_page *sp;
1906
1907         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1908                 int i;
1909                 u64 *pt;
1910
1911                 if (!test_bit(slot, &sp->slot_bitmap))
1912                         continue;
1913
1914                 pt = sp->spt;
1915                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1916                         /* avoid RMW */
1917                         if (pt[i] & PT_WRITABLE_MASK)
1918                                 pt[i] &= ~PT_WRITABLE_MASK;
1919         }
1920 }
1921
1922 void kvm_mmu_zap_all(struct kvm *kvm)
1923 {
1924         struct kvm_mmu_page *sp, *node;
1925
1926         spin_lock(&kvm->mmu_lock);
1927         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1928                 kvm_mmu_zap_page(kvm, sp);
1929         spin_unlock(&kvm->mmu_lock);
1930
1931         kvm_flush_remote_tlbs(kvm);
1932 }
1933
1934 void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
1935 {
1936         struct kvm_mmu_page *page;
1937
1938         page = container_of(kvm->arch.active_mmu_pages.prev,
1939                             struct kvm_mmu_page, link);
1940         kvm_mmu_zap_page(kvm, page);
1941 }
1942
1943 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
1944 {
1945         struct kvm *kvm;
1946         struct kvm *kvm_freed = NULL;
1947         int cache_count = 0;
1948
1949         spin_lock(&kvm_lock);
1950
1951         list_for_each_entry(kvm, &vm_list, vm_list) {
1952                 int npages;
1953
1954                 spin_lock(&kvm->mmu_lock);
1955                 npages = kvm->arch.n_alloc_mmu_pages -
1956                          kvm->arch.n_free_mmu_pages;
1957                 cache_count += npages;
1958                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
1959                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
1960                         cache_count--;
1961                         kvm_freed = kvm;
1962                 }
1963                 nr_to_scan--;
1964
1965                 spin_unlock(&kvm->mmu_lock);
1966         }
1967         if (kvm_freed)
1968                 list_move_tail(&kvm_freed->vm_list, &vm_list);
1969
1970         spin_unlock(&kvm_lock);
1971
1972         return cache_count;
1973 }
1974
1975 static struct shrinker mmu_shrinker = {
1976         .shrink = mmu_shrink,
1977         .seeks = DEFAULT_SEEKS * 10,
1978 };
1979
1980 void mmu_destroy_caches(void)
1981 {
1982         if (pte_chain_cache)
1983                 kmem_cache_destroy(pte_chain_cache);
1984         if (rmap_desc_cache)
1985                 kmem_cache_destroy(rmap_desc_cache);
1986         if (mmu_page_header_cache)
1987                 kmem_cache_destroy(mmu_page_header_cache);
1988 }
1989
1990 void kvm_mmu_module_exit(void)
1991 {
1992         mmu_destroy_caches();
1993         unregister_shrinker(&mmu_shrinker);
1994 }
1995
1996 int kvm_mmu_module_init(void)
1997 {
1998         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1999                                             sizeof(struct kvm_pte_chain),
2000                                             0, 0, NULL);
2001         if (!pte_chain_cache)
2002                 goto nomem;
2003         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2004                                             sizeof(struct kvm_rmap_desc),
2005                                             0, 0, NULL);
2006         if (!rmap_desc_cache)
2007                 goto nomem;
2008
2009         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2010                                                   sizeof(struct kvm_mmu_page),
2011                                                   0, 0, NULL);
2012         if (!mmu_page_header_cache)
2013                 goto nomem;
2014
2015         register_shrinker(&mmu_shrinker);
2016
2017         return 0;
2018
2019 nomem:
2020         mmu_destroy_caches();
2021         return -ENOMEM;
2022 }
2023
2024 /*
2025  * Caculate mmu pages needed for kvm.
2026  */
2027 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2028 {
2029         int i;
2030         unsigned int nr_mmu_pages;
2031         unsigned int  nr_pages = 0;
2032
2033         for (i = 0; i < kvm->nmemslots; i++)
2034                 nr_pages += kvm->memslots[i].npages;
2035
2036         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2037         nr_mmu_pages = max(nr_mmu_pages,
2038                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2039
2040         return nr_mmu_pages;
2041 }
2042
2043 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2044                                 unsigned len)
2045 {
2046         if (len > buffer->len)
2047                 return NULL;
2048         return buffer->ptr;
2049 }
2050
2051 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2052                                 unsigned len)
2053 {
2054         void *ret;
2055
2056         ret = pv_mmu_peek_buffer(buffer, len);
2057         if (!ret)
2058                 return ret;
2059         buffer->ptr += len;
2060         buffer->len -= len;
2061         buffer->processed += len;
2062         return ret;
2063 }
2064
2065 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2066                              gpa_t addr, gpa_t value)
2067 {
2068         int bytes = 8;
2069         int r;
2070
2071         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2072                 bytes = 4;
2073
2074         r = mmu_topup_memory_caches(vcpu);
2075         if (r)
2076                 return r;
2077
2078         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2079                 return -EFAULT;
2080
2081         return 1;
2082 }
2083
2084 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2085 {
2086         kvm_x86_ops->tlb_flush(vcpu);
2087         return 1;
2088 }
2089
2090 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2091 {
2092         spin_lock(&vcpu->kvm->mmu_lock);
2093         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2094         spin_unlock(&vcpu->kvm->mmu_lock);
2095         return 1;
2096 }
2097
2098 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2099                              struct kvm_pv_mmu_op_buffer *buffer)
2100 {
2101         struct kvm_mmu_op_header *header;
2102
2103         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2104         if (!header)
2105                 return 0;
2106         switch (header->op) {
2107         case KVM_MMU_OP_WRITE_PTE: {
2108                 struct kvm_mmu_op_write_pte *wpte;
2109
2110                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2111                 if (!wpte)
2112                         return 0;
2113                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2114                                         wpte->pte_val);
2115         }
2116         case KVM_MMU_OP_FLUSH_TLB: {
2117                 struct kvm_mmu_op_flush_tlb *ftlb;
2118
2119                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2120                 if (!ftlb)
2121                         return 0;
2122                 return kvm_pv_mmu_flush_tlb(vcpu);
2123         }
2124         case KVM_MMU_OP_RELEASE_PT: {
2125                 struct kvm_mmu_op_release_pt *rpt;
2126
2127                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2128                 if (!rpt)
2129                         return 0;
2130                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2131         }
2132         default: return 0;
2133         }
2134 }
2135
2136 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2137                   gpa_t addr, unsigned long *ret)
2138 {
2139         int r;
2140         struct kvm_pv_mmu_op_buffer buffer;
2141
2142         buffer.ptr = buffer.buf;
2143         buffer.len = min_t(unsigned long, bytes, sizeof buffer.buf);
2144         buffer.processed = 0;
2145
2146         r = kvm_read_guest(vcpu->kvm, addr, buffer.buf, buffer.len);
2147         if (r)
2148                 goto out;
2149
2150         while (buffer.len) {
2151                 r = kvm_pv_mmu_op_one(vcpu, &buffer);
2152                 if (r < 0)
2153                         goto out;
2154                 if (r == 0)
2155                         break;
2156         }
2157
2158         r = 1;
2159 out:
2160         *ret = buffer.processed;
2161         return r;
2162 }
2163
2164 #ifdef AUDIT
2165
2166 static const char *audit_msg;
2167
2168 static gva_t canonicalize(gva_t gva)
2169 {
2170 #ifdef CONFIG_X86_64
2171         gva = (long long)(gva << 16) >> 16;
2172 #endif
2173         return gva;
2174 }
2175
2176 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2177                                 gva_t va, int level)
2178 {
2179         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2180         int i;
2181         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2182
2183         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2184                 u64 ent = pt[i];
2185
2186                 if (ent == shadow_trap_nonpresent_pte)
2187                         continue;
2188
2189                 va = canonicalize(va);
2190                 if (level > 1) {
2191                         if (ent == shadow_notrap_nonpresent_pte)
2192                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2193                                        " in nonleaf level: levels %d gva %lx"
2194                                        " level %d pte %llx\n", audit_msg,
2195                                        vcpu->arch.mmu.root_level, va, level, ent);
2196
2197                         audit_mappings_page(vcpu, ent, va, level - 1);
2198                 } else {
2199                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2200                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2201
2202                         if (is_shadow_present_pte(ent)
2203                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2204                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2205                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2206                                        audit_msg, vcpu->arch.mmu.root_level,
2207                                        va, gpa, hpa, ent,
2208                                        is_shadow_present_pte(ent));
2209                         else if (ent == shadow_notrap_nonpresent_pte
2210                                  && !is_error_hpa(hpa))
2211                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2212                                        " valid guest gva %lx\n", audit_msg, va);
2213                         kvm_release_pfn_clean(pfn);
2214
2215                 }
2216         }
2217 }
2218
2219 static void audit_mappings(struct kvm_vcpu *vcpu)
2220 {
2221         unsigned i;
2222
2223         if (vcpu->arch.mmu.root_level == 4)
2224                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2225         else
2226                 for (i = 0; i < 4; ++i)
2227                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2228                                 audit_mappings_page(vcpu,
2229                                                     vcpu->arch.mmu.pae_root[i],
2230                                                     i << 30,
2231                                                     2);
2232 }
2233
2234 static int count_rmaps(struct kvm_vcpu *vcpu)
2235 {
2236         int nmaps = 0;
2237         int i, j, k;
2238
2239         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2240                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2241                 struct kvm_rmap_desc *d;
2242
2243                 for (j = 0; j < m->npages; ++j) {
2244                         unsigned long *rmapp = &m->rmap[j];
2245
2246                         if (!*rmapp)
2247                                 continue;
2248                         if (!(*rmapp & 1)) {
2249                                 ++nmaps;
2250                                 continue;
2251                         }
2252                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2253                         while (d) {
2254                                 for (k = 0; k < RMAP_EXT; ++k)
2255                                         if (d->shadow_ptes[k])
2256                                                 ++nmaps;
2257                                         else
2258                                                 break;
2259                                 d = d->more;
2260                         }
2261                 }
2262         }
2263         return nmaps;
2264 }
2265
2266 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2267 {
2268         int nmaps = 0;
2269         struct kvm_mmu_page *sp;
2270         int i;
2271
2272         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2273                 u64 *pt = sp->spt;
2274
2275                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2276                         continue;
2277
2278                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2279                         u64 ent = pt[i];
2280
2281                         if (!(ent & PT_PRESENT_MASK))
2282                                 continue;
2283                         if (!(ent & PT_WRITABLE_MASK))
2284                                 continue;
2285                         ++nmaps;
2286                 }
2287         }
2288         return nmaps;
2289 }
2290
2291 static void audit_rmap(struct kvm_vcpu *vcpu)
2292 {
2293         int n_rmap = count_rmaps(vcpu);
2294         int n_actual = count_writable_mappings(vcpu);
2295
2296         if (n_rmap != n_actual)
2297                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2298                        __func__, audit_msg, n_rmap, n_actual);
2299 }
2300
2301 static void audit_write_protection(struct kvm_vcpu *vcpu)
2302 {
2303         struct kvm_mmu_page *sp;
2304         struct kvm_memory_slot *slot;
2305         unsigned long *rmapp;
2306         gfn_t gfn;
2307
2308         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2309                 if (sp->role.metaphysical)
2310                         continue;
2311
2312                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2313                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2314                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2315                 if (*rmapp)
2316                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2317                                " mappings: gfn %lx role %x\n",
2318                                __func__, audit_msg, sp->gfn,
2319                                sp->role.word);
2320         }
2321 }
2322
2323 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2324 {
2325         int olddbg = dbg;
2326
2327         dbg = 0;
2328         audit_msg = msg;
2329         audit_rmap(vcpu);
2330         audit_write_protection(vcpu);
2331         audit_mappings(vcpu);
2332         dbg = olddbg;
2333 }
2334
2335 #endif