err.no Git - linux-2.6/blob - arch/x86/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "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 static u64 __read_mostly shadow_base_present_pte;
 156 static u64 __read_mostly shadow_nx_mask;
 157 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 158 static u64 __read_mostly shadow_user_mask;
 159 static u64 __read_mostly shadow_accessed_mask;
 160 static u64 __read_mostly shadow_dirty_mask;
 161
 162 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 163 {
 164         shadow_trap_nonpresent_pte = trap_pte;
 165         shadow_notrap_nonpresent_pte = notrap_pte;
 166 }
 167 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 168
 169 void kvm_mmu_set_base_ptes(u64 base_pte)
 170 {
 171         shadow_base_present_pte = base_pte;
 172 }
 173 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 174
 175 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 176                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 177 {
 178         shadow_user_mask = user_mask;
 179         shadow_accessed_mask = accessed_mask;
 180         shadow_dirty_mask = dirty_mask;
 181         shadow_nx_mask = nx_mask;
 182         shadow_x_mask = x_mask;
 183 }
 184 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 185
 186 static int is_write_protection(struct kvm_vcpu *vcpu)
 187 {
 188         return vcpu->arch.cr0 & X86_CR0_WP;
 189 }
 190
 191 static int is_cpuid_PSE36(void)
 192 {
 193         return 1;
 194 }
 195
 196 static int is_nx(struct kvm_vcpu *vcpu)
 197 {
 198         return vcpu->arch.shadow_efer & EFER_NX;
 199 }
 200
 201 static int is_present_pte(unsigned long pte)
 202 {
 203         return pte & PT_PRESENT_MASK;
 204 }
 205
 206 static int is_shadow_present_pte(u64 pte)
 207 {
 208         return pte != shadow_trap_nonpresent_pte
 209                 && pte != shadow_notrap_nonpresent_pte;
 210 }
 211
 212 static int is_large_pte(u64 pte)
 213 {
 214         return pte & PT_PAGE_SIZE_MASK;
 215 }
 216
 217 static int is_writeble_pte(unsigned long pte)
 218 {
 219         return pte & PT_WRITABLE_MASK;
 220 }
 221
 222 static int is_dirty_pte(unsigned long pte)
 223 {
 224         return pte & shadow_dirty_mask;
 225 }
 226
 227 static int is_rmap_pte(u64 pte)
 228 {
 229         return is_shadow_present_pte(pte);
 230 }
 231
 232 static pfn_t spte_to_pfn(u64 pte)
 233 {
 234         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 235 }
 236
 237 static gfn_t pse36_gfn_delta(u32 gpte)
 238 {
 239         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 240
 241         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 242 }
 243
 244 static void set_shadow_pte(u64 *sptep, u64 spte)
 245 {
 246 #ifdef CONFIG_X86_64
 247         set_64bit((unsigned long *)sptep, spte);
 248 #else
 249         set_64bit((unsigned long long *)sptep, spte);
 250 #endif
 251 }
 252
 253 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 254                                   struct kmem_cache *base_cache, int min)
 255 {
 256         void *obj;
 257
 258         if (cache->nobjs >= min)
 259                 return 0;
 260         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 261                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 262                 if (!obj)
 263                         return -ENOMEM;
 264                 cache->objects[cache->nobjs++] = obj;
 265         }
 266         return 0;
 267 }
 268
 269 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 270 {
 271         while (mc->nobjs)
 272                 kfree(mc->objects[--mc->nobjs]);
 273 }
 274
 275 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 276                                        int min)
 277 {
 278         struct page *page;
 279
 280         if (cache->nobjs >= min)
 281                 return 0;
 282         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 283                 page = alloc_page(GFP_KERNEL);
 284                 if (!page)
 285                         return -ENOMEM;
 286                 set_page_private(page, 0);
 287                 cache->objects[cache->nobjs++] = page_address(page);
 288         }
 289         return 0;
 290 }
 291
 292 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 293 {
 294         while (mc->nobjs)
 295                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 296 }
 297
 298 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 299 {
 300         int r;
 301
 302         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 303                                    pte_chain_cache, 4);
 304         if (r)
 305                 goto out;
 306         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 307                                    rmap_desc_cache, 1);
 308         if (r)
 309                 goto out;
 310         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 311         if (r)
 312                 goto out;
 313         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 314                                    mmu_page_header_cache, 4);
 315 out:
 316         return r;
 317 }
 318
 319 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 320 {
 321         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 322         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 323         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 324         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 325 }
 326
 327 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 328                                     size_t size)
 329 {
 330         void *p;
 331
 332         BUG_ON(!mc->nobjs);
 333         p = mc->objects[--mc->nobjs];
 334         memset(p, 0, size);
 335         return p;
 336 }
 337
 338 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 339 {
 340         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 341                                       sizeof(struct kvm_pte_chain));
 342 }
 343
 344 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 345 {
 346         kfree(pc);
 347 }
 348
 349 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 350 {
 351         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 352                                       sizeof(struct kvm_rmap_desc));
 353 }
 354
 355 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 356 {
 357         kfree(rd);
 358 }
 359
 360 /*
 361  * Return the pointer to the largepage write count for a given
 362  * gfn, handling slots that are not large page aligned.
 363  */
 364 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 365 {
 366         unsigned long idx;
 367
 368         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 369               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 370         return &slot->lpage_info[idx].write_count;
 371 }
 372
 373 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 374 {
 375         int *write_count;
 376
 377         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 378         *write_count += 1;
 379         WARN_ON(*write_count > KVM_PAGES_PER_HPAGE);
 380 }
 381
 382 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 383 {
 384         int *write_count;
 385
 386         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 387         *write_count -= 1;
 388         WARN_ON(*write_count < 0);
 389 }
 390
 391 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 392 {
 393         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
 394         int *largepage_idx;
 395
 396         if (slot) {
 397                 largepage_idx = slot_largepage_idx(gfn, slot);
 398                 return *largepage_idx;
 399         }
 400
 401         return 1;
 402 }
 403
 404 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 405 {
 406         struct vm_area_struct *vma;
 407         unsigned long addr;
 408
 409         addr = gfn_to_hva(kvm, gfn);
 410         if (kvm_is_error_hva(addr))
 411                 return 0;
 412
 413         vma = find_vma(current->mm, addr);
 414         if (vma && is_vm_hugetlb_page(vma))
 415                 return 1;
 416
 417         return 0;
 418 }
 419
 420 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 421 {
 422         struct kvm_memory_slot *slot;
 423
 424         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 425                 return 0;
 426
 427         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 428                 return 0;
 429
 430         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 431         if (slot && slot->dirty_bitmap)
 432                 return 0;
 433
 434         return 1;
 435 }
 436
 437 /*
 438  * Take gfn and return the reverse mapping to it.
 439  * Note: gfn must be unaliased before this function get called
 440  */
 441
 442 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 443 {
 444         struct kvm_memory_slot *slot;
 445         unsigned long idx;
 446
 447         slot = gfn_to_memslot(kvm, gfn);
 448         if (!lpage)
 449                 return &slot->rmap[gfn - slot->base_gfn];
 450
 451         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 452               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 453
 454         return &slot->lpage_info[idx].rmap_pde;
 455 }
 456
 457 /*
 458  * Reverse mapping data structures:
 459  *
 460  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 461  * that points to page_address(page).
 462  *
 463  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 464  * containing more mappings.
 465  */
 466 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 467 {
 468         struct kvm_mmu_page *sp;
 469         struct kvm_rmap_desc *desc;
 470         unsigned long *rmapp;
 471         int i;
 472
 473         if (!is_rmap_pte(*spte))
 474                 return;
 475         gfn = unalias_gfn(vcpu->kvm, gfn);
 476         sp = page_header(__pa(spte));
 477         sp->gfns[spte - sp->spt] = gfn;
 478         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 479         if (!*rmapp) {
 480                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 481                 *rmapp = (unsigned long)spte;
 482         } else if (!(*rmapp & 1)) {
 483                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 484                 desc = mmu_alloc_rmap_desc(vcpu);
 485                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 486                 desc->shadow_ptes[1] = spte;
 487                 *rmapp = (unsigned long)desc | 1;
 488         } else {
 489                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 490                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 491                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 492                         desc = desc->more;
 493                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 494                         desc->more = mmu_alloc_rmap_desc(vcpu);
 495                         desc = desc->more;
 496                 }
 497                 for (i = 0; desc->shadow_ptes[i]; ++i)
 498                         ;
 499                 desc->shadow_ptes[i] = spte;
 500         }
 501 }
 502
 503 static void rmap_desc_remove_entry(unsigned long *rmapp,
 504                                    struct kvm_rmap_desc *desc,
 505                                    int i,
 506                                    struct kvm_rmap_desc *prev_desc)
 507 {
 508         int j;
 509
 510         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 511                 ;
 512         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 513         desc->shadow_ptes[j] = NULL;
 514         if (j != 0)
 515                 return;
 516         if (!prev_desc && !desc->more)
 517                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 518         else
 519                 if (prev_desc)
 520                         prev_desc->more = desc->more;
 521                 else
 522                         *rmapp = (unsigned long)desc->more | 1;
 523         mmu_free_rmap_desc(desc);
 524 }
 525
 526 static void rmap_remove(struct kvm *kvm, u64 *spte)
 527 {
 528         struct kvm_rmap_desc *desc;
 529         struct kvm_rmap_desc *prev_desc;
 530         struct kvm_mmu_page *sp;
 531         pfn_t pfn;
 532         unsigned long *rmapp;
 533         int i;
 534
 535         if (!is_rmap_pte(*spte))
 536                 return;
 537         sp = page_header(__pa(spte));
 538         pfn = spte_to_pfn(*spte);
 539         if (*spte & shadow_accessed_mask)
 540                 kvm_set_pfn_accessed(pfn);
 541         if (is_writeble_pte(*spte))
 542                 kvm_release_pfn_dirty(pfn);
 543         else
 544                 kvm_release_pfn_clean(pfn);
 545         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 546         if (!*rmapp) {
 547                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 548                 BUG();
 549         } else if (!(*rmapp & 1)) {
 550                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 551                 if ((u64 *)*rmapp != spte) {
 552                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 553                                spte, *spte);
 554                         BUG();
 555                 }
 556                 *rmapp = 0;
 557         } else {
 558                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 559                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 560                 prev_desc = NULL;
 561                 while (desc) {
 562                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 563                                 if (desc->shadow_ptes[i] == spte) {
 564                                         rmap_desc_remove_entry(rmapp,
 565                                                                desc, i,
 566                                                                prev_desc);
 567                                         return;
 568                                 }
 569                         prev_desc = desc;
 570                         desc = desc->more;
 571                 }
 572                 BUG();
 573         }
 574 }
 575
 576 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 577 {
 578         struct kvm_rmap_desc *desc;
 579         struct kvm_rmap_desc *prev_desc;
 580         u64 *prev_spte;
 581         int i;
 582
 583         if (!*rmapp)
 584                 return NULL;
 585         else if (!(*rmapp & 1)) {
 586                 if (!spte)
 587                         return (u64 *)*rmapp;
 588                 return NULL;
 589         }
 590         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 591         prev_desc = NULL;
 592         prev_spte = NULL;
 593         while (desc) {
 594                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 595                         if (prev_spte == spte)
 596                                 return desc->shadow_ptes[i];
 597                         prev_spte = desc->shadow_ptes[i];
 598                 }
 599                 desc = desc->more;
 600         }
 601         return NULL;
 602 }
 603
 604 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 605 {
 606         unsigned long *rmapp;
 607         u64 *spte;
 608         int write_protected = 0;
 609
 610         gfn = unalias_gfn(kvm, gfn);
 611         rmapp = gfn_to_rmap(kvm, gfn, 0);
 612
 613         spte = rmap_next(kvm, rmapp, NULL);
 614         while (spte) {
 615                 BUG_ON(!spte);
 616                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 617                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 618                 if (is_writeble_pte(*spte)) {
 619                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 620                         write_protected = 1;
 621                 }
 622                 spte = rmap_next(kvm, rmapp, spte);
 623         }
 624         if (write_protected) {
 625                 pfn_t pfn;
 626
 627                 spte = rmap_next(kvm, rmapp, NULL);
 628                 pfn = spte_to_pfn(*spte);
 629                 kvm_set_pfn_dirty(pfn);
 630         }
 631
 632         /* check for huge page mappings */
 633         rmapp = gfn_to_rmap(kvm, gfn, 1);
 634         spte = rmap_next(kvm, rmapp, NULL);
 635         while (spte) {
 636                 BUG_ON(!spte);
 637                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 638                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 639                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 640                 if (is_writeble_pte(*spte)) {
 641                         rmap_remove(kvm, spte);
 642                         --kvm->stat.lpages;
 643                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 644                         write_protected = 1;
 645                 }
 646                 spte = rmap_next(kvm, rmapp, spte);
 647         }
 648
 649         if (write_protected)
 650                 kvm_flush_remote_tlbs(kvm);
 651
 652         account_shadowed(kvm, gfn);
 653 }
 654
 655 #ifdef MMU_DEBUG
 656 static int is_empty_shadow_page(u64 *spt)
 657 {
 658         u64 *pos;
 659         u64 *end;
 660
 661         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 662                 if (*pos != shadow_trap_nonpresent_pte) {
 663                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 664                                pos, *pos);
 665                         return 0;
 666                 }
 667         return 1;
 668 }
 669 #endif
 670
 671 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 672 {
 673         ASSERT(is_empty_shadow_page(sp->spt));
 674         list_del(&sp->link);
 675         __free_page(virt_to_page(sp->spt));
 676         __free_page(virt_to_page(sp->gfns));
 677         kfree(sp);
 678         ++kvm->arch.n_free_mmu_pages;
 679 }
 680
 681 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 682 {
 683         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 684 }
 685
 686 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 687                                                u64 *parent_pte)
 688 {
 689         struct kvm_mmu_page *sp;
 690
 691         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 692         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 693         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 694         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 695         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 696         ASSERT(is_empty_shadow_page(sp->spt));
 697         sp->slot_bitmap = 0;
 698         sp->multimapped = 0;
 699         sp->parent_pte = parent_pte;
 700         --vcpu->kvm->arch.n_free_mmu_pages;
 701         return sp;
 702 }
 703
 704 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 705                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 706 {
 707         struct kvm_pte_chain *pte_chain;
 708         struct hlist_node *node;
 709         int i;
 710
 711         if (!parent_pte)
 712                 return;
 713         if (!sp->multimapped) {
 714                 u64 *old = sp->parent_pte;
 715
 716                 if (!old) {
 717                         sp->parent_pte = parent_pte;
 718                         return;
 719                 }
 720                 sp->multimapped = 1;
 721                 pte_chain = mmu_alloc_pte_chain(vcpu);
 722                 INIT_HLIST_HEAD(&sp->parent_ptes);
 723                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 724                 pte_chain->parent_ptes[0] = old;
 725         }
 726         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 727                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 728                         continue;
 729                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 730                         if (!pte_chain->parent_ptes[i]) {
 731                                 pte_chain->parent_ptes[i] = parent_pte;
 732                                 return;
 733                         }
 734         }
 735         pte_chain = mmu_alloc_pte_chain(vcpu);
 736         BUG_ON(!pte_chain);
 737         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 738         pte_chain->parent_ptes[0] = parent_pte;
 739 }
 740
 741 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 742                                        u64 *parent_pte)
 743 {
 744         struct kvm_pte_chain *pte_chain;
 745         struct hlist_node *node;
 746         int i;
 747
 748         if (!sp->multimapped) {
 749                 BUG_ON(sp->parent_pte != parent_pte);
 750                 sp->parent_pte = NULL;
 751                 return;
 752         }
 753         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 754                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 755                         if (!pte_chain->parent_ptes[i])
 756                                 break;
 757                         if (pte_chain->parent_ptes[i] != parent_pte)
 758                                 continue;
 759                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 760                                 && pte_chain->parent_ptes[i + 1]) {
 761                                 pte_chain->parent_ptes[i]
 762                                         = pte_chain->parent_ptes[i + 1];
 763                                 ++i;
 764                         }
 765                         pte_chain->parent_ptes[i] = NULL;
 766                         if (i == 0) {
 767                                 hlist_del(&pte_chain->link);
 768                                 mmu_free_pte_chain(pte_chain);
 769                                 if (hlist_empty(&sp->parent_ptes)) {
 770                                         sp->multimapped = 0;
 771                                         sp->parent_pte = NULL;
 772                                 }
 773                         }
 774                         return;
 775                 }
 776         BUG();
 777 }
 778
 779 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
 780 {
 781         unsigned index;
 782         struct hlist_head *bucket;
 783         struct kvm_mmu_page *sp;
 784         struct hlist_node *node;
 785
 786         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 787         index = kvm_page_table_hashfn(gfn);
 788         bucket = &kvm->arch.mmu_page_hash[index];
 789         hlist_for_each_entry(sp, node, bucket, hash_link)
 790                 if (sp->gfn == gfn && !sp->role.metaphysical
 791                     && !sp->role.invalid) {
 792                         pgprintk("%s: found role %x\n",
 793                                  __func__, sp->role.word);
 794                         return sp;
 795                 }
 796         return NULL;
 797 }
 798
 799 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 800                                              gfn_t gfn,
 801                                              gva_t gaddr,
 802                                              unsigned level,
 803                                              int metaphysical,
 804                                              unsigned access,
 805                                              u64 *parent_pte)
 806 {
 807         union kvm_mmu_page_role role;
 808         unsigned index;
 809         unsigned quadrant;
 810         struct hlist_head *bucket;
 811         struct kvm_mmu_page *sp;
 812         struct hlist_node *node;
 813
 814         role.word = 0;
 815         role.glevels = vcpu->arch.mmu.root_level;
 816         role.level = level;
 817         role.metaphysical = metaphysical;
 818         role.access = access;
 819         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
 820                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 821                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 822                 role.quadrant = quadrant;
 823         }
 824         pgprintk("%s: looking gfn %lx role %x\n", __func__,
 825                  gfn, role.word);
 826         index = kvm_page_table_hashfn(gfn);
 827         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
 828         hlist_for_each_entry(sp, node, bucket, hash_link)
 829                 if (sp->gfn == gfn && sp->role.word == role.word) {
 830                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
 831                         pgprintk("%s: found\n", __func__);
 832                         return sp;
 833                 }
 834         ++vcpu->kvm->stat.mmu_cache_miss;
 835         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
 836         if (!sp)
 837                 return sp;
 838         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
 839         sp->gfn = gfn;
 840         sp->role = role;
 841         hlist_add_head(&sp->hash_link, bucket);
 842         if (!metaphysical)
 843                 rmap_write_protect(vcpu->kvm, gfn);
 844         vcpu->arch.mmu.prefetch_page(vcpu, sp);
 845         return sp;
 846 }
 847
 848 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 849                                          struct kvm_mmu_page *sp)
 850 {
 851         unsigned i;
 852         u64 *pt;
 853         u64 ent;
 854
 855         pt = sp->spt;
 856
 857         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
 858                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 859                         if (is_shadow_present_pte(pt[i]))
 860                                 rmap_remove(kvm, &pt[i]);
 861                         pt[i] = shadow_trap_nonpresent_pte;
 862                 }
 863                 kvm_flush_remote_tlbs(kvm);
 864                 return;
 865         }
 866
 867         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 868                 ent = pt[i];
 869
 870                 if (is_shadow_present_pte(ent)) {
 871                         if (!is_large_pte(ent)) {
 872                                 ent &= PT64_BASE_ADDR_MASK;
 873                                 mmu_page_remove_parent_pte(page_header(ent),
 874                                                            &pt[i]);
 875                         } else {
 876                                 --kvm->stat.lpages;
 877                                 rmap_remove(kvm, &pt[i]);
 878                         }
 879                 }
 880                 pt[i] = shadow_trap_nonpresent_pte;
 881         }
 882         kvm_flush_remote_tlbs(kvm);
 883 }
 884
 885 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
 886 {
 887         mmu_page_remove_parent_pte(sp, parent_pte);
 888 }
 889
 890 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 891 {
 892         int i;
 893
 894         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 895                 if (kvm->vcpus[i])
 896                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
 897 }
 898
 899 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 900 {
 901         u64 *parent_pte;
 902
 903         ++kvm->stat.mmu_shadow_zapped;
 904         while (sp->multimapped || sp->parent_pte) {
 905                 if (!sp->multimapped)
 906                         parent_pte = sp->parent_pte;
 907                 else {
 908                         struct kvm_pte_chain *chain;
 909
 910                         chain = container_of(sp->parent_ptes.first,
 911                                              struct kvm_pte_chain, link);
 912                         parent_pte = chain->parent_ptes[0];
 913                 }
 914                 BUG_ON(!parent_pte);
 915                 kvm_mmu_put_page(sp, parent_pte);
 916                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 917         }
 918         kvm_mmu_page_unlink_children(kvm, sp);
 919         if (!sp->root_count) {
 920                 if (!sp->role.metaphysical)
 921                         unaccount_shadowed(kvm, sp->gfn);
 922                 hlist_del(&sp->hash_link);
 923                 kvm_mmu_free_page(kvm, sp);
 924         } else {
 925                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
 926                 sp->role.invalid = 1;
 927                 kvm_reload_remote_mmus(kvm);
 928         }
 929         kvm_mmu_reset_last_pte_updated(kvm);
 930 }
 931
 932 /*
 933  * Changing the number of mmu pages allocated to the vm
 934  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 935  */
 936 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 937 {
 938         /*
 939          * If we set the number of mmu pages to be smaller be than the
 940          * number of actived pages , we must to free some mmu pages before we
 941          * change the value
 942          */
 943
 944         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
 945             kvm_nr_mmu_pages) {
 946                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
 947                                        - kvm->arch.n_free_mmu_pages;
 948
 949                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 950                         struct kvm_mmu_page *page;
 951
 952                         page = container_of(kvm->arch.active_mmu_pages.prev,
 953                                             struct kvm_mmu_page, link);
 954                         kvm_mmu_zap_page(kvm, page);
 955                         n_used_mmu_pages--;
 956                 }
 957                 kvm->arch.n_free_mmu_pages = 0;
 958         }
 959         else
 960                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
 961                                          - kvm->arch.n_alloc_mmu_pages;
 962
 963         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
 964 }
 965
 966 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 967 {
 968         unsigned index;
 969         struct hlist_head *bucket;
 970         struct kvm_mmu_page *sp;
 971         struct hlist_node *node, *n;
 972         int r;
 973
 974         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 975         r = 0;
 976         index = kvm_page_table_hashfn(gfn);
 977         bucket = &kvm->arch.mmu_page_hash[index];
 978         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
 979                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 980                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
 981                                  sp->role.word);
 982                         kvm_mmu_zap_page(kvm, sp);
 983                         r = 1;
 984                 }
 985         return r;
 986 }
 987
 988 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
 989 {
 990         struct kvm_mmu_page *sp;
 991
 992         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
 993                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
 994                 kvm_mmu_zap_page(kvm, sp);
 995         }
 996 }
 997
 998 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
 999 {
1000         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1001         struct kvm_mmu_page *sp = page_header(__pa(pte));
1002
1003         __set_bit(slot, &sp->slot_bitmap);
1004 }
1005
1006 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1007 {
1008         struct page *page;
1009
1010         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1011
1012         if (gpa == UNMAPPED_GVA)
1013                 return NULL;
1014
1015         down_read(&current->mm->mmap_sem);
1016         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1017         up_read(&current->mm->mmap_sem);
1018
1019         return page;
1020 }
1021
1022 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1023                          unsigned pt_access, unsigned pte_access,
1024                          int user_fault, int write_fault, int dirty,
1025                          int *ptwrite, int largepage, gfn_t gfn,
1026                          pfn_t pfn, bool speculative)
1027 {
1028         u64 spte;
1029         int was_rmapped = 0;
1030         int was_writeble = is_writeble_pte(*shadow_pte);
1031
1032         pgprintk("%s: spte %llx access %x write_fault %d"
1033                  " user_fault %d gfn %lx\n",
1034                  __func__, *shadow_pte, pt_access,
1035                  write_fault, user_fault, gfn);
1036
1037         if (is_rmap_pte(*shadow_pte)) {
1038                 /*
1039                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1040                  * the parent of the now unreachable PTE.
1041                  */
1042                 if (largepage && !is_large_pte(*shadow_pte)) {
1043                         struct kvm_mmu_page *child;
1044                         u64 pte = *shadow_pte;
1045
1046                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1047                         mmu_page_remove_parent_pte(child, shadow_pte);
1048                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1049                         pgprintk("hfn old %lx new %lx\n",
1050                                  spte_to_pfn(*shadow_pte), pfn);
1051                         rmap_remove(vcpu->kvm, shadow_pte);
1052                 } else {
1053                         if (largepage)
1054                                 was_rmapped = is_large_pte(*shadow_pte);
1055                         else
1056                                 was_rmapped = 1;
1057                 }
1058         }
1059
1060         /*
1061          * We don't set the accessed bit, since we sometimes want to see
1062          * whether the guest actually used the pte (in order to detect
1063          * demand paging).
1064          */
1065         spte = shadow_base_present_pte | shadow_dirty_mask;
1066         if (!speculative)
1067                 pte_access |= PT_ACCESSED_MASK;
1068         if (!dirty)
1069                 pte_access &= ~ACC_WRITE_MASK;
1070         if (pte_access & ACC_EXEC_MASK)
1071                 spte |= shadow_x_mask;
1072         else
1073                 spte |= shadow_nx_mask;
1074         if (pte_access & ACC_USER_MASK)
1075                 spte |= shadow_user_mask;
1076         if (largepage)
1077                 spte |= PT_PAGE_SIZE_MASK;
1078
1079         spte |= (u64)pfn << PAGE_SHIFT;
1080
1081         if ((pte_access & ACC_WRITE_MASK)
1082             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1083                 struct kvm_mmu_page *shadow;
1084
1085                 spte |= PT_WRITABLE_MASK;
1086                 if (user_fault) {
1087                         mmu_unshadow(vcpu->kvm, gfn);
1088                         goto unshadowed;
1089                 }
1090
1091                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1092                 if (shadow ||
1093                    (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
1094                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1095                                  __func__, gfn);
1096                         pte_access &= ~ACC_WRITE_MASK;
1097                         if (is_writeble_pte(spte)) {
1098                                 spte &= ~PT_WRITABLE_MASK;
1099                                 kvm_x86_ops->tlb_flush(vcpu);
1100                         }
1101                         if (write_fault)
1102                                 *ptwrite = 1;
1103                 }
1104         }
1105
1106 unshadowed:
1107
1108         if (pte_access & ACC_WRITE_MASK)
1109                 mark_page_dirty(vcpu->kvm, gfn);
1110
1111         pgprintk("%s: setting spte %llx\n", __func__, spte);
1112         pgprintk("instantiating %s PTE (%s) at %d (%llx) addr %llx\n",
1113                  (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
1114                  (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
1115         set_shadow_pte(shadow_pte, spte);
1116         if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
1117             && (spte & PT_PRESENT_MASK))
1118                 ++vcpu->kvm->stat.lpages;
1119
1120         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1121         if (!was_rmapped) {
1122                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1123                 if (!is_rmap_pte(*shadow_pte))
1124                         kvm_release_pfn_clean(pfn);
1125         } else {
1126                 if (was_writeble)
1127                         kvm_release_pfn_dirty(pfn);
1128                 else
1129                         kvm_release_pfn_clean(pfn);
1130         }
1131         if (!ptwrite || !*ptwrite)
1132                 vcpu->arch.last_pte_updated = shadow_pte;
1133 }
1134
1135 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1136 {
1137 }
1138
1139 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1140                            int largepage, gfn_t gfn, pfn_t pfn,
1141                            int level)
1142 {
1143         hpa_t table_addr = vcpu->arch.mmu.root_hpa;
1144         int pt_write = 0;
1145
1146         for (; ; level--) {
1147                 u32 index = PT64_INDEX(v, level);
1148                 u64 *table;
1149
1150                 ASSERT(VALID_PAGE(table_addr));
1151                 table = __va(table_addr);
1152
1153                 if (level == 1) {
1154                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1155                                      0, write, 1, &pt_write, 0, gfn, pfn, false);
1156                         return pt_write;
1157                 }
1158
1159                 if (largepage && level == 2) {
1160                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1161                                      0, write, 1, &pt_write, 1, gfn, pfn, false);
1162                         return pt_write;
1163                 }
1164
1165                 if (table[index] == shadow_trap_nonpresent_pte) {
1166                         struct kvm_mmu_page *new_table;
1167                         gfn_t pseudo_gfn;
1168
1169                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1170                                 >> PAGE_SHIFT;
1171                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1172                                                      v, level - 1,
1173                                                      1, ACC_ALL, &table[index]);
1174                         if (!new_table) {
1175                                 pgprintk("nonpaging_map: ENOMEM\n");
1176                                 kvm_release_pfn_clean(pfn);
1177                                 return -ENOMEM;
1178                         }
1179
1180                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
1181                                 | PT_WRITABLE_MASK | shadow_user_mask;
1182                 }
1183                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1184         }
1185 }
1186
1187 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1188 {
1189         int r;
1190         int largepage = 0;
1191         pfn_t pfn;
1192
1193         down_read(&current->mm->mmap_sem);
1194         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1195                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1196                 largepage = 1;
1197         }
1198
1199         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1200         up_read(&current->mm->mmap_sem);
1201
1202         /* mmio */
1203         if (is_error_pfn(pfn)) {
1204                 kvm_release_pfn_clean(pfn);
1205                 return 1;
1206         }
1207
1208         spin_lock(&vcpu->kvm->mmu_lock);
1209         kvm_mmu_free_some_pages(vcpu);
1210         r = __direct_map(vcpu, v, write, largepage, gfn, pfn,
1211                          PT32E_ROOT_LEVEL);
1212         spin_unlock(&vcpu->kvm->mmu_lock);
1213
1214
1215         return r;
1216 }
1217
1218
1219 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1220                                     struct kvm_mmu_page *sp)
1221 {
1222         int i;
1223
1224         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1225                 sp->spt[i] = shadow_trap_nonpresent_pte;
1226 }
1227
1228 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1229 {
1230         int i;
1231         struct kvm_mmu_page *sp;
1232
1233         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1234                 return;
1235         spin_lock(&vcpu->kvm->mmu_lock);
1236 #ifdef CONFIG_X86_64
1237         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1238                 hpa_t root = vcpu->arch.mmu.root_hpa;
1239
1240                 sp = page_header(root);
1241                 --sp->root_count;
1242                 if (!sp->root_count && sp->role.invalid)
1243                         kvm_mmu_zap_page(vcpu->kvm, sp);
1244                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1245                 spin_unlock(&vcpu->kvm->mmu_lock);
1246                 return;
1247         }
1248 #endif
1249         for (i = 0; i < 4; ++i) {
1250                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1251
1252                 if (root) {
1253                         root &= PT64_BASE_ADDR_MASK;
1254                         sp = page_header(root);
1255                         --sp->root_count;
1256                         if (!sp->root_count && sp->role.invalid)
1257                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1258                 }
1259                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1260         }
1261         spin_unlock(&vcpu->kvm->mmu_lock);
1262         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1263 }
1264
1265 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1266 {
1267         int i;
1268         gfn_t root_gfn;
1269         struct kvm_mmu_page *sp;
1270         int metaphysical = 0;
1271
1272         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1273
1274 #ifdef CONFIG_X86_64
1275         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1276                 hpa_t root = vcpu->arch.mmu.root_hpa;
1277
1278                 ASSERT(!VALID_PAGE(root));
1279                 if (tdp_enabled)
1280                         metaphysical = 1;
1281                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1282                                       PT64_ROOT_LEVEL, metaphysical,
1283                                       ACC_ALL, NULL);
1284                 root = __pa(sp->spt);
1285                 ++sp->root_count;
1286                 vcpu->arch.mmu.root_hpa = root;
1287                 return;
1288         }
1289 #endif
1290         metaphysical = !is_paging(vcpu);
1291         if (tdp_enabled)
1292                 metaphysical = 1;
1293         for (i = 0; i < 4; ++i) {
1294                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1295
1296                 ASSERT(!VALID_PAGE(root));
1297                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1298                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1299                                 vcpu->arch.mmu.pae_root[i] = 0;
1300                                 continue;
1301                         }
1302                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1303                 } else if (vcpu->arch.mmu.root_level == 0)
1304                         root_gfn = 0;
1305                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1306                                       PT32_ROOT_LEVEL, metaphysical,
1307                                       ACC_ALL, NULL);
1308                 root = __pa(sp->spt);
1309                 ++sp->root_count;
1310                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1311         }
1312         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1313 }
1314
1315 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1316 {
1317         return vaddr;
1318 }
1319
1320 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1321                                 u32 error_code)
1322 {
1323         gfn_t gfn;
1324         int r;
1325
1326         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1327         r = mmu_topup_memory_caches(vcpu);
1328         if (r)
1329                 return r;
1330
1331         ASSERT(vcpu);
1332         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1333
1334         gfn = gva >> PAGE_SHIFT;
1335
1336         return nonpaging_map(vcpu, gva & PAGE_MASK,
1337                              error_code & PFERR_WRITE_MASK, gfn);
1338 }
1339
1340 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1341                                 u32 error_code)
1342 {
1343         pfn_t pfn;
1344         int r;
1345         int largepage = 0;
1346         gfn_t gfn = gpa >> PAGE_SHIFT;
1347
1348         ASSERT(vcpu);
1349         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1350
1351         r = mmu_topup_memory_caches(vcpu);
1352         if (r)
1353                 return r;
1354
1355         down_read(&current->mm->mmap_sem);
1356         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1357                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1358                 largepage = 1;
1359         }
1360         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1361         up_read(&current->mm->mmap_sem);
1362         if (is_error_pfn(pfn)) {
1363                 kvm_release_pfn_clean(pfn);
1364                 return 1;
1365         }
1366         spin_lock(&vcpu->kvm->mmu_lock);
1367         kvm_mmu_free_some_pages(vcpu);
1368         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1369                          largepage, gfn, pfn, kvm_x86_ops->get_tdp_level());
1370         spin_unlock(&vcpu->kvm->mmu_lock);
1371
1372         return r;
1373 }
1374
1375 static void nonpaging_free(struct kvm_vcpu *vcpu)
1376 {
1377         mmu_free_roots(vcpu);
1378 }
1379
1380 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1381 {
1382         struct kvm_mmu *context = &vcpu->arch.mmu;
1383
1384         context->new_cr3 = nonpaging_new_cr3;
1385         context->page_fault = nonpaging_page_fault;
1386         context->gva_to_gpa = nonpaging_gva_to_gpa;
1387         context->free = nonpaging_free;
1388         context->prefetch_page = nonpaging_prefetch_page;
1389         context->root_level = 0;
1390         context->shadow_root_level = PT32E_ROOT_LEVEL;
1391         context->root_hpa = INVALID_PAGE;
1392         return 0;
1393 }
1394
1395 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1396 {
1397         ++vcpu->stat.tlb_flush;
1398         kvm_x86_ops->tlb_flush(vcpu);
1399 }
1400
1401 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1402 {
1403         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1404         mmu_free_roots(vcpu);
1405 }
1406
1407 static void inject_page_fault(struct kvm_vcpu *vcpu,
1408                               u64 addr,
1409                               u32 err_code)
1410 {
1411         kvm_inject_page_fault(vcpu, addr, err_code);
1412 }
1413
1414 static void paging_free(struct kvm_vcpu *vcpu)
1415 {
1416         nonpaging_free(vcpu);
1417 }
1418
1419 #define PTTYPE 64
1420 #include "paging_tmpl.h"
1421 #undef PTTYPE
1422
1423 #define PTTYPE 32
1424 #include "paging_tmpl.h"
1425 #undef PTTYPE
1426
1427 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1428 {
1429         struct kvm_mmu *context = &vcpu->arch.mmu;
1430
1431         ASSERT(is_pae(vcpu));
1432         context->new_cr3 = paging_new_cr3;
1433         context->page_fault = paging64_page_fault;
1434         context->gva_to_gpa = paging64_gva_to_gpa;
1435         context->prefetch_page = paging64_prefetch_page;
1436         context->free = paging_free;
1437         context->root_level = level;
1438         context->shadow_root_level = level;
1439         context->root_hpa = INVALID_PAGE;
1440         return 0;
1441 }
1442
1443 static int paging64_init_context(struct kvm_vcpu *vcpu)
1444 {
1445         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1446 }
1447
1448 static int paging32_init_context(struct kvm_vcpu *vcpu)
1449 {
1450         struct kvm_mmu *context = &vcpu->arch.mmu;
1451
1452         context->new_cr3 = paging_new_cr3;
1453         context->page_fault = paging32_page_fault;
1454         context->gva_to_gpa = paging32_gva_to_gpa;
1455         context->free = paging_free;
1456         context->prefetch_page = paging32_prefetch_page;
1457         context->root_level = PT32_ROOT_LEVEL;
1458         context->shadow_root_level = PT32E_ROOT_LEVEL;
1459         context->root_hpa = INVALID_PAGE;
1460         return 0;
1461 }
1462
1463 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1464 {
1465         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1466 }
1467
1468 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1469 {
1470         struct kvm_mmu *context = &vcpu->arch.mmu;
1471
1472         context->new_cr3 = nonpaging_new_cr3;
1473         context->page_fault = tdp_page_fault;
1474         context->free = nonpaging_free;
1475         context->prefetch_page = nonpaging_prefetch_page;
1476         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
1477         context->root_hpa = INVALID_PAGE;
1478
1479         if (!is_paging(vcpu)) {
1480                 context->gva_to_gpa = nonpaging_gva_to_gpa;
1481                 context->root_level = 0;
1482         } else if (is_long_mode(vcpu)) {
1483                 context->gva_to_gpa = paging64_gva_to_gpa;
1484                 context->root_level = PT64_ROOT_LEVEL;
1485         } else if (is_pae(vcpu)) {
1486                 context->gva_to_gpa = paging64_gva_to_gpa;
1487                 context->root_level = PT32E_ROOT_LEVEL;
1488         } else {
1489                 context->gva_to_gpa = paging32_gva_to_gpa;
1490                 context->root_level = PT32_ROOT_LEVEL;
1491         }
1492
1493         return 0;
1494 }
1495
1496 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1497 {
1498         ASSERT(vcpu);
1499         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1500
1501         if (!is_paging(vcpu))
1502                 return nonpaging_init_context(vcpu);
1503         else if (is_long_mode(vcpu))
1504                 return paging64_init_context(vcpu);
1505         else if (is_pae(vcpu))
1506                 return paging32E_init_context(vcpu);
1507         else
1508                 return paging32_init_context(vcpu);
1509 }
1510
1511 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1512 {
1513         vcpu->arch.update_pte.pfn = bad_pfn;
1514
1515         if (tdp_enabled)
1516                 return init_kvm_tdp_mmu(vcpu);
1517         else
1518                 return init_kvm_softmmu(vcpu);
1519 }
1520
1521 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1522 {
1523         ASSERT(vcpu);
1524         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1525                 vcpu->arch.mmu.free(vcpu);
1526                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1527         }
1528 }
1529
1530 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1531 {
1532         destroy_kvm_mmu(vcpu);
1533         return init_kvm_mmu(vcpu);
1534 }
1535 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1536
1537 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1538 {
1539         int r;
1540
1541         r = mmu_topup_memory_caches(vcpu);
1542         if (r)
1543                 goto out;
1544         spin_lock(&vcpu->kvm->mmu_lock);
1545         kvm_mmu_free_some_pages(vcpu);
1546         mmu_alloc_roots(vcpu);
1547         spin_unlock(&vcpu->kvm->mmu_lock);
1548         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1549         kvm_mmu_flush_tlb(vcpu);
1550 out:
1551         return r;
1552 }
1553 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1554
1555 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1556 {
1557         mmu_free_roots(vcpu);
1558 }
1559
1560 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1561                                   struct kvm_mmu_page *sp,
1562                                   u64 *spte)
1563 {
1564         u64 pte;
1565         struct kvm_mmu_page *child;
1566
1567         pte = *spte;
1568         if (is_shadow_present_pte(pte)) {
1569                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1570                     is_large_pte(pte))
1571                         rmap_remove(vcpu->kvm, spte);
1572                 else {
1573                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1574                         mmu_page_remove_parent_pte(child, spte);
1575                 }
1576         }
1577         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1578         if (is_large_pte(pte))
1579                 --vcpu->kvm->stat.lpages;
1580 }
1581
1582 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1583                                   struct kvm_mmu_page *sp,
1584                                   u64 *spte,
1585                                   const void *new)
1586 {
1587         if ((sp->role.level != PT_PAGE_TABLE_LEVEL)
1588             && !vcpu->arch.update_pte.largepage) {
1589                 ++vcpu->kvm->stat.mmu_pde_zapped;
1590                 return;
1591         }
1592
1593         ++vcpu->kvm->stat.mmu_pte_updated;
1594         if (sp->role.glevels == PT32_ROOT_LEVEL)
1595                 paging32_update_pte(vcpu, sp, spte, new);
1596         else
1597                 paging64_update_pte(vcpu, sp, spte, new);
1598 }
1599
1600 static bool need_remote_flush(u64 old, u64 new)
1601 {
1602         if (!is_shadow_present_pte(old))
1603                 return false;
1604         if (!is_shadow_present_pte(new))
1605                 return true;
1606         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1607                 return true;
1608         old ^= PT64_NX_MASK;
1609         new ^= PT64_NX_MASK;
1610         return (old & ~new & PT64_PERM_MASK) != 0;
1611 }
1612
1613 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1614 {
1615         if (need_remote_flush(old, new))
1616                 kvm_flush_remote_tlbs(vcpu->kvm);
1617         else
1618                 kvm_mmu_flush_tlb(vcpu);
1619 }
1620
1621 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1622 {
1623         u64 *spte = vcpu->arch.last_pte_updated;
1624
1625         return !!(spte && (*spte & shadow_accessed_mask));
1626 }
1627
1628 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1629                                           const u8 *new, int bytes)
1630 {
1631         gfn_t gfn;
1632         int r;
1633         u64 gpte = 0;
1634         pfn_t pfn;
1635
1636         vcpu->arch.update_pte.largepage = 0;
1637
1638         if (bytes != 4 && bytes != 8)
1639                 return;
1640
1641         /*
1642          * Assume that the pte write on a page table of the same type
1643          * as the current vcpu paging mode.  This is nearly always true
1644          * (might be false while changing modes).  Note it is verified later
1645          * by update_pte().
1646          */
1647         if (is_pae(vcpu)) {
1648                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1649                 if ((bytes == 4) && (gpa % 4 == 0)) {
1650                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1651                         if (r)
1652                                 return;
1653                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1654                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1655                         memcpy((void *)&gpte, new, 8);
1656                 }
1657         } else {
1658                 if ((bytes == 4) && (gpa % 4 == 0))
1659                         memcpy((void *)&gpte, new, 4);
1660         }
1661         if (!is_present_pte(gpte))
1662                 return;
1663         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1664
1665         down_read(&current->mm->mmap_sem);
1666         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1667                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1668                 vcpu->arch.update_pte.largepage = 1;
1669         }
1670         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1671         up_read(&current->mm->mmap_sem);
1672
1673         if (is_error_pfn(pfn)) {
1674                 kvm_release_pfn_clean(pfn);
1675                 return;
1676         }
1677         vcpu->arch.update_pte.gfn = gfn;
1678         vcpu->arch.update_pte.pfn = pfn;
1679 }
1680
1681 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1682                        const u8 *new, int bytes)
1683 {
1684         gfn_t gfn = gpa >> PAGE_SHIFT;
1685         struct kvm_mmu_page *sp;
1686         struct hlist_node *node, *n;
1687         struct hlist_head *bucket;
1688         unsigned index;
1689         u64 entry, gentry;
1690         u64 *spte;
1691         unsigned offset = offset_in_page(gpa);
1692         unsigned pte_size;
1693         unsigned page_offset;
1694         unsigned misaligned;
1695         unsigned quadrant;
1696         int level;
1697         int flooded = 0;
1698         int npte;
1699         int r;
1700
1701         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1702         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1703         spin_lock(&vcpu->kvm->mmu_lock);
1704         kvm_mmu_free_some_pages(vcpu);
1705         ++vcpu->kvm->stat.mmu_pte_write;
1706         kvm_mmu_audit(vcpu, "pre pte write");
1707         if (gfn == vcpu->arch.last_pt_write_gfn
1708             && !last_updated_pte_accessed(vcpu)) {
1709                 ++vcpu->arch.last_pt_write_count;
1710                 if (vcpu->arch.last_pt_write_count >= 3)
1711                         flooded = 1;
1712         } else {
1713                 vcpu->arch.last_pt_write_gfn = gfn;
1714                 vcpu->arch.last_pt_write_count = 1;
1715                 vcpu->arch.last_pte_updated = NULL;
1716         }
1717         index = kvm_page_table_hashfn(gfn);
1718         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1719         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1720                 if (sp->gfn != gfn || sp->role.metaphysical)
1721                         continue;
1722                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1723                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1724                 misaligned |= bytes < 4;
1725                 if (misaligned || flooded) {
1726                         /*
1727                          * Misaligned accesses are too much trouble to fix
1728                          * up; also, they usually indicate a page is not used
1729                          * as a page table.
1730                          *
1731                          * If we're seeing too many writes to a page,
1732                          * it may no longer be a page table, or we may be
1733                          * forking, in which case it is better to unmap the
1734                          * page.
1735                          */
1736                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1737                                  gpa, bytes, sp->role.word);
1738                         kvm_mmu_zap_page(vcpu->kvm, sp);
1739                         ++vcpu->kvm->stat.mmu_flooded;
1740                         continue;
1741                 }
1742                 page_offset = offset;
1743                 level = sp->role.level;
1744                 npte = 1;
1745                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1746                         page_offset <<= 1;      /* 32->64 */
1747                         /*
1748                          * A 32-bit pde maps 4MB while the shadow pdes map
1749                          * only 2MB.  So we need to double the offset again
1750                          * and zap two pdes instead of one.
1751                          */
1752                         if (level == PT32_ROOT_LEVEL) {
1753                                 page_offset &= ~7; /* kill rounding error */
1754                                 page_offset <<= 1;
1755                                 npte = 2;
1756                         }
1757                         quadrant = page_offset >> PAGE_SHIFT;
1758                         page_offset &= ~PAGE_MASK;
1759                         if (quadrant != sp->role.quadrant)
1760                                 continue;
1761                 }
1762                 spte = &sp->spt[page_offset / sizeof(*spte)];
1763                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1764                         gentry = 0;
1765                         r = kvm_read_guest_atomic(vcpu->kvm,
1766                                                   gpa & ~(u64)(pte_size - 1),
1767                                                   &gentry, pte_size);
1768                         new = (const void *)&gentry;
1769                         if (r < 0)
1770                                 new = NULL;
1771                 }
1772                 while (npte--) {
1773                         entry = *spte;
1774                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1775                         if (new)
1776                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1777                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1778                         ++spte;
1779                 }
1780         }
1781         kvm_mmu_audit(vcpu, "post pte write");
1782         spin_unlock(&vcpu->kvm->mmu_lock);
1783         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
1784                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
1785                 vcpu->arch.update_pte.pfn = bad_pfn;
1786         }
1787 }
1788
1789 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1790 {
1791         gpa_t gpa;
1792         int r;
1793
1794         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1795
1796         spin_lock(&vcpu->kvm->mmu_lock);
1797         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1798         spin_unlock(&vcpu->kvm->mmu_lock);
1799         return r;
1800 }
1801
1802 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1803 {
1804         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1805                 struct kvm_mmu_page *sp;
1806
1807                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1808                                   struct kvm_mmu_page, link);
1809                 kvm_mmu_zap_page(vcpu->kvm, sp);
1810                 ++vcpu->kvm->stat.mmu_recycled;
1811         }
1812 }
1813
1814 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1815 {
1816         int r;
1817         enum emulation_result er;
1818
1819         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1820         if (r < 0)
1821                 goto out;
1822
1823         if (!r) {
1824                 r = 1;
1825                 goto out;
1826         }
1827
1828         r = mmu_topup_memory_caches(vcpu);
1829         if (r)
1830                 goto out;
1831
1832         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1833
1834         switch (er) {
1835         case EMULATE_DONE:
1836                 return 1;
1837         case EMULATE_DO_MMIO:
1838                 ++vcpu->stat.mmio_exits;
1839                 return 0;
1840         case EMULATE_FAIL:
1841                 kvm_report_emulation_failure(vcpu, "pagetable");
1842                 return 1;
1843         default:
1844                 BUG();
1845         }
1846 out:
1847         return r;
1848 }
1849 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1850
1851 void kvm_enable_tdp(void)
1852 {
1853         tdp_enabled = true;
1854 }
1855 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1856
1857 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1858 {
1859         struct kvm_mmu_page *sp;
1860
1861         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1862                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1863                                   struct kvm_mmu_page, link);
1864                 kvm_mmu_zap_page(vcpu->kvm, sp);
1865         }
1866         free_page((unsigned long)vcpu->arch.mmu.pae_root);
1867 }
1868
1869 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1870 {
1871         struct page *page;
1872         int i;
1873
1874         ASSERT(vcpu);
1875
1876         if (vcpu->kvm->arch.n_requested_mmu_pages)
1877                 vcpu->kvm->arch.n_free_mmu_pages =
1878                                         vcpu->kvm->arch.n_requested_mmu_pages;
1879         else
1880                 vcpu->kvm->arch.n_free_mmu_pages =
1881                                         vcpu->kvm->arch.n_alloc_mmu_pages;
1882         /*
1883          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1884          * Therefore we need to allocate shadow page tables in the first
1885          * 4GB of memory, which happens to fit the DMA32 zone.
1886          */
1887         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1888         if (!page)
1889                 goto error_1;
1890         vcpu->arch.mmu.pae_root = page_address(page);
1891         for (i = 0; i < 4; ++i)
1892                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1893
1894         return 0;
1895
1896 error_1:
1897         free_mmu_pages(vcpu);
1898         return -ENOMEM;
1899 }
1900
1901 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1902 {
1903         ASSERT(vcpu);
1904         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1905
1906         return alloc_mmu_pages(vcpu);
1907 }
1908
1909 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1910 {
1911         ASSERT(vcpu);
1912         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1913
1914         return init_kvm_mmu(vcpu);
1915 }
1916
1917 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1918 {
1919         ASSERT(vcpu);
1920
1921         destroy_kvm_mmu(vcpu);
1922         free_mmu_pages(vcpu);
1923         mmu_free_memory_caches(vcpu);
1924 }
1925
1926 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1927 {
1928         struct kvm_mmu_page *sp;
1929
1930         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1931                 int i;
1932                 u64 *pt;
1933
1934                 if (!test_bit(slot, &sp->slot_bitmap))
1935                         continue;
1936
1937                 pt = sp->spt;
1938                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1939                         /* avoid RMW */
1940                         if (pt[i] & PT_WRITABLE_MASK)
1941                                 pt[i] &= ~PT_WRITABLE_MASK;
1942         }
1943 }
1944
1945 void kvm_mmu_zap_all(struct kvm *kvm)
1946 {
1947         struct kvm_mmu_page *sp, *node;
1948
1949         spin_lock(&kvm->mmu_lock);
1950         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1951                 kvm_mmu_zap_page(kvm, sp);
1952         spin_unlock(&kvm->mmu_lock);
1953
1954         kvm_flush_remote_tlbs(kvm);
1955 }
1956
1957 void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
1958 {
1959         struct kvm_mmu_page *page;
1960
1961         page = container_of(kvm->arch.active_mmu_pages.prev,
1962                             struct kvm_mmu_page, link);
1963         kvm_mmu_zap_page(kvm, page);
1964 }
1965
1966 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
1967 {
1968         struct kvm *kvm;
1969         struct kvm *kvm_freed = NULL;
1970         int cache_count = 0;
1971
1972         spin_lock(&kvm_lock);
1973
1974         list_for_each_entry(kvm, &vm_list, vm_list) {
1975                 int npages;
1976
1977                 spin_lock(&kvm->mmu_lock);
1978                 npages = kvm->arch.n_alloc_mmu_pages -
1979                          kvm->arch.n_free_mmu_pages;
1980                 cache_count += npages;
1981                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
1982                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
1983                         cache_count--;
1984                         kvm_freed = kvm;
1985                 }
1986                 nr_to_scan--;
1987
1988                 spin_unlock(&kvm->mmu_lock);
1989         }
1990         if (kvm_freed)
1991                 list_move_tail(&kvm_freed->vm_list, &vm_list);
1992
1993         spin_unlock(&kvm_lock);
1994
1995         return cache_count;
1996 }
1997
1998 static struct shrinker mmu_shrinker = {
1999         .shrink = mmu_shrink,
2000         .seeks = DEFAULT_SEEKS * 10,
2001 };
2002
2003 void mmu_destroy_caches(void)
2004 {
2005         if (pte_chain_cache)
2006                 kmem_cache_destroy(pte_chain_cache);
2007         if (rmap_desc_cache)
2008                 kmem_cache_destroy(rmap_desc_cache);
2009         if (mmu_page_header_cache)
2010                 kmem_cache_destroy(mmu_page_header_cache);
2011 }
2012
2013 void kvm_mmu_module_exit(void)
2014 {
2015         mmu_destroy_caches();
2016         unregister_shrinker(&mmu_shrinker);
2017 }
2018
2019 int kvm_mmu_module_init(void)
2020 {
2021         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2022                                             sizeof(struct kvm_pte_chain),
2023                                             0, 0, NULL);
2024         if (!pte_chain_cache)
2025                 goto nomem;
2026         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2027                                             sizeof(struct kvm_rmap_desc),
2028                                             0, 0, NULL);
2029         if (!rmap_desc_cache)
2030                 goto nomem;
2031
2032         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2033                                                   sizeof(struct kvm_mmu_page),
2034                                                   0, 0, NULL);
2035         if (!mmu_page_header_cache)
2036                 goto nomem;
2037
2038         register_shrinker(&mmu_shrinker);
2039
2040         return 0;
2041
2042 nomem:
2043         mmu_destroy_caches();
2044         return -ENOMEM;
2045 }
2046
2047 /*
2048  * Caculate mmu pages needed for kvm.
2049  */
2050 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2051 {
2052         int i;
2053         unsigned int nr_mmu_pages;
2054         unsigned int  nr_pages = 0;
2055
2056         for (i = 0; i < kvm->nmemslots; i++)
2057                 nr_pages += kvm->memslots[i].npages;
2058
2059         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2060         nr_mmu_pages = max(nr_mmu_pages,
2061                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2062
2063         return nr_mmu_pages;
2064 }
2065
2066 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2067                                 unsigned len)
2068 {
2069         if (len > buffer->len)
2070                 return NULL;
2071         return buffer->ptr;
2072 }
2073
2074 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2075                                 unsigned len)
2076 {
2077         void *ret;
2078
2079         ret = pv_mmu_peek_buffer(buffer, len);
2080         if (!ret)
2081                 return ret;
2082         buffer->ptr += len;
2083         buffer->len -= len;
2084         buffer->processed += len;
2085         return ret;
2086 }
2087
2088 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2089                              gpa_t addr, gpa_t value)
2090 {
2091         int bytes = 8;
2092         int r;
2093
2094         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2095                 bytes = 4;
2096
2097         r = mmu_topup_memory_caches(vcpu);
2098         if (r)
2099                 return r;
2100
2101         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2102                 return -EFAULT;
2103
2104         return 1;
2105 }
2106
2107 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2108 {
2109         kvm_x86_ops->tlb_flush(vcpu);
2110         return 1;
2111 }
2112
2113 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2114 {
2115         spin_lock(&vcpu->kvm->mmu_lock);
2116         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2117         spin_unlock(&vcpu->kvm->mmu_lock);
2118         return 1;
2119 }
2120
2121 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2122                              struct kvm_pv_mmu_op_buffer *buffer)
2123 {
2124         struct kvm_mmu_op_header *header;
2125
2126         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2127         if (!header)
2128                 return 0;
2129         switch (header->op) {
2130         case KVM_MMU_OP_WRITE_PTE: {
2131                 struct kvm_mmu_op_write_pte *wpte;
2132
2133                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2134                 if (!wpte)
2135                         return 0;
2136                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2137                                         wpte->pte_val);
2138         }
2139         case KVM_MMU_OP_FLUSH_TLB: {
2140                 struct kvm_mmu_op_flush_tlb *ftlb;
2141
2142                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2143                 if (!ftlb)
2144                         return 0;
2145                 return kvm_pv_mmu_flush_tlb(vcpu);
2146         }
2147         case KVM_MMU_OP_RELEASE_PT: {
2148                 struct kvm_mmu_op_release_pt *rpt;
2149
2150                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2151                 if (!rpt)
2152                         return 0;
2153                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2154         }
2155         default: return 0;
2156         }
2157 }
2158
2159 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2160                   gpa_t addr, unsigned long *ret)
2161 {
2162         int r;
2163         struct kvm_pv_mmu_op_buffer buffer;
2164
2165         buffer.ptr = buffer.buf;
2166         buffer.len = min_t(unsigned long, bytes, sizeof buffer.buf);
2167         buffer.processed = 0;
2168
2169         r = kvm_read_guest(vcpu->kvm, addr, buffer.buf, buffer.len);
2170         if (r)
2171                 goto out;
2172
2173         while (buffer.len) {
2174                 r = kvm_pv_mmu_op_one(vcpu, &buffer);
2175                 if (r < 0)
2176                         goto out;
2177                 if (r == 0)
2178                         break;
2179         }
2180
2181         r = 1;
2182 out:
2183         *ret = buffer.processed;
2184         return r;
2185 }
2186
2187 #ifdef AUDIT
2188
2189 static const char *audit_msg;
2190
2191 static gva_t canonicalize(gva_t gva)
2192 {
2193 #ifdef CONFIG_X86_64
2194         gva = (long long)(gva << 16) >> 16;
2195 #endif
2196         return gva;
2197 }
2198
2199 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2200                                 gva_t va, int level)
2201 {
2202         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2203         int i;
2204         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2205
2206         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2207                 u64 ent = pt[i];
2208
2209                 if (ent == shadow_trap_nonpresent_pte)
2210                         continue;
2211
2212                 va = canonicalize(va);
2213                 if (level > 1) {
2214                         if (ent == shadow_notrap_nonpresent_pte)
2215                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2216                                        " in nonleaf level: levels %d gva %lx"
2217                                        " level %d pte %llx\n", audit_msg,
2218                                        vcpu->arch.mmu.root_level, va, level, ent);
2219
2220                         audit_mappings_page(vcpu, ent, va, level - 1);
2221                 } else {
2222                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2223                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2224
2225                         if (is_shadow_present_pte(ent)
2226                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2227                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2228                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2229                                        audit_msg, vcpu->arch.mmu.root_level,
2230                                        va, gpa, hpa, ent,
2231                                        is_shadow_present_pte(ent));
2232                         else if (ent == shadow_notrap_nonpresent_pte
2233                                  && !is_error_hpa(hpa))
2234                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2235                                        " valid guest gva %lx\n", audit_msg, va);
2236                         kvm_release_pfn_clean(pfn);
2237
2238                 }
2239         }
2240 }
2241
2242 static void audit_mappings(struct kvm_vcpu *vcpu)
2243 {
2244         unsigned i;
2245
2246         if (vcpu->arch.mmu.root_level == 4)
2247                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2248         else
2249                 for (i = 0; i < 4; ++i)
2250                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2251                                 audit_mappings_page(vcpu,
2252                                                     vcpu->arch.mmu.pae_root[i],
2253                                                     i << 30,
2254                                                     2);
2255 }
2256
2257 static int count_rmaps(struct kvm_vcpu *vcpu)
2258 {
2259         int nmaps = 0;
2260         int i, j, k;
2261
2262         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2263                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2264                 struct kvm_rmap_desc *d;
2265
2266                 for (j = 0; j < m->npages; ++j) {
2267                         unsigned long *rmapp = &m->rmap[j];
2268
2269                         if (!*rmapp)
2270                                 continue;
2271                         if (!(*rmapp & 1)) {
2272                                 ++nmaps;
2273                                 continue;
2274                         }
2275                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2276                         while (d) {
2277                                 for (k = 0; k < RMAP_EXT; ++k)
2278                                         if (d->shadow_ptes[k])
2279                                                 ++nmaps;
2280                                         else
2281                                                 break;
2282                                 d = d->more;
2283                         }
2284                 }
2285         }
2286         return nmaps;
2287 }
2288
2289 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2290 {
2291         int nmaps = 0;
2292         struct kvm_mmu_page *sp;
2293         int i;
2294
2295         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2296                 u64 *pt = sp->spt;
2297
2298                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2299                         continue;
2300
2301                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2302                         u64 ent = pt[i];
2303
2304                         if (!(ent & PT_PRESENT_MASK))
2305                                 continue;
2306                         if (!(ent & PT_WRITABLE_MASK))
2307                                 continue;
2308                         ++nmaps;
2309                 }
2310         }
2311         return nmaps;
2312 }
2313
2314 static void audit_rmap(struct kvm_vcpu *vcpu)
2315 {
2316         int n_rmap = count_rmaps(vcpu);
2317         int n_actual = count_writable_mappings(vcpu);
2318
2319         if (n_rmap != n_actual)
2320                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2321                        __func__, audit_msg, n_rmap, n_actual);
2322 }
2323
2324 static void audit_write_protection(struct kvm_vcpu *vcpu)
2325 {
2326         struct kvm_mmu_page *sp;
2327         struct kvm_memory_slot *slot;
2328         unsigned long *rmapp;
2329         gfn_t gfn;
2330
2331         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2332                 if (sp->role.metaphysical)
2333                         continue;
2334
2335                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2336                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2337                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2338                 if (*rmapp)
2339                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2340                                " mappings: gfn %lx role %x\n",
2341                                __func__, audit_msg, sp->gfn,
2342                                sp->role.word);
2343         }
2344 }
2345
2346 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2347 {
2348         int olddbg = dbg;
2349
2350         dbg = 0;
2351         audit_msg = msg;
2352         audit_rmap(vcpu);
2353         audit_write_protection(vcpu);
2354         audit_mappings(vcpu);
2355         dbg = olddbg;
2356 }
2357
2358 #endif