arch/x86/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "vmx.h"
  21 #include "mmu.h"
  22
  23 #include <linux/kvm_host.h>
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29 #include <linux/swap.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/compiler.h>
  32
  33 #include <asm/page.h>
  34 #include <asm/cmpxchg.h>
  35 #include <asm/io.h>
  36
  37 /*
  38  * When setting this variable to true it enables Two-Dimensional-Paging
  39  * where the hardware walks 2 page tables:
  40  * 1. the guest-virtual to guest-physical
  41  * 2. while doing 1. it walks guest-physical to host-physical
  42  * If the hardware supports that we don't need to do shadow paging.
  43  */
  44 bool tdp_enabled = false;
  45
  46 #undef MMU_DEBUG
  47
  48 #undef AUDIT
  49
  50 #ifdef AUDIT
  51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  52 #else
  53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  54 #endif
  55
  56 #ifdef MMU_DEBUG
  57
  58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  60
  61 #else
  62
  63 #define pgprintk(x...) do { } while (0)
  64 #define rmap_printk(x...) do { } while (0)
  65
  66 #endif
  67
  68 #if defined(MMU_DEBUG) || defined(AUDIT)
  69 static int dbg = 0;
  70 module_param(dbg, bool, 0644);
  71 #endif
  72
  73 #ifndef MMU_DEBUG
  74 #define ASSERT(x) do { } while (0)
  75 #else
  76 #define ASSERT(x)                                                       \
  77         if (!(x)) {                                                     \
  78                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  79                        __FILE__, __LINE__, #x);                         \
  80         }
  81 #endif
  82
  83 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  84 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  85
  86 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  87
  88 #define PT64_LEVEL_BITS 9
  89
  90 #define PT64_LEVEL_SHIFT(level) \
  91                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  92
  93 #define PT64_LEVEL_MASK(level) \
  94                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  95
  96 #define PT64_INDEX(address, level)\
  97         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
  98
  99
 100 #define PT32_LEVEL_BITS 10
 101
 102 #define PT32_LEVEL_SHIFT(level) \
 103                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 104
 105 #define PT32_LEVEL_MASK(level) \
 106                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 107
 108 #define PT32_INDEX(address, level)\
 109         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 110
 111
 112 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 113 #define PT64_DIR_BASE_ADDR_MASK \
 114         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 115
 116 #define PT32_BASE_ADDR_MASK PAGE_MASK
 117 #define PT32_DIR_BASE_ADDR_MASK \
 118         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 119
 120 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 121                         | PT64_NX_MASK)
 122
 123 #define PFERR_PRESENT_MASK (1U << 0)
 124 #define PFERR_WRITE_MASK (1U << 1)
 125 #define PFERR_USER_MASK (1U << 2)
 126 #define PFERR_FETCH_MASK (1U << 4)
 127
 128 #define PT_DIRECTORY_LEVEL 2
 129 #define PT_PAGE_TABLE_LEVEL 1
 130
 131 #define RMAP_EXT 4
 132
 133 #define ACC_EXEC_MASK    1
 134 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 135 #define ACC_USER_MASK    PT_USER_MASK
 136 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 137
 138 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 139
 140 struct kvm_rmap_desc {
 141         u64 *shadow_ptes[RMAP_EXT];
 142         struct kvm_rmap_desc *more;
 143 };
 144
 145 struct kvm_shadow_walk {
 146         int (*entry)(struct kvm_shadow_walk *walk, struct kvm_vcpu *vcpu,
 147                      u64 addr, u64 *spte, int level);
 148 };
 149
 150 struct kvm_unsync_walk {
 151         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 152 };
 153
 154 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 155
 156 static struct kmem_cache *pte_chain_cache;
 157 static struct kmem_cache *rmap_desc_cache;
 158 static struct kmem_cache *mmu_page_header_cache;
 159
 160 static u64 __read_mostly shadow_trap_nonpresent_pte;
 161 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 162 static u64 __read_mostly shadow_base_present_pte;
 163 static u64 __read_mostly shadow_nx_mask;
 164 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 165 static u64 __read_mostly shadow_user_mask;
 166 static u64 __read_mostly shadow_accessed_mask;
 167 static u64 __read_mostly shadow_dirty_mask;
 168
 169 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 170 {
 171         shadow_trap_nonpresent_pte = trap_pte;
 172         shadow_notrap_nonpresent_pte = notrap_pte;
 173 }
 174 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 175
 176 void kvm_mmu_set_base_ptes(u64 base_pte)
 177 {
 178         shadow_base_present_pte = base_pte;
 179 }
 180 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 181
 182 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 183                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 184 {
 185         shadow_user_mask = user_mask;
 186         shadow_accessed_mask = accessed_mask;
 187         shadow_dirty_mask = dirty_mask;
 188         shadow_nx_mask = nx_mask;
 189         shadow_x_mask = x_mask;
 190 }
 191 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 192
 193 static int is_write_protection(struct kvm_vcpu *vcpu)
 194 {
 195         return vcpu->arch.cr0 & X86_CR0_WP;
 196 }
 197
 198 static int is_cpuid_PSE36(void)
 199 {
 200         return 1;
 201 }
 202
 203 static int is_nx(struct kvm_vcpu *vcpu)
 204 {
 205         return vcpu->arch.shadow_efer & EFER_NX;
 206 }
 207
 208 static int is_present_pte(unsigned long pte)
 209 {
 210         return pte & PT_PRESENT_MASK;
 211 }
 212
 213 static int is_shadow_present_pte(u64 pte)
 214 {
 215         return pte != shadow_trap_nonpresent_pte
 216                 && pte != shadow_notrap_nonpresent_pte;
 217 }
 218
 219 static int is_large_pte(u64 pte)
 220 {
 221         return pte & PT_PAGE_SIZE_MASK;
 222 }
 223
 224 static int is_writeble_pte(unsigned long pte)
 225 {
 226         return pte & PT_WRITABLE_MASK;
 227 }
 228
 229 static int is_dirty_pte(unsigned long pte)
 230 {
 231         return pte & shadow_dirty_mask;
 232 }
 233
 234 static int is_rmap_pte(u64 pte)
 235 {
 236         return is_shadow_present_pte(pte);
 237 }
 238
 239 static pfn_t spte_to_pfn(u64 pte)
 240 {
 241         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 242 }
 243
 244 static gfn_t pse36_gfn_delta(u32 gpte)
 245 {
 246         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 247
 248         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 249 }
 250
 251 static void set_shadow_pte(u64 *sptep, u64 spte)
 252 {
 253 #ifdef CONFIG_X86_64
 254         set_64bit((unsigned long *)sptep, spte);
 255 #else
 256         set_64bit((unsigned long long *)sptep, spte);
 257 #endif
 258 }
 259
 260 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 261                                   struct kmem_cache *base_cache, int min)
 262 {
 263         void *obj;
 264
 265         if (cache->nobjs >= min)
 266                 return 0;
 267         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 268                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 269                 if (!obj)
 270                         return -ENOMEM;
 271                 cache->objects[cache->nobjs++] = obj;
 272         }
 273         return 0;
 274 }
 275
 276 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 277 {
 278         while (mc->nobjs)
 279                 kfree(mc->objects[--mc->nobjs]);
 280 }
 281
 282 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 283                                        int min)
 284 {
 285         struct page *page;
 286
 287         if (cache->nobjs >= min)
 288                 return 0;
 289         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 290                 page = alloc_page(GFP_KERNEL);
 291                 if (!page)
 292                         return -ENOMEM;
 293                 set_page_private(page, 0);
 294                 cache->objects[cache->nobjs++] = page_address(page);
 295         }
 296         return 0;
 297 }
 298
 299 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 300 {
 301         while (mc->nobjs)
 302                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 303 }
 304
 305 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 306 {
 307         int r;
 308
 309         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 310                                    pte_chain_cache, 4);
 311         if (r)
 312                 goto out;
 313         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 314                                    rmap_desc_cache, 1);
 315         if (r)
 316                 goto out;
 317         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 318         if (r)
 319                 goto out;
 320         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 321                                    mmu_page_header_cache, 4);
 322 out:
 323         return r;
 324 }
 325
 326 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 327 {
 328         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 329         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 330         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 331         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 332 }
 333
 334 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 335                                     size_t size)
 336 {
 337         void *p;
 338
 339         BUG_ON(!mc->nobjs);
 340         p = mc->objects[--mc->nobjs];
 341         memset(p, 0, size);
 342         return p;
 343 }
 344
 345 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 346 {
 347         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 348                                       sizeof(struct kvm_pte_chain));
 349 }
 350
 351 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 352 {
 353         kfree(pc);
 354 }
 355
 356 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 357 {
 358         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 359                                       sizeof(struct kvm_rmap_desc));
 360 }
 361
 362 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 363 {
 364         kfree(rd);
 365 }
 366
 367 /*
 368  * Return the pointer to the largepage write count for a given
 369  * gfn, handling slots that are not large page aligned.
 370  */
 371 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 372 {
 373         unsigned long idx;
 374
 375         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 376               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 377         return &slot->lpage_info[idx].write_count;
 378 }
 379
 380 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 381 {
 382         int *write_count;
 383
 384         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 385         *write_count += 1;
 386 }
 387
 388 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 389 {
 390         int *write_count;
 391
 392         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 393         *write_count -= 1;
 394         WARN_ON(*write_count < 0);
 395 }
 396
 397 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 398 {
 399         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
 400         int *largepage_idx;
 401
 402         if (slot) {
 403                 largepage_idx = slot_largepage_idx(gfn, slot);
 404                 return *largepage_idx;
 405         }
 406
 407         return 1;
 408 }
 409
 410 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 411 {
 412         struct vm_area_struct *vma;
 413         unsigned long addr;
 414         int ret = 0;
 415
 416         addr = gfn_to_hva(kvm, gfn);
 417         if (kvm_is_error_hva(addr))
 418                 return ret;
 419
 420         down_read(&current->mm->mmap_sem);
 421         vma = find_vma(current->mm, addr);
 422         if (vma && is_vm_hugetlb_page(vma))
 423                 ret = 1;
 424         up_read(&current->mm->mmap_sem);
 425
 426         return ret;
 427 }
 428
 429 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 430 {
 431         struct kvm_memory_slot *slot;
 432
 433         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 434                 return 0;
 435
 436         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 437                 return 0;
 438
 439         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 440         if (slot && slot->dirty_bitmap)
 441                 return 0;
 442
 443         return 1;
 444 }
 445
 446 /*
 447  * Take gfn and return the reverse mapping to it.
 448  * Note: gfn must be unaliased before this function get called
 449  */
 450
 451 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 452 {
 453         struct kvm_memory_slot *slot;
 454         unsigned long idx;
 455
 456         slot = gfn_to_memslot(kvm, gfn);
 457         if (!lpage)
 458                 return &slot->rmap[gfn - slot->base_gfn];
 459
 460         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 461               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 462
 463         return &slot->lpage_info[idx].rmap_pde;
 464 }
 465
 466 /*
 467  * Reverse mapping data structures:
 468  *
 469  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 470  * that points to page_address(page).
 471  *
 472  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 473  * containing more mappings.
 474  */
 475 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 476 {
 477         struct kvm_mmu_page *sp;
 478         struct kvm_rmap_desc *desc;
 479         unsigned long *rmapp;
 480         int i;
 481
 482         if (!is_rmap_pte(*spte))
 483                 return;
 484         gfn = unalias_gfn(vcpu->kvm, gfn);
 485         sp = page_header(__pa(spte));
 486         sp->gfns[spte - sp->spt] = gfn;
 487         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 488         if (!*rmapp) {
 489                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 490                 *rmapp = (unsigned long)spte;
 491         } else if (!(*rmapp & 1)) {
 492                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 493                 desc = mmu_alloc_rmap_desc(vcpu);
 494                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 495                 desc->shadow_ptes[1] = spte;
 496                 *rmapp = (unsigned long)desc | 1;
 497         } else {
 498                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 499                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 500                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 501                         desc = desc->more;
 502                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 503                         desc->more = mmu_alloc_rmap_desc(vcpu);
 504                         desc = desc->more;
 505                 }
 506                 for (i = 0; desc->shadow_ptes[i]; ++i)
 507                         ;
 508                 desc->shadow_ptes[i] = spte;
 509         }
 510 }
 511
 512 static void rmap_desc_remove_entry(unsigned long *rmapp,
 513                                    struct kvm_rmap_desc *desc,
 514                                    int i,
 515                                    struct kvm_rmap_desc *prev_desc)
 516 {
 517         int j;
 518
 519         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 520                 ;
 521         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 522         desc->shadow_ptes[j] = NULL;
 523         if (j != 0)
 524                 return;
 525         if (!prev_desc && !desc->more)
 526                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 527         else
 528                 if (prev_desc)
 529                         prev_desc->more = desc->more;
 530                 else
 531                         *rmapp = (unsigned long)desc->more | 1;
 532         mmu_free_rmap_desc(desc);
 533 }
 534
 535 static void rmap_remove(struct kvm *kvm, u64 *spte)
 536 {
 537         struct kvm_rmap_desc *desc;
 538         struct kvm_rmap_desc *prev_desc;
 539         struct kvm_mmu_page *sp;
 540         pfn_t pfn;
 541         unsigned long *rmapp;
 542         int i;
 543
 544         if (!is_rmap_pte(*spte))
 545                 return;
 546         sp = page_header(__pa(spte));
 547         pfn = spte_to_pfn(*spte);
 548         if (*spte & shadow_accessed_mask)
 549                 kvm_set_pfn_accessed(pfn);
 550         if (is_writeble_pte(*spte))
 551                 kvm_release_pfn_dirty(pfn);
 552         else
 553                 kvm_release_pfn_clean(pfn);
 554         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 555         if (!*rmapp) {
 556                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 557                 BUG();
 558         } else if (!(*rmapp & 1)) {
 559                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 560                 if ((u64 *)*rmapp != spte) {
 561                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 562                                spte, *spte);
 563                         BUG();
 564                 }
 565                 *rmapp = 0;
 566         } else {
 567                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 568                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 569                 prev_desc = NULL;
 570                 while (desc) {
 571                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 572                                 if (desc->shadow_ptes[i] == spte) {
 573                                         rmap_desc_remove_entry(rmapp,
 574                                                                desc, i,
 575                                                                prev_desc);
 576                                         return;
 577                                 }
 578                         prev_desc = desc;
 579                         desc = desc->more;
 580                 }
 581                 BUG();
 582         }
 583 }
 584
 585 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 586 {
 587         struct kvm_rmap_desc *desc;
 588         struct kvm_rmap_desc *prev_desc;
 589         u64 *prev_spte;
 590         int i;
 591
 592         if (!*rmapp)
 593                 return NULL;
 594         else if (!(*rmapp & 1)) {
 595                 if (!spte)
 596                         return (u64 *)*rmapp;
 597                 return NULL;
 598         }
 599         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 600         prev_desc = NULL;
 601         prev_spte = NULL;
 602         while (desc) {
 603                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 604                         if (prev_spte == spte)
 605                                 return desc->shadow_ptes[i];
 606                         prev_spte = desc->shadow_ptes[i];
 607                 }
 608                 desc = desc->more;
 609         }
 610         return NULL;
 611 }
 612
 613 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 614 {
 615         unsigned long *rmapp;
 616         u64 *spte;
 617         int write_protected = 0;
 618
 619         gfn = unalias_gfn(kvm, gfn);
 620         rmapp = gfn_to_rmap(kvm, gfn, 0);
 621
 622         spte = rmap_next(kvm, rmapp, NULL);
 623         while (spte) {
 624                 BUG_ON(!spte);
 625                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 626                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 627                 if (is_writeble_pte(*spte)) {
 628                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 629                         write_protected = 1;
 630                 }
 631                 spte = rmap_next(kvm, rmapp, spte);
 632         }
 633         if (write_protected) {
 634                 pfn_t pfn;
 635
 636                 spte = rmap_next(kvm, rmapp, NULL);
 637                 pfn = spte_to_pfn(*spte);
 638                 kvm_set_pfn_dirty(pfn);
 639         }
 640
 641         /* check for huge page mappings */
 642         rmapp = gfn_to_rmap(kvm, gfn, 1);
 643         spte = rmap_next(kvm, rmapp, NULL);
 644         while (spte) {
 645                 BUG_ON(!spte);
 646                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 647                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 648                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 649                 if (is_writeble_pte(*spte)) {
 650                         rmap_remove(kvm, spte);
 651                         --kvm->stat.lpages;
 652                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 653                         spte = NULL;
 654                         write_protected = 1;
 655                 }
 656                 spte = rmap_next(kvm, rmapp, spte);
 657         }
 658
 659         if (write_protected)
 660                 kvm_flush_remote_tlbs(kvm);
 661 }
 662
 663 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 664 {
 665         u64 *spte;
 666         int need_tlb_flush = 0;
 667
 668         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 669                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 670                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 671                 rmap_remove(kvm, spte);
 672                 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 673                 need_tlb_flush = 1;
 674         }
 675         return need_tlb_flush;
 676 }
 677
 678 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 679                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 680 {
 681         int i;
 682         int retval = 0;
 683
 684         /*
 685          * If mmap_sem isn't taken, we can look the memslots with only
 686          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 687          */
 688         for (i = 0; i < kvm->nmemslots; i++) {
 689                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 690                 unsigned long start = memslot->userspace_addr;
 691                 unsigned long end;
 692
 693                 /* mmu_lock protects userspace_addr */
 694                 if (!start)
 695                         continue;
 696
 697                 end = start + (memslot->npages << PAGE_SHIFT);
 698                 if (hva >= start && hva < end) {
 699                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 700                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 701                         retval |= handler(kvm,
 702                                           &memslot->lpage_info[
 703                                                   gfn_offset /
 704                                                   KVM_PAGES_PER_HPAGE].rmap_pde);
 705                 }
 706         }
 707
 708         return retval;
 709 }
 710
 711 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 712 {
 713         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 714 }
 715
 716 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 717 {
 718         u64 *spte;
 719         int young = 0;
 720
 721         /* always return old for EPT */
 722         if (!shadow_accessed_mask)
 723                 return 0;
 724
 725         spte = rmap_next(kvm, rmapp, NULL);
 726         while (spte) {
 727                 int _young;
 728                 u64 _spte = *spte;
 729                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 730                 _young = _spte & PT_ACCESSED_MASK;
 731                 if (_young) {
 732                         young = 1;
 733                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 734                 }
 735                 spte = rmap_next(kvm, rmapp, spte);
 736         }
 737         return young;
 738 }
 739
 740 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 741 {
 742         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 743 }
 744
 745 #ifdef MMU_DEBUG
 746 static int is_empty_shadow_page(u64 *spt)
 747 {
 748         u64 *pos;
 749         u64 *end;
 750
 751         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 752                 if (is_shadow_present_pte(*pos)) {
 753                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 754                                pos, *pos);
 755                         return 0;
 756                 }
 757         return 1;
 758 }
 759 #endif
 760
 761 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 762 {
 763         ASSERT(is_empty_shadow_page(sp->spt));
 764         list_del(&sp->link);
 765         __free_page(virt_to_page(sp->spt));
 766         __free_page(virt_to_page(sp->gfns));
 767         kfree(sp);
 768         ++kvm->arch.n_free_mmu_pages;
 769 }
 770
 771 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 772 {
 773         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 774 }
 775
 776 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 777                                                u64 *parent_pte)
 778 {
 779         struct kvm_mmu_page *sp;
 780
 781         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 782         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 783         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 784         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 785         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 786         ASSERT(is_empty_shadow_page(sp->spt));
 787         sp->slot_bitmap = 0;
 788         sp->multimapped = 0;
 789         sp->parent_pte = parent_pte;
 790         --vcpu->kvm->arch.n_free_mmu_pages;
 791         return sp;
 792 }
 793
 794 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 795                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 796 {
 797         struct kvm_pte_chain *pte_chain;
 798         struct hlist_node *node;
 799         int i;
 800
 801         if (!parent_pte)
 802                 return;
 803         if (!sp->multimapped) {
 804                 u64 *old = sp->parent_pte;
 805
 806                 if (!old) {
 807                         sp->parent_pte = parent_pte;
 808                         return;
 809                 }
 810                 sp->multimapped = 1;
 811                 pte_chain = mmu_alloc_pte_chain(vcpu);
 812                 INIT_HLIST_HEAD(&sp->parent_ptes);
 813                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 814                 pte_chain->parent_ptes[0] = old;
 815         }
 816         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 817                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 818                         continue;
 819                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 820                         if (!pte_chain->parent_ptes[i]) {
 821                                 pte_chain->parent_ptes[i] = parent_pte;
 822                                 return;
 823                         }
 824         }
 825         pte_chain = mmu_alloc_pte_chain(vcpu);
 826         BUG_ON(!pte_chain);
 827         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 828         pte_chain->parent_ptes[0] = parent_pte;
 829 }
 830
 831 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 832                                        u64 *parent_pte)
 833 {
 834         struct kvm_pte_chain *pte_chain;
 835         struct hlist_node *node;
 836         int i;
 837
 838         if (!sp->multimapped) {
 839                 BUG_ON(sp->parent_pte != parent_pte);
 840                 sp->parent_pte = NULL;
 841                 return;
 842         }
 843         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 844                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 845                         if (!pte_chain->parent_ptes[i])
 846                                 break;
 847                         if (pte_chain->parent_ptes[i] != parent_pte)
 848                                 continue;
 849                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 850                                 && pte_chain->parent_ptes[i + 1]) {
 851                                 pte_chain->parent_ptes[i]
 852                                         = pte_chain->parent_ptes[i + 1];
 853                                 ++i;
 854                         }
 855                         pte_chain->parent_ptes[i] = NULL;
 856                         if (i == 0) {
 857                                 hlist_del(&pte_chain->link);
 858                                 mmu_free_pte_chain(pte_chain);
 859                                 if (hlist_empty(&sp->parent_ptes)) {
 860                                         sp->multimapped = 0;
 861                                         sp->parent_pte = NULL;
 862                                 }
 863                         }
 864                         return;
 865                 }
 866         BUG();
 867 }
 868
 869
 870 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 871                             mmu_parent_walk_fn fn)
 872 {
 873         struct kvm_pte_chain *pte_chain;
 874         struct hlist_node *node;
 875         struct kvm_mmu_page *parent_sp;
 876         int i;
 877
 878         if (!sp->multimapped && sp->parent_pte) {
 879                 parent_sp = page_header(__pa(sp->parent_pte));
 880                 fn(vcpu, parent_sp);
 881                 mmu_parent_walk(vcpu, parent_sp, fn);
 882                 return;
 883         }
 884         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 885                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 886                         if (!pte_chain->parent_ptes[i])
 887                                 break;
 888                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 889                         fn(vcpu, parent_sp);
 890                         mmu_parent_walk(vcpu, parent_sp, fn);
 891                 }
 892 }
 893
 894 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 895 {
 896         unsigned int index;
 897         struct kvm_mmu_page *sp = page_header(__pa(spte));
 898
 899         index = spte - sp->spt;
 900         __set_bit(index, sp->unsync_child_bitmap);
 901         sp->unsync_children = 1;
 902 }
 903
 904 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 905 {
 906         struct kvm_pte_chain *pte_chain;
 907         struct hlist_node *node;
 908         int i;
 909
 910         if (!sp->parent_pte)
 911                 return;
 912
 913         if (!sp->multimapped) {
 914                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 915                 return;
 916         }
 917
 918         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 919                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 920                         if (!pte_chain->parent_ptes[i])
 921                                 break;
 922                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 923                 }
 924 }
 925
 926 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 927 {
 928         sp->unsync_children = 1;
 929         kvm_mmu_update_parents_unsync(sp);
 930         return 1;
 931 }
 932
 933 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 934                                         struct kvm_mmu_page *sp)
 935 {
 936         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 937         kvm_mmu_update_parents_unsync(sp);
 938 }
 939
 940 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 941                                     struct kvm_mmu_page *sp)
 942 {
 943         int i;
 944
 945         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 946                 sp->spt[i] = shadow_trap_nonpresent_pte;
 947 }
 948
 949 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
 950                                struct kvm_mmu_page *sp)
 951 {
 952         return 1;
 953 }
 954
 955 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
 956 {
 957 }
 958
 959 #define for_each_unsync_children(bitmap, idx)           \
 960         for (idx = find_first_bit(bitmap, 512);         \
 961              idx < 512;                                 \
 962              idx = find_next_bit(bitmap, 512, idx+1))
 963
 964 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
 965                            struct kvm_unsync_walk *walker)
 966 {
 967         int i, ret;
 968
 969         if (!sp->unsync_children)
 970                 return 0;
 971
 972         for_each_unsync_children(sp->unsync_child_bitmap, i) {
 973                 u64 ent = sp->spt[i];
 974
 975                 if (is_shadow_present_pte(ent)) {
 976                         struct kvm_mmu_page *child;
 977                         child = page_header(ent & PT64_BASE_ADDR_MASK);
 978
 979                         if (child->unsync_children) {
 980                                 ret = mmu_unsync_walk(child, walker);
 981                                 if (ret)
 982                                         return ret;
 983                                 __clear_bit(i, sp->unsync_child_bitmap);
 984                         }
 985
 986                         if (child->unsync) {
 987                                 ret = walker->entry(child, walker);
 988                                 __clear_bit(i, sp->unsync_child_bitmap);
 989                                 if (ret)
 990                                         return ret;
 991                         }
 992                 }
 993         }
 994
 995         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
 996                 sp->unsync_children = 0;
 997
 998         return 0;
 999 }
1000
1001 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1002 {
1003         unsigned index;
1004         struct hlist_head *bucket;
1005         struct kvm_mmu_page *sp;
1006         struct hlist_node *node;
1007
1008         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1009         index = kvm_page_table_hashfn(gfn);
1010         bucket = &kvm->arch.mmu_page_hash[index];
1011         hlist_for_each_entry(sp, node, bucket, hash_link)
1012                 if (sp->gfn == gfn && !sp->role.metaphysical
1013                     && !sp->role.invalid) {
1014                         pgprintk("%s: found role %x\n",
1015                                  __func__, sp->role.word);
1016                         return sp;
1017                 }
1018         return NULL;
1019 }
1020
1021 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1022 {
1023         WARN_ON(!sp->unsync);
1024         sp->unsync = 0;
1025         --kvm->stat.mmu_unsync;
1026 }
1027
1028 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1029
1030 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1031 {
1032         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1033                 kvm_mmu_zap_page(vcpu->kvm, sp);
1034                 return 1;
1035         }
1036
1037         rmap_write_protect(vcpu->kvm, sp->gfn);
1038         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1039                 kvm_mmu_zap_page(vcpu->kvm, sp);
1040                 return 1;
1041         }
1042
1043         kvm_mmu_flush_tlb(vcpu);
1044         kvm_unlink_unsync_page(vcpu->kvm, sp);
1045         return 0;
1046 }
1047
1048 struct sync_walker {
1049         struct kvm_vcpu *vcpu;
1050         struct kvm_unsync_walk walker;
1051 };
1052
1053 static int mmu_sync_fn(struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk)
1054 {
1055         struct sync_walker *sync_walk = container_of(walk, struct sync_walker,
1056                                                      walker);
1057         struct kvm_vcpu *vcpu = sync_walk->vcpu;
1058
1059         kvm_sync_page(vcpu, sp);
1060         return (need_resched() || spin_needbreak(&vcpu->kvm->mmu_lock));
1061 }
1062
1063 static void mmu_sync_children(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1064 {
1065         struct sync_walker walker = {
1066                 .walker = { .entry = mmu_sync_fn, },
1067                 .vcpu = vcpu,
1068         };
1069
1070         while (mmu_unsync_walk(sp, &walker.walker))
1071                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1072 }
1073
1074 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1075                                              gfn_t gfn,
1076                                              gva_t gaddr,
1077                                              unsigned level,
1078                                              int metaphysical,
1079                                              unsigned access,
1080                                              u64 *parent_pte)
1081 {
1082         union kvm_mmu_page_role role;
1083         unsigned index;
1084         unsigned quadrant;
1085         struct hlist_head *bucket;
1086         struct kvm_mmu_page *sp;
1087         struct hlist_node *node, *tmp;
1088
1089         role.word = 0;
1090         role.glevels = vcpu->arch.mmu.root_level;
1091         role.level = level;
1092         role.metaphysical = metaphysical;
1093         role.access = access;
1094         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1095                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1096                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1097                 role.quadrant = quadrant;
1098         }
1099         pgprintk("%s: looking gfn %lx role %x\n", __func__,
1100                  gfn, role.word);
1101         index = kvm_page_table_hashfn(gfn);
1102         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1103         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1104                 if (sp->gfn == gfn) {
1105                         if (sp->unsync)
1106                                 if (kvm_sync_page(vcpu, sp))
1107                                         continue;
1108
1109                         if (sp->role.word != role.word)
1110                                 continue;
1111
1112                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1113                         if (sp->unsync_children) {
1114                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1115                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1116                         }
1117                         pgprintk("%s: found\n", __func__);
1118                         return sp;
1119                 }
1120         ++vcpu->kvm->stat.mmu_cache_miss;
1121         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1122         if (!sp)
1123                 return sp;
1124         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1125         sp->gfn = gfn;
1126         sp->role = role;
1127         hlist_add_head(&sp->hash_link, bucket);
1128         if (!metaphysical) {
1129                 rmap_write_protect(vcpu->kvm, gfn);
1130                 account_shadowed(vcpu->kvm, gfn);
1131         }
1132         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1133                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1134         else
1135                 nonpaging_prefetch_page(vcpu, sp);
1136         return sp;
1137 }
1138
1139 static int walk_shadow(struct kvm_shadow_walk *walker,
1140                        struct kvm_vcpu *vcpu, u64 addr)
1141 {
1142         hpa_t shadow_addr;
1143         int level;
1144         int r;
1145         u64 *sptep;
1146         unsigned index;
1147
1148         shadow_addr = vcpu->arch.mmu.root_hpa;
1149         level = vcpu->arch.mmu.shadow_root_level;
1150         if (level == PT32E_ROOT_LEVEL) {
1151                 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1152                 shadow_addr &= PT64_BASE_ADDR_MASK;
1153                 --level;
1154         }
1155
1156         while (level >= PT_PAGE_TABLE_LEVEL) {
1157                 index = SHADOW_PT_INDEX(addr, level);
1158                 sptep = ((u64 *)__va(shadow_addr)) + index;
1159                 r = walker->entry(walker, vcpu, addr, sptep, level);
1160                 if (r)
1161                         return r;
1162                 shadow_addr = *sptep & PT64_BASE_ADDR_MASK;
1163                 --level;
1164         }
1165         return 0;
1166 }
1167
1168 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1169                                          struct kvm_mmu_page *sp)
1170 {
1171         unsigned i;
1172         u64 *pt;
1173         u64 ent;
1174
1175         pt = sp->spt;
1176
1177         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1178                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1179                         if (is_shadow_present_pte(pt[i]))
1180                                 rmap_remove(kvm, &pt[i]);
1181                         pt[i] = shadow_trap_nonpresent_pte;
1182                 }
1183                 return;
1184         }
1185
1186         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1187                 ent = pt[i];
1188
1189                 if (is_shadow_present_pte(ent)) {
1190                         if (!is_large_pte(ent)) {
1191                                 ent &= PT64_BASE_ADDR_MASK;
1192                                 mmu_page_remove_parent_pte(page_header(ent),
1193                                                            &pt[i]);
1194                         } else {
1195                                 --kvm->stat.lpages;
1196                                 rmap_remove(kvm, &pt[i]);
1197                         }
1198                 }
1199                 pt[i] = shadow_trap_nonpresent_pte;
1200         }
1201 }
1202
1203 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1204 {
1205         mmu_page_remove_parent_pte(sp, parent_pte);
1206 }
1207
1208 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1209 {
1210         int i;
1211
1212         for (i = 0; i < KVM_MAX_VCPUS; ++i)
1213                 if (kvm->vcpus[i])
1214                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
1215 }
1216
1217 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1218 {
1219         u64 *parent_pte;
1220
1221         while (sp->multimapped || sp->parent_pte) {
1222                 if (!sp->multimapped)
1223                         parent_pte = sp->parent_pte;
1224                 else {
1225                         struct kvm_pte_chain *chain;
1226
1227                         chain = container_of(sp->parent_ptes.first,
1228                                              struct kvm_pte_chain, link);
1229                         parent_pte = chain->parent_ptes[0];
1230                 }
1231                 BUG_ON(!parent_pte);
1232                 kvm_mmu_put_page(sp, parent_pte);
1233                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1234         }
1235 }
1236
1237 struct zap_walker {
1238         struct kvm_unsync_walk walker;
1239         struct kvm *kvm;
1240         int zapped;
1241 };
1242
1243 static int mmu_zap_fn(struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk)
1244 {
1245         struct zap_walker *zap_walk = container_of(walk, struct zap_walker,
1246                                                      walker);
1247         kvm_mmu_zap_page(zap_walk->kvm, sp);
1248         zap_walk->zapped = 1;
1249         return 0;
1250 }
1251
1252 static int mmu_zap_unsync_children(struct kvm *kvm, struct kvm_mmu_page *sp)
1253 {
1254         struct zap_walker walker = {
1255                 .walker = { .entry = mmu_zap_fn, },
1256                 .kvm = kvm,
1257                 .zapped = 0,
1258         };
1259
1260         if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1261                 return 0;
1262         mmu_unsync_walk(sp, &walker.walker);
1263         return walker.zapped;
1264 }
1265
1266 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1267 {
1268         int ret;
1269         ++kvm->stat.mmu_shadow_zapped;
1270         ret = mmu_zap_unsync_children(kvm, sp);
1271         kvm_mmu_page_unlink_children(kvm, sp);
1272         kvm_mmu_unlink_parents(kvm, sp);
1273         kvm_flush_remote_tlbs(kvm);
1274         if (!sp->role.invalid && !sp->role.metaphysical)
1275                 unaccount_shadowed(kvm, sp->gfn);
1276         if (sp->unsync)
1277                 kvm_unlink_unsync_page(kvm, sp);
1278         if (!sp->root_count) {
1279                 hlist_del(&sp->hash_link);
1280                 kvm_mmu_free_page(kvm, sp);
1281         } else {
1282                 sp->role.invalid = 1;
1283                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1284                 kvm_reload_remote_mmus(kvm);
1285         }
1286         kvm_mmu_reset_last_pte_updated(kvm);
1287         return ret;
1288 }
1289
1290 /*
1291  * Changing the number of mmu pages allocated to the vm
1292  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1293  */
1294 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1295 {
1296         /*
1297          * If we set the number of mmu pages to be smaller be than the
1298          * number of actived pages , we must to free some mmu pages before we
1299          * change the value
1300          */
1301
1302         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1303             kvm_nr_mmu_pages) {
1304                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1305                                        - kvm->arch.n_free_mmu_pages;
1306
1307                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1308                         struct kvm_mmu_page *page;
1309
1310                         page = container_of(kvm->arch.active_mmu_pages.prev,
1311                                             struct kvm_mmu_page, link);
1312                         kvm_mmu_zap_page(kvm, page);
1313                         n_used_mmu_pages--;
1314                 }
1315                 kvm->arch.n_free_mmu_pages = 0;
1316         }
1317         else
1318                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1319                                          - kvm->arch.n_alloc_mmu_pages;
1320
1321         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1322 }
1323
1324 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1325 {
1326         unsigned index;
1327         struct hlist_head *bucket;
1328         struct kvm_mmu_page *sp;
1329         struct hlist_node *node, *n;
1330         int r;
1331
1332         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1333         r = 0;
1334         index = kvm_page_table_hashfn(gfn);
1335         bucket = &kvm->arch.mmu_page_hash[index];
1336         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1337                 if (sp->gfn == gfn && !sp->role.metaphysical) {
1338                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1339                                  sp->role.word);
1340                         r = 1;
1341                         if (kvm_mmu_zap_page(kvm, sp))
1342                                 n = bucket->first;
1343                 }
1344         return r;
1345 }
1346
1347 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1348 {
1349         struct kvm_mmu_page *sp;
1350
1351         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1352                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1353                 kvm_mmu_zap_page(kvm, sp);
1354         }
1355 }
1356
1357 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1358 {
1359         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1360         struct kvm_mmu_page *sp = page_header(__pa(pte));
1361
1362         __set_bit(slot, &sp->slot_bitmap);
1363 }
1364
1365 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1366 {
1367         int i;
1368         u64 *pt = sp->spt;
1369
1370         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1371                 return;
1372
1373         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1374                 if (pt[i] == shadow_notrap_nonpresent_pte)
1375                         set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1376         }
1377 }
1378
1379 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1380 {
1381         struct page *page;
1382
1383         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1384
1385         if (gpa == UNMAPPED_GVA)
1386                 return NULL;
1387
1388         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1389
1390         return page;
1391 }
1392
1393 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1394 {
1395         unsigned index;
1396         struct hlist_head *bucket;
1397         struct kvm_mmu_page *s;
1398         struct hlist_node *node, *n;
1399
1400         index = kvm_page_table_hashfn(sp->gfn);
1401         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1402         /* don't unsync if pagetable is shadowed with multiple roles */
1403         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1404                 if (s->gfn != sp->gfn || s->role.metaphysical)
1405                         continue;
1406                 if (s->role.word != sp->role.word)
1407                         return 1;
1408         }
1409         kvm_mmu_mark_parents_unsync(vcpu, sp);
1410         ++vcpu->kvm->stat.mmu_unsync;
1411         sp->unsync = 1;
1412         mmu_convert_notrap(sp);
1413         return 0;
1414 }
1415
1416 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1417                                   bool can_unsync)
1418 {
1419         struct kvm_mmu_page *shadow;
1420
1421         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1422         if (shadow) {
1423                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1424                         return 1;
1425                 if (shadow->unsync)
1426                         return 0;
1427                 if (can_unsync)
1428                         return kvm_unsync_page(vcpu, shadow);
1429                 return 1;
1430         }
1431         return 0;
1432 }
1433
1434 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1435                     unsigned pte_access, int user_fault,
1436                     int write_fault, int dirty, int largepage,
1437                     gfn_t gfn, pfn_t pfn, bool speculative,
1438                     bool can_unsync)
1439 {
1440         u64 spte;
1441         int ret = 0;
1442         /*
1443          * We don't set the accessed bit, since we sometimes want to see
1444          * whether the guest actually used the pte (in order to detect
1445          * demand paging).
1446          */
1447         spte = shadow_base_present_pte | shadow_dirty_mask;
1448         if (!speculative)
1449                 spte |= shadow_accessed_mask;
1450         if (!dirty)
1451                 pte_access &= ~ACC_WRITE_MASK;
1452         if (pte_access & ACC_EXEC_MASK)
1453                 spte |= shadow_x_mask;
1454         else
1455                 spte |= shadow_nx_mask;
1456         if (pte_access & ACC_USER_MASK)
1457                 spte |= shadow_user_mask;
1458         if (largepage)
1459                 spte |= PT_PAGE_SIZE_MASK;
1460
1461         spte |= (u64)pfn << PAGE_SHIFT;
1462
1463         if ((pte_access & ACC_WRITE_MASK)
1464             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1465
1466                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1467                         ret = 1;
1468                         spte = shadow_trap_nonpresent_pte;
1469                         goto set_pte;
1470                 }
1471
1472                 spte |= PT_WRITABLE_MASK;
1473
1474                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1475                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1476                                  __func__, gfn);
1477                         ret = 1;
1478                         pte_access &= ~ACC_WRITE_MASK;
1479                         if (is_writeble_pte(spte))
1480                                 spte &= ~PT_WRITABLE_MASK;
1481                 }
1482         }
1483
1484         if (pte_access & ACC_WRITE_MASK)
1485                 mark_page_dirty(vcpu->kvm, gfn);
1486
1487 set_pte:
1488         set_shadow_pte(shadow_pte, spte);
1489         return ret;
1490 }
1491
1492 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1493                          unsigned pt_access, unsigned pte_access,
1494                          int user_fault, int write_fault, int dirty,
1495                          int *ptwrite, int largepage, gfn_t gfn,
1496                          pfn_t pfn, bool speculative)
1497 {
1498         int was_rmapped = 0;
1499         int was_writeble = is_writeble_pte(*shadow_pte);
1500
1501         pgprintk("%s: spte %llx access %x write_fault %d"
1502                  " user_fault %d gfn %lx\n",
1503                  __func__, *shadow_pte, pt_access,
1504                  write_fault, user_fault, gfn);
1505
1506         if (is_rmap_pte(*shadow_pte)) {
1507                 /*
1508                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1509                  * the parent of the now unreachable PTE.
1510                  */
1511                 if (largepage && !is_large_pte(*shadow_pte)) {
1512                         struct kvm_mmu_page *child;
1513                         u64 pte = *shadow_pte;
1514
1515                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1516                         mmu_page_remove_parent_pte(child, shadow_pte);
1517                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1518                         pgprintk("hfn old %lx new %lx\n",
1519                                  spte_to_pfn(*shadow_pte), pfn);
1520                         rmap_remove(vcpu->kvm, shadow_pte);
1521                 } else {
1522                         if (largepage)
1523                                 was_rmapped = is_large_pte(*shadow_pte);
1524                         else
1525                                 was_rmapped = 1;
1526                 }
1527         }
1528         if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1529                       dirty, largepage, gfn, pfn, speculative, true)) {
1530                 if (write_fault)
1531                         *ptwrite = 1;
1532                 kvm_x86_ops->tlb_flush(vcpu);
1533         }
1534
1535         pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1536         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1537                  is_large_pte(*shadow_pte)? "2MB" : "4kB",
1538                  is_present_pte(*shadow_pte)?"RW":"R", gfn,
1539                  *shadow_pte, shadow_pte);
1540         if (!was_rmapped && is_large_pte(*shadow_pte))
1541                 ++vcpu->kvm->stat.lpages;
1542
1543         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1544         if (!was_rmapped) {
1545                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1546                 if (!is_rmap_pte(*shadow_pte))
1547                         kvm_release_pfn_clean(pfn);
1548         } else {
1549                 if (was_writeble)
1550                         kvm_release_pfn_dirty(pfn);
1551                 else
1552                         kvm_release_pfn_clean(pfn);
1553         }
1554         if (speculative) {
1555                 vcpu->arch.last_pte_updated = shadow_pte;
1556                 vcpu->arch.last_pte_gfn = gfn;
1557         }
1558 }
1559
1560 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1561 {
1562 }
1563
1564 struct direct_shadow_walk {
1565         struct kvm_shadow_walk walker;
1566         pfn_t pfn;
1567         int write;
1568         int largepage;
1569         int pt_write;
1570 };
1571
1572 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1573                             struct kvm_vcpu *vcpu,
1574                             u64 addr, u64 *sptep, int level)
1575 {
1576         struct direct_shadow_walk *walk =
1577                 container_of(_walk, struct direct_shadow_walk, walker);
1578         struct kvm_mmu_page *sp;
1579         gfn_t pseudo_gfn;
1580         gfn_t gfn = addr >> PAGE_SHIFT;
1581
1582         if (level == PT_PAGE_TABLE_LEVEL
1583             || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1584                 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1585                              0, walk->write, 1, &walk->pt_write,
1586                              walk->largepage, gfn, walk->pfn, false);
1587                 ++vcpu->stat.pf_fixed;
1588                 return 1;
1589         }
1590
1591         if (*sptep == shadow_trap_nonpresent_pte) {
1592                 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1593                 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1594                                       1, ACC_ALL, sptep);
1595                 if (!sp) {
1596                         pgprintk("nonpaging_map: ENOMEM\n");
1597                         kvm_release_pfn_clean(walk->pfn);
1598                         return -ENOMEM;
1599                 }
1600
1601                 set_shadow_pte(sptep,
1602                                __pa(sp->spt)
1603                                | PT_PRESENT_MASK | PT_WRITABLE_MASK
1604                                | shadow_user_mask | shadow_x_mask);
1605         }
1606         return 0;
1607 }
1608
1609 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1610                         int largepage, gfn_t gfn, pfn_t pfn)
1611 {
1612         int r;
1613         struct direct_shadow_walk walker = {
1614                 .walker = { .entry = direct_map_entry, },
1615                 .pfn = pfn,
1616                 .largepage = largepage,
1617                 .write = write,
1618                 .pt_write = 0,
1619         };
1620
1621         r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1622         if (r < 0)
1623                 return r;
1624         return walker.pt_write;
1625 }
1626
1627 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1628 {
1629         int r;
1630         int largepage = 0;
1631         pfn_t pfn;
1632         unsigned long mmu_seq;
1633
1634         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1635                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1636                 largepage = 1;
1637         }
1638
1639         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1640         smp_rmb();
1641         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1642
1643         /* mmio */
1644         if (is_error_pfn(pfn)) {
1645                 kvm_release_pfn_clean(pfn);
1646                 return 1;
1647         }
1648
1649         spin_lock(&vcpu->kvm->mmu_lock);
1650         if (mmu_notifier_retry(vcpu, mmu_seq))
1651                 goto out_unlock;
1652         kvm_mmu_free_some_pages(vcpu);
1653         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1654         spin_unlock(&vcpu->kvm->mmu_lock);
1655
1656
1657         return r;
1658
1659 out_unlock:
1660         spin_unlock(&vcpu->kvm->mmu_lock);
1661         kvm_release_pfn_clean(pfn);
1662         return 0;
1663 }
1664
1665
1666 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1667 {
1668         int i;
1669         struct kvm_mmu_page *sp;
1670
1671         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1672                 return;
1673         spin_lock(&vcpu->kvm->mmu_lock);
1674         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1675                 hpa_t root = vcpu->arch.mmu.root_hpa;
1676
1677                 sp = page_header(root);
1678                 --sp->root_count;
1679                 if (!sp->root_count && sp->role.invalid)
1680                         kvm_mmu_zap_page(vcpu->kvm, sp);
1681                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1682                 spin_unlock(&vcpu->kvm->mmu_lock);
1683                 return;
1684         }
1685         for (i = 0; i < 4; ++i) {
1686                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1687
1688                 if (root) {
1689                         root &= PT64_BASE_ADDR_MASK;
1690                         sp = page_header(root);
1691                         --sp->root_count;
1692                         if (!sp->root_count && sp->role.invalid)
1693                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1694                 }
1695                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1696         }
1697         spin_unlock(&vcpu->kvm->mmu_lock);
1698         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1699 }
1700
1701 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1702 {
1703         int i;
1704         gfn_t root_gfn;
1705         struct kvm_mmu_page *sp;
1706         int metaphysical = 0;
1707
1708         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1709
1710         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1711                 hpa_t root = vcpu->arch.mmu.root_hpa;
1712
1713                 ASSERT(!VALID_PAGE(root));
1714                 if (tdp_enabled)
1715                         metaphysical = 1;
1716                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1717                                       PT64_ROOT_LEVEL, metaphysical,
1718                                       ACC_ALL, NULL);
1719                 root = __pa(sp->spt);
1720                 ++sp->root_count;
1721                 vcpu->arch.mmu.root_hpa = root;
1722                 return;
1723         }
1724         metaphysical = !is_paging(vcpu);
1725         if (tdp_enabled)
1726                 metaphysical = 1;
1727         for (i = 0; i < 4; ++i) {
1728                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1729
1730                 ASSERT(!VALID_PAGE(root));
1731                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1732                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1733                                 vcpu->arch.mmu.pae_root[i] = 0;
1734                                 continue;
1735                         }
1736                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1737                 } else if (vcpu->arch.mmu.root_level == 0)
1738                         root_gfn = 0;
1739                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1740                                       PT32_ROOT_LEVEL, metaphysical,
1741                                       ACC_ALL, NULL);
1742                 root = __pa(sp->spt);
1743                 ++sp->root_count;
1744                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1745         }
1746         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1747 }
1748
1749 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
1750 {
1751         int i;
1752         struct kvm_mmu_page *sp;
1753
1754         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1755                 return;
1756         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1757                 hpa_t root = vcpu->arch.mmu.root_hpa;
1758                 sp = page_header(root);
1759                 mmu_sync_children(vcpu, sp);
1760                 return;
1761         }
1762         for (i = 0; i < 4; ++i) {
1763                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1764
1765                 if (root) {
1766                         root &= PT64_BASE_ADDR_MASK;
1767                         sp = page_header(root);
1768                         mmu_sync_children(vcpu, sp);
1769                 }
1770         }
1771 }
1772
1773 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
1774 {
1775         spin_lock(&vcpu->kvm->mmu_lock);
1776         mmu_sync_roots(vcpu);
1777         spin_unlock(&vcpu->kvm->mmu_lock);
1778 }
1779
1780 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1781 {
1782         return vaddr;
1783 }
1784
1785 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1786                                 u32 error_code)
1787 {
1788         gfn_t gfn;
1789         int r;
1790
1791         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1792         r = mmu_topup_memory_caches(vcpu);
1793         if (r)
1794                 return r;
1795
1796         ASSERT(vcpu);
1797         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1798
1799         gfn = gva >> PAGE_SHIFT;
1800
1801         return nonpaging_map(vcpu, gva & PAGE_MASK,
1802                              error_code & PFERR_WRITE_MASK, gfn);
1803 }
1804
1805 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1806                                 u32 error_code)
1807 {
1808         pfn_t pfn;
1809         int r;
1810         int largepage = 0;
1811         gfn_t gfn = gpa >> PAGE_SHIFT;
1812         unsigned long mmu_seq;
1813
1814         ASSERT(vcpu);
1815         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1816
1817         r = mmu_topup_memory_caches(vcpu);
1818         if (r)
1819                 return r;
1820
1821         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1822                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1823                 largepage = 1;
1824         }
1825         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1826         smp_rmb();
1827         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1828         if (is_error_pfn(pfn)) {
1829                 kvm_release_pfn_clean(pfn);
1830                 return 1;
1831         }
1832         spin_lock(&vcpu->kvm->mmu_lock);
1833         if (mmu_notifier_retry(vcpu, mmu_seq))
1834                 goto out_unlock;
1835         kvm_mmu_free_some_pages(vcpu);
1836         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1837                          largepage, gfn, pfn);
1838         spin_unlock(&vcpu->kvm->mmu_lock);
1839
1840         return r;
1841
1842 out_unlock:
1843         spin_unlock(&vcpu->kvm->mmu_lock);
1844         kvm_release_pfn_clean(pfn);
1845         return 0;
1846 }
1847
1848 static void nonpaging_free(struct kvm_vcpu *vcpu)
1849 {
1850         mmu_free_roots(vcpu);
1851 }
1852
1853 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1854 {
1855         struct kvm_mmu *context = &vcpu->arch.mmu;
1856
1857         context->new_cr3 = nonpaging_new_cr3;
1858         context->page_fault = nonpaging_page_fault;
1859         context->gva_to_gpa = nonpaging_gva_to_gpa;
1860         context->free = nonpaging_free;
1861         context->prefetch_page = nonpaging_prefetch_page;
1862         context->sync_page = nonpaging_sync_page;
1863         context->invlpg = nonpaging_invlpg;
1864         context->root_level = 0;
1865         context->shadow_root_level = PT32E_ROOT_LEVEL;
1866         context->root_hpa = INVALID_PAGE;
1867         return 0;
1868 }
1869
1870 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1871 {
1872         ++vcpu->stat.tlb_flush;
1873         kvm_x86_ops->tlb_flush(vcpu);
1874 }
1875
1876 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1877 {
1878         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1879         mmu_free_roots(vcpu);
1880 }
1881
1882 static void inject_page_fault(struct kvm_vcpu *vcpu,
1883                               u64 addr,
1884                               u32 err_code)
1885 {
1886         kvm_inject_page_fault(vcpu, addr, err_code);
1887 }
1888
1889 static void paging_free(struct kvm_vcpu *vcpu)
1890 {
1891         nonpaging_free(vcpu);
1892 }
1893
1894 #define PTTYPE 64
1895 #include "paging_tmpl.h"
1896 #undef PTTYPE
1897
1898 #define PTTYPE 32
1899 #include "paging_tmpl.h"
1900 #undef PTTYPE
1901
1902 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1903 {
1904         struct kvm_mmu *context = &vcpu->arch.mmu;
1905
1906         ASSERT(is_pae(vcpu));
1907         context->new_cr3 = paging_new_cr3;
1908         context->page_fault = paging64_page_fault;
1909         context->gva_to_gpa = paging64_gva_to_gpa;
1910         context->prefetch_page = paging64_prefetch_page;
1911         context->sync_page = paging64_sync_page;
1912         context->invlpg = paging64_invlpg;
1913         context->free = paging_free;
1914         context->root_level = level;
1915         context->shadow_root_level = level;
1916         context->root_hpa = INVALID_PAGE;
1917         return 0;
1918 }
1919
1920 static int paging64_init_context(struct kvm_vcpu *vcpu)
1921 {
1922         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1923 }
1924
1925 static int paging32_init_context(struct kvm_vcpu *vcpu)
1926 {
1927         struct kvm_mmu *context = &vcpu->arch.mmu;
1928
1929         context->new_cr3 = paging_new_cr3;
1930         context->page_fault = paging32_page_fault;
1931         context->gva_to_gpa = paging32_gva_to_gpa;
1932         context->free = paging_free;
1933         context->prefetch_page = paging32_prefetch_page;
1934         context->sync_page = paging32_sync_page;
1935         context->invlpg = paging32_invlpg;
1936         context->root_level = PT32_ROOT_LEVEL;
1937         context->shadow_root_level = PT32E_ROOT_LEVEL;
1938         context->root_hpa = INVALID_PAGE;
1939         return 0;
1940 }
1941
1942 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1943 {
1944         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1945 }
1946
1947 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1948 {
1949         struct kvm_mmu *context = &vcpu->arch.mmu;
1950
1951         context->new_cr3 = nonpaging_new_cr3;
1952         context->page_fault = tdp_page_fault;
1953         context->free = nonpaging_free;
1954         context->prefetch_page = nonpaging_prefetch_page;
1955         context->sync_page = nonpaging_sync_page;
1956         context->invlpg = nonpaging_invlpg;
1957         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
1958         context->root_hpa = INVALID_PAGE;
1959
1960         if (!is_paging(vcpu)) {
1961                 context->gva_to_gpa = nonpaging_gva_to_gpa;
1962                 context->root_level = 0;
1963         } else if (is_long_mode(vcpu)) {
1964                 context->gva_to_gpa = paging64_gva_to_gpa;
1965                 context->root_level = PT64_ROOT_LEVEL;
1966         } else if (is_pae(vcpu)) {
1967                 context->gva_to_gpa = paging64_gva_to_gpa;
1968                 context->root_level = PT32E_ROOT_LEVEL;
1969         } else {
1970                 context->gva_to_gpa = paging32_gva_to_gpa;
1971                 context->root_level = PT32_ROOT_LEVEL;
1972         }
1973
1974         return 0;
1975 }
1976
1977 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1978 {
1979         ASSERT(vcpu);
1980         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1981
1982         if (!is_paging(vcpu))
1983                 return nonpaging_init_context(vcpu);
1984         else if (is_long_mode(vcpu))
1985                 return paging64_init_context(vcpu);
1986         else if (is_pae(vcpu))
1987                 return paging32E_init_context(vcpu);
1988         else
1989                 return paging32_init_context(vcpu);
1990 }
1991
1992 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1993 {
1994         vcpu->arch.update_pte.pfn = bad_pfn;
1995
1996         if (tdp_enabled)
1997                 return init_kvm_tdp_mmu(vcpu);
1998         else
1999                 return init_kvm_softmmu(vcpu);
2000 }
2001
2002 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2003 {
2004         ASSERT(vcpu);
2005         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2006                 vcpu->arch.mmu.free(vcpu);
2007                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2008         }
2009 }
2010
2011 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2012 {
2013         destroy_kvm_mmu(vcpu);
2014         return init_kvm_mmu(vcpu);
2015 }
2016 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2017
2018 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2019 {
2020         int r;
2021
2022         r = mmu_topup_memory_caches(vcpu);
2023         if (r)
2024                 goto out;
2025         spin_lock(&vcpu->kvm->mmu_lock);
2026         kvm_mmu_free_some_pages(vcpu);
2027         mmu_alloc_roots(vcpu);
2028         mmu_sync_roots(vcpu);
2029         spin_unlock(&vcpu->kvm->mmu_lock);
2030         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2031         kvm_mmu_flush_tlb(vcpu);
2032 out:
2033         return r;
2034 }
2035 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2036
2037 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2038 {
2039         mmu_free_roots(vcpu);
2040 }
2041
2042 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2043                                   struct kvm_mmu_page *sp,
2044                                   u64 *spte)
2045 {
2046         u64 pte;
2047         struct kvm_mmu_page *child;
2048
2049         pte = *spte;
2050         if (is_shadow_present_pte(pte)) {
2051                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2052                     is_large_pte(pte))
2053                         rmap_remove(vcpu->kvm, spte);
2054                 else {
2055                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2056                         mmu_page_remove_parent_pte(child, spte);
2057                 }
2058         }
2059         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2060         if (is_large_pte(pte))
2061                 --vcpu->kvm->stat.lpages;
2062 }
2063
2064 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2065                                   struct kvm_mmu_page *sp,
2066                                   u64 *spte,
2067                                   const void *new)
2068 {
2069         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2070                 if (!vcpu->arch.update_pte.largepage ||
2071                     sp->role.glevels == PT32_ROOT_LEVEL) {
2072                         ++vcpu->kvm->stat.mmu_pde_zapped;
2073                         return;
2074                 }
2075         }
2076
2077         ++vcpu->kvm->stat.mmu_pte_updated;
2078         if (sp->role.glevels == PT32_ROOT_LEVEL)
2079                 paging32_update_pte(vcpu, sp, spte, new);
2080         else
2081                 paging64_update_pte(vcpu, sp, spte, new);
2082 }
2083
2084 static bool need_remote_flush(u64 old, u64 new)
2085 {
2086         if (!is_shadow_present_pte(old))
2087                 return false;
2088         if (!is_shadow_present_pte(new))
2089                 return true;
2090         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2091                 return true;
2092         old ^= PT64_NX_MASK;
2093         new ^= PT64_NX_MASK;
2094         return (old & ~new & PT64_PERM_MASK) != 0;
2095 }
2096
2097 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2098 {
2099         if (need_remote_flush(old, new))
2100                 kvm_flush_remote_tlbs(vcpu->kvm);
2101         else
2102                 kvm_mmu_flush_tlb(vcpu);
2103 }
2104
2105 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2106 {
2107         u64 *spte = vcpu->arch.last_pte_updated;
2108
2109         return !!(spte && (*spte & shadow_accessed_mask));
2110 }
2111
2112 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2113                                           const u8 *new, int bytes)
2114 {
2115         gfn_t gfn;
2116         int r;
2117         u64 gpte = 0;
2118         pfn_t pfn;
2119
2120         vcpu->arch.update_pte.largepage = 0;
2121
2122         if (bytes != 4 && bytes != 8)
2123                 return;
2124
2125         /*
2126          * Assume that the pte write on a page table of the same type
2127          * as the current vcpu paging mode.  This is nearly always true
2128          * (might be false while changing modes).  Note it is verified later
2129          * by update_pte().
2130          */
2131         if (is_pae(vcpu)) {
2132                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2133                 if ((bytes == 4) && (gpa % 4 == 0)) {
2134                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2135                         if (r)
2136                                 return;
2137                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2138                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2139                         memcpy((void *)&gpte, new, 8);
2140                 }
2141         } else {
2142                 if ((bytes == 4) && (gpa % 4 == 0))
2143                         memcpy((void *)&gpte, new, 4);
2144         }
2145         if (!is_present_pte(gpte))
2146                 return;
2147         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2148
2149         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2150                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2151                 vcpu->arch.update_pte.largepage = 1;
2152         }
2153         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2154         smp_rmb();
2155         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2156
2157         if (is_error_pfn(pfn)) {
2158                 kvm_release_pfn_clean(pfn);
2159                 return;
2160         }
2161         vcpu->arch.update_pte.gfn = gfn;
2162         vcpu->arch.update_pte.pfn = pfn;
2163 }
2164
2165 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2166 {
2167         u64 *spte = vcpu->arch.last_pte_updated;
2168
2169         if (spte
2170             && vcpu->arch.last_pte_gfn == gfn
2171             && shadow_accessed_mask
2172             && !(*spte & shadow_accessed_mask)
2173             && is_shadow_present_pte(*spte))
2174                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2175 }
2176
2177 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2178                        const u8 *new, int bytes)
2179 {
2180         gfn_t gfn = gpa >> PAGE_SHIFT;
2181         struct kvm_mmu_page *sp;
2182         struct hlist_node *node, *n;
2183         struct hlist_head *bucket;
2184         unsigned index;
2185         u64 entry, gentry;
2186         u64 *spte;
2187         unsigned offset = offset_in_page(gpa);
2188         unsigned pte_size;
2189         unsigned page_offset;
2190         unsigned misaligned;
2191         unsigned quadrant;
2192         int level;
2193         int flooded = 0;
2194         int npte;
2195         int r;
2196
2197         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2198         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2199         spin_lock(&vcpu->kvm->mmu_lock);
2200         kvm_mmu_access_page(vcpu, gfn);
2201         kvm_mmu_free_some_pages(vcpu);
2202         ++vcpu->kvm->stat.mmu_pte_write;
2203         kvm_mmu_audit(vcpu, "pre pte write");
2204         if (gfn == vcpu->arch.last_pt_write_gfn
2205             && !last_updated_pte_accessed(vcpu)) {
2206                 ++vcpu->arch.last_pt_write_count;
2207                 if (vcpu->arch.last_pt_write_count >= 3)
2208                         flooded = 1;
2209         } else {
2210                 vcpu->arch.last_pt_write_gfn = gfn;
2211                 vcpu->arch.last_pt_write_count = 1;
2212                 vcpu->arch.last_pte_updated = NULL;
2213         }
2214         index = kvm_page_table_hashfn(gfn);
2215         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2216         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2217                 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
2218                         continue;
2219                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2220                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2221                 misaligned |= bytes < 4;
2222                 if (misaligned || flooded) {
2223                         /*
2224                          * Misaligned accesses are too much trouble to fix
2225                          * up; also, they usually indicate a page is not used
2226                          * as a page table.
2227                          *
2228                          * If we're seeing too many writes to a page,
2229                          * it may no longer be a page table, or we may be
2230                          * forking, in which case it is better to unmap the
2231                          * page.
2232                          */
2233                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2234                                  gpa, bytes, sp->role.word);
2235                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2236                                 n = bucket->first;
2237                         ++vcpu->kvm->stat.mmu_flooded;
2238                         continue;
2239                 }
2240                 page_offset = offset;
2241                 level = sp->role.level;
2242                 npte = 1;
2243                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2244                         page_offset <<= 1;      /* 32->64 */
2245                         /*
2246                          * A 32-bit pde maps 4MB while the shadow pdes map
2247                          * only 2MB.  So we need to double the offset again
2248                          * and zap two pdes instead of one.
2249                          */
2250                         if (level == PT32_ROOT_LEVEL) {
2251                                 page_offset &= ~7; /* kill rounding error */
2252                                 page_offset <<= 1;
2253                                 npte = 2;
2254                         }
2255                         quadrant = page_offset >> PAGE_SHIFT;
2256                         page_offset &= ~PAGE_MASK;
2257                         if (quadrant != sp->role.quadrant)
2258                                 continue;
2259                 }
2260                 spte = &sp->spt[page_offset / sizeof(*spte)];
2261                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2262                         gentry = 0;
2263                         r = kvm_read_guest_atomic(vcpu->kvm,
2264                                                   gpa & ~(u64)(pte_size - 1),
2265                                                   &gentry, pte_size);
2266                         new = (const void *)&gentry;
2267                         if (r < 0)
2268                                 new = NULL;
2269                 }
2270                 while (npte--) {
2271                         entry = *spte;
2272                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2273                         if (new)
2274                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2275                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2276                         ++spte;
2277                 }
2278         }
2279         kvm_mmu_audit(vcpu, "post pte write");
2280         spin_unlock(&vcpu->kvm->mmu_lock);
2281         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2282                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2283                 vcpu->arch.update_pte.pfn = bad_pfn;
2284         }
2285 }
2286
2287 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2288 {
2289         gpa_t gpa;
2290         int r;
2291
2292         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2293
2294         spin_lock(&vcpu->kvm->mmu_lock);
2295         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2296         spin_unlock(&vcpu->kvm->mmu_lock);
2297         return r;
2298 }
2299 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2300
2301 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2302 {
2303         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2304                 struct kvm_mmu_page *sp;
2305
2306                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2307                                   struct kvm_mmu_page, link);
2308                 kvm_mmu_zap_page(vcpu->kvm, sp);
2309                 ++vcpu->kvm->stat.mmu_recycled;
2310         }
2311 }
2312
2313 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2314 {
2315         int r;
2316         enum emulation_result er;
2317
2318         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2319         if (r < 0)
2320                 goto out;
2321
2322         if (!r) {
2323                 r = 1;
2324                 goto out;
2325         }
2326
2327         r = mmu_topup_memory_caches(vcpu);
2328         if (r)
2329                 goto out;
2330
2331         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2332
2333         switch (er) {
2334         case EMULATE_DONE:
2335                 return 1;
2336         case EMULATE_DO_MMIO:
2337                 ++vcpu->stat.mmio_exits;
2338                 return 0;
2339         case EMULATE_FAIL:
2340                 kvm_report_emulation_failure(vcpu, "pagetable");
2341                 return 1;
2342         default:
2343                 BUG();
2344         }
2345 out:
2346         return r;
2347 }
2348 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2349
2350 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2351 {
2352         spin_lock(&vcpu->kvm->mmu_lock);
2353         vcpu->arch.mmu.invlpg(vcpu, gva);
2354         spin_unlock(&vcpu->kvm->mmu_lock);
2355         kvm_mmu_flush_tlb(vcpu);
2356         ++vcpu->stat.invlpg;
2357 }
2358 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2359
2360 void kvm_enable_tdp(void)
2361 {
2362         tdp_enabled = true;
2363 }
2364 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2365
2366 void kvm_disable_tdp(void)
2367 {
2368         tdp_enabled = false;
2369 }
2370 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2371
2372 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2373 {
2374         struct kvm_mmu_page *sp;
2375
2376         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2377                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2378                                   struct kvm_mmu_page, link);
2379                 kvm_mmu_zap_page(vcpu->kvm, sp);
2380                 cond_resched();
2381         }
2382         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2383 }
2384
2385 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2386 {
2387         struct page *page;
2388         int i;
2389
2390         ASSERT(vcpu);
2391
2392         if (vcpu->kvm->arch.n_requested_mmu_pages)
2393                 vcpu->kvm->arch.n_free_mmu_pages =
2394                                         vcpu->kvm->arch.n_requested_mmu_pages;
2395         else
2396                 vcpu->kvm->arch.n_free_mmu_pages =
2397                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2398         /*
2399          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2400          * Therefore we need to allocate shadow page tables in the first
2401          * 4GB of memory, which happens to fit the DMA32 zone.
2402          */
2403         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2404         if (!page)
2405                 goto error_1;
2406         vcpu->arch.mmu.pae_root = page_address(page);
2407         for (i = 0; i < 4; ++i)
2408                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2409
2410         return 0;
2411
2412 error_1:
2413         free_mmu_pages(vcpu);
2414         return -ENOMEM;
2415 }
2416
2417 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2418 {
2419         ASSERT(vcpu);
2420         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2421
2422         return alloc_mmu_pages(vcpu);
2423 }
2424
2425 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2426 {
2427         ASSERT(vcpu);
2428         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2429
2430         return init_kvm_mmu(vcpu);
2431 }
2432
2433 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2434 {
2435         ASSERT(vcpu);
2436
2437         destroy_kvm_mmu(vcpu);
2438         free_mmu_pages(vcpu);
2439         mmu_free_memory_caches(vcpu);
2440 }
2441
2442 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2443 {
2444         struct kvm_mmu_page *sp;
2445
2446         spin_lock(&kvm->mmu_lock);
2447         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2448                 int i;
2449                 u64 *pt;
2450
2451                 if (!test_bit(slot, &sp->slot_bitmap))
2452                         continue;
2453
2454                 pt = sp->spt;
2455                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2456                         /* avoid RMW */
2457                         if (pt[i] & PT_WRITABLE_MASK)
2458                                 pt[i] &= ~PT_WRITABLE_MASK;
2459         }
2460         kvm_flush_remote_tlbs(kvm);
2461         spin_unlock(&kvm->mmu_lock);
2462 }
2463
2464 void kvm_mmu_zap_all(struct kvm *kvm)
2465 {
2466         struct kvm_mmu_page *sp, *node;
2467
2468         spin_lock(&kvm->mmu_lock);
2469         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2470                 if (kvm_mmu_zap_page(kvm, sp))
2471                         node = container_of(kvm->arch.active_mmu_pages.next,
2472                                             struct kvm_mmu_page, link);
2473         spin_unlock(&kvm->mmu_lock);
2474
2475         kvm_flush_remote_tlbs(kvm);
2476 }
2477
2478 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2479 {
2480         struct kvm_mmu_page *page;
2481
2482         page = container_of(kvm->arch.active_mmu_pages.prev,
2483                             struct kvm_mmu_page, link);
2484         kvm_mmu_zap_page(kvm, page);
2485 }
2486
2487 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2488 {
2489         struct kvm *kvm;
2490         struct kvm *kvm_freed = NULL;
2491         int cache_count = 0;
2492
2493         spin_lock(&kvm_lock);
2494
2495         list_for_each_entry(kvm, &vm_list, vm_list) {
2496                 int npages;
2497
2498                 if (!down_read_trylock(&kvm->slots_lock))
2499                         continue;
2500                 spin_lock(&kvm->mmu_lock);
2501                 npages = kvm->arch.n_alloc_mmu_pages -
2502                          kvm->arch.n_free_mmu_pages;
2503                 cache_count += npages;
2504                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2505                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2506                         cache_count--;
2507                         kvm_freed = kvm;
2508                 }
2509                 nr_to_scan--;
2510
2511                 spin_unlock(&kvm->mmu_lock);
2512                 up_read(&kvm->slots_lock);
2513         }
2514         if (kvm_freed)
2515                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2516
2517         spin_unlock(&kvm_lock);
2518
2519         return cache_count;
2520 }
2521
2522 static struct shrinker mmu_shrinker = {
2523         .shrink = mmu_shrink,
2524         .seeks = DEFAULT_SEEKS * 10,
2525 };
2526
2527 static void mmu_destroy_caches(void)
2528 {
2529         if (pte_chain_cache)
2530                 kmem_cache_destroy(pte_chain_cache);
2531         if (rmap_desc_cache)
2532                 kmem_cache_destroy(rmap_desc_cache);
2533         if (mmu_page_header_cache)
2534                 kmem_cache_destroy(mmu_page_header_cache);
2535 }
2536
2537 void kvm_mmu_module_exit(void)
2538 {
2539         mmu_destroy_caches();
2540         unregister_shrinker(&mmu_shrinker);
2541 }
2542
2543 int kvm_mmu_module_init(void)
2544 {
2545         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2546                                             sizeof(struct kvm_pte_chain),
2547                                             0, 0, NULL);
2548         if (!pte_chain_cache)
2549                 goto nomem;
2550         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2551                                             sizeof(struct kvm_rmap_desc),
2552                                             0, 0, NULL);
2553         if (!rmap_desc_cache)
2554                 goto nomem;
2555
2556         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2557                                                   sizeof(struct kvm_mmu_page),
2558                                                   0, 0, NULL);
2559         if (!mmu_page_header_cache)
2560                 goto nomem;
2561
2562         register_shrinker(&mmu_shrinker);
2563
2564         return 0;
2565
2566 nomem:
2567         mmu_destroy_caches();
2568         return -ENOMEM;
2569 }
2570
2571 /*
2572  * Caculate mmu pages needed for kvm.
2573  */
2574 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2575 {
2576         int i;
2577         unsigned int nr_mmu_pages;
2578         unsigned int  nr_pages = 0;
2579
2580         for (i = 0; i < kvm->nmemslots; i++)
2581                 nr_pages += kvm->memslots[i].npages;
2582
2583         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2584         nr_mmu_pages = max(nr_mmu_pages,
2585                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2586
2587         return nr_mmu_pages;
2588 }
2589
2590 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2591                                 unsigned len)
2592 {
2593         if (len > buffer->len)
2594                 return NULL;
2595         return buffer->ptr;
2596 }
2597
2598 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2599                                 unsigned len)
2600 {
2601         void *ret;
2602
2603         ret = pv_mmu_peek_buffer(buffer, len);
2604         if (!ret)
2605                 return ret;
2606         buffer->ptr += len;
2607         buffer->len -= len;
2608         buffer->processed += len;
2609         return ret;
2610 }
2611
2612 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2613                              gpa_t addr, gpa_t value)
2614 {
2615         int bytes = 8;
2616         int r;
2617
2618         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2619                 bytes = 4;
2620
2621         r = mmu_topup_memory_caches(vcpu);
2622         if (r)
2623                 return r;
2624
2625         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2626                 return -EFAULT;
2627
2628         return 1;
2629 }
2630
2631 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2632 {
2633         kvm_x86_ops->tlb_flush(vcpu);
2634         return 1;
2635 }
2636
2637 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2638 {
2639         spin_lock(&vcpu->kvm->mmu_lock);
2640         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2641         spin_unlock(&vcpu->kvm->mmu_lock);
2642         return 1;
2643 }
2644
2645 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2646                              struct kvm_pv_mmu_op_buffer *buffer)
2647 {
2648         struct kvm_mmu_op_header *header;
2649
2650         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2651         if (!header)
2652                 return 0;
2653         switch (header->op) {
2654         case KVM_MMU_OP_WRITE_PTE: {
2655                 struct kvm_mmu_op_write_pte *wpte;
2656
2657                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2658                 if (!wpte)
2659                         return 0;
2660                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2661                                         wpte->pte_val);
2662         }
2663         case KVM_MMU_OP_FLUSH_TLB: {
2664                 struct kvm_mmu_op_flush_tlb *ftlb;
2665
2666                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2667                 if (!ftlb)
2668                         return 0;
2669                 return kvm_pv_mmu_flush_tlb(vcpu);
2670         }
2671         case KVM_MMU_OP_RELEASE_PT: {
2672                 struct kvm_mmu_op_release_pt *rpt;
2673
2674                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2675                 if (!rpt)
2676                         return 0;
2677                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2678         }
2679         default: return 0;
2680         }
2681 }
2682
2683 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2684                   gpa_t addr, unsigned long *ret)
2685 {
2686         int r;
2687         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2688
2689         buffer->ptr = buffer->buf;
2690         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2691         buffer->processed = 0;
2692
2693         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2694         if (r)
2695                 goto out;
2696
2697         while (buffer->len) {
2698                 r = kvm_pv_mmu_op_one(vcpu, buffer);
2699                 if (r < 0)
2700                         goto out;
2701                 if (r == 0)
2702                         break;
2703         }
2704
2705         r = 1;
2706 out:
2707         *ret = buffer->processed;
2708         return r;
2709 }
2710
2711 #ifdef AUDIT
2712
2713 static const char *audit_msg;
2714
2715 static gva_t canonicalize(gva_t gva)
2716 {
2717 #ifdef CONFIG_X86_64
2718         gva = (long long)(gva << 16) >> 16;
2719 #endif
2720         return gva;
2721 }
2722
2723 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2724                                 gva_t va, int level)
2725 {
2726         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2727         int i;
2728         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2729
2730         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2731                 u64 ent = pt[i];
2732
2733                 if (ent == shadow_trap_nonpresent_pte)
2734                         continue;
2735
2736                 va = canonicalize(va);
2737                 if (level > 1) {
2738                         if (ent == shadow_notrap_nonpresent_pte)
2739                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2740                                        " in nonleaf level: levels %d gva %lx"
2741                                        " level %d pte %llx\n", audit_msg,
2742                                        vcpu->arch.mmu.root_level, va, level, ent);
2743
2744                         audit_mappings_page(vcpu, ent, va, level - 1);
2745                 } else {
2746                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2747                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2748
2749                         if (is_shadow_present_pte(ent)
2750                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2751                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2752                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2753                                        audit_msg, vcpu->arch.mmu.root_level,
2754                                        va, gpa, hpa, ent,
2755                                        is_shadow_present_pte(ent));
2756                         else if (ent == shadow_notrap_nonpresent_pte
2757                                  && !is_error_hpa(hpa))
2758                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2759                                        " valid guest gva %lx\n", audit_msg, va);
2760                         kvm_release_pfn_clean(pfn);
2761
2762                 }
2763         }
2764 }
2765
2766 static void audit_mappings(struct kvm_vcpu *vcpu)
2767 {
2768         unsigned i;
2769
2770         if (vcpu->arch.mmu.root_level == 4)
2771                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2772         else
2773                 for (i = 0; i < 4; ++i)
2774                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2775                                 audit_mappings_page(vcpu,
2776                                                     vcpu->arch.mmu.pae_root[i],
2777                                                     i << 30,
2778                                                     2);
2779 }
2780
2781 static int count_rmaps(struct kvm_vcpu *vcpu)
2782 {
2783         int nmaps = 0;
2784         int i, j, k;
2785
2786         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2787                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2788                 struct kvm_rmap_desc *d;
2789
2790                 for (j = 0; j < m->npages; ++j) {
2791                         unsigned long *rmapp = &m->rmap[j];
2792
2793                         if (!*rmapp)
2794                                 continue;
2795                         if (!(*rmapp & 1)) {
2796                                 ++nmaps;
2797                                 continue;
2798                         }
2799                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2800                         while (d) {
2801                                 for (k = 0; k < RMAP_EXT; ++k)
2802                                         if (d->shadow_ptes[k])
2803                                                 ++nmaps;
2804                                         else
2805                                                 break;
2806                                 d = d->more;
2807                         }
2808                 }
2809         }
2810         return nmaps;
2811 }
2812
2813 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2814 {
2815         int nmaps = 0;
2816         struct kvm_mmu_page *sp;
2817         int i;
2818
2819         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2820                 u64 *pt = sp->spt;
2821
2822                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2823                         continue;
2824
2825                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2826                         u64 ent = pt[i];
2827
2828                         if (!(ent & PT_PRESENT_MASK))
2829                                 continue;
2830                         if (!(ent & PT_WRITABLE_MASK))
2831                                 continue;
2832                         ++nmaps;
2833                 }
2834         }
2835         return nmaps;
2836 }
2837
2838 static void audit_rmap(struct kvm_vcpu *vcpu)
2839 {
2840         int n_rmap = count_rmaps(vcpu);
2841         int n_actual = count_writable_mappings(vcpu);
2842
2843         if (n_rmap != n_actual)
2844                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2845                        __func__, audit_msg, n_rmap, n_actual);
2846 }
2847
2848 static void audit_write_protection(struct kvm_vcpu *vcpu)
2849 {
2850         struct kvm_mmu_page *sp;
2851         struct kvm_memory_slot *slot;
2852         unsigned long *rmapp;
2853         gfn_t gfn;
2854
2855         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2856                 if (sp->role.metaphysical)
2857                         continue;
2858
2859                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2860                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2861                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2862                 if (*rmapp)
2863                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2864                                " mappings: gfn %lx role %x\n",
2865                                __func__, audit_msg, sp->gfn,
2866                                sp->role.word);
2867         }
2868 }
2869
2870 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2871 {
2872         int olddbg = dbg;
2873
2874         dbg = 0;
2875         audit_msg = msg;
2876         audit_rmap(vcpu);
2877         audit_write_protection(vcpu);
2878         audit_mappings(vcpu);
2879         dbg = olddbg;
2880 }
2881
2882 #endif