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

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "vmx.h"
  21 #include "mmu.h"
  22
  23 #include <linux/kvm_host.h>
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29 #include <linux/swap.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/compiler.h>
  32
  33 #include <asm/page.h>
  34 #include <asm/cmpxchg.h>
  35 #include <asm/io.h>
  36
  37 /*
  38  * When setting this variable to true it enables Two-Dimensional-Paging
  39  * where the hardware walks 2 page tables:
  40  * 1. the guest-virtual to guest-physical
  41  * 2. while doing 1. it walks guest-physical to host-physical
  42  * If the hardware supports that we don't need to do shadow paging.
  43  */
  44 bool tdp_enabled = false;
  45
  46 #undef MMU_DEBUG
  47
  48 #undef AUDIT
  49
  50 #ifdef AUDIT
  51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  52 #else
  53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  54 #endif
  55
  56 #ifdef MMU_DEBUG
  57
  58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  60
  61 #else
  62
  63 #define pgprintk(x...) do { } while (0)
  64 #define rmap_printk(x...) do { } while (0)
  65
  66 #endif
  67
  68 #if defined(MMU_DEBUG) || defined(AUDIT)
  69 static int dbg = 0;
  70 module_param(dbg, bool, 0644);
  71 #endif
  72
  73 static int oos_shadow = 1;
  74 module_param(oos_shadow, bool, 0644);
  75
  76 #ifndef MMU_DEBUG
  77 #define ASSERT(x) do { } while (0)
  78 #else
  79 #define ASSERT(x)                                                       \
  80         if (!(x)) {                                                     \
  81                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  82                        __FILE__, __LINE__, #x);                         \
  83         }
  84 #endif
  85
  86 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  87 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  88
  89 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  90
  91 #define PT64_LEVEL_BITS 9
  92
  93 #define PT64_LEVEL_SHIFT(level) \
  94                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  95
  96 #define PT64_LEVEL_MASK(level) \
  97                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  98
  99 #define PT64_INDEX(address, level)\
 100         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 101
 102
 103 #define PT32_LEVEL_BITS 10
 104
 105 #define PT32_LEVEL_SHIFT(level) \
 106                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 107
 108 #define PT32_LEVEL_MASK(level) \
 109                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 110
 111 #define PT32_INDEX(address, level)\
 112         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 113
 114
 115 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 116 #define PT64_DIR_BASE_ADDR_MASK \
 117         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 118
 119 #define PT32_BASE_ADDR_MASK PAGE_MASK
 120 #define PT32_DIR_BASE_ADDR_MASK \
 121         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 122
 123 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 124                         | PT64_NX_MASK)
 125
 126 #define PFERR_PRESENT_MASK (1U << 0)
 127 #define PFERR_WRITE_MASK (1U << 1)
 128 #define PFERR_USER_MASK (1U << 2)
 129 #define PFERR_FETCH_MASK (1U << 4)
 130
 131 #define PT_DIRECTORY_LEVEL 2
 132 #define PT_PAGE_TABLE_LEVEL 1
 133
 134 #define RMAP_EXT 4
 135
 136 #define ACC_EXEC_MASK    1
 137 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 138 #define ACC_USER_MASK    PT_USER_MASK
 139 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 140
 141 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 142
 143 struct kvm_rmap_desc {
 144         u64 *shadow_ptes[RMAP_EXT];
 145         struct kvm_rmap_desc *more;
 146 };
 147
 148 struct kvm_shadow_walk {
 149         int (*entry)(struct kvm_shadow_walk *walk, struct kvm_vcpu *vcpu,
 150                      u64 addr, u64 *spte, int level);
 151 };
 152
 153 struct kvm_unsync_walk {
 154         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 155 };
 156
 157 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 158
 159 static struct kmem_cache *pte_chain_cache;
 160 static struct kmem_cache *rmap_desc_cache;
 161 static struct kmem_cache *mmu_page_header_cache;
 162
 163 static u64 __read_mostly shadow_trap_nonpresent_pte;
 164 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 165 static u64 __read_mostly shadow_base_present_pte;
 166 static u64 __read_mostly shadow_nx_mask;
 167 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 168 static u64 __read_mostly shadow_user_mask;
 169 static u64 __read_mostly shadow_accessed_mask;
 170 static u64 __read_mostly shadow_dirty_mask;
 171 static u64 __read_mostly shadow_mt_mask;
 172
 173 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 174 {
 175         shadow_trap_nonpresent_pte = trap_pte;
 176         shadow_notrap_nonpresent_pte = notrap_pte;
 177 }
 178 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 179
 180 void kvm_mmu_set_base_ptes(u64 base_pte)
 181 {
 182         shadow_base_present_pte = base_pte;
 183 }
 184 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 185
 186 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 187                 u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 mt_mask)
 188 {
 189         shadow_user_mask = user_mask;
 190         shadow_accessed_mask = accessed_mask;
 191         shadow_dirty_mask = dirty_mask;
 192         shadow_nx_mask = nx_mask;
 193         shadow_x_mask = x_mask;
 194         shadow_mt_mask = mt_mask;
 195 }
 196 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 197
 198 static int is_write_protection(struct kvm_vcpu *vcpu)
 199 {
 200         return vcpu->arch.cr0 & X86_CR0_WP;
 201 }
 202
 203 static int is_cpuid_PSE36(void)
 204 {
 205         return 1;
 206 }
 207
 208 static int is_nx(struct kvm_vcpu *vcpu)
 209 {
 210         return vcpu->arch.shadow_efer & EFER_NX;
 211 }
 212
 213 static int is_present_pte(unsigned long pte)
 214 {
 215         return pte & PT_PRESENT_MASK;
 216 }
 217
 218 static int is_shadow_present_pte(u64 pte)
 219 {
 220         return pte != shadow_trap_nonpresent_pte
 221                 && pte != shadow_notrap_nonpresent_pte;
 222 }
 223
 224 static int is_large_pte(u64 pte)
 225 {
 226         return pte & PT_PAGE_SIZE_MASK;
 227 }
 228
 229 static int is_writeble_pte(unsigned long pte)
 230 {
 231         return pte & PT_WRITABLE_MASK;
 232 }
 233
 234 static int is_dirty_pte(unsigned long pte)
 235 {
 236         return pte & shadow_dirty_mask;
 237 }
 238
 239 static int is_rmap_pte(u64 pte)
 240 {
 241         return is_shadow_present_pte(pte);
 242 }
 243
 244 static pfn_t spte_to_pfn(u64 pte)
 245 {
 246         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 247 }
 248
 249 static gfn_t pse36_gfn_delta(u32 gpte)
 250 {
 251         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 252
 253         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 254 }
 255
 256 static void set_shadow_pte(u64 *sptep, u64 spte)
 257 {
 258 #ifdef CONFIG_X86_64
 259         set_64bit((unsigned long *)sptep, spte);
 260 #else
 261         set_64bit((unsigned long long *)sptep, spte);
 262 #endif
 263 }
 264
 265 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 266                                   struct kmem_cache *base_cache, int min)
 267 {
 268         void *obj;
 269
 270         if (cache->nobjs >= min)
 271                 return 0;
 272         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 273                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 274                 if (!obj)
 275                         return -ENOMEM;
 276                 cache->objects[cache->nobjs++] = obj;
 277         }
 278         return 0;
 279 }
 280
 281 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 282 {
 283         while (mc->nobjs)
 284                 kfree(mc->objects[--mc->nobjs]);
 285 }
 286
 287 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 288                                        int min)
 289 {
 290         struct page *page;
 291
 292         if (cache->nobjs >= min)
 293                 return 0;
 294         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 295                 page = alloc_page(GFP_KERNEL);
 296                 if (!page)
 297                         return -ENOMEM;
 298                 set_page_private(page, 0);
 299                 cache->objects[cache->nobjs++] = page_address(page);
 300         }
 301         return 0;
 302 }
 303
 304 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 305 {
 306         while (mc->nobjs)
 307                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 308 }
 309
 310 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 311 {
 312         int r;
 313
 314         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 315                                    pte_chain_cache, 4);
 316         if (r)
 317                 goto out;
 318         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 319                                    rmap_desc_cache, 4);
 320         if (r)
 321                 goto out;
 322         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 323         if (r)
 324                 goto out;
 325         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 326                                    mmu_page_header_cache, 4);
 327 out:
 328         return r;
 329 }
 330
 331 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 332 {
 333         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 334         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 335         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 336         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 337 }
 338
 339 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 340                                     size_t size)
 341 {
 342         void *p;
 343
 344         BUG_ON(!mc->nobjs);
 345         p = mc->objects[--mc->nobjs];
 346         memset(p, 0, size);
 347         return p;
 348 }
 349
 350 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 351 {
 352         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 353                                       sizeof(struct kvm_pte_chain));
 354 }
 355
 356 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 357 {
 358         kfree(pc);
 359 }
 360
 361 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 362 {
 363         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 364                                       sizeof(struct kvm_rmap_desc));
 365 }
 366
 367 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 368 {
 369         kfree(rd);
 370 }
 371
 372 /*
 373  * Return the pointer to the largepage write count for a given
 374  * gfn, handling slots that are not large page aligned.
 375  */
 376 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 377 {
 378         unsigned long idx;
 379
 380         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 381               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 382         return &slot->lpage_info[idx].write_count;
 383 }
 384
 385 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 386 {
 387         int *write_count;
 388
 389         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 390         *write_count += 1;
 391 }
 392
 393 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 394 {
 395         int *write_count;
 396
 397         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 398         *write_count -= 1;
 399         WARN_ON(*write_count < 0);
 400 }
 401
 402 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 403 {
 404         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
 405         int *largepage_idx;
 406
 407         if (slot) {
 408                 largepage_idx = slot_largepage_idx(gfn, slot);
 409                 return *largepage_idx;
 410         }
 411
 412         return 1;
 413 }
 414
 415 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 416 {
 417         struct vm_area_struct *vma;
 418         unsigned long addr;
 419         int ret = 0;
 420
 421         addr = gfn_to_hva(kvm, gfn);
 422         if (kvm_is_error_hva(addr))
 423                 return ret;
 424
 425         down_read(&current->mm->mmap_sem);
 426         vma = find_vma(current->mm, addr);
 427         if (vma && is_vm_hugetlb_page(vma))
 428                 ret = 1;
 429         up_read(&current->mm->mmap_sem);
 430
 431         return ret;
 432 }
 433
 434 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 435 {
 436         struct kvm_memory_slot *slot;
 437
 438         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 439                 return 0;
 440
 441         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 442                 return 0;
 443
 444         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 445         if (slot && slot->dirty_bitmap)
 446                 return 0;
 447
 448         return 1;
 449 }
 450
 451 /*
 452  * Take gfn and return the reverse mapping to it.
 453  * Note: gfn must be unaliased before this function get called
 454  */
 455
 456 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 457 {
 458         struct kvm_memory_slot *slot;
 459         unsigned long idx;
 460
 461         slot = gfn_to_memslot(kvm, gfn);
 462         if (!lpage)
 463                 return &slot->rmap[gfn - slot->base_gfn];
 464
 465         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 466               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 467
 468         return &slot->lpage_info[idx].rmap_pde;
 469 }
 470
 471 /*
 472  * Reverse mapping data structures:
 473  *
 474  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 475  * that points to page_address(page).
 476  *
 477  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 478  * containing more mappings.
 479  */
 480 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 481 {
 482         struct kvm_mmu_page *sp;
 483         struct kvm_rmap_desc *desc;
 484         unsigned long *rmapp;
 485         int i;
 486
 487         if (!is_rmap_pte(*spte))
 488                 return;
 489         gfn = unalias_gfn(vcpu->kvm, gfn);
 490         sp = page_header(__pa(spte));
 491         sp->gfns[spte - sp->spt] = gfn;
 492         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 493         if (!*rmapp) {
 494                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 495                 *rmapp = (unsigned long)spte;
 496         } else if (!(*rmapp & 1)) {
 497                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 498                 desc = mmu_alloc_rmap_desc(vcpu);
 499                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 500                 desc->shadow_ptes[1] = spte;
 501                 *rmapp = (unsigned long)desc | 1;
 502         } else {
 503                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 504                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 505                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 506                         desc = desc->more;
 507                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 508                         desc->more = mmu_alloc_rmap_desc(vcpu);
 509                         desc = desc->more;
 510                 }
 511                 for (i = 0; desc->shadow_ptes[i]; ++i)
 512                         ;
 513                 desc->shadow_ptes[i] = spte;
 514         }
 515 }
 516
 517 static void rmap_desc_remove_entry(unsigned long *rmapp,
 518                                    struct kvm_rmap_desc *desc,
 519                                    int i,
 520                                    struct kvm_rmap_desc *prev_desc)
 521 {
 522         int j;
 523
 524         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 525                 ;
 526         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 527         desc->shadow_ptes[j] = NULL;
 528         if (j != 0)
 529                 return;
 530         if (!prev_desc && !desc->more)
 531                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 532         else
 533                 if (prev_desc)
 534                         prev_desc->more = desc->more;
 535                 else
 536                         *rmapp = (unsigned long)desc->more | 1;
 537         mmu_free_rmap_desc(desc);
 538 }
 539
 540 static void rmap_remove(struct kvm *kvm, u64 *spte)
 541 {
 542         struct kvm_rmap_desc *desc;
 543         struct kvm_rmap_desc *prev_desc;
 544         struct kvm_mmu_page *sp;
 545         pfn_t pfn;
 546         unsigned long *rmapp;
 547         int i;
 548
 549         if (!is_rmap_pte(*spte))
 550                 return;
 551         sp = page_header(__pa(spte));
 552         pfn = spte_to_pfn(*spte);
 553         if (*spte & shadow_accessed_mask)
 554                 kvm_set_pfn_accessed(pfn);
 555         if (is_writeble_pte(*spte))
 556                 kvm_release_pfn_dirty(pfn);
 557         else
 558                 kvm_release_pfn_clean(pfn);
 559         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 560         if (!*rmapp) {
 561                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 562                 BUG();
 563         } else if (!(*rmapp & 1)) {
 564                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 565                 if ((u64 *)*rmapp != spte) {
 566                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 567                                spte, *spte);
 568                         BUG();
 569                 }
 570                 *rmapp = 0;
 571         } else {
 572                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 573                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 574                 prev_desc = NULL;
 575                 while (desc) {
 576                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 577                                 if (desc->shadow_ptes[i] == spte) {
 578                                         rmap_desc_remove_entry(rmapp,
 579                                                                desc, i,
 580                                                                prev_desc);
 581                                         return;
 582                                 }
 583                         prev_desc = desc;
 584                         desc = desc->more;
 585                 }
 586                 BUG();
 587         }
 588 }
 589
 590 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 591 {
 592         struct kvm_rmap_desc *desc;
 593         struct kvm_rmap_desc *prev_desc;
 594         u64 *prev_spte;
 595         int i;
 596
 597         if (!*rmapp)
 598                 return NULL;
 599         else if (!(*rmapp & 1)) {
 600                 if (!spte)
 601                         return (u64 *)*rmapp;
 602                 return NULL;
 603         }
 604         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 605         prev_desc = NULL;
 606         prev_spte = NULL;
 607         while (desc) {
 608                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 609                         if (prev_spte == spte)
 610                                 return desc->shadow_ptes[i];
 611                         prev_spte = desc->shadow_ptes[i];
 612                 }
 613                 desc = desc->more;
 614         }
 615         return NULL;
 616 }
 617
 618 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 619 {
 620         unsigned long *rmapp;
 621         u64 *spte;
 622         int write_protected = 0;
 623
 624         gfn = unalias_gfn(kvm, gfn);
 625         rmapp = gfn_to_rmap(kvm, gfn, 0);
 626
 627         spte = rmap_next(kvm, rmapp, NULL);
 628         while (spte) {
 629                 BUG_ON(!spte);
 630                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 631                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 632                 if (is_writeble_pte(*spte)) {
 633                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 634                         write_protected = 1;
 635                 }
 636                 spte = rmap_next(kvm, rmapp, spte);
 637         }
 638         if (write_protected) {
 639                 pfn_t pfn;
 640
 641                 spte = rmap_next(kvm, rmapp, NULL);
 642                 pfn = spte_to_pfn(*spte);
 643                 kvm_set_pfn_dirty(pfn);
 644         }
 645
 646         /* check for huge page mappings */
 647         rmapp = gfn_to_rmap(kvm, gfn, 1);
 648         spte = rmap_next(kvm, rmapp, NULL);
 649         while (spte) {
 650                 BUG_ON(!spte);
 651                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 652                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 653                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 654                 if (is_writeble_pte(*spte)) {
 655                         rmap_remove(kvm, spte);
 656                         --kvm->stat.lpages;
 657                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 658                         spte = NULL;
 659                         write_protected = 1;
 660                 }
 661                 spte = rmap_next(kvm, rmapp, spte);
 662         }
 663
 664         if (write_protected)
 665                 kvm_flush_remote_tlbs(kvm);
 666 }
 667
 668 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 669 {
 670         u64 *spte;
 671         int need_tlb_flush = 0;
 672
 673         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 674                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 675                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 676                 rmap_remove(kvm, spte);
 677                 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 678                 need_tlb_flush = 1;
 679         }
 680         return need_tlb_flush;
 681 }
 682
 683 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 684                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 685 {
 686         int i;
 687         int retval = 0;
 688
 689         /*
 690          * If mmap_sem isn't taken, we can look the memslots with only
 691          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 692          */
 693         for (i = 0; i < kvm->nmemslots; i++) {
 694                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 695                 unsigned long start = memslot->userspace_addr;
 696                 unsigned long end;
 697
 698                 /* mmu_lock protects userspace_addr */
 699                 if (!start)
 700                         continue;
 701
 702                 end = start + (memslot->npages << PAGE_SHIFT);
 703                 if (hva >= start && hva < end) {
 704                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 705                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 706                         retval |= handler(kvm,
 707                                           &memslot->lpage_info[
 708                                                   gfn_offset /
 709                                                   KVM_PAGES_PER_HPAGE].rmap_pde);
 710                 }
 711         }
 712
 713         return retval;
 714 }
 715
 716 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 717 {
 718         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 719 }
 720
 721 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 722 {
 723         u64 *spte;
 724         int young = 0;
 725
 726         /* always return old for EPT */
 727         if (!shadow_accessed_mask)
 728                 return 0;
 729
 730         spte = rmap_next(kvm, rmapp, NULL);
 731         while (spte) {
 732                 int _young;
 733                 u64 _spte = *spte;
 734                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 735                 _young = _spte & PT_ACCESSED_MASK;
 736                 if (_young) {
 737                         young = 1;
 738                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 739                 }
 740                 spte = rmap_next(kvm, rmapp, spte);
 741         }
 742         return young;
 743 }
 744
 745 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 746 {
 747         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 748 }
 749
 750 #ifdef MMU_DEBUG
 751 static int is_empty_shadow_page(u64 *spt)
 752 {
 753         u64 *pos;
 754         u64 *end;
 755
 756         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 757                 if (is_shadow_present_pte(*pos)) {
 758                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 759                                pos, *pos);
 760                         return 0;
 761                 }
 762         return 1;
 763 }
 764 #endif
 765
 766 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 767 {
 768         ASSERT(is_empty_shadow_page(sp->spt));
 769         list_del(&sp->link);
 770         __free_page(virt_to_page(sp->spt));
 771         __free_page(virt_to_page(sp->gfns));
 772         kfree(sp);
 773         ++kvm->arch.n_free_mmu_pages;
 774 }
 775
 776 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 777 {
 778         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 779 }
 780
 781 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 782                                                u64 *parent_pte)
 783 {
 784         struct kvm_mmu_page *sp;
 785
 786         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 787         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 788         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 789         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 790         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 791         ASSERT(is_empty_shadow_page(sp->spt));
 792         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 793         sp->multimapped = 0;
 794         sp->parent_pte = parent_pte;
 795         --vcpu->kvm->arch.n_free_mmu_pages;
 796         return sp;
 797 }
 798
 799 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 800                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 801 {
 802         struct kvm_pte_chain *pte_chain;
 803         struct hlist_node *node;
 804         int i;
 805
 806         if (!parent_pte)
 807                 return;
 808         if (!sp->multimapped) {
 809                 u64 *old = sp->parent_pte;
 810
 811                 if (!old) {
 812                         sp->parent_pte = parent_pte;
 813                         return;
 814                 }
 815                 sp->multimapped = 1;
 816                 pte_chain = mmu_alloc_pte_chain(vcpu);
 817                 INIT_HLIST_HEAD(&sp->parent_ptes);
 818                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 819                 pte_chain->parent_ptes[0] = old;
 820         }
 821         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 822                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 823                         continue;
 824                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 825                         if (!pte_chain->parent_ptes[i]) {
 826                                 pte_chain->parent_ptes[i] = parent_pte;
 827                                 return;
 828                         }
 829         }
 830         pte_chain = mmu_alloc_pte_chain(vcpu);
 831         BUG_ON(!pte_chain);
 832         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 833         pte_chain->parent_ptes[0] = parent_pte;
 834 }
 835
 836 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 837                                        u64 *parent_pte)
 838 {
 839         struct kvm_pte_chain *pte_chain;
 840         struct hlist_node *node;
 841         int i;
 842
 843         if (!sp->multimapped) {
 844                 BUG_ON(sp->parent_pte != parent_pte);
 845                 sp->parent_pte = NULL;
 846                 return;
 847         }
 848         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 849                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 850                         if (!pte_chain->parent_ptes[i])
 851                                 break;
 852                         if (pte_chain->parent_ptes[i] != parent_pte)
 853                                 continue;
 854                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 855                                 && pte_chain->parent_ptes[i + 1]) {
 856                                 pte_chain->parent_ptes[i]
 857                                         = pte_chain->parent_ptes[i + 1];
 858                                 ++i;
 859                         }
 860                         pte_chain->parent_ptes[i] = NULL;
 861                         if (i == 0) {
 862                                 hlist_del(&pte_chain->link);
 863                                 mmu_free_pte_chain(pte_chain);
 864                                 if (hlist_empty(&sp->parent_ptes)) {
 865                                         sp->multimapped = 0;
 866                                         sp->parent_pte = NULL;
 867                                 }
 868                         }
 869                         return;
 870                 }
 871         BUG();
 872 }
 873
 874
 875 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 876                             mmu_parent_walk_fn fn)
 877 {
 878         struct kvm_pte_chain *pte_chain;
 879         struct hlist_node *node;
 880         struct kvm_mmu_page *parent_sp;
 881         int i;
 882
 883         if (!sp->multimapped && sp->parent_pte) {
 884                 parent_sp = page_header(__pa(sp->parent_pte));
 885                 fn(vcpu, parent_sp);
 886                 mmu_parent_walk(vcpu, parent_sp, fn);
 887                 return;
 888         }
 889         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 890                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 891                         if (!pte_chain->parent_ptes[i])
 892                                 break;
 893                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 894                         fn(vcpu, parent_sp);
 895                         mmu_parent_walk(vcpu, parent_sp, fn);
 896                 }
 897 }
 898
 899 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 900 {
 901         unsigned int index;
 902         struct kvm_mmu_page *sp = page_header(__pa(spte));
 903
 904         index = spte - sp->spt;
 905         __set_bit(index, sp->unsync_child_bitmap);
 906         sp->unsync_children = 1;
 907 }
 908
 909 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 910 {
 911         struct kvm_pte_chain *pte_chain;
 912         struct hlist_node *node;
 913         int i;
 914
 915         if (!sp->parent_pte)
 916                 return;
 917
 918         if (!sp->multimapped) {
 919                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 920                 return;
 921         }
 922
 923         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 924                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 925                         if (!pte_chain->parent_ptes[i])
 926                                 break;
 927                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 928                 }
 929 }
 930
 931 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 932 {
 933         sp->unsync_children = 1;
 934         kvm_mmu_update_parents_unsync(sp);
 935         return 1;
 936 }
 937
 938 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 939                                         struct kvm_mmu_page *sp)
 940 {
 941         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 942         kvm_mmu_update_parents_unsync(sp);
 943 }
 944
 945 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 946                                     struct kvm_mmu_page *sp)
 947 {
 948         int i;
 949
 950         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 951                 sp->spt[i] = shadow_trap_nonpresent_pte;
 952 }
 953
 954 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
 955                                struct kvm_mmu_page *sp)
 956 {
 957         return 1;
 958 }
 959
 960 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
 961 {
 962 }
 963
 964 #define for_each_unsync_children(bitmap, idx)           \
 965         for (idx = find_first_bit(bitmap, 512);         \
 966              idx < 512;                                 \
 967              idx = find_next_bit(bitmap, 512, idx+1))
 968
 969 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
 970                            struct kvm_unsync_walk *walker)
 971 {
 972         int i, ret;
 973
 974         if (!sp->unsync_children)
 975                 return 0;
 976
 977         for_each_unsync_children(sp->unsync_child_bitmap, i) {
 978                 u64 ent = sp->spt[i];
 979
 980                 if (is_shadow_present_pte(ent)) {
 981                         struct kvm_mmu_page *child;
 982                         child = page_header(ent & PT64_BASE_ADDR_MASK);
 983
 984                         if (child->unsync_children) {
 985                                 ret = mmu_unsync_walk(child, walker);
 986                                 if (ret)
 987                                         return ret;
 988                                 __clear_bit(i, sp->unsync_child_bitmap);
 989                         }
 990
 991                         if (child->unsync) {
 992                                 ret = walker->entry(child, walker);
 993                                 __clear_bit(i, sp->unsync_child_bitmap);
 994                                 if (ret)
 995                                         return ret;
 996                         }
 997                 }
 998         }
 999
1000         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1001                 sp->unsync_children = 0;
1002
1003         return 0;
1004 }
1005
1006 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1007 {
1008         unsigned index;
1009         struct hlist_head *bucket;
1010         struct kvm_mmu_page *sp;
1011         struct hlist_node *node;
1012
1013         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1014         index = kvm_page_table_hashfn(gfn);
1015         bucket = &kvm->arch.mmu_page_hash[index];
1016         hlist_for_each_entry(sp, node, bucket, hash_link)
1017                 if (sp->gfn == gfn && !sp->role.metaphysical
1018                     && !sp->role.invalid) {
1019                         pgprintk("%s: found role %x\n",
1020                                  __func__, sp->role.word);
1021                         return sp;
1022                 }
1023         return NULL;
1024 }
1025
1026 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1027 {
1028         WARN_ON(!sp->unsync);
1029         sp->unsync = 0;
1030         --kvm->stat.mmu_unsync;
1031 }
1032
1033 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1034
1035 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1036 {
1037         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1038                 kvm_mmu_zap_page(vcpu->kvm, sp);
1039                 return 1;
1040         }
1041
1042         rmap_write_protect(vcpu->kvm, sp->gfn);
1043         kvm_unlink_unsync_page(vcpu->kvm, sp);
1044         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1045                 kvm_mmu_zap_page(vcpu->kvm, sp);
1046                 return 1;
1047         }
1048
1049         kvm_mmu_flush_tlb(vcpu);
1050         return 0;
1051 }
1052
1053 struct sync_walker {
1054         struct kvm_vcpu *vcpu;
1055         struct kvm_unsync_walk walker;
1056 };
1057
1058 static int mmu_sync_fn(struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk)
1059 {
1060         struct sync_walker *sync_walk = container_of(walk, struct sync_walker,
1061                                                      walker);
1062         struct kvm_vcpu *vcpu = sync_walk->vcpu;
1063
1064         kvm_sync_page(vcpu, sp);
1065         return (need_resched() || spin_needbreak(&vcpu->kvm->mmu_lock));
1066 }
1067
1068 static void mmu_sync_children(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1069 {
1070         struct sync_walker walker = {
1071                 .walker = { .entry = mmu_sync_fn, },
1072                 .vcpu = vcpu,
1073         };
1074
1075         while (mmu_unsync_walk(sp, &walker.walker))
1076                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1077 }
1078
1079 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1080                                              gfn_t gfn,
1081                                              gva_t gaddr,
1082                                              unsigned level,
1083                                              int metaphysical,
1084                                              unsigned access,
1085                                              u64 *parent_pte)
1086 {
1087         union kvm_mmu_page_role role;
1088         unsigned index;
1089         unsigned quadrant;
1090         struct hlist_head *bucket;
1091         struct kvm_mmu_page *sp;
1092         struct hlist_node *node, *tmp;
1093
1094         role.word = 0;
1095         role.glevels = vcpu->arch.mmu.root_level;
1096         role.level = level;
1097         role.metaphysical = metaphysical;
1098         role.access = access;
1099         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1100                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1101                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1102                 role.quadrant = quadrant;
1103         }
1104         pgprintk("%s: looking gfn %lx role %x\n", __func__,
1105                  gfn, role.word);
1106         index = kvm_page_table_hashfn(gfn);
1107         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1108         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1109                 if (sp->gfn == gfn) {
1110                         if (sp->unsync)
1111                                 if (kvm_sync_page(vcpu, sp))
1112                                         continue;
1113
1114                         if (sp->role.word != role.word)
1115                                 continue;
1116
1117                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1118                         if (sp->unsync_children) {
1119                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1120                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1121                         }
1122                         pgprintk("%s: found\n", __func__);
1123                         return sp;
1124                 }
1125         ++vcpu->kvm->stat.mmu_cache_miss;
1126         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1127         if (!sp)
1128                 return sp;
1129         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1130         sp->gfn = gfn;
1131         sp->role = role;
1132         hlist_add_head(&sp->hash_link, bucket);
1133         if (!metaphysical) {
1134                 rmap_write_protect(vcpu->kvm, gfn);
1135                 account_shadowed(vcpu->kvm, gfn);
1136         }
1137         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1138                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1139         else
1140                 nonpaging_prefetch_page(vcpu, sp);
1141         return sp;
1142 }
1143
1144 static int walk_shadow(struct kvm_shadow_walk *walker,
1145                        struct kvm_vcpu *vcpu, u64 addr)
1146 {
1147         hpa_t shadow_addr;
1148         int level;
1149         int r;
1150         u64 *sptep;
1151         unsigned index;
1152
1153         shadow_addr = vcpu->arch.mmu.root_hpa;
1154         level = vcpu->arch.mmu.shadow_root_level;
1155         if (level == PT32E_ROOT_LEVEL) {
1156                 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1157                 shadow_addr &= PT64_BASE_ADDR_MASK;
1158                 --level;
1159         }
1160
1161         while (level >= PT_PAGE_TABLE_LEVEL) {
1162                 index = SHADOW_PT_INDEX(addr, level);
1163                 sptep = ((u64 *)__va(shadow_addr)) + index;
1164                 r = walker->entry(walker, vcpu, addr, sptep, level);
1165                 if (r)
1166                         return r;
1167                 shadow_addr = *sptep & PT64_BASE_ADDR_MASK;
1168                 --level;
1169         }
1170         return 0;
1171 }
1172
1173 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1174                                          struct kvm_mmu_page *sp)
1175 {
1176         unsigned i;
1177         u64 *pt;
1178         u64 ent;
1179
1180         pt = sp->spt;
1181
1182         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1183                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1184                         if (is_shadow_present_pte(pt[i]))
1185                                 rmap_remove(kvm, &pt[i]);
1186                         pt[i] = shadow_trap_nonpresent_pte;
1187                 }
1188                 return;
1189         }
1190
1191         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1192                 ent = pt[i];
1193
1194                 if (is_shadow_present_pte(ent)) {
1195                         if (!is_large_pte(ent)) {
1196                                 ent &= PT64_BASE_ADDR_MASK;
1197                                 mmu_page_remove_parent_pte(page_header(ent),
1198                                                            &pt[i]);
1199                         } else {
1200                                 --kvm->stat.lpages;
1201                                 rmap_remove(kvm, &pt[i]);
1202                         }
1203                 }
1204                 pt[i] = shadow_trap_nonpresent_pte;
1205         }
1206 }
1207
1208 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1209 {
1210         mmu_page_remove_parent_pte(sp, parent_pte);
1211 }
1212
1213 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1214 {
1215         int i;
1216
1217         for (i = 0; i < KVM_MAX_VCPUS; ++i)
1218                 if (kvm->vcpus[i])
1219                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
1220 }
1221
1222 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1223 {
1224         u64 *parent_pte;
1225
1226         while (sp->multimapped || sp->parent_pte) {
1227                 if (!sp->multimapped)
1228                         parent_pte = sp->parent_pte;
1229                 else {
1230                         struct kvm_pte_chain *chain;
1231
1232                         chain = container_of(sp->parent_ptes.first,
1233                                              struct kvm_pte_chain, link);
1234                         parent_pte = chain->parent_ptes[0];
1235                 }
1236                 BUG_ON(!parent_pte);
1237                 kvm_mmu_put_page(sp, parent_pte);
1238                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1239         }
1240 }
1241
1242 struct zap_walker {
1243         struct kvm_unsync_walk walker;
1244         struct kvm *kvm;
1245         int zapped;
1246 };
1247
1248 static int mmu_zap_fn(struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk)
1249 {
1250         struct zap_walker *zap_walk = container_of(walk, struct zap_walker,
1251                                                      walker);
1252         kvm_mmu_zap_page(zap_walk->kvm, sp);
1253         zap_walk->zapped = 1;
1254         return 0;
1255 }
1256
1257 static int mmu_zap_unsync_children(struct kvm *kvm, struct kvm_mmu_page *sp)
1258 {
1259         struct zap_walker walker = {
1260                 .walker = { .entry = mmu_zap_fn, },
1261                 .kvm = kvm,
1262                 .zapped = 0,
1263         };
1264
1265         if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1266                 return 0;
1267         mmu_unsync_walk(sp, &walker.walker);
1268         return walker.zapped;
1269 }
1270
1271 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1272 {
1273         int ret;
1274         ++kvm->stat.mmu_shadow_zapped;
1275         ret = mmu_zap_unsync_children(kvm, sp);
1276         kvm_mmu_page_unlink_children(kvm, sp);
1277         kvm_mmu_unlink_parents(kvm, sp);
1278         kvm_flush_remote_tlbs(kvm);
1279         if (!sp->role.invalid && !sp->role.metaphysical)
1280                 unaccount_shadowed(kvm, sp->gfn);
1281         if (sp->unsync)
1282                 kvm_unlink_unsync_page(kvm, sp);
1283         if (!sp->root_count) {
1284                 hlist_del(&sp->hash_link);
1285                 kvm_mmu_free_page(kvm, sp);
1286         } else {
1287                 sp->role.invalid = 1;
1288                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1289                 kvm_reload_remote_mmus(kvm);
1290         }
1291         kvm_mmu_reset_last_pte_updated(kvm);
1292         return ret;
1293 }
1294
1295 /*
1296  * Changing the number of mmu pages allocated to the vm
1297  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1298  */
1299 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1300 {
1301         /*
1302          * If we set the number of mmu pages to be smaller be than the
1303          * number of actived pages , we must to free some mmu pages before we
1304          * change the value
1305          */
1306
1307         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1308             kvm_nr_mmu_pages) {
1309                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1310                                        - kvm->arch.n_free_mmu_pages;
1311
1312                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1313                         struct kvm_mmu_page *page;
1314
1315                         page = container_of(kvm->arch.active_mmu_pages.prev,
1316                                             struct kvm_mmu_page, link);
1317                         kvm_mmu_zap_page(kvm, page);
1318                         n_used_mmu_pages--;
1319                 }
1320                 kvm->arch.n_free_mmu_pages = 0;
1321         }
1322         else
1323                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1324                                          - kvm->arch.n_alloc_mmu_pages;
1325
1326         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1327 }
1328
1329 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1330 {
1331         unsigned index;
1332         struct hlist_head *bucket;
1333         struct kvm_mmu_page *sp;
1334         struct hlist_node *node, *n;
1335         int r;
1336
1337         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1338         r = 0;
1339         index = kvm_page_table_hashfn(gfn);
1340         bucket = &kvm->arch.mmu_page_hash[index];
1341         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1342                 if (sp->gfn == gfn && !sp->role.metaphysical) {
1343                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1344                                  sp->role.word);
1345                         r = 1;
1346                         if (kvm_mmu_zap_page(kvm, sp))
1347                                 n = bucket->first;
1348                 }
1349         return r;
1350 }
1351
1352 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1353 {
1354         struct kvm_mmu_page *sp;
1355
1356         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1357                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1358                 kvm_mmu_zap_page(kvm, sp);
1359         }
1360 }
1361
1362 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1363 {
1364         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1365         struct kvm_mmu_page *sp = page_header(__pa(pte));
1366
1367         __set_bit(slot, sp->slot_bitmap);
1368 }
1369
1370 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1371 {
1372         int i;
1373         u64 *pt = sp->spt;
1374
1375         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1376                 return;
1377
1378         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1379                 if (pt[i] == shadow_notrap_nonpresent_pte)
1380                         set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1381         }
1382 }
1383
1384 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1385 {
1386         struct page *page;
1387
1388         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1389
1390         if (gpa == UNMAPPED_GVA)
1391                 return NULL;
1392
1393         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1394
1395         return page;
1396 }
1397
1398 /*
1399  * The function is based on mtrr_type_lookup() in
1400  * arch/x86/kernel/cpu/mtrr/generic.c
1401  */
1402 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1403                          u64 start, u64 end)
1404 {
1405         int i;
1406         u64 base, mask;
1407         u8 prev_match, curr_match;
1408         int num_var_ranges = KVM_NR_VAR_MTRR;
1409
1410         if (!mtrr_state->enabled)
1411                 return 0xFF;
1412
1413         /* Make end inclusive end, instead of exclusive */
1414         end--;
1415
1416         /* Look in fixed ranges. Just return the type as per start */
1417         if (mtrr_state->have_fixed && (start < 0x100000)) {
1418                 int idx;
1419
1420                 if (start < 0x80000) {
1421                         idx = 0;
1422                         idx += (start >> 16);
1423                         return mtrr_state->fixed_ranges[idx];
1424                 } else if (start < 0xC0000) {
1425                         idx = 1 * 8;
1426                         idx += ((start - 0x80000) >> 14);
1427                         return mtrr_state->fixed_ranges[idx];
1428                 } else if (start < 0x1000000) {
1429                         idx = 3 * 8;
1430                         idx += ((start - 0xC0000) >> 12);
1431                         return mtrr_state->fixed_ranges[idx];
1432                 }
1433         }
1434
1435         /*
1436          * Look in variable ranges
1437          * Look of multiple ranges matching this address and pick type
1438          * as per MTRR precedence
1439          */
1440         if (!(mtrr_state->enabled & 2))
1441                 return mtrr_state->def_type;
1442
1443         prev_match = 0xFF;
1444         for (i = 0; i < num_var_ranges; ++i) {
1445                 unsigned short start_state, end_state;
1446
1447                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1448                         continue;
1449
1450                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1451                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1452                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1453                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1454
1455                 start_state = ((start & mask) == (base & mask));
1456                 end_state = ((end & mask) == (base & mask));
1457                 if (start_state != end_state)
1458                         return 0xFE;
1459
1460                 if ((start & mask) != (base & mask))
1461                         continue;
1462
1463                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1464                 if (prev_match == 0xFF) {
1465                         prev_match = curr_match;
1466                         continue;
1467                 }
1468
1469                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1470                     curr_match == MTRR_TYPE_UNCACHABLE)
1471                         return MTRR_TYPE_UNCACHABLE;
1472
1473                 if ((prev_match == MTRR_TYPE_WRBACK &&
1474                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1475                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1476                      curr_match == MTRR_TYPE_WRBACK)) {
1477                         prev_match = MTRR_TYPE_WRTHROUGH;
1478                         curr_match = MTRR_TYPE_WRTHROUGH;
1479                 }
1480
1481                 if (prev_match != curr_match)
1482                         return MTRR_TYPE_UNCACHABLE;
1483         }
1484
1485         if (prev_match != 0xFF)
1486                 return prev_match;
1487
1488         return mtrr_state->def_type;
1489 }
1490
1491 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1492 {
1493         u8 mtrr;
1494
1495         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1496                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1497         if (mtrr == 0xfe || mtrr == 0xff)
1498                 mtrr = MTRR_TYPE_WRBACK;
1499         return mtrr;
1500 }
1501
1502 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1503 {
1504         unsigned index;
1505         struct hlist_head *bucket;
1506         struct kvm_mmu_page *s;
1507         struct hlist_node *node, *n;
1508
1509         index = kvm_page_table_hashfn(sp->gfn);
1510         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1511         /* don't unsync if pagetable is shadowed with multiple roles */
1512         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1513                 if (s->gfn != sp->gfn || s->role.metaphysical)
1514                         continue;
1515                 if (s->role.word != sp->role.word)
1516                         return 1;
1517         }
1518         kvm_mmu_mark_parents_unsync(vcpu, sp);
1519         ++vcpu->kvm->stat.mmu_unsync;
1520         sp->unsync = 1;
1521         mmu_convert_notrap(sp);
1522         return 0;
1523 }
1524
1525 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1526                                   bool can_unsync)
1527 {
1528         struct kvm_mmu_page *shadow;
1529
1530         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1531         if (shadow) {
1532                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1533                         return 1;
1534                 if (shadow->unsync)
1535                         return 0;
1536                 if (can_unsync && oos_shadow)
1537                         return kvm_unsync_page(vcpu, shadow);
1538                 return 1;
1539         }
1540         return 0;
1541 }
1542
1543 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1544                     unsigned pte_access, int user_fault,
1545                     int write_fault, int dirty, int largepage,
1546                     gfn_t gfn, pfn_t pfn, bool speculative,
1547                     bool can_unsync)
1548 {
1549         u64 spte;
1550         int ret = 0;
1551         u64 mt_mask = shadow_mt_mask;
1552
1553         /*
1554          * We don't set the accessed bit, since we sometimes want to see
1555          * whether the guest actually used the pte (in order to detect
1556          * demand paging).
1557          */
1558         spte = shadow_base_present_pte | shadow_dirty_mask;
1559         if (!speculative)
1560                 spte |= shadow_accessed_mask;
1561         if (!dirty)
1562                 pte_access &= ~ACC_WRITE_MASK;
1563         if (pte_access & ACC_EXEC_MASK)
1564                 spte |= shadow_x_mask;
1565         else
1566                 spte |= shadow_nx_mask;
1567         if (pte_access & ACC_USER_MASK)
1568                 spte |= shadow_user_mask;
1569         if (largepage)
1570                 spte |= PT_PAGE_SIZE_MASK;
1571         if (mt_mask) {
1572                 mt_mask = get_memory_type(vcpu, gfn) <<
1573                           kvm_x86_ops->get_mt_mask_shift();
1574                 spte |= mt_mask;
1575         }
1576
1577         spte |= (u64)pfn << PAGE_SHIFT;
1578
1579         if ((pte_access & ACC_WRITE_MASK)
1580             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1581
1582                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1583                         ret = 1;
1584                         spte = shadow_trap_nonpresent_pte;
1585                         goto set_pte;
1586                 }
1587
1588                 spte |= PT_WRITABLE_MASK;
1589
1590                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1591                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1592                                  __func__, gfn);
1593                         ret = 1;
1594                         pte_access &= ~ACC_WRITE_MASK;
1595                         if (is_writeble_pte(spte))
1596                                 spte &= ~PT_WRITABLE_MASK;
1597                 }
1598         }
1599
1600         if (pte_access & ACC_WRITE_MASK)
1601                 mark_page_dirty(vcpu->kvm, gfn);
1602
1603 set_pte:
1604         set_shadow_pte(shadow_pte, spte);
1605         return ret;
1606 }
1607
1608 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1609                          unsigned pt_access, unsigned pte_access,
1610                          int user_fault, int write_fault, int dirty,
1611                          int *ptwrite, int largepage, gfn_t gfn,
1612                          pfn_t pfn, bool speculative)
1613 {
1614         int was_rmapped = 0;
1615         int was_writeble = is_writeble_pte(*shadow_pte);
1616
1617         pgprintk("%s: spte %llx access %x write_fault %d"
1618                  " user_fault %d gfn %lx\n",
1619                  __func__, *shadow_pte, pt_access,
1620                  write_fault, user_fault, gfn);
1621
1622         if (is_rmap_pte(*shadow_pte)) {
1623                 /*
1624                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1625                  * the parent of the now unreachable PTE.
1626                  */
1627                 if (largepage && !is_large_pte(*shadow_pte)) {
1628                         struct kvm_mmu_page *child;
1629                         u64 pte = *shadow_pte;
1630
1631                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1632                         mmu_page_remove_parent_pte(child, shadow_pte);
1633                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1634                         pgprintk("hfn old %lx new %lx\n",
1635                                  spte_to_pfn(*shadow_pte), pfn);
1636                         rmap_remove(vcpu->kvm, shadow_pte);
1637                 } else {
1638                         if (largepage)
1639                                 was_rmapped = is_large_pte(*shadow_pte);
1640                         else
1641                                 was_rmapped = 1;
1642                 }
1643         }
1644         if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1645                       dirty, largepage, gfn, pfn, speculative, true)) {
1646                 if (write_fault)
1647                         *ptwrite = 1;
1648                 kvm_x86_ops->tlb_flush(vcpu);
1649         }
1650
1651         pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1652         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1653                  is_large_pte(*shadow_pte)? "2MB" : "4kB",
1654                  is_present_pte(*shadow_pte)?"RW":"R", gfn,
1655                  *shadow_pte, shadow_pte);
1656         if (!was_rmapped && is_large_pte(*shadow_pte))
1657                 ++vcpu->kvm->stat.lpages;
1658
1659         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1660         if (!was_rmapped) {
1661                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1662                 if (!is_rmap_pte(*shadow_pte))
1663                         kvm_release_pfn_clean(pfn);
1664         } else {
1665                 if (was_writeble)
1666                         kvm_release_pfn_dirty(pfn);
1667                 else
1668                         kvm_release_pfn_clean(pfn);
1669         }
1670         if (speculative) {
1671                 vcpu->arch.last_pte_updated = shadow_pte;
1672                 vcpu->arch.last_pte_gfn = gfn;
1673         }
1674 }
1675
1676 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1677 {
1678 }
1679
1680 struct direct_shadow_walk {
1681         struct kvm_shadow_walk walker;
1682         pfn_t pfn;
1683         int write;
1684         int largepage;
1685         int pt_write;
1686 };
1687
1688 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1689                             struct kvm_vcpu *vcpu,
1690                             u64 addr, u64 *sptep, int level)
1691 {
1692         struct direct_shadow_walk *walk =
1693                 container_of(_walk, struct direct_shadow_walk, walker);
1694         struct kvm_mmu_page *sp;
1695         gfn_t pseudo_gfn;
1696         gfn_t gfn = addr >> PAGE_SHIFT;
1697
1698         if (level == PT_PAGE_TABLE_LEVEL
1699             || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1700                 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1701                              0, walk->write, 1, &walk->pt_write,
1702                              walk->largepage, gfn, walk->pfn, false);
1703                 ++vcpu->stat.pf_fixed;
1704                 return 1;
1705         }
1706
1707         if (*sptep == shadow_trap_nonpresent_pte) {
1708                 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1709                 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1710                                       1, ACC_ALL, sptep);
1711                 if (!sp) {
1712                         pgprintk("nonpaging_map: ENOMEM\n");
1713                         kvm_release_pfn_clean(walk->pfn);
1714                         return -ENOMEM;
1715                 }
1716
1717                 set_shadow_pte(sptep,
1718                                __pa(sp->spt)
1719                                | PT_PRESENT_MASK | PT_WRITABLE_MASK
1720                                | shadow_user_mask | shadow_x_mask);
1721         }
1722         return 0;
1723 }
1724
1725 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1726                         int largepage, gfn_t gfn, pfn_t pfn)
1727 {
1728         int r;
1729         struct direct_shadow_walk walker = {
1730                 .walker = { .entry = direct_map_entry, },
1731                 .pfn = pfn,
1732                 .largepage = largepage,
1733                 .write = write,
1734                 .pt_write = 0,
1735         };
1736
1737         r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1738         if (r < 0)
1739                 return r;
1740         return walker.pt_write;
1741 }
1742
1743 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1744 {
1745         int r;
1746         int largepage = 0;
1747         pfn_t pfn;
1748         unsigned long mmu_seq;
1749
1750         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1751                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1752                 largepage = 1;
1753         }
1754
1755         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1756         smp_rmb();
1757         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1758
1759         /* mmio */
1760         if (is_error_pfn(pfn)) {
1761                 kvm_release_pfn_clean(pfn);
1762                 return 1;
1763         }
1764
1765         spin_lock(&vcpu->kvm->mmu_lock);
1766         if (mmu_notifier_retry(vcpu, mmu_seq))
1767                 goto out_unlock;
1768         kvm_mmu_free_some_pages(vcpu);
1769         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1770         spin_unlock(&vcpu->kvm->mmu_lock);
1771
1772
1773         return r;
1774
1775 out_unlock:
1776         spin_unlock(&vcpu->kvm->mmu_lock);
1777         kvm_release_pfn_clean(pfn);
1778         return 0;
1779 }
1780
1781
1782 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1783 {
1784         int i;
1785         struct kvm_mmu_page *sp;
1786
1787         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1788                 return;
1789         spin_lock(&vcpu->kvm->mmu_lock);
1790         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1791                 hpa_t root = vcpu->arch.mmu.root_hpa;
1792
1793                 sp = page_header(root);
1794                 --sp->root_count;
1795                 if (!sp->root_count && sp->role.invalid)
1796                         kvm_mmu_zap_page(vcpu->kvm, sp);
1797                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1798                 spin_unlock(&vcpu->kvm->mmu_lock);
1799                 return;
1800         }
1801         for (i = 0; i < 4; ++i) {
1802                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1803
1804                 if (root) {
1805                         root &= PT64_BASE_ADDR_MASK;
1806                         sp = page_header(root);
1807                         --sp->root_count;
1808                         if (!sp->root_count && sp->role.invalid)
1809                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1810                 }
1811                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1812         }
1813         spin_unlock(&vcpu->kvm->mmu_lock);
1814         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1815 }
1816
1817 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1818 {
1819         int i;
1820         gfn_t root_gfn;
1821         struct kvm_mmu_page *sp;
1822         int metaphysical = 0;
1823
1824         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1825
1826         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1827                 hpa_t root = vcpu->arch.mmu.root_hpa;
1828
1829                 ASSERT(!VALID_PAGE(root));
1830                 if (tdp_enabled)
1831                         metaphysical = 1;
1832                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1833                                       PT64_ROOT_LEVEL, metaphysical,
1834                                       ACC_ALL, NULL);
1835                 root = __pa(sp->spt);
1836                 ++sp->root_count;
1837                 vcpu->arch.mmu.root_hpa = root;
1838                 return;
1839         }
1840         metaphysical = !is_paging(vcpu);
1841         if (tdp_enabled)
1842                 metaphysical = 1;
1843         for (i = 0; i < 4; ++i) {
1844                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1845
1846                 ASSERT(!VALID_PAGE(root));
1847                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1848                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1849                                 vcpu->arch.mmu.pae_root[i] = 0;
1850                                 continue;
1851                         }
1852                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1853                 } else if (vcpu->arch.mmu.root_level == 0)
1854                         root_gfn = 0;
1855                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1856                                       PT32_ROOT_LEVEL, metaphysical,
1857                                       ACC_ALL, NULL);
1858                 root = __pa(sp->spt);
1859                 ++sp->root_count;
1860                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1861         }
1862         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1863 }
1864
1865 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
1866 {
1867         int i;
1868         struct kvm_mmu_page *sp;
1869
1870         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1871                 return;
1872         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1873                 hpa_t root = vcpu->arch.mmu.root_hpa;
1874                 sp = page_header(root);
1875                 mmu_sync_children(vcpu, sp);
1876                 return;
1877         }
1878         for (i = 0; i < 4; ++i) {
1879                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1880
1881                 if (root) {
1882                         root &= PT64_BASE_ADDR_MASK;
1883                         sp = page_header(root);
1884                         mmu_sync_children(vcpu, sp);
1885                 }
1886         }
1887 }
1888
1889 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
1890 {
1891         spin_lock(&vcpu->kvm->mmu_lock);
1892         mmu_sync_roots(vcpu);
1893         spin_unlock(&vcpu->kvm->mmu_lock);
1894 }
1895
1896 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1897 {
1898         return vaddr;
1899 }
1900
1901 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1902                                 u32 error_code)
1903 {
1904         gfn_t gfn;
1905         int r;
1906
1907         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1908         r = mmu_topup_memory_caches(vcpu);
1909         if (r)
1910                 return r;
1911
1912         ASSERT(vcpu);
1913         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1914
1915         gfn = gva >> PAGE_SHIFT;
1916
1917         return nonpaging_map(vcpu, gva & PAGE_MASK,
1918                              error_code & PFERR_WRITE_MASK, gfn);
1919 }
1920
1921 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1922                                 u32 error_code)
1923 {
1924         pfn_t pfn;
1925         int r;
1926         int largepage = 0;
1927         gfn_t gfn = gpa >> PAGE_SHIFT;
1928         unsigned long mmu_seq;
1929
1930         ASSERT(vcpu);
1931         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1932
1933         r = mmu_topup_memory_caches(vcpu);
1934         if (r)
1935                 return r;
1936
1937         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1938                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1939                 largepage = 1;
1940         }
1941         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1942         smp_rmb();
1943         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1944         if (is_error_pfn(pfn)) {
1945                 kvm_release_pfn_clean(pfn);
1946                 return 1;
1947         }
1948         spin_lock(&vcpu->kvm->mmu_lock);
1949         if (mmu_notifier_retry(vcpu, mmu_seq))
1950                 goto out_unlock;
1951         kvm_mmu_free_some_pages(vcpu);
1952         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1953                          largepage, gfn, pfn);
1954         spin_unlock(&vcpu->kvm->mmu_lock);
1955
1956         return r;
1957
1958 out_unlock:
1959         spin_unlock(&vcpu->kvm->mmu_lock);
1960         kvm_release_pfn_clean(pfn);
1961         return 0;
1962 }
1963
1964 static void nonpaging_free(struct kvm_vcpu *vcpu)
1965 {
1966         mmu_free_roots(vcpu);
1967 }
1968
1969 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1970 {
1971         struct kvm_mmu *context = &vcpu->arch.mmu;
1972
1973         context->new_cr3 = nonpaging_new_cr3;
1974         context->page_fault = nonpaging_page_fault;
1975         context->gva_to_gpa = nonpaging_gva_to_gpa;
1976         context->free = nonpaging_free;
1977         context->prefetch_page = nonpaging_prefetch_page;
1978         context->sync_page = nonpaging_sync_page;
1979         context->invlpg = nonpaging_invlpg;
1980         context->root_level = 0;
1981         context->shadow_root_level = PT32E_ROOT_LEVEL;
1982         context->root_hpa = INVALID_PAGE;
1983         return 0;
1984 }
1985
1986 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1987 {
1988         ++vcpu->stat.tlb_flush;
1989         kvm_x86_ops->tlb_flush(vcpu);
1990 }
1991
1992 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1993 {
1994         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1995         mmu_free_roots(vcpu);
1996 }
1997
1998 static void inject_page_fault(struct kvm_vcpu *vcpu,
1999                               u64 addr,
2000                               u32 err_code)
2001 {
2002         kvm_inject_page_fault(vcpu, addr, err_code);
2003 }
2004
2005 static void paging_free(struct kvm_vcpu *vcpu)
2006 {
2007         nonpaging_free(vcpu);
2008 }
2009
2010 #define PTTYPE 64
2011 #include "paging_tmpl.h"
2012 #undef PTTYPE
2013
2014 #define PTTYPE 32
2015 #include "paging_tmpl.h"
2016 #undef PTTYPE
2017
2018 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2019 {
2020         struct kvm_mmu *context = &vcpu->arch.mmu;
2021
2022         ASSERT(is_pae(vcpu));
2023         context->new_cr3 = paging_new_cr3;
2024         context->page_fault = paging64_page_fault;
2025         context->gva_to_gpa = paging64_gva_to_gpa;
2026         context->prefetch_page = paging64_prefetch_page;
2027         context->sync_page = paging64_sync_page;
2028         context->invlpg = paging64_invlpg;
2029         context->free = paging_free;
2030         context->root_level = level;
2031         context->shadow_root_level = level;
2032         context->root_hpa = INVALID_PAGE;
2033         return 0;
2034 }
2035
2036 static int paging64_init_context(struct kvm_vcpu *vcpu)
2037 {
2038         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2039 }
2040
2041 static int paging32_init_context(struct kvm_vcpu *vcpu)
2042 {
2043         struct kvm_mmu *context = &vcpu->arch.mmu;
2044
2045         context->new_cr3 = paging_new_cr3;
2046         context->page_fault = paging32_page_fault;
2047         context->gva_to_gpa = paging32_gva_to_gpa;
2048         context->free = paging_free;
2049         context->prefetch_page = paging32_prefetch_page;
2050         context->sync_page = paging32_sync_page;
2051         context->invlpg = paging32_invlpg;
2052         context->root_level = PT32_ROOT_LEVEL;
2053         context->shadow_root_level = PT32E_ROOT_LEVEL;
2054         context->root_hpa = INVALID_PAGE;
2055         return 0;
2056 }
2057
2058 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2059 {
2060         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2061 }
2062
2063 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2064 {
2065         struct kvm_mmu *context = &vcpu->arch.mmu;
2066
2067         context->new_cr3 = nonpaging_new_cr3;
2068         context->page_fault = tdp_page_fault;
2069         context->free = nonpaging_free;
2070         context->prefetch_page = nonpaging_prefetch_page;
2071         context->sync_page = nonpaging_sync_page;
2072         context->invlpg = nonpaging_invlpg;
2073         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2074         context->root_hpa = INVALID_PAGE;
2075
2076         if (!is_paging(vcpu)) {
2077                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2078                 context->root_level = 0;
2079         } else if (is_long_mode(vcpu)) {
2080                 context->gva_to_gpa = paging64_gva_to_gpa;
2081                 context->root_level = PT64_ROOT_LEVEL;
2082         } else if (is_pae(vcpu)) {
2083                 context->gva_to_gpa = paging64_gva_to_gpa;
2084                 context->root_level = PT32E_ROOT_LEVEL;
2085         } else {
2086                 context->gva_to_gpa = paging32_gva_to_gpa;
2087                 context->root_level = PT32_ROOT_LEVEL;
2088         }
2089
2090         return 0;
2091 }
2092
2093 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2094 {
2095         ASSERT(vcpu);
2096         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2097
2098         if (!is_paging(vcpu))
2099                 return nonpaging_init_context(vcpu);
2100         else if (is_long_mode(vcpu))
2101                 return paging64_init_context(vcpu);
2102         else if (is_pae(vcpu))
2103                 return paging32E_init_context(vcpu);
2104         else
2105                 return paging32_init_context(vcpu);
2106 }
2107
2108 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2109 {
2110         vcpu->arch.update_pte.pfn = bad_pfn;
2111
2112         if (tdp_enabled)
2113                 return init_kvm_tdp_mmu(vcpu);
2114         else
2115                 return init_kvm_softmmu(vcpu);
2116 }
2117
2118 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2119 {
2120         ASSERT(vcpu);
2121         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2122                 vcpu->arch.mmu.free(vcpu);
2123                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2124         }
2125 }
2126
2127 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2128 {
2129         destroy_kvm_mmu(vcpu);
2130         return init_kvm_mmu(vcpu);
2131 }
2132 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2133
2134 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2135 {
2136         int r;
2137
2138         r = mmu_topup_memory_caches(vcpu);
2139         if (r)
2140                 goto out;
2141         spin_lock(&vcpu->kvm->mmu_lock);
2142         kvm_mmu_free_some_pages(vcpu);
2143         mmu_alloc_roots(vcpu);
2144         mmu_sync_roots(vcpu);
2145         spin_unlock(&vcpu->kvm->mmu_lock);
2146         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2147         kvm_mmu_flush_tlb(vcpu);
2148 out:
2149         return r;
2150 }
2151 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2152
2153 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2154 {
2155         mmu_free_roots(vcpu);
2156 }
2157
2158 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2159                                   struct kvm_mmu_page *sp,
2160                                   u64 *spte)
2161 {
2162         u64 pte;
2163         struct kvm_mmu_page *child;
2164
2165         pte = *spte;
2166         if (is_shadow_present_pte(pte)) {
2167                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2168                     is_large_pte(pte))
2169                         rmap_remove(vcpu->kvm, spte);
2170                 else {
2171                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2172                         mmu_page_remove_parent_pte(child, spte);
2173                 }
2174         }
2175         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2176         if (is_large_pte(pte))
2177                 --vcpu->kvm->stat.lpages;
2178 }
2179
2180 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2181                                   struct kvm_mmu_page *sp,
2182                                   u64 *spte,
2183                                   const void *new)
2184 {
2185         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2186                 if (!vcpu->arch.update_pte.largepage ||
2187                     sp->role.glevels == PT32_ROOT_LEVEL) {
2188                         ++vcpu->kvm->stat.mmu_pde_zapped;
2189                         return;
2190                 }
2191         }
2192
2193         ++vcpu->kvm->stat.mmu_pte_updated;
2194         if (sp->role.glevels == PT32_ROOT_LEVEL)
2195                 paging32_update_pte(vcpu, sp, spte, new);
2196         else
2197                 paging64_update_pte(vcpu, sp, spte, new);
2198 }
2199
2200 static bool need_remote_flush(u64 old, u64 new)
2201 {
2202         if (!is_shadow_present_pte(old))
2203                 return false;
2204         if (!is_shadow_present_pte(new))
2205                 return true;
2206         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2207                 return true;
2208         old ^= PT64_NX_MASK;
2209         new ^= PT64_NX_MASK;
2210         return (old & ~new & PT64_PERM_MASK) != 0;
2211 }
2212
2213 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2214 {
2215         if (need_remote_flush(old, new))
2216                 kvm_flush_remote_tlbs(vcpu->kvm);
2217         else
2218                 kvm_mmu_flush_tlb(vcpu);
2219 }
2220
2221 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2222 {
2223         u64 *spte = vcpu->arch.last_pte_updated;
2224
2225         return !!(spte && (*spte & shadow_accessed_mask));
2226 }
2227
2228 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2229                                           const u8 *new, int bytes)
2230 {
2231         gfn_t gfn;
2232         int r;
2233         u64 gpte = 0;
2234         pfn_t pfn;
2235
2236         vcpu->arch.update_pte.largepage = 0;
2237
2238         if (bytes != 4 && bytes != 8)
2239                 return;
2240
2241         /*
2242          * Assume that the pte write on a page table of the same type
2243          * as the current vcpu paging mode.  This is nearly always true
2244          * (might be false while changing modes).  Note it is verified later
2245          * by update_pte().
2246          */
2247         if (is_pae(vcpu)) {
2248                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2249                 if ((bytes == 4) && (gpa % 4 == 0)) {
2250                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2251                         if (r)
2252                                 return;
2253                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2254                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2255                         memcpy((void *)&gpte, new, 8);
2256                 }
2257         } else {
2258                 if ((bytes == 4) && (gpa % 4 == 0))
2259                         memcpy((void *)&gpte, new, 4);
2260         }
2261         if (!is_present_pte(gpte))
2262                 return;
2263         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2264
2265         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2266                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2267                 vcpu->arch.update_pte.largepage = 1;
2268         }
2269         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2270         smp_rmb();
2271         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2272
2273         if (is_error_pfn(pfn)) {
2274                 kvm_release_pfn_clean(pfn);
2275                 return;
2276         }
2277         vcpu->arch.update_pte.gfn = gfn;
2278         vcpu->arch.update_pte.pfn = pfn;
2279 }
2280
2281 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2282 {
2283         u64 *spte = vcpu->arch.last_pte_updated;
2284
2285         if (spte
2286             && vcpu->arch.last_pte_gfn == gfn
2287             && shadow_accessed_mask
2288             && !(*spte & shadow_accessed_mask)
2289             && is_shadow_present_pte(*spte))
2290                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2291 }
2292
2293 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2294                        const u8 *new, int bytes)
2295 {
2296         gfn_t gfn = gpa >> PAGE_SHIFT;
2297         struct kvm_mmu_page *sp;
2298         struct hlist_node *node, *n;
2299         struct hlist_head *bucket;
2300         unsigned index;
2301         u64 entry, gentry;
2302         u64 *spte;
2303         unsigned offset = offset_in_page(gpa);
2304         unsigned pte_size;
2305         unsigned page_offset;
2306         unsigned misaligned;
2307         unsigned quadrant;
2308         int level;
2309         int flooded = 0;
2310         int npte;
2311         int r;
2312
2313         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2314         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2315         spin_lock(&vcpu->kvm->mmu_lock);
2316         kvm_mmu_access_page(vcpu, gfn);
2317         kvm_mmu_free_some_pages(vcpu);
2318         ++vcpu->kvm->stat.mmu_pte_write;
2319         kvm_mmu_audit(vcpu, "pre pte write");
2320         if (gfn == vcpu->arch.last_pt_write_gfn
2321             && !last_updated_pte_accessed(vcpu)) {
2322                 ++vcpu->arch.last_pt_write_count;
2323                 if (vcpu->arch.last_pt_write_count >= 3)
2324                         flooded = 1;
2325         } else {
2326                 vcpu->arch.last_pt_write_gfn = gfn;
2327                 vcpu->arch.last_pt_write_count = 1;
2328                 vcpu->arch.last_pte_updated = NULL;
2329         }
2330         index = kvm_page_table_hashfn(gfn);
2331         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2332         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2333                 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
2334                         continue;
2335                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2336                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2337                 misaligned |= bytes < 4;
2338                 if (misaligned || flooded) {
2339                         /*
2340                          * Misaligned accesses are too much trouble to fix
2341                          * up; also, they usually indicate a page is not used
2342                          * as a page table.
2343                          *
2344                          * If we're seeing too many writes to a page,
2345                          * it may no longer be a page table, or we may be
2346                          * forking, in which case it is better to unmap the
2347                          * page.
2348                          */
2349                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2350                                  gpa, bytes, sp->role.word);
2351                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2352                                 n = bucket->first;
2353                         ++vcpu->kvm->stat.mmu_flooded;
2354                         continue;
2355                 }
2356                 page_offset = offset;
2357                 level = sp->role.level;
2358                 npte = 1;
2359                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2360                         page_offset <<= 1;      /* 32->64 */
2361                         /*
2362                          * A 32-bit pde maps 4MB while the shadow pdes map
2363                          * only 2MB.  So we need to double the offset again
2364                          * and zap two pdes instead of one.
2365                          */
2366                         if (level == PT32_ROOT_LEVEL) {
2367                                 page_offset &= ~7; /* kill rounding error */
2368                                 page_offset <<= 1;
2369                                 npte = 2;
2370                         }
2371                         quadrant = page_offset >> PAGE_SHIFT;
2372                         page_offset &= ~PAGE_MASK;
2373                         if (quadrant != sp->role.quadrant)
2374                                 continue;
2375                 }
2376                 spte = &sp->spt[page_offset / sizeof(*spte)];
2377                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2378                         gentry = 0;
2379                         r = kvm_read_guest_atomic(vcpu->kvm,
2380                                                   gpa & ~(u64)(pte_size - 1),
2381                                                   &gentry, pte_size);
2382                         new = (const void *)&gentry;
2383                         if (r < 0)
2384                                 new = NULL;
2385                 }
2386                 while (npte--) {
2387                         entry = *spte;
2388                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2389                         if (new)
2390                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2391                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2392                         ++spte;
2393                 }
2394         }
2395         kvm_mmu_audit(vcpu, "post pte write");
2396         spin_unlock(&vcpu->kvm->mmu_lock);
2397         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2398                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2399                 vcpu->arch.update_pte.pfn = bad_pfn;
2400         }
2401 }
2402
2403 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2404 {
2405         gpa_t gpa;
2406         int r;
2407
2408         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2409
2410         spin_lock(&vcpu->kvm->mmu_lock);
2411         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2412         spin_unlock(&vcpu->kvm->mmu_lock);
2413         return r;
2414 }
2415 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2416
2417 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2418 {
2419         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2420                 struct kvm_mmu_page *sp;
2421
2422                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2423                                   struct kvm_mmu_page, link);
2424                 kvm_mmu_zap_page(vcpu->kvm, sp);
2425                 ++vcpu->kvm->stat.mmu_recycled;
2426         }
2427 }
2428
2429 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2430 {
2431         int r;
2432         enum emulation_result er;
2433
2434         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2435         if (r < 0)
2436                 goto out;
2437
2438         if (!r) {
2439                 r = 1;
2440                 goto out;
2441         }
2442
2443         r = mmu_topup_memory_caches(vcpu);
2444         if (r)
2445                 goto out;
2446
2447         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2448
2449         switch (er) {
2450         case EMULATE_DONE:
2451                 return 1;
2452         case EMULATE_DO_MMIO:
2453                 ++vcpu->stat.mmio_exits;
2454                 return 0;
2455         case EMULATE_FAIL:
2456                 kvm_report_emulation_failure(vcpu, "pagetable");
2457                 return 1;
2458         default:
2459                 BUG();
2460         }
2461 out:
2462         return r;
2463 }
2464 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2465
2466 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2467 {
2468         spin_lock(&vcpu->kvm->mmu_lock);
2469         vcpu->arch.mmu.invlpg(vcpu, gva);
2470         spin_unlock(&vcpu->kvm->mmu_lock);
2471         kvm_mmu_flush_tlb(vcpu);
2472         ++vcpu->stat.invlpg;
2473 }
2474 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2475
2476 void kvm_enable_tdp(void)
2477 {
2478         tdp_enabled = true;
2479 }
2480 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2481
2482 void kvm_disable_tdp(void)
2483 {
2484         tdp_enabled = false;
2485 }
2486 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2487
2488 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2489 {
2490         struct kvm_mmu_page *sp;
2491
2492         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2493                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2494                                   struct kvm_mmu_page, link);
2495                 kvm_mmu_zap_page(vcpu->kvm, sp);
2496                 cond_resched();
2497         }
2498         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2499 }
2500
2501 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2502 {
2503         struct page *page;
2504         int i;
2505
2506         ASSERT(vcpu);
2507
2508         if (vcpu->kvm->arch.n_requested_mmu_pages)
2509                 vcpu->kvm->arch.n_free_mmu_pages =
2510                                         vcpu->kvm->arch.n_requested_mmu_pages;
2511         else
2512                 vcpu->kvm->arch.n_free_mmu_pages =
2513                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2514         /*
2515          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2516          * Therefore we need to allocate shadow page tables in the first
2517          * 4GB of memory, which happens to fit the DMA32 zone.
2518          */
2519         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2520         if (!page)
2521                 goto error_1;
2522         vcpu->arch.mmu.pae_root = page_address(page);
2523         for (i = 0; i < 4; ++i)
2524                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2525
2526         return 0;
2527
2528 error_1:
2529         free_mmu_pages(vcpu);
2530         return -ENOMEM;
2531 }
2532
2533 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2534 {
2535         ASSERT(vcpu);
2536         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2537
2538         return alloc_mmu_pages(vcpu);
2539 }
2540
2541 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2542 {
2543         ASSERT(vcpu);
2544         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2545
2546         return init_kvm_mmu(vcpu);
2547 }
2548
2549 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2550 {
2551         ASSERT(vcpu);
2552
2553         destroy_kvm_mmu(vcpu);
2554         free_mmu_pages(vcpu);
2555         mmu_free_memory_caches(vcpu);
2556 }
2557
2558 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2559 {
2560         struct kvm_mmu_page *sp;
2561
2562         spin_lock(&kvm->mmu_lock);
2563         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2564                 int i;
2565                 u64 *pt;
2566
2567                 if (!test_bit(slot, sp->slot_bitmap))
2568                         continue;
2569
2570                 pt = sp->spt;
2571                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2572                         /* avoid RMW */
2573                         if (pt[i] & PT_WRITABLE_MASK)
2574                                 pt[i] &= ~PT_WRITABLE_MASK;
2575         }
2576         kvm_flush_remote_tlbs(kvm);
2577         spin_unlock(&kvm->mmu_lock);
2578 }
2579
2580 void kvm_mmu_zap_all(struct kvm *kvm)
2581 {
2582         struct kvm_mmu_page *sp, *node;
2583
2584         spin_lock(&kvm->mmu_lock);
2585         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2586                 if (kvm_mmu_zap_page(kvm, sp))
2587                         node = container_of(kvm->arch.active_mmu_pages.next,
2588                                             struct kvm_mmu_page, link);
2589         spin_unlock(&kvm->mmu_lock);
2590
2591         kvm_flush_remote_tlbs(kvm);
2592 }
2593
2594 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2595 {
2596         struct kvm_mmu_page *page;
2597
2598         page = container_of(kvm->arch.active_mmu_pages.prev,
2599                             struct kvm_mmu_page, link);
2600         kvm_mmu_zap_page(kvm, page);
2601 }
2602
2603 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2604 {
2605         struct kvm *kvm;
2606         struct kvm *kvm_freed = NULL;
2607         int cache_count = 0;
2608
2609         spin_lock(&kvm_lock);
2610
2611         list_for_each_entry(kvm, &vm_list, vm_list) {
2612                 int npages;
2613
2614                 if (!down_read_trylock(&kvm->slots_lock))
2615                         continue;
2616                 spin_lock(&kvm->mmu_lock);
2617                 npages = kvm->arch.n_alloc_mmu_pages -
2618                          kvm->arch.n_free_mmu_pages;
2619                 cache_count += npages;
2620                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2621                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2622                         cache_count--;
2623                         kvm_freed = kvm;
2624                 }
2625                 nr_to_scan--;
2626
2627                 spin_unlock(&kvm->mmu_lock);
2628                 up_read(&kvm->slots_lock);
2629         }
2630         if (kvm_freed)
2631                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2632
2633         spin_unlock(&kvm_lock);
2634
2635         return cache_count;
2636 }
2637
2638 static struct shrinker mmu_shrinker = {
2639         .shrink = mmu_shrink,
2640         .seeks = DEFAULT_SEEKS * 10,
2641 };
2642
2643 static void mmu_destroy_caches(void)
2644 {
2645         if (pte_chain_cache)
2646                 kmem_cache_destroy(pte_chain_cache);
2647         if (rmap_desc_cache)
2648                 kmem_cache_destroy(rmap_desc_cache);
2649         if (mmu_page_header_cache)
2650                 kmem_cache_destroy(mmu_page_header_cache);
2651 }
2652
2653 void kvm_mmu_module_exit(void)
2654 {
2655         mmu_destroy_caches();
2656         unregister_shrinker(&mmu_shrinker);
2657 }
2658
2659 int kvm_mmu_module_init(void)
2660 {
2661         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2662                                             sizeof(struct kvm_pte_chain),
2663                                             0, 0, NULL);
2664         if (!pte_chain_cache)
2665                 goto nomem;
2666         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2667                                             sizeof(struct kvm_rmap_desc),
2668                                             0, 0, NULL);
2669         if (!rmap_desc_cache)
2670                 goto nomem;
2671
2672         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2673                                                   sizeof(struct kvm_mmu_page),
2674                                                   0, 0, NULL);
2675         if (!mmu_page_header_cache)
2676                 goto nomem;
2677
2678         register_shrinker(&mmu_shrinker);
2679
2680         return 0;
2681
2682 nomem:
2683         mmu_destroy_caches();
2684         return -ENOMEM;
2685 }
2686
2687 /*
2688  * Caculate mmu pages needed for kvm.
2689  */
2690 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2691 {
2692         int i;
2693         unsigned int nr_mmu_pages;
2694         unsigned int  nr_pages = 0;
2695
2696         for (i = 0; i < kvm->nmemslots; i++)
2697                 nr_pages += kvm->memslots[i].npages;
2698
2699         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2700         nr_mmu_pages = max(nr_mmu_pages,
2701                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2702
2703         return nr_mmu_pages;
2704 }
2705
2706 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2707                                 unsigned len)
2708 {
2709         if (len > buffer->len)
2710                 return NULL;
2711         return buffer->ptr;
2712 }
2713
2714 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2715                                 unsigned len)
2716 {
2717         void *ret;
2718
2719         ret = pv_mmu_peek_buffer(buffer, len);
2720         if (!ret)
2721                 return ret;
2722         buffer->ptr += len;
2723         buffer->len -= len;
2724         buffer->processed += len;
2725         return ret;
2726 }
2727
2728 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2729                              gpa_t addr, gpa_t value)
2730 {
2731         int bytes = 8;
2732         int r;
2733
2734         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2735                 bytes = 4;
2736
2737         r = mmu_topup_memory_caches(vcpu);
2738         if (r)
2739                 return r;
2740
2741         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2742                 return -EFAULT;
2743
2744         return 1;
2745 }
2746
2747 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2748 {
2749         kvm_x86_ops->tlb_flush(vcpu);
2750         set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
2751         return 1;
2752 }
2753
2754 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2755 {
2756         spin_lock(&vcpu->kvm->mmu_lock);
2757         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2758         spin_unlock(&vcpu->kvm->mmu_lock);
2759         return 1;
2760 }
2761
2762 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2763                              struct kvm_pv_mmu_op_buffer *buffer)
2764 {
2765         struct kvm_mmu_op_header *header;
2766
2767         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2768         if (!header)
2769                 return 0;
2770         switch (header->op) {
2771         case KVM_MMU_OP_WRITE_PTE: {
2772                 struct kvm_mmu_op_write_pte *wpte;
2773
2774                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2775                 if (!wpte)
2776                         return 0;
2777                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2778                                         wpte->pte_val);
2779         }
2780         case KVM_MMU_OP_FLUSH_TLB: {
2781                 struct kvm_mmu_op_flush_tlb *ftlb;
2782
2783                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2784                 if (!ftlb)
2785                         return 0;
2786                 return kvm_pv_mmu_flush_tlb(vcpu);
2787         }
2788         case KVM_MMU_OP_RELEASE_PT: {
2789                 struct kvm_mmu_op_release_pt *rpt;
2790
2791                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2792                 if (!rpt)
2793                         return 0;
2794                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2795         }
2796         default: return 0;
2797         }
2798 }
2799
2800 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2801                   gpa_t addr, unsigned long *ret)
2802 {
2803         int r;
2804         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2805
2806         buffer->ptr = buffer->buf;
2807         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2808         buffer->processed = 0;
2809
2810         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2811         if (r)
2812                 goto out;
2813
2814         while (buffer->len) {
2815                 r = kvm_pv_mmu_op_one(vcpu, buffer);
2816                 if (r < 0)
2817                         goto out;
2818                 if (r == 0)
2819                         break;
2820         }
2821
2822         r = 1;
2823 out:
2824         *ret = buffer->processed;
2825         return r;
2826 }
2827
2828 #ifdef AUDIT
2829
2830 static const char *audit_msg;
2831
2832 static gva_t canonicalize(gva_t gva)
2833 {
2834 #ifdef CONFIG_X86_64
2835         gva = (long long)(gva << 16) >> 16;
2836 #endif
2837         return gva;
2838 }
2839
2840 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2841                                 gva_t va, int level)
2842 {
2843         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2844         int i;
2845         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2846
2847         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2848                 u64 ent = pt[i];
2849
2850                 if (ent == shadow_trap_nonpresent_pte)
2851                         continue;
2852
2853                 va = canonicalize(va);
2854                 if (level > 1) {
2855                         if (ent == shadow_notrap_nonpresent_pte)
2856                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2857                                        " in nonleaf level: levels %d gva %lx"
2858                                        " level %d pte %llx\n", audit_msg,
2859                                        vcpu->arch.mmu.root_level, va, level, ent);
2860
2861                         audit_mappings_page(vcpu, ent, va, level - 1);
2862                 } else {
2863                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2864                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2865
2866                         if (is_shadow_present_pte(ent)
2867                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2868                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2869                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2870                                        audit_msg, vcpu->arch.mmu.root_level,
2871                                        va, gpa, hpa, ent,
2872                                        is_shadow_present_pte(ent));
2873                         else if (ent == shadow_notrap_nonpresent_pte
2874                                  && !is_error_hpa(hpa))
2875                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2876                                        " valid guest gva %lx\n", audit_msg, va);
2877                         kvm_release_pfn_clean(pfn);
2878
2879                 }
2880         }
2881 }
2882
2883 static void audit_mappings(struct kvm_vcpu *vcpu)
2884 {
2885         unsigned i;
2886
2887         if (vcpu->arch.mmu.root_level == 4)
2888                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2889         else
2890                 for (i = 0; i < 4; ++i)
2891                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2892                                 audit_mappings_page(vcpu,
2893                                                     vcpu->arch.mmu.pae_root[i],
2894                                                     i << 30,
2895                                                     2);
2896 }
2897
2898 static int count_rmaps(struct kvm_vcpu *vcpu)
2899 {
2900         int nmaps = 0;
2901         int i, j, k;
2902
2903         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2904                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2905                 struct kvm_rmap_desc *d;
2906
2907                 for (j = 0; j < m->npages; ++j) {
2908                         unsigned long *rmapp = &m->rmap[j];
2909
2910                         if (!*rmapp)
2911                                 continue;
2912                         if (!(*rmapp & 1)) {
2913                                 ++nmaps;
2914                                 continue;
2915                         }
2916                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2917                         while (d) {
2918                                 for (k = 0; k < RMAP_EXT; ++k)
2919                                         if (d->shadow_ptes[k])
2920                                                 ++nmaps;
2921                                         else
2922                                                 break;
2923                                 d = d->more;
2924                         }
2925                 }
2926         }
2927         return nmaps;
2928 }
2929
2930 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2931 {
2932         int nmaps = 0;
2933         struct kvm_mmu_page *sp;
2934         int i;
2935
2936         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2937                 u64 *pt = sp->spt;
2938
2939                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2940                         continue;
2941
2942                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2943                         u64 ent = pt[i];
2944
2945                         if (!(ent & PT_PRESENT_MASK))
2946                                 continue;
2947                         if (!(ent & PT_WRITABLE_MASK))
2948                                 continue;
2949                         ++nmaps;
2950                 }
2951         }
2952         return nmaps;
2953 }
2954
2955 static void audit_rmap(struct kvm_vcpu *vcpu)
2956 {
2957         int n_rmap = count_rmaps(vcpu);
2958         int n_actual = count_writable_mappings(vcpu);
2959
2960         if (n_rmap != n_actual)
2961                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2962                        __func__, audit_msg, n_rmap, n_actual);
2963 }
2964
2965 static void audit_write_protection(struct kvm_vcpu *vcpu)
2966 {
2967         struct kvm_mmu_page *sp;
2968         struct kvm_memory_slot *slot;
2969         unsigned long *rmapp;
2970         gfn_t gfn;
2971
2972         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2973                 if (sp->role.metaphysical)
2974                         continue;
2975
2976                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2977                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2978                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2979                 if (*rmapp)
2980                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2981                                " mappings: gfn %lx role %x\n",
2982                                __func__, audit_msg, sp->gfn,
2983                                sp->role.word);
2984         }
2985 }
2986
2987 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2988 {
2989         int olddbg = dbg;
2990
2991         dbg = 0;
2992         audit_msg = msg;
2993         audit_rmap(vcpu);
2994         audit_write_protection(vcpu);
2995         audit_mappings(vcpu);
2996         dbg = olddbg;
2997 }
2998
2999 #endif