git.oblomov.eu Git - linux-2.6/blob - arch/x86/kvm/mmu.c

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