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

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "mmu.h"
  21
  22 #include <linux/kvm_host.h>
  23 #include <linux/types.h>
  24 #include <linux/string.h>
  25 #include <linux/mm.h>
  26 #include <linux/highmem.h>
  27 #include <linux/module.h>
  28 #include <linux/swap.h>
  29 #include <linux/hugetlb.h>
  30 #include <linux/compiler.h>
  31
  32 #include <asm/page.h>
  33 #include <asm/cmpxchg.h>
  34 #include <asm/io.h>
  35 #include <asm/vmx.h>
  36
  37 /*
  38  * When setting this variable to true it enables Two-Dimensional-Paging
  39  * where the hardware walks 2 page tables:
  40  * 1. the guest-virtual to guest-physical
  41  * 2. while doing 1. it walks guest-physical to host-physical
  42  * If the hardware supports that we don't need to do shadow paging.
  43  */
  44 bool tdp_enabled = false;
  45
  46 #undef MMU_DEBUG
  47
  48 #undef AUDIT
  49
  50 #ifdef AUDIT
  51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  52 #else
  53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  54 #endif
  55
  56 #ifdef MMU_DEBUG
  57
  58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  60
  61 #else
  62
  63 #define pgprintk(x...) do { } while (0)
  64 #define rmap_printk(x...) do { } while (0)
  65
  66 #endif
  67
  68 #if defined(MMU_DEBUG) || defined(AUDIT)
  69 static int dbg = 0;
  70 module_param(dbg, bool, 0644);
  71 #endif
  72
  73 static int oos_shadow = 1;
  74 module_param(oos_shadow, bool, 0644);
  75
  76 #ifndef MMU_DEBUG
  77 #define ASSERT(x) do { } while (0)
  78 #else
  79 #define ASSERT(x)                                                       \
  80         if (!(x)) {                                                     \
  81                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  82                        __FILE__, __LINE__, #x);                         \
  83         }
  84 #endif
  85
  86 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  87 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  88
  89 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  90
  91 #define PT64_LEVEL_BITS 9
  92
  93 #define PT64_LEVEL_SHIFT(level) \
  94                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  95
  96 #define PT64_LEVEL_MASK(level) \
  97                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  98
  99 #define PT64_INDEX(address, level)\
 100         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 101
 102
 103 #define PT32_LEVEL_BITS 10
 104
 105 #define PT32_LEVEL_SHIFT(level) \
 106                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 107
 108 #define PT32_LEVEL_MASK(level) \
 109                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 110
 111 #define PT32_INDEX(address, level)\
 112         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 113
 114
 115 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 116 #define PT64_DIR_BASE_ADDR_MASK \
 117         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 118
 119 #define PT32_BASE_ADDR_MASK PAGE_MASK
 120 #define PT32_DIR_BASE_ADDR_MASK \
 121         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 122
 123 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 124                         | PT64_NX_MASK)
 125
 126 #define PFERR_PRESENT_MASK (1U << 0)
 127 #define PFERR_WRITE_MASK (1U << 1)
 128 #define PFERR_USER_MASK (1U << 2)
 129 #define PFERR_FETCH_MASK (1U << 4)
 130
 131 #define PT_DIRECTORY_LEVEL 2
 132 #define PT_PAGE_TABLE_LEVEL 1
 133
 134 #define RMAP_EXT 4
 135
 136 #define ACC_EXEC_MASK    1
 137 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 138 #define ACC_USER_MASK    PT_USER_MASK
 139 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 140
 141 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 142
 143 struct kvm_rmap_desc {
 144         u64 *shadow_ptes[RMAP_EXT];
 145         struct kvm_rmap_desc *more;
 146 };
 147
 148 struct kvm_shadow_walk {
 149         int (*entry)(struct kvm_shadow_walk *walk, struct kvm_vcpu *vcpu,
 150                      u64 addr, u64 *spte, int level);
 151 };
 152
 153 struct kvm_unsync_walk {
 154         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 155 };
 156
 157 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 158
 159 static struct kmem_cache *pte_chain_cache;
 160 static struct kmem_cache *rmap_desc_cache;
 161 static struct kmem_cache *mmu_page_header_cache;
 162
 163 static u64 __read_mostly shadow_trap_nonpresent_pte;
 164 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 165 static u64 __read_mostly shadow_base_present_pte;
 166 static u64 __read_mostly shadow_nx_mask;
 167 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 168 static u64 __read_mostly shadow_user_mask;
 169 static u64 __read_mostly shadow_accessed_mask;
 170 static u64 __read_mostly shadow_dirty_mask;
 171 static u64 __read_mostly shadow_mt_mask;
 172
 173 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 174 {
 175         shadow_trap_nonpresent_pte = trap_pte;
 176         shadow_notrap_nonpresent_pte = notrap_pte;
 177 }
 178 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 179
 180 void kvm_mmu_set_base_ptes(u64 base_pte)
 181 {
 182         shadow_base_present_pte = base_pte;
 183 }
 184 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 185
 186 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 187                 u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 mt_mask)
 188 {
 189         shadow_user_mask = user_mask;
 190         shadow_accessed_mask = accessed_mask;
 191         shadow_dirty_mask = dirty_mask;
 192         shadow_nx_mask = nx_mask;
 193         shadow_x_mask = x_mask;
 194         shadow_mt_mask = mt_mask;
 195 }
 196 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 197
 198 static int is_write_protection(struct kvm_vcpu *vcpu)
 199 {
 200         return vcpu->arch.cr0 & X86_CR0_WP;
 201 }
 202
 203 static int is_cpuid_PSE36(void)
 204 {
 205         return 1;
 206 }
 207
 208 static int is_nx(struct kvm_vcpu *vcpu)
 209 {
 210         return vcpu->arch.shadow_efer & EFER_NX;
 211 }
 212
 213 static int is_present_pte(unsigned long pte)
 214 {
 215         return pte & PT_PRESENT_MASK;
 216 }
 217
 218 static int is_shadow_present_pte(u64 pte)
 219 {
 220         return pte != shadow_trap_nonpresent_pte
 221                 && pte != shadow_notrap_nonpresent_pte;
 222 }
 223
 224 static int is_large_pte(u64 pte)
 225 {
 226         return pte & PT_PAGE_SIZE_MASK;
 227 }
 228
 229 static int is_writeble_pte(unsigned long pte)
 230 {
 231         return pte & PT_WRITABLE_MASK;
 232 }
 233
 234 static int is_dirty_pte(unsigned long pte)
 235 {
 236         return pte & shadow_dirty_mask;
 237 }
 238
 239 static int is_rmap_pte(u64 pte)
 240 {
 241         return is_shadow_present_pte(pte);
 242 }
 243
 244 static pfn_t spte_to_pfn(u64 pte)
 245 {
 246         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 247 }
 248
 249 static gfn_t pse36_gfn_delta(u32 gpte)
 250 {
 251         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 252
 253         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 254 }
 255
 256 static void set_shadow_pte(u64 *sptep, u64 spte)
 257 {
 258 #ifdef CONFIG_X86_64
 259         set_64bit((unsigned long *)sptep, spte);
 260 #else
 261         set_64bit((unsigned long long *)sptep, spte);
 262 #endif
 263 }
 264
 265 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 266                                   struct kmem_cache *base_cache, int min)
 267 {
 268         void *obj;
 269
 270         if (cache->nobjs >= min)
 271                 return 0;
 272         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 273                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 274                 if (!obj)
 275                         return -ENOMEM;
 276                 cache->objects[cache->nobjs++] = obj;
 277         }
 278         return 0;
 279 }
 280
 281 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 282 {
 283         while (mc->nobjs)
 284                 kfree(mc->objects[--mc->nobjs]);
 285 }
 286
 287 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 288                                        int min)
 289 {
 290         struct page *page;
 291
 292         if (cache->nobjs >= min)
 293                 return 0;
 294         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 295                 page = alloc_page(GFP_KERNEL);
 296                 if (!page)
 297                         return -ENOMEM;
 298                 set_page_private(page, 0);
 299                 cache->objects[cache->nobjs++] = page_address(page);
 300         }
 301         return 0;
 302 }
 303
 304 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 305 {
 306         while (mc->nobjs)
 307                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 308 }
 309
 310 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 311 {
 312         int r;
 313
 314         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 315                                    pte_chain_cache, 4);
 316         if (r)
 317                 goto out;
 318         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 319                                    rmap_desc_cache, 4);
 320         if (r)
 321                 goto out;
 322         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 323         if (r)
 324                 goto out;
 325         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 326                                    mmu_page_header_cache, 4);
 327 out:
 328         return r;
 329 }
 330
 331 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 332 {
 333         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 334         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 335         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 336         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 337 }
 338
 339 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 340                                     size_t size)
 341 {
 342         void *p;
 343
 344         BUG_ON(!mc->nobjs);
 345         p = mc->objects[--mc->nobjs];
 346         memset(p, 0, size);
 347         return p;
 348 }
 349
 350 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 351 {
 352         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 353                                       sizeof(struct kvm_pte_chain));
 354 }
 355
 356 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 357 {
 358         kfree(pc);
 359 }
 360
 361 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 362 {
 363         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 364                                       sizeof(struct kvm_rmap_desc));
 365 }
 366
 367 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 368 {
 369         kfree(rd);
 370 }
 371
 372 /*
 373  * Return the pointer to the largepage write count for a given
 374  * gfn, handling slots that are not large page aligned.
 375  */
 376 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 377 {
 378         unsigned long idx;
 379
 380         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 381               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 382         return &slot->lpage_info[idx].write_count;
 383 }
 384
 385 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 386 {
 387         int *write_count;
 388
 389         gfn = unalias_gfn(kvm, gfn);
 390         write_count = slot_largepage_idx(gfn,
 391                                          gfn_to_memslot_unaliased(kvm, gfn));
 392         *write_count += 1;
 393 }
 394
 395 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 396 {
 397         int *write_count;
 398
 399         gfn = unalias_gfn(kvm, gfn);
 400         write_count = slot_largepage_idx(gfn,
 401                                          gfn_to_memslot_unaliased(kvm, gfn));
 402         *write_count -= 1;
 403         WARN_ON(*write_count < 0);
 404 }
 405
 406 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 407 {
 408         struct kvm_memory_slot *slot;
 409         int *largepage_idx;
 410
 411         gfn = unalias_gfn(kvm, gfn);
 412         slot = gfn_to_memslot_unaliased(kvm, gfn);
 413         if (slot) {
 414                 largepage_idx = slot_largepage_idx(gfn, slot);
 415                 return *largepage_idx;
 416         }
 417
 418         return 1;
 419 }
 420
 421 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 422 {
 423         struct vm_area_struct *vma;
 424         unsigned long addr;
 425         int ret = 0;
 426
 427         addr = gfn_to_hva(kvm, gfn);
 428         if (kvm_is_error_hva(addr))
 429                 return ret;
 430
 431         down_read(&current->mm->mmap_sem);
 432         vma = find_vma(current->mm, addr);
 433         if (vma && is_vm_hugetlb_page(vma))
 434                 ret = 1;
 435         up_read(&current->mm->mmap_sem);
 436
 437         return ret;
 438 }
 439
 440 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 441 {
 442         struct kvm_memory_slot *slot;
 443
 444         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 445                 return 0;
 446
 447         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 448                 return 0;
 449
 450         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 451         if (slot && slot->dirty_bitmap)
 452                 return 0;
 453
 454         return 1;
 455 }
 456
 457 /*
 458  * Take gfn and return the reverse mapping to it.
 459  * Note: gfn must be unaliased before this function get called
 460  */
 461
 462 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 463 {
 464         struct kvm_memory_slot *slot;
 465         unsigned long idx;
 466
 467         slot = gfn_to_memslot(kvm, gfn);
 468         if (!lpage)
 469                 return &slot->rmap[gfn - slot->base_gfn];
 470
 471         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 472               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 473
 474         return &slot->lpage_info[idx].rmap_pde;
 475 }
 476
 477 /*
 478  * Reverse mapping data structures:
 479  *
 480  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 481  * that points to page_address(page).
 482  *
 483  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 484  * containing more mappings.
 485  */
 486 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 487 {
 488         struct kvm_mmu_page *sp;
 489         struct kvm_rmap_desc *desc;
 490         unsigned long *rmapp;
 491         int i;
 492
 493         if (!is_rmap_pte(*spte))
 494                 return;
 495         gfn = unalias_gfn(vcpu->kvm, gfn);
 496         sp = page_header(__pa(spte));
 497         sp->gfns[spte - sp->spt] = gfn;
 498         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 499         if (!*rmapp) {
 500                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 501                 *rmapp = (unsigned long)spte;
 502         } else if (!(*rmapp & 1)) {
 503                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 504                 desc = mmu_alloc_rmap_desc(vcpu);
 505                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 506                 desc->shadow_ptes[1] = spte;
 507                 *rmapp = (unsigned long)desc | 1;
 508         } else {
 509                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 510                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 511                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 512                         desc = desc->more;
 513                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 514                         desc->more = mmu_alloc_rmap_desc(vcpu);
 515                         desc = desc->more;
 516                 }
 517                 for (i = 0; desc->shadow_ptes[i]; ++i)
 518                         ;
 519                 desc->shadow_ptes[i] = spte;
 520         }
 521 }
 522
 523 static void rmap_desc_remove_entry(unsigned long *rmapp,
 524                                    struct kvm_rmap_desc *desc,
 525                                    int i,
 526                                    struct kvm_rmap_desc *prev_desc)
 527 {
 528         int j;
 529
 530         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 531                 ;
 532         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 533         desc->shadow_ptes[j] = NULL;
 534         if (j != 0)
 535                 return;
 536         if (!prev_desc && !desc->more)
 537                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 538         else
 539                 if (prev_desc)
 540                         prev_desc->more = desc->more;
 541                 else
 542                         *rmapp = (unsigned long)desc->more | 1;
 543         mmu_free_rmap_desc(desc);
 544 }
 545
 546 static void rmap_remove(struct kvm *kvm, u64 *spte)
 547 {
 548         struct kvm_rmap_desc *desc;
 549         struct kvm_rmap_desc *prev_desc;
 550         struct kvm_mmu_page *sp;
 551         pfn_t pfn;
 552         unsigned long *rmapp;
 553         int i;
 554
 555         if (!is_rmap_pte(*spte))
 556                 return;
 557         sp = page_header(__pa(spte));
 558         pfn = spte_to_pfn(*spte);
 559         if (*spte & shadow_accessed_mask)
 560                 kvm_set_pfn_accessed(pfn);
 561         if (is_writeble_pte(*spte))
 562                 kvm_release_pfn_dirty(pfn);
 563         else
 564                 kvm_release_pfn_clean(pfn);
 565         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 566         if (!*rmapp) {
 567                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 568                 BUG();
 569         } else if (!(*rmapp & 1)) {
 570                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 571                 if ((u64 *)*rmapp != spte) {
 572                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 573                                spte, *spte);
 574                         BUG();
 575                 }
 576                 *rmapp = 0;
 577         } else {
 578                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 579                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 580                 prev_desc = NULL;
 581                 while (desc) {
 582                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 583                                 if (desc->shadow_ptes[i] == spte) {
 584                                         rmap_desc_remove_entry(rmapp,
 585                                                                desc, i,
 586                                                                prev_desc);
 587                                         return;
 588                                 }
 589                         prev_desc = desc;
 590                         desc = desc->more;
 591                 }
 592                 BUG();
 593         }
 594 }
 595
 596 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 597 {
 598         struct kvm_rmap_desc *desc;
 599         struct kvm_rmap_desc *prev_desc;
 600         u64 *prev_spte;
 601         int i;
 602
 603         if (!*rmapp)
 604                 return NULL;
 605         else if (!(*rmapp & 1)) {
 606                 if (!spte)
 607                         return (u64 *)*rmapp;
 608                 return NULL;
 609         }
 610         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 611         prev_desc = NULL;
 612         prev_spte = NULL;
 613         while (desc) {
 614                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 615                         if (prev_spte == spte)
 616                                 return desc->shadow_ptes[i];
 617                         prev_spte = desc->shadow_ptes[i];
 618                 }
 619                 desc = desc->more;
 620         }
 621         return NULL;
 622 }
 623
 624 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 625 {
 626         unsigned long *rmapp;
 627         u64 *spte;
 628         int write_protected = 0;
 629
 630         gfn = unalias_gfn(kvm, gfn);
 631         rmapp = gfn_to_rmap(kvm, gfn, 0);
 632
 633         spte = rmap_next(kvm, rmapp, NULL);
 634         while (spte) {
 635                 BUG_ON(!spte);
 636                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 637                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 638                 if (is_writeble_pte(*spte)) {
 639                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 640                         write_protected = 1;
 641                 }
 642                 spte = rmap_next(kvm, rmapp, spte);
 643         }
 644         if (write_protected) {
 645                 pfn_t pfn;
 646
 647                 spte = rmap_next(kvm, rmapp, NULL);
 648                 pfn = spte_to_pfn(*spte);
 649                 kvm_set_pfn_dirty(pfn);
 650         }
 651
 652         /* check for huge page mappings */
 653         rmapp = gfn_to_rmap(kvm, gfn, 1);
 654         spte = rmap_next(kvm, rmapp, NULL);
 655         while (spte) {
 656                 BUG_ON(!spte);
 657                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 658                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 659                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 660                 if (is_writeble_pte(*spte)) {
 661                         rmap_remove(kvm, spte);
 662                         --kvm->stat.lpages;
 663                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 664                         spte = NULL;
 665                         write_protected = 1;
 666                 }
 667                 spte = rmap_next(kvm, rmapp, spte);
 668         }
 669
 670         return write_protected;
 671 }
 672
 673 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 674 {
 675         u64 *spte;
 676         int need_tlb_flush = 0;
 677
 678         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 679                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 680                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 681                 rmap_remove(kvm, spte);
 682                 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 683                 need_tlb_flush = 1;
 684         }
 685         return need_tlb_flush;
 686 }
 687
 688 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 689                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 690 {
 691         int i;
 692         int retval = 0;
 693
 694         /*
 695          * If mmap_sem isn't taken, we can look the memslots with only
 696          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 697          */
 698         for (i = 0; i < kvm->nmemslots; i++) {
 699                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 700                 unsigned long start = memslot->userspace_addr;
 701                 unsigned long end;
 702
 703                 /* mmu_lock protects userspace_addr */
 704                 if (!start)
 705                         continue;
 706
 707                 end = start + (memslot->npages << PAGE_SHIFT);
 708                 if (hva >= start && hva < end) {
 709                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 710                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 711                         retval |= handler(kvm,
 712                                           &memslot->lpage_info[
 713                                                   gfn_offset /
 714                                                   KVM_PAGES_PER_HPAGE].rmap_pde);
 715                 }
 716         }
 717
 718         return retval;
 719 }
 720
 721 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 722 {
 723         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 724 }
 725
 726 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 727 {
 728         u64 *spte;
 729         int young = 0;
 730
 731         /* always return old for EPT */
 732         if (!shadow_accessed_mask)
 733                 return 0;
 734
 735         spte = rmap_next(kvm, rmapp, NULL);
 736         while (spte) {
 737                 int _young;
 738                 u64 _spte = *spte;
 739                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 740                 _young = _spte & PT_ACCESSED_MASK;
 741                 if (_young) {
 742                         young = 1;
 743                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 744                 }
 745                 spte = rmap_next(kvm, rmapp, spte);
 746         }
 747         return young;
 748 }
 749
 750 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 751 {
 752         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 753 }
 754
 755 #ifdef MMU_DEBUG
 756 static int is_empty_shadow_page(u64 *spt)
 757 {
 758         u64 *pos;
 759         u64 *end;
 760
 761         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 762                 if (is_shadow_present_pte(*pos)) {
 763                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 764                                pos, *pos);
 765                         return 0;
 766                 }
 767         return 1;
 768 }
 769 #endif
 770
 771 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 772 {
 773         ASSERT(is_empty_shadow_page(sp->spt));
 774         list_del(&sp->link);
 775         __free_page(virt_to_page(sp->spt));
 776         __free_page(virt_to_page(sp->gfns));
 777         kfree(sp);
 778         ++kvm->arch.n_free_mmu_pages;
 779 }
 780
 781 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 782 {
 783         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 784 }
 785
 786 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 787                                                u64 *parent_pte)
 788 {
 789         struct kvm_mmu_page *sp;
 790
 791         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 792         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 793         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 794         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 795         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 796         ASSERT(is_empty_shadow_page(sp->spt));
 797         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 798         sp->multimapped = 0;
 799         sp->parent_pte = parent_pte;
 800         --vcpu->kvm->arch.n_free_mmu_pages;
 801         return sp;
 802 }
 803
 804 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 805                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 806 {
 807         struct kvm_pte_chain *pte_chain;
 808         struct hlist_node *node;
 809         int i;
 810
 811         if (!parent_pte)
 812                 return;
 813         if (!sp->multimapped) {
 814                 u64 *old = sp->parent_pte;
 815
 816                 if (!old) {
 817                         sp->parent_pte = parent_pte;
 818                         return;
 819                 }
 820                 sp->multimapped = 1;
 821                 pte_chain = mmu_alloc_pte_chain(vcpu);
 822                 INIT_HLIST_HEAD(&sp->parent_ptes);
 823                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 824                 pte_chain->parent_ptes[0] = old;
 825         }
 826         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 827                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 828                         continue;
 829                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 830                         if (!pte_chain->parent_ptes[i]) {
 831                                 pte_chain->parent_ptes[i] = parent_pte;
 832                                 return;
 833                         }
 834         }
 835         pte_chain = mmu_alloc_pte_chain(vcpu);
 836         BUG_ON(!pte_chain);
 837         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 838         pte_chain->parent_ptes[0] = parent_pte;
 839 }
 840
 841 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 842                                        u64 *parent_pte)
 843 {
 844         struct kvm_pte_chain *pte_chain;
 845         struct hlist_node *node;
 846         int i;
 847
 848         if (!sp->multimapped) {
 849                 BUG_ON(sp->parent_pte != parent_pte);
 850                 sp->parent_pte = NULL;
 851                 return;
 852         }
 853         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 854                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 855                         if (!pte_chain->parent_ptes[i])
 856                                 break;
 857                         if (pte_chain->parent_ptes[i] != parent_pte)
 858                                 continue;
 859                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 860                                 && pte_chain->parent_ptes[i + 1]) {
 861                                 pte_chain->parent_ptes[i]
 862                                         = pte_chain->parent_ptes[i + 1];
 863                                 ++i;
 864                         }
 865                         pte_chain->parent_ptes[i] = NULL;
 866                         if (i == 0) {
 867                                 hlist_del(&pte_chain->link);
 868                                 mmu_free_pte_chain(pte_chain);
 869                                 if (hlist_empty(&sp->parent_ptes)) {
 870                                         sp->multimapped = 0;
 871                                         sp->parent_pte = NULL;
 872                                 }
 873                         }
 874                         return;
 875                 }
 876         BUG();
 877 }
 878
 879
 880 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 881                             mmu_parent_walk_fn fn)
 882 {
 883         struct kvm_pte_chain *pte_chain;
 884         struct hlist_node *node;
 885         struct kvm_mmu_page *parent_sp;
 886         int i;
 887
 888         if (!sp->multimapped && sp->parent_pte) {
 889                 parent_sp = page_header(__pa(sp->parent_pte));
 890                 fn(vcpu, parent_sp);
 891                 mmu_parent_walk(vcpu, parent_sp, fn);
 892                 return;
 893         }
 894         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 895                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 896                         if (!pte_chain->parent_ptes[i])
 897                                 break;
 898                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 899                         fn(vcpu, parent_sp);
 900                         mmu_parent_walk(vcpu, parent_sp, fn);
 901                 }
 902 }
 903
 904 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 905 {
 906         unsigned int index;
 907         struct kvm_mmu_page *sp = page_header(__pa(spte));
 908
 909         index = spte - sp->spt;
 910         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
 911                 sp->unsync_children++;
 912         WARN_ON(!sp->unsync_children);
 913 }
 914
 915 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 916 {
 917         struct kvm_pte_chain *pte_chain;
 918         struct hlist_node *node;
 919         int i;
 920
 921         if (!sp->parent_pte)
 922                 return;
 923
 924         if (!sp->multimapped) {
 925                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 926                 return;
 927         }
 928
 929         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 930                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 931                         if (!pte_chain->parent_ptes[i])
 932                                 break;
 933                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 934                 }
 935 }
 936
 937 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 938 {
 939         kvm_mmu_update_parents_unsync(sp);
 940         return 1;
 941 }
 942
 943 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 944                                         struct kvm_mmu_page *sp)
 945 {
 946         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 947         kvm_mmu_update_parents_unsync(sp);
 948 }
 949
 950 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 951                                     struct kvm_mmu_page *sp)
 952 {
 953         int i;
 954
 955         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 956                 sp->spt[i] = shadow_trap_nonpresent_pte;
 957 }
 958
 959 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
 960                                struct kvm_mmu_page *sp)
 961 {
 962         return 1;
 963 }
 964
 965 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
 966 {
 967 }
 968
 969 #define KVM_PAGE_ARRAY_NR 16
 970
 971 struct kvm_mmu_pages {
 972         struct mmu_page_and_offset {
 973                 struct kvm_mmu_page *sp;
 974                 unsigned int idx;
 975         } page[KVM_PAGE_ARRAY_NR];
 976         unsigned int nr;
 977 };
 978
 979 #define for_each_unsync_children(bitmap, idx)           \
 980         for (idx = find_first_bit(bitmap, 512);         \
 981              idx < 512;                                 \
 982              idx = find_next_bit(bitmap, 512, idx+1))
 983
 984 int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
 985                    int idx)
 986 {
 987         int i;
 988
 989         if (sp->unsync)
 990                 for (i=0; i < pvec->nr; i++)
 991                         if (pvec->page[i].sp == sp)
 992                                 return 0;
 993
 994         pvec->page[pvec->nr].sp = sp;
 995         pvec->page[pvec->nr].idx = idx;
 996         pvec->nr++;
 997         return (pvec->nr == KVM_PAGE_ARRAY_NR);
 998 }
 999
1000 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1001                            struct kvm_mmu_pages *pvec)
1002 {
1003         int i, ret, nr_unsync_leaf = 0;
1004
1005         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1006                 u64 ent = sp->spt[i];
1007
1008                 if (is_shadow_present_pte(ent)) {
1009                         struct kvm_mmu_page *child;
1010                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1011
1012                         if (child->unsync_children) {
1013                                 if (mmu_pages_add(pvec, child, i))
1014                                         return -ENOSPC;
1015
1016                                 ret = __mmu_unsync_walk(child, pvec);
1017                                 if (!ret)
1018                                         __clear_bit(i, sp->unsync_child_bitmap);
1019                                 else if (ret > 0)
1020                                         nr_unsync_leaf += ret;
1021                                 else
1022                                         return ret;
1023                         }
1024
1025                         if (child->unsync) {
1026                                 nr_unsync_leaf++;
1027                                 if (mmu_pages_add(pvec, child, i))
1028                                         return -ENOSPC;
1029                         }
1030                 }
1031         }
1032
1033         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1034                 sp->unsync_children = 0;
1035
1036         return nr_unsync_leaf;
1037 }
1038
1039 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1040                            struct kvm_mmu_pages *pvec)
1041 {
1042         if (!sp->unsync_children)
1043                 return 0;
1044
1045         mmu_pages_add(pvec, sp, 0);
1046         return __mmu_unsync_walk(sp, pvec);
1047 }
1048
1049 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1050 {
1051         unsigned index;
1052         struct hlist_head *bucket;
1053         struct kvm_mmu_page *sp;
1054         struct hlist_node *node;
1055
1056         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1057         index = kvm_page_table_hashfn(gfn);
1058         bucket = &kvm->arch.mmu_page_hash[index];
1059         hlist_for_each_entry(sp, node, bucket, hash_link)
1060                 if (sp->gfn == gfn && !sp->role.metaphysical
1061                     && !sp->role.invalid) {
1062                         pgprintk("%s: found role %x\n",
1063                                  __func__, sp->role.word);
1064                         return sp;
1065                 }
1066         return NULL;
1067 }
1068
1069 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1070 {
1071         WARN_ON(!sp->unsync);
1072         sp->unsync = 0;
1073         --kvm->stat.mmu_unsync;
1074 }
1075
1076 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1077
1078 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1079 {
1080         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1081                 kvm_mmu_zap_page(vcpu->kvm, sp);
1082                 return 1;
1083         }
1084
1085         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1086                 kvm_flush_remote_tlbs(vcpu->kvm);
1087         kvm_unlink_unsync_page(vcpu->kvm, sp);
1088         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1089                 kvm_mmu_zap_page(vcpu->kvm, sp);
1090                 return 1;
1091         }
1092
1093         kvm_mmu_flush_tlb(vcpu);
1094         return 0;
1095 }
1096
1097 struct mmu_page_path {
1098         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1099         unsigned int idx[PT64_ROOT_LEVEL-1];
1100 };
1101
1102 #define for_each_sp(pvec, sp, parents, i)                       \
1103                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1104                         sp = pvec.page[i].sp;                   \
1105                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1106                         i = mmu_pages_next(&pvec, &parents, i))
1107
1108 int mmu_pages_next(struct kvm_mmu_pages *pvec, struct mmu_page_path *parents,
1109                    int i)
1110 {
1111         int n;
1112
1113         for (n = i+1; n < pvec->nr; n++) {
1114                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1115
1116                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1117                         parents->idx[0] = pvec->page[n].idx;
1118                         return n;
1119                 }
1120
1121                 parents->parent[sp->role.level-2] = sp;
1122                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1123         }
1124
1125         return n;
1126 }
1127
1128 void mmu_pages_clear_parents(struct mmu_page_path *parents)
1129 {
1130         struct kvm_mmu_page *sp;
1131         unsigned int level = 0;
1132
1133         do {
1134                 unsigned int idx = parents->idx[level];
1135
1136                 sp = parents->parent[level];
1137                 if (!sp)
1138                         return;
1139
1140                 --sp->unsync_children;
1141                 WARN_ON((int)sp->unsync_children < 0);
1142                 __clear_bit(idx, sp->unsync_child_bitmap);
1143                 level++;
1144         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1145 }
1146
1147 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1148                                struct mmu_page_path *parents,
1149                                struct kvm_mmu_pages *pvec)
1150 {
1151         parents->parent[parent->role.level-1] = NULL;
1152         pvec->nr = 0;
1153 }
1154
1155 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1156                               struct kvm_mmu_page *parent)
1157 {
1158         int i;
1159         struct kvm_mmu_page *sp;
1160         struct mmu_page_path parents;
1161         struct kvm_mmu_pages pages;
1162
1163         kvm_mmu_pages_init(parent, &parents, &pages);
1164         while (mmu_unsync_walk(parent, &pages)) {
1165                 int protected = 0;
1166
1167                 for_each_sp(pages, sp, parents, i)
1168                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1169
1170                 if (protected)
1171                         kvm_flush_remote_tlbs(vcpu->kvm);
1172
1173                 for_each_sp(pages, sp, parents, i) {
1174                         kvm_sync_page(vcpu, sp);
1175                         mmu_pages_clear_parents(&parents);
1176                 }
1177                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1178                 kvm_mmu_pages_init(parent, &parents, &pages);
1179         }
1180 }
1181
1182 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1183                                              gfn_t gfn,
1184                                              gva_t gaddr,
1185                                              unsigned level,
1186                                              int metaphysical,
1187                                              unsigned access,
1188                                              u64 *parent_pte)
1189 {
1190         union kvm_mmu_page_role role;
1191         unsigned index;
1192         unsigned quadrant;
1193         struct hlist_head *bucket;
1194         struct kvm_mmu_page *sp;
1195         struct hlist_node *node, *tmp;
1196
1197         role.word = 0;
1198         role.glevels = vcpu->arch.mmu.root_level;
1199         role.level = level;
1200         role.metaphysical = metaphysical;
1201         role.access = access;
1202         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1203                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1204                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1205                 role.quadrant = quadrant;
1206         }
1207         pgprintk("%s: looking gfn %lx role %x\n", __func__,
1208                  gfn, role.word);
1209         index = kvm_page_table_hashfn(gfn);
1210         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1211         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1212                 if (sp->gfn == gfn) {
1213                         if (sp->unsync)
1214                                 if (kvm_sync_page(vcpu, sp))
1215                                         continue;
1216
1217                         if (sp->role.word != role.word)
1218                                 continue;
1219
1220                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1221                         if (sp->unsync_children) {
1222                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1223                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1224                         }
1225                         pgprintk("%s: found\n", __func__);
1226                         return sp;
1227                 }
1228         ++vcpu->kvm->stat.mmu_cache_miss;
1229         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1230         if (!sp)
1231                 return sp;
1232         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1233         sp->gfn = gfn;
1234         sp->role = role;
1235         hlist_add_head(&sp->hash_link, bucket);
1236         if (!metaphysical) {
1237                 if (rmap_write_protect(vcpu->kvm, gfn))
1238                         kvm_flush_remote_tlbs(vcpu->kvm);
1239                 account_shadowed(vcpu->kvm, gfn);
1240         }
1241         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1242                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1243         else
1244                 nonpaging_prefetch_page(vcpu, sp);
1245         return sp;
1246 }
1247
1248 static int walk_shadow(struct kvm_shadow_walk *walker,
1249                        struct kvm_vcpu *vcpu, u64 addr)
1250 {
1251         hpa_t shadow_addr;
1252         int level;
1253         int r;
1254         u64 *sptep;
1255         unsigned index;
1256
1257         shadow_addr = vcpu->arch.mmu.root_hpa;
1258         level = vcpu->arch.mmu.shadow_root_level;
1259         if (level == PT32E_ROOT_LEVEL) {
1260                 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1261                 shadow_addr &= PT64_BASE_ADDR_MASK;
1262                 --level;
1263         }
1264
1265         while (level >= PT_PAGE_TABLE_LEVEL) {
1266                 index = SHADOW_PT_INDEX(addr, level);
1267                 sptep = ((u64 *)__va(shadow_addr)) + index;
1268                 r = walker->entry(walker, vcpu, addr, sptep, level);
1269                 if (r)
1270                         return r;
1271                 shadow_addr = *sptep & PT64_BASE_ADDR_MASK;
1272                 --level;
1273         }
1274         return 0;
1275 }
1276
1277 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1278                                          struct kvm_mmu_page *sp)
1279 {
1280         unsigned i;
1281         u64 *pt;
1282         u64 ent;
1283
1284         pt = sp->spt;
1285
1286         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1287                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1288                         if (is_shadow_present_pte(pt[i]))
1289                                 rmap_remove(kvm, &pt[i]);
1290                         pt[i] = shadow_trap_nonpresent_pte;
1291                 }
1292                 return;
1293         }
1294
1295         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1296                 ent = pt[i];
1297
1298                 if (is_shadow_present_pte(ent)) {
1299                         if (!is_large_pte(ent)) {
1300                                 ent &= PT64_BASE_ADDR_MASK;
1301                                 mmu_page_remove_parent_pte(page_header(ent),
1302                                                            &pt[i]);
1303                         } else {
1304                                 --kvm->stat.lpages;
1305                                 rmap_remove(kvm, &pt[i]);
1306                         }
1307                 }
1308                 pt[i] = shadow_trap_nonpresent_pte;
1309         }
1310 }
1311
1312 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1313 {
1314         mmu_page_remove_parent_pte(sp, parent_pte);
1315 }
1316
1317 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1318 {
1319         int i;
1320
1321         for (i = 0; i < KVM_MAX_VCPUS; ++i)
1322                 if (kvm->vcpus[i])
1323                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
1324 }
1325
1326 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1327 {
1328         u64 *parent_pte;
1329
1330         while (sp->multimapped || sp->parent_pte) {
1331                 if (!sp->multimapped)
1332                         parent_pte = sp->parent_pte;
1333                 else {
1334                         struct kvm_pte_chain *chain;
1335
1336                         chain = container_of(sp->parent_ptes.first,
1337                                              struct kvm_pte_chain, link);
1338                         parent_pte = chain->parent_ptes[0];
1339                 }
1340                 BUG_ON(!parent_pte);
1341                 kvm_mmu_put_page(sp, parent_pte);
1342                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1343         }
1344 }
1345
1346 static int mmu_zap_unsync_children(struct kvm *kvm,
1347                                    struct kvm_mmu_page *parent)
1348 {
1349         int i, zapped = 0;
1350         struct mmu_page_path parents;
1351         struct kvm_mmu_pages pages;
1352
1353         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1354                 return 0;
1355
1356         kvm_mmu_pages_init(parent, &parents, &pages);
1357         while (mmu_unsync_walk(parent, &pages)) {
1358                 struct kvm_mmu_page *sp;
1359
1360                 for_each_sp(pages, sp, parents, i) {
1361                         kvm_mmu_zap_page(kvm, sp);
1362                         mmu_pages_clear_parents(&parents);
1363                 }
1364                 zapped += pages.nr;
1365                 kvm_mmu_pages_init(parent, &parents, &pages);
1366         }
1367
1368         return zapped;
1369 }
1370
1371 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1372 {
1373         int ret;
1374         ++kvm->stat.mmu_shadow_zapped;
1375         ret = mmu_zap_unsync_children(kvm, sp);
1376         kvm_mmu_page_unlink_children(kvm, sp);
1377         kvm_mmu_unlink_parents(kvm, sp);
1378         kvm_flush_remote_tlbs(kvm);
1379         if (!sp->role.invalid && !sp->role.metaphysical)
1380                 unaccount_shadowed(kvm, sp->gfn);
1381         if (sp->unsync)
1382                 kvm_unlink_unsync_page(kvm, sp);
1383         if (!sp->root_count) {
1384                 hlist_del(&sp->hash_link);
1385                 kvm_mmu_free_page(kvm, sp);
1386         } else {
1387                 sp->role.invalid = 1;
1388                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1389                 kvm_reload_remote_mmus(kvm);
1390         }
1391         kvm_mmu_reset_last_pte_updated(kvm);
1392         return ret;
1393 }
1394
1395 /*
1396  * Changing the number of mmu pages allocated to the vm
1397  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1398  */
1399 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1400 {
1401         /*
1402          * If we set the number of mmu pages to be smaller be than the
1403          * number of actived pages , we must to free some mmu pages before we
1404          * change the value
1405          */
1406
1407         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1408             kvm_nr_mmu_pages) {
1409                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1410                                        - kvm->arch.n_free_mmu_pages;
1411
1412                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1413                         struct kvm_mmu_page *page;
1414
1415                         page = container_of(kvm->arch.active_mmu_pages.prev,
1416                                             struct kvm_mmu_page, link);
1417                         kvm_mmu_zap_page(kvm, page);
1418                         n_used_mmu_pages--;
1419                 }
1420                 kvm->arch.n_free_mmu_pages = 0;
1421         }
1422         else
1423                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1424                                          - kvm->arch.n_alloc_mmu_pages;
1425
1426         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1427 }
1428
1429 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1430 {
1431         unsigned index;
1432         struct hlist_head *bucket;
1433         struct kvm_mmu_page *sp;
1434         struct hlist_node *node, *n;
1435         int r;
1436
1437         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1438         r = 0;
1439         index = kvm_page_table_hashfn(gfn);
1440         bucket = &kvm->arch.mmu_page_hash[index];
1441         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1442                 if (sp->gfn == gfn && !sp->role.metaphysical) {
1443                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1444                                  sp->role.word);
1445                         r = 1;
1446                         if (kvm_mmu_zap_page(kvm, sp))
1447                                 n = bucket->first;
1448                 }
1449         return r;
1450 }
1451
1452 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1453 {
1454         struct kvm_mmu_page *sp;
1455
1456         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1457                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1458                 kvm_mmu_zap_page(kvm, sp);
1459         }
1460 }
1461
1462 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1463 {
1464         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1465         struct kvm_mmu_page *sp = page_header(__pa(pte));
1466
1467         __set_bit(slot, sp->slot_bitmap);
1468 }
1469
1470 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1471 {
1472         int i;
1473         u64 *pt = sp->spt;
1474
1475         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1476                 return;
1477
1478         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1479                 if (pt[i] == shadow_notrap_nonpresent_pte)
1480                         set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1481         }
1482 }
1483
1484 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1485 {
1486         struct page *page;
1487
1488         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1489
1490         if (gpa == UNMAPPED_GVA)
1491                 return NULL;
1492
1493         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1494
1495         return page;
1496 }
1497
1498 /*
1499  * The function is based on mtrr_type_lookup() in
1500  * arch/x86/kernel/cpu/mtrr/generic.c
1501  */
1502 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1503                          u64 start, u64 end)
1504 {
1505         int i;
1506         u64 base, mask;
1507         u8 prev_match, curr_match;
1508         int num_var_ranges = KVM_NR_VAR_MTRR;
1509
1510         if (!mtrr_state->enabled)
1511                 return 0xFF;
1512
1513         /* Make end inclusive end, instead of exclusive */
1514         end--;
1515
1516         /* Look in fixed ranges. Just return the type as per start */
1517         if (mtrr_state->have_fixed && (start < 0x100000)) {
1518                 int idx;
1519
1520                 if (start < 0x80000) {
1521                         idx = 0;
1522                         idx += (start >> 16);
1523                         return mtrr_state->fixed_ranges[idx];
1524                 } else if (start < 0xC0000) {
1525                         idx = 1 * 8;
1526                         idx += ((start - 0x80000) >> 14);
1527                         return mtrr_state->fixed_ranges[idx];
1528                 } else if (start < 0x1000000) {
1529                         idx = 3 * 8;
1530                         idx += ((start - 0xC0000) >> 12);
1531                         return mtrr_state->fixed_ranges[idx];
1532                 }
1533         }
1534
1535         /*
1536          * Look in variable ranges
1537          * Look of multiple ranges matching this address and pick type
1538          * as per MTRR precedence
1539          */
1540         if (!(mtrr_state->enabled & 2))
1541                 return mtrr_state->def_type;
1542
1543         prev_match = 0xFF;
1544         for (i = 0; i < num_var_ranges; ++i) {
1545                 unsigned short start_state, end_state;
1546
1547                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1548                         continue;
1549
1550                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1551                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1552                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1553                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1554
1555                 start_state = ((start & mask) == (base & mask));
1556                 end_state = ((end & mask) == (base & mask));
1557                 if (start_state != end_state)
1558                         return 0xFE;
1559
1560                 if ((start & mask) != (base & mask))
1561                         continue;
1562
1563                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1564                 if (prev_match == 0xFF) {
1565                         prev_match = curr_match;
1566                         continue;
1567                 }
1568
1569                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1570                     curr_match == MTRR_TYPE_UNCACHABLE)
1571                         return MTRR_TYPE_UNCACHABLE;
1572
1573                 if ((prev_match == MTRR_TYPE_WRBACK &&
1574                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1575                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1576                      curr_match == MTRR_TYPE_WRBACK)) {
1577                         prev_match = MTRR_TYPE_WRTHROUGH;
1578                         curr_match = MTRR_TYPE_WRTHROUGH;
1579                 }
1580
1581                 if (prev_match != curr_match)
1582                         return MTRR_TYPE_UNCACHABLE;
1583         }
1584
1585         if (prev_match != 0xFF)
1586                 return prev_match;
1587
1588         return mtrr_state->def_type;
1589 }
1590
1591 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1592 {
1593         u8 mtrr;
1594
1595         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1596                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1597         if (mtrr == 0xfe || mtrr == 0xff)
1598                 mtrr = MTRR_TYPE_WRBACK;
1599         return mtrr;
1600 }
1601
1602 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1603 {
1604         unsigned index;
1605         struct hlist_head *bucket;
1606         struct kvm_mmu_page *s;
1607         struct hlist_node *node, *n;
1608
1609         index = kvm_page_table_hashfn(sp->gfn);
1610         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1611         /* don't unsync if pagetable is shadowed with multiple roles */
1612         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1613                 if (s->gfn != sp->gfn || s->role.metaphysical)
1614                         continue;
1615                 if (s->role.word != sp->role.word)
1616                         return 1;
1617         }
1618         kvm_mmu_mark_parents_unsync(vcpu, sp);
1619         ++vcpu->kvm->stat.mmu_unsync;
1620         sp->unsync = 1;
1621         mmu_convert_notrap(sp);
1622         return 0;
1623 }
1624
1625 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1626                                   bool can_unsync)
1627 {
1628         struct kvm_mmu_page *shadow;
1629
1630         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1631         if (shadow) {
1632                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1633                         return 1;
1634                 if (shadow->unsync)
1635                         return 0;
1636                 if (can_unsync && oos_shadow)
1637                         return kvm_unsync_page(vcpu, shadow);
1638                 return 1;
1639         }
1640         return 0;
1641 }
1642
1643 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1644                     unsigned pte_access, int user_fault,
1645                     int write_fault, int dirty, int largepage,
1646                     gfn_t gfn, pfn_t pfn, bool speculative,
1647                     bool can_unsync)
1648 {
1649         u64 spte;
1650         int ret = 0;
1651         u64 mt_mask = shadow_mt_mask;
1652
1653         /*
1654          * We don't set the accessed bit, since we sometimes want to see
1655          * whether the guest actually used the pte (in order to detect
1656          * demand paging).
1657          */
1658         spte = shadow_base_present_pte | shadow_dirty_mask;
1659         if (!speculative)
1660                 spte |= shadow_accessed_mask;
1661         if (!dirty)
1662                 pte_access &= ~ACC_WRITE_MASK;
1663         if (pte_access & ACC_EXEC_MASK)
1664                 spte |= shadow_x_mask;
1665         else
1666                 spte |= shadow_nx_mask;
1667         if (pte_access & ACC_USER_MASK)
1668                 spte |= shadow_user_mask;
1669         if (largepage)
1670                 spte |= PT_PAGE_SIZE_MASK;
1671         if (mt_mask) {
1672                 mt_mask = get_memory_type(vcpu, gfn) <<
1673                           kvm_x86_ops->get_mt_mask_shift();
1674                 spte |= mt_mask;
1675         }
1676
1677         spte |= (u64)pfn << PAGE_SHIFT;
1678
1679         if ((pte_access & ACC_WRITE_MASK)
1680             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1681
1682                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1683                         ret = 1;
1684                         spte = shadow_trap_nonpresent_pte;
1685                         goto set_pte;
1686                 }
1687
1688                 spte |= PT_WRITABLE_MASK;
1689
1690                 /*
1691                  * Optimization: for pte sync, if spte was writable the hash
1692                  * lookup is unnecessary (and expensive). Write protection
1693                  * is responsibility of mmu_get_page / kvm_sync_page.
1694                  * Same reasoning can be applied to dirty page accounting.
1695                  */
1696                 if (!can_unsync && is_writeble_pte(*shadow_pte))
1697                         goto set_pte;
1698
1699                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1700                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1701                                  __func__, gfn);
1702                         ret = 1;
1703                         pte_access &= ~ACC_WRITE_MASK;
1704                         if (is_writeble_pte(spte))
1705                                 spte &= ~PT_WRITABLE_MASK;
1706                 }
1707         }
1708
1709         if (pte_access & ACC_WRITE_MASK)
1710                 mark_page_dirty(vcpu->kvm, gfn);
1711
1712 set_pte:
1713         set_shadow_pte(shadow_pte, spte);
1714         return ret;
1715 }
1716
1717 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1718                          unsigned pt_access, unsigned pte_access,
1719                          int user_fault, int write_fault, int dirty,
1720                          int *ptwrite, int largepage, gfn_t gfn,
1721                          pfn_t pfn, bool speculative)
1722 {
1723         int was_rmapped = 0;
1724         int was_writeble = is_writeble_pte(*shadow_pte);
1725
1726         pgprintk("%s: spte %llx access %x write_fault %d"
1727                  " user_fault %d gfn %lx\n",
1728                  __func__, *shadow_pte, pt_access,
1729                  write_fault, user_fault, gfn);
1730
1731         if (is_rmap_pte(*shadow_pte)) {
1732                 /*
1733                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1734                  * the parent of the now unreachable PTE.
1735                  */
1736                 if (largepage && !is_large_pte(*shadow_pte)) {
1737                         struct kvm_mmu_page *child;
1738                         u64 pte = *shadow_pte;
1739
1740                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1741                         mmu_page_remove_parent_pte(child, shadow_pte);
1742                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1743                         pgprintk("hfn old %lx new %lx\n",
1744                                  spte_to_pfn(*shadow_pte), pfn);
1745                         rmap_remove(vcpu->kvm, shadow_pte);
1746                 } else {
1747                         if (largepage)
1748                                 was_rmapped = is_large_pte(*shadow_pte);
1749                         else
1750                                 was_rmapped = 1;
1751                 }
1752         }
1753         if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1754                       dirty, largepage, gfn, pfn, speculative, true)) {
1755                 if (write_fault)
1756                         *ptwrite = 1;
1757                 kvm_x86_ops->tlb_flush(vcpu);
1758         }
1759
1760         pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1761         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1762                  is_large_pte(*shadow_pte)? "2MB" : "4kB",
1763                  is_present_pte(*shadow_pte)?"RW":"R", gfn,
1764                  *shadow_pte, shadow_pte);
1765         if (!was_rmapped && is_large_pte(*shadow_pte))
1766                 ++vcpu->kvm->stat.lpages;
1767
1768         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1769         if (!was_rmapped) {
1770                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1771                 if (!is_rmap_pte(*shadow_pte))
1772                         kvm_release_pfn_clean(pfn);
1773         } else {
1774                 if (was_writeble)
1775                         kvm_release_pfn_dirty(pfn);
1776                 else
1777                         kvm_release_pfn_clean(pfn);
1778         }
1779         if (speculative) {
1780                 vcpu->arch.last_pte_updated = shadow_pte;
1781                 vcpu->arch.last_pte_gfn = gfn;
1782         }
1783 }
1784
1785 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1786 {
1787 }
1788
1789 struct direct_shadow_walk {
1790         struct kvm_shadow_walk walker;
1791         pfn_t pfn;
1792         int write;
1793         int largepage;
1794         int pt_write;
1795 };
1796
1797 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1798                             struct kvm_vcpu *vcpu,
1799                             u64 addr, u64 *sptep, int level)
1800 {
1801         struct direct_shadow_walk *walk =
1802                 container_of(_walk, struct direct_shadow_walk, walker);
1803         struct kvm_mmu_page *sp;
1804         gfn_t pseudo_gfn;
1805         gfn_t gfn = addr >> PAGE_SHIFT;
1806
1807         if (level == PT_PAGE_TABLE_LEVEL
1808             || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1809                 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1810                              0, walk->write, 1, &walk->pt_write,
1811                              walk->largepage, gfn, walk->pfn, false);
1812                 ++vcpu->stat.pf_fixed;
1813                 return 1;
1814         }
1815
1816         if (*sptep == shadow_trap_nonpresent_pte) {
1817                 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1818                 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1819                                       1, ACC_ALL, sptep);
1820                 if (!sp) {
1821                         pgprintk("nonpaging_map: ENOMEM\n");
1822                         kvm_release_pfn_clean(walk->pfn);
1823                         return -ENOMEM;
1824                 }
1825
1826                 set_shadow_pte(sptep,
1827                                __pa(sp->spt)
1828                                | PT_PRESENT_MASK | PT_WRITABLE_MASK
1829                                | shadow_user_mask | shadow_x_mask);
1830         }
1831         return 0;
1832 }
1833
1834 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1835                         int largepage, gfn_t gfn, pfn_t pfn)
1836 {
1837         int r;
1838         struct direct_shadow_walk walker = {
1839                 .walker = { .entry = direct_map_entry, },
1840                 .pfn = pfn,
1841                 .largepage = largepage,
1842                 .write = write,
1843                 .pt_write = 0,
1844         };
1845
1846         r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1847         if (r < 0)
1848                 return r;
1849         return walker.pt_write;
1850 }
1851
1852 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1853 {
1854         int r;
1855         int largepage = 0;
1856         pfn_t pfn;
1857         unsigned long mmu_seq;
1858
1859         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1860                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1861                 largepage = 1;
1862         }
1863
1864         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1865         smp_rmb();
1866         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1867
1868         /* mmio */
1869         if (is_error_pfn(pfn)) {
1870                 kvm_release_pfn_clean(pfn);
1871                 return 1;
1872         }
1873
1874         spin_lock(&vcpu->kvm->mmu_lock);
1875         if (mmu_notifier_retry(vcpu, mmu_seq))
1876                 goto out_unlock;
1877         kvm_mmu_free_some_pages(vcpu);
1878         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1879         spin_unlock(&vcpu->kvm->mmu_lock);
1880
1881
1882         return r;
1883
1884 out_unlock:
1885         spin_unlock(&vcpu->kvm->mmu_lock);
1886         kvm_release_pfn_clean(pfn);
1887         return 0;
1888 }
1889
1890
1891 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1892 {
1893         int i;
1894         struct kvm_mmu_page *sp;
1895
1896         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1897                 return;
1898         spin_lock(&vcpu->kvm->mmu_lock);
1899         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1900                 hpa_t root = vcpu->arch.mmu.root_hpa;
1901
1902                 sp = page_header(root);
1903                 --sp->root_count;
1904                 if (!sp->root_count && sp->role.invalid)
1905                         kvm_mmu_zap_page(vcpu->kvm, sp);
1906                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1907                 spin_unlock(&vcpu->kvm->mmu_lock);
1908                 return;
1909         }
1910         for (i = 0; i < 4; ++i) {
1911                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1912
1913                 if (root) {
1914                         root &= PT64_BASE_ADDR_MASK;
1915                         sp = page_header(root);
1916                         --sp->root_count;
1917                         if (!sp->root_count && sp->role.invalid)
1918                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1919                 }
1920                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1921         }
1922         spin_unlock(&vcpu->kvm->mmu_lock);
1923         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1924 }
1925
1926 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1927 {
1928         int i;
1929         gfn_t root_gfn;
1930         struct kvm_mmu_page *sp;
1931         int metaphysical = 0;
1932
1933         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1934
1935         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1936                 hpa_t root = vcpu->arch.mmu.root_hpa;
1937
1938                 ASSERT(!VALID_PAGE(root));
1939                 if (tdp_enabled)
1940                         metaphysical = 1;
1941                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1942                                       PT64_ROOT_LEVEL, metaphysical,
1943                                       ACC_ALL, NULL);
1944                 root = __pa(sp->spt);
1945                 ++sp->root_count;
1946                 vcpu->arch.mmu.root_hpa = root;
1947                 return;
1948         }
1949         metaphysical = !is_paging(vcpu);
1950         if (tdp_enabled)
1951                 metaphysical = 1;
1952         for (i = 0; i < 4; ++i) {
1953                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1954
1955                 ASSERT(!VALID_PAGE(root));
1956                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1957                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1958                                 vcpu->arch.mmu.pae_root[i] = 0;
1959                                 continue;
1960                         }
1961                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1962                 } else if (vcpu->arch.mmu.root_level == 0)
1963                         root_gfn = 0;
1964                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1965                                       PT32_ROOT_LEVEL, metaphysical,
1966                                       ACC_ALL, NULL);
1967                 root = __pa(sp->spt);
1968                 ++sp->root_count;
1969                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1970         }
1971         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1972 }
1973
1974 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
1975 {
1976         int i;
1977         struct kvm_mmu_page *sp;
1978
1979         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1980                 return;
1981         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1982                 hpa_t root = vcpu->arch.mmu.root_hpa;
1983                 sp = page_header(root);
1984                 mmu_sync_children(vcpu, sp);
1985                 return;
1986         }
1987         for (i = 0; i < 4; ++i) {
1988                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1989
1990                 if (root) {
1991                         root &= PT64_BASE_ADDR_MASK;
1992                         sp = page_header(root);
1993                         mmu_sync_children(vcpu, sp);
1994                 }
1995         }
1996 }
1997
1998 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
1999 {
2000         spin_lock(&vcpu->kvm->mmu_lock);
2001         mmu_sync_roots(vcpu);
2002         spin_unlock(&vcpu->kvm->mmu_lock);
2003 }
2004
2005 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2006 {
2007         return vaddr;
2008 }
2009
2010 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2011                                 u32 error_code)
2012 {
2013         gfn_t gfn;
2014         int r;
2015
2016         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2017         r = mmu_topup_memory_caches(vcpu);
2018         if (r)
2019                 return r;
2020
2021         ASSERT(vcpu);
2022         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2023
2024         gfn = gva >> PAGE_SHIFT;
2025
2026         return nonpaging_map(vcpu, gva & PAGE_MASK,
2027                              error_code & PFERR_WRITE_MASK, gfn);
2028 }
2029
2030 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2031                                 u32 error_code)
2032 {
2033         pfn_t pfn;
2034         int r;
2035         int largepage = 0;
2036         gfn_t gfn = gpa >> PAGE_SHIFT;
2037         unsigned long mmu_seq;
2038
2039         ASSERT(vcpu);
2040         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2041
2042         r = mmu_topup_memory_caches(vcpu);
2043         if (r)
2044                 return r;
2045
2046         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
2047                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2048                 largepage = 1;
2049         }
2050         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2051         smp_rmb();
2052         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2053         if (is_error_pfn(pfn)) {
2054                 kvm_release_pfn_clean(pfn);
2055                 return 1;
2056         }
2057         spin_lock(&vcpu->kvm->mmu_lock);
2058         if (mmu_notifier_retry(vcpu, mmu_seq))
2059                 goto out_unlock;
2060         kvm_mmu_free_some_pages(vcpu);
2061         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2062                          largepage, gfn, pfn);
2063         spin_unlock(&vcpu->kvm->mmu_lock);
2064
2065         return r;
2066
2067 out_unlock:
2068         spin_unlock(&vcpu->kvm->mmu_lock);
2069         kvm_release_pfn_clean(pfn);
2070         return 0;
2071 }
2072
2073 static void nonpaging_free(struct kvm_vcpu *vcpu)
2074 {
2075         mmu_free_roots(vcpu);
2076 }
2077
2078 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2079 {
2080         struct kvm_mmu *context = &vcpu->arch.mmu;
2081
2082         context->new_cr3 = nonpaging_new_cr3;
2083         context->page_fault = nonpaging_page_fault;
2084         context->gva_to_gpa = nonpaging_gva_to_gpa;
2085         context->free = nonpaging_free;
2086         context->prefetch_page = nonpaging_prefetch_page;
2087         context->sync_page = nonpaging_sync_page;
2088         context->invlpg = nonpaging_invlpg;
2089         context->root_level = 0;
2090         context->shadow_root_level = PT32E_ROOT_LEVEL;
2091         context->root_hpa = INVALID_PAGE;
2092         return 0;
2093 }
2094
2095 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2096 {
2097         ++vcpu->stat.tlb_flush;
2098         kvm_x86_ops->tlb_flush(vcpu);
2099 }
2100
2101 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2102 {
2103         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2104         mmu_free_roots(vcpu);
2105 }
2106
2107 static void inject_page_fault(struct kvm_vcpu *vcpu,
2108                               u64 addr,
2109                               u32 err_code)
2110 {
2111         kvm_inject_page_fault(vcpu, addr, err_code);
2112 }
2113
2114 static void paging_free(struct kvm_vcpu *vcpu)
2115 {
2116         nonpaging_free(vcpu);
2117 }
2118
2119 #define PTTYPE 64
2120 #include "paging_tmpl.h"
2121 #undef PTTYPE
2122
2123 #define PTTYPE 32
2124 #include "paging_tmpl.h"
2125 #undef PTTYPE
2126
2127 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2128 {
2129         struct kvm_mmu *context = &vcpu->arch.mmu;
2130
2131         ASSERT(is_pae(vcpu));
2132         context->new_cr3 = paging_new_cr3;
2133         context->page_fault = paging64_page_fault;
2134         context->gva_to_gpa = paging64_gva_to_gpa;
2135         context->prefetch_page = paging64_prefetch_page;
2136         context->sync_page = paging64_sync_page;
2137         context->invlpg = paging64_invlpg;
2138         context->free = paging_free;
2139         context->root_level = level;
2140         context->shadow_root_level = level;
2141         context->root_hpa = INVALID_PAGE;
2142         return 0;
2143 }
2144
2145 static int paging64_init_context(struct kvm_vcpu *vcpu)
2146 {
2147         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2148 }
2149
2150 static int paging32_init_context(struct kvm_vcpu *vcpu)
2151 {
2152         struct kvm_mmu *context = &vcpu->arch.mmu;
2153
2154         context->new_cr3 = paging_new_cr3;
2155         context->page_fault = paging32_page_fault;
2156         context->gva_to_gpa = paging32_gva_to_gpa;
2157         context->free = paging_free;
2158         context->prefetch_page = paging32_prefetch_page;
2159         context->sync_page = paging32_sync_page;
2160         context->invlpg = paging32_invlpg;
2161         context->root_level = PT32_ROOT_LEVEL;
2162         context->shadow_root_level = PT32E_ROOT_LEVEL;
2163         context->root_hpa = INVALID_PAGE;
2164         return 0;
2165 }
2166
2167 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2168 {
2169         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2170 }
2171
2172 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2173 {
2174         struct kvm_mmu *context = &vcpu->arch.mmu;
2175
2176         context->new_cr3 = nonpaging_new_cr3;
2177         context->page_fault = tdp_page_fault;
2178         context->free = nonpaging_free;
2179         context->prefetch_page = nonpaging_prefetch_page;
2180         context->sync_page = nonpaging_sync_page;
2181         context->invlpg = nonpaging_invlpg;
2182         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2183         context->root_hpa = INVALID_PAGE;
2184
2185         if (!is_paging(vcpu)) {
2186                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2187                 context->root_level = 0;
2188         } else if (is_long_mode(vcpu)) {
2189                 context->gva_to_gpa = paging64_gva_to_gpa;
2190                 context->root_level = PT64_ROOT_LEVEL;
2191         } else if (is_pae(vcpu)) {
2192                 context->gva_to_gpa = paging64_gva_to_gpa;
2193                 context->root_level = PT32E_ROOT_LEVEL;
2194         } else {
2195                 context->gva_to_gpa = paging32_gva_to_gpa;
2196                 context->root_level = PT32_ROOT_LEVEL;
2197         }
2198
2199         return 0;
2200 }
2201
2202 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2203 {
2204         ASSERT(vcpu);
2205         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2206
2207         if (!is_paging(vcpu))
2208                 return nonpaging_init_context(vcpu);
2209         else if (is_long_mode(vcpu))
2210                 return paging64_init_context(vcpu);
2211         else if (is_pae(vcpu))
2212                 return paging32E_init_context(vcpu);
2213         else
2214                 return paging32_init_context(vcpu);
2215 }
2216
2217 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2218 {
2219         vcpu->arch.update_pte.pfn = bad_pfn;
2220
2221         if (tdp_enabled)
2222                 return init_kvm_tdp_mmu(vcpu);
2223         else
2224                 return init_kvm_softmmu(vcpu);
2225 }
2226
2227 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2228 {
2229         ASSERT(vcpu);
2230         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2231                 vcpu->arch.mmu.free(vcpu);
2232                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2233         }
2234 }
2235
2236 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2237 {
2238         destroy_kvm_mmu(vcpu);
2239         return init_kvm_mmu(vcpu);
2240 }
2241 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2242
2243 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2244 {
2245         int r;
2246
2247         r = mmu_topup_memory_caches(vcpu);
2248         if (r)
2249                 goto out;
2250         spin_lock(&vcpu->kvm->mmu_lock);
2251         kvm_mmu_free_some_pages(vcpu);
2252         mmu_alloc_roots(vcpu);
2253         mmu_sync_roots(vcpu);
2254         spin_unlock(&vcpu->kvm->mmu_lock);
2255         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2256         kvm_mmu_flush_tlb(vcpu);
2257 out:
2258         return r;
2259 }
2260 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2261
2262 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2263 {
2264         mmu_free_roots(vcpu);
2265 }
2266
2267 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2268                                   struct kvm_mmu_page *sp,
2269                                   u64 *spte)
2270 {
2271         u64 pte;
2272         struct kvm_mmu_page *child;
2273
2274         pte = *spte;
2275         if (is_shadow_present_pte(pte)) {
2276                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2277                     is_large_pte(pte))
2278                         rmap_remove(vcpu->kvm, spte);
2279                 else {
2280                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2281                         mmu_page_remove_parent_pte(child, spte);
2282                 }
2283         }
2284         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2285         if (is_large_pte(pte))
2286                 --vcpu->kvm->stat.lpages;
2287 }
2288
2289 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2290                                   struct kvm_mmu_page *sp,
2291                                   u64 *spte,
2292                                   const void *new)
2293 {
2294         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2295                 if (!vcpu->arch.update_pte.largepage ||
2296                     sp->role.glevels == PT32_ROOT_LEVEL) {
2297                         ++vcpu->kvm->stat.mmu_pde_zapped;
2298                         return;
2299                 }
2300         }
2301
2302         ++vcpu->kvm->stat.mmu_pte_updated;
2303         if (sp->role.glevels == PT32_ROOT_LEVEL)
2304                 paging32_update_pte(vcpu, sp, spte, new);
2305         else
2306                 paging64_update_pte(vcpu, sp, spte, new);
2307 }
2308
2309 static bool need_remote_flush(u64 old, u64 new)
2310 {
2311         if (!is_shadow_present_pte(old))
2312                 return false;
2313         if (!is_shadow_present_pte(new))
2314                 return true;
2315         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2316                 return true;
2317         old ^= PT64_NX_MASK;
2318         new ^= PT64_NX_MASK;
2319         return (old & ~new & PT64_PERM_MASK) != 0;
2320 }
2321
2322 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2323 {
2324         if (need_remote_flush(old, new))
2325                 kvm_flush_remote_tlbs(vcpu->kvm);
2326         else
2327                 kvm_mmu_flush_tlb(vcpu);
2328 }
2329
2330 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2331 {
2332         u64 *spte = vcpu->arch.last_pte_updated;
2333
2334         return !!(spte && (*spte & shadow_accessed_mask));
2335 }
2336
2337 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2338                                           const u8 *new, int bytes)
2339 {
2340         gfn_t gfn;
2341         int r;
2342         u64 gpte = 0;
2343         pfn_t pfn;
2344
2345         vcpu->arch.update_pte.largepage = 0;
2346
2347         if (bytes != 4 && bytes != 8)
2348                 return;
2349
2350         /*
2351          * Assume that the pte write on a page table of the same type
2352          * as the current vcpu paging mode.  This is nearly always true
2353          * (might be false while changing modes).  Note it is verified later
2354          * by update_pte().
2355          */
2356         if (is_pae(vcpu)) {
2357                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2358                 if ((bytes == 4) && (gpa % 4 == 0)) {
2359                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2360                         if (r)
2361                                 return;
2362                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2363                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2364                         memcpy((void *)&gpte, new, 8);
2365                 }
2366         } else {
2367                 if ((bytes == 4) && (gpa % 4 == 0))
2368                         memcpy((void *)&gpte, new, 4);
2369         }
2370         if (!is_present_pte(gpte))
2371                 return;
2372         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2373
2374         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2375                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2376                 vcpu->arch.update_pte.largepage = 1;
2377         }
2378         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2379         smp_rmb();
2380         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2381
2382         if (is_error_pfn(pfn)) {
2383                 kvm_release_pfn_clean(pfn);
2384                 return;
2385         }
2386         vcpu->arch.update_pte.gfn = gfn;
2387         vcpu->arch.update_pte.pfn = pfn;
2388 }
2389
2390 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2391 {
2392         u64 *spte = vcpu->arch.last_pte_updated;
2393
2394         if (spte
2395             && vcpu->arch.last_pte_gfn == gfn
2396             && shadow_accessed_mask
2397             && !(*spte & shadow_accessed_mask)
2398             && is_shadow_present_pte(*spte))
2399                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2400 }
2401
2402 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2403                        const u8 *new, int bytes)
2404 {
2405         gfn_t gfn = gpa >> PAGE_SHIFT;
2406         struct kvm_mmu_page *sp;
2407         struct hlist_node *node, *n;
2408         struct hlist_head *bucket;
2409         unsigned index;
2410         u64 entry, gentry;
2411         u64 *spte;
2412         unsigned offset = offset_in_page(gpa);
2413         unsigned pte_size;
2414         unsigned page_offset;
2415         unsigned misaligned;
2416         unsigned quadrant;
2417         int level;
2418         int flooded = 0;
2419         int npte;
2420         int r;
2421
2422         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2423         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2424         spin_lock(&vcpu->kvm->mmu_lock);
2425         kvm_mmu_access_page(vcpu, gfn);
2426         kvm_mmu_free_some_pages(vcpu);
2427         ++vcpu->kvm->stat.mmu_pte_write;
2428         kvm_mmu_audit(vcpu, "pre pte write");
2429         if (gfn == vcpu->arch.last_pt_write_gfn
2430             && !last_updated_pte_accessed(vcpu)) {
2431                 ++vcpu->arch.last_pt_write_count;
2432                 if (vcpu->arch.last_pt_write_count >= 3)
2433                         flooded = 1;
2434         } else {
2435                 vcpu->arch.last_pt_write_gfn = gfn;
2436                 vcpu->arch.last_pt_write_count = 1;
2437                 vcpu->arch.last_pte_updated = NULL;
2438         }
2439         index = kvm_page_table_hashfn(gfn);
2440         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2441         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2442                 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
2443                         continue;
2444                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2445                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2446                 misaligned |= bytes < 4;
2447                 if (misaligned || flooded) {
2448                         /*
2449                          * Misaligned accesses are too much trouble to fix
2450                          * up; also, they usually indicate a page is not used
2451                          * as a page table.
2452                          *
2453                          * If we're seeing too many writes to a page,
2454                          * it may no longer be a page table, or we may be
2455                          * forking, in which case it is better to unmap the
2456                          * page.
2457                          */
2458                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2459                                  gpa, bytes, sp->role.word);
2460                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2461                                 n = bucket->first;
2462                         ++vcpu->kvm->stat.mmu_flooded;
2463                         continue;
2464                 }
2465                 page_offset = offset;
2466                 level = sp->role.level;
2467                 npte = 1;
2468                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2469                         page_offset <<= 1;      /* 32->64 */
2470                         /*
2471                          * A 32-bit pde maps 4MB while the shadow pdes map
2472                          * only 2MB.  So we need to double the offset again
2473                          * and zap two pdes instead of one.
2474                          */
2475                         if (level == PT32_ROOT_LEVEL) {
2476                                 page_offset &= ~7; /* kill rounding error */
2477                                 page_offset <<= 1;
2478                                 npte = 2;
2479                         }
2480                         quadrant = page_offset >> PAGE_SHIFT;
2481                         page_offset &= ~PAGE_MASK;
2482                         if (quadrant != sp->role.quadrant)
2483                                 continue;
2484                 }
2485                 spte = &sp->spt[page_offset / sizeof(*spte)];
2486                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2487                         gentry = 0;
2488                         r = kvm_read_guest_atomic(vcpu->kvm,
2489                                                   gpa & ~(u64)(pte_size - 1),
2490                                                   &gentry, pte_size);
2491                         new = (const void *)&gentry;
2492                         if (r < 0)
2493                                 new = NULL;
2494                 }
2495                 while (npte--) {
2496                         entry = *spte;
2497                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2498                         if (new)
2499                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2500                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2501                         ++spte;
2502                 }
2503         }
2504         kvm_mmu_audit(vcpu, "post pte write");
2505         spin_unlock(&vcpu->kvm->mmu_lock);
2506         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2507                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2508                 vcpu->arch.update_pte.pfn = bad_pfn;
2509         }
2510 }
2511
2512 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2513 {
2514         gpa_t gpa;
2515         int r;
2516
2517         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2518
2519         spin_lock(&vcpu->kvm->mmu_lock);
2520         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2521         spin_unlock(&vcpu->kvm->mmu_lock);
2522         return r;
2523 }
2524 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2525
2526 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2527 {
2528         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2529                 struct kvm_mmu_page *sp;
2530
2531                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2532                                   struct kvm_mmu_page, link);
2533                 kvm_mmu_zap_page(vcpu->kvm, sp);
2534                 ++vcpu->kvm->stat.mmu_recycled;
2535         }
2536 }
2537
2538 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2539 {
2540         int r;
2541         enum emulation_result er;
2542
2543         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2544         if (r < 0)
2545                 goto out;
2546
2547         if (!r) {
2548                 r = 1;
2549                 goto out;
2550         }
2551
2552         r = mmu_topup_memory_caches(vcpu);
2553         if (r)
2554                 goto out;
2555
2556         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2557
2558         switch (er) {
2559         case EMULATE_DONE:
2560                 return 1;
2561         case EMULATE_DO_MMIO:
2562                 ++vcpu->stat.mmio_exits;
2563                 return 0;
2564         case EMULATE_FAIL:
2565                 kvm_report_emulation_failure(vcpu, "pagetable");
2566                 return 1;
2567         default:
2568                 BUG();
2569         }
2570 out:
2571         return r;
2572 }
2573 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2574
2575 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2576 {
2577         spin_lock(&vcpu->kvm->mmu_lock);
2578         vcpu->arch.mmu.invlpg(vcpu, gva);
2579         spin_unlock(&vcpu->kvm->mmu_lock);
2580         kvm_mmu_flush_tlb(vcpu);
2581         ++vcpu->stat.invlpg;
2582 }
2583 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2584
2585 void kvm_enable_tdp(void)
2586 {
2587         tdp_enabled = true;
2588 }
2589 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2590
2591 void kvm_disable_tdp(void)
2592 {
2593         tdp_enabled = false;
2594 }
2595 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2596
2597 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2598 {
2599         struct kvm_mmu_page *sp;
2600
2601         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2602                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2603                                   struct kvm_mmu_page, link);
2604                 kvm_mmu_zap_page(vcpu->kvm, sp);
2605                 cond_resched();
2606         }
2607         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2608 }
2609
2610 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2611 {
2612         struct page *page;
2613         int i;
2614
2615         ASSERT(vcpu);
2616
2617         if (vcpu->kvm->arch.n_requested_mmu_pages)
2618                 vcpu->kvm->arch.n_free_mmu_pages =
2619                                         vcpu->kvm->arch.n_requested_mmu_pages;
2620         else
2621                 vcpu->kvm->arch.n_free_mmu_pages =
2622                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2623         /*
2624          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2625          * Therefore we need to allocate shadow page tables in the first
2626          * 4GB of memory, which happens to fit the DMA32 zone.
2627          */
2628         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2629         if (!page)
2630                 goto error_1;
2631         vcpu->arch.mmu.pae_root = page_address(page);
2632         for (i = 0; i < 4; ++i)
2633                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2634
2635         return 0;
2636
2637 error_1:
2638         free_mmu_pages(vcpu);
2639         return -ENOMEM;
2640 }
2641
2642 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2643 {
2644         ASSERT(vcpu);
2645         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2646
2647         return alloc_mmu_pages(vcpu);
2648 }
2649
2650 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2651 {
2652         ASSERT(vcpu);
2653         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2654
2655         return init_kvm_mmu(vcpu);
2656 }
2657
2658 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2659 {
2660         ASSERT(vcpu);
2661
2662         destroy_kvm_mmu(vcpu);
2663         free_mmu_pages(vcpu);
2664         mmu_free_memory_caches(vcpu);
2665 }
2666
2667 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2668 {
2669         struct kvm_mmu_page *sp;
2670
2671         spin_lock(&kvm->mmu_lock);
2672         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2673                 int i;
2674                 u64 *pt;
2675
2676                 if (!test_bit(slot, sp->slot_bitmap))
2677                         continue;
2678
2679                 pt = sp->spt;
2680                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2681                         /* avoid RMW */
2682                         if (pt[i] & PT_WRITABLE_MASK)
2683                                 pt[i] &= ~PT_WRITABLE_MASK;
2684         }
2685         kvm_flush_remote_tlbs(kvm);
2686         spin_unlock(&kvm->mmu_lock);
2687 }
2688
2689 void kvm_mmu_zap_all(struct kvm *kvm)
2690 {
2691         struct kvm_mmu_page *sp, *node;
2692
2693         spin_lock(&kvm->mmu_lock);
2694         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2695                 if (kvm_mmu_zap_page(kvm, sp))
2696                         node = container_of(kvm->arch.active_mmu_pages.next,
2697                                             struct kvm_mmu_page, link);
2698         spin_unlock(&kvm->mmu_lock);
2699
2700         kvm_flush_remote_tlbs(kvm);
2701 }
2702
2703 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2704 {
2705         struct kvm_mmu_page *page;
2706
2707         page = container_of(kvm->arch.active_mmu_pages.prev,
2708                             struct kvm_mmu_page, link);
2709         kvm_mmu_zap_page(kvm, page);
2710 }
2711
2712 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2713 {
2714         struct kvm *kvm;
2715         struct kvm *kvm_freed = NULL;
2716         int cache_count = 0;
2717
2718         spin_lock(&kvm_lock);
2719
2720         list_for_each_entry(kvm, &vm_list, vm_list) {
2721                 int npages;
2722
2723                 if (!down_read_trylock(&kvm->slots_lock))
2724                         continue;
2725                 spin_lock(&kvm->mmu_lock);
2726                 npages = kvm->arch.n_alloc_mmu_pages -
2727                          kvm->arch.n_free_mmu_pages;
2728                 cache_count += npages;
2729                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2730                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2731                         cache_count--;
2732                         kvm_freed = kvm;
2733                 }
2734                 nr_to_scan--;
2735
2736                 spin_unlock(&kvm->mmu_lock);
2737                 up_read(&kvm->slots_lock);
2738         }
2739         if (kvm_freed)
2740                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2741
2742         spin_unlock(&kvm_lock);
2743
2744         return cache_count;
2745 }
2746
2747 static struct shrinker mmu_shrinker = {
2748         .shrink = mmu_shrink,
2749         .seeks = DEFAULT_SEEKS * 10,
2750 };
2751
2752 static void mmu_destroy_caches(void)
2753 {
2754         if (pte_chain_cache)
2755                 kmem_cache_destroy(pte_chain_cache);
2756         if (rmap_desc_cache)
2757                 kmem_cache_destroy(rmap_desc_cache);
2758         if (mmu_page_header_cache)
2759                 kmem_cache_destroy(mmu_page_header_cache);
2760 }
2761
2762 void kvm_mmu_module_exit(void)
2763 {
2764         mmu_destroy_caches();
2765         unregister_shrinker(&mmu_shrinker);
2766 }
2767
2768 int kvm_mmu_module_init(void)
2769 {
2770         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2771                                             sizeof(struct kvm_pte_chain),
2772                                             0, 0, NULL);
2773         if (!pte_chain_cache)
2774                 goto nomem;
2775         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2776                                             sizeof(struct kvm_rmap_desc),
2777                                             0, 0, NULL);
2778         if (!rmap_desc_cache)
2779                 goto nomem;
2780
2781         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2782                                                   sizeof(struct kvm_mmu_page),
2783                                                   0, 0, NULL);
2784         if (!mmu_page_header_cache)
2785                 goto nomem;
2786
2787         register_shrinker(&mmu_shrinker);
2788
2789         return 0;
2790
2791 nomem:
2792         mmu_destroy_caches();
2793         return -ENOMEM;
2794 }
2795
2796 /*
2797  * Caculate mmu pages needed for kvm.
2798  */
2799 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2800 {
2801         int i;
2802         unsigned int nr_mmu_pages;
2803         unsigned int  nr_pages = 0;
2804
2805         for (i = 0; i < kvm->nmemslots; i++)
2806                 nr_pages += kvm->memslots[i].npages;
2807
2808         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2809         nr_mmu_pages = max(nr_mmu_pages,
2810                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2811
2812         return nr_mmu_pages;
2813 }
2814
2815 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2816                                 unsigned len)
2817 {
2818         if (len > buffer->len)
2819                 return NULL;
2820         return buffer->ptr;
2821 }
2822
2823 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2824                                 unsigned len)
2825 {
2826         void *ret;
2827
2828         ret = pv_mmu_peek_buffer(buffer, len);
2829         if (!ret)
2830                 return ret;
2831         buffer->ptr += len;
2832         buffer->len -= len;
2833         buffer->processed += len;
2834         return ret;
2835 }
2836
2837 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2838                              gpa_t addr, gpa_t value)
2839 {
2840         int bytes = 8;
2841         int r;
2842
2843         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2844                 bytes = 4;
2845
2846         r = mmu_topup_memory_caches(vcpu);
2847         if (r)
2848                 return r;
2849
2850         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2851                 return -EFAULT;
2852
2853         return 1;
2854 }
2855
2856 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2857 {
2858         kvm_x86_ops->tlb_flush(vcpu);
2859         set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
2860         return 1;
2861 }
2862
2863 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2864 {
2865         spin_lock(&vcpu->kvm->mmu_lock);
2866         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2867         spin_unlock(&vcpu->kvm->mmu_lock);
2868         return 1;
2869 }
2870
2871 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2872                              struct kvm_pv_mmu_op_buffer *buffer)
2873 {
2874         struct kvm_mmu_op_header *header;
2875
2876         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2877         if (!header)
2878                 return 0;
2879         switch (header->op) {
2880         case KVM_MMU_OP_WRITE_PTE: {
2881                 struct kvm_mmu_op_write_pte *wpte;
2882
2883                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2884                 if (!wpte)
2885                         return 0;
2886                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2887                                         wpte->pte_val);
2888         }
2889         case KVM_MMU_OP_FLUSH_TLB: {
2890                 struct kvm_mmu_op_flush_tlb *ftlb;
2891
2892                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2893                 if (!ftlb)
2894                         return 0;
2895                 return kvm_pv_mmu_flush_tlb(vcpu);
2896         }
2897         case KVM_MMU_OP_RELEASE_PT: {
2898                 struct kvm_mmu_op_release_pt *rpt;
2899
2900                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2901                 if (!rpt)
2902                         return 0;
2903                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2904         }
2905         default: return 0;
2906         }
2907 }
2908
2909 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2910                   gpa_t addr, unsigned long *ret)
2911 {
2912         int r;
2913         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2914
2915         buffer->ptr = buffer->buf;
2916         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2917         buffer->processed = 0;
2918
2919         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2920         if (r)
2921                 goto out;
2922
2923         while (buffer->len) {
2924                 r = kvm_pv_mmu_op_one(vcpu, buffer);
2925                 if (r < 0)
2926                         goto out;
2927                 if (r == 0)
2928                         break;
2929         }
2930
2931         r = 1;
2932 out:
2933         *ret = buffer->processed;
2934         return r;
2935 }
2936
2937 #ifdef AUDIT
2938
2939 static const char *audit_msg;
2940
2941 static gva_t canonicalize(gva_t gva)
2942 {
2943 #ifdef CONFIG_X86_64
2944         gva = (long long)(gva << 16) >> 16;
2945 #endif
2946         return gva;
2947 }
2948
2949 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2950                                 gva_t va, int level)
2951 {
2952         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2953         int i;
2954         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2955
2956         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2957                 u64 ent = pt[i];
2958
2959                 if (ent == shadow_trap_nonpresent_pte)
2960                         continue;
2961
2962                 va = canonicalize(va);
2963                 if (level > 1) {
2964                         if (ent == shadow_notrap_nonpresent_pte)
2965                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2966                                        " in nonleaf level: levels %d gva %lx"
2967                                        " level %d pte %llx\n", audit_msg,
2968                                        vcpu->arch.mmu.root_level, va, level, ent);
2969
2970                         audit_mappings_page(vcpu, ent, va, level - 1);
2971                 } else {
2972                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2973                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2974
2975                         if (is_shadow_present_pte(ent)
2976                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2977                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2978                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2979                                        audit_msg, vcpu->arch.mmu.root_level,
2980                                        va, gpa, hpa, ent,
2981                                        is_shadow_present_pte(ent));
2982                         else if (ent == shadow_notrap_nonpresent_pte
2983                                  && !is_error_hpa(hpa))
2984                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2985                                        " valid guest gva %lx\n", audit_msg, va);
2986                         kvm_release_pfn_clean(pfn);
2987
2988                 }
2989         }
2990 }
2991
2992 static void audit_mappings(struct kvm_vcpu *vcpu)
2993 {
2994         unsigned i;
2995
2996         if (vcpu->arch.mmu.root_level == 4)
2997                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2998         else
2999                 for (i = 0; i < 4; ++i)
3000                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3001                                 audit_mappings_page(vcpu,
3002                                                     vcpu->arch.mmu.pae_root[i],
3003                                                     i << 30,
3004                                                     2);
3005 }
3006
3007 static int count_rmaps(struct kvm_vcpu *vcpu)
3008 {
3009         int nmaps = 0;
3010         int i, j, k;
3011
3012         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3013                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3014                 struct kvm_rmap_desc *d;
3015
3016                 for (j = 0; j < m->npages; ++j) {
3017                         unsigned long *rmapp = &m->rmap[j];
3018
3019                         if (!*rmapp)
3020                                 continue;
3021                         if (!(*rmapp & 1)) {
3022                                 ++nmaps;
3023                                 continue;
3024                         }
3025                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3026                         while (d) {
3027                                 for (k = 0; k < RMAP_EXT; ++k)
3028                                         if (d->shadow_ptes[k])
3029                                                 ++nmaps;
3030                                         else
3031                                                 break;
3032                                 d = d->more;
3033                         }
3034                 }
3035         }
3036         return nmaps;
3037 }
3038
3039 static int count_writable_mappings(struct kvm_vcpu *vcpu)
3040 {
3041         int nmaps = 0;
3042         struct kvm_mmu_page *sp;
3043         int i;
3044
3045         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3046                 u64 *pt = sp->spt;
3047
3048                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3049                         continue;
3050
3051                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3052                         u64 ent = pt[i];
3053
3054                         if (!(ent & PT_PRESENT_MASK))
3055                                 continue;
3056                         if (!(ent & PT_WRITABLE_MASK))
3057                                 continue;
3058                         ++nmaps;
3059                 }
3060         }
3061         return nmaps;
3062 }
3063
3064 static void audit_rmap(struct kvm_vcpu *vcpu)
3065 {
3066         int n_rmap = count_rmaps(vcpu);
3067         int n_actual = count_writable_mappings(vcpu);
3068
3069         if (n_rmap != n_actual)
3070                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
3071                        __func__, audit_msg, n_rmap, n_actual);
3072 }
3073
3074 static void audit_write_protection(struct kvm_vcpu *vcpu)
3075 {
3076         struct kvm_mmu_page *sp;
3077         struct kvm_memory_slot *slot;
3078         unsigned long *rmapp;
3079         gfn_t gfn;
3080
3081         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3082                 if (sp->role.metaphysical)
3083                         continue;
3084
3085                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3086                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3087                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3088                 if (*rmapp)
3089                         printk(KERN_ERR "%s: (%s) shadow page has writable"
3090                                " mappings: gfn %lx role %x\n",
3091                                __func__, audit_msg, sp->gfn,
3092                                sp->role.word);
3093         }
3094 }
3095
3096 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3097 {
3098         int olddbg = dbg;
3099
3100         dbg = 0;
3101         audit_msg = msg;
3102         audit_rmap(vcpu);
3103         audit_write_protection(vcpu);
3104         audit_mappings(vcpu);
3105         dbg = olddbg;
3106 }
3107
3108 #endif