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 #include "kvm_cache_regs.h"
  22
  23 #include <linux/kvm_host.h>
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29 #include <linux/swap.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/compiler.h>
  32
  33 #include <asm/page.h>
  34 #include <asm/cmpxchg.h>
  35 #include <asm/io.h>
  36 #include <asm/vmx.h>
  37
  38 /*
  39  * When setting this variable to true it enables Two-Dimensional-Paging
  40  * where the hardware walks 2 page tables:
  41  * 1. the guest-virtual to guest-physical
  42  * 2. while doing 1. it walks guest-physical to host-physical
  43  * If the hardware supports that we don't need to do shadow paging.
  44  */
  45 bool tdp_enabled = false;
  46
  47 #undef MMU_DEBUG
  48
  49 #undef AUDIT
  50
  51 #ifdef AUDIT
  52 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  53 #else
  54 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  55 #endif
  56
  57 #ifdef MMU_DEBUG
  58
  59 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  60 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  61
  62 #else
  63
  64 #define pgprintk(x...) do { } while (0)
  65 #define rmap_printk(x...) do { } while (0)
  66
  67 #endif
  68
  69 #if defined(MMU_DEBUG) || defined(AUDIT)
  70 static int dbg = 0;
  71 module_param(dbg, bool, 0644);
  72 #endif
  73
  74 static int oos_shadow = 1;
  75 module_param(oos_shadow, bool, 0644);
  76
  77 #ifndef MMU_DEBUG
  78 #define ASSERT(x) do { } while (0)
  79 #else
  80 #define ASSERT(x)                                                       \
  81         if (!(x)) {                                                     \
  82                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  83                        __FILE__, __LINE__, #x);                         \
  84         }
  85 #endif
  86
  87 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  88 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  89
  90 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  91
  92 #define PT64_LEVEL_BITS 9
  93
  94 #define PT64_LEVEL_SHIFT(level) \
  95                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  96
  97 #define PT64_LEVEL_MASK(level) \
  98                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  99
 100 #define PT64_INDEX(address, level)\
 101         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 102
 103
 104 #define PT32_LEVEL_BITS 10
 105
 106 #define PT32_LEVEL_SHIFT(level) \
 107                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 108
 109 #define PT32_LEVEL_MASK(level) \
 110                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 111 #define PT32_LVL_OFFSET_MASK(level) \
 112         (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 113                                                 * PT32_LEVEL_BITS))) - 1))
 114
 115 #define PT32_INDEX(address, level)\
 116         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 117
 118
 119 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 120 #define PT64_DIR_BASE_ADDR_MASK \
 121         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 122 #define PT64_LVL_ADDR_MASK(level) \
 123         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 124                                                 * PT64_LEVEL_BITS))) - 1))
 125 #define PT64_LVL_OFFSET_MASK(level) \
 126         (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 127                                                 * PT64_LEVEL_BITS))) - 1))
 128
 129 #define PT32_BASE_ADDR_MASK PAGE_MASK
 130 #define PT32_DIR_BASE_ADDR_MASK \
 131         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 132 #define PT32_LVL_ADDR_MASK(level) \
 133         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 134                                             * PT32_LEVEL_BITS))) - 1))
 135
 136 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 137                         | PT64_NX_MASK)
 138
 139 #define PFERR_PRESENT_MASK (1U << 0)
 140 #define PFERR_WRITE_MASK (1U << 1)
 141 #define PFERR_USER_MASK (1U << 2)
 142 #define PFERR_RSVD_MASK (1U << 3)
 143 #define PFERR_FETCH_MASK (1U << 4)
 144
 145 #define PT_PDPE_LEVEL 3
 146 #define PT_DIRECTORY_LEVEL 2
 147 #define PT_PAGE_TABLE_LEVEL 1
 148
 149 #define RMAP_EXT 4
 150
 151 #define ACC_EXEC_MASK    1
 152 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 153 #define ACC_USER_MASK    PT_USER_MASK
 154 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 155
 156 #define CREATE_TRACE_POINTS
 157 #include "mmutrace.h"
 158
 159 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 160
 161 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 162
 163 struct kvm_rmap_desc {
 164         u64 *sptes[RMAP_EXT];
 165         struct kvm_rmap_desc *more;
 166 };
 167
 168 struct kvm_shadow_walk_iterator {
 169         u64 addr;
 170         hpa_t shadow_addr;
 171         int level;
 172         u64 *sptep;
 173         unsigned index;
 174 };
 175
 176 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
 177         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
 178              shadow_walk_okay(&(_walker));                      \
 179              shadow_walk_next(&(_walker)))
 180
 181
 182 struct kvm_unsync_walk {
 183         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 184 };
 185
 186 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 187
 188 static struct kmem_cache *pte_chain_cache;
 189 static struct kmem_cache *rmap_desc_cache;
 190 static struct kmem_cache *mmu_page_header_cache;
 191
 192 static u64 __read_mostly shadow_trap_nonpresent_pte;
 193 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 194 static u64 __read_mostly shadow_base_present_pte;
 195 static u64 __read_mostly shadow_nx_mask;
 196 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 197 static u64 __read_mostly shadow_user_mask;
 198 static u64 __read_mostly shadow_accessed_mask;
 199 static u64 __read_mostly shadow_dirty_mask;
 200
 201 static inline u64 rsvd_bits(int s, int e)
 202 {
 203         return ((1ULL << (e - s + 1)) - 1) << s;
 204 }
 205
 206 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 207 {
 208         shadow_trap_nonpresent_pte = trap_pte;
 209         shadow_notrap_nonpresent_pte = notrap_pte;
 210 }
 211 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 212
 213 void kvm_mmu_set_base_ptes(u64 base_pte)
 214 {
 215         shadow_base_present_pte = base_pte;
 216 }
 217 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 218
 219 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 220                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 221 {
 222         shadow_user_mask = user_mask;
 223         shadow_accessed_mask = accessed_mask;
 224         shadow_dirty_mask = dirty_mask;
 225         shadow_nx_mask = nx_mask;
 226         shadow_x_mask = x_mask;
 227 }
 228 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 229
 230 static int is_write_protection(struct kvm_vcpu *vcpu)
 231 {
 232         return vcpu->arch.cr0 & X86_CR0_WP;
 233 }
 234
 235 static int is_cpuid_PSE36(void)
 236 {
 237         return 1;
 238 }
 239
 240 static int is_nx(struct kvm_vcpu *vcpu)
 241 {
 242         return vcpu->arch.shadow_efer & EFER_NX;
 243 }
 244
 245 static int is_shadow_present_pte(u64 pte)
 246 {
 247         return pte != shadow_trap_nonpresent_pte
 248                 && pte != shadow_notrap_nonpresent_pte;
 249 }
 250
 251 static int is_large_pte(u64 pte)
 252 {
 253         return pte & PT_PAGE_SIZE_MASK;
 254 }
 255
 256 static int is_writeble_pte(unsigned long pte)
 257 {
 258         return pte & PT_WRITABLE_MASK;
 259 }
 260
 261 static int is_dirty_gpte(unsigned long pte)
 262 {
 263         return pte & PT_DIRTY_MASK;
 264 }
 265
 266 static int is_rmap_spte(u64 pte)
 267 {
 268         return is_shadow_present_pte(pte);
 269 }
 270
 271 static int is_last_spte(u64 pte, int level)
 272 {
 273         if (level == PT_PAGE_TABLE_LEVEL)
 274                 return 1;
 275         if (is_large_pte(pte))
 276                 return 1;
 277         return 0;
 278 }
 279
 280 static pfn_t spte_to_pfn(u64 pte)
 281 {
 282         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 283 }
 284
 285 static gfn_t pse36_gfn_delta(u32 gpte)
 286 {
 287         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 288
 289         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 290 }
 291
 292 static void __set_spte(u64 *sptep, u64 spte)
 293 {
 294 #ifdef CONFIG_X86_64
 295         set_64bit((unsigned long *)sptep, spte);
 296 #else
 297         set_64bit((unsigned long long *)sptep, spte);
 298 #endif
 299 }
 300
 301 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 302                                   struct kmem_cache *base_cache, int min)
 303 {
 304         void *obj;
 305
 306         if (cache->nobjs >= min)
 307                 return 0;
 308         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 309                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 310                 if (!obj)
 311                         return -ENOMEM;
 312                 cache->objects[cache->nobjs++] = obj;
 313         }
 314         return 0;
 315 }
 316
 317 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 318 {
 319         while (mc->nobjs)
 320                 kfree(mc->objects[--mc->nobjs]);
 321 }
 322
 323 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 324                                        int min)
 325 {
 326         struct page *page;
 327
 328         if (cache->nobjs >= min)
 329                 return 0;
 330         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 331                 page = alloc_page(GFP_KERNEL);
 332                 if (!page)
 333                         return -ENOMEM;
 334                 set_page_private(page, 0);
 335                 cache->objects[cache->nobjs++] = page_address(page);
 336         }
 337         return 0;
 338 }
 339
 340 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 341 {
 342         while (mc->nobjs)
 343                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 344 }
 345
 346 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 347 {
 348         int r;
 349
 350         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 351                                    pte_chain_cache, 4);
 352         if (r)
 353                 goto out;
 354         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 355                                    rmap_desc_cache, 4);
 356         if (r)
 357                 goto out;
 358         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 359         if (r)
 360                 goto out;
 361         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 362                                    mmu_page_header_cache, 4);
 363 out:
 364         return r;
 365 }
 366
 367 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 368 {
 369         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 370         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 371         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 372         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 373 }
 374
 375 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 376                                     size_t size)
 377 {
 378         void *p;
 379
 380         BUG_ON(!mc->nobjs);
 381         p = mc->objects[--mc->nobjs];
 382         return p;
 383 }
 384
 385 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 386 {
 387         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 388                                       sizeof(struct kvm_pte_chain));
 389 }
 390
 391 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 392 {
 393         kfree(pc);
 394 }
 395
 396 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 397 {
 398         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 399                                       sizeof(struct kvm_rmap_desc));
 400 }
 401
 402 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 403 {
 404         kfree(rd);
 405 }
 406
 407 /*
 408  * Return the pointer to the largepage write count for a given
 409  * gfn, handling slots that are not large page aligned.
 410  */
 411 static int *slot_largepage_idx(gfn_t gfn,
 412                                struct kvm_memory_slot *slot,
 413                                int level)
 414 {
 415         unsigned long idx;
 416
 417         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 418               (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 419         return &slot->lpage_info[level - 2][idx].write_count;
 420 }
 421
 422 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 423 {
 424         struct kvm_memory_slot *slot;
 425         int *write_count;
 426         int i;
 427
 428         gfn = unalias_gfn(kvm, gfn);
 429
 430         slot = gfn_to_memslot_unaliased(kvm, gfn);
 431         for (i = PT_DIRECTORY_LEVEL;
 432              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 433                 write_count   = slot_largepage_idx(gfn, slot, i);
 434                 *write_count += 1;
 435         }
 436 }
 437
 438 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 439 {
 440         struct kvm_memory_slot *slot;
 441         int *write_count;
 442         int i;
 443
 444         gfn = unalias_gfn(kvm, gfn);
 445         for (i = PT_DIRECTORY_LEVEL;
 446              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 447                 slot          = gfn_to_memslot_unaliased(kvm, gfn);
 448                 write_count   = slot_largepage_idx(gfn, slot, i);
 449                 *write_count -= 1;
 450                 WARN_ON(*write_count < 0);
 451         }
 452 }
 453
 454 static int has_wrprotected_page(struct kvm *kvm,
 455                                 gfn_t gfn,
 456                                 int level)
 457 {
 458         struct kvm_memory_slot *slot;
 459         int *largepage_idx;
 460
 461         gfn = unalias_gfn(kvm, gfn);
 462         slot = gfn_to_memslot_unaliased(kvm, gfn);
 463         if (slot) {
 464                 largepage_idx = slot_largepage_idx(gfn, slot, level);
 465                 return *largepage_idx;
 466         }
 467
 468         return 1;
 469 }
 470
 471 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
 472 {
 473         unsigned long page_size = PAGE_SIZE;
 474         struct vm_area_struct *vma;
 475         unsigned long addr;
 476         int i, ret = 0;
 477
 478         addr = gfn_to_hva(kvm, gfn);
 479         if (kvm_is_error_hva(addr))
 480                 return PT_PAGE_TABLE_LEVEL;
 481
 482         down_read(&current->mm->mmap_sem);
 483         vma = find_vma(current->mm, addr);
 484         if (!vma)
 485                 goto out;
 486
 487         page_size = vma_kernel_pagesize(vma);
 488
 489 out:
 490         up_read(&current->mm->mmap_sem);
 491
 492         for (i = PT_PAGE_TABLE_LEVEL;
 493              i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
 494                 if (page_size >= KVM_HPAGE_SIZE(i))
 495                         ret = i;
 496                 else
 497                         break;
 498         }
 499
 500         return ret;
 501 }
 502
 503 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 504 {
 505         struct kvm_memory_slot *slot;
 506         int host_level;
 507         int level = PT_PAGE_TABLE_LEVEL;
 508
 509         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 510         if (slot && slot->dirty_bitmap)
 511                 return PT_PAGE_TABLE_LEVEL;
 512
 513         host_level = host_mapping_level(vcpu->kvm, large_gfn);
 514
 515         if (host_level == PT_PAGE_TABLE_LEVEL)
 516                 return host_level;
 517
 518         for (level = PT_DIRECTORY_LEVEL; level <= host_level; ++level)
 519                 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
 520                         break;
 521
 522         return level - 1;
 523 }
 524
 525 /*
 526  * Take gfn and return the reverse mapping to it.
 527  * Note: gfn must be unaliased before this function get called
 528  */
 529
 530 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
 531 {
 532         struct kvm_memory_slot *slot;
 533         unsigned long idx;
 534
 535         slot = gfn_to_memslot(kvm, gfn);
 536         if (likely(level == PT_PAGE_TABLE_LEVEL))
 537                 return &slot->rmap[gfn - slot->base_gfn];
 538
 539         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 540                 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 541
 542         return &slot->lpage_info[level - 2][idx].rmap_pde;
 543 }
 544
 545 /*
 546  * Reverse mapping data structures:
 547  *
 548  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 549  * that points to page_address(page).
 550  *
 551  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 552  * containing more mappings.
 553  *
 554  * Returns the number of rmap entries before the spte was added or zero if
 555  * the spte was not added.
 556  *
 557  */
 558 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 559 {
 560         struct kvm_mmu_page *sp;
 561         struct kvm_rmap_desc *desc;
 562         unsigned long *rmapp;
 563         int i, count = 0;
 564
 565         if (!is_rmap_spte(*spte))
 566                 return count;
 567         gfn = unalias_gfn(vcpu->kvm, gfn);
 568         sp = page_header(__pa(spte));
 569         sp->gfns[spte - sp->spt] = gfn;
 570         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 571         if (!*rmapp) {
 572                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 573                 *rmapp = (unsigned long)spte;
 574         } else if (!(*rmapp & 1)) {
 575                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 576                 desc = mmu_alloc_rmap_desc(vcpu);
 577                 desc->sptes[0] = (u64 *)*rmapp;
 578                 desc->sptes[1] = spte;
 579                 *rmapp = (unsigned long)desc | 1;
 580         } else {
 581                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 582                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 583                 while (desc->sptes[RMAP_EXT-1] && desc->more) {
 584                         desc = desc->more;
 585                         count += RMAP_EXT;
 586                 }
 587                 if (desc->sptes[RMAP_EXT-1]) {
 588                         desc->more = mmu_alloc_rmap_desc(vcpu);
 589                         desc = desc->more;
 590                 }
 591                 for (i = 0; desc->sptes[i]; ++i)
 592                         ;
 593                 desc->sptes[i] = spte;
 594         }
 595         return count;
 596 }
 597
 598 static void rmap_desc_remove_entry(unsigned long *rmapp,
 599                                    struct kvm_rmap_desc *desc,
 600                                    int i,
 601                                    struct kvm_rmap_desc *prev_desc)
 602 {
 603         int j;
 604
 605         for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
 606                 ;
 607         desc->sptes[i] = desc->sptes[j];
 608         desc->sptes[j] = NULL;
 609         if (j != 0)
 610                 return;
 611         if (!prev_desc && !desc->more)
 612                 *rmapp = (unsigned long)desc->sptes[0];
 613         else
 614                 if (prev_desc)
 615                         prev_desc->more = desc->more;
 616                 else
 617                         *rmapp = (unsigned long)desc->more | 1;
 618         mmu_free_rmap_desc(desc);
 619 }
 620
 621 static void rmap_remove(struct kvm *kvm, u64 *spte)
 622 {
 623         struct kvm_rmap_desc *desc;
 624         struct kvm_rmap_desc *prev_desc;
 625         struct kvm_mmu_page *sp;
 626         pfn_t pfn;
 627         unsigned long *rmapp;
 628         int i;
 629
 630         if (!is_rmap_spte(*spte))
 631                 return;
 632         sp = page_header(__pa(spte));
 633         pfn = spte_to_pfn(*spte);
 634         if (*spte & shadow_accessed_mask)
 635                 kvm_set_pfn_accessed(pfn);
 636         if (is_writeble_pte(*spte))
 637                 kvm_set_pfn_dirty(pfn);
 638         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], sp->role.level);
 639         if (!*rmapp) {
 640                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 641                 BUG();
 642         } else if (!(*rmapp & 1)) {
 643                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 644                 if ((u64 *)*rmapp != spte) {
 645                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 646                                spte, *spte);
 647                         BUG();
 648                 }
 649                 *rmapp = 0;
 650         } else {
 651                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 652                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 653                 prev_desc = NULL;
 654                 while (desc) {
 655                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
 656                                 if (desc->sptes[i] == spte) {
 657                                         rmap_desc_remove_entry(rmapp,
 658                                                                desc, i,
 659                                                                prev_desc);
 660                                         return;
 661                                 }
 662                         prev_desc = desc;
 663                         desc = desc->more;
 664                 }
 665                 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
 666                 BUG();
 667         }
 668 }
 669
 670 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 671 {
 672         struct kvm_rmap_desc *desc;
 673         struct kvm_rmap_desc *prev_desc;
 674         u64 *prev_spte;
 675         int i;
 676
 677         if (!*rmapp)
 678                 return NULL;
 679         else if (!(*rmapp & 1)) {
 680                 if (!spte)
 681                         return (u64 *)*rmapp;
 682                 return NULL;
 683         }
 684         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 685         prev_desc = NULL;
 686         prev_spte = NULL;
 687         while (desc) {
 688                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
 689                         if (prev_spte == spte)
 690                                 return desc->sptes[i];
 691                         prev_spte = desc->sptes[i];
 692                 }
 693                 desc = desc->more;
 694         }
 695         return NULL;
 696 }
 697
 698 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 699 {
 700         unsigned long *rmapp;
 701         u64 *spte;
 702         int i, write_protected = 0;
 703
 704         gfn = unalias_gfn(kvm, gfn);
 705         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
 706
 707         spte = rmap_next(kvm, rmapp, NULL);
 708         while (spte) {
 709                 BUG_ON(!spte);
 710                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 711                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 712                 if (is_writeble_pte(*spte)) {
 713                         __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
 714                         write_protected = 1;
 715                 }
 716                 spte = rmap_next(kvm, rmapp, spte);
 717         }
 718         if (write_protected) {
 719                 pfn_t pfn;
 720
 721                 spte = rmap_next(kvm, rmapp, NULL);
 722                 pfn = spte_to_pfn(*spte);
 723                 kvm_set_pfn_dirty(pfn);
 724         }
 725
 726         /* check for huge page mappings */
 727         for (i = PT_DIRECTORY_LEVEL;
 728              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 729                 rmapp = gfn_to_rmap(kvm, gfn, i);
 730                 spte = rmap_next(kvm, rmapp, NULL);
 731                 while (spte) {
 732                         BUG_ON(!spte);
 733                         BUG_ON(!(*spte & PT_PRESENT_MASK));
 734                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 735                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 736                         if (is_writeble_pte(*spte)) {
 737                                 rmap_remove(kvm, spte);
 738                                 --kvm->stat.lpages;
 739                                 __set_spte(spte, shadow_trap_nonpresent_pte);
 740                                 spte = NULL;
 741                                 write_protected = 1;
 742                         }
 743                         spte = rmap_next(kvm, rmapp, spte);
 744                 }
 745         }
 746
 747         return write_protected;
 748 }
 749
 750 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 751                            unsigned long data)
 752 {
 753         u64 *spte;
 754         int need_tlb_flush = 0;
 755
 756         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 757                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 758                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 759                 rmap_remove(kvm, spte);
 760                 __set_spte(spte, shadow_trap_nonpresent_pte);
 761                 need_tlb_flush = 1;
 762         }
 763         return need_tlb_flush;
 764 }
 765
 766 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
 767                              unsigned long data)
 768 {
 769         int need_flush = 0;
 770         u64 *spte, new_spte;
 771         pte_t *ptep = (pte_t *)data;
 772         pfn_t new_pfn;
 773
 774         WARN_ON(pte_huge(*ptep));
 775         new_pfn = pte_pfn(*ptep);
 776         spte = rmap_next(kvm, rmapp, NULL);
 777         while (spte) {
 778                 BUG_ON(!is_shadow_present_pte(*spte));
 779                 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
 780                 need_flush = 1;
 781                 if (pte_write(*ptep)) {
 782                         rmap_remove(kvm, spte);
 783                         __set_spte(spte, shadow_trap_nonpresent_pte);
 784                         spte = rmap_next(kvm, rmapp, NULL);
 785                 } else {
 786                         new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
 787                         new_spte |= (u64)new_pfn << PAGE_SHIFT;
 788
 789                         new_spte &= ~PT_WRITABLE_MASK;
 790                         new_spte &= ~SPTE_HOST_WRITEABLE;
 791                         if (is_writeble_pte(*spte))
 792                                 kvm_set_pfn_dirty(spte_to_pfn(*spte));
 793                         __set_spte(spte, new_spte);
 794                         spte = rmap_next(kvm, rmapp, spte);
 795                 }
 796         }
 797         if (need_flush)
 798                 kvm_flush_remote_tlbs(kvm);
 799
 800         return 0;
 801 }
 802
 803 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 804                           unsigned long data,
 805                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 806                                          unsigned long data))
 807 {
 808         int i, j;
 809         int retval = 0;
 810         struct kvm_memslots *slots = kvm->memslots;
 811
 812         /*
 813          * If mmap_sem isn't taken, we can look the memslots with only
 814          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 815          */
 816         for (i = 0; i < slots->nmemslots; i++) {
 817                 struct kvm_memory_slot *memslot = &slots->memslots[i];
 818                 unsigned long start = memslot->userspace_addr;
 819                 unsigned long end;
 820
 821                 /* mmu_lock protects userspace_addr */
 822                 if (!start)
 823                         continue;
 824
 825                 end = start + (memslot->npages << PAGE_SHIFT);
 826                 if (hva >= start && hva < end) {
 827                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 828
 829                         retval |= handler(kvm, &memslot->rmap[gfn_offset],
 830                                           data);
 831
 832                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
 833                                 int idx = gfn_offset;
 834                                 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
 835                                 retval |= handler(kvm,
 836                                         &memslot->lpage_info[j][idx].rmap_pde,
 837                                         data);
 838                         }
 839                 }
 840         }
 841
 842         return retval;
 843 }
 844
 845 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 846 {
 847         return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
 848 }
 849
 850 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
 851 {
 852         kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
 853 }
 854
 855 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 856                          unsigned long data)
 857 {
 858         u64 *spte;
 859         int young = 0;
 860
 861         /* always return old for EPT */
 862         if (!shadow_accessed_mask)
 863                 return 0;
 864
 865         spte = rmap_next(kvm, rmapp, NULL);
 866         while (spte) {
 867                 int _young;
 868                 u64 _spte = *spte;
 869                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 870                 _young = _spte & PT_ACCESSED_MASK;
 871                 if (_young) {
 872                         young = 1;
 873                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 874                 }
 875                 spte = rmap_next(kvm, rmapp, spte);
 876         }
 877         return young;
 878 }
 879
 880 #define RMAP_RECYCLE_THRESHOLD 1000
 881
 882 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 883 {
 884         unsigned long *rmapp;
 885         struct kvm_mmu_page *sp;
 886
 887         sp = page_header(__pa(spte));
 888
 889         gfn = unalias_gfn(vcpu->kvm, gfn);
 890         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 891
 892         kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
 893         kvm_flush_remote_tlbs(vcpu->kvm);
 894 }
 895
 896 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 897 {
 898         return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
 899 }
 900
 901 #ifdef MMU_DEBUG
 902 static int is_empty_shadow_page(u64 *spt)
 903 {
 904         u64 *pos;
 905         u64 *end;
 906
 907         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 908                 if (is_shadow_present_pte(*pos)) {
 909                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 910                                pos, *pos);
 911                         return 0;
 912                 }
 913         return 1;
 914 }
 915 #endif
 916
 917 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 918 {
 919         ASSERT(is_empty_shadow_page(sp->spt));
 920         list_del(&sp->link);
 921         __free_page(virt_to_page(sp->spt));
 922         __free_page(virt_to_page(sp->gfns));
 923         kfree(sp);
 924         ++kvm->arch.n_free_mmu_pages;
 925 }
 926
 927 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 928 {
 929         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 930 }
 931
 932 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 933                                                u64 *parent_pte)
 934 {
 935         struct kvm_mmu_page *sp;
 936
 937         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 938         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 939         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 940         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 941         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 942         INIT_LIST_HEAD(&sp->oos_link);
 943         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 944         sp->multimapped = 0;
 945         sp->parent_pte = parent_pte;
 946         --vcpu->kvm->arch.n_free_mmu_pages;
 947         return sp;
 948 }
 949
 950 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 951                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 952 {
 953         struct kvm_pte_chain *pte_chain;
 954         struct hlist_node *node;
 955         int i;
 956
 957         if (!parent_pte)
 958                 return;
 959         if (!sp->multimapped) {
 960                 u64 *old = sp->parent_pte;
 961
 962                 if (!old) {
 963                         sp->parent_pte = parent_pte;
 964                         return;
 965                 }
 966                 sp->multimapped = 1;
 967                 pte_chain = mmu_alloc_pte_chain(vcpu);
 968                 INIT_HLIST_HEAD(&sp->parent_ptes);
 969                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 970                 pte_chain->parent_ptes[0] = old;
 971         }
 972         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 973                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 974                         continue;
 975                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 976                         if (!pte_chain->parent_ptes[i]) {
 977                                 pte_chain->parent_ptes[i] = parent_pte;
 978                                 return;
 979                         }
 980         }
 981         pte_chain = mmu_alloc_pte_chain(vcpu);
 982         BUG_ON(!pte_chain);
 983         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 984         pte_chain->parent_ptes[0] = parent_pte;
 985 }
 986
 987 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 988                                        u64 *parent_pte)
 989 {
 990         struct kvm_pte_chain *pte_chain;
 991         struct hlist_node *node;
 992         int i;
 993
 994         if (!sp->multimapped) {
 995                 BUG_ON(sp->parent_pte != parent_pte);
 996                 sp->parent_pte = NULL;
 997                 return;
 998         }
 999         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1000                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1001                         if (!pte_chain->parent_ptes[i])
1002                                 break;
1003                         if (pte_chain->parent_ptes[i] != parent_pte)
1004                                 continue;
1005                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
1006                                 && pte_chain->parent_ptes[i + 1]) {
1007                                 pte_chain->parent_ptes[i]
1008                                         = pte_chain->parent_ptes[i + 1];
1009                                 ++i;
1010                         }
1011                         pte_chain->parent_ptes[i] = NULL;
1012                         if (i == 0) {
1013                                 hlist_del(&pte_chain->link);
1014                                 mmu_free_pte_chain(pte_chain);
1015                                 if (hlist_empty(&sp->parent_ptes)) {
1016                                         sp->multimapped = 0;
1017                                         sp->parent_pte = NULL;
1018                                 }
1019                         }
1020                         return;
1021                 }
1022         BUG();
1023 }
1024
1025
1026 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1027                             mmu_parent_walk_fn fn)
1028 {
1029         struct kvm_pte_chain *pte_chain;
1030         struct hlist_node *node;
1031         struct kvm_mmu_page *parent_sp;
1032         int i;
1033
1034         if (!sp->multimapped && sp->parent_pte) {
1035                 parent_sp = page_header(__pa(sp->parent_pte));
1036                 fn(vcpu, parent_sp);
1037                 mmu_parent_walk(vcpu, parent_sp, fn);
1038                 return;
1039         }
1040         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1041                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1042                         if (!pte_chain->parent_ptes[i])
1043                                 break;
1044                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
1045                         fn(vcpu, parent_sp);
1046                         mmu_parent_walk(vcpu, parent_sp, fn);
1047                 }
1048 }
1049
1050 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1051 {
1052         unsigned int index;
1053         struct kvm_mmu_page *sp = page_header(__pa(spte));
1054
1055         index = spte - sp->spt;
1056         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1057                 sp->unsync_children++;
1058         WARN_ON(!sp->unsync_children);
1059 }
1060
1061 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1062 {
1063         struct kvm_pte_chain *pte_chain;
1064         struct hlist_node *node;
1065         int i;
1066
1067         if (!sp->parent_pte)
1068                 return;
1069
1070         if (!sp->multimapped) {
1071                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1072                 return;
1073         }
1074
1075         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1076                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1077                         if (!pte_chain->parent_ptes[i])
1078                                 break;
1079                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1080                 }
1081 }
1082
1083 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1084 {
1085         kvm_mmu_update_parents_unsync(sp);
1086         return 1;
1087 }
1088
1089 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
1090                                         struct kvm_mmu_page *sp)
1091 {
1092         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
1093         kvm_mmu_update_parents_unsync(sp);
1094 }
1095
1096 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1097                                     struct kvm_mmu_page *sp)
1098 {
1099         int i;
1100
1101         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1102                 sp->spt[i] = shadow_trap_nonpresent_pte;
1103 }
1104
1105 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1106                                struct kvm_mmu_page *sp)
1107 {
1108         return 1;
1109 }
1110
1111 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1112 {
1113 }
1114
1115 #define KVM_PAGE_ARRAY_NR 16
1116
1117 struct kvm_mmu_pages {
1118         struct mmu_page_and_offset {
1119                 struct kvm_mmu_page *sp;
1120                 unsigned int idx;
1121         } page[KVM_PAGE_ARRAY_NR];
1122         unsigned int nr;
1123 };
1124
1125 #define for_each_unsync_children(bitmap, idx)           \
1126         for (idx = find_first_bit(bitmap, 512);         \
1127              idx < 512;                                 \
1128              idx = find_next_bit(bitmap, 512, idx+1))
1129
1130 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1131                          int idx)
1132 {
1133         int i;
1134
1135         if (sp->unsync)
1136                 for (i=0; i < pvec->nr; i++)
1137                         if (pvec->page[i].sp == sp)
1138                                 return 0;
1139
1140         pvec->page[pvec->nr].sp = sp;
1141         pvec->page[pvec->nr].idx = idx;
1142         pvec->nr++;
1143         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1144 }
1145
1146 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1147                            struct kvm_mmu_pages *pvec)
1148 {
1149         int i, ret, nr_unsync_leaf = 0;
1150
1151         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1152                 u64 ent = sp->spt[i];
1153
1154                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1155                         struct kvm_mmu_page *child;
1156                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1157
1158                         if (child->unsync_children) {
1159                                 if (mmu_pages_add(pvec, child, i))
1160                                         return -ENOSPC;
1161
1162                                 ret = __mmu_unsync_walk(child, pvec);
1163                                 if (!ret)
1164                                         __clear_bit(i, sp->unsync_child_bitmap);
1165                                 else if (ret > 0)
1166                                         nr_unsync_leaf += ret;
1167                                 else
1168                                         return ret;
1169                         }
1170
1171                         if (child->unsync) {
1172                                 nr_unsync_leaf++;
1173                                 if (mmu_pages_add(pvec, child, i))
1174                                         return -ENOSPC;
1175                         }
1176                 }
1177         }
1178
1179         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1180                 sp->unsync_children = 0;
1181
1182         return nr_unsync_leaf;
1183 }
1184
1185 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1186                            struct kvm_mmu_pages *pvec)
1187 {
1188         if (!sp->unsync_children)
1189                 return 0;
1190
1191         mmu_pages_add(pvec, sp, 0);
1192         return __mmu_unsync_walk(sp, pvec);
1193 }
1194
1195 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1196 {
1197         unsigned index;
1198         struct hlist_head *bucket;
1199         struct kvm_mmu_page *sp;
1200         struct hlist_node *node;
1201
1202         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1203         index = kvm_page_table_hashfn(gfn);
1204         bucket = &kvm->arch.mmu_page_hash[index];
1205         hlist_for_each_entry(sp, node, bucket, hash_link)
1206                 if (sp->gfn == gfn && !sp->role.direct
1207                     && !sp->role.invalid) {
1208                         pgprintk("%s: found role %x\n",
1209                                  __func__, sp->role.word);
1210                         return sp;
1211                 }
1212         return NULL;
1213 }
1214
1215 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1216 {
1217         WARN_ON(!sp->unsync);
1218         sp->unsync = 0;
1219         --kvm->stat.mmu_unsync;
1220 }
1221
1222 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1223
1224 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1225 {
1226         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1227                 kvm_mmu_zap_page(vcpu->kvm, sp);
1228                 return 1;
1229         }
1230
1231         trace_kvm_mmu_sync_page(sp);
1232         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1233                 kvm_flush_remote_tlbs(vcpu->kvm);
1234         kvm_unlink_unsync_page(vcpu->kvm, sp);
1235         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1236                 kvm_mmu_zap_page(vcpu->kvm, sp);
1237                 return 1;
1238         }
1239
1240         kvm_mmu_flush_tlb(vcpu);
1241         return 0;
1242 }
1243
1244 struct mmu_page_path {
1245         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1246         unsigned int idx[PT64_ROOT_LEVEL-1];
1247 };
1248
1249 #define for_each_sp(pvec, sp, parents, i)                       \
1250                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1251                         sp = pvec.page[i].sp;                   \
1252                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1253                         i = mmu_pages_next(&pvec, &parents, i))
1254
1255 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1256                           struct mmu_page_path *parents,
1257                           int i)
1258 {
1259         int n;
1260
1261         for (n = i+1; n < pvec->nr; n++) {
1262                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1263
1264                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1265                         parents->idx[0] = pvec->page[n].idx;
1266                         return n;
1267                 }
1268
1269                 parents->parent[sp->role.level-2] = sp;
1270                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1271         }
1272
1273         return n;
1274 }
1275
1276 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1277 {
1278         struct kvm_mmu_page *sp;
1279         unsigned int level = 0;
1280
1281         do {
1282                 unsigned int idx = parents->idx[level];
1283
1284                 sp = parents->parent[level];
1285                 if (!sp)
1286                         return;
1287
1288                 --sp->unsync_children;
1289                 WARN_ON((int)sp->unsync_children < 0);
1290                 __clear_bit(idx, sp->unsync_child_bitmap);
1291                 level++;
1292         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1293 }
1294
1295 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1296                                struct mmu_page_path *parents,
1297                                struct kvm_mmu_pages *pvec)
1298 {
1299         parents->parent[parent->role.level-1] = NULL;
1300         pvec->nr = 0;
1301 }
1302
1303 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1304                               struct kvm_mmu_page *parent)
1305 {
1306         int i;
1307         struct kvm_mmu_page *sp;
1308         struct mmu_page_path parents;
1309         struct kvm_mmu_pages pages;
1310
1311         kvm_mmu_pages_init(parent, &parents, &pages);
1312         while (mmu_unsync_walk(parent, &pages)) {
1313                 int protected = 0;
1314
1315                 for_each_sp(pages, sp, parents, i)
1316                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1317
1318                 if (protected)
1319                         kvm_flush_remote_tlbs(vcpu->kvm);
1320
1321                 for_each_sp(pages, sp, parents, i) {
1322                         kvm_sync_page(vcpu, sp);
1323                         mmu_pages_clear_parents(&parents);
1324                 }
1325                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1326                 kvm_mmu_pages_init(parent, &parents, &pages);
1327         }
1328 }
1329
1330 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1331                                              gfn_t gfn,
1332                                              gva_t gaddr,
1333                                              unsigned level,
1334                                              int direct,
1335                                              unsigned access,
1336                                              u64 *parent_pte)
1337 {
1338         union kvm_mmu_page_role role;
1339         unsigned index;
1340         unsigned quadrant;
1341         struct hlist_head *bucket;
1342         struct kvm_mmu_page *sp;
1343         struct hlist_node *node, *tmp;
1344
1345         role = vcpu->arch.mmu.base_role;
1346         role.level = level;
1347         role.direct = direct;
1348         role.access = access;
1349         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1350                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1351                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1352                 role.quadrant = quadrant;
1353         }
1354         index = kvm_page_table_hashfn(gfn);
1355         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1356         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1357                 if (sp->gfn == gfn) {
1358                         if (sp->unsync)
1359                                 if (kvm_sync_page(vcpu, sp))
1360                                         continue;
1361
1362                         if (sp->role.word != role.word)
1363                                 continue;
1364
1365                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1366                         if (sp->unsync_children) {
1367                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1368                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1369                         }
1370                         trace_kvm_mmu_get_page(sp, false);
1371                         return sp;
1372                 }
1373         ++vcpu->kvm->stat.mmu_cache_miss;
1374         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1375         if (!sp)
1376                 return sp;
1377         sp->gfn = gfn;
1378         sp->role = role;
1379         hlist_add_head(&sp->hash_link, bucket);
1380         if (!direct) {
1381                 if (rmap_write_protect(vcpu->kvm, gfn))
1382                         kvm_flush_remote_tlbs(vcpu->kvm);
1383                 account_shadowed(vcpu->kvm, gfn);
1384         }
1385         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1386                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1387         else
1388                 nonpaging_prefetch_page(vcpu, sp);
1389         trace_kvm_mmu_get_page(sp, true);
1390         return sp;
1391 }
1392
1393 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1394                              struct kvm_vcpu *vcpu, u64 addr)
1395 {
1396         iterator->addr = addr;
1397         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1398         iterator->level = vcpu->arch.mmu.shadow_root_level;
1399         if (iterator->level == PT32E_ROOT_LEVEL) {
1400                 iterator->shadow_addr
1401                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1402                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1403                 --iterator->level;
1404                 if (!iterator->shadow_addr)
1405                         iterator->level = 0;
1406         }
1407 }
1408
1409 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1410 {
1411         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1412                 return false;
1413
1414         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1415                 if (is_large_pte(*iterator->sptep))
1416                         return false;
1417
1418         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1419         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1420         return true;
1421 }
1422
1423 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1424 {
1425         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1426         --iterator->level;
1427 }
1428
1429 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1430                                          struct kvm_mmu_page *sp)
1431 {
1432         unsigned i;
1433         u64 *pt;
1434         u64 ent;
1435
1436         pt = sp->spt;
1437
1438         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1439                 ent = pt[i];
1440
1441                 if (is_shadow_present_pte(ent)) {
1442                         if (!is_last_spte(ent, sp->role.level)) {
1443                                 ent &= PT64_BASE_ADDR_MASK;
1444                                 mmu_page_remove_parent_pte(page_header(ent),
1445                                                            &pt[i]);
1446                         } else {
1447                                 if (is_large_pte(ent))
1448                                         --kvm->stat.lpages;
1449                                 rmap_remove(kvm, &pt[i]);
1450                         }
1451                 }
1452                 pt[i] = shadow_trap_nonpresent_pte;
1453         }
1454 }
1455
1456 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1457 {
1458         mmu_page_remove_parent_pte(sp, parent_pte);
1459 }
1460
1461 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1462 {
1463         int i;
1464         struct kvm_vcpu *vcpu;
1465
1466         kvm_for_each_vcpu(i, vcpu, kvm)
1467                 vcpu->arch.last_pte_updated = NULL;
1468 }
1469
1470 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1471 {
1472         u64 *parent_pte;
1473
1474         while (sp->multimapped || sp->parent_pte) {
1475                 if (!sp->multimapped)
1476                         parent_pte = sp->parent_pte;
1477                 else {
1478                         struct kvm_pte_chain *chain;
1479
1480                         chain = container_of(sp->parent_ptes.first,
1481                                              struct kvm_pte_chain, link);
1482                         parent_pte = chain->parent_ptes[0];
1483                 }
1484                 BUG_ON(!parent_pte);
1485                 kvm_mmu_put_page(sp, parent_pte);
1486                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1487         }
1488 }
1489
1490 static int mmu_zap_unsync_children(struct kvm *kvm,
1491                                    struct kvm_mmu_page *parent)
1492 {
1493         int i, zapped = 0;
1494         struct mmu_page_path parents;
1495         struct kvm_mmu_pages pages;
1496
1497         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1498                 return 0;
1499
1500         kvm_mmu_pages_init(parent, &parents, &pages);
1501         while (mmu_unsync_walk(parent, &pages)) {
1502                 struct kvm_mmu_page *sp;
1503
1504                 for_each_sp(pages, sp, parents, i) {
1505                         kvm_mmu_zap_page(kvm, sp);
1506                         mmu_pages_clear_parents(&parents);
1507                 }
1508                 zapped += pages.nr;
1509                 kvm_mmu_pages_init(parent, &parents, &pages);
1510         }
1511
1512         return zapped;
1513 }
1514
1515 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1516 {
1517         int ret;
1518
1519         trace_kvm_mmu_zap_page(sp);
1520         ++kvm->stat.mmu_shadow_zapped;
1521         ret = mmu_zap_unsync_children(kvm, sp);
1522         kvm_mmu_page_unlink_children(kvm, sp);
1523         kvm_mmu_unlink_parents(kvm, sp);
1524         kvm_flush_remote_tlbs(kvm);
1525         if (!sp->role.invalid && !sp->role.direct)
1526                 unaccount_shadowed(kvm, sp->gfn);
1527         if (sp->unsync)
1528                 kvm_unlink_unsync_page(kvm, sp);
1529         if (!sp->root_count) {
1530                 hlist_del(&sp->hash_link);
1531                 kvm_mmu_free_page(kvm, sp);
1532         } else {
1533                 sp->role.invalid = 1;
1534                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1535                 kvm_reload_remote_mmus(kvm);
1536         }
1537         kvm_mmu_reset_last_pte_updated(kvm);
1538         return ret;
1539 }
1540
1541 /*
1542  * Changing the number of mmu pages allocated to the vm
1543  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1544  */
1545 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1546 {
1547         int used_pages;
1548
1549         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1550         used_pages = max(0, used_pages);
1551
1552         /*
1553          * If we set the number of mmu pages to be smaller be than the
1554          * number of actived pages , we must to free some mmu pages before we
1555          * change the value
1556          */
1557
1558         if (used_pages > kvm_nr_mmu_pages) {
1559                 while (used_pages > kvm_nr_mmu_pages) {
1560                         struct kvm_mmu_page *page;
1561
1562                         page = container_of(kvm->arch.active_mmu_pages.prev,
1563                                             struct kvm_mmu_page, link);
1564                         kvm_mmu_zap_page(kvm, page);
1565                         used_pages--;
1566                 }
1567                 kvm->arch.n_free_mmu_pages = 0;
1568         }
1569         else
1570                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1571                                          - kvm->arch.n_alloc_mmu_pages;
1572
1573         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1574 }
1575
1576 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1577 {
1578         unsigned index;
1579         struct hlist_head *bucket;
1580         struct kvm_mmu_page *sp;
1581         struct hlist_node *node, *n;
1582         int r;
1583
1584         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1585         r = 0;
1586         index = kvm_page_table_hashfn(gfn);
1587         bucket = &kvm->arch.mmu_page_hash[index];
1588         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1589                 if (sp->gfn == gfn && !sp->role.direct) {
1590                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1591                                  sp->role.word);
1592                         r = 1;
1593                         if (kvm_mmu_zap_page(kvm, sp))
1594                                 n = bucket->first;
1595                 }
1596         return r;
1597 }
1598
1599 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1600 {
1601         unsigned index;
1602         struct hlist_head *bucket;
1603         struct kvm_mmu_page *sp;
1604         struct hlist_node *node, *nn;
1605
1606         index = kvm_page_table_hashfn(gfn);
1607         bucket = &kvm->arch.mmu_page_hash[index];
1608         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1609                 if (sp->gfn == gfn && !sp->role.direct
1610                     && !sp->role.invalid) {
1611                         pgprintk("%s: zap %lx %x\n",
1612                                  __func__, gfn, sp->role.word);
1613                         kvm_mmu_zap_page(kvm, sp);
1614                 }
1615         }
1616 }
1617
1618 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1619 {
1620         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1621         struct kvm_mmu_page *sp = page_header(__pa(pte));
1622
1623         __set_bit(slot, sp->slot_bitmap);
1624 }
1625
1626 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1627 {
1628         int i;
1629         u64 *pt = sp->spt;
1630
1631         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1632                 return;
1633
1634         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1635                 if (pt[i] == shadow_notrap_nonpresent_pte)
1636                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1637         }
1638 }
1639
1640 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1641 {
1642         struct page *page;
1643
1644         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1645
1646         if (gpa == UNMAPPED_GVA)
1647                 return NULL;
1648
1649         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1650
1651         return page;
1652 }
1653
1654 /*
1655  * The function is based on mtrr_type_lookup() in
1656  * arch/x86/kernel/cpu/mtrr/generic.c
1657  */
1658 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1659                          u64 start, u64 end)
1660 {
1661         int i;
1662         u64 base, mask;
1663         u8 prev_match, curr_match;
1664         int num_var_ranges = KVM_NR_VAR_MTRR;
1665
1666         if (!mtrr_state->enabled)
1667                 return 0xFF;
1668
1669         /* Make end inclusive end, instead of exclusive */
1670         end--;
1671
1672         /* Look in fixed ranges. Just return the type as per start */
1673         if (mtrr_state->have_fixed && (start < 0x100000)) {
1674                 int idx;
1675
1676                 if (start < 0x80000) {
1677                         idx = 0;
1678                         idx += (start >> 16);
1679                         return mtrr_state->fixed_ranges[idx];
1680                 } else if (start < 0xC0000) {
1681                         idx = 1 * 8;
1682                         idx += ((start - 0x80000) >> 14);
1683                         return mtrr_state->fixed_ranges[idx];
1684                 } else if (start < 0x1000000) {
1685                         idx = 3 * 8;
1686                         idx += ((start - 0xC0000) >> 12);
1687                         return mtrr_state->fixed_ranges[idx];
1688                 }
1689         }
1690
1691         /*
1692          * Look in variable ranges
1693          * Look of multiple ranges matching this address and pick type
1694          * as per MTRR precedence
1695          */
1696         if (!(mtrr_state->enabled & 2))
1697                 return mtrr_state->def_type;
1698
1699         prev_match = 0xFF;
1700         for (i = 0; i < num_var_ranges; ++i) {
1701                 unsigned short start_state, end_state;
1702
1703                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1704                         continue;
1705
1706                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1707                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1708                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1709                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1710
1711                 start_state = ((start & mask) == (base & mask));
1712                 end_state = ((end & mask) == (base & mask));
1713                 if (start_state != end_state)
1714                         return 0xFE;
1715
1716                 if ((start & mask) != (base & mask))
1717                         continue;
1718
1719                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1720                 if (prev_match == 0xFF) {
1721                         prev_match = curr_match;
1722                         continue;
1723                 }
1724
1725                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1726                     curr_match == MTRR_TYPE_UNCACHABLE)
1727                         return MTRR_TYPE_UNCACHABLE;
1728
1729                 if ((prev_match == MTRR_TYPE_WRBACK &&
1730                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1731                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1732                      curr_match == MTRR_TYPE_WRBACK)) {
1733                         prev_match = MTRR_TYPE_WRTHROUGH;
1734                         curr_match = MTRR_TYPE_WRTHROUGH;
1735                 }
1736
1737                 if (prev_match != curr_match)
1738                         return MTRR_TYPE_UNCACHABLE;
1739         }
1740
1741         if (prev_match != 0xFF)
1742                 return prev_match;
1743
1744         return mtrr_state->def_type;
1745 }
1746
1747 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1748 {
1749         u8 mtrr;
1750
1751         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1752                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1753         if (mtrr == 0xfe || mtrr == 0xff)
1754                 mtrr = MTRR_TYPE_WRBACK;
1755         return mtrr;
1756 }
1757 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1758
1759 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1760 {
1761         unsigned index;
1762         struct hlist_head *bucket;
1763         struct kvm_mmu_page *s;
1764         struct hlist_node *node, *n;
1765
1766         trace_kvm_mmu_unsync_page(sp);
1767         index = kvm_page_table_hashfn(sp->gfn);
1768         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1769         /* don't unsync if pagetable is shadowed with multiple roles */
1770         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1771                 if (s->gfn != sp->gfn || s->role.direct)
1772                         continue;
1773                 if (s->role.word != sp->role.word)
1774                         return 1;
1775         }
1776         ++vcpu->kvm->stat.mmu_unsync;
1777         sp->unsync = 1;
1778
1779         kvm_mmu_mark_parents_unsync(vcpu, sp);
1780
1781         mmu_convert_notrap(sp);
1782         return 0;
1783 }
1784
1785 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1786                                   bool can_unsync)
1787 {
1788         struct kvm_mmu_page *shadow;
1789
1790         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1791         if (shadow) {
1792                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1793                         return 1;
1794                 if (shadow->unsync)
1795                         return 0;
1796                 if (can_unsync && oos_shadow)
1797                         return kvm_unsync_page(vcpu, shadow);
1798                 return 1;
1799         }
1800         return 0;
1801 }
1802
1803 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1804                     unsigned pte_access, int user_fault,
1805                     int write_fault, int dirty, int level,
1806                     gfn_t gfn, pfn_t pfn, bool speculative,
1807                     bool can_unsync, bool reset_host_protection)
1808 {
1809         u64 spte;
1810         int ret = 0;
1811
1812         /*
1813          * We don't set the accessed bit, since we sometimes want to see
1814          * whether the guest actually used the pte (in order to detect
1815          * demand paging).
1816          */
1817         spte = shadow_base_present_pte | shadow_dirty_mask;
1818         if (!speculative)
1819                 spte |= shadow_accessed_mask;
1820         if (!dirty)
1821                 pte_access &= ~ACC_WRITE_MASK;
1822         if (pte_access & ACC_EXEC_MASK)
1823                 spte |= shadow_x_mask;
1824         else
1825                 spte |= shadow_nx_mask;
1826         if (pte_access & ACC_USER_MASK)
1827                 spte |= shadow_user_mask;
1828         if (level > PT_PAGE_TABLE_LEVEL)
1829                 spte |= PT_PAGE_SIZE_MASK;
1830         if (tdp_enabled)
1831                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1832                         kvm_is_mmio_pfn(pfn));
1833
1834         if (reset_host_protection)
1835                 spte |= SPTE_HOST_WRITEABLE;
1836
1837         spte |= (u64)pfn << PAGE_SHIFT;
1838
1839         if ((pte_access & ACC_WRITE_MASK)
1840             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1841
1842                 if (level > PT_PAGE_TABLE_LEVEL &&
1843                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1844                         ret = 1;
1845                         spte = shadow_trap_nonpresent_pte;
1846                         goto set_pte;
1847                 }
1848
1849                 spte |= PT_WRITABLE_MASK;
1850
1851                 /*
1852                  * Optimization: for pte sync, if spte was writable the hash
1853                  * lookup is unnecessary (and expensive). Write protection
1854                  * is responsibility of mmu_get_page / kvm_sync_page.
1855                  * Same reasoning can be applied to dirty page accounting.
1856                  */
1857                 if (!can_unsync && is_writeble_pte(*sptep))
1858                         goto set_pte;
1859
1860                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1861                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1862                                  __func__, gfn);
1863                         ret = 1;
1864                         pte_access &= ~ACC_WRITE_MASK;
1865                         if (is_writeble_pte(spte))
1866                                 spte &= ~PT_WRITABLE_MASK;
1867                 }
1868         }
1869
1870         if (pte_access & ACC_WRITE_MASK)
1871                 mark_page_dirty(vcpu->kvm, gfn);
1872
1873 set_pte:
1874         __set_spte(sptep, spte);
1875         return ret;
1876 }
1877
1878 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1879                          unsigned pt_access, unsigned pte_access,
1880                          int user_fault, int write_fault, int dirty,
1881                          int *ptwrite, int level, gfn_t gfn,
1882                          pfn_t pfn, bool speculative,
1883                          bool reset_host_protection)
1884 {
1885         int was_rmapped = 0;
1886         int was_writeble = is_writeble_pte(*sptep);
1887         int rmap_count;
1888
1889         pgprintk("%s: spte %llx access %x write_fault %d"
1890                  " user_fault %d gfn %lx\n",
1891                  __func__, *sptep, pt_access,
1892                  write_fault, user_fault, gfn);
1893
1894         if (is_rmap_spte(*sptep)) {
1895                 /*
1896                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1897                  * the parent of the now unreachable PTE.
1898                  */
1899                 if (level > PT_PAGE_TABLE_LEVEL &&
1900                     !is_large_pte(*sptep)) {
1901                         struct kvm_mmu_page *child;
1902                         u64 pte = *sptep;
1903
1904                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1905                         mmu_page_remove_parent_pte(child, sptep);
1906                 } else if (pfn != spte_to_pfn(*sptep)) {
1907                         pgprintk("hfn old %lx new %lx\n",
1908                                  spte_to_pfn(*sptep), pfn);
1909                         rmap_remove(vcpu->kvm, sptep);
1910                 } else
1911                         was_rmapped = 1;
1912         }
1913
1914         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1915                       dirty, level, gfn, pfn, speculative, true,
1916                       reset_host_protection)) {
1917                 if (write_fault)
1918                         *ptwrite = 1;
1919                 kvm_x86_ops->tlb_flush(vcpu);
1920         }
1921
1922         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1923         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1924                  is_large_pte(*sptep)? "2MB" : "4kB",
1925                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1926                  *sptep, sptep);
1927         if (!was_rmapped && is_large_pte(*sptep))
1928                 ++vcpu->kvm->stat.lpages;
1929
1930         page_header_update_slot(vcpu->kvm, sptep, gfn);
1931         if (!was_rmapped) {
1932                 rmap_count = rmap_add(vcpu, sptep, gfn);
1933                 kvm_release_pfn_clean(pfn);
1934                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1935                         rmap_recycle(vcpu, sptep, gfn);
1936         } else {
1937                 if (was_writeble)
1938                         kvm_release_pfn_dirty(pfn);
1939                 else
1940                         kvm_release_pfn_clean(pfn);
1941         }
1942         if (speculative) {
1943                 vcpu->arch.last_pte_updated = sptep;
1944                 vcpu->arch.last_pte_gfn = gfn;
1945         }
1946 }
1947
1948 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1949 {
1950 }
1951
1952 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1953                         int level, gfn_t gfn, pfn_t pfn)
1954 {
1955         struct kvm_shadow_walk_iterator iterator;
1956         struct kvm_mmu_page *sp;
1957         int pt_write = 0;
1958         gfn_t pseudo_gfn;
1959
1960         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1961                 if (iterator.level == level) {
1962                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1963                                      0, write, 1, &pt_write,
1964                                      level, gfn, pfn, false, true);
1965                         ++vcpu->stat.pf_fixed;
1966                         break;
1967                 }
1968
1969                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1970                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1971                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1972                                               iterator.level - 1,
1973                                               1, ACC_ALL, iterator.sptep);
1974                         if (!sp) {
1975                                 pgprintk("nonpaging_map: ENOMEM\n");
1976                                 kvm_release_pfn_clean(pfn);
1977                                 return -ENOMEM;
1978                         }
1979
1980                         __set_spte(iterator.sptep,
1981                                    __pa(sp->spt)
1982                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1983                                    | shadow_user_mask | shadow_x_mask);
1984                 }
1985         }
1986         return pt_write;
1987 }
1988
1989 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1990 {
1991         int r;
1992         int level;
1993         pfn_t pfn;
1994         unsigned long mmu_seq;
1995
1996         level = mapping_level(vcpu, gfn);
1997
1998         /*
1999          * This path builds a PAE pagetable - so we can map 2mb pages at
2000          * maximum. Therefore check if the level is larger than that.
2001          */
2002         if (level > PT_DIRECTORY_LEVEL)
2003                 level = PT_DIRECTORY_LEVEL;
2004
2005         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2006
2007         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2008         smp_rmb();
2009         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2010
2011         /* mmio */
2012         if (is_error_pfn(pfn)) {
2013                 kvm_release_pfn_clean(pfn);
2014                 return 1;
2015         }
2016
2017         spin_lock(&vcpu->kvm->mmu_lock);
2018         if (mmu_notifier_retry(vcpu, mmu_seq))
2019                 goto out_unlock;
2020         kvm_mmu_free_some_pages(vcpu);
2021         r = __direct_map(vcpu, v, write, level, gfn, pfn);
2022         spin_unlock(&vcpu->kvm->mmu_lock);
2023
2024
2025         return r;
2026
2027 out_unlock:
2028         spin_unlock(&vcpu->kvm->mmu_lock);
2029         kvm_release_pfn_clean(pfn);
2030         return 0;
2031 }
2032
2033
2034 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2035 {
2036         int i;
2037         struct kvm_mmu_page *sp;
2038
2039         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2040                 return;
2041         spin_lock(&vcpu->kvm->mmu_lock);
2042         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2043                 hpa_t root = vcpu->arch.mmu.root_hpa;
2044
2045                 sp = page_header(root);
2046                 --sp->root_count;
2047                 if (!sp->root_count && sp->role.invalid)
2048                         kvm_mmu_zap_page(vcpu->kvm, sp);
2049                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2050                 spin_unlock(&vcpu->kvm->mmu_lock);
2051                 return;
2052         }
2053         for (i = 0; i < 4; ++i) {
2054                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2055
2056                 if (root) {
2057                         root &= PT64_BASE_ADDR_MASK;
2058                         sp = page_header(root);
2059                         --sp->root_count;
2060                         if (!sp->root_count && sp->role.invalid)
2061                                 kvm_mmu_zap_page(vcpu->kvm, sp);
2062                 }
2063                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2064         }
2065         spin_unlock(&vcpu->kvm->mmu_lock);
2066         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2067 }
2068
2069 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2070 {
2071         int ret = 0;
2072
2073         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2074                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2075                 ret = 1;
2076         }
2077
2078         return ret;
2079 }
2080
2081 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2082 {
2083         int i;
2084         gfn_t root_gfn;
2085         struct kvm_mmu_page *sp;
2086         int direct = 0;
2087         u64 pdptr;
2088
2089         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2090
2091         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2092                 hpa_t root = vcpu->arch.mmu.root_hpa;
2093
2094                 ASSERT(!VALID_PAGE(root));
2095                 if (tdp_enabled)
2096                         direct = 1;
2097                 if (mmu_check_root(vcpu, root_gfn))
2098                         return 1;
2099                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2100                                       PT64_ROOT_LEVEL, direct,
2101                                       ACC_ALL, NULL);
2102                 root = __pa(sp->spt);
2103                 ++sp->root_count;
2104                 vcpu->arch.mmu.root_hpa = root;
2105                 return 0;
2106         }
2107         direct = !is_paging(vcpu);
2108         if (tdp_enabled)
2109                 direct = 1;
2110         for (i = 0; i < 4; ++i) {
2111                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2112
2113                 ASSERT(!VALID_PAGE(root));
2114                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2115                         pdptr = kvm_pdptr_read(vcpu, i);
2116                         if (!is_present_gpte(pdptr)) {
2117                                 vcpu->arch.mmu.pae_root[i] = 0;
2118                                 continue;
2119                         }
2120                         root_gfn = pdptr >> PAGE_SHIFT;
2121                 } else if (vcpu->arch.mmu.root_level == 0)
2122                         root_gfn = 0;
2123                 if (mmu_check_root(vcpu, root_gfn))
2124                         return 1;
2125                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2126                                       PT32_ROOT_LEVEL, direct,
2127                                       ACC_ALL, NULL);
2128                 root = __pa(sp->spt);
2129                 ++sp->root_count;
2130                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2131         }
2132         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2133         return 0;
2134 }
2135
2136 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2137 {
2138         int i;
2139         struct kvm_mmu_page *sp;
2140
2141         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2142                 return;
2143         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2144                 hpa_t root = vcpu->arch.mmu.root_hpa;
2145                 sp = page_header(root);
2146                 mmu_sync_children(vcpu, sp);
2147                 return;
2148         }
2149         for (i = 0; i < 4; ++i) {
2150                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2151
2152                 if (root && VALID_PAGE(root)) {
2153                         root &= PT64_BASE_ADDR_MASK;
2154                         sp = page_header(root);
2155                         mmu_sync_children(vcpu, sp);
2156                 }
2157         }
2158 }
2159
2160 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2161 {
2162         spin_lock(&vcpu->kvm->mmu_lock);
2163         mmu_sync_roots(vcpu);
2164         spin_unlock(&vcpu->kvm->mmu_lock);
2165 }
2166
2167 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2168 {
2169         return vaddr;
2170 }
2171
2172 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2173                                 u32 error_code)
2174 {
2175         gfn_t gfn;
2176         int r;
2177
2178         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2179         r = mmu_topup_memory_caches(vcpu);
2180         if (r)
2181                 return r;
2182
2183         ASSERT(vcpu);
2184         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2185
2186         gfn = gva >> PAGE_SHIFT;
2187
2188         return nonpaging_map(vcpu, gva & PAGE_MASK,
2189                              error_code & PFERR_WRITE_MASK, gfn);
2190 }
2191
2192 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2193                                 u32 error_code)
2194 {
2195         pfn_t pfn;
2196         int r;
2197         int level;
2198         gfn_t gfn = gpa >> PAGE_SHIFT;
2199         unsigned long mmu_seq;
2200
2201         ASSERT(vcpu);
2202         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2203
2204         r = mmu_topup_memory_caches(vcpu);
2205         if (r)
2206                 return r;
2207
2208         level = mapping_level(vcpu, gfn);
2209
2210         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2211
2212         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2213         smp_rmb();
2214         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2215         if (is_error_pfn(pfn)) {
2216                 kvm_release_pfn_clean(pfn);
2217                 return 1;
2218         }
2219         spin_lock(&vcpu->kvm->mmu_lock);
2220         if (mmu_notifier_retry(vcpu, mmu_seq))
2221                 goto out_unlock;
2222         kvm_mmu_free_some_pages(vcpu);
2223         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2224                          level, gfn, pfn);
2225         spin_unlock(&vcpu->kvm->mmu_lock);
2226
2227         return r;
2228
2229 out_unlock:
2230         spin_unlock(&vcpu->kvm->mmu_lock);
2231         kvm_release_pfn_clean(pfn);
2232         return 0;
2233 }
2234
2235 static void nonpaging_free(struct kvm_vcpu *vcpu)
2236 {
2237         mmu_free_roots(vcpu);
2238 }
2239
2240 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2241 {
2242         struct kvm_mmu *context = &vcpu->arch.mmu;
2243
2244         context->new_cr3 = nonpaging_new_cr3;
2245         context->page_fault = nonpaging_page_fault;
2246         context->gva_to_gpa = nonpaging_gva_to_gpa;
2247         context->free = nonpaging_free;
2248         context->prefetch_page = nonpaging_prefetch_page;
2249         context->sync_page = nonpaging_sync_page;
2250         context->invlpg = nonpaging_invlpg;
2251         context->root_level = 0;
2252         context->shadow_root_level = PT32E_ROOT_LEVEL;
2253         context->root_hpa = INVALID_PAGE;
2254         return 0;
2255 }
2256
2257 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2258 {
2259         ++vcpu->stat.tlb_flush;
2260         kvm_x86_ops->tlb_flush(vcpu);
2261 }
2262
2263 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2264 {
2265         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2266         mmu_free_roots(vcpu);
2267 }
2268
2269 static void inject_page_fault(struct kvm_vcpu *vcpu,
2270                               u64 addr,
2271                               u32 err_code)
2272 {
2273         kvm_inject_page_fault(vcpu, addr, err_code);
2274 }
2275
2276 static void paging_free(struct kvm_vcpu *vcpu)
2277 {
2278         nonpaging_free(vcpu);
2279 }
2280
2281 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2282 {
2283         int bit7;
2284
2285         bit7 = (gpte >> 7) & 1;
2286         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2287 }
2288
2289 #define PTTYPE 64
2290 #include "paging_tmpl.h"
2291 #undef PTTYPE
2292
2293 #define PTTYPE 32
2294 #include "paging_tmpl.h"
2295 #undef PTTYPE
2296
2297 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2298 {
2299         struct kvm_mmu *context = &vcpu->arch.mmu;
2300         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2301         u64 exb_bit_rsvd = 0;
2302
2303         if (!is_nx(vcpu))
2304                 exb_bit_rsvd = rsvd_bits(63, 63);
2305         switch (level) {
2306         case PT32_ROOT_LEVEL:
2307                 /* no rsvd bits for 2 level 4K page table entries */
2308                 context->rsvd_bits_mask[0][1] = 0;
2309                 context->rsvd_bits_mask[0][0] = 0;
2310                 if (is_cpuid_PSE36())
2311                         /* 36bits PSE 4MB page */
2312                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2313                 else
2314                         /* 32 bits PSE 4MB page */
2315                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2316                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2317                 break;
2318         case PT32E_ROOT_LEVEL:
2319                 context->rsvd_bits_mask[0][2] =
2320                         rsvd_bits(maxphyaddr, 63) |
2321                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2322                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2323                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2324                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2325                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2326                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2327                         rsvd_bits(maxphyaddr, 62) |
2328                         rsvd_bits(13, 20);              /* large page */
2329                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2330                 break;
2331         case PT64_ROOT_LEVEL:
2332                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2333                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2334                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2335                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2336                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2337                         rsvd_bits(maxphyaddr, 51);
2338                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2339                         rsvd_bits(maxphyaddr, 51);
2340                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2341                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2342                         rsvd_bits(maxphyaddr, 51) |
2343                         rsvd_bits(13, 29);
2344                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2345                         rsvd_bits(maxphyaddr, 51) |
2346                         rsvd_bits(13, 20);              /* large page */
2347                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2348                 break;
2349         }
2350 }
2351
2352 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2353 {
2354         struct kvm_mmu *context = &vcpu->arch.mmu;
2355
2356         ASSERT(is_pae(vcpu));
2357         context->new_cr3 = paging_new_cr3;
2358         context->page_fault = paging64_page_fault;
2359         context->gva_to_gpa = paging64_gva_to_gpa;
2360         context->prefetch_page = paging64_prefetch_page;
2361         context->sync_page = paging64_sync_page;
2362         context->invlpg = paging64_invlpg;
2363         context->free = paging_free;
2364         context->root_level = level;
2365         context->shadow_root_level = level;
2366         context->root_hpa = INVALID_PAGE;
2367         return 0;
2368 }
2369
2370 static int paging64_init_context(struct kvm_vcpu *vcpu)
2371 {
2372         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2373         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2374 }
2375
2376 static int paging32_init_context(struct kvm_vcpu *vcpu)
2377 {
2378         struct kvm_mmu *context = &vcpu->arch.mmu;
2379
2380         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2381         context->new_cr3 = paging_new_cr3;
2382         context->page_fault = paging32_page_fault;
2383         context->gva_to_gpa = paging32_gva_to_gpa;
2384         context->free = paging_free;
2385         context->prefetch_page = paging32_prefetch_page;
2386         context->sync_page = paging32_sync_page;
2387         context->invlpg = paging32_invlpg;
2388         context->root_level = PT32_ROOT_LEVEL;
2389         context->shadow_root_level = PT32E_ROOT_LEVEL;
2390         context->root_hpa = INVALID_PAGE;
2391         return 0;
2392 }
2393
2394 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2395 {
2396         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2397         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2398 }
2399
2400 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2401 {
2402         struct kvm_mmu *context = &vcpu->arch.mmu;
2403
2404         context->new_cr3 = nonpaging_new_cr3;
2405         context->page_fault = tdp_page_fault;
2406         context->free = nonpaging_free;
2407         context->prefetch_page = nonpaging_prefetch_page;
2408         context->sync_page = nonpaging_sync_page;
2409         context->invlpg = nonpaging_invlpg;
2410         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2411         context->root_hpa = INVALID_PAGE;
2412
2413         if (!is_paging(vcpu)) {
2414                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2415                 context->root_level = 0;
2416         } else if (is_long_mode(vcpu)) {
2417                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2418                 context->gva_to_gpa = paging64_gva_to_gpa;
2419                 context->root_level = PT64_ROOT_LEVEL;
2420         } else if (is_pae(vcpu)) {
2421                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2422                 context->gva_to_gpa = paging64_gva_to_gpa;
2423                 context->root_level = PT32E_ROOT_LEVEL;
2424         } else {
2425                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2426                 context->gva_to_gpa = paging32_gva_to_gpa;
2427                 context->root_level = PT32_ROOT_LEVEL;
2428         }
2429
2430         return 0;
2431 }
2432
2433 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2434 {
2435         int r;
2436
2437         ASSERT(vcpu);
2438         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2439
2440         if (!is_paging(vcpu))
2441                 r = nonpaging_init_context(vcpu);
2442         else if (is_long_mode(vcpu))
2443                 r = paging64_init_context(vcpu);
2444         else if (is_pae(vcpu))
2445                 r = paging32E_init_context(vcpu);
2446         else
2447                 r = paging32_init_context(vcpu);
2448
2449         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2450
2451         return r;
2452 }
2453
2454 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2455 {
2456         vcpu->arch.update_pte.pfn = bad_pfn;
2457
2458         if (tdp_enabled)
2459                 return init_kvm_tdp_mmu(vcpu);
2460         else
2461                 return init_kvm_softmmu(vcpu);
2462 }
2463
2464 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2465 {
2466         ASSERT(vcpu);
2467         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2468                 vcpu->arch.mmu.free(vcpu);
2469                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2470         }
2471 }
2472
2473 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2474 {
2475         destroy_kvm_mmu(vcpu);
2476         return init_kvm_mmu(vcpu);
2477 }
2478 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2479
2480 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2481 {
2482         int r;
2483
2484         r = mmu_topup_memory_caches(vcpu);
2485         if (r)
2486                 goto out;
2487         spin_lock(&vcpu->kvm->mmu_lock);
2488         kvm_mmu_free_some_pages(vcpu);
2489         r = mmu_alloc_roots(vcpu);
2490         mmu_sync_roots(vcpu);
2491         spin_unlock(&vcpu->kvm->mmu_lock);
2492         if (r)
2493                 goto out;
2494         /* set_cr3() should ensure TLB has been flushed */
2495         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2496 out:
2497         return r;
2498 }
2499 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2500
2501 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2502 {
2503         mmu_free_roots(vcpu);
2504 }
2505
2506 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2507                                   struct kvm_mmu_page *sp,
2508                                   u64 *spte)
2509 {
2510         u64 pte;
2511         struct kvm_mmu_page *child;
2512
2513         pte = *spte;
2514         if (is_shadow_present_pte(pte)) {
2515                 if (is_last_spte(pte, sp->role.level))
2516                         rmap_remove(vcpu->kvm, spte);
2517                 else {
2518                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2519                         mmu_page_remove_parent_pte(child, spte);
2520                 }
2521         }
2522         __set_spte(spte, shadow_trap_nonpresent_pte);
2523         if (is_large_pte(pte))
2524                 --vcpu->kvm->stat.lpages;
2525 }
2526
2527 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2528                                   struct kvm_mmu_page *sp,
2529                                   u64 *spte,
2530                                   const void *new)
2531 {
2532         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2533                 ++vcpu->kvm->stat.mmu_pde_zapped;
2534                 return;
2535         }
2536
2537         ++vcpu->kvm->stat.mmu_pte_updated;
2538         if (sp->role.glevels == PT32_ROOT_LEVEL)
2539                 paging32_update_pte(vcpu, sp, spte, new);
2540         else
2541                 paging64_update_pte(vcpu, sp, spte, new);
2542 }
2543
2544 static bool need_remote_flush(u64 old, u64 new)
2545 {
2546         if (!is_shadow_present_pte(old))
2547                 return false;
2548         if (!is_shadow_present_pte(new))
2549                 return true;
2550         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2551                 return true;
2552         old ^= PT64_NX_MASK;
2553         new ^= PT64_NX_MASK;
2554         return (old & ~new & PT64_PERM_MASK) != 0;
2555 }
2556
2557 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2558 {
2559         if (need_remote_flush(old, new))
2560                 kvm_flush_remote_tlbs(vcpu->kvm);
2561         else
2562                 kvm_mmu_flush_tlb(vcpu);
2563 }
2564
2565 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2566 {
2567         u64 *spte = vcpu->arch.last_pte_updated;
2568
2569         return !!(spte && (*spte & shadow_accessed_mask));
2570 }
2571
2572 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2573                                           const u8 *new, int bytes)
2574 {
2575         gfn_t gfn;
2576         int r;
2577         u64 gpte = 0;
2578         pfn_t pfn;
2579
2580         if (bytes != 4 && bytes != 8)
2581                 return;
2582
2583         /*
2584          * Assume that the pte write on a page table of the same type
2585          * as the current vcpu paging mode.  This is nearly always true
2586          * (might be false while changing modes).  Note it is verified later
2587          * by update_pte().
2588          */
2589         if (is_pae(vcpu)) {
2590                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2591                 if ((bytes == 4) && (gpa % 4 == 0)) {
2592                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2593                         if (r)
2594                                 return;
2595                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2596                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2597                         memcpy((void *)&gpte, new, 8);
2598                 }
2599         } else {
2600                 if ((bytes == 4) && (gpa % 4 == 0))
2601                         memcpy((void *)&gpte, new, 4);
2602         }
2603         if (!is_present_gpte(gpte))
2604                 return;
2605         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2606
2607         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2608         smp_rmb();
2609         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2610
2611         if (is_error_pfn(pfn)) {
2612                 kvm_release_pfn_clean(pfn);
2613                 return;
2614         }
2615         vcpu->arch.update_pte.gfn = gfn;
2616         vcpu->arch.update_pte.pfn = pfn;
2617 }
2618
2619 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2620 {
2621         u64 *spte = vcpu->arch.last_pte_updated;
2622
2623         if (spte
2624             && vcpu->arch.last_pte_gfn == gfn
2625             && shadow_accessed_mask
2626             && !(*spte & shadow_accessed_mask)
2627             && is_shadow_present_pte(*spte))
2628                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2629 }
2630
2631 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2632                        const u8 *new, int bytes,
2633                        bool guest_initiated)
2634 {
2635         gfn_t gfn = gpa >> PAGE_SHIFT;
2636         struct kvm_mmu_page *sp;
2637         struct hlist_node *node, *n;
2638         struct hlist_head *bucket;
2639         unsigned index;
2640         u64 entry, gentry;
2641         u64 *spte;
2642         unsigned offset = offset_in_page(gpa);
2643         unsigned pte_size;
2644         unsigned page_offset;
2645         unsigned misaligned;
2646         unsigned quadrant;
2647         int level;
2648         int flooded = 0;
2649         int npte;
2650         int r;
2651
2652         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2653         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2654         spin_lock(&vcpu->kvm->mmu_lock);
2655         kvm_mmu_access_page(vcpu, gfn);
2656         kvm_mmu_free_some_pages(vcpu);
2657         ++vcpu->kvm->stat.mmu_pte_write;
2658         kvm_mmu_audit(vcpu, "pre pte write");
2659         if (guest_initiated) {
2660                 if (gfn == vcpu->arch.last_pt_write_gfn
2661                     && !last_updated_pte_accessed(vcpu)) {
2662                         ++vcpu->arch.last_pt_write_count;
2663                         if (vcpu->arch.last_pt_write_count >= 3)
2664                                 flooded = 1;
2665                 } else {
2666                         vcpu->arch.last_pt_write_gfn = gfn;
2667                         vcpu->arch.last_pt_write_count = 1;
2668                         vcpu->arch.last_pte_updated = NULL;
2669                 }
2670         }
2671         index = kvm_page_table_hashfn(gfn);
2672         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2673         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2674                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2675                         continue;
2676                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2677                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2678                 misaligned |= bytes < 4;
2679                 if (misaligned || flooded) {
2680                         /*
2681                          * Misaligned accesses are too much trouble to fix
2682                          * up; also, they usually indicate a page is not used
2683                          * as a page table.
2684                          *
2685                          * If we're seeing too many writes to a page,
2686                          * it may no longer be a page table, or we may be
2687                          * forking, in which case it is better to unmap the
2688                          * page.
2689                          */
2690                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2691                                  gpa, bytes, sp->role.word);
2692                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2693                                 n = bucket->first;
2694                         ++vcpu->kvm->stat.mmu_flooded;
2695                         continue;
2696                 }
2697                 page_offset = offset;
2698                 level = sp->role.level;
2699                 npte = 1;
2700                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2701                         page_offset <<= 1;      /* 32->64 */
2702                         /*
2703                          * A 32-bit pde maps 4MB while the shadow pdes map
2704                          * only 2MB.  So we need to double the offset again
2705                          * and zap two pdes instead of one.
2706                          */
2707                         if (level == PT32_ROOT_LEVEL) {
2708                                 page_offset &= ~7; /* kill rounding error */
2709                                 page_offset <<= 1;
2710                                 npte = 2;
2711                         }
2712                         quadrant = page_offset >> PAGE_SHIFT;
2713                         page_offset &= ~PAGE_MASK;
2714                         if (quadrant != sp->role.quadrant)
2715                                 continue;
2716                 }
2717                 spte = &sp->spt[page_offset / sizeof(*spte)];
2718                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2719                         gentry = 0;
2720                         r = kvm_read_guest_atomic(vcpu->kvm,
2721                                                   gpa & ~(u64)(pte_size - 1),
2722                                                   &gentry, pte_size);
2723                         new = (const void *)&gentry;
2724                         if (r < 0)
2725                                 new = NULL;
2726                 }
2727                 while (npte--) {
2728                         entry = *spte;
2729                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2730                         if (new)
2731                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2732                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2733                         ++spte;
2734                 }
2735         }
2736         kvm_mmu_audit(vcpu, "post pte write");
2737         spin_unlock(&vcpu->kvm->mmu_lock);
2738         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2739                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2740                 vcpu->arch.update_pte.pfn = bad_pfn;
2741         }
2742 }
2743
2744 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2745 {
2746         gpa_t gpa;
2747         int r;
2748
2749         if (tdp_enabled)
2750                 return 0;
2751
2752         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2753
2754         spin_lock(&vcpu->kvm->mmu_lock);
2755         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2756         spin_unlock(&vcpu->kvm->mmu_lock);
2757         return r;
2758 }
2759 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2760
2761 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2762 {
2763         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2764                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2765                 struct kvm_mmu_page *sp;
2766
2767                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2768                                   struct kvm_mmu_page, link);
2769                 kvm_mmu_zap_page(vcpu->kvm, sp);
2770                 ++vcpu->kvm->stat.mmu_recycled;
2771         }
2772 }
2773
2774 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2775 {
2776         int r;
2777         enum emulation_result er;
2778
2779         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2780         if (r < 0)
2781                 goto out;
2782
2783         if (!r) {
2784                 r = 1;
2785                 goto out;
2786         }
2787
2788         r = mmu_topup_memory_caches(vcpu);
2789         if (r)
2790                 goto out;
2791
2792         er = emulate_instruction(vcpu, cr2, error_code, 0);
2793
2794         switch (er) {
2795         case EMULATE_DONE:
2796                 return 1;
2797         case EMULATE_DO_MMIO:
2798                 ++vcpu->stat.mmio_exits;
2799                 return 0;
2800         case EMULATE_FAIL:
2801                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2802                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2803                 vcpu->run->internal.ndata = 0;
2804                 return 0;
2805         default:
2806                 BUG();
2807         }
2808 out:
2809         return r;
2810 }
2811 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2812
2813 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2814 {
2815         vcpu->arch.mmu.invlpg(vcpu, gva);
2816         kvm_mmu_flush_tlb(vcpu);
2817         ++vcpu->stat.invlpg;
2818 }
2819 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2820
2821 void kvm_enable_tdp(void)
2822 {
2823         tdp_enabled = true;
2824 }
2825 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2826
2827 void kvm_disable_tdp(void)
2828 {
2829         tdp_enabled = false;
2830 }
2831 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2832
2833 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2834 {
2835         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2836 }
2837
2838 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2839 {
2840         struct page *page;
2841         int i;
2842
2843         ASSERT(vcpu);
2844
2845         /*
2846          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2847          * Therefore we need to allocate shadow page tables in the first
2848          * 4GB of memory, which happens to fit the DMA32 zone.
2849          */
2850         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2851         if (!page)
2852                 goto error_1;
2853         vcpu->arch.mmu.pae_root = page_address(page);
2854         for (i = 0; i < 4; ++i)
2855                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2856
2857         return 0;
2858
2859 error_1:
2860         free_mmu_pages(vcpu);
2861         return -ENOMEM;
2862 }
2863
2864 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2865 {
2866         ASSERT(vcpu);
2867         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2868
2869         return alloc_mmu_pages(vcpu);
2870 }
2871
2872 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2873 {
2874         ASSERT(vcpu);
2875         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2876
2877         return init_kvm_mmu(vcpu);
2878 }
2879
2880 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2881 {
2882         ASSERT(vcpu);
2883
2884         destroy_kvm_mmu(vcpu);
2885         free_mmu_pages(vcpu);
2886         mmu_free_memory_caches(vcpu);
2887 }
2888
2889 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2890 {
2891         struct kvm_mmu_page *sp;
2892
2893         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2894                 int i;
2895                 u64 *pt;
2896
2897                 if (!test_bit(slot, sp->slot_bitmap))
2898                         continue;
2899
2900                 pt = sp->spt;
2901                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2902                         /* avoid RMW */
2903                         if (pt[i] & PT_WRITABLE_MASK)
2904                                 pt[i] &= ~PT_WRITABLE_MASK;
2905         }
2906         kvm_flush_remote_tlbs(kvm);
2907 }
2908
2909 void kvm_mmu_zap_all(struct kvm *kvm)
2910 {
2911         struct kvm_mmu_page *sp, *node;
2912
2913         spin_lock(&kvm->mmu_lock);
2914         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2915                 if (kvm_mmu_zap_page(kvm, sp))
2916                         node = container_of(kvm->arch.active_mmu_pages.next,
2917                                             struct kvm_mmu_page, link);
2918         spin_unlock(&kvm->mmu_lock);
2919
2920         kvm_flush_remote_tlbs(kvm);
2921 }
2922
2923 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2924 {
2925         struct kvm_mmu_page *page;
2926
2927         page = container_of(kvm->arch.active_mmu_pages.prev,
2928                             struct kvm_mmu_page, link);
2929         kvm_mmu_zap_page(kvm, page);
2930 }
2931
2932 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2933 {
2934         struct kvm *kvm;
2935         struct kvm *kvm_freed = NULL;
2936         int cache_count = 0;
2937
2938         spin_lock(&kvm_lock);
2939
2940         list_for_each_entry(kvm, &vm_list, vm_list) {
2941                 int npages;
2942
2943                 if (!down_read_trylock(&kvm->slots_lock))
2944                         continue;
2945                 spin_lock(&kvm->mmu_lock);
2946                 npages = kvm->arch.n_alloc_mmu_pages -
2947                          kvm->arch.n_free_mmu_pages;
2948                 cache_count += npages;
2949                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2950                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2951                         cache_count--;
2952                         kvm_freed = kvm;
2953                 }
2954                 nr_to_scan--;
2955
2956                 spin_unlock(&kvm->mmu_lock);
2957                 up_read(&kvm->slots_lock);
2958         }
2959         if (kvm_freed)
2960                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2961
2962         spin_unlock(&kvm_lock);
2963
2964         return cache_count;
2965 }
2966
2967 static struct shrinker mmu_shrinker = {
2968         .shrink = mmu_shrink,
2969         .seeks = DEFAULT_SEEKS * 10,
2970 };
2971
2972 static void mmu_destroy_caches(void)
2973 {
2974         if (pte_chain_cache)
2975                 kmem_cache_destroy(pte_chain_cache);
2976         if (rmap_desc_cache)
2977                 kmem_cache_destroy(rmap_desc_cache);
2978         if (mmu_page_header_cache)
2979                 kmem_cache_destroy(mmu_page_header_cache);
2980 }
2981
2982 void kvm_mmu_module_exit(void)
2983 {
2984         mmu_destroy_caches();
2985         unregister_shrinker(&mmu_shrinker);
2986 }
2987
2988 int kvm_mmu_module_init(void)
2989 {
2990         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2991                                             sizeof(struct kvm_pte_chain),
2992                                             0, 0, NULL);
2993         if (!pte_chain_cache)
2994                 goto nomem;
2995         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2996                                             sizeof(struct kvm_rmap_desc),
2997                                             0, 0, NULL);
2998         if (!rmap_desc_cache)
2999                 goto nomem;
3000
3001         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
3002                                                   sizeof(struct kvm_mmu_page),
3003                                                   0, 0, NULL);
3004         if (!mmu_page_header_cache)
3005                 goto nomem;
3006
3007         register_shrinker(&mmu_shrinker);
3008
3009         return 0;
3010
3011 nomem:
3012         mmu_destroy_caches();
3013         return -ENOMEM;
3014 }
3015
3016 /*
3017  * Caculate mmu pages needed for kvm.
3018  */
3019 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3020 {
3021         int i;
3022         unsigned int nr_mmu_pages;
3023         unsigned int  nr_pages = 0;
3024
3025         for (i = 0; i < kvm->memslots->nmemslots; i++)
3026                 nr_pages += kvm->memslots->memslots[i].npages;
3027
3028         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3029         nr_mmu_pages = max(nr_mmu_pages,
3030                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3031
3032         return nr_mmu_pages;
3033 }
3034
3035 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3036                                 unsigned len)
3037 {
3038         if (len > buffer->len)
3039                 return NULL;
3040         return buffer->ptr;
3041 }
3042
3043 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3044                                 unsigned len)
3045 {
3046         void *ret;
3047
3048         ret = pv_mmu_peek_buffer(buffer, len);
3049         if (!ret)
3050                 return ret;
3051         buffer->ptr += len;
3052         buffer->len -= len;
3053         buffer->processed += len;
3054         return ret;
3055 }
3056
3057 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3058                              gpa_t addr, gpa_t value)
3059 {
3060         int bytes = 8;
3061         int r;
3062
3063         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3064                 bytes = 4;
3065
3066         r = mmu_topup_memory_caches(vcpu);
3067         if (r)
3068                 return r;
3069
3070         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3071                 return -EFAULT;
3072
3073         return 1;
3074 }
3075
3076 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3077 {
3078         kvm_set_cr3(vcpu, vcpu->arch.cr3);
3079         return 1;
3080 }
3081
3082 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3083 {
3084         spin_lock(&vcpu->kvm->mmu_lock);
3085         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3086         spin_unlock(&vcpu->kvm->mmu_lock);
3087         return 1;
3088 }
3089
3090 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3091                              struct kvm_pv_mmu_op_buffer *buffer)
3092 {
3093         struct kvm_mmu_op_header *header;
3094
3095         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3096         if (!header)
3097                 return 0;
3098         switch (header->op) {
3099         case KVM_MMU_OP_WRITE_PTE: {
3100                 struct kvm_mmu_op_write_pte *wpte;
3101
3102                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3103                 if (!wpte)
3104                         return 0;
3105                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3106                                         wpte->pte_val);
3107         }
3108         case KVM_MMU_OP_FLUSH_TLB: {
3109                 struct kvm_mmu_op_flush_tlb *ftlb;
3110
3111                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3112                 if (!ftlb)
3113                         return 0;
3114                 return kvm_pv_mmu_flush_tlb(vcpu);
3115         }
3116         case KVM_MMU_OP_RELEASE_PT: {
3117                 struct kvm_mmu_op_release_pt *rpt;
3118
3119                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3120                 if (!rpt)
3121                         return 0;
3122                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3123         }
3124         default: return 0;
3125         }
3126 }
3127
3128 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3129                   gpa_t addr, unsigned long *ret)
3130 {
3131         int r;
3132         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3133
3134         buffer->ptr = buffer->buf;
3135         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3136         buffer->processed = 0;
3137
3138         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3139         if (r)
3140                 goto out;
3141
3142         while (buffer->len) {
3143                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3144                 if (r < 0)
3145                         goto out;
3146                 if (r == 0)
3147                         break;
3148         }
3149
3150         r = 1;
3151 out:
3152         *ret = buffer->processed;
3153         return r;
3154 }
3155
3156 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3157 {
3158         struct kvm_shadow_walk_iterator iterator;
3159         int nr_sptes = 0;
3160
3161         spin_lock(&vcpu->kvm->mmu_lock);
3162         for_each_shadow_entry(vcpu, addr, iterator) {
3163                 sptes[iterator.level-1] = *iterator.sptep;
3164                 nr_sptes++;
3165                 if (!is_shadow_present_pte(*iterator.sptep))
3166                         break;
3167         }
3168         spin_unlock(&vcpu->kvm->mmu_lock);
3169
3170         return nr_sptes;
3171 }
3172 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3173
3174 #ifdef AUDIT
3175
3176 static const char *audit_msg;
3177
3178 static gva_t canonicalize(gva_t gva)
3179 {
3180 #ifdef CONFIG_X86_64
3181         gva = (long long)(gva << 16) >> 16;
3182 #endif
3183         return gva;
3184 }
3185
3186
3187 typedef void (*inspect_spte_fn) (struct kvm *kvm, struct kvm_mmu_page *sp,
3188                                  u64 *sptep);
3189
3190 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3191                             inspect_spte_fn fn)
3192 {
3193         int i;
3194
3195         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3196                 u64 ent = sp->spt[i];
3197
3198                 if (is_shadow_present_pte(ent)) {
3199                         if (!is_last_spte(ent, sp->role.level)) {
3200                                 struct kvm_mmu_page *child;
3201                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3202                                 __mmu_spte_walk(kvm, child, fn);
3203                         } else
3204                                 fn(kvm, sp, &sp->spt[i]);
3205                 }
3206         }
3207 }
3208
3209 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3210 {
3211         int i;
3212         struct kvm_mmu_page *sp;
3213
3214         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3215                 return;
3216         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3217                 hpa_t root = vcpu->arch.mmu.root_hpa;
3218                 sp = page_header(root);
3219                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3220                 return;
3221         }
3222         for (i = 0; i < 4; ++i) {
3223                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3224
3225                 if (root && VALID_PAGE(root)) {
3226                         root &= PT64_BASE_ADDR_MASK;
3227                         sp = page_header(root);
3228                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3229                 }
3230         }
3231         return;
3232 }
3233
3234 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3235                                 gva_t va, int level)
3236 {
3237         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3238         int i;
3239         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3240
3241         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3242                 u64 ent = pt[i];
3243
3244                 if (ent == shadow_trap_nonpresent_pte)
3245                         continue;
3246
3247                 va = canonicalize(va);
3248                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3249                         audit_mappings_page(vcpu, ent, va, level - 1);
3250                 else {
3251                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3252                         gfn_t gfn = gpa >> PAGE_SHIFT;
3253                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3254                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3255
3256                         if (is_error_pfn(pfn)) {
3257                                 kvm_release_pfn_clean(pfn);
3258                                 continue;
3259                         }
3260
3261                         if (is_shadow_present_pte(ent)
3262                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3263                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3264                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3265                                        audit_msg, vcpu->arch.mmu.root_level,
3266                                        va, gpa, hpa, ent,
3267                                        is_shadow_present_pte(ent));
3268                         else if (ent == shadow_notrap_nonpresent_pte
3269                                  && !is_error_hpa(hpa))
3270                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3271                                        " valid guest gva %lx\n", audit_msg, va);
3272                         kvm_release_pfn_clean(pfn);
3273
3274                 }
3275         }
3276 }
3277
3278 static void audit_mappings(struct kvm_vcpu *vcpu)
3279 {
3280         unsigned i;
3281
3282         if (vcpu->arch.mmu.root_level == 4)
3283                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3284         else
3285                 for (i = 0; i < 4; ++i)
3286                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3287                                 audit_mappings_page(vcpu,
3288                                                     vcpu->arch.mmu.pae_root[i],
3289                                                     i << 30,
3290                                                     2);
3291 }
3292
3293 static int count_rmaps(struct kvm_vcpu *vcpu)
3294 {
3295         int nmaps = 0;
3296         int i, j, k;
3297
3298         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3299                 struct kvm_memory_slot *m = &vcpu->kvm->memslots->memslots[i];
3300                 struct kvm_rmap_desc *d;
3301
3302                 for (j = 0; j < m->npages; ++j) {
3303                         unsigned long *rmapp = &m->rmap[j];
3304
3305                         if (!*rmapp)
3306                                 continue;
3307                         if (!(*rmapp & 1)) {
3308                                 ++nmaps;
3309                                 continue;
3310                         }
3311                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3312                         while (d) {
3313                                 for (k = 0; k < RMAP_EXT; ++k)
3314                                         if (d->sptes[k])
3315                                                 ++nmaps;
3316                                         else
3317                                                 break;
3318                                 d = d->more;
3319                         }
3320                 }
3321         }
3322         return nmaps;
3323 }
3324
3325 void inspect_spte_has_rmap(struct kvm *kvm, struct kvm_mmu_page *sp, u64 *sptep)
3326 {
3327         unsigned long *rmapp;
3328         struct kvm_mmu_page *rev_sp;
3329         gfn_t gfn;
3330
3331         if (*sptep & PT_WRITABLE_MASK) {
3332                 rev_sp = page_header(__pa(sptep));
3333                 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3334
3335                 if (!gfn_to_memslot(kvm, gfn)) {
3336                         if (!printk_ratelimit())
3337                                 return;
3338                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3339                                          audit_msg, gfn);
3340                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3341                                         audit_msg, sptep - rev_sp->spt,
3342                                         rev_sp->gfn);
3343                         dump_stack();
3344                         return;
3345                 }
3346
3347                 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3348                                     is_large_pte(*sptep));
3349                 if (!*rmapp) {
3350                         if (!printk_ratelimit())
3351                                 return;
3352                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3353                                          audit_msg, *sptep);
3354                         dump_stack();
3355                 }
3356         }
3357
3358 }
3359
3360 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3361 {
3362         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3363 }
3364
3365 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3366 {
3367         struct kvm_mmu_page *sp;
3368         int i;
3369
3370         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3371                 u64 *pt = sp->spt;
3372
3373                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3374                         continue;
3375
3376                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3377                         u64 ent = pt[i];
3378
3379                         if (!(ent & PT_PRESENT_MASK))
3380                                 continue;
3381                         if (!(ent & PT_WRITABLE_MASK))
3382                                 continue;
3383                         inspect_spte_has_rmap(vcpu->kvm, sp, &pt[i]);
3384                 }
3385         }
3386         return;
3387 }
3388
3389 static void audit_rmap(struct kvm_vcpu *vcpu)
3390 {
3391         check_writable_mappings_rmap(vcpu);
3392         count_rmaps(vcpu);
3393 }
3394
3395 static void audit_write_protection(struct kvm_vcpu *vcpu)
3396 {
3397         struct kvm_mmu_page *sp;
3398         struct kvm_memory_slot *slot;
3399         unsigned long *rmapp;
3400         u64 *spte;
3401         gfn_t gfn;
3402
3403         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3404                 if (sp->role.direct)
3405                         continue;
3406                 if (sp->unsync)
3407                         continue;
3408
3409                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3410                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3411                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3412
3413                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3414                 while (spte) {
3415                         if (*spte & PT_WRITABLE_MASK)
3416                                 printk(KERN_ERR "%s: (%s) shadow page has "
3417                                 "writable mappings: gfn %lx role %x\n",
3418                                __func__, audit_msg, sp->gfn,
3419                                sp->role.word);
3420                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3421                 }
3422         }
3423 }
3424
3425 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3426 {
3427         int olddbg = dbg;
3428
3429         dbg = 0;
3430         audit_msg = msg;
3431         audit_rmap(vcpu);
3432         audit_write_protection(vcpu);
3433         if (strcmp("pre pte write", audit_msg) != 0)
3434                 audit_mappings(vcpu);
3435         audit_writable_sptes_have_rmaps(vcpu);
3436         dbg = olddbg;
3437 }
3438
3439 #endif