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 "x86.h"
  22 #include "kvm_cache_regs.h"
  23
  24 #include <linux/kvm_host.h>
  25 #include <linux/types.h>
  26 #include <linux/string.h>
  27 #include <linux/mm.h>
  28 #include <linux/highmem.h>
  29 #include <linux/module.h>
  30 #include <linux/swap.h>
  31 #include <linux/hugetlb.h>
  32 #include <linux/compiler.h>
  33 #include <linux/srcu.h>
  34 #include <linux/slab.h>
  35
  36 #include <asm/page.h>
  37 #include <asm/cmpxchg.h>
  38 #include <asm/io.h>
  39 #include <asm/vmx.h>
  40
  41 /*
  42  * When setting this variable to true it enables Two-Dimensional-Paging
  43  * where the hardware walks 2 page tables:
  44  * 1. the guest-virtual to guest-physical
  45  * 2. while doing 1. it walks guest-physical to host-physical
  46  * If the hardware supports that we don't need to do shadow paging.
  47  */
  48 bool tdp_enabled = false;
  49
  50 #undef MMU_DEBUG
  51
  52 #undef AUDIT
  53
  54 #ifdef AUDIT
  55 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  56 #else
  57 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  58 #endif
  59
  60 #ifdef MMU_DEBUG
  61
  62 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  63 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  64
  65 #else
  66
  67 #define pgprintk(x...) do { } while (0)
  68 #define rmap_printk(x...) do { } while (0)
  69
  70 #endif
  71
  72 #if defined(MMU_DEBUG) || defined(AUDIT)
  73 static int dbg = 0;
  74 module_param(dbg, bool, 0644);
  75 #endif
  76
  77 static int oos_shadow = 1;
  78 module_param(oos_shadow, bool, 0644);
  79
  80 #ifndef MMU_DEBUG
  81 #define ASSERT(x) do { } while (0)
  82 #else
  83 #define ASSERT(x)                                                       \
  84         if (!(x)) {                                                     \
  85                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  86                        __FILE__, __LINE__, #x);                         \
  87         }
  88 #endif
  89
  90 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  91 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  92
  93 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  94
  95 #define PT64_LEVEL_BITS 9
  96
  97 #define PT64_LEVEL_SHIFT(level) \
  98                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  99
 100 #define PT64_LEVEL_MASK(level) \
 101                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 102
 103 #define PT64_INDEX(address, level)\
 104         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 105
 106
 107 #define PT32_LEVEL_BITS 10
 108
 109 #define PT32_LEVEL_SHIFT(level) \
 110                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 111
 112 #define PT32_LEVEL_MASK(level) \
 113                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 114 #define PT32_LVL_OFFSET_MASK(level) \
 115         (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 116                                                 * PT32_LEVEL_BITS))) - 1))
 117
 118 #define PT32_INDEX(address, level)\
 119         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 120
 121
 122 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 123 #define PT64_DIR_BASE_ADDR_MASK \
 124         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 125 #define PT64_LVL_ADDR_MASK(level) \
 126         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 127                                                 * PT64_LEVEL_BITS))) - 1))
 128 #define PT64_LVL_OFFSET_MASK(level) \
 129         (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 130                                                 * PT64_LEVEL_BITS))) - 1))
 131
 132 #define PT32_BASE_ADDR_MASK PAGE_MASK
 133 #define PT32_DIR_BASE_ADDR_MASK \
 134         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 135 #define PT32_LVL_ADDR_MASK(level) \
 136         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 137                                             * PT32_LEVEL_BITS))) - 1))
 138
 139 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 140                         | PT64_NX_MASK)
 141
 142 #define RMAP_EXT 4
 143
 144 #define ACC_EXEC_MASK    1
 145 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 146 #define ACC_USER_MASK    PT_USER_MASK
 147 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 148
 149 #include <trace/events/kvm.h>
 150
 151 #define CREATE_TRACE_POINTS
 152 #include "mmutrace.h"
 153
 154 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 155
 156 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 157
 158 struct kvm_rmap_desc {
 159         u64 *sptes[RMAP_EXT];
 160         struct kvm_rmap_desc *more;
 161 };
 162
 163 struct kvm_shadow_walk_iterator {
 164         u64 addr;
 165         hpa_t shadow_addr;
 166         int level;
 167         u64 *sptep;
 168         unsigned index;
 169 };
 170
 171 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
 172         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
 173              shadow_walk_okay(&(_walker));                      \
 174              shadow_walk_next(&(_walker)))
 175
 176 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 177
 178 static struct kmem_cache *pte_chain_cache;
 179 static struct kmem_cache *rmap_desc_cache;
 180 static struct kmem_cache *mmu_page_header_cache;
 181
 182 static u64 __read_mostly shadow_trap_nonpresent_pte;
 183 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 184 static u64 __read_mostly shadow_base_present_pte;
 185 static u64 __read_mostly shadow_nx_mask;
 186 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 187 static u64 __read_mostly shadow_user_mask;
 188 static u64 __read_mostly shadow_accessed_mask;
 189 static u64 __read_mostly shadow_dirty_mask;
 190
 191 static inline u64 rsvd_bits(int s, int e)
 192 {
 193         return ((1ULL << (e - s + 1)) - 1) << s;
 194 }
 195
 196 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 197 {
 198         shadow_trap_nonpresent_pte = trap_pte;
 199         shadow_notrap_nonpresent_pte = notrap_pte;
 200 }
 201 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 202
 203 void kvm_mmu_set_base_ptes(u64 base_pte)
 204 {
 205         shadow_base_present_pte = base_pte;
 206 }
 207 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 208
 209 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 210                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 211 {
 212         shadow_user_mask = user_mask;
 213         shadow_accessed_mask = accessed_mask;
 214         shadow_dirty_mask = dirty_mask;
 215         shadow_nx_mask = nx_mask;
 216         shadow_x_mask = x_mask;
 217 }
 218 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 219
 220 static int is_write_protection(struct kvm_vcpu *vcpu)
 221 {
 222         return kvm_read_cr0_bits(vcpu, X86_CR0_WP);
 223 }
 224
 225 static int is_cpuid_PSE36(void)
 226 {
 227         return 1;
 228 }
 229
 230 static int is_nx(struct kvm_vcpu *vcpu)
 231 {
 232         return vcpu->arch.efer & EFER_NX;
 233 }
 234
 235 static int is_shadow_present_pte(u64 pte)
 236 {
 237         return pte != shadow_trap_nonpresent_pte
 238                 && pte != shadow_notrap_nonpresent_pte;
 239 }
 240
 241 static int is_large_pte(u64 pte)
 242 {
 243         return pte & PT_PAGE_SIZE_MASK;
 244 }
 245
 246 static int is_writable_pte(unsigned long pte)
 247 {
 248         return pte & PT_WRITABLE_MASK;
 249 }
 250
 251 static int is_dirty_gpte(unsigned long pte)
 252 {
 253         return pte & PT_DIRTY_MASK;
 254 }
 255
 256 static int is_rmap_spte(u64 pte)
 257 {
 258         return is_shadow_present_pte(pte);
 259 }
 260
 261 static int is_last_spte(u64 pte, int level)
 262 {
 263         if (level == PT_PAGE_TABLE_LEVEL)
 264                 return 1;
 265         if (is_large_pte(pte))
 266                 return 1;
 267         return 0;
 268 }
 269
 270 static pfn_t spte_to_pfn(u64 pte)
 271 {
 272         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 273 }
 274
 275 static gfn_t pse36_gfn_delta(u32 gpte)
 276 {
 277         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 278
 279         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 280 }
 281
 282 static void __set_spte(u64 *sptep, u64 spte)
 283 {
 284 #ifdef CONFIG_X86_64
 285         set_64bit((unsigned long *)sptep, spte);
 286 #else
 287         set_64bit((unsigned long long *)sptep, spte);
 288 #endif
 289 }
 290
 291 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 292                                   struct kmem_cache *base_cache, int min)
 293 {
 294         void *obj;
 295
 296         if (cache->nobjs >= min)
 297                 return 0;
 298         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 299                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 300                 if (!obj)
 301                         return -ENOMEM;
 302                 cache->objects[cache->nobjs++] = obj;
 303         }
 304         return 0;
 305 }
 306
 307 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 308 {
 309         while (mc->nobjs)
 310                 kfree(mc->objects[--mc->nobjs]);
 311 }
 312
 313 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 314                                        int min)
 315 {
 316         struct page *page;
 317
 318         if (cache->nobjs >= min)
 319                 return 0;
 320         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 321                 page = alloc_page(GFP_KERNEL);
 322                 if (!page)
 323                         return -ENOMEM;
 324                 cache->objects[cache->nobjs++] = page_address(page);
 325         }
 326         return 0;
 327 }
 328
 329 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 330 {
 331         while (mc->nobjs)
 332                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 333 }
 334
 335 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 336 {
 337         int r;
 338
 339         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 340                                    pte_chain_cache, 4);
 341         if (r)
 342                 goto out;
 343         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 344                                    rmap_desc_cache, 4);
 345         if (r)
 346                 goto out;
 347         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 348         if (r)
 349                 goto out;
 350         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 351                                    mmu_page_header_cache, 4);
 352 out:
 353         return r;
 354 }
 355
 356 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 357 {
 358         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 359         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 360         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 361         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 362 }
 363
 364 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 365                                     size_t size)
 366 {
 367         void *p;
 368
 369         BUG_ON(!mc->nobjs);
 370         p = mc->objects[--mc->nobjs];
 371         return p;
 372 }
 373
 374 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 375 {
 376         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 377                                       sizeof(struct kvm_pte_chain));
 378 }
 379
 380 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 381 {
 382         kfree(pc);
 383 }
 384
 385 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 386 {
 387         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 388                                       sizeof(struct kvm_rmap_desc));
 389 }
 390
 391 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 392 {
 393         kfree(rd);
 394 }
 395
 396 /*
 397  * Return the pointer to the largepage write count for a given
 398  * gfn, handling slots that are not large page aligned.
 399  */
 400 static int *slot_largepage_idx(gfn_t gfn,
 401                                struct kvm_memory_slot *slot,
 402                                int level)
 403 {
 404         unsigned long idx;
 405
 406         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 407               (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 408         return &slot->lpage_info[level - 2][idx].write_count;
 409 }
 410
 411 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 412 {
 413         struct kvm_memory_slot *slot;
 414         int *write_count;
 415         int i;
 416
 417         gfn = unalias_gfn(kvm, gfn);
 418
 419         slot = gfn_to_memslot_unaliased(kvm, gfn);
 420         for (i = PT_DIRECTORY_LEVEL;
 421              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 422                 write_count   = slot_largepage_idx(gfn, slot, i);
 423                 *write_count += 1;
 424         }
 425 }
 426
 427 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 428 {
 429         struct kvm_memory_slot *slot;
 430         int *write_count;
 431         int i;
 432
 433         gfn = unalias_gfn(kvm, gfn);
 434         for (i = PT_DIRECTORY_LEVEL;
 435              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 436                 slot          = gfn_to_memslot_unaliased(kvm, gfn);
 437                 write_count   = slot_largepage_idx(gfn, slot, i);
 438                 *write_count -= 1;
 439                 WARN_ON(*write_count < 0);
 440         }
 441 }
 442
 443 static int has_wrprotected_page(struct kvm *kvm,
 444                                 gfn_t gfn,
 445                                 int level)
 446 {
 447         struct kvm_memory_slot *slot;
 448         int *largepage_idx;
 449
 450         gfn = unalias_gfn(kvm, gfn);
 451         slot = gfn_to_memslot_unaliased(kvm, gfn);
 452         if (slot) {
 453                 largepage_idx = slot_largepage_idx(gfn, slot, level);
 454                 return *largepage_idx;
 455         }
 456
 457         return 1;
 458 }
 459
 460 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
 461 {
 462         unsigned long page_size;
 463         int i, ret = 0;
 464
 465         page_size = kvm_host_page_size(kvm, gfn);
 466
 467         for (i = PT_PAGE_TABLE_LEVEL;
 468              i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
 469                 if (page_size >= KVM_HPAGE_SIZE(i))
 470                         ret = i;
 471                 else
 472                         break;
 473         }
 474
 475         return ret;
 476 }
 477
 478 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 479 {
 480         struct kvm_memory_slot *slot;
 481         int host_level, level, max_level;
 482
 483         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 484         if (slot && slot->dirty_bitmap)
 485                 return PT_PAGE_TABLE_LEVEL;
 486
 487         host_level = host_mapping_level(vcpu->kvm, large_gfn);
 488
 489         if (host_level == PT_PAGE_TABLE_LEVEL)
 490                 return host_level;
 491
 492         max_level = kvm_x86_ops->get_lpage_level() < host_level ?
 493                 kvm_x86_ops->get_lpage_level() : host_level;
 494
 495         for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
 496                 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
 497                         break;
 498
 499         return level - 1;
 500 }
 501
 502 /*
 503  * Take gfn and return the reverse mapping to it.
 504  * Note: gfn must be unaliased before this function get called
 505  */
 506
 507 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
 508 {
 509         struct kvm_memory_slot *slot;
 510         unsigned long idx;
 511
 512         slot = gfn_to_memslot(kvm, gfn);
 513         if (likely(level == PT_PAGE_TABLE_LEVEL))
 514                 return &slot->rmap[gfn - slot->base_gfn];
 515
 516         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 517                 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 518
 519         return &slot->lpage_info[level - 2][idx].rmap_pde;
 520 }
 521
 522 /*
 523  * Reverse mapping data structures:
 524  *
 525  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 526  * that points to page_address(page).
 527  *
 528  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 529  * containing more mappings.
 530  *
 531  * Returns the number of rmap entries before the spte was added or zero if
 532  * the spte was not added.
 533  *
 534  */
 535 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 536 {
 537         struct kvm_mmu_page *sp;
 538         struct kvm_rmap_desc *desc;
 539         unsigned long *rmapp;
 540         int i, count = 0;
 541
 542         if (!is_rmap_spte(*spte))
 543                 return count;
 544         gfn = unalias_gfn(vcpu->kvm, gfn);
 545         sp = page_header(__pa(spte));
 546         sp->gfns[spte - sp->spt] = gfn;
 547         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 548         if (!*rmapp) {
 549                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 550                 *rmapp = (unsigned long)spte;
 551         } else if (!(*rmapp & 1)) {
 552                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 553                 desc = mmu_alloc_rmap_desc(vcpu);
 554                 desc->sptes[0] = (u64 *)*rmapp;
 555                 desc->sptes[1] = spte;
 556                 *rmapp = (unsigned long)desc | 1;
 557         } else {
 558                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 559                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 560                 while (desc->sptes[RMAP_EXT-1] && desc->more) {
 561                         desc = desc->more;
 562                         count += RMAP_EXT;
 563                 }
 564                 if (desc->sptes[RMAP_EXT-1]) {
 565                         desc->more = mmu_alloc_rmap_desc(vcpu);
 566                         desc = desc->more;
 567                 }
 568                 for (i = 0; desc->sptes[i]; ++i)
 569                         ;
 570                 desc->sptes[i] = spte;
 571         }
 572         return count;
 573 }
 574
 575 static void rmap_desc_remove_entry(unsigned long *rmapp,
 576                                    struct kvm_rmap_desc *desc,
 577                                    int i,
 578                                    struct kvm_rmap_desc *prev_desc)
 579 {
 580         int j;
 581
 582         for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
 583                 ;
 584         desc->sptes[i] = desc->sptes[j];
 585         desc->sptes[j] = NULL;
 586         if (j != 0)
 587                 return;
 588         if (!prev_desc && !desc->more)
 589                 *rmapp = (unsigned long)desc->sptes[0];
 590         else
 591                 if (prev_desc)
 592                         prev_desc->more = desc->more;
 593                 else
 594                         *rmapp = (unsigned long)desc->more | 1;
 595         mmu_free_rmap_desc(desc);
 596 }
 597
 598 static void rmap_remove(struct kvm *kvm, u64 *spte)
 599 {
 600         struct kvm_rmap_desc *desc;
 601         struct kvm_rmap_desc *prev_desc;
 602         struct kvm_mmu_page *sp;
 603         pfn_t pfn;
 604         unsigned long *rmapp;
 605         int i;
 606
 607         if (!is_rmap_spte(*spte))
 608                 return;
 609         sp = page_header(__pa(spte));
 610         pfn = spte_to_pfn(*spte);
 611         if (*spte & shadow_accessed_mask)
 612                 kvm_set_pfn_accessed(pfn);
 613         if (is_writable_pte(*spte))
 614                 kvm_set_pfn_dirty(pfn);
 615         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], sp->role.level);
 616         if (!*rmapp) {
 617                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 618                 BUG();
 619         } else if (!(*rmapp & 1)) {
 620                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 621                 if ((u64 *)*rmapp != spte) {
 622                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 623                                spte, *spte);
 624                         BUG();
 625                 }
 626                 *rmapp = 0;
 627         } else {
 628                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 629                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 630                 prev_desc = NULL;
 631                 while (desc) {
 632                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
 633                                 if (desc->sptes[i] == spte) {
 634                                         rmap_desc_remove_entry(rmapp,
 635                                                                desc, i,
 636                                                                prev_desc);
 637                                         return;
 638                                 }
 639                         prev_desc = desc;
 640                         desc = desc->more;
 641                 }
 642                 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
 643                 BUG();
 644         }
 645 }
 646
 647 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 648 {
 649         struct kvm_rmap_desc *desc;
 650         struct kvm_rmap_desc *prev_desc;
 651         u64 *prev_spte;
 652         int i;
 653
 654         if (!*rmapp)
 655                 return NULL;
 656         else if (!(*rmapp & 1)) {
 657                 if (!spte)
 658                         return (u64 *)*rmapp;
 659                 return NULL;
 660         }
 661         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 662         prev_desc = NULL;
 663         prev_spte = NULL;
 664         while (desc) {
 665                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
 666                         if (prev_spte == spte)
 667                                 return desc->sptes[i];
 668                         prev_spte = desc->sptes[i];
 669                 }
 670                 desc = desc->more;
 671         }
 672         return NULL;
 673 }
 674
 675 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 676 {
 677         unsigned long *rmapp;
 678         u64 *spte;
 679         int i, write_protected = 0;
 680
 681         gfn = unalias_gfn(kvm, gfn);
 682         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
 683
 684         spte = rmap_next(kvm, rmapp, NULL);
 685         while (spte) {
 686                 BUG_ON(!spte);
 687                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 688                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 689                 if (is_writable_pte(*spte)) {
 690                         __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
 691                         write_protected = 1;
 692                 }
 693                 spte = rmap_next(kvm, rmapp, spte);
 694         }
 695         if (write_protected) {
 696                 pfn_t pfn;
 697
 698                 spte = rmap_next(kvm, rmapp, NULL);
 699                 pfn = spte_to_pfn(*spte);
 700                 kvm_set_pfn_dirty(pfn);
 701         }
 702
 703         /* check for huge page mappings */
 704         for (i = PT_DIRECTORY_LEVEL;
 705              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 706                 rmapp = gfn_to_rmap(kvm, gfn, i);
 707                 spte = rmap_next(kvm, rmapp, NULL);
 708                 while (spte) {
 709                         BUG_ON(!spte);
 710                         BUG_ON(!(*spte & PT_PRESENT_MASK));
 711                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 712                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 713                         if (is_writable_pte(*spte)) {
 714                                 rmap_remove(kvm, spte);
 715                                 --kvm->stat.lpages;
 716                                 __set_spte(spte, shadow_trap_nonpresent_pte);
 717                                 spte = NULL;
 718                                 write_protected = 1;
 719                         }
 720                         spte = rmap_next(kvm, rmapp, spte);
 721                 }
 722         }
 723
 724         return write_protected;
 725 }
 726
 727 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 728                            unsigned long data)
 729 {
 730         u64 *spte;
 731         int need_tlb_flush = 0;
 732
 733         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 734                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 735                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 736                 rmap_remove(kvm, spte);
 737                 __set_spte(spte, shadow_trap_nonpresent_pte);
 738                 need_tlb_flush = 1;
 739         }
 740         return need_tlb_flush;
 741 }
 742
 743 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
 744                              unsigned long data)
 745 {
 746         int need_flush = 0;
 747         u64 *spte, new_spte;
 748         pte_t *ptep = (pte_t *)data;
 749         pfn_t new_pfn;
 750
 751         WARN_ON(pte_huge(*ptep));
 752         new_pfn = pte_pfn(*ptep);
 753         spte = rmap_next(kvm, rmapp, NULL);
 754         while (spte) {
 755                 BUG_ON(!is_shadow_present_pte(*spte));
 756                 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
 757                 need_flush = 1;
 758                 if (pte_write(*ptep)) {
 759                         rmap_remove(kvm, spte);
 760                         __set_spte(spte, shadow_trap_nonpresent_pte);
 761                         spte = rmap_next(kvm, rmapp, NULL);
 762                 } else {
 763                         new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
 764                         new_spte |= (u64)new_pfn << PAGE_SHIFT;
 765
 766                         new_spte &= ~PT_WRITABLE_MASK;
 767                         new_spte &= ~SPTE_HOST_WRITEABLE;
 768                         if (is_writable_pte(*spte))
 769                                 kvm_set_pfn_dirty(spte_to_pfn(*spte));
 770                         __set_spte(spte, new_spte);
 771                         spte = rmap_next(kvm, rmapp, spte);
 772                 }
 773         }
 774         if (need_flush)
 775                 kvm_flush_remote_tlbs(kvm);
 776
 777         return 0;
 778 }
 779
 780 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 781                           unsigned long data,
 782                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 783                                          unsigned long data))
 784 {
 785         int i, j;
 786         int ret;
 787         int retval = 0;
 788         struct kvm_memslots *slots;
 789
 790         slots = rcu_dereference(kvm->memslots);
 791
 792         for (i = 0; i < slots->nmemslots; i++) {
 793                 struct kvm_memory_slot *memslot = &slots->memslots[i];
 794                 unsigned long start = memslot->userspace_addr;
 795                 unsigned long end;
 796
 797                 end = start + (memslot->npages << PAGE_SHIFT);
 798                 if (hva >= start && hva < end) {
 799                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 800
 801                         ret = handler(kvm, &memslot->rmap[gfn_offset], data);
 802
 803                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
 804                                 int idx = gfn_offset;
 805                                 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
 806                                 ret |= handler(kvm,
 807                                         &memslot->lpage_info[j][idx].rmap_pde,
 808                                         data);
 809                         }
 810                         trace_kvm_age_page(hva, memslot, ret);
 811                         retval |= ret;
 812                 }
 813         }
 814
 815         return retval;
 816 }
 817
 818 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 819 {
 820         return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
 821 }
 822
 823 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
 824 {
 825         kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
 826 }
 827
 828 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 829                          unsigned long data)
 830 {
 831         u64 *spte;
 832         int young = 0;
 833
 834         /*
 835          * Emulate the accessed bit for EPT, by checking if this page has
 836          * an EPT mapping, and clearing it if it does. On the next access,
 837          * a new EPT mapping will be established.
 838          * This has some overhead, but not as much as the cost of swapping
 839          * out actively used pages or breaking up actively used hugepages.
 840          */
 841         if (!shadow_accessed_mask)
 842                 return kvm_unmap_rmapp(kvm, rmapp, data);
 843
 844         spte = rmap_next(kvm, rmapp, NULL);
 845         while (spte) {
 846                 int _young;
 847                 u64 _spte = *spte;
 848                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 849                 _young = _spte & PT_ACCESSED_MASK;
 850                 if (_young) {
 851                         young = 1;
 852                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 853                 }
 854                 spte = rmap_next(kvm, rmapp, spte);
 855         }
 856         return young;
 857 }
 858
 859 #define RMAP_RECYCLE_THRESHOLD 1000
 860
 861 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 862 {
 863         unsigned long *rmapp;
 864         struct kvm_mmu_page *sp;
 865
 866         sp = page_header(__pa(spte));
 867
 868         gfn = unalias_gfn(vcpu->kvm, gfn);
 869         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 870
 871         kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
 872         kvm_flush_remote_tlbs(vcpu->kvm);
 873 }
 874
 875 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 876 {
 877         return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
 878 }
 879
 880 #ifdef MMU_DEBUG
 881 static int is_empty_shadow_page(u64 *spt)
 882 {
 883         u64 *pos;
 884         u64 *end;
 885
 886         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 887                 if (is_shadow_present_pte(*pos)) {
 888                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 889                                pos, *pos);
 890                         return 0;
 891                 }
 892         return 1;
 893 }
 894 #endif
 895
 896 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 897 {
 898         ASSERT(is_empty_shadow_page(sp->spt));
 899         list_del(&sp->link);
 900         __free_page(virt_to_page(sp->spt));
 901         __free_page(virt_to_page(sp->gfns));
 902         kfree(sp);
 903         ++kvm->arch.n_free_mmu_pages;
 904 }
 905
 906 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 907 {
 908         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 909 }
 910
 911 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 912                                                u64 *parent_pte)
 913 {
 914         struct kvm_mmu_page *sp;
 915
 916         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 917         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 918         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 919         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 920         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 921         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 922         sp->multimapped = 0;
 923         sp->parent_pte = parent_pte;
 924         --vcpu->kvm->arch.n_free_mmu_pages;
 925         return sp;
 926 }
 927
 928 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 929                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 930 {
 931         struct kvm_pte_chain *pte_chain;
 932         struct hlist_node *node;
 933         int i;
 934
 935         if (!parent_pte)
 936                 return;
 937         if (!sp->multimapped) {
 938                 u64 *old = sp->parent_pte;
 939
 940                 if (!old) {
 941                         sp->parent_pte = parent_pte;
 942                         return;
 943                 }
 944                 sp->multimapped = 1;
 945                 pte_chain = mmu_alloc_pte_chain(vcpu);
 946                 INIT_HLIST_HEAD(&sp->parent_ptes);
 947                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 948                 pte_chain->parent_ptes[0] = old;
 949         }
 950         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 951                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 952                         continue;
 953                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 954                         if (!pte_chain->parent_ptes[i]) {
 955                                 pte_chain->parent_ptes[i] = parent_pte;
 956                                 return;
 957                         }
 958         }
 959         pte_chain = mmu_alloc_pte_chain(vcpu);
 960         BUG_ON(!pte_chain);
 961         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 962         pte_chain->parent_ptes[0] = parent_pte;
 963 }
 964
 965 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 966                                        u64 *parent_pte)
 967 {
 968         struct kvm_pte_chain *pte_chain;
 969         struct hlist_node *node;
 970         int i;
 971
 972         if (!sp->multimapped) {
 973                 BUG_ON(sp->parent_pte != parent_pte);
 974                 sp->parent_pte = NULL;
 975                 return;
 976         }
 977         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 978                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 979                         if (!pte_chain->parent_ptes[i])
 980                                 break;
 981                         if (pte_chain->parent_ptes[i] != parent_pte)
 982                                 continue;
 983                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 984                                 && pte_chain->parent_ptes[i + 1]) {
 985                                 pte_chain->parent_ptes[i]
 986                                         = pte_chain->parent_ptes[i + 1];
 987                                 ++i;
 988                         }
 989                         pte_chain->parent_ptes[i] = NULL;
 990                         if (i == 0) {
 991                                 hlist_del(&pte_chain->link);
 992                                 mmu_free_pte_chain(pte_chain);
 993                                 if (hlist_empty(&sp->parent_ptes)) {
 994                                         sp->multimapped = 0;
 995                                         sp->parent_pte = NULL;
 996                                 }
 997                         }
 998                         return;
 999                 }
1000         BUG();
1001 }
1002
1003
1004 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1005                             mmu_parent_walk_fn fn)
1006 {
1007         struct kvm_pte_chain *pte_chain;
1008         struct hlist_node *node;
1009         struct kvm_mmu_page *parent_sp;
1010         int i;
1011
1012         if (!sp->multimapped && sp->parent_pte) {
1013                 parent_sp = page_header(__pa(sp->parent_pte));
1014                 fn(vcpu, parent_sp);
1015                 mmu_parent_walk(vcpu, parent_sp, fn);
1016                 return;
1017         }
1018         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1019                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1020                         if (!pte_chain->parent_ptes[i])
1021                                 break;
1022                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
1023                         fn(vcpu, parent_sp);
1024                         mmu_parent_walk(vcpu, parent_sp, fn);
1025                 }
1026 }
1027
1028 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1029 {
1030         unsigned int index;
1031         struct kvm_mmu_page *sp = page_header(__pa(spte));
1032
1033         index = spte - sp->spt;
1034         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1035                 sp->unsync_children++;
1036         WARN_ON(!sp->unsync_children);
1037 }
1038
1039 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1040 {
1041         struct kvm_pte_chain *pte_chain;
1042         struct hlist_node *node;
1043         int i;
1044
1045         if (!sp->parent_pte)
1046                 return;
1047
1048         if (!sp->multimapped) {
1049                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1050                 return;
1051         }
1052
1053         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1054                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1055                         if (!pte_chain->parent_ptes[i])
1056                                 break;
1057                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1058                 }
1059 }
1060
1061 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1062 {
1063         kvm_mmu_update_parents_unsync(sp);
1064         return 1;
1065 }
1066
1067 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
1068                                         struct kvm_mmu_page *sp)
1069 {
1070         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
1071         kvm_mmu_update_parents_unsync(sp);
1072 }
1073
1074 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1075                                     struct kvm_mmu_page *sp)
1076 {
1077         int i;
1078
1079         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1080                 sp->spt[i] = shadow_trap_nonpresent_pte;
1081 }
1082
1083 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1084                                struct kvm_mmu_page *sp)
1085 {
1086         return 1;
1087 }
1088
1089 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1090 {
1091 }
1092
1093 #define KVM_PAGE_ARRAY_NR 16
1094
1095 struct kvm_mmu_pages {
1096         struct mmu_page_and_offset {
1097                 struct kvm_mmu_page *sp;
1098                 unsigned int idx;
1099         } page[KVM_PAGE_ARRAY_NR];
1100         unsigned int nr;
1101 };
1102
1103 #define for_each_unsync_children(bitmap, idx)           \
1104         for (idx = find_first_bit(bitmap, 512);         \
1105              idx < 512;                                 \
1106              idx = find_next_bit(bitmap, 512, idx+1))
1107
1108 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1109                          int idx)
1110 {
1111         int i;
1112
1113         if (sp->unsync)
1114                 for (i=0; i < pvec->nr; i++)
1115                         if (pvec->page[i].sp == sp)
1116                                 return 0;
1117
1118         pvec->page[pvec->nr].sp = sp;
1119         pvec->page[pvec->nr].idx = idx;
1120         pvec->nr++;
1121         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1122 }
1123
1124 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1125                            struct kvm_mmu_pages *pvec)
1126 {
1127         int i, ret, nr_unsync_leaf = 0;
1128
1129         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1130                 u64 ent = sp->spt[i];
1131
1132                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1133                         struct kvm_mmu_page *child;
1134                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1135
1136                         if (child->unsync_children) {
1137                                 if (mmu_pages_add(pvec, child, i))
1138                                         return -ENOSPC;
1139
1140                                 ret = __mmu_unsync_walk(child, pvec);
1141                                 if (!ret)
1142                                         __clear_bit(i, sp->unsync_child_bitmap);
1143                                 else if (ret > 0)
1144                                         nr_unsync_leaf += ret;
1145                                 else
1146                                         return ret;
1147                         }
1148
1149                         if (child->unsync) {
1150                                 nr_unsync_leaf++;
1151                                 if (mmu_pages_add(pvec, child, i))
1152                                         return -ENOSPC;
1153                         }
1154                 }
1155         }
1156
1157         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1158                 sp->unsync_children = 0;
1159
1160         return nr_unsync_leaf;
1161 }
1162
1163 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1164                            struct kvm_mmu_pages *pvec)
1165 {
1166         if (!sp->unsync_children)
1167                 return 0;
1168
1169         mmu_pages_add(pvec, sp, 0);
1170         return __mmu_unsync_walk(sp, pvec);
1171 }
1172
1173 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1174 {
1175         unsigned index;
1176         struct hlist_head *bucket;
1177         struct kvm_mmu_page *sp;
1178         struct hlist_node *node;
1179
1180         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1181         index = kvm_page_table_hashfn(gfn);
1182         bucket = &kvm->arch.mmu_page_hash[index];
1183         hlist_for_each_entry(sp, node, bucket, hash_link)
1184                 if (sp->gfn == gfn && !sp->role.direct
1185                     && !sp->role.invalid) {
1186                         pgprintk("%s: found role %x\n",
1187                                  __func__, sp->role.word);
1188                         return sp;
1189                 }
1190         return NULL;
1191 }
1192
1193 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1194 {
1195         WARN_ON(!sp->unsync);
1196         sp->unsync = 0;
1197         --kvm->stat.mmu_unsync;
1198 }
1199
1200 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1201
1202 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1203 {
1204         if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1205                 kvm_mmu_zap_page(vcpu->kvm, sp);
1206                 return 1;
1207         }
1208
1209         trace_kvm_mmu_sync_page(sp);
1210         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1211                 kvm_flush_remote_tlbs(vcpu->kvm);
1212         kvm_unlink_unsync_page(vcpu->kvm, sp);
1213         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1214                 kvm_mmu_zap_page(vcpu->kvm, sp);
1215                 return 1;
1216         }
1217
1218         kvm_mmu_flush_tlb(vcpu);
1219         return 0;
1220 }
1221
1222 struct mmu_page_path {
1223         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1224         unsigned int idx[PT64_ROOT_LEVEL-1];
1225 };
1226
1227 #define for_each_sp(pvec, sp, parents, i)                       \
1228                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1229                         sp = pvec.page[i].sp;                   \
1230                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1231                         i = mmu_pages_next(&pvec, &parents, i))
1232
1233 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1234                           struct mmu_page_path *parents,
1235                           int i)
1236 {
1237         int n;
1238
1239         for (n = i+1; n < pvec->nr; n++) {
1240                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1241
1242                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1243                         parents->idx[0] = pvec->page[n].idx;
1244                         return n;
1245                 }
1246
1247                 parents->parent[sp->role.level-2] = sp;
1248                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1249         }
1250
1251         return n;
1252 }
1253
1254 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1255 {
1256         struct kvm_mmu_page *sp;
1257         unsigned int level = 0;
1258
1259         do {
1260                 unsigned int idx = parents->idx[level];
1261
1262                 sp = parents->parent[level];
1263                 if (!sp)
1264                         return;
1265
1266                 --sp->unsync_children;
1267                 WARN_ON((int)sp->unsync_children < 0);
1268                 __clear_bit(idx, sp->unsync_child_bitmap);
1269                 level++;
1270         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1271 }
1272
1273 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1274                                struct mmu_page_path *parents,
1275                                struct kvm_mmu_pages *pvec)
1276 {
1277         parents->parent[parent->role.level-1] = NULL;
1278         pvec->nr = 0;
1279 }
1280
1281 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1282                               struct kvm_mmu_page *parent)
1283 {
1284         int i;
1285         struct kvm_mmu_page *sp;
1286         struct mmu_page_path parents;
1287         struct kvm_mmu_pages pages;
1288
1289         kvm_mmu_pages_init(parent, &parents, &pages);
1290         while (mmu_unsync_walk(parent, &pages)) {
1291                 int protected = 0;
1292
1293                 for_each_sp(pages, sp, parents, i)
1294                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1295
1296                 if (protected)
1297                         kvm_flush_remote_tlbs(vcpu->kvm);
1298
1299                 for_each_sp(pages, sp, parents, i) {
1300                         kvm_sync_page(vcpu, sp);
1301                         mmu_pages_clear_parents(&parents);
1302                 }
1303                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1304                 kvm_mmu_pages_init(parent, &parents, &pages);
1305         }
1306 }
1307
1308 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1309                                              gfn_t gfn,
1310                                              gva_t gaddr,
1311                                              unsigned level,
1312                                              int direct,
1313                                              unsigned access,
1314                                              u64 *parent_pte)
1315 {
1316         union kvm_mmu_page_role role;
1317         unsigned index;
1318         unsigned quadrant;
1319         struct hlist_head *bucket;
1320         struct kvm_mmu_page *sp;
1321         struct hlist_node *node, *tmp;
1322
1323         role = vcpu->arch.mmu.base_role;
1324         role.level = level;
1325         role.direct = direct;
1326         if (role.direct)
1327                 role.cr4_pae = 0;
1328         role.access = access;
1329         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1330                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1331                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1332                 role.quadrant = quadrant;
1333         }
1334         index = kvm_page_table_hashfn(gfn);
1335         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1336         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1337                 if (sp->gfn == gfn) {
1338                         if (sp->unsync)
1339                                 if (kvm_sync_page(vcpu, sp))
1340                                         continue;
1341
1342                         if (sp->role.word != role.word)
1343                                 continue;
1344
1345                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1346                         if (sp->unsync_children) {
1347                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1348                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1349                         }
1350                         trace_kvm_mmu_get_page(sp, false);
1351                         return sp;
1352                 }
1353         ++vcpu->kvm->stat.mmu_cache_miss;
1354         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1355         if (!sp)
1356                 return sp;
1357         sp->gfn = gfn;
1358         sp->role = role;
1359         hlist_add_head(&sp->hash_link, bucket);
1360         if (!direct) {
1361                 if (rmap_write_protect(vcpu->kvm, gfn))
1362                         kvm_flush_remote_tlbs(vcpu->kvm);
1363                 account_shadowed(vcpu->kvm, gfn);
1364         }
1365         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1366                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1367         else
1368                 nonpaging_prefetch_page(vcpu, sp);
1369         trace_kvm_mmu_get_page(sp, true);
1370         return sp;
1371 }
1372
1373 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1374                              struct kvm_vcpu *vcpu, u64 addr)
1375 {
1376         iterator->addr = addr;
1377         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1378         iterator->level = vcpu->arch.mmu.shadow_root_level;
1379         if (iterator->level == PT32E_ROOT_LEVEL) {
1380                 iterator->shadow_addr
1381                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1382                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1383                 --iterator->level;
1384                 if (!iterator->shadow_addr)
1385                         iterator->level = 0;
1386         }
1387 }
1388
1389 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1390 {
1391         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1392                 return false;
1393
1394         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1395                 if (is_large_pte(*iterator->sptep))
1396                         return false;
1397
1398         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1399         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1400         return true;
1401 }
1402
1403 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1404 {
1405         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1406         --iterator->level;
1407 }
1408
1409 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1410                                          struct kvm_mmu_page *sp)
1411 {
1412         unsigned i;
1413         u64 *pt;
1414         u64 ent;
1415
1416         pt = sp->spt;
1417
1418         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1419                 ent = pt[i];
1420
1421                 if (is_shadow_present_pte(ent)) {
1422                         if (!is_last_spte(ent, sp->role.level)) {
1423                                 ent &= PT64_BASE_ADDR_MASK;
1424                                 mmu_page_remove_parent_pte(page_header(ent),
1425                                                            &pt[i]);
1426                         } else {
1427                                 if (is_large_pte(ent))
1428                                         --kvm->stat.lpages;
1429                                 rmap_remove(kvm, &pt[i]);
1430                         }
1431                 }
1432                 pt[i] = shadow_trap_nonpresent_pte;
1433         }
1434 }
1435
1436 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1437 {
1438         mmu_page_remove_parent_pte(sp, parent_pte);
1439 }
1440
1441 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1442 {
1443         int i;
1444         struct kvm_vcpu *vcpu;
1445
1446         kvm_for_each_vcpu(i, vcpu, kvm)
1447                 vcpu->arch.last_pte_updated = NULL;
1448 }
1449
1450 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1451 {
1452         u64 *parent_pte;
1453
1454         while (sp->multimapped || sp->parent_pte) {
1455                 if (!sp->multimapped)
1456                         parent_pte = sp->parent_pte;
1457                 else {
1458                         struct kvm_pte_chain *chain;
1459
1460                         chain = container_of(sp->parent_ptes.first,
1461                                              struct kvm_pte_chain, link);
1462                         parent_pte = chain->parent_ptes[0];
1463                 }
1464                 BUG_ON(!parent_pte);
1465                 kvm_mmu_put_page(sp, parent_pte);
1466                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1467         }
1468 }
1469
1470 static int mmu_zap_unsync_children(struct kvm *kvm,
1471                                    struct kvm_mmu_page *parent)
1472 {
1473         int i, zapped = 0;
1474         struct mmu_page_path parents;
1475         struct kvm_mmu_pages pages;
1476
1477         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1478                 return 0;
1479
1480         kvm_mmu_pages_init(parent, &parents, &pages);
1481         while (mmu_unsync_walk(parent, &pages)) {
1482                 struct kvm_mmu_page *sp;
1483
1484                 for_each_sp(pages, sp, parents, i) {
1485                         kvm_mmu_zap_page(kvm, sp);
1486                         mmu_pages_clear_parents(&parents);
1487                         zapped++;
1488                 }
1489                 kvm_mmu_pages_init(parent, &parents, &pages);
1490         }
1491
1492         return zapped;
1493 }
1494
1495 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1496 {
1497         int ret;
1498
1499         trace_kvm_mmu_zap_page(sp);
1500         ++kvm->stat.mmu_shadow_zapped;
1501         ret = mmu_zap_unsync_children(kvm, sp);
1502         kvm_mmu_page_unlink_children(kvm, sp);
1503         kvm_mmu_unlink_parents(kvm, sp);
1504         kvm_flush_remote_tlbs(kvm);
1505         if (!sp->role.invalid && !sp->role.direct)
1506                 unaccount_shadowed(kvm, sp->gfn);
1507         if (sp->unsync)
1508                 kvm_unlink_unsync_page(kvm, sp);
1509         if (!sp->root_count) {
1510                 hlist_del(&sp->hash_link);
1511                 kvm_mmu_free_page(kvm, sp);
1512         } else {
1513                 sp->role.invalid = 1;
1514                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1515                 kvm_reload_remote_mmus(kvm);
1516         }
1517         kvm_mmu_reset_last_pte_updated(kvm);
1518         return ret;
1519 }
1520
1521 /*
1522  * Changing the number of mmu pages allocated to the vm
1523  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1524  */
1525 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1526 {
1527         int used_pages;
1528
1529         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1530         used_pages = max(0, used_pages);
1531
1532         /*
1533          * If we set the number of mmu pages to be smaller be than the
1534          * number of actived pages , we must to free some mmu pages before we
1535          * change the value
1536          */
1537
1538         if (used_pages > kvm_nr_mmu_pages) {
1539                 while (used_pages > kvm_nr_mmu_pages &&
1540                         !list_empty(&kvm->arch.active_mmu_pages)) {
1541                         struct kvm_mmu_page *page;
1542
1543                         page = container_of(kvm->arch.active_mmu_pages.prev,
1544                                             struct kvm_mmu_page, link);
1545                         used_pages -= kvm_mmu_zap_page(kvm, page);
1546                         used_pages--;
1547                 }
1548                 kvm_nr_mmu_pages = used_pages;
1549                 kvm->arch.n_free_mmu_pages = 0;
1550         }
1551         else
1552                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1553                                          - kvm->arch.n_alloc_mmu_pages;
1554
1555         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1556 }
1557
1558 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1559 {
1560         unsigned index;
1561         struct hlist_head *bucket;
1562         struct kvm_mmu_page *sp;
1563         struct hlist_node *node, *n;
1564         int r;
1565
1566         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1567         r = 0;
1568         index = kvm_page_table_hashfn(gfn);
1569         bucket = &kvm->arch.mmu_page_hash[index];
1570         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1571                 if (sp->gfn == gfn && !sp->role.direct) {
1572                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1573                                  sp->role.word);
1574                         r = 1;
1575                         if (kvm_mmu_zap_page(kvm, sp))
1576                                 n = bucket->first;
1577                 }
1578         return r;
1579 }
1580
1581 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1582 {
1583         unsigned index;
1584         struct hlist_head *bucket;
1585         struct kvm_mmu_page *sp;
1586         struct hlist_node *node, *nn;
1587
1588         index = kvm_page_table_hashfn(gfn);
1589         bucket = &kvm->arch.mmu_page_hash[index];
1590         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1591                 if (sp->gfn == gfn && !sp->role.direct
1592                     && !sp->role.invalid) {
1593                         pgprintk("%s: zap %lx %x\n",
1594                                  __func__, gfn, sp->role.word);
1595                         if (kvm_mmu_zap_page(kvm, sp))
1596                                 nn = bucket->first;
1597                 }
1598         }
1599 }
1600
1601 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1602 {
1603         int slot = memslot_id(kvm, gfn);
1604         struct kvm_mmu_page *sp = page_header(__pa(pte));
1605
1606         __set_bit(slot, sp->slot_bitmap);
1607 }
1608
1609 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1610 {
1611         int i;
1612         u64 *pt = sp->spt;
1613
1614         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1615                 return;
1616
1617         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1618                 if (pt[i] == shadow_notrap_nonpresent_pte)
1619                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1620         }
1621 }
1622
1623 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1624 {
1625         struct page *page;
1626
1627         gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
1628
1629         if (gpa == UNMAPPED_GVA)
1630                 return NULL;
1631
1632         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1633
1634         return page;
1635 }
1636
1637 /*
1638  * The function is based on mtrr_type_lookup() in
1639  * arch/x86/kernel/cpu/mtrr/generic.c
1640  */
1641 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1642                          u64 start, u64 end)
1643 {
1644         int i;
1645         u64 base, mask;
1646         u8 prev_match, curr_match;
1647         int num_var_ranges = KVM_NR_VAR_MTRR;
1648
1649         if (!mtrr_state->enabled)
1650                 return 0xFF;
1651
1652         /* Make end inclusive end, instead of exclusive */
1653         end--;
1654
1655         /* Look in fixed ranges. Just return the type as per start */
1656         if (mtrr_state->have_fixed && (start < 0x100000)) {
1657                 int idx;
1658
1659                 if (start < 0x80000) {
1660                         idx = 0;
1661                         idx += (start >> 16);
1662                         return mtrr_state->fixed_ranges[idx];
1663                 } else if (start < 0xC0000) {
1664                         idx = 1 * 8;
1665                         idx += ((start - 0x80000) >> 14);
1666                         return mtrr_state->fixed_ranges[idx];
1667                 } else if (start < 0x1000000) {
1668                         idx = 3 * 8;
1669                         idx += ((start - 0xC0000) >> 12);
1670                         return mtrr_state->fixed_ranges[idx];
1671                 }
1672         }
1673
1674         /*
1675          * Look in variable ranges
1676          * Look of multiple ranges matching this address and pick type
1677          * as per MTRR precedence
1678          */
1679         if (!(mtrr_state->enabled & 2))
1680                 return mtrr_state->def_type;
1681
1682         prev_match = 0xFF;
1683         for (i = 0; i < num_var_ranges; ++i) {
1684                 unsigned short start_state, end_state;
1685
1686                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1687                         continue;
1688
1689                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1690                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1691                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1692                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1693
1694                 start_state = ((start & mask) == (base & mask));
1695                 end_state = ((end & mask) == (base & mask));
1696                 if (start_state != end_state)
1697                         return 0xFE;
1698
1699                 if ((start & mask) != (base & mask))
1700                         continue;
1701
1702                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1703                 if (prev_match == 0xFF) {
1704                         prev_match = curr_match;
1705                         continue;
1706                 }
1707
1708                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1709                     curr_match == MTRR_TYPE_UNCACHABLE)
1710                         return MTRR_TYPE_UNCACHABLE;
1711
1712                 if ((prev_match == MTRR_TYPE_WRBACK &&
1713                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1714                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1715                      curr_match == MTRR_TYPE_WRBACK)) {
1716                         prev_match = MTRR_TYPE_WRTHROUGH;
1717                         curr_match = MTRR_TYPE_WRTHROUGH;
1718                 }
1719
1720                 if (prev_match != curr_match)
1721                         return MTRR_TYPE_UNCACHABLE;
1722         }
1723
1724         if (prev_match != 0xFF)
1725                 return prev_match;
1726
1727         return mtrr_state->def_type;
1728 }
1729
1730 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1731 {
1732         u8 mtrr;
1733
1734         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1735                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1736         if (mtrr == 0xfe || mtrr == 0xff)
1737                 mtrr = MTRR_TYPE_WRBACK;
1738         return mtrr;
1739 }
1740 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1741
1742 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1743 {
1744         unsigned index;
1745         struct hlist_head *bucket;
1746         struct kvm_mmu_page *s;
1747         struct hlist_node *node, *n;
1748
1749         trace_kvm_mmu_unsync_page(sp);
1750         index = kvm_page_table_hashfn(sp->gfn);
1751         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1752         /* don't unsync if pagetable is shadowed with multiple roles */
1753         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1754                 if (s->gfn != sp->gfn || s->role.direct)
1755                         continue;
1756                 if (s->role.word != sp->role.word)
1757                         return 1;
1758         }
1759         ++vcpu->kvm->stat.mmu_unsync;
1760         sp->unsync = 1;
1761
1762         kvm_mmu_mark_parents_unsync(vcpu, sp);
1763
1764         mmu_convert_notrap(sp);
1765         return 0;
1766 }
1767
1768 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1769                                   bool can_unsync)
1770 {
1771         struct kvm_mmu_page *shadow;
1772
1773         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1774         if (shadow) {
1775                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1776                         return 1;
1777                 if (shadow->unsync)
1778                         return 0;
1779                 if (can_unsync && oos_shadow)
1780                         return kvm_unsync_page(vcpu, shadow);
1781                 return 1;
1782         }
1783         return 0;
1784 }
1785
1786 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1787                     unsigned pte_access, int user_fault,
1788                     int write_fault, int dirty, int level,
1789                     gfn_t gfn, pfn_t pfn, bool speculative,
1790                     bool can_unsync, bool reset_host_protection)
1791 {
1792         u64 spte;
1793         int ret = 0;
1794
1795         /*
1796          * We don't set the accessed bit, since we sometimes want to see
1797          * whether the guest actually used the pte (in order to detect
1798          * demand paging).
1799          */
1800         spte = shadow_base_present_pte | shadow_dirty_mask;
1801         if (!speculative)
1802                 spte |= shadow_accessed_mask;
1803         if (!dirty)
1804                 pte_access &= ~ACC_WRITE_MASK;
1805         if (pte_access & ACC_EXEC_MASK)
1806                 spte |= shadow_x_mask;
1807         else
1808                 spte |= shadow_nx_mask;
1809         if (pte_access & ACC_USER_MASK)
1810                 spte |= shadow_user_mask;
1811         if (level > PT_PAGE_TABLE_LEVEL)
1812                 spte |= PT_PAGE_SIZE_MASK;
1813         if (tdp_enabled)
1814                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1815                         kvm_is_mmio_pfn(pfn));
1816
1817         if (reset_host_protection)
1818                 spte |= SPTE_HOST_WRITEABLE;
1819
1820         spte |= (u64)pfn << PAGE_SHIFT;
1821
1822         if ((pte_access & ACC_WRITE_MASK)
1823             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1824
1825                 if (level > PT_PAGE_TABLE_LEVEL &&
1826                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1827                         ret = 1;
1828                         spte = shadow_trap_nonpresent_pte;
1829                         goto set_pte;
1830                 }
1831
1832                 spte |= PT_WRITABLE_MASK;
1833
1834                 /*
1835                  * Optimization: for pte sync, if spte was writable the hash
1836                  * lookup is unnecessary (and expensive). Write protection
1837                  * is responsibility of mmu_get_page / kvm_sync_page.
1838                  * Same reasoning can be applied to dirty page accounting.
1839                  */
1840                 if (!can_unsync && is_writable_pte(*sptep))
1841                         goto set_pte;
1842
1843                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1844                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1845                                  __func__, gfn);
1846                         ret = 1;
1847                         pte_access &= ~ACC_WRITE_MASK;
1848                         if (is_writable_pte(spte))
1849                                 spte &= ~PT_WRITABLE_MASK;
1850                 }
1851         }
1852
1853         if (pte_access & ACC_WRITE_MASK)
1854                 mark_page_dirty(vcpu->kvm, gfn);
1855
1856 set_pte:
1857         __set_spte(sptep, spte);
1858         return ret;
1859 }
1860
1861 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1862                          unsigned pt_access, unsigned pte_access,
1863                          int user_fault, int write_fault, int dirty,
1864                          int *ptwrite, int level, gfn_t gfn,
1865                          pfn_t pfn, bool speculative,
1866                          bool reset_host_protection)
1867 {
1868         int was_rmapped = 0;
1869         int was_writable = is_writable_pte(*sptep);
1870         int rmap_count;
1871
1872         pgprintk("%s: spte %llx access %x write_fault %d"
1873                  " user_fault %d gfn %lx\n",
1874                  __func__, *sptep, pt_access,
1875                  write_fault, user_fault, gfn);
1876
1877         if (is_rmap_spte(*sptep)) {
1878                 /*
1879                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1880                  * the parent of the now unreachable PTE.
1881                  */
1882                 if (level > PT_PAGE_TABLE_LEVEL &&
1883                     !is_large_pte(*sptep)) {
1884                         struct kvm_mmu_page *child;
1885                         u64 pte = *sptep;
1886
1887                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1888                         mmu_page_remove_parent_pte(child, sptep);
1889                 } else if (pfn != spte_to_pfn(*sptep)) {
1890                         pgprintk("hfn old %lx new %lx\n",
1891                                  spte_to_pfn(*sptep), pfn);
1892                         rmap_remove(vcpu->kvm, sptep);
1893                 } else
1894                         was_rmapped = 1;
1895         }
1896
1897         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1898                       dirty, level, gfn, pfn, speculative, true,
1899                       reset_host_protection)) {
1900                 if (write_fault)
1901                         *ptwrite = 1;
1902                 kvm_x86_ops->tlb_flush(vcpu);
1903         }
1904
1905         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1906         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1907                  is_large_pte(*sptep)? "2MB" : "4kB",
1908                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1909                  *sptep, sptep);
1910         if (!was_rmapped && is_large_pte(*sptep))
1911                 ++vcpu->kvm->stat.lpages;
1912
1913         page_header_update_slot(vcpu->kvm, sptep, gfn);
1914         if (!was_rmapped) {
1915                 rmap_count = rmap_add(vcpu, sptep, gfn);
1916                 kvm_release_pfn_clean(pfn);
1917                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1918                         rmap_recycle(vcpu, sptep, gfn);
1919         } else {
1920                 if (was_writable)
1921                         kvm_release_pfn_dirty(pfn);
1922                 else
1923                         kvm_release_pfn_clean(pfn);
1924         }
1925         if (speculative) {
1926                 vcpu->arch.last_pte_updated = sptep;
1927                 vcpu->arch.last_pte_gfn = gfn;
1928         }
1929 }
1930
1931 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1932 {
1933 }
1934
1935 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1936                         int level, gfn_t gfn, pfn_t pfn)
1937 {
1938         struct kvm_shadow_walk_iterator iterator;
1939         struct kvm_mmu_page *sp;
1940         int pt_write = 0;
1941         gfn_t pseudo_gfn;
1942
1943         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1944                 if (iterator.level == level) {
1945                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1946                                      0, write, 1, &pt_write,
1947                                      level, gfn, pfn, false, true);
1948                         ++vcpu->stat.pf_fixed;
1949                         break;
1950                 }
1951
1952                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1953                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1954                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1955                                               iterator.level - 1,
1956                                               1, ACC_ALL, iterator.sptep);
1957                         if (!sp) {
1958                                 pgprintk("nonpaging_map: ENOMEM\n");
1959                                 kvm_release_pfn_clean(pfn);
1960                                 return -ENOMEM;
1961                         }
1962
1963                         __set_spte(iterator.sptep,
1964                                    __pa(sp->spt)
1965                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1966                                    | shadow_user_mask | shadow_x_mask);
1967                 }
1968         }
1969         return pt_write;
1970 }
1971
1972 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1973 {
1974         int r;
1975         int level;
1976         pfn_t pfn;
1977         unsigned long mmu_seq;
1978
1979         level = mapping_level(vcpu, gfn);
1980
1981         /*
1982          * This path builds a PAE pagetable - so we can map 2mb pages at
1983          * maximum. Therefore check if the level is larger than that.
1984          */
1985         if (level > PT_DIRECTORY_LEVEL)
1986                 level = PT_DIRECTORY_LEVEL;
1987
1988         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
1989
1990         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1991         smp_rmb();
1992         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1993
1994         /* mmio */
1995         if (is_error_pfn(pfn)) {
1996                 kvm_release_pfn_clean(pfn);
1997                 return 1;
1998         }
1999
2000         spin_lock(&vcpu->kvm->mmu_lock);
2001         if (mmu_notifier_retry(vcpu, mmu_seq))
2002                 goto out_unlock;
2003         kvm_mmu_free_some_pages(vcpu);
2004         r = __direct_map(vcpu, v, write, level, gfn, pfn);
2005         spin_unlock(&vcpu->kvm->mmu_lock);
2006
2007
2008         return r;
2009
2010 out_unlock:
2011         spin_unlock(&vcpu->kvm->mmu_lock);
2012         kvm_release_pfn_clean(pfn);
2013         return 0;
2014 }
2015
2016
2017 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2018 {
2019         int i;
2020         struct kvm_mmu_page *sp;
2021
2022         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2023                 return;
2024         spin_lock(&vcpu->kvm->mmu_lock);
2025         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2026                 hpa_t root = vcpu->arch.mmu.root_hpa;
2027
2028                 sp = page_header(root);
2029                 --sp->root_count;
2030                 if (!sp->root_count && sp->role.invalid)
2031                         kvm_mmu_zap_page(vcpu->kvm, sp);
2032                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2033                 spin_unlock(&vcpu->kvm->mmu_lock);
2034                 return;
2035         }
2036         for (i = 0; i < 4; ++i) {
2037                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2038
2039                 if (root) {
2040                         root &= PT64_BASE_ADDR_MASK;
2041                         sp = page_header(root);
2042                         --sp->root_count;
2043                         if (!sp->root_count && sp->role.invalid)
2044                                 kvm_mmu_zap_page(vcpu->kvm, sp);
2045                 }
2046                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2047         }
2048         spin_unlock(&vcpu->kvm->mmu_lock);
2049         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2050 }
2051
2052 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2053 {
2054         int ret = 0;
2055
2056         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2057                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2058                 ret = 1;
2059         }
2060
2061         return ret;
2062 }
2063
2064 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2065 {
2066         int i;
2067         gfn_t root_gfn;
2068         struct kvm_mmu_page *sp;
2069         int direct = 0;
2070         u64 pdptr;
2071
2072         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2073
2074         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2075                 hpa_t root = vcpu->arch.mmu.root_hpa;
2076
2077                 ASSERT(!VALID_PAGE(root));
2078                 if (tdp_enabled)
2079                         direct = 1;
2080                 if (mmu_check_root(vcpu, root_gfn))
2081                         return 1;
2082                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2083                                       PT64_ROOT_LEVEL, direct,
2084                                       ACC_ALL, NULL);
2085                 root = __pa(sp->spt);
2086                 ++sp->root_count;
2087                 vcpu->arch.mmu.root_hpa = root;
2088                 return 0;
2089         }
2090         direct = !is_paging(vcpu);
2091         if (tdp_enabled)
2092                 direct = 1;
2093         for (i = 0; i < 4; ++i) {
2094                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2095
2096                 ASSERT(!VALID_PAGE(root));
2097                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2098                         pdptr = kvm_pdptr_read(vcpu, i);
2099                         if (!is_present_gpte(pdptr)) {
2100                                 vcpu->arch.mmu.pae_root[i] = 0;
2101                                 continue;
2102                         }
2103                         root_gfn = pdptr >> PAGE_SHIFT;
2104                 } else if (vcpu->arch.mmu.root_level == 0)
2105                         root_gfn = 0;
2106                 if (mmu_check_root(vcpu, root_gfn))
2107                         return 1;
2108                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2109                                       PT32_ROOT_LEVEL, direct,
2110                                       ACC_ALL, NULL);
2111                 root = __pa(sp->spt);
2112                 ++sp->root_count;
2113                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2114         }
2115         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2116         return 0;
2117 }
2118
2119 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2120 {
2121         int i;
2122         struct kvm_mmu_page *sp;
2123
2124         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2125                 return;
2126         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2127                 hpa_t root = vcpu->arch.mmu.root_hpa;
2128                 sp = page_header(root);
2129                 mmu_sync_children(vcpu, sp);
2130                 return;
2131         }
2132         for (i = 0; i < 4; ++i) {
2133                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2134
2135                 if (root && VALID_PAGE(root)) {
2136                         root &= PT64_BASE_ADDR_MASK;
2137                         sp = page_header(root);
2138                         mmu_sync_children(vcpu, sp);
2139                 }
2140         }
2141 }
2142
2143 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2144 {
2145         spin_lock(&vcpu->kvm->mmu_lock);
2146         mmu_sync_roots(vcpu);
2147         spin_unlock(&vcpu->kvm->mmu_lock);
2148 }
2149
2150 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2151                                   u32 access, u32 *error)
2152 {
2153         if (error)
2154                 *error = 0;
2155         return vaddr;
2156 }
2157
2158 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2159                                 u32 error_code)
2160 {
2161         gfn_t gfn;
2162         int r;
2163
2164         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2165         r = mmu_topup_memory_caches(vcpu);
2166         if (r)
2167                 return r;
2168
2169         ASSERT(vcpu);
2170         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2171
2172         gfn = gva >> PAGE_SHIFT;
2173
2174         return nonpaging_map(vcpu, gva & PAGE_MASK,
2175                              error_code & PFERR_WRITE_MASK, gfn);
2176 }
2177
2178 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2179                                 u32 error_code)
2180 {
2181         pfn_t pfn;
2182         int r;
2183         int level;
2184         gfn_t gfn = gpa >> PAGE_SHIFT;
2185         unsigned long mmu_seq;
2186
2187         ASSERT(vcpu);
2188         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2189
2190         r = mmu_topup_memory_caches(vcpu);
2191         if (r)
2192                 return r;
2193
2194         level = mapping_level(vcpu, gfn);
2195
2196         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2197
2198         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2199         smp_rmb();
2200         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2201         if (is_error_pfn(pfn)) {
2202                 kvm_release_pfn_clean(pfn);
2203                 return 1;
2204         }
2205         spin_lock(&vcpu->kvm->mmu_lock);
2206         if (mmu_notifier_retry(vcpu, mmu_seq))
2207                 goto out_unlock;
2208         kvm_mmu_free_some_pages(vcpu);
2209         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2210                          level, gfn, pfn);
2211         spin_unlock(&vcpu->kvm->mmu_lock);
2212
2213         return r;
2214
2215 out_unlock:
2216         spin_unlock(&vcpu->kvm->mmu_lock);
2217         kvm_release_pfn_clean(pfn);
2218         return 0;
2219 }
2220
2221 static void nonpaging_free(struct kvm_vcpu *vcpu)
2222 {
2223         mmu_free_roots(vcpu);
2224 }
2225
2226 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2227 {
2228         struct kvm_mmu *context = &vcpu->arch.mmu;
2229
2230         context->new_cr3 = nonpaging_new_cr3;
2231         context->page_fault = nonpaging_page_fault;
2232         context->gva_to_gpa = nonpaging_gva_to_gpa;
2233         context->free = nonpaging_free;
2234         context->prefetch_page = nonpaging_prefetch_page;
2235         context->sync_page = nonpaging_sync_page;
2236         context->invlpg = nonpaging_invlpg;
2237         context->root_level = 0;
2238         context->shadow_root_level = PT32E_ROOT_LEVEL;
2239         context->root_hpa = INVALID_PAGE;
2240         return 0;
2241 }
2242
2243 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2244 {
2245         ++vcpu->stat.tlb_flush;
2246         kvm_x86_ops->tlb_flush(vcpu);
2247 }
2248
2249 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2250 {
2251         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2252         mmu_free_roots(vcpu);
2253 }
2254
2255 static void inject_page_fault(struct kvm_vcpu *vcpu,
2256                               u64 addr,
2257                               u32 err_code)
2258 {
2259         kvm_inject_page_fault(vcpu, addr, err_code);
2260 }
2261
2262 static void paging_free(struct kvm_vcpu *vcpu)
2263 {
2264         nonpaging_free(vcpu);
2265 }
2266
2267 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2268 {
2269         int bit7;
2270
2271         bit7 = (gpte >> 7) & 1;
2272         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2273 }
2274
2275 #define PTTYPE 64
2276 #include "paging_tmpl.h"
2277 #undef PTTYPE
2278
2279 #define PTTYPE 32
2280 #include "paging_tmpl.h"
2281 #undef PTTYPE
2282
2283 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2284 {
2285         struct kvm_mmu *context = &vcpu->arch.mmu;
2286         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2287         u64 exb_bit_rsvd = 0;
2288
2289         if (!is_nx(vcpu))
2290                 exb_bit_rsvd = rsvd_bits(63, 63);
2291         switch (level) {
2292         case PT32_ROOT_LEVEL:
2293                 /* no rsvd bits for 2 level 4K page table entries */
2294                 context->rsvd_bits_mask[0][1] = 0;
2295                 context->rsvd_bits_mask[0][0] = 0;
2296                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2297
2298                 if (!is_pse(vcpu)) {
2299                         context->rsvd_bits_mask[1][1] = 0;
2300                         break;
2301                 }
2302
2303                 if (is_cpuid_PSE36())
2304                         /* 36bits PSE 4MB page */
2305                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2306                 else
2307                         /* 32 bits PSE 4MB page */
2308                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2309                 break;
2310         case PT32E_ROOT_LEVEL:
2311                 context->rsvd_bits_mask[0][2] =
2312                         rsvd_bits(maxphyaddr, 63) |
2313                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2314                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2315                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2316                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2317                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2318                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2319                         rsvd_bits(maxphyaddr, 62) |
2320                         rsvd_bits(13, 20);              /* large page */
2321                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2322                 break;
2323         case PT64_ROOT_LEVEL:
2324                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2325                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2326                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2327                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2328                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2329                         rsvd_bits(maxphyaddr, 51);
2330                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2331                         rsvd_bits(maxphyaddr, 51);
2332                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2333                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2334                         rsvd_bits(maxphyaddr, 51) |
2335                         rsvd_bits(13, 29);
2336                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2337                         rsvd_bits(maxphyaddr, 51) |
2338                         rsvd_bits(13, 20);              /* large page */
2339                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2340                 break;
2341         }
2342 }
2343
2344 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2345 {
2346         struct kvm_mmu *context = &vcpu->arch.mmu;
2347
2348         ASSERT(is_pae(vcpu));
2349         context->new_cr3 = paging_new_cr3;
2350         context->page_fault = paging64_page_fault;
2351         context->gva_to_gpa = paging64_gva_to_gpa;
2352         context->prefetch_page = paging64_prefetch_page;
2353         context->sync_page = paging64_sync_page;
2354         context->invlpg = paging64_invlpg;
2355         context->free = paging_free;
2356         context->root_level = level;
2357         context->shadow_root_level = level;
2358         context->root_hpa = INVALID_PAGE;
2359         return 0;
2360 }
2361
2362 static int paging64_init_context(struct kvm_vcpu *vcpu)
2363 {
2364         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2365         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2366 }
2367
2368 static int paging32_init_context(struct kvm_vcpu *vcpu)
2369 {
2370         struct kvm_mmu *context = &vcpu->arch.mmu;
2371
2372         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2373         context->new_cr3 = paging_new_cr3;
2374         context->page_fault = paging32_page_fault;
2375         context->gva_to_gpa = paging32_gva_to_gpa;
2376         context->free = paging_free;
2377         context->prefetch_page = paging32_prefetch_page;
2378         context->sync_page = paging32_sync_page;
2379         context->invlpg = paging32_invlpg;
2380         context->root_level = PT32_ROOT_LEVEL;
2381         context->shadow_root_level = PT32E_ROOT_LEVEL;
2382         context->root_hpa = INVALID_PAGE;
2383         return 0;
2384 }
2385
2386 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2387 {
2388         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2389         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2390 }
2391
2392 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2393 {
2394         struct kvm_mmu *context = &vcpu->arch.mmu;
2395
2396         context->new_cr3 = nonpaging_new_cr3;
2397         context->page_fault = tdp_page_fault;
2398         context->free = nonpaging_free;
2399         context->prefetch_page = nonpaging_prefetch_page;
2400         context->sync_page = nonpaging_sync_page;
2401         context->invlpg = nonpaging_invlpg;
2402         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2403         context->root_hpa = INVALID_PAGE;
2404
2405         if (!is_paging(vcpu)) {
2406                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2407                 context->root_level = 0;
2408         } else if (is_long_mode(vcpu)) {
2409                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2410                 context->gva_to_gpa = paging64_gva_to_gpa;
2411                 context->root_level = PT64_ROOT_LEVEL;
2412         } else if (is_pae(vcpu)) {
2413                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2414                 context->gva_to_gpa = paging64_gva_to_gpa;
2415                 context->root_level = PT32E_ROOT_LEVEL;
2416         } else {
2417                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2418                 context->gva_to_gpa = paging32_gva_to_gpa;
2419                 context->root_level = PT32_ROOT_LEVEL;
2420         }
2421
2422         return 0;
2423 }
2424
2425 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2426 {
2427         int r;
2428
2429         ASSERT(vcpu);
2430         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2431
2432         if (!is_paging(vcpu))
2433                 r = nonpaging_init_context(vcpu);
2434         else if (is_long_mode(vcpu))
2435                 r = paging64_init_context(vcpu);
2436         else if (is_pae(vcpu))
2437                 r = paging32E_init_context(vcpu);
2438         else
2439                 r = paging32_init_context(vcpu);
2440
2441         vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
2442
2443         return r;
2444 }
2445
2446 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2447 {
2448         vcpu->arch.update_pte.pfn = bad_pfn;
2449
2450         if (tdp_enabled)
2451                 return init_kvm_tdp_mmu(vcpu);
2452         else
2453                 return init_kvm_softmmu(vcpu);
2454 }
2455
2456 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2457 {
2458         ASSERT(vcpu);
2459         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2460                 vcpu->arch.mmu.free(vcpu);
2461                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2462         }
2463 }
2464
2465 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2466 {
2467         destroy_kvm_mmu(vcpu);
2468         return init_kvm_mmu(vcpu);
2469 }
2470 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2471
2472 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2473 {
2474         int r;
2475
2476         r = mmu_topup_memory_caches(vcpu);
2477         if (r)
2478                 goto out;
2479         spin_lock(&vcpu->kvm->mmu_lock);
2480         kvm_mmu_free_some_pages(vcpu);
2481         r = mmu_alloc_roots(vcpu);
2482         mmu_sync_roots(vcpu);
2483         spin_unlock(&vcpu->kvm->mmu_lock);
2484         if (r)
2485                 goto out;
2486         /* set_cr3() should ensure TLB has been flushed */
2487         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2488 out:
2489         return r;
2490 }
2491 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2492
2493 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2494 {
2495         mmu_free_roots(vcpu);
2496 }
2497
2498 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2499                                   struct kvm_mmu_page *sp,
2500                                   u64 *spte)
2501 {
2502         u64 pte;
2503         struct kvm_mmu_page *child;
2504
2505         pte = *spte;
2506         if (is_shadow_present_pte(pte)) {
2507                 if (is_last_spte(pte, sp->role.level))
2508                         rmap_remove(vcpu->kvm, spte);
2509                 else {
2510                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2511                         mmu_page_remove_parent_pte(child, spte);
2512                 }
2513         }
2514         __set_spte(spte, shadow_trap_nonpresent_pte);
2515         if (is_large_pte(pte))
2516                 --vcpu->kvm->stat.lpages;
2517 }
2518
2519 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2520                                   struct kvm_mmu_page *sp,
2521                                   u64 *spte,
2522                                   const void *new)
2523 {
2524         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2525                 ++vcpu->kvm->stat.mmu_pde_zapped;
2526                 return;
2527         }
2528
2529         ++vcpu->kvm->stat.mmu_pte_updated;
2530         if (!sp->role.cr4_pae)
2531                 paging32_update_pte(vcpu, sp, spte, new);
2532         else
2533                 paging64_update_pte(vcpu, sp, spte, new);
2534 }
2535
2536 static bool need_remote_flush(u64 old, u64 new)
2537 {
2538         if (!is_shadow_present_pte(old))
2539                 return false;
2540         if (!is_shadow_present_pte(new))
2541                 return true;
2542         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2543                 return true;
2544         old ^= PT64_NX_MASK;
2545         new ^= PT64_NX_MASK;
2546         return (old & ~new & PT64_PERM_MASK) != 0;
2547 }
2548
2549 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2550 {
2551         if (need_remote_flush(old, new))
2552                 kvm_flush_remote_tlbs(vcpu->kvm);
2553         else
2554                 kvm_mmu_flush_tlb(vcpu);
2555 }
2556
2557 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2558 {
2559         u64 *spte = vcpu->arch.last_pte_updated;
2560
2561         return !!(spte && (*spte & shadow_accessed_mask));
2562 }
2563
2564 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2565                                           u64 gpte)
2566 {
2567         gfn_t gfn;
2568         pfn_t pfn;
2569
2570         if (!is_present_gpte(gpte))
2571                 return;
2572         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2573
2574         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2575         smp_rmb();
2576         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2577
2578         if (is_error_pfn(pfn)) {
2579                 kvm_release_pfn_clean(pfn);
2580                 return;
2581         }
2582         vcpu->arch.update_pte.gfn = gfn;
2583         vcpu->arch.update_pte.pfn = pfn;
2584 }
2585
2586 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2587 {
2588         u64 *spte = vcpu->arch.last_pte_updated;
2589
2590         if (spte
2591             && vcpu->arch.last_pte_gfn == gfn
2592             && shadow_accessed_mask
2593             && !(*spte & shadow_accessed_mask)
2594             && is_shadow_present_pte(*spte))
2595                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2596 }
2597
2598 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2599                        const u8 *new, int bytes,
2600                        bool guest_initiated)
2601 {
2602         gfn_t gfn = gpa >> PAGE_SHIFT;
2603         struct kvm_mmu_page *sp;
2604         struct hlist_node *node, *n;
2605         struct hlist_head *bucket;
2606         unsigned index;
2607         u64 entry, gentry;
2608         u64 *spte;
2609         unsigned offset = offset_in_page(gpa);
2610         unsigned pte_size;
2611         unsigned page_offset;
2612         unsigned misaligned;
2613         unsigned quadrant;
2614         int level;
2615         int flooded = 0;
2616         int npte;
2617         int r;
2618         int invlpg_counter;
2619
2620         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2621
2622         invlpg_counter = atomic_read(&vcpu->kvm->arch.invlpg_counter);
2623
2624         /*
2625          * Assume that the pte write on a page table of the same type
2626          * as the current vcpu paging mode.  This is nearly always true
2627          * (might be false while changing modes).  Note it is verified later
2628          * by update_pte().
2629          */
2630         if ((is_pae(vcpu) && bytes == 4) || !new) {
2631                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2632                 if (is_pae(vcpu)) {
2633                         gpa &= ~(gpa_t)7;
2634                         bytes = 8;
2635                 }
2636                 r = kvm_read_guest(vcpu->kvm, gpa, &gentry, min(bytes, 8));
2637                 if (r)
2638                         gentry = 0;
2639                 new = (const u8 *)&gentry;
2640         }
2641
2642         switch (bytes) {
2643         case 4:
2644                 gentry = *(const u32 *)new;
2645                 break;
2646         case 8:
2647                 gentry = *(const u64 *)new;
2648                 break;
2649         default:
2650                 gentry = 0;
2651                 break;
2652         }
2653
2654         mmu_guess_page_from_pte_write(vcpu, gpa, gentry);
2655         spin_lock(&vcpu->kvm->mmu_lock);
2656         if (atomic_read(&vcpu->kvm->arch.invlpg_counter) != invlpg_counter)
2657                 gentry = 0;
2658         kvm_mmu_access_page(vcpu, gfn);
2659         kvm_mmu_free_some_pages(vcpu);
2660         ++vcpu->kvm->stat.mmu_pte_write;
2661         kvm_mmu_audit(vcpu, "pre pte write");
2662         if (guest_initiated) {
2663                 if (gfn == vcpu->arch.last_pt_write_gfn
2664                     && !last_updated_pte_accessed(vcpu)) {
2665                         ++vcpu->arch.last_pt_write_count;
2666                         if (vcpu->arch.last_pt_write_count >= 3)
2667                                 flooded = 1;
2668                 } else {
2669                         vcpu->arch.last_pt_write_gfn = gfn;
2670                         vcpu->arch.last_pt_write_count = 1;
2671                         vcpu->arch.last_pte_updated = NULL;
2672                 }
2673         }
2674         index = kvm_page_table_hashfn(gfn);
2675         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2676         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2677                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2678                         continue;
2679                 pte_size = sp->role.cr4_pae ? 8 : 4;
2680                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2681                 misaligned |= bytes < 4;
2682                 if (misaligned || flooded) {
2683                         /*
2684                          * Misaligned accesses are too much trouble to fix
2685                          * up; also, they usually indicate a page is not used
2686                          * as a page table.
2687                          *
2688                          * If we're seeing too many writes to a page,
2689                          * it may no longer be a page table, or we may be
2690                          * forking, in which case it is better to unmap the
2691                          * page.
2692                          */
2693                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2694                                  gpa, bytes, sp->role.word);
2695                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2696                                 n = bucket->first;
2697                         ++vcpu->kvm->stat.mmu_flooded;
2698                         continue;
2699                 }
2700                 page_offset = offset;
2701                 level = sp->role.level;
2702                 npte = 1;
2703                 if (!sp->role.cr4_pae) {
2704                         page_offset <<= 1;      /* 32->64 */
2705                         /*
2706                          * A 32-bit pde maps 4MB while the shadow pdes map
2707                          * only 2MB.  So we need to double the offset again
2708                          * and zap two pdes instead of one.
2709                          */
2710                         if (level == PT32_ROOT_LEVEL) {
2711                                 page_offset &= ~7; /* kill rounding error */
2712                                 page_offset <<= 1;
2713                                 npte = 2;
2714                         }
2715                         quadrant = page_offset >> PAGE_SHIFT;
2716                         page_offset &= ~PAGE_MASK;
2717                         if (quadrant != sp->role.quadrant)
2718                                 continue;
2719                 }
2720                 spte = &sp->spt[page_offset / sizeof(*spte)];
2721                 while (npte--) {
2722                         entry = *spte;
2723                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2724                         if (gentry)
2725                                 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
2726                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2727                         ++spte;
2728                 }
2729         }
2730         kvm_mmu_audit(vcpu, "post pte write");
2731         spin_unlock(&vcpu->kvm->mmu_lock);
2732         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2733                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2734                 vcpu->arch.update_pte.pfn = bad_pfn;
2735         }
2736 }
2737
2738 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2739 {
2740         gpa_t gpa;
2741         int r;
2742
2743         if (tdp_enabled)
2744                 return 0;
2745
2746         gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2747
2748         spin_lock(&vcpu->kvm->mmu_lock);
2749         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2750         spin_unlock(&vcpu->kvm->mmu_lock);
2751         return r;
2752 }
2753 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2754
2755 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2756 {
2757         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2758                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2759                 struct kvm_mmu_page *sp;
2760
2761                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2762                                   struct kvm_mmu_page, link);
2763                 kvm_mmu_zap_page(vcpu->kvm, sp);
2764                 ++vcpu->kvm->stat.mmu_recycled;
2765         }
2766 }
2767
2768 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2769 {
2770         int r;
2771         enum emulation_result er;
2772
2773         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2774         if (r < 0)
2775                 goto out;
2776
2777         if (!r) {
2778                 r = 1;
2779                 goto out;
2780         }
2781
2782         r = mmu_topup_memory_caches(vcpu);
2783         if (r)
2784                 goto out;
2785
2786         er = emulate_instruction(vcpu, cr2, error_code, 0);
2787
2788         switch (er) {
2789         case EMULATE_DONE:
2790                 return 1;
2791         case EMULATE_DO_MMIO:
2792                 ++vcpu->stat.mmio_exits;
2793                 return 0;
2794         case EMULATE_FAIL:
2795                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2796                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2797                 vcpu->run->internal.ndata = 0;
2798                 return 0;
2799         default:
2800                 BUG();
2801         }
2802 out:
2803         return r;
2804 }
2805 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2806
2807 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2808 {
2809         vcpu->arch.mmu.invlpg(vcpu, gva);
2810         kvm_mmu_flush_tlb(vcpu);
2811         ++vcpu->stat.invlpg;
2812 }
2813 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2814
2815 void kvm_enable_tdp(void)
2816 {
2817         tdp_enabled = true;
2818 }
2819 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2820
2821 void kvm_disable_tdp(void)
2822 {
2823         tdp_enabled = false;
2824 }
2825 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2826
2827 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2828 {
2829         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2830 }
2831
2832 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2833 {
2834         struct page *page;
2835         int i;
2836
2837         ASSERT(vcpu);
2838
2839         /*
2840          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2841          * Therefore we need to allocate shadow page tables in the first
2842          * 4GB of memory, which happens to fit the DMA32 zone.
2843          */
2844         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2845         if (!page)
2846                 return -ENOMEM;
2847
2848         vcpu->arch.mmu.pae_root = page_address(page);
2849         for (i = 0; i < 4; ++i)
2850                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2851
2852         return 0;
2853 }
2854
2855 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2856 {
2857         ASSERT(vcpu);
2858         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2859
2860         return alloc_mmu_pages(vcpu);
2861 }
2862
2863 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2864 {
2865         ASSERT(vcpu);
2866         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2867
2868         return init_kvm_mmu(vcpu);
2869 }
2870
2871 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2872 {
2873         ASSERT(vcpu);
2874
2875         destroy_kvm_mmu(vcpu);
2876         free_mmu_pages(vcpu);
2877         mmu_free_memory_caches(vcpu);
2878 }
2879
2880 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2881 {
2882         struct kvm_mmu_page *sp;
2883
2884         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2885                 int i;
2886                 u64 *pt;
2887
2888                 if (!test_bit(slot, sp->slot_bitmap))
2889                         continue;
2890
2891                 pt = sp->spt;
2892                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2893                         /* avoid RMW */
2894                         if (pt[i] & PT_WRITABLE_MASK)
2895                                 pt[i] &= ~PT_WRITABLE_MASK;
2896         }
2897         kvm_flush_remote_tlbs(kvm);
2898 }
2899
2900 void kvm_mmu_zap_all(struct kvm *kvm)
2901 {
2902         struct kvm_mmu_page *sp, *node;
2903
2904         spin_lock(&kvm->mmu_lock);
2905         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2906                 if (kvm_mmu_zap_page(kvm, sp))
2907                         node = container_of(kvm->arch.active_mmu_pages.next,
2908                                             struct kvm_mmu_page, link);
2909         spin_unlock(&kvm->mmu_lock);
2910
2911         kvm_flush_remote_tlbs(kvm);
2912 }
2913
2914 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2915 {
2916         struct kvm_mmu_page *page;
2917
2918         page = container_of(kvm->arch.active_mmu_pages.prev,
2919                             struct kvm_mmu_page, link);
2920         kvm_mmu_zap_page(kvm, page);
2921 }
2922
2923 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2924 {
2925         struct kvm *kvm;
2926         struct kvm *kvm_freed = NULL;
2927         int cache_count = 0;
2928
2929         spin_lock(&kvm_lock);
2930
2931         list_for_each_entry(kvm, &vm_list, vm_list) {
2932                 int npages, idx;
2933
2934                 idx = srcu_read_lock(&kvm->srcu);
2935                 spin_lock(&kvm->mmu_lock);
2936                 npages = kvm->arch.n_alloc_mmu_pages -
2937                          kvm->arch.n_free_mmu_pages;
2938                 cache_count += npages;
2939                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2940                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2941                         cache_count--;
2942                         kvm_freed = kvm;
2943                 }
2944                 nr_to_scan--;
2945
2946                 spin_unlock(&kvm->mmu_lock);
2947                 srcu_read_unlock(&kvm->srcu, idx);
2948         }
2949         if (kvm_freed)
2950                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2951
2952         spin_unlock(&kvm_lock);
2953
2954         return cache_count;
2955 }
2956
2957 static struct shrinker mmu_shrinker = {
2958         .shrink = mmu_shrink,
2959         .seeks = DEFAULT_SEEKS * 10,
2960 };
2961
2962 static void mmu_destroy_caches(void)
2963 {
2964         if (pte_chain_cache)
2965                 kmem_cache_destroy(pte_chain_cache);
2966         if (rmap_desc_cache)
2967                 kmem_cache_destroy(rmap_desc_cache);
2968         if (mmu_page_header_cache)
2969                 kmem_cache_destroy(mmu_page_header_cache);
2970 }
2971
2972 void kvm_mmu_module_exit(void)
2973 {
2974         mmu_destroy_caches();
2975         unregister_shrinker(&mmu_shrinker);
2976 }
2977
2978 int kvm_mmu_module_init(void)
2979 {
2980         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2981                                             sizeof(struct kvm_pte_chain),
2982                                             0, 0, NULL);
2983         if (!pte_chain_cache)
2984                 goto nomem;
2985         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2986                                             sizeof(struct kvm_rmap_desc),
2987                                             0, 0, NULL);
2988         if (!rmap_desc_cache)
2989                 goto nomem;
2990
2991         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2992                                                   sizeof(struct kvm_mmu_page),
2993                                                   0, 0, NULL);
2994         if (!mmu_page_header_cache)
2995                 goto nomem;
2996
2997         register_shrinker(&mmu_shrinker);
2998
2999         return 0;
3000
3001 nomem:
3002         mmu_destroy_caches();
3003         return -ENOMEM;
3004 }
3005
3006 /*
3007  * Caculate mmu pages needed for kvm.
3008  */
3009 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3010 {
3011         int i;
3012         unsigned int nr_mmu_pages;
3013         unsigned int  nr_pages = 0;
3014         struct kvm_memslots *slots;
3015
3016         slots = rcu_dereference(kvm->memslots);
3017         for (i = 0; i < slots->nmemslots; i++)
3018                 nr_pages += slots->memslots[i].npages;
3019
3020         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3021         nr_mmu_pages = max(nr_mmu_pages,
3022                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3023
3024         return nr_mmu_pages;
3025 }
3026
3027 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3028                                 unsigned len)
3029 {
3030         if (len > buffer->len)
3031                 return NULL;
3032         return buffer->ptr;
3033 }
3034
3035 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3036                                 unsigned len)
3037 {
3038         void *ret;
3039
3040         ret = pv_mmu_peek_buffer(buffer, len);
3041         if (!ret)
3042                 return ret;
3043         buffer->ptr += len;
3044         buffer->len -= len;
3045         buffer->processed += len;
3046         return ret;
3047 }
3048
3049 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3050                              gpa_t addr, gpa_t value)
3051 {
3052         int bytes = 8;
3053         int r;
3054
3055         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3056                 bytes = 4;
3057
3058         r = mmu_topup_memory_caches(vcpu);
3059         if (r)
3060                 return r;
3061
3062         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3063                 return -EFAULT;
3064
3065         return 1;
3066 }
3067
3068 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3069 {
3070         kvm_set_cr3(vcpu, vcpu->arch.cr3);
3071         return 1;
3072 }
3073
3074 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3075 {
3076         spin_lock(&vcpu->kvm->mmu_lock);
3077         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3078         spin_unlock(&vcpu->kvm->mmu_lock);
3079         return 1;
3080 }
3081
3082 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3083                              struct kvm_pv_mmu_op_buffer *buffer)
3084 {
3085         struct kvm_mmu_op_header *header;
3086
3087         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3088         if (!header)
3089                 return 0;
3090         switch (header->op) {
3091         case KVM_MMU_OP_WRITE_PTE: {
3092                 struct kvm_mmu_op_write_pte *wpte;
3093
3094                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3095                 if (!wpte)
3096                         return 0;
3097                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3098                                         wpte->pte_val);
3099         }
3100         case KVM_MMU_OP_FLUSH_TLB: {
3101                 struct kvm_mmu_op_flush_tlb *ftlb;
3102
3103                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3104                 if (!ftlb)
3105                         return 0;
3106                 return kvm_pv_mmu_flush_tlb(vcpu);
3107         }
3108         case KVM_MMU_OP_RELEASE_PT: {
3109                 struct kvm_mmu_op_release_pt *rpt;
3110
3111                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3112                 if (!rpt)
3113                         return 0;
3114                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3115         }
3116         default: return 0;
3117         }
3118 }
3119
3120 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3121                   gpa_t addr, unsigned long *ret)
3122 {
3123         int r;
3124         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3125
3126         buffer->ptr = buffer->buf;
3127         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3128         buffer->processed = 0;
3129
3130         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3131         if (r)
3132                 goto out;
3133
3134         while (buffer->len) {
3135                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3136                 if (r < 0)
3137                         goto out;
3138                 if (r == 0)
3139                         break;
3140         }
3141
3142         r = 1;
3143 out:
3144         *ret = buffer->processed;
3145         return r;
3146 }
3147
3148 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3149 {
3150         struct kvm_shadow_walk_iterator iterator;
3151         int nr_sptes = 0;
3152
3153         spin_lock(&vcpu->kvm->mmu_lock);
3154         for_each_shadow_entry(vcpu, addr, iterator) {
3155                 sptes[iterator.level-1] = *iterator.sptep;
3156                 nr_sptes++;
3157                 if (!is_shadow_present_pte(*iterator.sptep))
3158                         break;
3159         }
3160         spin_unlock(&vcpu->kvm->mmu_lock);
3161
3162         return nr_sptes;
3163 }
3164 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3165
3166 #ifdef AUDIT
3167
3168 static const char *audit_msg;
3169
3170 static gva_t canonicalize(gva_t gva)
3171 {
3172 #ifdef CONFIG_X86_64
3173         gva = (long long)(gva << 16) >> 16;
3174 #endif
3175         return gva;
3176 }
3177
3178
3179 typedef void (*inspect_spte_fn) (struct kvm *kvm, u64 *sptep);
3180
3181 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3182                             inspect_spte_fn fn)
3183 {
3184         int i;
3185
3186         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3187                 u64 ent = sp->spt[i];
3188
3189                 if (is_shadow_present_pte(ent)) {
3190                         if (!is_last_spte(ent, sp->role.level)) {
3191                                 struct kvm_mmu_page *child;
3192                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3193                                 __mmu_spte_walk(kvm, child, fn);
3194                         } else
3195                                 fn(kvm, &sp->spt[i]);
3196                 }
3197         }
3198 }
3199
3200 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3201 {
3202         int i;
3203         struct kvm_mmu_page *sp;
3204
3205         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3206                 return;
3207         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3208                 hpa_t root = vcpu->arch.mmu.root_hpa;
3209                 sp = page_header(root);
3210                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3211                 return;
3212         }
3213         for (i = 0; i < 4; ++i) {
3214                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3215
3216                 if (root && VALID_PAGE(root)) {
3217                         root &= PT64_BASE_ADDR_MASK;
3218                         sp = page_header(root);
3219                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3220                 }
3221         }
3222         return;
3223 }
3224
3225 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3226                                 gva_t va, int level)
3227 {
3228         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3229         int i;
3230         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3231
3232         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3233                 u64 ent = pt[i];
3234
3235                 if (ent == shadow_trap_nonpresent_pte)
3236                         continue;
3237
3238                 va = canonicalize(va);
3239                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3240                         audit_mappings_page(vcpu, ent, va, level - 1);
3241                 else {
3242                         gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, va, NULL);
3243                         gfn_t gfn = gpa >> PAGE_SHIFT;
3244                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3245                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3246
3247                         if (is_error_pfn(pfn)) {
3248                                 kvm_release_pfn_clean(pfn);
3249                                 continue;
3250                         }
3251
3252                         if (is_shadow_present_pte(ent)
3253                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3254                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3255                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3256                                        audit_msg, vcpu->arch.mmu.root_level,
3257                                        va, gpa, hpa, ent,
3258                                        is_shadow_present_pte(ent));
3259                         else if (ent == shadow_notrap_nonpresent_pte
3260                                  && !is_error_hpa(hpa))
3261                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3262                                        " valid guest gva %lx\n", audit_msg, va);
3263                         kvm_release_pfn_clean(pfn);
3264
3265                 }
3266         }
3267 }
3268
3269 static void audit_mappings(struct kvm_vcpu *vcpu)
3270 {
3271         unsigned i;
3272
3273         if (vcpu->arch.mmu.root_level == 4)
3274                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3275         else
3276                 for (i = 0; i < 4; ++i)
3277                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3278                                 audit_mappings_page(vcpu,
3279                                                     vcpu->arch.mmu.pae_root[i],
3280                                                     i << 30,
3281                                                     2);
3282 }
3283
3284 static int count_rmaps(struct kvm_vcpu *vcpu)
3285 {
3286         struct kvm *kvm = vcpu->kvm;
3287         struct kvm_memslots *slots;
3288         int nmaps = 0;
3289         int i, j, k, idx;
3290
3291         idx = srcu_read_lock(&kvm->srcu);
3292         slots = rcu_dereference(kvm->memslots);
3293         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3294                 struct kvm_memory_slot *m = &slots->memslots[i];
3295                 struct kvm_rmap_desc *d;
3296
3297                 for (j = 0; j < m->npages; ++j) {
3298                         unsigned long *rmapp = &m->rmap[j];
3299
3300                         if (!*rmapp)
3301                                 continue;
3302                         if (!(*rmapp & 1)) {
3303                                 ++nmaps;
3304                                 continue;
3305                         }
3306                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3307                         while (d) {
3308                                 for (k = 0; k < RMAP_EXT; ++k)
3309                                         if (d->sptes[k])
3310                                                 ++nmaps;
3311                                         else
3312                                                 break;
3313                                 d = d->more;
3314                         }
3315                 }
3316         }
3317         srcu_read_unlock(&kvm->srcu, idx);
3318         return nmaps;
3319 }
3320
3321 void inspect_spte_has_rmap(struct kvm *kvm, u64 *sptep)
3322 {
3323         unsigned long *rmapp;
3324         struct kvm_mmu_page *rev_sp;
3325         gfn_t gfn;
3326
3327         if (*sptep & PT_WRITABLE_MASK) {
3328                 rev_sp = page_header(__pa(sptep));
3329                 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3330
3331                 if (!gfn_to_memslot(kvm, gfn)) {
3332                         if (!printk_ratelimit())
3333                                 return;
3334                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3335                                          audit_msg, gfn);
3336                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3337                                audit_msg, (long int)(sptep - rev_sp->spt),
3338                                         rev_sp->gfn);
3339                         dump_stack();
3340                         return;
3341                 }
3342
3343                 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3344                                     rev_sp->role.level);
3345                 if (!*rmapp) {
3346                         if (!printk_ratelimit())
3347                                 return;
3348                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3349                                          audit_msg, *sptep);
3350                         dump_stack();
3351                 }
3352         }
3353
3354 }
3355
3356 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3357 {
3358         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3359 }
3360
3361 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3362 {
3363         struct kvm_mmu_page *sp;
3364         int i;
3365
3366         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3367                 u64 *pt = sp->spt;
3368
3369                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3370                         continue;
3371
3372                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3373                         u64 ent = pt[i];
3374
3375                         if (!(ent & PT_PRESENT_MASK))
3376                                 continue;
3377                         if (!(ent & PT_WRITABLE_MASK))
3378                                 continue;
3379                         inspect_spte_has_rmap(vcpu->kvm, &pt[i]);
3380                 }
3381         }
3382         return;
3383 }
3384
3385 static void audit_rmap(struct kvm_vcpu *vcpu)
3386 {
3387         check_writable_mappings_rmap(vcpu);
3388         count_rmaps(vcpu);
3389 }
3390
3391 static void audit_write_protection(struct kvm_vcpu *vcpu)
3392 {
3393         struct kvm_mmu_page *sp;
3394         struct kvm_memory_slot *slot;
3395         unsigned long *rmapp;
3396         u64 *spte;
3397         gfn_t gfn;
3398
3399         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3400                 if (sp->role.direct)
3401                         continue;
3402                 if (sp->unsync)
3403                         continue;
3404
3405                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3406                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3407                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3408
3409                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3410                 while (spte) {
3411                         if (*spte & PT_WRITABLE_MASK)
3412                                 printk(KERN_ERR "%s: (%s) shadow page has "
3413                                 "writable mappings: gfn %lx role %x\n",
3414                                __func__, audit_msg, sp->gfn,
3415                                sp->role.word);
3416                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3417                 }
3418         }
3419 }
3420
3421 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3422 {
3423         int olddbg = dbg;
3424
3425         dbg = 0;
3426         audit_msg = msg;
3427         audit_rmap(vcpu);
3428         audit_write_protection(vcpu);
3429         if (strcmp("pre pte write", audit_msg) != 0)
3430                 audit_mappings(vcpu);
3431         audit_writable_sptes_have_rmaps(vcpu);
3432         dbg = olddbg;
3433 }
3434
3435 #endif