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_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 bool 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         slot = gfn_to_memslot_unaliased(kvm, gfn);
 435         for (i = PT_DIRECTORY_LEVEL;
 436              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 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         u64 *prev_spte;
 651         int i;
 652
 653         if (!*rmapp)
 654                 return NULL;
 655         else if (!(*rmapp & 1)) {
 656                 if (!spte)
 657                         return (u64 *)*rmapp;
 658                 return NULL;
 659         }
 660         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 661         prev_spte = NULL;
 662         while (desc) {
 663                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
 664                         if (prev_spte == spte)
 665                                 return desc->sptes[i];
 666                         prev_spte = desc->sptes[i];
 667                 }
 668                 desc = desc->more;
 669         }
 670         return NULL;
 671 }
 672
 673 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 674 {
 675         unsigned long *rmapp;
 676         u64 *spte;
 677         int i, write_protected = 0;
 678
 679         gfn = unalias_gfn(kvm, gfn);
 680         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
 681
 682         spte = rmap_next(kvm, rmapp, NULL);
 683         while (spte) {
 684                 BUG_ON(!spte);
 685                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 686                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 687                 if (is_writable_pte(*spte)) {
 688                         __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
 689                         write_protected = 1;
 690                 }
 691                 spte = rmap_next(kvm, rmapp, spte);
 692         }
 693         if (write_protected) {
 694                 pfn_t pfn;
 695
 696                 spte = rmap_next(kvm, rmapp, NULL);
 697                 pfn = spte_to_pfn(*spte);
 698                 kvm_set_pfn_dirty(pfn);
 699         }
 700
 701         /* check for huge page mappings */
 702         for (i = PT_DIRECTORY_LEVEL;
 703              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 704                 rmapp = gfn_to_rmap(kvm, gfn, i);
 705                 spte = rmap_next(kvm, rmapp, NULL);
 706                 while (spte) {
 707                         BUG_ON(!spte);
 708                         BUG_ON(!(*spte & PT_PRESENT_MASK));
 709                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 710                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 711                         if (is_writable_pte(*spte)) {
 712                                 rmap_remove(kvm, spte);
 713                                 --kvm->stat.lpages;
 714                                 __set_spte(spte, shadow_trap_nonpresent_pte);
 715                                 spte = NULL;
 716                                 write_protected = 1;
 717                         }
 718                         spte = rmap_next(kvm, rmapp, spte);
 719                 }
 720         }
 721
 722         return write_protected;
 723 }
 724
 725 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 726                            unsigned long data)
 727 {
 728         u64 *spte;
 729         int need_tlb_flush = 0;
 730
 731         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 732                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 733                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 734                 rmap_remove(kvm, spte);
 735                 __set_spte(spte, shadow_trap_nonpresent_pte);
 736                 need_tlb_flush = 1;
 737         }
 738         return need_tlb_flush;
 739 }
 740
 741 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
 742                              unsigned long data)
 743 {
 744         int need_flush = 0;
 745         u64 *spte, new_spte;
 746         pte_t *ptep = (pte_t *)data;
 747         pfn_t new_pfn;
 748
 749         WARN_ON(pte_huge(*ptep));
 750         new_pfn = pte_pfn(*ptep);
 751         spte = rmap_next(kvm, rmapp, NULL);
 752         while (spte) {
 753                 BUG_ON(!is_shadow_present_pte(*spte));
 754                 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
 755                 need_flush = 1;
 756                 if (pte_write(*ptep)) {
 757                         rmap_remove(kvm, spte);
 758                         __set_spte(spte, shadow_trap_nonpresent_pte);
 759                         spte = rmap_next(kvm, rmapp, NULL);
 760                 } else {
 761                         new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
 762                         new_spte |= (u64)new_pfn << PAGE_SHIFT;
 763
 764                         new_spte &= ~PT_WRITABLE_MASK;
 765                         new_spte &= ~SPTE_HOST_WRITEABLE;
 766                         if (is_writable_pte(*spte))
 767                                 kvm_set_pfn_dirty(spte_to_pfn(*spte));
 768                         __set_spte(spte, new_spte);
 769                         spte = rmap_next(kvm, rmapp, spte);
 770                 }
 771         }
 772         if (need_flush)
 773                 kvm_flush_remote_tlbs(kvm);
 774
 775         return 0;
 776 }
 777
 778 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 779                           unsigned long data,
 780                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 781                                          unsigned long data))
 782 {
 783         int i, j;
 784         int ret;
 785         int retval = 0;
 786         struct kvm_memslots *slots;
 787
 788         slots = kvm_memslots(kvm);
 789
 790         for (i = 0; i < slots->nmemslots; i++) {
 791                 struct kvm_memory_slot *memslot = &slots->memslots[i];
 792                 unsigned long start = memslot->userspace_addr;
 793                 unsigned long end;
 794
 795                 end = start + (memslot->npages << PAGE_SHIFT);
 796                 if (hva >= start && hva < end) {
 797                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 798
 799                         ret = handler(kvm, &memslot->rmap[gfn_offset], data);
 800
 801                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
 802                                 int idx = gfn_offset;
 803                                 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
 804                                 ret |= handler(kvm,
 805                                         &memslot->lpage_info[j][idx].rmap_pde,
 806                                         data);
 807                         }
 808                         trace_kvm_age_page(hva, memslot, ret);
 809                         retval |= ret;
 810                 }
 811         }
 812
 813         return retval;
 814 }
 815
 816 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 817 {
 818         return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
 819 }
 820
 821 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
 822 {
 823         kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
 824 }
 825
 826 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 827                          unsigned long data)
 828 {
 829         u64 *spte;
 830         int young = 0;
 831
 832         /*
 833          * Emulate the accessed bit for EPT, by checking if this page has
 834          * an EPT mapping, and clearing it if it does. On the next access,
 835          * a new EPT mapping will be established.
 836          * This has some overhead, but not as much as the cost of swapping
 837          * out actively used pages or breaking up actively used hugepages.
 838          */
 839         if (!shadow_accessed_mask)
 840                 return kvm_unmap_rmapp(kvm, rmapp, data);
 841
 842         spte = rmap_next(kvm, rmapp, NULL);
 843         while (spte) {
 844                 int _young;
 845                 u64 _spte = *spte;
 846                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 847                 _young = _spte & PT_ACCESSED_MASK;
 848                 if (_young) {
 849                         young = 1;
 850                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 851                 }
 852                 spte = rmap_next(kvm, rmapp, spte);
 853         }
 854         return young;
 855 }
 856
 857 #define RMAP_RECYCLE_THRESHOLD 1000
 858
 859 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 860 {
 861         unsigned long *rmapp;
 862         struct kvm_mmu_page *sp;
 863
 864         sp = page_header(__pa(spte));
 865
 866         gfn = unalias_gfn(vcpu->kvm, gfn);
 867         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 868
 869         kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
 870         kvm_flush_remote_tlbs(vcpu->kvm);
 871 }
 872
 873 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 874 {
 875         return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
 876 }
 877
 878 #ifdef MMU_DEBUG
 879 static int is_empty_shadow_page(u64 *spt)
 880 {
 881         u64 *pos;
 882         u64 *end;
 883
 884         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 885                 if (is_shadow_present_pte(*pos)) {
 886                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 887                                pos, *pos);
 888                         return 0;
 889                 }
 890         return 1;
 891 }
 892 #endif
 893
 894 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 895 {
 896         ASSERT(is_empty_shadow_page(sp->spt));
 897         list_del(&sp->link);
 898         __free_page(virt_to_page(sp->spt));
 899         __free_page(virt_to_page(sp->gfns));
 900         kfree(sp);
 901         ++kvm->arch.n_free_mmu_pages;
 902 }
 903
 904 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 905 {
 906         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 907 }
 908
 909 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 910                                                u64 *parent_pte)
 911 {
 912         struct kvm_mmu_page *sp;
 913
 914         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 915         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 916         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 917         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 918         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 919         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 920         sp->multimapped = 0;
 921         sp->parent_pte = parent_pte;
 922         --vcpu->kvm->arch.n_free_mmu_pages;
 923         return sp;
 924 }
 925
 926 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 927                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 928 {
 929         struct kvm_pte_chain *pte_chain;
 930         struct hlist_node *node;
 931         int i;
 932
 933         if (!parent_pte)
 934                 return;
 935         if (!sp->multimapped) {
 936                 u64 *old = sp->parent_pte;
 937
 938                 if (!old) {
 939                         sp->parent_pte = parent_pte;
 940                         return;
 941                 }
 942                 sp->multimapped = 1;
 943                 pte_chain = mmu_alloc_pte_chain(vcpu);
 944                 INIT_HLIST_HEAD(&sp->parent_ptes);
 945                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 946                 pte_chain->parent_ptes[0] = old;
 947         }
 948         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 949                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 950                         continue;
 951                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 952                         if (!pte_chain->parent_ptes[i]) {
 953                                 pte_chain->parent_ptes[i] = parent_pte;
 954                                 return;
 955                         }
 956         }
 957         pte_chain = mmu_alloc_pte_chain(vcpu);
 958         BUG_ON(!pte_chain);
 959         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 960         pte_chain->parent_ptes[0] = parent_pte;
 961 }
 962
 963 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 964                                        u64 *parent_pte)
 965 {
 966         struct kvm_pte_chain *pte_chain;
 967         struct hlist_node *node;
 968         int i;
 969
 970         if (!sp->multimapped) {
 971                 BUG_ON(sp->parent_pte != parent_pte);
 972                 sp->parent_pte = NULL;
 973                 return;
 974         }
 975         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 976                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 977                         if (!pte_chain->parent_ptes[i])
 978                                 break;
 979                         if (pte_chain->parent_ptes[i] != parent_pte)
 980                                 continue;
 981                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 982                                 && pte_chain->parent_ptes[i + 1]) {
 983                                 pte_chain->parent_ptes[i]
 984                                         = pte_chain->parent_ptes[i + 1];
 985                                 ++i;
 986                         }
 987                         pte_chain->parent_ptes[i] = NULL;
 988                         if (i == 0) {
 989                                 hlist_del(&pte_chain->link);
 990                                 mmu_free_pte_chain(pte_chain);
 991                                 if (hlist_empty(&sp->parent_ptes)) {
 992                                         sp->multimapped = 0;
 993                                         sp->parent_pte = NULL;
 994                                 }
 995                         }
 996                         return;
 997                 }
 998         BUG();
 999 }
1000
1001
1002 static void mmu_parent_walk(struct kvm_mmu_page *sp, mmu_parent_walk_fn fn)
1003 {
1004         struct kvm_pte_chain *pte_chain;
1005         struct hlist_node *node;
1006         struct kvm_mmu_page *parent_sp;
1007         int i;
1008
1009         if (!sp->multimapped && sp->parent_pte) {
1010                 parent_sp = page_header(__pa(sp->parent_pte));
1011                 fn(parent_sp);
1012                 mmu_parent_walk(parent_sp, fn);
1013                 return;
1014         }
1015         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1016                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1017                         if (!pte_chain->parent_ptes[i])
1018                                 break;
1019                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
1020                         fn(parent_sp);
1021                         mmu_parent_walk(parent_sp, fn);
1022                 }
1023 }
1024
1025 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1026 {
1027         unsigned int index;
1028         struct kvm_mmu_page *sp = page_header(__pa(spte));
1029
1030         index = spte - sp->spt;
1031         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1032                 sp->unsync_children++;
1033         WARN_ON(!sp->unsync_children);
1034 }
1035
1036 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1037 {
1038         struct kvm_pte_chain *pte_chain;
1039         struct hlist_node *node;
1040         int i;
1041
1042         if (!sp->parent_pte)
1043                 return;
1044
1045         if (!sp->multimapped) {
1046                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1047                 return;
1048         }
1049
1050         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1051                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1052                         if (!pte_chain->parent_ptes[i])
1053                                 break;
1054                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1055                 }
1056 }
1057
1058 static int unsync_walk_fn(struct kvm_mmu_page *sp)
1059 {
1060         kvm_mmu_update_parents_unsync(sp);
1061         return 1;
1062 }
1063
1064 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1065 {
1066         mmu_parent_walk(sp, unsync_walk_fn);
1067         kvm_mmu_update_parents_unsync(sp);
1068 }
1069
1070 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1071                                     struct kvm_mmu_page *sp)
1072 {
1073         int i;
1074
1075         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1076                 sp->spt[i] = shadow_trap_nonpresent_pte;
1077 }
1078
1079 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1080                                struct kvm_mmu_page *sp)
1081 {
1082         return 1;
1083 }
1084
1085 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1086 {
1087 }
1088
1089 #define KVM_PAGE_ARRAY_NR 16
1090
1091 struct kvm_mmu_pages {
1092         struct mmu_page_and_offset {
1093                 struct kvm_mmu_page *sp;
1094                 unsigned int idx;
1095         } page[KVM_PAGE_ARRAY_NR];
1096         unsigned int nr;
1097 };
1098
1099 #define for_each_unsync_children(bitmap, idx)           \
1100         for (idx = find_first_bit(bitmap, 512);         \
1101              idx < 512;                                 \
1102              idx = find_next_bit(bitmap, 512, idx+1))
1103
1104 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1105                          int idx)
1106 {
1107         int i;
1108
1109         if (sp->unsync)
1110                 for (i=0; i < pvec->nr; i++)
1111                         if (pvec->page[i].sp == sp)
1112                                 return 0;
1113
1114         pvec->page[pvec->nr].sp = sp;
1115         pvec->page[pvec->nr].idx = idx;
1116         pvec->nr++;
1117         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1118 }
1119
1120 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1121                            struct kvm_mmu_pages *pvec)
1122 {
1123         int i, ret, nr_unsync_leaf = 0;
1124
1125         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1126                 u64 ent = sp->spt[i];
1127
1128                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1129                         struct kvm_mmu_page *child;
1130                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1131
1132                         if (child->unsync_children) {
1133                                 if (mmu_pages_add(pvec, child, i))
1134                                         return -ENOSPC;
1135
1136                                 ret = __mmu_unsync_walk(child, pvec);
1137                                 if (!ret)
1138                                         __clear_bit(i, sp->unsync_child_bitmap);
1139                                 else if (ret > 0)
1140                                         nr_unsync_leaf += ret;
1141                                 else
1142                                         return ret;
1143                         }
1144
1145                         if (child->unsync) {
1146                                 nr_unsync_leaf++;
1147                                 if (mmu_pages_add(pvec, child, i))
1148                                         return -ENOSPC;
1149                         }
1150                 }
1151         }
1152
1153         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1154                 sp->unsync_children = 0;
1155
1156         return nr_unsync_leaf;
1157 }
1158
1159 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1160                            struct kvm_mmu_pages *pvec)
1161 {
1162         if (!sp->unsync_children)
1163                 return 0;
1164
1165         mmu_pages_add(pvec, sp, 0);
1166         return __mmu_unsync_walk(sp, pvec);
1167 }
1168
1169 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1170 {
1171         unsigned index;
1172         struct hlist_head *bucket;
1173         struct kvm_mmu_page *sp;
1174         struct hlist_node *node;
1175
1176         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1177         index = kvm_page_table_hashfn(gfn);
1178         bucket = &kvm->arch.mmu_page_hash[index];
1179         hlist_for_each_entry(sp, node, bucket, hash_link)
1180                 if (sp->gfn == gfn && !sp->role.direct
1181                     && !sp->role.invalid) {
1182                         pgprintk("%s: found role %x\n",
1183                                  __func__, sp->role.word);
1184                         return sp;
1185                 }
1186         return NULL;
1187 }
1188
1189 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1190 {
1191         WARN_ON(!sp->unsync);
1192         trace_kvm_mmu_sync_page(sp);
1193         sp->unsync = 0;
1194         --kvm->stat.mmu_unsync;
1195 }
1196
1197 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1198
1199 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1200 {
1201         if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1202                 kvm_mmu_zap_page(vcpu->kvm, sp);
1203                 return 1;
1204         }
1205
1206         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1207                 kvm_flush_remote_tlbs(vcpu->kvm);
1208         kvm_unlink_unsync_page(vcpu->kvm, sp);
1209         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1210                 kvm_mmu_zap_page(vcpu->kvm, sp);
1211                 return 1;
1212         }
1213
1214         kvm_mmu_flush_tlb(vcpu);
1215         return 0;
1216 }
1217
1218 struct mmu_page_path {
1219         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1220         unsigned int idx[PT64_ROOT_LEVEL-1];
1221 };
1222
1223 #define for_each_sp(pvec, sp, parents, i)                       \
1224                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1225                         sp = pvec.page[i].sp;                   \
1226                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1227                         i = mmu_pages_next(&pvec, &parents, i))
1228
1229 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1230                           struct mmu_page_path *parents,
1231                           int i)
1232 {
1233         int n;
1234
1235         for (n = i+1; n < pvec->nr; n++) {
1236                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1237
1238                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1239                         parents->idx[0] = pvec->page[n].idx;
1240                         return n;
1241                 }
1242
1243                 parents->parent[sp->role.level-2] = sp;
1244                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1245         }
1246
1247         return n;
1248 }
1249
1250 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1251 {
1252         struct kvm_mmu_page *sp;
1253         unsigned int level = 0;
1254
1255         do {
1256                 unsigned int idx = parents->idx[level];
1257
1258                 sp = parents->parent[level];
1259                 if (!sp)
1260                         return;
1261
1262                 --sp->unsync_children;
1263                 WARN_ON((int)sp->unsync_children < 0);
1264                 __clear_bit(idx, sp->unsync_child_bitmap);
1265                 level++;
1266         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1267 }
1268
1269 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1270                                struct mmu_page_path *parents,
1271                                struct kvm_mmu_pages *pvec)
1272 {
1273         parents->parent[parent->role.level-1] = NULL;
1274         pvec->nr = 0;
1275 }
1276
1277 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1278                               struct kvm_mmu_page *parent)
1279 {
1280         int i;
1281         struct kvm_mmu_page *sp;
1282         struct mmu_page_path parents;
1283         struct kvm_mmu_pages pages;
1284
1285         kvm_mmu_pages_init(parent, &parents, &pages);
1286         while (mmu_unsync_walk(parent, &pages)) {
1287                 int protected = 0;
1288
1289                 for_each_sp(pages, sp, parents, i)
1290                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1291
1292                 if (protected)
1293                         kvm_flush_remote_tlbs(vcpu->kvm);
1294
1295                 for_each_sp(pages, sp, parents, i) {
1296                         kvm_sync_page(vcpu, sp);
1297                         mmu_pages_clear_parents(&parents);
1298                 }
1299                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1300                 kvm_mmu_pages_init(parent, &parents, &pages);
1301         }
1302 }
1303
1304 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1305                                              gfn_t gfn,
1306                                              gva_t gaddr,
1307                                              unsigned level,
1308                                              int direct,
1309                                              unsigned access,
1310                                              u64 *parent_pte)
1311 {
1312         union kvm_mmu_page_role role;
1313         unsigned index;
1314         unsigned quadrant;
1315         struct hlist_head *bucket;
1316         struct kvm_mmu_page *sp;
1317         struct hlist_node *node, *tmp;
1318
1319         role = vcpu->arch.mmu.base_role;
1320         role.level = level;
1321         role.direct = direct;
1322         if (role.direct)
1323                 role.cr4_pae = 0;
1324         role.access = access;
1325         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1326                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1327                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1328                 role.quadrant = quadrant;
1329         }
1330         index = kvm_page_table_hashfn(gfn);
1331         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1332         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1333                 if (sp->gfn == gfn) {
1334                         if (sp->unsync)
1335                                 if (kvm_sync_page(vcpu, sp))
1336                                         continue;
1337
1338                         if (sp->role.word != role.word)
1339                                 continue;
1340
1341                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1342                         if (sp->unsync_children) {
1343                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1344                                 kvm_mmu_mark_parents_unsync(sp);
1345                         }
1346                         trace_kvm_mmu_get_page(sp, false);
1347                         return sp;
1348                 }
1349         ++vcpu->kvm->stat.mmu_cache_miss;
1350         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1351         if (!sp)
1352                 return sp;
1353         sp->gfn = gfn;
1354         sp->role = role;
1355         hlist_add_head(&sp->hash_link, bucket);
1356         if (!direct) {
1357                 if (rmap_write_protect(vcpu->kvm, gfn))
1358                         kvm_flush_remote_tlbs(vcpu->kvm);
1359                 account_shadowed(vcpu->kvm, gfn);
1360         }
1361         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1362                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1363         else
1364                 nonpaging_prefetch_page(vcpu, sp);
1365         trace_kvm_mmu_get_page(sp, true);
1366         return sp;
1367 }
1368
1369 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1370                              struct kvm_vcpu *vcpu, u64 addr)
1371 {
1372         iterator->addr = addr;
1373         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1374         iterator->level = vcpu->arch.mmu.shadow_root_level;
1375         if (iterator->level == PT32E_ROOT_LEVEL) {
1376                 iterator->shadow_addr
1377                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1378                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1379                 --iterator->level;
1380                 if (!iterator->shadow_addr)
1381                         iterator->level = 0;
1382         }
1383 }
1384
1385 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1386 {
1387         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1388                 return false;
1389
1390         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1391                 if (is_large_pte(*iterator->sptep))
1392                         return false;
1393
1394         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1395         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1396         return true;
1397 }
1398
1399 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1400 {
1401         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1402         --iterator->level;
1403 }
1404
1405 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1406                                          struct kvm_mmu_page *sp)
1407 {
1408         unsigned i;
1409         u64 *pt;
1410         u64 ent;
1411
1412         pt = sp->spt;
1413
1414         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1415                 ent = pt[i];
1416
1417                 if (is_shadow_present_pte(ent)) {
1418                         if (!is_last_spte(ent, sp->role.level)) {
1419                                 ent &= PT64_BASE_ADDR_MASK;
1420                                 mmu_page_remove_parent_pte(page_header(ent),
1421                                                            &pt[i]);
1422                         } else {
1423                                 if (is_large_pte(ent))
1424                                         --kvm->stat.lpages;
1425                                 rmap_remove(kvm, &pt[i]);
1426                         }
1427                 }
1428                 pt[i] = shadow_trap_nonpresent_pte;
1429         }
1430 }
1431
1432 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1433 {
1434         mmu_page_remove_parent_pte(sp, parent_pte);
1435 }
1436
1437 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1438 {
1439         int i;
1440         struct kvm_vcpu *vcpu;
1441
1442         kvm_for_each_vcpu(i, vcpu, kvm)
1443                 vcpu->arch.last_pte_updated = NULL;
1444 }
1445
1446 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1447 {
1448         u64 *parent_pte;
1449
1450         while (sp->multimapped || sp->parent_pte) {
1451                 if (!sp->multimapped)
1452                         parent_pte = sp->parent_pte;
1453                 else {
1454                         struct kvm_pte_chain *chain;
1455
1456                         chain = container_of(sp->parent_ptes.first,
1457                                              struct kvm_pte_chain, link);
1458                         parent_pte = chain->parent_ptes[0];
1459                 }
1460                 BUG_ON(!parent_pte);
1461                 kvm_mmu_put_page(sp, parent_pte);
1462                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1463         }
1464 }
1465
1466 static int mmu_zap_unsync_children(struct kvm *kvm,
1467                                    struct kvm_mmu_page *parent)
1468 {
1469         int i, zapped = 0;
1470         struct mmu_page_path parents;
1471         struct kvm_mmu_pages pages;
1472
1473         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1474                 return 0;
1475
1476         kvm_mmu_pages_init(parent, &parents, &pages);
1477         while (mmu_unsync_walk(parent, &pages)) {
1478                 struct kvm_mmu_page *sp;
1479
1480                 for_each_sp(pages, sp, parents, i) {
1481                         kvm_mmu_zap_page(kvm, sp);
1482                         mmu_pages_clear_parents(&parents);
1483                         zapped++;
1484                 }
1485                 kvm_mmu_pages_init(parent, &parents, &pages);
1486         }
1487
1488         return zapped;
1489 }
1490
1491 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1492 {
1493         int ret;
1494
1495         trace_kvm_mmu_zap_page(sp);
1496         ++kvm->stat.mmu_shadow_zapped;
1497         ret = mmu_zap_unsync_children(kvm, sp);
1498         kvm_mmu_page_unlink_children(kvm, sp);
1499         kvm_mmu_unlink_parents(kvm, sp);
1500         kvm_flush_remote_tlbs(kvm);
1501         if (!sp->role.invalid && !sp->role.direct)
1502                 unaccount_shadowed(kvm, sp->gfn);
1503         if (sp->unsync)
1504                 kvm_unlink_unsync_page(kvm, sp);
1505         if (!sp->root_count) {
1506                 hlist_del(&sp->hash_link);
1507                 kvm_mmu_free_page(kvm, sp);
1508         } else {
1509                 sp->role.invalid = 1;
1510                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1511                 kvm_reload_remote_mmus(kvm);
1512         }
1513         kvm_mmu_reset_last_pte_updated(kvm);
1514         return ret;
1515 }
1516
1517 /*
1518  * Changing the number of mmu pages allocated to the vm
1519  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1520  */
1521 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1522 {
1523         int used_pages;
1524
1525         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1526         used_pages = max(0, used_pages);
1527
1528         /*
1529          * If we set the number of mmu pages to be smaller be than the
1530          * number of actived pages , we must to free some mmu pages before we
1531          * change the value
1532          */
1533
1534         if (used_pages > kvm_nr_mmu_pages) {
1535                 while (used_pages > kvm_nr_mmu_pages &&
1536                         !list_empty(&kvm->arch.active_mmu_pages)) {
1537                         struct kvm_mmu_page *page;
1538
1539                         page = container_of(kvm->arch.active_mmu_pages.prev,
1540                                             struct kvm_mmu_page, link);
1541                         used_pages -= kvm_mmu_zap_page(kvm, page);
1542                         used_pages--;
1543                 }
1544                 kvm_nr_mmu_pages = used_pages;
1545                 kvm->arch.n_free_mmu_pages = 0;
1546         }
1547         else
1548                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1549                                          - kvm->arch.n_alloc_mmu_pages;
1550
1551         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1552 }
1553
1554 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1555 {
1556         unsigned index;
1557         struct hlist_head *bucket;
1558         struct kvm_mmu_page *sp;
1559         struct hlist_node *node, *n;
1560         int r;
1561
1562         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1563         r = 0;
1564         index = kvm_page_table_hashfn(gfn);
1565         bucket = &kvm->arch.mmu_page_hash[index];
1566 restart:
1567         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1568                 if (sp->gfn == gfn && !sp->role.direct) {
1569                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1570                                  sp->role.word);
1571                         r = 1;
1572                         if (kvm_mmu_zap_page(kvm, sp))
1573                                 goto restart;
1574                 }
1575         return r;
1576 }
1577
1578 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1579 {
1580         unsigned index;
1581         struct hlist_head *bucket;
1582         struct kvm_mmu_page *sp;
1583         struct hlist_node *node, *nn;
1584
1585         index = kvm_page_table_hashfn(gfn);
1586         bucket = &kvm->arch.mmu_page_hash[index];
1587 restart:
1588         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1589                 if (sp->gfn == gfn && !sp->role.direct
1590                     && !sp->role.invalid) {
1591                         pgprintk("%s: zap %lx %x\n",
1592                                  __func__, gfn, sp->role.word);
1593                         if (kvm_mmu_zap_page(kvm, sp))
1594                                 goto restart;
1595                 }
1596         }
1597 }
1598
1599 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1600 {
1601         int slot = memslot_id(kvm, gfn);
1602         struct kvm_mmu_page *sp = page_header(__pa(pte));
1603
1604         __set_bit(slot, sp->slot_bitmap);
1605 }
1606
1607 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1608 {
1609         int i;
1610         u64 *pt = sp->spt;
1611
1612         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1613                 return;
1614
1615         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1616                 if (pt[i] == shadow_notrap_nonpresent_pte)
1617                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1618         }
1619 }
1620
1621 /*
1622  * The function is based on mtrr_type_lookup() in
1623  * arch/x86/kernel/cpu/mtrr/generic.c
1624  */
1625 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1626                          u64 start, u64 end)
1627 {
1628         int i;
1629         u64 base, mask;
1630         u8 prev_match, curr_match;
1631         int num_var_ranges = KVM_NR_VAR_MTRR;
1632
1633         if (!mtrr_state->enabled)
1634                 return 0xFF;
1635
1636         /* Make end inclusive end, instead of exclusive */
1637         end--;
1638
1639         /* Look in fixed ranges. Just return the type as per start */
1640         if (mtrr_state->have_fixed && (start < 0x100000)) {
1641                 int idx;
1642
1643                 if (start < 0x80000) {
1644                         idx = 0;
1645                         idx += (start >> 16);
1646                         return mtrr_state->fixed_ranges[idx];
1647                 } else if (start < 0xC0000) {
1648                         idx = 1 * 8;
1649                         idx += ((start - 0x80000) >> 14);
1650                         return mtrr_state->fixed_ranges[idx];
1651                 } else if (start < 0x1000000) {
1652                         idx = 3 * 8;
1653                         idx += ((start - 0xC0000) >> 12);
1654                         return mtrr_state->fixed_ranges[idx];
1655                 }
1656         }
1657
1658         /*
1659          * Look in variable ranges
1660          * Look of multiple ranges matching this address and pick type
1661          * as per MTRR precedence
1662          */
1663         if (!(mtrr_state->enabled & 2))
1664                 return mtrr_state->def_type;
1665
1666         prev_match = 0xFF;
1667         for (i = 0; i < num_var_ranges; ++i) {
1668                 unsigned short start_state, end_state;
1669
1670                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1671                         continue;
1672
1673                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1674                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1675                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1676                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1677
1678                 start_state = ((start & mask) == (base & mask));
1679                 end_state = ((end & mask) == (base & mask));
1680                 if (start_state != end_state)
1681                         return 0xFE;
1682
1683                 if ((start & mask) != (base & mask))
1684                         continue;
1685
1686                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1687                 if (prev_match == 0xFF) {
1688                         prev_match = curr_match;
1689                         continue;
1690                 }
1691
1692                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1693                     curr_match == MTRR_TYPE_UNCACHABLE)
1694                         return MTRR_TYPE_UNCACHABLE;
1695
1696                 if ((prev_match == MTRR_TYPE_WRBACK &&
1697                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1698                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1699                      curr_match == MTRR_TYPE_WRBACK)) {
1700                         prev_match = MTRR_TYPE_WRTHROUGH;
1701                         curr_match = MTRR_TYPE_WRTHROUGH;
1702                 }
1703
1704                 if (prev_match != curr_match)
1705                         return MTRR_TYPE_UNCACHABLE;
1706         }
1707
1708         if (prev_match != 0xFF)
1709                 return prev_match;
1710
1711         return mtrr_state->def_type;
1712 }
1713
1714 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1715 {
1716         u8 mtrr;
1717
1718         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1719                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1720         if (mtrr == 0xfe || mtrr == 0xff)
1721                 mtrr = MTRR_TYPE_WRBACK;
1722         return mtrr;
1723 }
1724 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1725
1726 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1727 {
1728         unsigned index;
1729         struct hlist_head *bucket;
1730         struct kvm_mmu_page *s;
1731         struct hlist_node *node, *n;
1732
1733         index = kvm_page_table_hashfn(sp->gfn);
1734         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1735         /* don't unsync if pagetable is shadowed with multiple roles */
1736         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1737                 if (s->gfn != sp->gfn || s->role.direct)
1738                         continue;
1739                 if (s->role.word != sp->role.word)
1740                         return 1;
1741         }
1742         trace_kvm_mmu_unsync_page(sp);
1743         ++vcpu->kvm->stat.mmu_unsync;
1744         sp->unsync = 1;
1745
1746         kvm_mmu_mark_parents_unsync(sp);
1747
1748         mmu_convert_notrap(sp);
1749         return 0;
1750 }
1751
1752 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1753                                   bool can_unsync)
1754 {
1755         struct kvm_mmu_page *shadow;
1756
1757         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1758         if (shadow) {
1759                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1760                         return 1;
1761                 if (shadow->unsync)
1762                         return 0;
1763                 if (can_unsync && oos_shadow)
1764                         return kvm_unsync_page(vcpu, shadow);
1765                 return 1;
1766         }
1767         return 0;
1768 }
1769
1770 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1771                     unsigned pte_access, int user_fault,
1772                     int write_fault, int dirty, int level,
1773                     gfn_t gfn, pfn_t pfn, bool speculative,
1774                     bool can_unsync, bool reset_host_protection)
1775 {
1776         u64 spte;
1777         int ret = 0;
1778
1779         /*
1780          * We don't set the accessed bit, since we sometimes want to see
1781          * whether the guest actually used the pte (in order to detect
1782          * demand paging).
1783          */
1784         spte = shadow_base_present_pte | shadow_dirty_mask;
1785         if (!speculative)
1786                 spte |= shadow_accessed_mask;
1787         if (!dirty)
1788                 pte_access &= ~ACC_WRITE_MASK;
1789         if (pte_access & ACC_EXEC_MASK)
1790                 spte |= shadow_x_mask;
1791         else
1792                 spte |= shadow_nx_mask;
1793         if (pte_access & ACC_USER_MASK)
1794                 spte |= shadow_user_mask;
1795         if (level > PT_PAGE_TABLE_LEVEL)
1796                 spte |= PT_PAGE_SIZE_MASK;
1797         if (tdp_enabled)
1798                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1799                         kvm_is_mmio_pfn(pfn));
1800
1801         if (reset_host_protection)
1802                 spte |= SPTE_HOST_WRITEABLE;
1803
1804         spte |= (u64)pfn << PAGE_SHIFT;
1805
1806         if ((pte_access & ACC_WRITE_MASK)
1807             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1808
1809                 if (level > PT_PAGE_TABLE_LEVEL &&
1810                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1811                         ret = 1;
1812                         spte = shadow_trap_nonpresent_pte;
1813                         goto set_pte;
1814                 }
1815
1816                 spte |= PT_WRITABLE_MASK;
1817
1818                 /*
1819                  * Optimization: for pte sync, if spte was writable the hash
1820                  * lookup is unnecessary (and expensive). Write protection
1821                  * is responsibility of mmu_get_page / kvm_sync_page.
1822                  * Same reasoning can be applied to dirty page accounting.
1823                  */
1824                 if (!can_unsync && is_writable_pte(*sptep))
1825                         goto set_pte;
1826
1827                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1828                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1829                                  __func__, gfn);
1830                         ret = 1;
1831                         pte_access &= ~ACC_WRITE_MASK;
1832                         if (is_writable_pte(spte))
1833                                 spte &= ~PT_WRITABLE_MASK;
1834                 }
1835         }
1836
1837         if (pte_access & ACC_WRITE_MASK)
1838                 mark_page_dirty(vcpu->kvm, gfn);
1839
1840 set_pte:
1841         __set_spte(sptep, spte);
1842         return ret;
1843 }
1844
1845 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1846                          unsigned pt_access, unsigned pte_access,
1847                          int user_fault, int write_fault, int dirty,
1848                          int *ptwrite, int level, gfn_t gfn,
1849                          pfn_t pfn, bool speculative,
1850                          bool reset_host_protection)
1851 {
1852         int was_rmapped = 0;
1853         int was_writable = is_writable_pte(*sptep);
1854         int rmap_count;
1855
1856         pgprintk("%s: spte %llx access %x write_fault %d"
1857                  " user_fault %d gfn %lx\n",
1858                  __func__, *sptep, pt_access,
1859                  write_fault, user_fault, gfn);
1860
1861         if (is_rmap_spte(*sptep)) {
1862                 /*
1863                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1864                  * the parent of the now unreachable PTE.
1865                  */
1866                 if (level > PT_PAGE_TABLE_LEVEL &&
1867                     !is_large_pte(*sptep)) {
1868                         struct kvm_mmu_page *child;
1869                         u64 pte = *sptep;
1870
1871                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1872                         mmu_page_remove_parent_pte(child, sptep);
1873                 } else if (pfn != spte_to_pfn(*sptep)) {
1874                         pgprintk("hfn old %lx new %lx\n",
1875                                  spte_to_pfn(*sptep), pfn);
1876                         rmap_remove(vcpu->kvm, sptep);
1877                 } else
1878                         was_rmapped = 1;
1879         }
1880
1881         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1882                       dirty, level, gfn, pfn, speculative, true,
1883                       reset_host_protection)) {
1884                 if (write_fault)
1885                         *ptwrite = 1;
1886                 kvm_x86_ops->tlb_flush(vcpu);
1887         }
1888
1889         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1890         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1891                  is_large_pte(*sptep)? "2MB" : "4kB",
1892                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1893                  *sptep, sptep);
1894         if (!was_rmapped && is_large_pte(*sptep))
1895                 ++vcpu->kvm->stat.lpages;
1896
1897         page_header_update_slot(vcpu->kvm, sptep, gfn);
1898         if (!was_rmapped) {
1899                 rmap_count = rmap_add(vcpu, sptep, gfn);
1900                 kvm_release_pfn_clean(pfn);
1901                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1902                         rmap_recycle(vcpu, sptep, gfn);
1903         } else {
1904                 if (was_writable)
1905                         kvm_release_pfn_dirty(pfn);
1906                 else
1907                         kvm_release_pfn_clean(pfn);
1908         }
1909         if (speculative) {
1910                 vcpu->arch.last_pte_updated = sptep;
1911                 vcpu->arch.last_pte_gfn = gfn;
1912         }
1913 }
1914
1915 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1916 {
1917 }
1918
1919 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1920                         int level, gfn_t gfn, pfn_t pfn)
1921 {
1922         struct kvm_shadow_walk_iterator iterator;
1923         struct kvm_mmu_page *sp;
1924         int pt_write = 0;
1925         gfn_t pseudo_gfn;
1926
1927         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1928                 if (iterator.level == level) {
1929                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1930                                      0, write, 1, &pt_write,
1931                                      level, gfn, pfn, false, true);
1932                         ++vcpu->stat.pf_fixed;
1933                         break;
1934                 }
1935
1936                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1937                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1938                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1939                                               iterator.level - 1,
1940                                               1, ACC_ALL, iterator.sptep);
1941                         if (!sp) {
1942                                 pgprintk("nonpaging_map: ENOMEM\n");
1943                                 kvm_release_pfn_clean(pfn);
1944                                 return -ENOMEM;
1945                         }
1946
1947                         __set_spte(iterator.sptep,
1948                                    __pa(sp->spt)
1949                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1950                                    | shadow_user_mask | shadow_x_mask);
1951                 }
1952         }
1953         return pt_write;
1954 }
1955
1956 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1957 {
1958         int r;
1959         int level;
1960         pfn_t pfn;
1961         unsigned long mmu_seq;
1962
1963         level = mapping_level(vcpu, gfn);
1964
1965         /*
1966          * This path builds a PAE pagetable - so we can map 2mb pages at
1967          * maximum. Therefore check if the level is larger than that.
1968          */
1969         if (level > PT_DIRECTORY_LEVEL)
1970                 level = PT_DIRECTORY_LEVEL;
1971
1972         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
1973
1974         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1975         smp_rmb();
1976         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1977
1978         /* mmio */
1979         if (is_error_pfn(pfn)) {
1980                 kvm_release_pfn_clean(pfn);
1981                 return 1;
1982         }
1983
1984         spin_lock(&vcpu->kvm->mmu_lock);
1985         if (mmu_notifier_retry(vcpu, mmu_seq))
1986                 goto out_unlock;
1987         kvm_mmu_free_some_pages(vcpu);
1988         r = __direct_map(vcpu, v, write, level, gfn, pfn);
1989         spin_unlock(&vcpu->kvm->mmu_lock);
1990
1991
1992         return r;
1993
1994 out_unlock:
1995         spin_unlock(&vcpu->kvm->mmu_lock);
1996         kvm_release_pfn_clean(pfn);
1997         return 0;
1998 }
1999
2000
2001 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2002 {
2003         int i;
2004         struct kvm_mmu_page *sp;
2005
2006         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2007                 return;
2008         spin_lock(&vcpu->kvm->mmu_lock);
2009         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2010                 hpa_t root = vcpu->arch.mmu.root_hpa;
2011
2012                 sp = page_header(root);
2013                 --sp->root_count;
2014                 if (!sp->root_count && sp->role.invalid)
2015                         kvm_mmu_zap_page(vcpu->kvm, sp);
2016                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2017                 spin_unlock(&vcpu->kvm->mmu_lock);
2018                 return;
2019         }
2020         for (i = 0; i < 4; ++i) {
2021                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2022
2023                 if (root) {
2024                         root &= PT64_BASE_ADDR_MASK;
2025                         sp = page_header(root);
2026                         --sp->root_count;
2027                         if (!sp->root_count && sp->role.invalid)
2028                                 kvm_mmu_zap_page(vcpu->kvm, sp);
2029                 }
2030                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2031         }
2032         spin_unlock(&vcpu->kvm->mmu_lock);
2033         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2034 }
2035
2036 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2037 {
2038         int ret = 0;
2039
2040         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2041                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2042                 ret = 1;
2043         }
2044
2045         return ret;
2046 }
2047
2048 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2049 {
2050         int i;
2051         gfn_t root_gfn;
2052         struct kvm_mmu_page *sp;
2053         int direct = 0;
2054         u64 pdptr;
2055
2056         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2057
2058         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2059                 hpa_t root = vcpu->arch.mmu.root_hpa;
2060
2061                 ASSERT(!VALID_PAGE(root));
2062                 if (mmu_check_root(vcpu, root_gfn))
2063                         return 1;
2064                 if (tdp_enabled) {
2065                         direct = 1;
2066                         root_gfn = 0;
2067                 }
2068                 spin_lock(&vcpu->kvm->mmu_lock);
2069                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2070                                       PT64_ROOT_LEVEL, direct,
2071                                       ACC_ALL, NULL);
2072                 root = __pa(sp->spt);
2073                 ++sp->root_count;
2074                 spin_unlock(&vcpu->kvm->mmu_lock);
2075                 vcpu->arch.mmu.root_hpa = root;
2076                 return 0;
2077         }
2078         direct = !is_paging(vcpu);
2079         for (i = 0; i < 4; ++i) {
2080                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2081
2082                 ASSERT(!VALID_PAGE(root));
2083                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2084                         pdptr = kvm_pdptr_read(vcpu, i);
2085                         if (!is_present_gpte(pdptr)) {
2086                                 vcpu->arch.mmu.pae_root[i] = 0;
2087                                 continue;
2088                         }
2089                         root_gfn = pdptr >> PAGE_SHIFT;
2090                 } else if (vcpu->arch.mmu.root_level == 0)
2091                         root_gfn = 0;
2092                 if (mmu_check_root(vcpu, root_gfn))
2093                         return 1;
2094                 if (tdp_enabled) {
2095                         direct = 1;
2096                         root_gfn = i << 30;
2097                 }
2098                 spin_lock(&vcpu->kvm->mmu_lock);
2099                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2100                                       PT32_ROOT_LEVEL, direct,
2101                                       ACC_ALL, NULL);
2102                 root = __pa(sp->spt);
2103                 ++sp->root_count;
2104                 spin_unlock(&vcpu->kvm->mmu_lock);
2105
2106                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2107         }
2108         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2109         return 0;
2110 }
2111
2112 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2113 {
2114         int i;
2115         struct kvm_mmu_page *sp;
2116
2117         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2118                 return;
2119         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2120                 hpa_t root = vcpu->arch.mmu.root_hpa;
2121                 sp = page_header(root);
2122                 mmu_sync_children(vcpu, sp);
2123                 return;
2124         }
2125         for (i = 0; i < 4; ++i) {
2126                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2127
2128                 if (root && VALID_PAGE(root)) {
2129                         root &= PT64_BASE_ADDR_MASK;
2130                         sp = page_header(root);
2131                         mmu_sync_children(vcpu, sp);
2132                 }
2133         }
2134 }
2135
2136 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2137 {
2138         spin_lock(&vcpu->kvm->mmu_lock);
2139         mmu_sync_roots(vcpu);
2140         spin_unlock(&vcpu->kvm->mmu_lock);
2141 }
2142
2143 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2144                                   u32 access, u32 *error)
2145 {
2146         if (error)
2147                 *error = 0;
2148         return vaddr;
2149 }
2150
2151 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2152                                 u32 error_code)
2153 {
2154         gfn_t gfn;
2155         int r;
2156
2157         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2158         r = mmu_topup_memory_caches(vcpu);
2159         if (r)
2160                 return r;
2161
2162         ASSERT(vcpu);
2163         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2164
2165         gfn = gva >> PAGE_SHIFT;
2166
2167         return nonpaging_map(vcpu, gva & PAGE_MASK,
2168                              error_code & PFERR_WRITE_MASK, gfn);
2169 }
2170
2171 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2172                                 u32 error_code)
2173 {
2174         pfn_t pfn;
2175         int r;
2176         int level;
2177         gfn_t gfn = gpa >> PAGE_SHIFT;
2178         unsigned long mmu_seq;
2179
2180         ASSERT(vcpu);
2181         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2182
2183         r = mmu_topup_memory_caches(vcpu);
2184         if (r)
2185                 return r;
2186
2187         level = mapping_level(vcpu, gfn);
2188
2189         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2190
2191         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2192         smp_rmb();
2193         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2194         if (is_error_pfn(pfn)) {
2195                 kvm_release_pfn_clean(pfn);
2196                 return 1;
2197         }
2198         spin_lock(&vcpu->kvm->mmu_lock);
2199         if (mmu_notifier_retry(vcpu, mmu_seq))
2200                 goto out_unlock;
2201         kvm_mmu_free_some_pages(vcpu);
2202         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2203                          level, gfn, pfn);
2204         spin_unlock(&vcpu->kvm->mmu_lock);
2205
2206         return r;
2207
2208 out_unlock:
2209         spin_unlock(&vcpu->kvm->mmu_lock);
2210         kvm_release_pfn_clean(pfn);
2211         return 0;
2212 }
2213
2214 static void nonpaging_free(struct kvm_vcpu *vcpu)
2215 {
2216         mmu_free_roots(vcpu);
2217 }
2218
2219 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2220 {
2221         struct kvm_mmu *context = &vcpu->arch.mmu;
2222
2223         context->new_cr3 = nonpaging_new_cr3;
2224         context->page_fault = nonpaging_page_fault;
2225         context->gva_to_gpa = nonpaging_gva_to_gpa;
2226         context->free = nonpaging_free;
2227         context->prefetch_page = nonpaging_prefetch_page;
2228         context->sync_page = nonpaging_sync_page;
2229         context->invlpg = nonpaging_invlpg;
2230         context->root_level = 0;
2231         context->shadow_root_level = PT32E_ROOT_LEVEL;
2232         context->root_hpa = INVALID_PAGE;
2233         return 0;
2234 }
2235
2236 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2237 {
2238         ++vcpu->stat.tlb_flush;
2239         kvm_x86_ops->tlb_flush(vcpu);
2240 }
2241
2242 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2243 {
2244         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2245         mmu_free_roots(vcpu);
2246 }
2247
2248 static void inject_page_fault(struct kvm_vcpu *vcpu,
2249                               u64 addr,
2250                               u32 err_code)
2251 {
2252         kvm_inject_page_fault(vcpu, addr, err_code);
2253 }
2254
2255 static void paging_free(struct kvm_vcpu *vcpu)
2256 {
2257         nonpaging_free(vcpu);
2258 }
2259
2260 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2261 {
2262         int bit7;
2263
2264         bit7 = (gpte >> 7) & 1;
2265         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2266 }
2267
2268 #define PTTYPE 64
2269 #include "paging_tmpl.h"
2270 #undef PTTYPE
2271
2272 #define PTTYPE 32
2273 #include "paging_tmpl.h"
2274 #undef PTTYPE
2275
2276 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2277 {
2278         struct kvm_mmu *context = &vcpu->arch.mmu;
2279         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2280         u64 exb_bit_rsvd = 0;
2281
2282         if (!is_nx(vcpu))
2283                 exb_bit_rsvd = rsvd_bits(63, 63);
2284         switch (level) {
2285         case PT32_ROOT_LEVEL:
2286                 /* no rsvd bits for 2 level 4K page table entries */
2287                 context->rsvd_bits_mask[0][1] = 0;
2288                 context->rsvd_bits_mask[0][0] = 0;
2289                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2290
2291                 if (!is_pse(vcpu)) {
2292                         context->rsvd_bits_mask[1][1] = 0;
2293                         break;
2294                 }
2295
2296                 if (is_cpuid_PSE36())
2297                         /* 36bits PSE 4MB page */
2298                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2299                 else
2300                         /* 32 bits PSE 4MB page */
2301                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2302                 break;
2303         case PT32E_ROOT_LEVEL:
2304                 context->rsvd_bits_mask[0][2] =
2305                         rsvd_bits(maxphyaddr, 63) |
2306                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2307                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2308                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2309                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2310                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2311                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2312                         rsvd_bits(maxphyaddr, 62) |
2313                         rsvd_bits(13, 20);              /* large page */
2314                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2315                 break;
2316         case PT64_ROOT_LEVEL:
2317                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2318                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2319                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2320                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2321                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2322                         rsvd_bits(maxphyaddr, 51);
2323                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2324                         rsvd_bits(maxphyaddr, 51);
2325                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2326                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2327                         rsvd_bits(maxphyaddr, 51) |
2328                         rsvd_bits(13, 29);
2329                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2330                         rsvd_bits(maxphyaddr, 51) |
2331                         rsvd_bits(13, 20);              /* large page */
2332                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2333                 break;
2334         }
2335 }
2336
2337 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2338 {
2339         struct kvm_mmu *context = &vcpu->arch.mmu;
2340
2341         ASSERT(is_pae(vcpu));
2342         context->new_cr3 = paging_new_cr3;
2343         context->page_fault = paging64_page_fault;
2344         context->gva_to_gpa = paging64_gva_to_gpa;
2345         context->prefetch_page = paging64_prefetch_page;
2346         context->sync_page = paging64_sync_page;
2347         context->invlpg = paging64_invlpg;
2348         context->free = paging_free;
2349         context->root_level = level;
2350         context->shadow_root_level = level;
2351         context->root_hpa = INVALID_PAGE;
2352         return 0;
2353 }
2354
2355 static int paging64_init_context(struct kvm_vcpu *vcpu)
2356 {
2357         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2358         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2359 }
2360
2361 static int paging32_init_context(struct kvm_vcpu *vcpu)
2362 {
2363         struct kvm_mmu *context = &vcpu->arch.mmu;
2364
2365         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2366         context->new_cr3 = paging_new_cr3;
2367         context->page_fault = paging32_page_fault;
2368         context->gva_to_gpa = paging32_gva_to_gpa;
2369         context->free = paging_free;
2370         context->prefetch_page = paging32_prefetch_page;
2371         context->sync_page = paging32_sync_page;
2372         context->invlpg = paging32_invlpg;
2373         context->root_level = PT32_ROOT_LEVEL;
2374         context->shadow_root_level = PT32E_ROOT_LEVEL;
2375         context->root_hpa = INVALID_PAGE;
2376         return 0;
2377 }
2378
2379 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2380 {
2381         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2382         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2383 }
2384
2385 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2386 {
2387         struct kvm_mmu *context = &vcpu->arch.mmu;
2388
2389         context->new_cr3 = nonpaging_new_cr3;
2390         context->page_fault = tdp_page_fault;
2391         context->free = nonpaging_free;
2392         context->prefetch_page = nonpaging_prefetch_page;
2393         context->sync_page = nonpaging_sync_page;
2394         context->invlpg = nonpaging_invlpg;
2395         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2396         context->root_hpa = INVALID_PAGE;
2397
2398         if (!is_paging(vcpu)) {
2399                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2400                 context->root_level = 0;
2401         } else if (is_long_mode(vcpu)) {
2402                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2403                 context->gva_to_gpa = paging64_gva_to_gpa;
2404                 context->root_level = PT64_ROOT_LEVEL;
2405         } else if (is_pae(vcpu)) {
2406                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2407                 context->gva_to_gpa = paging64_gva_to_gpa;
2408                 context->root_level = PT32E_ROOT_LEVEL;
2409         } else {
2410                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2411                 context->gva_to_gpa = paging32_gva_to_gpa;
2412                 context->root_level = PT32_ROOT_LEVEL;
2413         }
2414
2415         return 0;
2416 }
2417
2418 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2419 {
2420         int r;
2421
2422         ASSERT(vcpu);
2423         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2424
2425         if (!is_paging(vcpu))
2426                 r = nonpaging_init_context(vcpu);
2427         else if (is_long_mode(vcpu))
2428                 r = paging64_init_context(vcpu);
2429         else if (is_pae(vcpu))
2430                 r = paging32E_init_context(vcpu);
2431         else
2432                 r = paging32_init_context(vcpu);
2433
2434         vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
2435         vcpu->arch.mmu.base_role.cr0_wp = is_write_protection(vcpu);
2436
2437         return r;
2438 }
2439
2440 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2441 {
2442         vcpu->arch.update_pte.pfn = bad_pfn;
2443
2444         if (tdp_enabled)
2445                 return init_kvm_tdp_mmu(vcpu);
2446         else
2447                 return init_kvm_softmmu(vcpu);
2448 }
2449
2450 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2451 {
2452         ASSERT(vcpu);
2453         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2454                 vcpu->arch.mmu.free(vcpu);
2455                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2456         }
2457 }
2458
2459 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2460 {
2461         destroy_kvm_mmu(vcpu);
2462         return init_kvm_mmu(vcpu);
2463 }
2464 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2465
2466 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2467 {
2468         int r;
2469
2470         r = mmu_topup_memory_caches(vcpu);
2471         if (r)
2472                 goto out;
2473         spin_lock(&vcpu->kvm->mmu_lock);
2474         kvm_mmu_free_some_pages(vcpu);
2475         spin_unlock(&vcpu->kvm->mmu_lock);
2476         r = mmu_alloc_roots(vcpu);
2477         spin_lock(&vcpu->kvm->mmu_lock);
2478         mmu_sync_roots(vcpu);
2479         spin_unlock(&vcpu->kvm->mmu_lock);
2480         if (r)
2481                 goto out;
2482         /* set_cr3() should ensure TLB has been flushed */
2483         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2484 out:
2485         return r;
2486 }
2487 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2488
2489 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2490 {
2491         mmu_free_roots(vcpu);
2492 }
2493
2494 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2495                                   struct kvm_mmu_page *sp,
2496                                   u64 *spte)
2497 {
2498         u64 pte;
2499         struct kvm_mmu_page *child;
2500
2501         pte = *spte;
2502         if (is_shadow_present_pte(pte)) {
2503                 if (is_last_spte(pte, sp->role.level))
2504                         rmap_remove(vcpu->kvm, spte);
2505                 else {
2506                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2507                         mmu_page_remove_parent_pte(child, spte);
2508                 }
2509         }
2510         __set_spte(spte, shadow_trap_nonpresent_pte);
2511         if (is_large_pte(pte))
2512                 --vcpu->kvm->stat.lpages;
2513 }
2514
2515 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2516                                   struct kvm_mmu_page *sp,
2517                                   u64 *spte,
2518                                   const void *new)
2519 {
2520         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2521                 ++vcpu->kvm->stat.mmu_pde_zapped;
2522                 return;
2523         }
2524
2525         ++vcpu->kvm->stat.mmu_pte_updated;
2526         if (!sp->role.cr4_pae)
2527                 paging32_update_pte(vcpu, sp, spte, new);
2528         else
2529                 paging64_update_pte(vcpu, sp, spte, new);
2530 }
2531
2532 static bool need_remote_flush(u64 old, u64 new)
2533 {
2534         if (!is_shadow_present_pte(old))
2535                 return false;
2536         if (!is_shadow_present_pte(new))
2537                 return true;
2538         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2539                 return true;
2540         old ^= PT64_NX_MASK;
2541         new ^= PT64_NX_MASK;
2542         return (old & ~new & PT64_PERM_MASK) != 0;
2543 }
2544
2545 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2546 {
2547         if (need_remote_flush(old, new))
2548                 kvm_flush_remote_tlbs(vcpu->kvm);
2549         else
2550                 kvm_mmu_flush_tlb(vcpu);
2551 }
2552
2553 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2554 {
2555         u64 *spte = vcpu->arch.last_pte_updated;
2556
2557         return !!(spte && (*spte & shadow_accessed_mask));
2558 }
2559
2560 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2561                                           u64 gpte)
2562 {
2563         gfn_t gfn;
2564         pfn_t pfn;
2565
2566         if (!is_present_gpte(gpte))
2567                 return;
2568         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2569
2570         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2571         smp_rmb();
2572         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2573
2574         if (is_error_pfn(pfn)) {
2575                 kvm_release_pfn_clean(pfn);
2576                 return;
2577         }
2578         vcpu->arch.update_pte.gfn = gfn;
2579         vcpu->arch.update_pte.pfn = pfn;
2580 }
2581
2582 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2583 {
2584         u64 *spte = vcpu->arch.last_pte_updated;
2585
2586         if (spte
2587             && vcpu->arch.last_pte_gfn == gfn
2588             && shadow_accessed_mask
2589             && !(*spte & shadow_accessed_mask)
2590             && is_shadow_present_pte(*spte))
2591                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2592 }
2593
2594 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2595                        const u8 *new, int bytes,
2596                        bool guest_initiated)
2597 {
2598         gfn_t gfn = gpa >> PAGE_SHIFT;
2599         struct kvm_mmu_page *sp;
2600         struct hlist_node *node, *n;
2601         struct hlist_head *bucket;
2602         unsigned index;
2603         u64 entry, gentry;
2604         u64 *spte;
2605         unsigned offset = offset_in_page(gpa);
2606         unsigned pte_size;
2607         unsigned page_offset;
2608         unsigned misaligned;
2609         unsigned quadrant;
2610         int level;
2611         int flooded = 0;
2612         int npte;
2613         int r;
2614         int invlpg_counter;
2615
2616         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2617
2618         invlpg_counter = atomic_read(&vcpu->kvm->arch.invlpg_counter);
2619
2620         /*
2621          * Assume that the pte write on a page table of the same type
2622          * as the current vcpu paging mode.  This is nearly always true
2623          * (might be false while changing modes).  Note it is verified later
2624          * by update_pte().
2625          */
2626         if ((is_pae(vcpu) && bytes == 4) || !new) {
2627                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2628                 if (is_pae(vcpu)) {
2629                         gpa &= ~(gpa_t)7;
2630                         bytes = 8;
2631                 }
2632                 r = kvm_read_guest(vcpu->kvm, gpa, &gentry, min(bytes, 8));
2633                 if (r)
2634                         gentry = 0;
2635                 new = (const u8 *)&gentry;
2636         }
2637
2638         switch (bytes) {
2639         case 4:
2640                 gentry = *(const u32 *)new;
2641                 break;
2642         case 8:
2643                 gentry = *(const u64 *)new;
2644                 break;
2645         default:
2646                 gentry = 0;
2647                 break;
2648         }
2649
2650         mmu_guess_page_from_pte_write(vcpu, gpa, gentry);
2651         spin_lock(&vcpu->kvm->mmu_lock);
2652         if (atomic_read(&vcpu->kvm->arch.invlpg_counter) != invlpg_counter)
2653                 gentry = 0;
2654         kvm_mmu_access_page(vcpu, gfn);
2655         kvm_mmu_free_some_pages(vcpu);
2656         ++vcpu->kvm->stat.mmu_pte_write;
2657         kvm_mmu_audit(vcpu, "pre pte write");
2658         if (guest_initiated) {
2659                 if (gfn == vcpu->arch.last_pt_write_gfn
2660                     && !last_updated_pte_accessed(vcpu)) {
2661                         ++vcpu->arch.last_pt_write_count;
2662                         if (vcpu->arch.last_pt_write_count >= 3)
2663                                 flooded = 1;
2664                 } else {
2665                         vcpu->arch.last_pt_write_gfn = gfn;
2666                         vcpu->arch.last_pt_write_count = 1;
2667                         vcpu->arch.last_pte_updated = NULL;
2668                 }
2669         }
2670         index = kvm_page_table_hashfn(gfn);
2671         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2672
2673 restart:
2674         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2675                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2676                         continue;
2677                 pte_size = sp->role.cr4_pae ? 8 : 4;
2678                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2679                 misaligned |= bytes < 4;
2680                 if (misaligned || flooded) {
2681                         /*
2682                          * Misaligned accesses are too much trouble to fix
2683                          * up; also, they usually indicate a page is not used
2684                          * as a page table.
2685                          *
2686                          * If we're seeing too many writes to a page,
2687                          * it may no longer be a page table, or we may be
2688                          * forking, in which case it is better to unmap the
2689                          * page.
2690                          */
2691                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2692                                  gpa, bytes, sp->role.word);
2693                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2694                                 goto restart;
2695                         ++vcpu->kvm->stat.mmu_flooded;
2696                         continue;
2697                 }
2698                 page_offset = offset;
2699                 level = sp->role.level;
2700                 npte = 1;
2701                 if (!sp->role.cr4_pae) {
2702                         page_offset <<= 1;      /* 32->64 */
2703                         /*
2704                          * A 32-bit pde maps 4MB while the shadow pdes map
2705                          * only 2MB.  So we need to double the offset again
2706                          * and zap two pdes instead of one.
2707                          */
2708                         if (level == PT32_ROOT_LEVEL) {
2709                                 page_offset &= ~7; /* kill rounding error */
2710                                 page_offset <<= 1;
2711                                 npte = 2;
2712                         }
2713                         quadrant = page_offset >> PAGE_SHIFT;
2714                         page_offset &= ~PAGE_MASK;
2715                         if (quadrant != sp->role.quadrant)
2716                                 continue;
2717                 }
2718                 spte = &sp->spt[page_offset / sizeof(*spte)];
2719                 while (npte--) {
2720                         entry = *spte;
2721                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2722                         if (gentry)
2723                                 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
2724                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2725                         ++spte;
2726                 }
2727         }
2728         kvm_mmu_audit(vcpu, "post pte write");
2729         spin_unlock(&vcpu->kvm->mmu_lock);
2730         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2731                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2732                 vcpu->arch.update_pte.pfn = bad_pfn;
2733         }
2734 }
2735
2736 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2737 {
2738         gpa_t gpa;
2739         int r;
2740
2741         if (tdp_enabled)
2742                 return 0;
2743
2744         gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2745
2746         spin_lock(&vcpu->kvm->mmu_lock);
2747         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2748         spin_unlock(&vcpu->kvm->mmu_lock);
2749         return r;
2750 }
2751 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2752
2753 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2754 {
2755         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2756                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2757                 struct kvm_mmu_page *sp;
2758
2759                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2760                                   struct kvm_mmu_page, link);
2761                 kvm_mmu_zap_page(vcpu->kvm, sp);
2762                 ++vcpu->kvm->stat.mmu_recycled;
2763         }
2764 }
2765
2766 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2767 {
2768         int r;
2769         enum emulation_result er;
2770
2771         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2772         if (r < 0)
2773                 goto out;
2774
2775         if (!r) {
2776                 r = 1;
2777                 goto out;
2778         }
2779
2780         r = mmu_topup_memory_caches(vcpu);
2781         if (r)
2782                 goto out;
2783
2784         er = emulate_instruction(vcpu, cr2, error_code, 0);
2785
2786         switch (er) {
2787         case EMULATE_DONE:
2788                 return 1;
2789         case EMULATE_DO_MMIO:
2790                 ++vcpu->stat.mmio_exits;
2791                 return 0;
2792         case EMULATE_FAIL:
2793                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2794                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2795                 vcpu->run->internal.ndata = 0;
2796                 return 0;
2797         default:
2798                 BUG();
2799         }
2800 out:
2801         return r;
2802 }
2803 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2804
2805 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2806 {
2807         vcpu->arch.mmu.invlpg(vcpu, gva);
2808         kvm_mmu_flush_tlb(vcpu);
2809         ++vcpu->stat.invlpg;
2810 }
2811 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2812
2813 void kvm_enable_tdp(void)
2814 {
2815         tdp_enabled = true;
2816 }
2817 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2818
2819 void kvm_disable_tdp(void)
2820 {
2821         tdp_enabled = false;
2822 }
2823 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2824
2825 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2826 {
2827         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2828 }
2829
2830 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2831 {
2832         struct page *page;
2833         int i;
2834
2835         ASSERT(vcpu);
2836
2837         /*
2838          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2839          * Therefore we need to allocate shadow page tables in the first
2840          * 4GB of memory, which happens to fit the DMA32 zone.
2841          */
2842         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2843         if (!page)
2844                 return -ENOMEM;
2845
2846         vcpu->arch.mmu.pae_root = page_address(page);
2847         for (i = 0; i < 4; ++i)
2848                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2849
2850         return 0;
2851 }
2852
2853 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2854 {
2855         ASSERT(vcpu);
2856         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2857
2858         return alloc_mmu_pages(vcpu);
2859 }
2860
2861 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2862 {
2863         ASSERT(vcpu);
2864         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2865
2866         return init_kvm_mmu(vcpu);
2867 }
2868
2869 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2870 {
2871         ASSERT(vcpu);
2872
2873         destroy_kvm_mmu(vcpu);
2874         free_mmu_pages(vcpu);
2875         mmu_free_memory_caches(vcpu);
2876 }
2877
2878 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2879 {
2880         struct kvm_mmu_page *sp;
2881
2882         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2883                 int i;
2884                 u64 *pt;
2885
2886                 if (!test_bit(slot, sp->slot_bitmap))
2887                         continue;
2888
2889                 pt = sp->spt;
2890                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2891                         /* avoid RMW */
2892                         if (pt[i] & PT_WRITABLE_MASK)
2893                                 pt[i] &= ~PT_WRITABLE_MASK;
2894         }
2895         kvm_flush_remote_tlbs(kvm);
2896 }
2897
2898 void kvm_mmu_zap_all(struct kvm *kvm)
2899 {
2900         struct kvm_mmu_page *sp, *node;
2901
2902         spin_lock(&kvm->mmu_lock);
2903 restart:
2904         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2905                 if (kvm_mmu_zap_page(kvm, sp))
2906                         goto restart;
2907
2908         spin_unlock(&kvm->mmu_lock);
2909
2910         kvm_flush_remote_tlbs(kvm);
2911 }
2912
2913 static int kvm_mmu_remove_some_alloc_mmu_pages(struct kvm *kvm)
2914 {
2915         struct kvm_mmu_page *page;
2916
2917         page = container_of(kvm->arch.active_mmu_pages.prev,
2918                             struct kvm_mmu_page, link);
2919         return kvm_mmu_zap_page(kvm, page) + 1;
2920 }
2921
2922 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2923 {
2924         struct kvm *kvm;
2925         struct kvm *kvm_freed = NULL;
2926         int cache_count = 0;
2927
2928         spin_lock(&kvm_lock);
2929
2930         list_for_each_entry(kvm, &vm_list, vm_list) {
2931                 int npages, idx, freed_pages;
2932
2933                 idx = srcu_read_lock(&kvm->srcu);
2934                 spin_lock(&kvm->mmu_lock);
2935                 npages = kvm->arch.n_alloc_mmu_pages -
2936                          kvm->arch.n_free_mmu_pages;
2937                 cache_count += npages;
2938                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2939                         freed_pages = kvm_mmu_remove_some_alloc_mmu_pages(kvm);
2940                         cache_count -= freed_pages;
2941                         kvm_freed = kvm;
2942                 }
2943                 nr_to_scan--;
2944
2945                 spin_unlock(&kvm->mmu_lock);
2946                 srcu_read_unlock(&kvm->srcu, idx);
2947         }
2948         if (kvm_freed)
2949                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2950
2951         spin_unlock(&kvm_lock);
2952
2953         return cache_count;
2954 }
2955
2956 static struct shrinker mmu_shrinker = {
2957         .shrink = mmu_shrink,
2958         .seeks = DEFAULT_SEEKS * 10,
2959 };
2960
2961 static void mmu_destroy_caches(void)
2962 {
2963         if (pte_chain_cache)
2964                 kmem_cache_destroy(pte_chain_cache);
2965         if (rmap_desc_cache)
2966                 kmem_cache_destroy(rmap_desc_cache);
2967         if (mmu_page_header_cache)
2968                 kmem_cache_destroy(mmu_page_header_cache);
2969 }
2970
2971 void kvm_mmu_module_exit(void)
2972 {
2973         mmu_destroy_caches();
2974         unregister_shrinker(&mmu_shrinker);
2975 }
2976
2977 int kvm_mmu_module_init(void)
2978 {
2979         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2980                                             sizeof(struct kvm_pte_chain),
2981                                             0, 0, NULL);
2982         if (!pte_chain_cache)
2983                 goto nomem;
2984         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2985                                             sizeof(struct kvm_rmap_desc),
2986                                             0, 0, NULL);
2987         if (!rmap_desc_cache)
2988                 goto nomem;
2989
2990         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2991                                                   sizeof(struct kvm_mmu_page),
2992                                                   0, 0, NULL);
2993         if (!mmu_page_header_cache)
2994                 goto nomem;
2995
2996         register_shrinker(&mmu_shrinker);
2997
2998         return 0;
2999
3000 nomem:
3001         mmu_destroy_caches();
3002         return -ENOMEM;
3003 }
3004
3005 /*
3006  * Caculate mmu pages needed for kvm.
3007  */
3008 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3009 {
3010         int i;
3011         unsigned int nr_mmu_pages;
3012         unsigned int  nr_pages = 0;
3013         struct kvm_memslots *slots;
3014
3015         slots = kvm_memslots(kvm);
3016
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 = kvm_memslots(kvm);
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