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