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

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "mmu.h"
  21 #include "x86.h"
  22 #include "kvm_cache_regs.h"
  23
  24 #include <linux/kvm_host.h>
  25 #include <linux/types.h>
  26 #include <linux/string.h>
  27 #include <linux/mm.h>
  28 #include <linux/highmem.h>
  29 #include <linux/module.h>
  30 #include <linux/swap.h>
  31 #include <linux/hugetlb.h>
  32 #include <linux/compiler.h>
  33 #include <linux/srcu.h>
  34
  35 #include <asm/page.h>
  36 #include <asm/cmpxchg.h>
  37 #include <asm/io.h>
  38 #include <asm/vmx.h>
  39
  40 /*
  41  * When setting this variable to true it enables Two-Dimensional-Paging
  42  * where the hardware walks 2 page tables:
  43  * 1. the guest-virtual to guest-physical
  44  * 2. while doing 1. it walks guest-physical to host-physical
  45  * If the hardware supports that we don't need to do shadow paging.
  46  */
  47 bool tdp_enabled = false;
  48
  49 #undef MMU_DEBUG
  50
  51 #undef AUDIT
  52
  53 #ifdef AUDIT
  54 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  55 #else
  56 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  57 #endif
  58
  59 #ifdef MMU_DEBUG
  60
  61 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  62 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  63
  64 #else
  65
  66 #define pgprintk(x...) do { } while (0)
  67 #define rmap_printk(x...) do { } while (0)
  68
  69 #endif
  70
  71 #if defined(MMU_DEBUG) || defined(AUDIT)
  72 static int dbg = 0;
  73 module_param(dbg, bool, 0644);
  74 #endif
  75
  76 static int oos_shadow = 1;
  77 module_param(oos_shadow, bool, 0644);
  78
  79 #ifndef MMU_DEBUG
  80 #define ASSERT(x) do { } while (0)
  81 #else
  82 #define ASSERT(x)                                                       \
  83         if (!(x)) {                                                     \
  84                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  85                        __FILE__, __LINE__, #x);                         \
  86         }
  87 #endif
  88
  89 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  90 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  91
  92 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  93
  94 #define PT64_LEVEL_BITS 9
  95
  96 #define PT64_LEVEL_SHIFT(level) \
  97                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  98
  99 #define PT64_LEVEL_MASK(level) \
 100                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 101
 102 #define PT64_INDEX(address, level)\
 103         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 104
 105
 106 #define PT32_LEVEL_BITS 10
 107
 108 #define PT32_LEVEL_SHIFT(level) \
 109                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 110
 111 #define PT32_LEVEL_MASK(level) \
 112                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 113 #define PT32_LVL_OFFSET_MASK(level) \
 114         (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 115                                                 * PT32_LEVEL_BITS))) - 1))
 116
 117 #define PT32_INDEX(address, level)\
 118         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 119
 120
 121 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 122 #define PT64_DIR_BASE_ADDR_MASK \
 123         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 124 #define PT64_LVL_ADDR_MASK(level) \
 125         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 126                                                 * PT64_LEVEL_BITS))) - 1))
 127 #define PT64_LVL_OFFSET_MASK(level) \
 128         (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 129                                                 * PT64_LEVEL_BITS))) - 1))
 130
 131 #define PT32_BASE_ADDR_MASK PAGE_MASK
 132 #define PT32_DIR_BASE_ADDR_MASK \
 133         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 134 #define PT32_LVL_ADDR_MASK(level) \
 135         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 136                                             * PT32_LEVEL_BITS))) - 1))
 137
 138 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 139                         | PT64_NX_MASK)
 140
 141 #define PFERR_PRESENT_MASK (1U << 0)
 142 #define PFERR_WRITE_MASK (1U << 1)
 143 #define PFERR_USER_MASK (1U << 2)
 144 #define PFERR_RSVD_MASK (1U << 3)
 145 #define PFERR_FETCH_MASK (1U << 4)
 146
 147 #define RMAP_EXT 4
 148
 149 #define ACC_EXEC_MASK    1
 150 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 151 #define ACC_USER_MASK    PT_USER_MASK
 152 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 153
 154 #define CREATE_TRACE_POINTS
 155 #include "mmutrace.h"
 156
 157 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 158
 159 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 160
 161 struct kvm_rmap_desc {
 162         u64 *sptes[RMAP_EXT];
 163         struct kvm_rmap_desc *more;
 164 };
 165
 166 struct kvm_shadow_walk_iterator {
 167         u64 addr;
 168         hpa_t shadow_addr;
 169         int level;
 170         u64 *sptep;
 171         unsigned index;
 172 };
 173
 174 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
 175         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
 176              shadow_walk_okay(&(_walker));                      \
 177              shadow_walk_next(&(_walker)))
 178
 179
 180 struct kvm_unsync_walk {
 181         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 182 };
 183
 184 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 185
 186 static struct kmem_cache *pte_chain_cache;
 187 static struct kmem_cache *rmap_desc_cache;
 188 static struct kmem_cache *mmu_page_header_cache;
 189
 190 static u64 __read_mostly shadow_trap_nonpresent_pte;
 191 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 192 static u64 __read_mostly shadow_base_present_pte;
 193 static u64 __read_mostly shadow_nx_mask;
 194 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 195 static u64 __read_mostly shadow_user_mask;
 196 static u64 __read_mostly shadow_accessed_mask;
 197 static u64 __read_mostly shadow_dirty_mask;
 198
 199 static inline u64 rsvd_bits(int s, int e)
 200 {
 201         return ((1ULL << (e - s + 1)) - 1) << s;
 202 }
 203
 204 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 205 {
 206         shadow_trap_nonpresent_pte = trap_pte;
 207         shadow_notrap_nonpresent_pte = notrap_pte;
 208 }
 209 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 210
 211 void kvm_mmu_set_base_ptes(u64 base_pte)
 212 {
 213         shadow_base_present_pte = base_pte;
 214 }
 215 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 216
 217 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 218                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 219 {
 220         shadow_user_mask = user_mask;
 221         shadow_accessed_mask = accessed_mask;
 222         shadow_dirty_mask = dirty_mask;
 223         shadow_nx_mask = nx_mask;
 224         shadow_x_mask = x_mask;
 225 }
 226 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 227
 228 static int is_write_protection(struct kvm_vcpu *vcpu)
 229 {
 230         return kvm_read_cr0_bits(vcpu, X86_CR0_WP);
 231 }
 232
 233 static int is_cpuid_PSE36(void)
 234 {
 235         return 1;
 236 }
 237
 238 static int is_nx(struct kvm_vcpu *vcpu)
 239 {
 240         return vcpu->arch.efer & EFER_NX;
 241 }
 242
 243 static int is_shadow_present_pte(u64 pte)
 244 {
 245         return pte != shadow_trap_nonpresent_pte
 246                 && pte != shadow_notrap_nonpresent_pte;
 247 }
 248
 249 static int is_large_pte(u64 pte)
 250 {
 251         return pte & PT_PAGE_SIZE_MASK;
 252 }
 253
 254 static int is_writable_pte(unsigned long pte)
 255 {
 256         return pte & PT_WRITABLE_MASK;
 257 }
 258
 259 static int is_dirty_gpte(unsigned long pte)
 260 {
 261         return pte & PT_DIRTY_MASK;
 262 }
 263
 264 static int is_rmap_spte(u64 pte)
 265 {
 266         return is_shadow_present_pte(pte);
 267 }
 268
 269 static int is_last_spte(u64 pte, int level)
 270 {
 271         if (level == PT_PAGE_TABLE_LEVEL)
 272                 return 1;
 273         if (is_large_pte(pte))
 274                 return 1;
 275         return 0;
 276 }
 277
 278 static pfn_t spte_to_pfn(u64 pte)
 279 {
 280         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 281 }
 282
 283 static gfn_t pse36_gfn_delta(u32 gpte)
 284 {
 285         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 286
 287         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 288 }
 289
 290 static void __set_spte(u64 *sptep, u64 spte)
 291 {
 292 #ifdef CONFIG_X86_64
 293         set_64bit((unsigned long *)sptep, spte);
 294 #else
 295         set_64bit((unsigned long long *)sptep, spte);
 296 #endif
 297 }
 298
 299 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 300                                   struct kmem_cache *base_cache, int min)
 301 {
 302         void *obj;
 303
 304         if (cache->nobjs >= min)
 305                 return 0;
 306         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 307                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 308                 if (!obj)
 309                         return -ENOMEM;
 310                 cache->objects[cache->nobjs++] = obj;
 311         }
 312         return 0;
 313 }
 314
 315 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 316 {
 317         while (mc->nobjs)
 318                 kfree(mc->objects[--mc->nobjs]);
 319 }
 320
 321 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 322                                        int min)
 323 {
 324         struct page *page;
 325
 326         if (cache->nobjs >= min)
 327                 return 0;
 328         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 329                 page = alloc_page(GFP_KERNEL);
 330                 if (!page)
 331                         return -ENOMEM;
 332                 set_page_private(page, 0);
 333                 cache->objects[cache->nobjs++] = page_address(page);
 334         }
 335         return 0;
 336 }
 337
 338 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 339 {
 340         while (mc->nobjs)
 341                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 342 }
 343
 344 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 345 {
 346         int r;
 347
 348         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 349                                    pte_chain_cache, 4);
 350         if (r)
 351                 goto out;
 352         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 353                                    rmap_desc_cache, 4);
 354         if (r)
 355                 goto out;
 356         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 357         if (r)
 358                 goto out;
 359         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 360                                    mmu_page_header_cache, 4);
 361 out:
 362         return r;
 363 }
 364
 365 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 366 {
 367         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 368         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 369         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 370         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 371 }
 372
 373 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 374                                     size_t size)
 375 {
 376         void *p;
 377
 378         BUG_ON(!mc->nobjs);
 379         p = mc->objects[--mc->nobjs];
 380         return p;
 381 }
 382
 383 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 384 {
 385         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 386                                       sizeof(struct kvm_pte_chain));
 387 }
 388
 389 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 390 {
 391         kfree(pc);
 392 }
 393
 394 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 395 {
 396         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 397                                       sizeof(struct kvm_rmap_desc));
 398 }
 399
 400 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 401 {
 402         kfree(rd);
 403 }
 404
 405 /*
 406  * Return the pointer to the largepage write count for a given
 407  * gfn, handling slots that are not large page aligned.
 408  */
 409 static int *slot_largepage_idx(gfn_t gfn,
 410                                struct kvm_memory_slot *slot,
 411                                int level)
 412 {
 413         unsigned long idx;
 414
 415         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 416               (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 417         return &slot->lpage_info[level - 2][idx].write_count;
 418 }
 419
 420 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 421 {
 422         struct kvm_memory_slot *slot;
 423         int *write_count;
 424         int i;
 425
 426         gfn = unalias_gfn(kvm, gfn);
 427
 428         slot = gfn_to_memslot_unaliased(kvm, gfn);
 429         for (i = PT_DIRECTORY_LEVEL;
 430              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 431                 write_count   = slot_largepage_idx(gfn, slot, i);
 432                 *write_count += 1;
 433         }
 434 }
 435
 436 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 437 {
 438         struct kvm_memory_slot *slot;
 439         int *write_count;
 440         int i;
 441
 442         gfn = unalias_gfn(kvm, gfn);
 443         for (i = PT_DIRECTORY_LEVEL;
 444              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 445                 slot          = gfn_to_memslot_unaliased(kvm, gfn);
 446                 write_count   = slot_largepage_idx(gfn, slot, i);
 447                 *write_count -= 1;
 448                 WARN_ON(*write_count < 0);
 449         }
 450 }
 451
 452 static int has_wrprotected_page(struct kvm *kvm,
 453                                 gfn_t gfn,
 454                                 int level)
 455 {
 456         struct kvm_memory_slot *slot;
 457         int *largepage_idx;
 458
 459         gfn = unalias_gfn(kvm, gfn);
 460         slot = gfn_to_memslot_unaliased(kvm, gfn);
 461         if (slot) {
 462                 largepage_idx = slot_largepage_idx(gfn, slot, level);
 463                 return *largepage_idx;
 464         }
 465
 466         return 1;
 467 }
 468
 469 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
 470 {
 471         unsigned long page_size = PAGE_SIZE;
 472         struct vm_area_struct *vma;
 473         unsigned long addr;
 474         int i, ret = 0;
 475
 476         addr = gfn_to_hva(kvm, gfn);
 477         if (kvm_is_error_hva(addr))
 478                 return PT_PAGE_TABLE_LEVEL;
 479
 480         down_read(&current->mm->mmap_sem);
 481         vma = find_vma(current->mm, addr);
 482         if (!vma)
 483                 goto out;
 484
 485         page_size = vma_kernel_pagesize(vma);
 486
 487 out:
 488         up_read(&current->mm->mmap_sem);
 489
 490         for (i = PT_PAGE_TABLE_LEVEL;
 491              i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
 492                 if (page_size >= KVM_HPAGE_SIZE(i))
 493                         ret = i;
 494                 else
 495                         break;
 496         }
 497
 498         return ret;
 499 }
 500
 501 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 502 {
 503         struct kvm_memory_slot *slot;
 504         int host_level, level, max_level;
 505
 506         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 507         if (slot && slot->dirty_bitmap)
 508                 return PT_PAGE_TABLE_LEVEL;
 509
 510         host_level = host_mapping_level(vcpu->kvm, large_gfn);
 511
 512         if (host_level == PT_PAGE_TABLE_LEVEL)
 513                 return host_level;
 514
 515         max_level = kvm_x86_ops->get_lpage_level() < host_level ?
 516                 kvm_x86_ops->get_lpage_level() : host_level;
 517
 518         for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
 519                 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
 520                         break;
 521
 522         return level - 1;
 523 }
 524
 525 /*
 526  * Take gfn and return the reverse mapping to it.
 527  * Note: gfn must be unaliased before this function get called
 528  */
 529
 530 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
 531 {
 532         struct kvm_memory_slot *slot;
 533         unsigned long idx;
 534
 535         slot = gfn_to_memslot(kvm, gfn);
 536         if (likely(level == PT_PAGE_TABLE_LEVEL))
 537                 return &slot->rmap[gfn - slot->base_gfn];
 538
 539         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 540                 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 541
 542         return &slot->lpage_info[level - 2][idx].rmap_pde;
 543 }
 544
 545 /*
 546  * Reverse mapping data structures:
 547  *
 548  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 549  * that points to page_address(page).
 550  *
 551  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 552  * containing more mappings.
 553  *
 554  * Returns the number of rmap entries before the spte was added or zero if
 555  * the spte was not added.
 556  *
 557  */
 558 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 559 {
 560         struct kvm_mmu_page *sp;
 561         struct kvm_rmap_desc *desc;
 562         unsigned long *rmapp;
 563         int i, count = 0;
 564
 565         if (!is_rmap_spte(*spte))
 566                 return count;
 567         gfn = unalias_gfn(vcpu->kvm, gfn);
 568         sp = page_header(__pa(spte));
 569         sp->gfns[spte - sp->spt] = gfn;
 570         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 571         if (!*rmapp) {
 572                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 573                 *rmapp = (unsigned long)spte;
 574         } else if (!(*rmapp & 1)) {
 575                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 576                 desc = mmu_alloc_rmap_desc(vcpu);
 577                 desc->sptes[0] = (u64 *)*rmapp;
 578                 desc->sptes[1] = spte;
 579                 *rmapp = (unsigned long)desc | 1;
 580         } else {
 581                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 582                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 583                 while (desc->sptes[RMAP_EXT-1] && desc->more) {
 584                         desc = desc->more;
 585                         count += RMAP_EXT;
 586                 }
 587                 if (desc->sptes[RMAP_EXT-1]) {
 588                         desc->more = mmu_alloc_rmap_desc(vcpu);
 589                         desc = desc->more;
 590                 }
 591                 for (i = 0; desc->sptes[i]; ++i)
 592                         ;
 593                 desc->sptes[i] = spte;
 594         }
 595         return count;
 596 }
 597
 598 static void rmap_desc_remove_entry(unsigned long *rmapp,
 599                                    struct kvm_rmap_desc *desc,
 600                                    int i,
 601                                    struct kvm_rmap_desc *prev_desc)
 602 {
 603         int j;
 604
 605         for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
 606                 ;
 607         desc->sptes[i] = desc->sptes[j];
 608         desc->sptes[j] = NULL;
 609         if (j != 0)
 610                 return;
 611         if (!prev_desc && !desc->more)
 612                 *rmapp = (unsigned long)desc->sptes[0];
 613         else
 614                 if (prev_desc)
 615                         prev_desc->more = desc->more;
 616                 else
 617                         *rmapp = (unsigned long)desc->more | 1;
 618         mmu_free_rmap_desc(desc);
 619 }
 620
 621 static void rmap_remove(struct kvm *kvm, u64 *spte)
 622 {
 623         struct kvm_rmap_desc *desc;
 624         struct kvm_rmap_desc *prev_desc;
 625         struct kvm_mmu_page *sp;
 626         pfn_t pfn;
 627         unsigned long *rmapp;
 628         int i;
 629
 630         if (!is_rmap_spte(*spte))
 631                 return;
 632         sp = page_header(__pa(spte));
 633         pfn = spte_to_pfn(*spte);
 634         if (*spte & shadow_accessed_mask)
 635                 kvm_set_pfn_accessed(pfn);
 636         if (is_writable_pte(*spte))
 637                 kvm_set_pfn_dirty(pfn);
 638         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], sp->role.level);
 639         if (!*rmapp) {
 640                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 641                 BUG();
 642         } else if (!(*rmapp & 1)) {
 643                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 644                 if ((u64 *)*rmapp != spte) {
 645                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 646                                spte, *spte);
 647                         BUG();
 648                 }
 649                 *rmapp = 0;
 650         } else {
 651                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 652                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 653                 prev_desc = NULL;
 654                 while (desc) {
 655                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
 656                                 if (desc->sptes[i] == spte) {
 657                                         rmap_desc_remove_entry(rmapp,
 658                                                                desc, i,
 659                                                                prev_desc);
 660                                         return;
 661                                 }
 662                         prev_desc = desc;
 663                         desc = desc->more;
 664                 }
 665                 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
 666                 BUG();
 667         }
 668 }
 669
 670 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 671 {
 672         struct kvm_rmap_desc *desc;
 673         struct kvm_rmap_desc *prev_desc;
 674         u64 *prev_spte;
 675         int i;
 676
 677         if (!*rmapp)
 678                 return NULL;
 679         else if (!(*rmapp & 1)) {
 680                 if (!spte)
 681                         return (u64 *)*rmapp;
 682                 return NULL;
 683         }
 684         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 685         prev_desc = NULL;
 686         prev_spte = NULL;
 687         while (desc) {
 688                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
 689                         if (prev_spte == spte)
 690                                 return desc->sptes[i];
 691                         prev_spte = desc->sptes[i];
 692                 }
 693                 desc = desc->more;
 694         }
 695         return NULL;
 696 }
 697
 698 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 699 {
 700         unsigned long *rmapp;
 701         u64 *spte;
 702         int i, write_protected = 0;
 703
 704         gfn = unalias_gfn(kvm, gfn);
 705         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
 706
 707         spte = rmap_next(kvm, rmapp, NULL);
 708         while (spte) {
 709                 BUG_ON(!spte);
 710                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 711                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 712                 if (is_writable_pte(*spte)) {
 713                         __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
 714                         write_protected = 1;
 715                 }
 716                 spte = rmap_next(kvm, rmapp, spte);
 717         }
 718         if (write_protected) {
 719                 pfn_t pfn;
 720
 721                 spte = rmap_next(kvm, rmapp, NULL);
 722                 pfn = spte_to_pfn(*spte);
 723                 kvm_set_pfn_dirty(pfn);
 724         }
 725
 726         /* check for huge page mappings */
 727         for (i = PT_DIRECTORY_LEVEL;
 728              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 729                 rmapp = gfn_to_rmap(kvm, gfn, i);
 730                 spte = rmap_next(kvm, rmapp, NULL);
 731                 while (spte) {
 732                         BUG_ON(!spte);
 733                         BUG_ON(!(*spte & PT_PRESENT_MASK));
 734                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 735                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 736                         if (is_writable_pte(*spte)) {
 737                                 rmap_remove(kvm, spte);
 738                                 --kvm->stat.lpages;
 739                                 __set_spte(spte, shadow_trap_nonpresent_pte);
 740                                 spte = NULL;
 741                                 write_protected = 1;
 742                         }
 743                         spte = rmap_next(kvm, rmapp, spte);
 744                 }
 745         }
 746
 747         return write_protected;
 748 }
 749
 750 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 751                            unsigned long data)
 752 {
 753         u64 *spte;
 754         int need_tlb_flush = 0;
 755
 756         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 757                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 758                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 759                 rmap_remove(kvm, spte);
 760                 __set_spte(spte, shadow_trap_nonpresent_pte);
 761                 need_tlb_flush = 1;
 762         }
 763         return need_tlb_flush;
 764 }
 765
 766 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
 767                              unsigned long data)
 768 {
 769         int need_flush = 0;
 770         u64 *spte, new_spte;
 771         pte_t *ptep = (pte_t *)data;
 772         pfn_t new_pfn;
 773
 774         WARN_ON(pte_huge(*ptep));
 775         new_pfn = pte_pfn(*ptep);
 776         spte = rmap_next(kvm, rmapp, NULL);
 777         while (spte) {
 778                 BUG_ON(!is_shadow_present_pte(*spte));
 779                 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
 780                 need_flush = 1;
 781                 if (pte_write(*ptep)) {
 782                         rmap_remove(kvm, spte);
 783                         __set_spte(spte, shadow_trap_nonpresent_pte);
 784                         spte = rmap_next(kvm, rmapp, NULL);
 785                 } else {
 786                         new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
 787                         new_spte |= (u64)new_pfn << PAGE_SHIFT;
 788
 789                         new_spte &= ~PT_WRITABLE_MASK;
 790                         new_spte &= ~SPTE_HOST_WRITEABLE;
 791                         if (is_writable_pte(*spte))
 792                                 kvm_set_pfn_dirty(spte_to_pfn(*spte));
 793                         __set_spte(spte, new_spte);
 794                         spte = rmap_next(kvm, rmapp, spte);
 795                 }
 796         }
 797         if (need_flush)
 798                 kvm_flush_remote_tlbs(kvm);
 799
 800         return 0;
 801 }
 802
 803 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 804                           unsigned long data,
 805                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 806                                          unsigned long data))
 807 {
 808         int i, j;
 809         int retval = 0;
 810         struct kvm_memslots *slots;
 811
 812         slots = rcu_dereference(kvm->memslots);
 813
 814         for (i = 0; i < slots->nmemslots; i++) {
 815                 struct kvm_memory_slot *memslot = &slots->memslots[i];
 816                 unsigned long start = memslot->userspace_addr;
 817                 unsigned long end;
 818
 819                 end = start + (memslot->npages << PAGE_SHIFT);
 820                 if (hva >= start && hva < end) {
 821                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 822
 823                         retval |= handler(kvm, &memslot->rmap[gfn_offset],
 824                                           data);
 825
 826                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
 827                                 int idx = gfn_offset;
 828                                 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
 829                                 retval |= handler(kvm,
 830                                         &memslot->lpage_info[j][idx].rmap_pde,
 831                                         data);
 832                         }
 833                 }
 834         }
 835
 836         return retval;
 837 }
 838
 839 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 840 {
 841         return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
 842 }
 843
 844 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
 845 {
 846         kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
 847 }
 848
 849 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 850                          unsigned long data)
 851 {
 852         u64 *spte;
 853         int young = 0;
 854
 855         /* always return old for EPT */
 856         if (!shadow_accessed_mask)
 857                 return 0;
 858
 859         spte = rmap_next(kvm, rmapp, NULL);
 860         while (spte) {
 861                 int _young;
 862                 u64 _spte = *spte;
 863                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 864                 _young = _spte & PT_ACCESSED_MASK;
 865                 if (_young) {
 866                         young = 1;
 867                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 868                 }
 869                 spte = rmap_next(kvm, rmapp, spte);
 870         }
 871         return young;
 872 }
 873
 874 #define RMAP_RECYCLE_THRESHOLD 1000
 875
 876 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 877 {
 878         unsigned long *rmapp;
 879         struct kvm_mmu_page *sp;
 880
 881         sp = page_header(__pa(spte));
 882
 883         gfn = unalias_gfn(vcpu->kvm, gfn);
 884         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 885
 886         kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
 887         kvm_flush_remote_tlbs(vcpu->kvm);
 888 }
 889
 890 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 891 {
 892         return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
 893 }
 894
 895 #ifdef MMU_DEBUG
 896 static int is_empty_shadow_page(u64 *spt)
 897 {
 898         u64 *pos;
 899         u64 *end;
 900
 901         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 902                 if (is_shadow_present_pte(*pos)) {
 903                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 904                                pos, *pos);
 905                         return 0;
 906                 }
 907         return 1;
 908 }
 909 #endif
 910
 911 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 912 {
 913         ASSERT(is_empty_shadow_page(sp->spt));
 914         list_del(&sp->link);
 915         __free_page(virt_to_page(sp->spt));
 916         __free_page(virt_to_page(sp->gfns));
 917         kfree(sp);
 918         ++kvm->arch.n_free_mmu_pages;
 919 }
 920
 921 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 922 {
 923         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 924 }
 925
 926 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 927                                                u64 *parent_pte)
 928 {
 929         struct kvm_mmu_page *sp;
 930
 931         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 932         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 933         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 934         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 935         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 936         INIT_LIST_HEAD(&sp->oos_link);
 937         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 938         sp->multimapped = 0;
 939         sp->parent_pte = parent_pte;
 940         --vcpu->kvm->arch.n_free_mmu_pages;
 941         return sp;
 942 }
 943
 944 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 945                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 946 {
 947         struct kvm_pte_chain *pte_chain;
 948         struct hlist_node *node;
 949         int i;
 950
 951         if (!parent_pte)
 952                 return;
 953         if (!sp->multimapped) {
 954                 u64 *old = sp->parent_pte;
 955
 956                 if (!old) {
 957                         sp->parent_pte = parent_pte;
 958                         return;
 959                 }
 960                 sp->multimapped = 1;
 961                 pte_chain = mmu_alloc_pte_chain(vcpu);
 962                 INIT_HLIST_HEAD(&sp->parent_ptes);
 963                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 964                 pte_chain->parent_ptes[0] = old;
 965         }
 966         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 967                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 968                         continue;
 969                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 970                         if (!pte_chain->parent_ptes[i]) {
 971                                 pte_chain->parent_ptes[i] = parent_pte;
 972                                 return;
 973                         }
 974         }
 975         pte_chain = mmu_alloc_pte_chain(vcpu);
 976         BUG_ON(!pte_chain);
 977         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 978         pte_chain->parent_ptes[0] = parent_pte;
 979 }
 980
 981 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 982                                        u64 *parent_pte)
 983 {
 984         struct kvm_pte_chain *pte_chain;
 985         struct hlist_node *node;
 986         int i;
 987
 988         if (!sp->multimapped) {
 989                 BUG_ON(sp->parent_pte != parent_pte);
 990                 sp->parent_pte = NULL;
 991                 return;
 992         }
 993         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 994                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 995                         if (!pte_chain->parent_ptes[i])
 996                                 break;
 997                         if (pte_chain->parent_ptes[i] != parent_pte)
 998                                 continue;
 999                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
1000                                 && pte_chain->parent_ptes[i + 1]) {
1001                                 pte_chain->parent_ptes[i]
1002                                         = pte_chain->parent_ptes[i + 1];
1003                                 ++i;
1004                         }
1005                         pte_chain->parent_ptes[i] = NULL;
1006                         if (i == 0) {
1007                                 hlist_del(&pte_chain->link);
1008                                 mmu_free_pte_chain(pte_chain);
1009                                 if (hlist_empty(&sp->parent_ptes)) {
1010                                         sp->multimapped = 0;
1011                                         sp->parent_pte = NULL;
1012                                 }
1013                         }
1014                         return;
1015                 }
1016         BUG();
1017 }
1018
1019
1020 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1021                             mmu_parent_walk_fn fn)
1022 {
1023         struct kvm_pte_chain *pte_chain;
1024         struct hlist_node *node;
1025         struct kvm_mmu_page *parent_sp;
1026         int i;
1027
1028         if (!sp->multimapped && sp->parent_pte) {
1029                 parent_sp = page_header(__pa(sp->parent_pte));
1030                 fn(vcpu, parent_sp);
1031                 mmu_parent_walk(vcpu, parent_sp, fn);
1032                 return;
1033         }
1034         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1035                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1036                         if (!pte_chain->parent_ptes[i])
1037                                 break;
1038                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
1039                         fn(vcpu, parent_sp);
1040                         mmu_parent_walk(vcpu, parent_sp, fn);
1041                 }
1042 }
1043
1044 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1045 {
1046         unsigned int index;
1047         struct kvm_mmu_page *sp = page_header(__pa(spte));
1048
1049         index = spte - sp->spt;
1050         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1051                 sp->unsync_children++;
1052         WARN_ON(!sp->unsync_children);
1053 }
1054
1055 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1056 {
1057         struct kvm_pte_chain *pte_chain;
1058         struct hlist_node *node;
1059         int i;
1060
1061         if (!sp->parent_pte)
1062                 return;
1063
1064         if (!sp->multimapped) {
1065                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1066                 return;
1067         }
1068
1069         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1070                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1071                         if (!pte_chain->parent_ptes[i])
1072                                 break;
1073                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1074                 }
1075 }
1076
1077 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1078 {
1079         kvm_mmu_update_parents_unsync(sp);
1080         return 1;
1081 }
1082
1083 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
1084                                         struct kvm_mmu_page *sp)
1085 {
1086         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
1087         kvm_mmu_update_parents_unsync(sp);
1088 }
1089
1090 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1091                                     struct kvm_mmu_page *sp)
1092 {
1093         int i;
1094
1095         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1096                 sp->spt[i] = shadow_trap_nonpresent_pte;
1097 }
1098
1099 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1100                                struct kvm_mmu_page *sp)
1101 {
1102         return 1;
1103 }
1104
1105 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1106 {
1107 }
1108
1109 #define KVM_PAGE_ARRAY_NR 16
1110
1111 struct kvm_mmu_pages {
1112         struct mmu_page_and_offset {
1113                 struct kvm_mmu_page *sp;
1114                 unsigned int idx;
1115         } page[KVM_PAGE_ARRAY_NR];
1116         unsigned int nr;
1117 };
1118
1119 #define for_each_unsync_children(bitmap, idx)           \
1120         for (idx = find_first_bit(bitmap, 512);         \
1121              idx < 512;                                 \
1122              idx = find_next_bit(bitmap, 512, idx+1))
1123
1124 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1125                          int idx)
1126 {
1127         int i;
1128
1129         if (sp->unsync)
1130                 for (i=0; i < pvec->nr; i++)
1131                         if (pvec->page[i].sp == sp)
1132                                 return 0;
1133
1134         pvec->page[pvec->nr].sp = sp;
1135         pvec->page[pvec->nr].idx = idx;
1136         pvec->nr++;
1137         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1138 }
1139
1140 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1141                            struct kvm_mmu_pages *pvec)
1142 {
1143         int i, ret, nr_unsync_leaf = 0;
1144
1145         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1146                 u64 ent = sp->spt[i];
1147
1148                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1149                         struct kvm_mmu_page *child;
1150                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1151
1152                         if (child->unsync_children) {
1153                                 if (mmu_pages_add(pvec, child, i))
1154                                         return -ENOSPC;
1155
1156                                 ret = __mmu_unsync_walk(child, pvec);
1157                                 if (!ret)
1158                                         __clear_bit(i, sp->unsync_child_bitmap);
1159                                 else if (ret > 0)
1160                                         nr_unsync_leaf += ret;
1161                                 else
1162                                         return ret;
1163                         }
1164
1165                         if (child->unsync) {
1166                                 nr_unsync_leaf++;
1167                                 if (mmu_pages_add(pvec, child, i))
1168                                         return -ENOSPC;
1169                         }
1170                 }
1171         }
1172
1173         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1174                 sp->unsync_children = 0;
1175
1176         return nr_unsync_leaf;
1177 }
1178
1179 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1180                            struct kvm_mmu_pages *pvec)
1181 {
1182         if (!sp->unsync_children)
1183                 return 0;
1184
1185         mmu_pages_add(pvec, sp, 0);
1186         return __mmu_unsync_walk(sp, pvec);
1187 }
1188
1189 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1190 {
1191         unsigned index;
1192         struct hlist_head *bucket;
1193         struct kvm_mmu_page *sp;
1194         struct hlist_node *node;
1195
1196         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1197         index = kvm_page_table_hashfn(gfn);
1198         bucket = &kvm->arch.mmu_page_hash[index];
1199         hlist_for_each_entry(sp, node, bucket, hash_link)
1200                 if (sp->gfn == gfn && !sp->role.direct
1201                     && !sp->role.invalid) {
1202                         pgprintk("%s: found role %x\n",
1203                                  __func__, sp->role.word);
1204                         return sp;
1205                 }
1206         return NULL;
1207 }
1208
1209 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1210 {
1211         WARN_ON(!sp->unsync);
1212         sp->unsync = 0;
1213         --kvm->stat.mmu_unsync;
1214 }
1215
1216 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1217
1218 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1219 {
1220         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1221                 kvm_mmu_zap_page(vcpu->kvm, sp);
1222                 return 1;
1223         }
1224
1225         trace_kvm_mmu_sync_page(sp);
1226         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1227                 kvm_flush_remote_tlbs(vcpu->kvm);
1228         kvm_unlink_unsync_page(vcpu->kvm, sp);
1229         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1230                 kvm_mmu_zap_page(vcpu->kvm, sp);
1231                 return 1;
1232         }
1233
1234         kvm_mmu_flush_tlb(vcpu);
1235         return 0;
1236 }
1237
1238 struct mmu_page_path {
1239         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1240         unsigned int idx[PT64_ROOT_LEVEL-1];
1241 };
1242
1243 #define for_each_sp(pvec, sp, parents, i)                       \
1244                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1245                         sp = pvec.page[i].sp;                   \
1246                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1247                         i = mmu_pages_next(&pvec, &parents, i))
1248
1249 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1250                           struct mmu_page_path *parents,
1251                           int i)
1252 {
1253         int n;
1254
1255         for (n = i+1; n < pvec->nr; n++) {
1256                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1257
1258                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1259                         parents->idx[0] = pvec->page[n].idx;
1260                         return n;
1261                 }
1262
1263                 parents->parent[sp->role.level-2] = sp;
1264                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1265         }
1266
1267         return n;
1268 }
1269
1270 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1271 {
1272         struct kvm_mmu_page *sp;
1273         unsigned int level = 0;
1274
1275         do {
1276                 unsigned int idx = parents->idx[level];
1277
1278                 sp = parents->parent[level];
1279                 if (!sp)
1280                         return;
1281
1282                 --sp->unsync_children;
1283                 WARN_ON((int)sp->unsync_children < 0);
1284                 __clear_bit(idx, sp->unsync_child_bitmap);
1285                 level++;
1286         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1287 }
1288
1289 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1290                                struct mmu_page_path *parents,
1291                                struct kvm_mmu_pages *pvec)
1292 {
1293         parents->parent[parent->role.level-1] = NULL;
1294         pvec->nr = 0;
1295 }
1296
1297 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1298                               struct kvm_mmu_page *parent)
1299 {
1300         int i;
1301         struct kvm_mmu_page *sp;
1302         struct mmu_page_path parents;
1303         struct kvm_mmu_pages pages;
1304
1305         kvm_mmu_pages_init(parent, &parents, &pages);
1306         while (mmu_unsync_walk(parent, &pages)) {
1307                 int protected = 0;
1308
1309                 for_each_sp(pages, sp, parents, i)
1310                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1311
1312                 if (protected)
1313                         kvm_flush_remote_tlbs(vcpu->kvm);
1314
1315                 for_each_sp(pages, sp, parents, i) {
1316                         kvm_sync_page(vcpu, sp);
1317                         mmu_pages_clear_parents(&parents);
1318                 }
1319                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1320                 kvm_mmu_pages_init(parent, &parents, &pages);
1321         }
1322 }
1323
1324 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1325                                              gfn_t gfn,
1326                                              gva_t gaddr,
1327                                              unsigned level,
1328                                              int direct,
1329                                              unsigned access,
1330                                              u64 *parent_pte)
1331 {
1332         union kvm_mmu_page_role role;
1333         unsigned index;
1334         unsigned quadrant;
1335         struct hlist_head *bucket;
1336         struct kvm_mmu_page *sp;
1337         struct hlist_node *node, *tmp;
1338
1339         role = vcpu->arch.mmu.base_role;
1340         role.level = level;
1341         role.direct = direct;
1342         role.access = access;
1343         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1344                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1345                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1346                 role.quadrant = quadrant;
1347         }
1348         index = kvm_page_table_hashfn(gfn);
1349         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1350         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1351                 if (sp->gfn == gfn) {
1352                         if (sp->unsync)
1353                                 if (kvm_sync_page(vcpu, sp))
1354                                         continue;
1355
1356                         if (sp->role.word != role.word)
1357                                 continue;
1358
1359                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1360                         if (sp->unsync_children) {
1361                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1362                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1363                         }
1364                         trace_kvm_mmu_get_page(sp, false);
1365                         return sp;
1366                 }
1367         ++vcpu->kvm->stat.mmu_cache_miss;
1368         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1369         if (!sp)
1370                 return sp;
1371         sp->gfn = gfn;
1372         sp->role = role;
1373         hlist_add_head(&sp->hash_link, bucket);
1374         if (!direct) {
1375                 if (rmap_write_protect(vcpu->kvm, gfn))
1376                         kvm_flush_remote_tlbs(vcpu->kvm);
1377                 account_shadowed(vcpu->kvm, gfn);
1378         }
1379         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1380                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1381         else
1382                 nonpaging_prefetch_page(vcpu, sp);
1383         trace_kvm_mmu_get_page(sp, true);
1384         return sp;
1385 }
1386
1387 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1388                              struct kvm_vcpu *vcpu, u64 addr)
1389 {
1390         iterator->addr = addr;
1391         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1392         iterator->level = vcpu->arch.mmu.shadow_root_level;
1393         if (iterator->level == PT32E_ROOT_LEVEL) {
1394                 iterator->shadow_addr
1395                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1396                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1397                 --iterator->level;
1398                 if (!iterator->shadow_addr)
1399                         iterator->level = 0;
1400         }
1401 }
1402
1403 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1404 {
1405         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1406                 return false;
1407
1408         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1409                 if (is_large_pte(*iterator->sptep))
1410                         return false;
1411
1412         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1413         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1414         return true;
1415 }
1416
1417 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1418 {
1419         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1420         --iterator->level;
1421 }
1422
1423 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1424                                          struct kvm_mmu_page *sp)
1425 {
1426         unsigned i;
1427         u64 *pt;
1428         u64 ent;
1429
1430         pt = sp->spt;
1431
1432         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1433                 ent = pt[i];
1434
1435                 if (is_shadow_present_pte(ent)) {
1436                         if (!is_last_spte(ent, sp->role.level)) {
1437                                 ent &= PT64_BASE_ADDR_MASK;
1438                                 mmu_page_remove_parent_pte(page_header(ent),
1439                                                            &pt[i]);
1440                         } else {
1441                                 if (is_large_pte(ent))
1442                                         --kvm->stat.lpages;
1443                                 rmap_remove(kvm, &pt[i]);
1444                         }
1445                 }
1446                 pt[i] = shadow_trap_nonpresent_pte;
1447         }
1448 }
1449
1450 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1451 {
1452         mmu_page_remove_parent_pte(sp, parent_pte);
1453 }
1454
1455 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1456 {
1457         int i;
1458         struct kvm_vcpu *vcpu;
1459
1460         kvm_for_each_vcpu(i, vcpu, kvm)
1461                 vcpu->arch.last_pte_updated = NULL;
1462 }
1463
1464 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1465 {
1466         u64 *parent_pte;
1467
1468         while (sp->multimapped || sp->parent_pte) {
1469                 if (!sp->multimapped)
1470                         parent_pte = sp->parent_pte;
1471                 else {
1472                         struct kvm_pte_chain *chain;
1473
1474                         chain = container_of(sp->parent_ptes.first,
1475                                              struct kvm_pte_chain, link);
1476                         parent_pte = chain->parent_ptes[0];
1477                 }
1478                 BUG_ON(!parent_pte);
1479                 kvm_mmu_put_page(sp, parent_pte);
1480                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1481         }
1482 }
1483
1484 static int mmu_zap_unsync_children(struct kvm *kvm,
1485                                    struct kvm_mmu_page *parent)
1486 {
1487         int i, zapped = 0;
1488         struct mmu_page_path parents;
1489         struct kvm_mmu_pages pages;
1490
1491         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1492                 return 0;
1493
1494         kvm_mmu_pages_init(parent, &parents, &pages);
1495         while (mmu_unsync_walk(parent, &pages)) {
1496                 struct kvm_mmu_page *sp;
1497
1498                 for_each_sp(pages, sp, parents, i) {
1499                         kvm_mmu_zap_page(kvm, sp);
1500                         mmu_pages_clear_parents(&parents);
1501                 }
1502                 zapped += pages.nr;
1503                 kvm_mmu_pages_init(parent, &parents, &pages);
1504         }
1505
1506         return zapped;
1507 }
1508
1509 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1510 {
1511         int ret;
1512
1513         trace_kvm_mmu_zap_page(sp);
1514         ++kvm->stat.mmu_shadow_zapped;
1515         ret = mmu_zap_unsync_children(kvm, sp);
1516         kvm_mmu_page_unlink_children(kvm, sp);
1517         kvm_mmu_unlink_parents(kvm, sp);
1518         kvm_flush_remote_tlbs(kvm);
1519         if (!sp->role.invalid && !sp->role.direct)
1520                 unaccount_shadowed(kvm, sp->gfn);
1521         if (sp->unsync)
1522                 kvm_unlink_unsync_page(kvm, sp);
1523         if (!sp->root_count) {
1524                 hlist_del(&sp->hash_link);
1525                 kvm_mmu_free_page(kvm, sp);
1526         } else {
1527                 sp->role.invalid = 1;
1528                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1529                 kvm_reload_remote_mmus(kvm);
1530         }
1531         kvm_mmu_reset_last_pte_updated(kvm);
1532         return ret;
1533 }
1534
1535 /*
1536  * Changing the number of mmu pages allocated to the vm
1537  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1538  */
1539 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1540 {
1541         int used_pages;
1542
1543         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1544         used_pages = max(0, used_pages);
1545
1546         /*
1547          * If we set the number of mmu pages to be smaller be than the
1548          * number of actived pages , we must to free some mmu pages before we
1549          * change the value
1550          */
1551
1552         if (used_pages > kvm_nr_mmu_pages) {
1553                 while (used_pages > kvm_nr_mmu_pages) {
1554                         struct kvm_mmu_page *page;
1555
1556                         page = container_of(kvm->arch.active_mmu_pages.prev,
1557                                             struct kvm_mmu_page, link);
1558                         kvm_mmu_zap_page(kvm, page);
1559                         used_pages--;
1560                 }
1561                 kvm->arch.n_free_mmu_pages = 0;
1562         }
1563         else
1564                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1565                                          - kvm->arch.n_alloc_mmu_pages;
1566
1567         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1568 }
1569
1570 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1571 {
1572         unsigned index;
1573         struct hlist_head *bucket;
1574         struct kvm_mmu_page *sp;
1575         struct hlist_node *node, *n;
1576         int r;
1577
1578         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1579         r = 0;
1580         index = kvm_page_table_hashfn(gfn);
1581         bucket = &kvm->arch.mmu_page_hash[index];
1582         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1583                 if (sp->gfn == gfn && !sp->role.direct) {
1584                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1585                                  sp->role.word);
1586                         r = 1;
1587                         if (kvm_mmu_zap_page(kvm, sp))
1588                                 n = bucket->first;
1589                 }
1590         return r;
1591 }
1592
1593 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1594 {
1595         unsigned index;
1596         struct hlist_head *bucket;
1597         struct kvm_mmu_page *sp;
1598         struct hlist_node *node, *nn;
1599
1600         index = kvm_page_table_hashfn(gfn);
1601         bucket = &kvm->arch.mmu_page_hash[index];
1602         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1603                 if (sp->gfn == gfn && !sp->role.direct
1604                     && !sp->role.invalid) {
1605                         pgprintk("%s: zap %lx %x\n",
1606                                  __func__, gfn, sp->role.word);
1607                         kvm_mmu_zap_page(kvm, sp);
1608                 }
1609         }
1610 }
1611
1612 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1613 {
1614         int slot = memslot_id(kvm, gfn);
1615         struct kvm_mmu_page *sp = page_header(__pa(pte));
1616
1617         __set_bit(slot, sp->slot_bitmap);
1618 }
1619
1620 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1621 {
1622         int i;
1623         u64 *pt = sp->spt;
1624
1625         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1626                 return;
1627
1628         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1629                 if (pt[i] == shadow_notrap_nonpresent_pte)
1630                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1631         }
1632 }
1633
1634 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1635 {
1636         struct page *page;
1637
1638         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1639
1640         if (gpa == UNMAPPED_GVA)
1641                 return NULL;
1642
1643         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1644
1645         return page;
1646 }
1647
1648 /*
1649  * The function is based on mtrr_type_lookup() in
1650  * arch/x86/kernel/cpu/mtrr/generic.c
1651  */
1652 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1653                          u64 start, u64 end)
1654 {
1655         int i;
1656         u64 base, mask;
1657         u8 prev_match, curr_match;
1658         int num_var_ranges = KVM_NR_VAR_MTRR;
1659
1660         if (!mtrr_state->enabled)
1661                 return 0xFF;
1662
1663         /* Make end inclusive end, instead of exclusive */
1664         end--;
1665
1666         /* Look in fixed ranges. Just return the type as per start */
1667         if (mtrr_state->have_fixed && (start < 0x100000)) {
1668                 int idx;
1669
1670                 if (start < 0x80000) {
1671                         idx = 0;
1672                         idx += (start >> 16);
1673                         return mtrr_state->fixed_ranges[idx];
1674                 } else if (start < 0xC0000) {
1675                         idx = 1 * 8;
1676                         idx += ((start - 0x80000) >> 14);
1677                         return mtrr_state->fixed_ranges[idx];
1678                 } else if (start < 0x1000000) {
1679                         idx = 3 * 8;
1680                         idx += ((start - 0xC0000) >> 12);
1681                         return mtrr_state->fixed_ranges[idx];
1682                 }
1683         }
1684
1685         /*
1686          * Look in variable ranges
1687          * Look of multiple ranges matching this address and pick type
1688          * as per MTRR precedence
1689          */
1690         if (!(mtrr_state->enabled & 2))
1691                 return mtrr_state->def_type;
1692
1693         prev_match = 0xFF;
1694         for (i = 0; i < num_var_ranges; ++i) {
1695                 unsigned short start_state, end_state;
1696
1697                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1698                         continue;
1699
1700                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1701                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1702                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1703                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1704
1705                 start_state = ((start & mask) == (base & mask));
1706                 end_state = ((end & mask) == (base & mask));
1707                 if (start_state != end_state)
1708                         return 0xFE;
1709
1710                 if ((start & mask) != (base & mask))
1711                         continue;
1712
1713                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1714                 if (prev_match == 0xFF) {
1715                         prev_match = curr_match;
1716                         continue;
1717                 }
1718
1719                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1720                     curr_match == MTRR_TYPE_UNCACHABLE)
1721                         return MTRR_TYPE_UNCACHABLE;
1722
1723                 if ((prev_match == MTRR_TYPE_WRBACK &&
1724                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1725                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1726                      curr_match == MTRR_TYPE_WRBACK)) {
1727                         prev_match = MTRR_TYPE_WRTHROUGH;
1728                         curr_match = MTRR_TYPE_WRTHROUGH;
1729                 }
1730
1731                 if (prev_match != curr_match)
1732                         return MTRR_TYPE_UNCACHABLE;
1733         }
1734
1735         if (prev_match != 0xFF)
1736                 return prev_match;
1737
1738         return mtrr_state->def_type;
1739 }
1740
1741 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1742 {
1743         u8 mtrr;
1744
1745         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1746                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1747         if (mtrr == 0xfe || mtrr == 0xff)
1748                 mtrr = MTRR_TYPE_WRBACK;
1749         return mtrr;
1750 }
1751 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1752
1753 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1754 {
1755         unsigned index;
1756         struct hlist_head *bucket;
1757         struct kvm_mmu_page *s;
1758         struct hlist_node *node, *n;
1759
1760         trace_kvm_mmu_unsync_page(sp);
1761         index = kvm_page_table_hashfn(sp->gfn);
1762         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1763         /* don't unsync if pagetable is shadowed with multiple roles */
1764         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1765                 if (s->gfn != sp->gfn || s->role.direct)
1766                         continue;
1767                 if (s->role.word != sp->role.word)
1768                         return 1;
1769         }
1770         ++vcpu->kvm->stat.mmu_unsync;
1771         sp->unsync = 1;
1772
1773         kvm_mmu_mark_parents_unsync(vcpu, sp);
1774
1775         mmu_convert_notrap(sp);
1776         return 0;
1777 }
1778
1779 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1780                                   bool can_unsync)
1781 {
1782         struct kvm_mmu_page *shadow;
1783
1784         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1785         if (shadow) {
1786                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1787                         return 1;
1788                 if (shadow->unsync)
1789                         return 0;
1790                 if (can_unsync && oos_shadow)
1791                         return kvm_unsync_page(vcpu, shadow);
1792                 return 1;
1793         }
1794         return 0;
1795 }
1796
1797 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1798                     unsigned pte_access, int user_fault,
1799                     int write_fault, int dirty, int level,
1800                     gfn_t gfn, pfn_t pfn, bool speculative,
1801                     bool can_unsync, bool reset_host_protection)
1802 {
1803         u64 spte;
1804         int ret = 0;
1805
1806         /*
1807          * We don't set the accessed bit, since we sometimes want to see
1808          * whether the guest actually used the pte (in order to detect
1809          * demand paging).
1810          */
1811         spte = shadow_base_present_pte | shadow_dirty_mask;
1812         if (!speculative)
1813                 spte |= shadow_accessed_mask;
1814         if (!dirty)
1815                 pte_access &= ~ACC_WRITE_MASK;
1816         if (pte_access & ACC_EXEC_MASK)
1817                 spte |= shadow_x_mask;
1818         else
1819                 spte |= shadow_nx_mask;
1820         if (pte_access & ACC_USER_MASK)
1821                 spte |= shadow_user_mask;
1822         if (level > PT_PAGE_TABLE_LEVEL)
1823                 spte |= PT_PAGE_SIZE_MASK;
1824         if (tdp_enabled)
1825                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1826                         kvm_is_mmio_pfn(pfn));
1827
1828         if (reset_host_protection)
1829                 spte |= SPTE_HOST_WRITEABLE;
1830
1831         spte |= (u64)pfn << PAGE_SHIFT;
1832
1833         if ((pte_access & ACC_WRITE_MASK)
1834             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1835
1836                 if (level > PT_PAGE_TABLE_LEVEL &&
1837                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1838                         ret = 1;
1839                         spte = shadow_trap_nonpresent_pte;
1840                         goto set_pte;
1841                 }
1842
1843                 spte |= PT_WRITABLE_MASK;
1844
1845                 /*
1846                  * Optimization: for pte sync, if spte was writable the hash
1847                  * lookup is unnecessary (and expensive). Write protection
1848                  * is responsibility of mmu_get_page / kvm_sync_page.
1849                  * Same reasoning can be applied to dirty page accounting.
1850                  */
1851                 if (!can_unsync && is_writable_pte(*sptep))
1852                         goto set_pte;
1853
1854                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1855                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1856                                  __func__, gfn);
1857                         ret = 1;
1858                         pte_access &= ~ACC_WRITE_MASK;
1859                         if (is_writable_pte(spte))
1860                                 spte &= ~PT_WRITABLE_MASK;
1861                 }
1862         }
1863
1864         if (pte_access & ACC_WRITE_MASK)
1865                 mark_page_dirty(vcpu->kvm, gfn);
1866
1867 set_pte:
1868         __set_spte(sptep, spte);
1869         return ret;
1870 }
1871
1872 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1873                          unsigned pt_access, unsigned pte_access,
1874                          int user_fault, int write_fault, int dirty,
1875                          int *ptwrite, int level, gfn_t gfn,
1876                          pfn_t pfn, bool speculative,
1877                          bool reset_host_protection)
1878 {
1879         int was_rmapped = 0;
1880         int was_writable = is_writable_pte(*sptep);
1881         int rmap_count;
1882
1883         pgprintk("%s: spte %llx access %x write_fault %d"
1884                  " user_fault %d gfn %lx\n",
1885                  __func__, *sptep, pt_access,
1886                  write_fault, user_fault, gfn);
1887
1888         if (is_rmap_spte(*sptep)) {
1889                 /*
1890                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1891                  * the parent of the now unreachable PTE.
1892                  */
1893                 if (level > PT_PAGE_TABLE_LEVEL &&
1894                     !is_large_pte(*sptep)) {
1895                         struct kvm_mmu_page *child;
1896                         u64 pte = *sptep;
1897
1898                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1899                         mmu_page_remove_parent_pte(child, sptep);
1900                 } else if (pfn != spte_to_pfn(*sptep)) {
1901                         pgprintk("hfn old %lx new %lx\n",
1902                                  spte_to_pfn(*sptep), pfn);
1903                         rmap_remove(vcpu->kvm, sptep);
1904                 } else
1905                         was_rmapped = 1;
1906         }
1907
1908         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1909                       dirty, level, gfn, pfn, speculative, true,
1910                       reset_host_protection)) {
1911                 if (write_fault)
1912                         *ptwrite = 1;
1913                 kvm_x86_ops->tlb_flush(vcpu);
1914         }
1915
1916         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1917         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1918                  is_large_pte(*sptep)? "2MB" : "4kB",
1919                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1920                  *sptep, sptep);
1921         if (!was_rmapped && is_large_pte(*sptep))
1922                 ++vcpu->kvm->stat.lpages;
1923
1924         page_header_update_slot(vcpu->kvm, sptep, gfn);
1925         if (!was_rmapped) {
1926                 rmap_count = rmap_add(vcpu, sptep, gfn);
1927                 kvm_release_pfn_clean(pfn);
1928                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1929                         rmap_recycle(vcpu, sptep, gfn);
1930         } else {
1931                 if (was_writable)
1932                         kvm_release_pfn_dirty(pfn);
1933                 else
1934                         kvm_release_pfn_clean(pfn);
1935         }
1936         if (speculative) {
1937                 vcpu->arch.last_pte_updated = sptep;
1938                 vcpu->arch.last_pte_gfn = gfn;
1939         }
1940 }
1941
1942 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1943 {
1944 }
1945
1946 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1947                         int level, gfn_t gfn, pfn_t pfn)
1948 {
1949         struct kvm_shadow_walk_iterator iterator;
1950         struct kvm_mmu_page *sp;
1951         int pt_write = 0;
1952         gfn_t pseudo_gfn;
1953
1954         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1955                 if (iterator.level == level) {
1956                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1957                                      0, write, 1, &pt_write,
1958                                      level, gfn, pfn, false, true);
1959                         ++vcpu->stat.pf_fixed;
1960                         break;
1961                 }
1962
1963                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1964                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1965                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1966                                               iterator.level - 1,
1967                                               1, ACC_ALL, iterator.sptep);
1968                         if (!sp) {
1969                                 pgprintk("nonpaging_map: ENOMEM\n");
1970                                 kvm_release_pfn_clean(pfn);
1971                                 return -ENOMEM;
1972                         }
1973
1974                         __set_spte(iterator.sptep,
1975                                    __pa(sp->spt)
1976                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1977                                    | shadow_user_mask | shadow_x_mask);
1978                 }
1979         }
1980         return pt_write;
1981 }
1982
1983 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1984 {
1985         int r;
1986         int level;
1987         pfn_t pfn;
1988         unsigned long mmu_seq;
1989
1990         level = mapping_level(vcpu, gfn);
1991
1992         /*
1993          * This path builds a PAE pagetable - so we can map 2mb pages at
1994          * maximum. Therefore check if the level is larger than that.
1995          */
1996         if (level > PT_DIRECTORY_LEVEL)
1997                 level = PT_DIRECTORY_LEVEL;
1998
1999         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2000
2001         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2002         smp_rmb();
2003         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2004
2005         /* mmio */
2006         if (is_error_pfn(pfn)) {
2007                 kvm_release_pfn_clean(pfn);
2008                 return 1;
2009         }
2010
2011         spin_lock(&vcpu->kvm->mmu_lock);
2012         if (mmu_notifier_retry(vcpu, mmu_seq))
2013                 goto out_unlock;
2014         kvm_mmu_free_some_pages(vcpu);
2015         r = __direct_map(vcpu, v, write, level, gfn, pfn);
2016         spin_unlock(&vcpu->kvm->mmu_lock);
2017
2018
2019         return r;
2020
2021 out_unlock:
2022         spin_unlock(&vcpu->kvm->mmu_lock);
2023         kvm_release_pfn_clean(pfn);
2024         return 0;
2025 }
2026
2027
2028 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2029 {
2030         int i;
2031         struct kvm_mmu_page *sp;
2032
2033         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2034                 return;
2035         spin_lock(&vcpu->kvm->mmu_lock);
2036         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2037                 hpa_t root = vcpu->arch.mmu.root_hpa;
2038
2039                 sp = page_header(root);
2040                 --sp->root_count;
2041                 if (!sp->root_count && sp->role.invalid)
2042                         kvm_mmu_zap_page(vcpu->kvm, sp);
2043                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2044                 spin_unlock(&vcpu->kvm->mmu_lock);
2045                 return;
2046         }
2047         for (i = 0; i < 4; ++i) {
2048                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2049
2050                 if (root) {
2051                         root &= PT64_BASE_ADDR_MASK;
2052                         sp = page_header(root);
2053                         --sp->root_count;
2054                         if (!sp->root_count && sp->role.invalid)
2055                                 kvm_mmu_zap_page(vcpu->kvm, sp);
2056                 }
2057                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2058         }
2059         spin_unlock(&vcpu->kvm->mmu_lock);
2060         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2061 }
2062
2063 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2064 {
2065         int ret = 0;
2066
2067         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2068                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2069                 ret = 1;
2070         }
2071
2072         return ret;
2073 }
2074
2075 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2076 {
2077         int i;
2078         gfn_t root_gfn;
2079         struct kvm_mmu_page *sp;
2080         int direct = 0;
2081         u64 pdptr;
2082
2083         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2084
2085         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2086                 hpa_t root = vcpu->arch.mmu.root_hpa;
2087
2088                 ASSERT(!VALID_PAGE(root));
2089                 if (tdp_enabled)
2090                         direct = 1;
2091                 if (mmu_check_root(vcpu, root_gfn))
2092                         return 1;
2093                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2094                                       PT64_ROOT_LEVEL, direct,
2095                                       ACC_ALL, NULL);
2096                 root = __pa(sp->spt);
2097                 ++sp->root_count;
2098                 vcpu->arch.mmu.root_hpa = root;
2099                 return 0;
2100         }
2101         direct = !is_paging(vcpu);
2102         if (tdp_enabled)
2103                 direct = 1;
2104         for (i = 0; i < 4; ++i) {
2105                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2106
2107                 ASSERT(!VALID_PAGE(root));
2108                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2109                         pdptr = kvm_pdptr_read(vcpu, i);
2110                         if (!is_present_gpte(pdptr)) {
2111                                 vcpu->arch.mmu.pae_root[i] = 0;
2112                                 continue;
2113                         }
2114                         root_gfn = pdptr >> PAGE_SHIFT;
2115                 } else if (vcpu->arch.mmu.root_level == 0)
2116                         root_gfn = 0;
2117                 if (mmu_check_root(vcpu, root_gfn))
2118                         return 1;
2119                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2120                                       PT32_ROOT_LEVEL, direct,
2121                                       ACC_ALL, NULL);
2122                 root = __pa(sp->spt);
2123                 ++sp->root_count;
2124                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2125         }
2126         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2127         return 0;
2128 }
2129
2130 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2131 {
2132         int i;
2133         struct kvm_mmu_page *sp;
2134
2135         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2136                 return;
2137         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2138                 hpa_t root = vcpu->arch.mmu.root_hpa;
2139                 sp = page_header(root);
2140                 mmu_sync_children(vcpu, sp);
2141                 return;
2142         }
2143         for (i = 0; i < 4; ++i) {
2144                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2145
2146                 if (root && VALID_PAGE(root)) {
2147                         root &= PT64_BASE_ADDR_MASK;
2148                         sp = page_header(root);
2149                         mmu_sync_children(vcpu, sp);
2150                 }
2151         }
2152 }
2153
2154 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2155 {
2156         spin_lock(&vcpu->kvm->mmu_lock);
2157         mmu_sync_roots(vcpu);
2158         spin_unlock(&vcpu->kvm->mmu_lock);
2159 }
2160
2161 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2162 {
2163         return vaddr;
2164 }
2165
2166 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2167                                 u32 error_code)
2168 {
2169         gfn_t gfn;
2170         int r;
2171
2172         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2173         r = mmu_topup_memory_caches(vcpu);
2174         if (r)
2175                 return r;
2176
2177         ASSERT(vcpu);
2178         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2179
2180         gfn = gva >> PAGE_SHIFT;
2181
2182         return nonpaging_map(vcpu, gva & PAGE_MASK,
2183                              error_code & PFERR_WRITE_MASK, gfn);
2184 }
2185
2186 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2187                                 u32 error_code)
2188 {
2189         pfn_t pfn;
2190         int r;
2191         int level;
2192         gfn_t gfn = gpa >> PAGE_SHIFT;
2193         unsigned long mmu_seq;
2194
2195         ASSERT(vcpu);
2196         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2197
2198         r = mmu_topup_memory_caches(vcpu);
2199         if (r)
2200                 return r;
2201
2202         level = mapping_level(vcpu, gfn);
2203
2204         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2205
2206         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2207         smp_rmb();
2208         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2209         if (is_error_pfn(pfn)) {
2210                 kvm_release_pfn_clean(pfn);
2211                 return 1;
2212         }
2213         spin_lock(&vcpu->kvm->mmu_lock);
2214         if (mmu_notifier_retry(vcpu, mmu_seq))
2215                 goto out_unlock;
2216         kvm_mmu_free_some_pages(vcpu);
2217         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2218                          level, gfn, pfn);
2219         spin_unlock(&vcpu->kvm->mmu_lock);
2220
2221         return r;
2222
2223 out_unlock:
2224         spin_unlock(&vcpu->kvm->mmu_lock);
2225         kvm_release_pfn_clean(pfn);
2226         return 0;
2227 }
2228
2229 static void nonpaging_free(struct kvm_vcpu *vcpu)
2230 {
2231         mmu_free_roots(vcpu);
2232 }
2233
2234 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2235 {
2236         struct kvm_mmu *context = &vcpu->arch.mmu;
2237
2238         context->new_cr3 = nonpaging_new_cr3;
2239         context->page_fault = nonpaging_page_fault;
2240         context->gva_to_gpa = nonpaging_gva_to_gpa;
2241         context->free = nonpaging_free;
2242         context->prefetch_page = nonpaging_prefetch_page;
2243         context->sync_page = nonpaging_sync_page;
2244         context->invlpg = nonpaging_invlpg;
2245         context->root_level = 0;
2246         context->shadow_root_level = PT32E_ROOT_LEVEL;
2247         context->root_hpa = INVALID_PAGE;
2248         return 0;
2249 }
2250
2251 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2252 {
2253         ++vcpu->stat.tlb_flush;
2254         kvm_x86_ops->tlb_flush(vcpu);
2255 }
2256
2257 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2258 {
2259         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2260         mmu_free_roots(vcpu);
2261 }
2262
2263 static void inject_page_fault(struct kvm_vcpu *vcpu,
2264                               u64 addr,
2265                               u32 err_code)
2266 {
2267         kvm_inject_page_fault(vcpu, addr, err_code);
2268 }
2269
2270 static void paging_free(struct kvm_vcpu *vcpu)
2271 {
2272         nonpaging_free(vcpu);
2273 }
2274
2275 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2276 {
2277         int bit7;
2278
2279         bit7 = (gpte >> 7) & 1;
2280         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2281 }
2282
2283 #define PTTYPE 64
2284 #include "paging_tmpl.h"
2285 #undef PTTYPE
2286
2287 #define PTTYPE 32
2288 #include "paging_tmpl.h"
2289 #undef PTTYPE
2290
2291 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2292 {
2293         struct kvm_mmu *context = &vcpu->arch.mmu;
2294         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2295         u64 exb_bit_rsvd = 0;
2296
2297         if (!is_nx(vcpu))
2298                 exb_bit_rsvd = rsvd_bits(63, 63);
2299         switch (level) {
2300         case PT32_ROOT_LEVEL:
2301                 /* no rsvd bits for 2 level 4K page table entries */
2302                 context->rsvd_bits_mask[0][1] = 0;
2303                 context->rsvd_bits_mask[0][0] = 0;
2304                 if (is_cpuid_PSE36())
2305                         /* 36bits PSE 4MB page */
2306                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2307                 else
2308                         /* 32 bits PSE 4MB page */
2309                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2310                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2311                 break;
2312         case PT32E_ROOT_LEVEL:
2313                 context->rsvd_bits_mask[0][2] =
2314                         rsvd_bits(maxphyaddr, 63) |
2315                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2316                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2317                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2318                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2319                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2320                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2321                         rsvd_bits(maxphyaddr, 62) |
2322                         rsvd_bits(13, 20);              /* large page */
2323                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2324                 break;
2325         case PT64_ROOT_LEVEL:
2326                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2327                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2328                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2329                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2330                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2331                         rsvd_bits(maxphyaddr, 51);
2332                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2333                         rsvd_bits(maxphyaddr, 51);
2334                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2335                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2336                         rsvd_bits(maxphyaddr, 51) |
2337                         rsvd_bits(13, 29);
2338                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2339                         rsvd_bits(maxphyaddr, 51) |
2340                         rsvd_bits(13, 20);              /* large page */
2341                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2342                 break;
2343         }
2344 }
2345
2346 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2347 {
2348         struct kvm_mmu *context = &vcpu->arch.mmu;
2349
2350         ASSERT(is_pae(vcpu));
2351         context->new_cr3 = paging_new_cr3;
2352         context->page_fault = paging64_page_fault;
2353         context->gva_to_gpa = paging64_gva_to_gpa;
2354         context->prefetch_page = paging64_prefetch_page;
2355         context->sync_page = paging64_sync_page;
2356         context->invlpg = paging64_invlpg;
2357         context->free = paging_free;
2358         context->root_level = level;
2359         context->shadow_root_level = level;
2360         context->root_hpa = INVALID_PAGE;
2361         return 0;
2362 }
2363
2364 static int paging64_init_context(struct kvm_vcpu *vcpu)
2365 {
2366         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2367         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2368 }
2369
2370 static int paging32_init_context(struct kvm_vcpu *vcpu)
2371 {
2372         struct kvm_mmu *context = &vcpu->arch.mmu;
2373
2374         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2375         context->new_cr3 = paging_new_cr3;
2376         context->page_fault = paging32_page_fault;
2377         context->gva_to_gpa = paging32_gva_to_gpa;
2378         context->free = paging_free;
2379         context->prefetch_page = paging32_prefetch_page;
2380         context->sync_page = paging32_sync_page;
2381         context->invlpg = paging32_invlpg;
2382         context->root_level = PT32_ROOT_LEVEL;
2383         context->shadow_root_level = PT32E_ROOT_LEVEL;
2384         context->root_hpa = INVALID_PAGE;
2385         return 0;
2386 }
2387
2388 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2389 {
2390         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2391         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2392 }
2393
2394 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2395 {
2396         struct kvm_mmu *context = &vcpu->arch.mmu;
2397
2398         context->new_cr3 = nonpaging_new_cr3;
2399         context->page_fault = tdp_page_fault;
2400         context->free = nonpaging_free;
2401         context->prefetch_page = nonpaging_prefetch_page;
2402         context->sync_page = nonpaging_sync_page;
2403         context->invlpg = nonpaging_invlpg;
2404         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2405         context->root_hpa = INVALID_PAGE;
2406
2407         if (!is_paging(vcpu)) {
2408                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2409                 context->root_level = 0;
2410         } else if (is_long_mode(vcpu)) {
2411                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2412                 context->gva_to_gpa = paging64_gva_to_gpa;
2413                 context->root_level = PT64_ROOT_LEVEL;
2414         } else if (is_pae(vcpu)) {
2415                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2416                 context->gva_to_gpa = paging64_gva_to_gpa;
2417                 context->root_level = PT32E_ROOT_LEVEL;
2418         } else {
2419                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2420                 context->gva_to_gpa = paging32_gva_to_gpa;
2421                 context->root_level = PT32_ROOT_LEVEL;
2422         }
2423
2424         return 0;
2425 }
2426
2427 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2428 {
2429         int r;
2430
2431         ASSERT(vcpu);
2432         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2433
2434         if (!is_paging(vcpu))
2435                 r = nonpaging_init_context(vcpu);
2436         else if (is_long_mode(vcpu))
2437                 r = paging64_init_context(vcpu);
2438         else if (is_pae(vcpu))
2439                 r = paging32E_init_context(vcpu);
2440         else
2441                 r = paging32_init_context(vcpu);
2442
2443         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2444
2445         return r;
2446 }
2447
2448 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2449 {
2450         vcpu->arch.update_pte.pfn = bad_pfn;
2451
2452         if (tdp_enabled)
2453                 return init_kvm_tdp_mmu(vcpu);
2454         else
2455                 return init_kvm_softmmu(vcpu);
2456 }
2457
2458 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2459 {
2460         ASSERT(vcpu);
2461         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2462                 vcpu->arch.mmu.free(vcpu);
2463                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2464         }
2465 }
2466
2467 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2468 {
2469         destroy_kvm_mmu(vcpu);
2470         return init_kvm_mmu(vcpu);
2471 }
2472 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2473
2474 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2475 {
2476         int r;
2477
2478         r = mmu_topup_memory_caches(vcpu);
2479         if (r)
2480                 goto out;
2481         spin_lock(&vcpu->kvm->mmu_lock);
2482         kvm_mmu_free_some_pages(vcpu);
2483         r = mmu_alloc_roots(vcpu);
2484         mmu_sync_roots(vcpu);
2485         spin_unlock(&vcpu->kvm->mmu_lock);
2486         if (r)
2487                 goto out;
2488         /* set_cr3() should ensure TLB has been flushed */
2489         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2490 out:
2491         return r;
2492 }
2493 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2494
2495 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2496 {
2497         mmu_free_roots(vcpu);
2498 }
2499
2500 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2501                                   struct kvm_mmu_page *sp,
2502                                   u64 *spte)
2503 {
2504         u64 pte;
2505         struct kvm_mmu_page *child;
2506
2507         pte = *spte;
2508         if (is_shadow_present_pte(pte)) {
2509                 if (is_last_spte(pte, sp->role.level))
2510                         rmap_remove(vcpu->kvm, spte);
2511                 else {
2512                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2513                         mmu_page_remove_parent_pte(child, spte);
2514                 }
2515         }
2516         __set_spte(spte, shadow_trap_nonpresent_pte);
2517         if (is_large_pte(pte))
2518                 --vcpu->kvm->stat.lpages;
2519 }
2520
2521 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2522                                   struct kvm_mmu_page *sp,
2523                                   u64 *spte,
2524                                   const void *new)
2525 {
2526         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2527                 ++vcpu->kvm->stat.mmu_pde_zapped;
2528                 return;
2529         }
2530
2531         ++vcpu->kvm->stat.mmu_pte_updated;
2532         if (sp->role.glevels == PT32_ROOT_LEVEL)
2533                 paging32_update_pte(vcpu, sp, spte, new);
2534         else
2535                 paging64_update_pte(vcpu, sp, spte, new);
2536 }
2537
2538 static bool need_remote_flush(u64 old, u64 new)
2539 {
2540         if (!is_shadow_present_pte(old))
2541                 return false;
2542         if (!is_shadow_present_pte(new))
2543                 return true;
2544         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2545                 return true;
2546         old ^= PT64_NX_MASK;
2547         new ^= PT64_NX_MASK;
2548         return (old & ~new & PT64_PERM_MASK) != 0;
2549 }
2550
2551 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2552 {
2553         if (need_remote_flush(old, new))
2554                 kvm_flush_remote_tlbs(vcpu->kvm);
2555         else
2556                 kvm_mmu_flush_tlb(vcpu);
2557 }
2558
2559 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2560 {
2561         u64 *spte = vcpu->arch.last_pte_updated;
2562
2563         return !!(spte && (*spte & shadow_accessed_mask));
2564 }
2565
2566 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2567                                           const u8 *new, int bytes)
2568 {
2569         gfn_t gfn;
2570         int r;
2571         u64 gpte = 0;
2572         pfn_t pfn;
2573
2574         if (bytes != 4 && bytes != 8)
2575                 return;
2576
2577         /*
2578          * Assume that the pte write on a page table of the same type
2579          * as the current vcpu paging mode.  This is nearly always true
2580          * (might be false while changing modes).  Note it is verified later
2581          * by update_pte().
2582          */
2583         if (is_pae(vcpu)) {
2584                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2585                 if ((bytes == 4) && (gpa % 4 == 0)) {
2586                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2587                         if (r)
2588                                 return;
2589                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2590                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2591                         memcpy((void *)&gpte, new, 8);
2592                 }
2593         } else {
2594                 if ((bytes == 4) && (gpa % 4 == 0))
2595                         memcpy((void *)&gpte, new, 4);
2596         }
2597         if (!is_present_gpte(gpte))
2598                 return;
2599         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2600
2601         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2602         smp_rmb();
2603         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2604
2605         if (is_error_pfn(pfn)) {
2606                 kvm_release_pfn_clean(pfn);
2607                 return;
2608         }
2609         vcpu->arch.update_pte.gfn = gfn;
2610         vcpu->arch.update_pte.pfn = pfn;
2611 }
2612
2613 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2614 {
2615         u64 *spte = vcpu->arch.last_pte_updated;
2616
2617         if (spte
2618             && vcpu->arch.last_pte_gfn == gfn
2619             && shadow_accessed_mask
2620             && !(*spte & shadow_accessed_mask)
2621             && is_shadow_present_pte(*spte))
2622                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2623 }
2624
2625 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2626                        const u8 *new, int bytes,
2627                        bool guest_initiated)
2628 {
2629         gfn_t gfn = gpa >> PAGE_SHIFT;
2630         struct kvm_mmu_page *sp;
2631         struct hlist_node *node, *n;
2632         struct hlist_head *bucket;
2633         unsigned index;
2634         u64 entry, gentry;
2635         u64 *spte;
2636         unsigned offset = offset_in_page(gpa);
2637         unsigned pte_size;
2638         unsigned page_offset;
2639         unsigned misaligned;
2640         unsigned quadrant;
2641         int level;
2642         int flooded = 0;
2643         int npte;
2644         int r;
2645
2646         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2647         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2648         spin_lock(&vcpu->kvm->mmu_lock);
2649         kvm_mmu_access_page(vcpu, gfn);
2650         kvm_mmu_free_some_pages(vcpu);
2651         ++vcpu->kvm->stat.mmu_pte_write;
2652         kvm_mmu_audit(vcpu, "pre pte write");
2653         if (guest_initiated) {
2654                 if (gfn == vcpu->arch.last_pt_write_gfn
2655                     && !last_updated_pte_accessed(vcpu)) {
2656                         ++vcpu->arch.last_pt_write_count;
2657                         if (vcpu->arch.last_pt_write_count >= 3)
2658                                 flooded = 1;
2659                 } else {
2660                         vcpu->arch.last_pt_write_gfn = gfn;
2661                         vcpu->arch.last_pt_write_count = 1;
2662                         vcpu->arch.last_pte_updated = NULL;
2663                 }
2664         }
2665         index = kvm_page_table_hashfn(gfn);
2666         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2667         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2668                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2669                         continue;
2670                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2671                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2672                 misaligned |= bytes < 4;
2673                 if (misaligned || flooded) {
2674                         /*
2675                          * Misaligned accesses are too much trouble to fix
2676                          * up; also, they usually indicate a page is not used
2677                          * as a page table.
2678                          *
2679                          * If we're seeing too many writes to a page,
2680                          * it may no longer be a page table, or we may be
2681                          * forking, in which case it is better to unmap the
2682                          * page.
2683                          */
2684                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2685                                  gpa, bytes, sp->role.word);
2686                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2687                                 n = bucket->first;
2688                         ++vcpu->kvm->stat.mmu_flooded;
2689                         continue;
2690                 }
2691                 page_offset = offset;
2692                 level = sp->role.level;
2693                 npte = 1;
2694                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2695                         page_offset <<= 1;      /* 32->64 */
2696                         /*
2697                          * A 32-bit pde maps 4MB while the shadow pdes map
2698                          * only 2MB.  So we need to double the offset again
2699                          * and zap two pdes instead of one.
2700                          */
2701                         if (level == PT32_ROOT_LEVEL) {
2702                                 page_offset &= ~7; /* kill rounding error */
2703                                 page_offset <<= 1;
2704                                 npte = 2;
2705                         }
2706                         quadrant = page_offset >> PAGE_SHIFT;
2707                         page_offset &= ~PAGE_MASK;
2708                         if (quadrant != sp->role.quadrant)
2709                                 continue;
2710                 }
2711                 spte = &sp->spt[page_offset / sizeof(*spte)];
2712                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2713                         gentry = 0;
2714                         r = kvm_read_guest_atomic(vcpu->kvm,
2715                                                   gpa & ~(u64)(pte_size - 1),
2716                                                   &gentry, pte_size);
2717                         new = (const void *)&gentry;
2718                         if (r < 0)
2719                                 new = NULL;
2720                 }
2721                 while (npte--) {
2722                         entry = *spte;
2723                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2724                         if (new)
2725                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2726                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2727                         ++spte;
2728                 }
2729         }
2730         kvm_mmu_audit(vcpu, "post pte write");
2731         spin_unlock(&vcpu->kvm->mmu_lock);
2732         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2733                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2734                 vcpu->arch.update_pte.pfn = bad_pfn;
2735         }
2736 }
2737
2738 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2739 {
2740         gpa_t gpa;
2741         int r;
2742
2743         if (tdp_enabled)
2744                 return 0;
2745
2746         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2747
2748         spin_lock(&vcpu->kvm->mmu_lock);
2749         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2750         spin_unlock(&vcpu->kvm->mmu_lock);
2751         return r;
2752 }
2753 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2754
2755 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2756 {
2757         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2758                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2759                 struct kvm_mmu_page *sp;
2760
2761                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2762                                   struct kvm_mmu_page, link);
2763                 kvm_mmu_zap_page(vcpu->kvm, sp);
2764                 ++vcpu->kvm->stat.mmu_recycled;
2765         }
2766 }
2767
2768 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2769 {
2770         int r;
2771         enum emulation_result er;
2772
2773         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2774         if (r < 0)
2775                 goto out;
2776
2777         if (!r) {
2778                 r = 1;
2779                 goto out;
2780         }
2781
2782         r = mmu_topup_memory_caches(vcpu);
2783         if (r)
2784                 goto out;
2785
2786         er = emulate_instruction(vcpu, cr2, error_code, 0);
2787
2788         switch (er) {
2789         case EMULATE_DONE:
2790                 return 1;
2791         case EMULATE_DO_MMIO:
2792                 ++vcpu->stat.mmio_exits;
2793                 return 0;
2794         case EMULATE_FAIL:
2795                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2796                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2797                 vcpu->run->internal.ndata = 0;
2798                 return 0;
2799         default:
2800                 BUG();
2801         }
2802 out:
2803         return r;
2804 }
2805 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2806
2807 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2808 {
2809         vcpu->arch.mmu.invlpg(vcpu, gva);
2810         kvm_mmu_flush_tlb(vcpu);
2811         ++vcpu->stat.invlpg;
2812 }
2813 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2814
2815 void kvm_enable_tdp(void)
2816 {
2817         tdp_enabled = true;
2818 }
2819 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2820
2821 void kvm_disable_tdp(void)
2822 {
2823         tdp_enabled = false;
2824 }
2825 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2826
2827 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2828 {
2829         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2830 }
2831
2832 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2833 {
2834         struct page *page;
2835         int i;
2836
2837         ASSERT(vcpu);
2838
2839         /*
2840          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2841          * Therefore we need to allocate shadow page tables in the first
2842          * 4GB of memory, which happens to fit the DMA32 zone.
2843          */
2844         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2845         if (!page)
2846                 goto error_1;
2847         vcpu->arch.mmu.pae_root = page_address(page);
2848         for (i = 0; i < 4; ++i)
2849                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2850
2851         return 0;
2852
2853 error_1:
2854         free_mmu_pages(vcpu);
2855         return -ENOMEM;
2856 }
2857
2858 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2859 {
2860         ASSERT(vcpu);
2861         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2862
2863         return alloc_mmu_pages(vcpu);
2864 }
2865
2866 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2867 {
2868         ASSERT(vcpu);
2869         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2870
2871         return init_kvm_mmu(vcpu);
2872 }
2873
2874 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2875 {
2876         ASSERT(vcpu);
2877
2878         destroy_kvm_mmu(vcpu);
2879         free_mmu_pages(vcpu);
2880         mmu_free_memory_caches(vcpu);
2881 }
2882
2883 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2884 {
2885         struct kvm_mmu_page *sp;
2886
2887         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2888                 int i;
2889                 u64 *pt;
2890
2891                 if (!test_bit(slot, sp->slot_bitmap))
2892                         continue;
2893
2894                 pt = sp->spt;
2895                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2896                         /* avoid RMW */
2897                         if (pt[i] & PT_WRITABLE_MASK)
2898                                 pt[i] &= ~PT_WRITABLE_MASK;
2899         }
2900         kvm_flush_remote_tlbs(kvm);
2901 }
2902
2903 void kvm_mmu_zap_all(struct kvm *kvm)
2904 {
2905         struct kvm_mmu_page *sp, *node;
2906
2907         spin_lock(&kvm->mmu_lock);
2908         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2909                 if (kvm_mmu_zap_page(kvm, sp))
2910                         node = container_of(kvm->arch.active_mmu_pages.next,
2911                                             struct kvm_mmu_page, link);
2912         spin_unlock(&kvm->mmu_lock);
2913
2914         kvm_flush_remote_tlbs(kvm);
2915 }
2916
2917 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2918 {
2919         struct kvm_mmu_page *page;
2920
2921         page = container_of(kvm->arch.active_mmu_pages.prev,
2922                             struct kvm_mmu_page, link);
2923         kvm_mmu_zap_page(kvm, page);
2924 }
2925
2926 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2927 {
2928         struct kvm *kvm;
2929         struct kvm *kvm_freed = NULL;
2930         int cache_count = 0;
2931
2932         spin_lock(&kvm_lock);
2933
2934         list_for_each_entry(kvm, &vm_list, vm_list) {
2935                 int npages, idx;
2936
2937                 idx = srcu_read_lock(&kvm->srcu);
2938                 spin_lock(&kvm->mmu_lock);
2939                 npages = kvm->arch.n_alloc_mmu_pages -
2940                          kvm->arch.n_free_mmu_pages;
2941                 cache_count += npages;
2942                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2943                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2944                         cache_count--;
2945                         kvm_freed = kvm;
2946                 }
2947                 nr_to_scan--;
2948
2949                 spin_unlock(&kvm->mmu_lock);
2950                 srcu_read_unlock(&kvm->srcu, idx);
2951         }
2952         if (kvm_freed)
2953                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2954
2955         spin_unlock(&kvm_lock);
2956
2957         return cache_count;
2958 }
2959
2960 static struct shrinker mmu_shrinker = {
2961         .shrink = mmu_shrink,
2962         .seeks = DEFAULT_SEEKS * 10,
2963 };
2964
2965 static void mmu_destroy_caches(void)
2966 {
2967         if (pte_chain_cache)
2968                 kmem_cache_destroy(pte_chain_cache);
2969         if (rmap_desc_cache)
2970                 kmem_cache_destroy(rmap_desc_cache);
2971         if (mmu_page_header_cache)
2972                 kmem_cache_destroy(mmu_page_header_cache);
2973 }
2974
2975 void kvm_mmu_module_exit(void)
2976 {
2977         mmu_destroy_caches();
2978         unregister_shrinker(&mmu_shrinker);
2979 }
2980
2981 int kvm_mmu_module_init(void)
2982 {
2983         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2984                                             sizeof(struct kvm_pte_chain),
2985                                             0, 0, NULL);
2986         if (!pte_chain_cache)
2987                 goto nomem;
2988         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2989                                             sizeof(struct kvm_rmap_desc),
2990                                             0, 0, NULL);
2991         if (!rmap_desc_cache)
2992                 goto nomem;
2993
2994         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2995                                                   sizeof(struct kvm_mmu_page),
2996                                                   0, 0, NULL);
2997         if (!mmu_page_header_cache)
2998                 goto nomem;
2999
3000         register_shrinker(&mmu_shrinker);
3001
3002         return 0;
3003
3004 nomem:
3005         mmu_destroy_caches();
3006         return -ENOMEM;
3007 }
3008
3009 /*
3010  * Caculate mmu pages needed for kvm.
3011  */
3012 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3013 {
3014         int i;
3015         unsigned int nr_mmu_pages;
3016         unsigned int  nr_pages = 0;
3017         struct kvm_memslots *slots;
3018
3019         slots = rcu_dereference(kvm->memslots);
3020         for (i = 0; i < slots->nmemslots; i++)
3021                 nr_pages += slots->memslots[i].npages;
3022
3023         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3024         nr_mmu_pages = max(nr_mmu_pages,
3025                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3026
3027         return nr_mmu_pages;
3028 }
3029
3030 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3031                                 unsigned len)
3032 {
3033         if (len > buffer->len)
3034                 return NULL;
3035         return buffer->ptr;
3036 }
3037
3038 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3039                                 unsigned len)
3040 {
3041         void *ret;
3042
3043         ret = pv_mmu_peek_buffer(buffer, len);
3044         if (!ret)
3045                 return ret;
3046         buffer->ptr += len;
3047         buffer->len -= len;
3048         buffer->processed += len;
3049         return ret;
3050 }
3051
3052 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3053                              gpa_t addr, gpa_t value)
3054 {
3055         int bytes = 8;
3056         int r;
3057
3058         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3059                 bytes = 4;
3060
3061         r = mmu_topup_memory_caches(vcpu);
3062         if (r)
3063                 return r;
3064
3065         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3066                 return -EFAULT;
3067
3068         return 1;
3069 }
3070
3071 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3072 {
3073         kvm_set_cr3(vcpu, vcpu->arch.cr3);
3074         return 1;
3075 }
3076
3077 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3078 {
3079         spin_lock(&vcpu->kvm->mmu_lock);
3080         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3081         spin_unlock(&vcpu->kvm->mmu_lock);
3082         return 1;
3083 }
3084
3085 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3086                              struct kvm_pv_mmu_op_buffer *buffer)
3087 {
3088         struct kvm_mmu_op_header *header;
3089
3090         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3091         if (!header)
3092                 return 0;
3093         switch (header->op) {
3094         case KVM_MMU_OP_WRITE_PTE: {
3095                 struct kvm_mmu_op_write_pte *wpte;
3096
3097                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3098                 if (!wpte)
3099                         return 0;
3100                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3101                                         wpte->pte_val);
3102         }
3103         case KVM_MMU_OP_FLUSH_TLB: {
3104                 struct kvm_mmu_op_flush_tlb *ftlb;
3105
3106                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3107                 if (!ftlb)
3108                         return 0;
3109                 return kvm_pv_mmu_flush_tlb(vcpu);
3110         }
3111         case KVM_MMU_OP_RELEASE_PT: {
3112                 struct kvm_mmu_op_release_pt *rpt;
3113
3114                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3115                 if (!rpt)
3116                         return 0;
3117                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3118         }
3119         default: return 0;
3120         }
3121 }
3122
3123 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3124                   gpa_t addr, unsigned long *ret)
3125 {
3126         int r;
3127         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3128
3129         buffer->ptr = buffer->buf;
3130         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3131         buffer->processed = 0;
3132
3133         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3134         if (r)
3135                 goto out;
3136
3137         while (buffer->len) {
3138                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3139                 if (r < 0)
3140                         goto out;
3141                 if (r == 0)
3142                         break;
3143         }
3144
3145         r = 1;
3146 out:
3147         *ret = buffer->processed;
3148         return r;
3149 }
3150
3151 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3152 {
3153         struct kvm_shadow_walk_iterator iterator;
3154         int nr_sptes = 0;
3155
3156         spin_lock(&vcpu->kvm->mmu_lock);
3157         for_each_shadow_entry(vcpu, addr, iterator) {
3158                 sptes[iterator.level-1] = *iterator.sptep;
3159                 nr_sptes++;
3160                 if (!is_shadow_present_pte(*iterator.sptep))
3161                         break;
3162         }
3163         spin_unlock(&vcpu->kvm->mmu_lock);
3164
3165         return nr_sptes;
3166 }
3167 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3168
3169 #ifdef AUDIT
3170
3171 static const char *audit_msg;
3172
3173 static gva_t canonicalize(gva_t gva)
3174 {
3175 #ifdef CONFIG_X86_64
3176         gva = (long long)(gva << 16) >> 16;
3177 #endif
3178         return gva;
3179 }
3180
3181
3182 typedef void (*inspect_spte_fn) (struct kvm *kvm, struct kvm_mmu_page *sp,
3183                                  u64 *sptep);
3184
3185 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3186                             inspect_spte_fn fn)
3187 {
3188         int i;
3189
3190         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3191                 u64 ent = sp->spt[i];
3192
3193                 if (is_shadow_present_pte(ent)) {
3194                         if (!is_last_spte(ent, sp->role.level)) {
3195                                 struct kvm_mmu_page *child;
3196                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3197                                 __mmu_spte_walk(kvm, child, fn);
3198                         } else
3199                                 fn(kvm, sp, &sp->spt[i]);
3200                 }
3201         }
3202 }
3203
3204 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3205 {
3206         int i;
3207         struct kvm_mmu_page *sp;
3208
3209         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3210                 return;
3211         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3212                 hpa_t root = vcpu->arch.mmu.root_hpa;
3213                 sp = page_header(root);
3214                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3215                 return;
3216         }
3217         for (i = 0; i < 4; ++i) {
3218                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3219
3220                 if (root && VALID_PAGE(root)) {
3221                         root &= PT64_BASE_ADDR_MASK;
3222                         sp = page_header(root);
3223                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3224                 }
3225         }
3226         return;
3227 }
3228
3229 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3230                                 gva_t va, int level)
3231 {
3232         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3233         int i;
3234         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3235
3236         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3237                 u64 ent = pt[i];
3238
3239                 if (ent == shadow_trap_nonpresent_pte)
3240                         continue;
3241
3242                 va = canonicalize(va);
3243                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3244                         audit_mappings_page(vcpu, ent, va, level - 1);
3245                 else {
3246                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3247                         gfn_t gfn = gpa >> PAGE_SHIFT;
3248                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3249                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3250
3251                         if (is_error_pfn(pfn)) {
3252                                 kvm_release_pfn_clean(pfn);
3253                                 continue;
3254                         }
3255
3256                         if (is_shadow_present_pte(ent)
3257                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3258                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3259                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3260                                        audit_msg, vcpu->arch.mmu.root_level,
3261                                        va, gpa, hpa, ent,
3262                                        is_shadow_present_pte(ent));
3263                         else if (ent == shadow_notrap_nonpresent_pte
3264                                  && !is_error_hpa(hpa))
3265                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3266                                        " valid guest gva %lx\n", audit_msg, va);
3267                         kvm_release_pfn_clean(pfn);
3268
3269                 }
3270         }
3271 }
3272
3273 static void audit_mappings(struct kvm_vcpu *vcpu)
3274 {
3275         unsigned i;
3276
3277         if (vcpu->arch.mmu.root_level == 4)
3278                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3279         else
3280                 for (i = 0; i < 4; ++i)
3281                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3282                                 audit_mappings_page(vcpu,
3283                                                     vcpu->arch.mmu.pae_root[i],
3284                                                     i << 30,
3285                                                     2);
3286 }
3287
3288 static int count_rmaps(struct kvm_vcpu *vcpu)
3289 {
3290         int nmaps = 0;
3291         int i, j, k, idx;
3292
3293         idx = srcu_read_lock(&kvm->srcu);
3294         slots = rcu_dereference(kvm->memslots);
3295         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3296                 struct kvm_memory_slot *m = &slots->memslots[i];
3297                 struct kvm_rmap_desc *d;
3298
3299                 for (j = 0; j < m->npages; ++j) {
3300                         unsigned long *rmapp = &m->rmap[j];
3301
3302                         if (!*rmapp)
3303                                 continue;
3304                         if (!(*rmapp & 1)) {
3305                                 ++nmaps;
3306                                 continue;
3307                         }
3308                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3309                         while (d) {
3310                                 for (k = 0; k < RMAP_EXT; ++k)
3311                                         if (d->sptes[k])
3312                                                 ++nmaps;
3313                                         else
3314                                                 break;
3315                                 d = d->more;
3316                         }
3317                 }
3318         }
3319         srcu_read_unlock(&kvm->srcu, idx);
3320         return nmaps;
3321 }
3322
3323 void inspect_spte_has_rmap(struct kvm *kvm, struct kvm_mmu_page *sp, u64 *sptep)
3324 {
3325         unsigned long *rmapp;
3326         struct kvm_mmu_page *rev_sp;
3327         gfn_t gfn;
3328
3329         if (*sptep & PT_WRITABLE_MASK) {
3330                 rev_sp = page_header(__pa(sptep));
3331                 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3332
3333                 if (!gfn_to_memslot(kvm, gfn)) {
3334                         if (!printk_ratelimit())
3335                                 return;
3336                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3337                                          audit_msg, gfn);
3338                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3339                                         audit_msg, sptep - rev_sp->spt,
3340                                         rev_sp->gfn);
3341                         dump_stack();
3342                         return;
3343                 }
3344
3345                 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3346                                     is_large_pte(*sptep));
3347                 if (!*rmapp) {
3348                         if (!printk_ratelimit())
3349                                 return;
3350                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3351                                          audit_msg, *sptep);
3352                         dump_stack();
3353                 }
3354         }
3355
3356 }
3357
3358 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3359 {
3360         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3361 }
3362
3363 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3364 {
3365         struct kvm_mmu_page *sp;
3366         int i;
3367
3368         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3369                 u64 *pt = sp->spt;
3370
3371                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3372                         continue;
3373
3374                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3375                         u64 ent = pt[i];
3376
3377                         if (!(ent & PT_PRESENT_MASK))
3378                                 continue;
3379                         if (!(ent & PT_WRITABLE_MASK))
3380                                 continue;
3381                         inspect_spte_has_rmap(vcpu->kvm, sp, &pt[i]);
3382                 }
3383         }
3384         return;
3385 }
3386
3387 static void audit_rmap(struct kvm_vcpu *vcpu)
3388 {
3389         check_writable_mappings_rmap(vcpu);
3390         count_rmaps(vcpu);
3391 }
3392
3393 static void audit_write_protection(struct kvm_vcpu *vcpu)
3394 {
3395         struct kvm_mmu_page *sp;
3396         struct kvm_memory_slot *slot;
3397         unsigned long *rmapp;
3398         u64 *spte;
3399         gfn_t gfn;
3400
3401         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3402                 if (sp->role.direct)
3403                         continue;
3404                 if (sp->unsync)
3405                         continue;
3406
3407                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3408                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3409                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3410
3411                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3412                 while (spte) {
3413                         if (*spte & PT_WRITABLE_MASK)
3414                                 printk(KERN_ERR "%s: (%s) shadow page has "
3415                                 "writable mappings: gfn %lx role %x\n",
3416                                __func__, audit_msg, sp->gfn,
3417                                sp->role.word);
3418                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3419                 }
3420         }
3421 }
3422
3423 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3424 {
3425         int olddbg = dbg;
3426
3427         dbg = 0;
3428         audit_msg = msg;
3429         audit_rmap(vcpu);
3430         audit_write_protection(vcpu);
3431         if (strcmp("pre pte write", audit_msg) != 0)
3432                 audit_mappings(vcpu);
3433         audit_writable_sptes_have_rmaps(vcpu);
3434         dbg = olddbg;
3435 }
3436
3437 #endif