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

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "mmu.h"
  21 #include "kvm_cache_regs.h"
  22
  23 #include <linux/kvm_host.h>
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29 #include <linux/swap.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/compiler.h>
  32 #include <linux/srcu.h>
  33
  34 #include <asm/page.h>
  35 #include <asm/cmpxchg.h>
  36 #include <asm/io.h>
  37 #include <asm/vmx.h>
  38
  39 /*
  40  * When setting this variable to true it enables Two-Dimensional-Paging
  41  * where the hardware walks 2 page tables:
  42  * 1. the guest-virtual to guest-physical
  43  * 2. while doing 1. it walks guest-physical to host-physical
  44  * If the hardware supports that we don't need to do shadow paging.
  45  */
  46 bool tdp_enabled = false;
  47
  48 #undef MMU_DEBUG
  49
  50 #undef AUDIT
  51
  52 #ifdef AUDIT
  53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  54 #else
  55 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  56 #endif
  57
  58 #ifdef MMU_DEBUG
  59
  60 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  61 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  62
  63 #else
  64
  65 #define pgprintk(x...) do { } while (0)
  66 #define rmap_printk(x...) do { } while (0)
  67
  68 #endif
  69
  70 #if defined(MMU_DEBUG) || defined(AUDIT)
  71 static int dbg = 0;
  72 module_param(dbg, bool, 0644);
  73 #endif
  74
  75 static int oos_shadow = 1;
  76 module_param(oos_shadow, bool, 0644);
  77
  78 #ifndef MMU_DEBUG
  79 #define ASSERT(x) do { } while (0)
  80 #else
  81 #define ASSERT(x)                                                       \
  82         if (!(x)) {                                                     \
  83                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  84                        __FILE__, __LINE__, #x);                         \
  85         }
  86 #endif
  87
  88 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  89 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  90
  91 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  92
  93 #define PT64_LEVEL_BITS 9
  94
  95 #define PT64_LEVEL_SHIFT(level) \
  96                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  97
  98 #define PT64_LEVEL_MASK(level) \
  99                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 100
 101 #define PT64_INDEX(address, level)\
 102         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 103
 104
 105 #define PT32_LEVEL_BITS 10
 106
 107 #define PT32_LEVEL_SHIFT(level) \
 108                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 109
 110 #define PT32_LEVEL_MASK(level) \
 111                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 112 #define PT32_LVL_OFFSET_MASK(level) \
 113         (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 114                                                 * PT32_LEVEL_BITS))) - 1))
 115
 116 #define PT32_INDEX(address, level)\
 117         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 118
 119
 120 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 121 #define PT64_DIR_BASE_ADDR_MASK \
 122         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 123 #define PT64_LVL_ADDR_MASK(level) \
 124         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 125                                                 * PT64_LEVEL_BITS))) - 1))
 126 #define PT64_LVL_OFFSET_MASK(level) \
 127         (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 128                                                 * PT64_LEVEL_BITS))) - 1))
 129
 130 #define PT32_BASE_ADDR_MASK PAGE_MASK
 131 #define PT32_DIR_BASE_ADDR_MASK \
 132         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 133 #define PT32_LVL_ADDR_MASK(level) \
 134         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 135                                             * PT32_LEVEL_BITS))) - 1))
 136
 137 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 138                         | PT64_NX_MASK)
 139
 140 #define PFERR_PRESENT_MASK (1U << 0)
 141 #define PFERR_WRITE_MASK (1U << 1)
 142 #define PFERR_USER_MASK (1U << 2)
 143 #define PFERR_RSVD_MASK (1U << 3)
 144 #define PFERR_FETCH_MASK (1U << 4)
 145
 146 #define RMAP_EXT 4
 147
 148 #define ACC_EXEC_MASK    1
 149 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 150 #define ACC_USER_MASK    PT_USER_MASK
 151 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 152
 153 #define CREATE_TRACE_POINTS
 154 #include "mmutrace.h"
 155
 156 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 157
 158 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 159
 160 struct kvm_rmap_desc {
 161         u64 *sptes[RMAP_EXT];
 162         struct kvm_rmap_desc *more;
 163 };
 164
 165 struct kvm_shadow_walk_iterator {
 166         u64 addr;
 167         hpa_t shadow_addr;
 168         int level;
 169         u64 *sptep;
 170         unsigned index;
 171 };
 172
 173 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
 174         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
 175              shadow_walk_okay(&(_walker));                      \
 176              shadow_walk_next(&(_walker)))
 177
 178
 179 struct kvm_unsync_walk {
 180         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 181 };
 182
 183 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 184
 185 static struct kmem_cache *pte_chain_cache;
 186 static struct kmem_cache *rmap_desc_cache;
 187 static struct kmem_cache *mmu_page_header_cache;
 188
 189 static u64 __read_mostly shadow_trap_nonpresent_pte;
 190 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 191 static u64 __read_mostly shadow_base_present_pte;
 192 static u64 __read_mostly shadow_nx_mask;
 193 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 194 static u64 __read_mostly shadow_user_mask;
 195 static u64 __read_mostly shadow_accessed_mask;
 196 static u64 __read_mostly shadow_dirty_mask;
 197
 198 static inline u64 rsvd_bits(int s, int e)
 199 {
 200         return ((1ULL << (e - s + 1)) - 1) << s;
 201 }
 202
 203 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 204 {
 205         shadow_trap_nonpresent_pte = trap_pte;
 206         shadow_notrap_nonpresent_pte = notrap_pte;
 207 }
 208 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 209
 210 void kvm_mmu_set_base_ptes(u64 base_pte)
 211 {
 212         shadow_base_present_pte = base_pte;
 213 }
 214 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 215
 216 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 217                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 218 {
 219         shadow_user_mask = user_mask;
 220         shadow_accessed_mask = accessed_mask;
 221         shadow_dirty_mask = dirty_mask;
 222         shadow_nx_mask = nx_mask;
 223         shadow_x_mask = x_mask;
 224 }
 225 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 226
 227 static int is_write_protection(struct kvm_vcpu *vcpu)
 228 {
 229         return vcpu->arch.cr0 & X86_CR0_WP;
 230 }
 231
 232 static int is_cpuid_PSE36(void)
 233 {
 234         return 1;
 235 }
 236
 237 static int is_nx(struct kvm_vcpu *vcpu)
 238 {
 239         return vcpu->arch.shadow_efer & EFER_NX;
 240 }
 241
 242 static int is_shadow_present_pte(u64 pte)
 243 {
 244         return pte != shadow_trap_nonpresent_pte
 245                 && pte != shadow_notrap_nonpresent_pte;
 246 }
 247
 248 static int is_large_pte(u64 pte)
 249 {
 250         return pte & PT_PAGE_SIZE_MASK;
 251 }
 252
 253 static int is_writeble_pte(unsigned long pte)
 254 {
 255         return pte & PT_WRITABLE_MASK;
 256 }
 257
 258 static int is_dirty_gpte(unsigned long pte)
 259 {
 260         return pte & PT_DIRTY_MASK;
 261 }
 262
 263 static int is_rmap_spte(u64 pte)
 264 {
 265         return is_shadow_present_pte(pte);
 266 }
 267
 268 static int is_last_spte(u64 pte, int level)
 269 {
 270         if (level == PT_PAGE_TABLE_LEVEL)
 271                 return 1;
 272         if (is_large_pte(pte))
 273                 return 1;
 274         return 0;
 275 }
 276
 277 static pfn_t spte_to_pfn(u64 pte)
 278 {
 279         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 280 }
 281
 282 static gfn_t pse36_gfn_delta(u32 gpte)
 283 {
 284         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 285
 286         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 287 }
 288
 289 static void __set_spte(u64 *sptep, u64 spte)
 290 {
 291 #ifdef CONFIG_X86_64
 292         set_64bit((unsigned long *)sptep, spte);
 293 #else
 294         set_64bit((unsigned long long *)sptep, spte);
 295 #endif
 296 }
 297
 298 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 299                                   struct kmem_cache *base_cache, int min)
 300 {
 301         void *obj;
 302
 303         if (cache->nobjs >= min)
 304                 return 0;
 305         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 306                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 307                 if (!obj)
 308                         return -ENOMEM;
 309                 cache->objects[cache->nobjs++] = obj;
 310         }
 311         return 0;
 312 }
 313
 314 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 315 {
 316         while (mc->nobjs)
 317                 kfree(mc->objects[--mc->nobjs]);
 318 }
 319
 320 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 321                                        int min)
 322 {
 323         struct page *page;
 324
 325         if (cache->nobjs >= min)
 326                 return 0;
 327         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 328                 page = alloc_page(GFP_KERNEL);
 329                 if (!page)
 330                         return -ENOMEM;
 331                 set_page_private(page, 0);
 332                 cache->objects[cache->nobjs++] = page_address(page);
 333         }
 334         return 0;
 335 }
 336
 337 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 338 {
 339         while (mc->nobjs)
 340                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 341 }
 342
 343 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 344 {
 345         int r;
 346
 347         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 348                                    pte_chain_cache, 4);
 349         if (r)
 350                 goto out;
 351         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 352                                    rmap_desc_cache, 4);
 353         if (r)
 354                 goto out;
 355         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 356         if (r)
 357                 goto out;
 358         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 359                                    mmu_page_header_cache, 4);
 360 out:
 361         return r;
 362 }
 363
 364 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 365 {
 366         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 367         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 368         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 369         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 370 }
 371
 372 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 373                                     size_t size)
 374 {
 375         void *p;
 376
 377         BUG_ON(!mc->nobjs);
 378         p = mc->objects[--mc->nobjs];
 379         return p;
 380 }
 381
 382 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 383 {
 384         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 385                                       sizeof(struct kvm_pte_chain));
 386 }
 387
 388 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 389 {
 390         kfree(pc);
 391 }
 392
 393 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 394 {
 395         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 396                                       sizeof(struct kvm_rmap_desc));
 397 }
 398
 399 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 400 {
 401         kfree(rd);
 402 }
 403
 404 /*
 405  * Return the pointer to the largepage write count for a given
 406  * gfn, handling slots that are not large page aligned.
 407  */
 408 static int *slot_largepage_idx(gfn_t gfn,
 409                                struct kvm_memory_slot *slot,
 410                                int level)
 411 {
 412         unsigned long idx;
 413
 414         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 415               (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 416         return &slot->lpage_info[level - 2][idx].write_count;
 417 }
 418
 419 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 420 {
 421         struct kvm_memory_slot *slot;
 422         int *write_count;
 423         int i;
 424
 425         gfn = unalias_gfn(kvm, gfn);
 426
 427         slot = gfn_to_memslot_unaliased(kvm, gfn);
 428         for (i = PT_DIRECTORY_LEVEL;
 429              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 430                 write_count   = slot_largepage_idx(gfn, slot, i);
 431                 *write_count += 1;
 432         }
 433 }
 434
 435 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 436 {
 437         struct kvm_memory_slot *slot;
 438         int *write_count;
 439         int i;
 440
 441         gfn = unalias_gfn(kvm, gfn);
 442         for (i = PT_DIRECTORY_LEVEL;
 443              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 444                 slot          = gfn_to_memslot_unaliased(kvm, gfn);
 445                 write_count   = slot_largepage_idx(gfn, slot, i);
 446                 *write_count -= 1;
 447                 WARN_ON(*write_count < 0);
 448         }
 449 }
 450
 451 static int has_wrprotected_page(struct kvm *kvm,
 452                                 gfn_t gfn,
 453                                 int level)
 454 {
 455         struct kvm_memory_slot *slot;
 456         int *largepage_idx;
 457
 458         gfn = unalias_gfn(kvm, gfn);
 459         slot = gfn_to_memslot_unaliased(kvm, gfn);
 460         if (slot) {
 461                 largepage_idx = slot_largepage_idx(gfn, slot, level);
 462                 return *largepage_idx;
 463         }
 464
 465         return 1;
 466 }
 467
 468 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
 469 {
 470         unsigned long page_size = PAGE_SIZE;
 471         struct vm_area_struct *vma;
 472         unsigned long addr;
 473         int i, ret = 0;
 474
 475         addr = gfn_to_hva(kvm, gfn);
 476         if (kvm_is_error_hva(addr))
 477                 return PT_PAGE_TABLE_LEVEL;
 478
 479         down_read(&current->mm->mmap_sem);
 480         vma = find_vma(current->mm, addr);
 481         if (!vma)
 482                 goto out;
 483
 484         page_size = vma_kernel_pagesize(vma);
 485
 486 out:
 487         up_read(&current->mm->mmap_sem);
 488
 489         for (i = PT_PAGE_TABLE_LEVEL;
 490              i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
 491                 if (page_size >= KVM_HPAGE_SIZE(i))
 492                         ret = i;
 493                 else
 494                         break;
 495         }
 496
 497         return ret;
 498 }
 499
 500 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 501 {
 502         struct kvm_memory_slot *slot;
 503         int host_level;
 504         int level = PT_PAGE_TABLE_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         for (level = PT_DIRECTORY_LEVEL; level <= host_level; ++level)
 516                 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
 517                         break;
 518
 519         return level - 1;
 520 }
 521
 522 /*
 523  * Take gfn and return the reverse mapping to it.
 524  * Note: gfn must be unaliased before this function get called
 525  */
 526
 527 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
 528 {
 529         struct kvm_memory_slot *slot;
 530         unsigned long idx;
 531
 532         slot = gfn_to_memslot(kvm, gfn);
 533         if (likely(level == PT_PAGE_TABLE_LEVEL))
 534                 return &slot->rmap[gfn - slot->base_gfn];
 535
 536         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 537                 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 538
 539         return &slot->lpage_info[level - 2][idx].rmap_pde;
 540 }
 541
 542 /*
 543  * Reverse mapping data structures:
 544  *
 545  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 546  * that points to page_address(page).
 547  *
 548  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 549  * containing more mappings.
 550  *
 551  * Returns the number of rmap entries before the spte was added or zero if
 552  * the spte was not added.
 553  *
 554  */
 555 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 556 {
 557         struct kvm_mmu_page *sp;
 558         struct kvm_rmap_desc *desc;
 559         unsigned long *rmapp;
 560         int i, count = 0;
 561
 562         if (!is_rmap_spte(*spte))
 563                 return count;
 564         gfn = unalias_gfn(vcpu->kvm, gfn);
 565         sp = page_header(__pa(spte));
 566         sp->gfns[spte - sp->spt] = gfn;
 567         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 568         if (!*rmapp) {
 569                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 570                 *rmapp = (unsigned long)spte;
 571         } else if (!(*rmapp & 1)) {
 572                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 573                 desc = mmu_alloc_rmap_desc(vcpu);
 574                 desc->sptes[0] = (u64 *)*rmapp;
 575                 desc->sptes[1] = spte;
 576                 *rmapp = (unsigned long)desc | 1;
 577         } else {
 578                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 579                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 580                 while (desc->sptes[RMAP_EXT-1] && desc->more) {
 581                         desc = desc->more;
 582                         count += RMAP_EXT;
 583                 }
 584                 if (desc->sptes[RMAP_EXT-1]) {
 585                         desc->more = mmu_alloc_rmap_desc(vcpu);
 586                         desc = desc->more;
 587                 }
 588                 for (i = 0; desc->sptes[i]; ++i)
 589                         ;
 590                 desc->sptes[i] = spte;
 591         }
 592         return count;
 593 }
 594
 595 static void rmap_desc_remove_entry(unsigned long *rmapp,
 596                                    struct kvm_rmap_desc *desc,
 597                                    int i,
 598                                    struct kvm_rmap_desc *prev_desc)
 599 {
 600         int j;
 601
 602         for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
 603                 ;
 604         desc->sptes[i] = desc->sptes[j];
 605         desc->sptes[j] = NULL;
 606         if (j != 0)
 607                 return;
 608         if (!prev_desc && !desc->more)
 609                 *rmapp = (unsigned long)desc->sptes[0];
 610         else
 611                 if (prev_desc)
 612                         prev_desc->more = desc->more;
 613                 else
 614                         *rmapp = (unsigned long)desc->more | 1;
 615         mmu_free_rmap_desc(desc);
 616 }
 617
 618 static void rmap_remove(struct kvm *kvm, u64 *spte)
 619 {
 620         struct kvm_rmap_desc *desc;
 621         struct kvm_rmap_desc *prev_desc;
 622         struct kvm_mmu_page *sp;
 623         pfn_t pfn;
 624         unsigned long *rmapp;
 625         int i;
 626
 627         if (!is_rmap_spte(*spte))
 628                 return;
 629         sp = page_header(__pa(spte));
 630         pfn = spte_to_pfn(*spte);
 631         if (*spte & shadow_accessed_mask)
 632                 kvm_set_pfn_accessed(pfn);
 633         if (is_writeble_pte(*spte))
 634                 kvm_set_pfn_dirty(pfn);
 635         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], sp->role.level);
 636         if (!*rmapp) {
 637                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 638                 BUG();
 639         } else if (!(*rmapp & 1)) {
 640                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 641                 if ((u64 *)*rmapp != spte) {
 642                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 643                                spte, *spte);
 644                         BUG();
 645                 }
 646                 *rmapp = 0;
 647         } else {
 648                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 649                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 650                 prev_desc = NULL;
 651                 while (desc) {
 652                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
 653                                 if (desc->sptes[i] == spte) {
 654                                         rmap_desc_remove_entry(rmapp,
 655                                                                desc, i,
 656                                                                prev_desc);
 657                                         return;
 658                                 }
 659                         prev_desc = desc;
 660                         desc = desc->more;
 661                 }
 662                 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
 663                 BUG();
 664         }
 665 }
 666
 667 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 668 {
 669         struct kvm_rmap_desc *desc;
 670         struct kvm_rmap_desc *prev_desc;
 671         u64 *prev_spte;
 672         int i;
 673
 674         if (!*rmapp)
 675                 return NULL;
 676         else if (!(*rmapp & 1)) {
 677                 if (!spte)
 678                         return (u64 *)*rmapp;
 679                 return NULL;
 680         }
 681         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 682         prev_desc = NULL;
 683         prev_spte = NULL;
 684         while (desc) {
 685                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
 686                         if (prev_spte == spte)
 687                                 return desc->sptes[i];
 688                         prev_spte = desc->sptes[i];
 689                 }
 690                 desc = desc->more;
 691         }
 692         return NULL;
 693 }
 694
 695 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 696 {
 697         unsigned long *rmapp;
 698         u64 *spte;
 699         int i, write_protected = 0;
 700
 701         gfn = unalias_gfn(kvm, gfn);
 702         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
 703
 704         spte = rmap_next(kvm, rmapp, NULL);
 705         while (spte) {
 706                 BUG_ON(!spte);
 707                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 708                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 709                 if (is_writeble_pte(*spte)) {
 710                         __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
 711                         write_protected = 1;
 712                 }
 713                 spte = rmap_next(kvm, rmapp, spte);
 714         }
 715         if (write_protected) {
 716                 pfn_t pfn;
 717
 718                 spte = rmap_next(kvm, rmapp, NULL);
 719                 pfn = spte_to_pfn(*spte);
 720                 kvm_set_pfn_dirty(pfn);
 721         }
 722
 723         /* check for huge page mappings */
 724         for (i = PT_DIRECTORY_LEVEL;
 725              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 726                 rmapp = gfn_to_rmap(kvm, gfn, i);
 727                 spte = rmap_next(kvm, rmapp, NULL);
 728                 while (spte) {
 729                         BUG_ON(!spte);
 730                         BUG_ON(!(*spte & PT_PRESENT_MASK));
 731                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 732                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 733                         if (is_writeble_pte(*spte)) {
 734                                 rmap_remove(kvm, spte);
 735                                 --kvm->stat.lpages;
 736                                 __set_spte(spte, shadow_trap_nonpresent_pte);
 737                                 spte = NULL;
 738                                 write_protected = 1;
 739                         }
 740                         spte = rmap_next(kvm, rmapp, spte);
 741                 }
 742         }
 743
 744         return write_protected;
 745 }
 746
 747 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 748                            unsigned long data)
 749 {
 750         u64 *spte;
 751         int need_tlb_flush = 0;
 752
 753         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 754                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 755                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 756                 rmap_remove(kvm, spte);
 757                 __set_spte(spte, shadow_trap_nonpresent_pte);
 758                 need_tlb_flush = 1;
 759         }
 760         return need_tlb_flush;
 761 }
 762
 763 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
 764                              unsigned long data)
 765 {
 766         int need_flush = 0;
 767         u64 *spte, new_spte;
 768         pte_t *ptep = (pte_t *)data;
 769         pfn_t new_pfn;
 770
 771         WARN_ON(pte_huge(*ptep));
 772         new_pfn = pte_pfn(*ptep);
 773         spte = rmap_next(kvm, rmapp, NULL);
 774         while (spte) {
 775                 BUG_ON(!is_shadow_present_pte(*spte));
 776                 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
 777                 need_flush = 1;
 778                 if (pte_write(*ptep)) {
 779                         rmap_remove(kvm, spte);
 780                         __set_spte(spte, shadow_trap_nonpresent_pte);
 781                         spte = rmap_next(kvm, rmapp, NULL);
 782                 } else {
 783                         new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
 784                         new_spte |= (u64)new_pfn << PAGE_SHIFT;
 785
 786                         new_spte &= ~PT_WRITABLE_MASK;
 787                         new_spte &= ~SPTE_HOST_WRITEABLE;
 788                         if (is_writeble_pte(*spte))
 789                                 kvm_set_pfn_dirty(spte_to_pfn(*spte));
 790                         __set_spte(spte, new_spte);
 791                         spte = rmap_next(kvm, rmapp, spte);
 792                 }
 793         }
 794         if (need_flush)
 795                 kvm_flush_remote_tlbs(kvm);
 796
 797         return 0;
 798 }
 799
 800 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 801                           unsigned long data,
 802                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 803                                          unsigned long data))
 804 {
 805         int i, j;
 806         int retval = 0;
 807         struct kvm_memslots *slots;
 808
 809         slots = rcu_dereference(kvm->memslots);
 810
 811         for (i = 0; i < slots->nmemslots; i++) {
 812                 struct kvm_memory_slot *memslot = &slots->memslots[i];
 813                 unsigned long start = memslot->userspace_addr;
 814                 unsigned long end;
 815
 816                 end = start + (memslot->npages << PAGE_SHIFT);
 817                 if (hva >= start && hva < end) {
 818                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 819
 820                         retval |= handler(kvm, &memslot->rmap[gfn_offset],
 821                                           data);
 822
 823                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
 824                                 int idx = gfn_offset;
 825                                 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
 826                                 retval |= handler(kvm,
 827                                         &memslot->lpage_info[j][idx].rmap_pde,
 828                                         data);
 829                         }
 830                 }
 831         }
 832
 833         return retval;
 834 }
 835
 836 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 837 {
 838         return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
 839 }
 840
 841 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
 842 {
 843         kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
 844 }
 845
 846 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 847                          unsigned long data)
 848 {
 849         u64 *spte;
 850         int young = 0;
 851
 852         /* always return old for EPT */
 853         if (!shadow_accessed_mask)
 854                 return 0;
 855
 856         spte = rmap_next(kvm, rmapp, NULL);
 857         while (spte) {
 858                 int _young;
 859                 u64 _spte = *spte;
 860                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 861                 _young = _spte & PT_ACCESSED_MASK;
 862                 if (_young) {
 863                         young = 1;
 864                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 865                 }
 866                 spte = rmap_next(kvm, rmapp, spte);
 867         }
 868         return young;
 869 }
 870
 871 #define RMAP_RECYCLE_THRESHOLD 1000
 872
 873 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 874 {
 875         unsigned long *rmapp;
 876         struct kvm_mmu_page *sp;
 877
 878         sp = page_header(__pa(spte));
 879
 880         gfn = unalias_gfn(vcpu->kvm, gfn);
 881         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 882
 883         kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
 884         kvm_flush_remote_tlbs(vcpu->kvm);
 885 }
 886
 887 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 888 {
 889         return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
 890 }
 891
 892 #ifdef MMU_DEBUG
 893 static int is_empty_shadow_page(u64 *spt)
 894 {
 895         u64 *pos;
 896         u64 *end;
 897
 898         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 899                 if (is_shadow_present_pte(*pos)) {
 900                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 901                                pos, *pos);
 902                         return 0;
 903                 }
 904         return 1;
 905 }
 906 #endif
 907
 908 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 909 {
 910         ASSERT(is_empty_shadow_page(sp->spt));
 911         list_del(&sp->link);
 912         __free_page(virt_to_page(sp->spt));
 913         __free_page(virt_to_page(sp->gfns));
 914         kfree(sp);
 915         ++kvm->arch.n_free_mmu_pages;
 916 }
 917
 918 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 919 {
 920         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 921 }
 922
 923 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 924                                                u64 *parent_pte)
 925 {
 926         struct kvm_mmu_page *sp;
 927
 928         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 929         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 930         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 931         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 932         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 933         INIT_LIST_HEAD(&sp->oos_link);
 934         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 935         sp->multimapped = 0;
 936         sp->parent_pte = parent_pte;
 937         --vcpu->kvm->arch.n_free_mmu_pages;
 938         return sp;
 939 }
 940
 941 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 942                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 943 {
 944         struct kvm_pte_chain *pte_chain;
 945         struct hlist_node *node;
 946         int i;
 947
 948         if (!parent_pte)
 949                 return;
 950         if (!sp->multimapped) {
 951                 u64 *old = sp->parent_pte;
 952
 953                 if (!old) {
 954                         sp->parent_pte = parent_pte;
 955                         return;
 956                 }
 957                 sp->multimapped = 1;
 958                 pte_chain = mmu_alloc_pte_chain(vcpu);
 959                 INIT_HLIST_HEAD(&sp->parent_ptes);
 960                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 961                 pte_chain->parent_ptes[0] = old;
 962         }
 963         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 964                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 965                         continue;
 966                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 967                         if (!pte_chain->parent_ptes[i]) {
 968                                 pte_chain->parent_ptes[i] = parent_pte;
 969                                 return;
 970                         }
 971         }
 972         pte_chain = mmu_alloc_pte_chain(vcpu);
 973         BUG_ON(!pte_chain);
 974         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 975         pte_chain->parent_ptes[0] = parent_pte;
 976 }
 977
 978 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 979                                        u64 *parent_pte)
 980 {
 981         struct kvm_pte_chain *pte_chain;
 982         struct hlist_node *node;
 983         int i;
 984
 985         if (!sp->multimapped) {
 986                 BUG_ON(sp->parent_pte != parent_pte);
 987                 sp->parent_pte = NULL;
 988                 return;
 989         }
 990         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 991                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 992                         if (!pte_chain->parent_ptes[i])
 993                                 break;
 994                         if (pte_chain->parent_ptes[i] != parent_pte)
 995                                 continue;
 996                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 997                                 && pte_chain->parent_ptes[i + 1]) {
 998                                 pte_chain->parent_ptes[i]
 999                                         = pte_chain->parent_ptes[i + 1];
1000                                 ++i;
1001                         }
1002                         pte_chain->parent_ptes[i] = NULL;
1003                         if (i == 0) {
1004                                 hlist_del(&pte_chain->link);
1005                                 mmu_free_pte_chain(pte_chain);
1006                                 if (hlist_empty(&sp->parent_ptes)) {
1007                                         sp->multimapped = 0;
1008                                         sp->parent_pte = NULL;
1009                                 }
1010                         }
1011                         return;
1012                 }
1013         BUG();
1014 }
1015
1016
1017 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1018                             mmu_parent_walk_fn fn)
1019 {
1020         struct kvm_pte_chain *pte_chain;
1021         struct hlist_node *node;
1022         struct kvm_mmu_page *parent_sp;
1023         int i;
1024
1025         if (!sp->multimapped && sp->parent_pte) {
1026                 parent_sp = page_header(__pa(sp->parent_pte));
1027                 fn(vcpu, parent_sp);
1028                 mmu_parent_walk(vcpu, parent_sp, fn);
1029                 return;
1030         }
1031         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1032                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1033                         if (!pte_chain->parent_ptes[i])
1034                                 break;
1035                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
1036                         fn(vcpu, parent_sp);
1037                         mmu_parent_walk(vcpu, parent_sp, fn);
1038                 }
1039 }
1040
1041 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1042 {
1043         unsigned int index;
1044         struct kvm_mmu_page *sp = page_header(__pa(spte));
1045
1046         index = spte - sp->spt;
1047         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1048                 sp->unsync_children++;
1049         WARN_ON(!sp->unsync_children);
1050 }
1051
1052 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1053 {
1054         struct kvm_pte_chain *pte_chain;
1055         struct hlist_node *node;
1056         int i;
1057
1058         if (!sp->parent_pte)
1059                 return;
1060
1061         if (!sp->multimapped) {
1062                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1063                 return;
1064         }
1065
1066         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1067                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1068                         if (!pte_chain->parent_ptes[i])
1069                                 break;
1070                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1071                 }
1072 }
1073
1074 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1075 {
1076         kvm_mmu_update_parents_unsync(sp);
1077         return 1;
1078 }
1079
1080 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
1081                                         struct kvm_mmu_page *sp)
1082 {
1083         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
1084         kvm_mmu_update_parents_unsync(sp);
1085 }
1086
1087 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1088                                     struct kvm_mmu_page *sp)
1089 {
1090         int i;
1091
1092         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1093                 sp->spt[i] = shadow_trap_nonpresent_pte;
1094 }
1095
1096 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1097                                struct kvm_mmu_page *sp)
1098 {
1099         return 1;
1100 }
1101
1102 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1103 {
1104 }
1105
1106 #define KVM_PAGE_ARRAY_NR 16
1107
1108 struct kvm_mmu_pages {
1109         struct mmu_page_and_offset {
1110                 struct kvm_mmu_page *sp;
1111                 unsigned int idx;
1112         } page[KVM_PAGE_ARRAY_NR];
1113         unsigned int nr;
1114 };
1115
1116 #define for_each_unsync_children(bitmap, idx)           \
1117         for (idx = find_first_bit(bitmap, 512);         \
1118              idx < 512;                                 \
1119              idx = find_next_bit(bitmap, 512, idx+1))
1120
1121 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1122                          int idx)
1123 {
1124         int i;
1125
1126         if (sp->unsync)
1127                 for (i=0; i < pvec->nr; i++)
1128                         if (pvec->page[i].sp == sp)
1129                                 return 0;
1130
1131         pvec->page[pvec->nr].sp = sp;
1132         pvec->page[pvec->nr].idx = idx;
1133         pvec->nr++;
1134         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1135 }
1136
1137 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1138                            struct kvm_mmu_pages *pvec)
1139 {
1140         int i, ret, nr_unsync_leaf = 0;
1141
1142         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1143                 u64 ent = sp->spt[i];
1144
1145                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1146                         struct kvm_mmu_page *child;
1147                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1148
1149                         if (child->unsync_children) {
1150                                 if (mmu_pages_add(pvec, child, i))
1151                                         return -ENOSPC;
1152
1153                                 ret = __mmu_unsync_walk(child, pvec);
1154                                 if (!ret)
1155                                         __clear_bit(i, sp->unsync_child_bitmap);
1156                                 else if (ret > 0)
1157                                         nr_unsync_leaf += ret;
1158                                 else
1159                                         return ret;
1160                         }
1161
1162                         if (child->unsync) {
1163                                 nr_unsync_leaf++;
1164                                 if (mmu_pages_add(pvec, child, i))
1165                                         return -ENOSPC;
1166                         }
1167                 }
1168         }
1169
1170         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1171                 sp->unsync_children = 0;
1172
1173         return nr_unsync_leaf;
1174 }
1175
1176 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1177                            struct kvm_mmu_pages *pvec)
1178 {
1179         if (!sp->unsync_children)
1180                 return 0;
1181
1182         mmu_pages_add(pvec, sp, 0);
1183         return __mmu_unsync_walk(sp, pvec);
1184 }
1185
1186 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1187 {
1188         unsigned index;
1189         struct hlist_head *bucket;
1190         struct kvm_mmu_page *sp;
1191         struct hlist_node *node;
1192
1193         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1194         index = kvm_page_table_hashfn(gfn);
1195         bucket = &kvm->arch.mmu_page_hash[index];
1196         hlist_for_each_entry(sp, node, bucket, hash_link)
1197                 if (sp->gfn == gfn && !sp->role.direct
1198                     && !sp->role.invalid) {
1199                         pgprintk("%s: found role %x\n",
1200                                  __func__, sp->role.word);
1201                         return sp;
1202                 }
1203         return NULL;
1204 }
1205
1206 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1207 {
1208         WARN_ON(!sp->unsync);
1209         sp->unsync = 0;
1210         --kvm->stat.mmu_unsync;
1211 }
1212
1213 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1214
1215 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1216 {
1217         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1218                 kvm_mmu_zap_page(vcpu->kvm, sp);
1219                 return 1;
1220         }
1221
1222         trace_kvm_mmu_sync_page(sp);
1223         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1224                 kvm_flush_remote_tlbs(vcpu->kvm);
1225         kvm_unlink_unsync_page(vcpu->kvm, sp);
1226         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1227                 kvm_mmu_zap_page(vcpu->kvm, sp);
1228                 return 1;
1229         }
1230
1231         kvm_mmu_flush_tlb(vcpu);
1232         return 0;
1233 }
1234
1235 struct mmu_page_path {
1236         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1237         unsigned int idx[PT64_ROOT_LEVEL-1];
1238 };
1239
1240 #define for_each_sp(pvec, sp, parents, i)                       \
1241                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1242                         sp = pvec.page[i].sp;                   \
1243                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1244                         i = mmu_pages_next(&pvec, &parents, i))
1245
1246 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1247                           struct mmu_page_path *parents,
1248                           int i)
1249 {
1250         int n;
1251
1252         for (n = i+1; n < pvec->nr; n++) {
1253                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1254
1255                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1256                         parents->idx[0] = pvec->page[n].idx;
1257                         return n;
1258                 }
1259
1260                 parents->parent[sp->role.level-2] = sp;
1261                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1262         }
1263
1264         return n;
1265 }
1266
1267 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1268 {
1269         struct kvm_mmu_page *sp;
1270         unsigned int level = 0;
1271
1272         do {
1273                 unsigned int idx = parents->idx[level];
1274
1275                 sp = parents->parent[level];
1276                 if (!sp)
1277                         return;
1278
1279                 --sp->unsync_children;
1280                 WARN_ON((int)sp->unsync_children < 0);
1281                 __clear_bit(idx, sp->unsync_child_bitmap);
1282                 level++;
1283         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1284 }
1285
1286 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1287                                struct mmu_page_path *parents,
1288                                struct kvm_mmu_pages *pvec)
1289 {
1290         parents->parent[parent->role.level-1] = NULL;
1291         pvec->nr = 0;
1292 }
1293
1294 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1295                               struct kvm_mmu_page *parent)
1296 {
1297         int i;
1298         struct kvm_mmu_page *sp;
1299         struct mmu_page_path parents;
1300         struct kvm_mmu_pages pages;
1301
1302         kvm_mmu_pages_init(parent, &parents, &pages);
1303         while (mmu_unsync_walk(parent, &pages)) {
1304                 int protected = 0;
1305
1306                 for_each_sp(pages, sp, parents, i)
1307                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1308
1309                 if (protected)
1310                         kvm_flush_remote_tlbs(vcpu->kvm);
1311
1312                 for_each_sp(pages, sp, parents, i) {
1313                         kvm_sync_page(vcpu, sp);
1314                         mmu_pages_clear_parents(&parents);
1315                 }
1316                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1317                 kvm_mmu_pages_init(parent, &parents, &pages);
1318         }
1319 }
1320
1321 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1322                                              gfn_t gfn,
1323                                              gva_t gaddr,
1324                                              unsigned level,
1325                                              int direct,
1326                                              unsigned access,
1327                                              u64 *parent_pte)
1328 {
1329         union kvm_mmu_page_role role;
1330         unsigned index;
1331         unsigned quadrant;
1332         struct hlist_head *bucket;
1333         struct kvm_mmu_page *sp;
1334         struct hlist_node *node, *tmp;
1335
1336         role = vcpu->arch.mmu.base_role;
1337         role.level = level;
1338         role.direct = direct;
1339         role.access = access;
1340         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1341                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1342                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1343                 role.quadrant = quadrant;
1344         }
1345         index = kvm_page_table_hashfn(gfn);
1346         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1347         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1348                 if (sp->gfn == gfn) {
1349                         if (sp->unsync)
1350                                 if (kvm_sync_page(vcpu, sp))
1351                                         continue;
1352
1353                         if (sp->role.word != role.word)
1354                                 continue;
1355
1356                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1357                         if (sp->unsync_children) {
1358                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1359                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1360                         }
1361                         trace_kvm_mmu_get_page(sp, false);
1362                         return sp;
1363                 }
1364         ++vcpu->kvm->stat.mmu_cache_miss;
1365         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1366         if (!sp)
1367                 return sp;
1368         sp->gfn = gfn;
1369         sp->role = role;
1370         hlist_add_head(&sp->hash_link, bucket);
1371         if (!direct) {
1372                 if (rmap_write_protect(vcpu->kvm, gfn))
1373                         kvm_flush_remote_tlbs(vcpu->kvm);
1374                 account_shadowed(vcpu->kvm, gfn);
1375         }
1376         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1377                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1378         else
1379                 nonpaging_prefetch_page(vcpu, sp);
1380         trace_kvm_mmu_get_page(sp, true);
1381         return sp;
1382 }
1383
1384 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1385                              struct kvm_vcpu *vcpu, u64 addr)
1386 {
1387         iterator->addr = addr;
1388         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1389         iterator->level = vcpu->arch.mmu.shadow_root_level;
1390         if (iterator->level == PT32E_ROOT_LEVEL) {
1391                 iterator->shadow_addr
1392                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1393                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1394                 --iterator->level;
1395                 if (!iterator->shadow_addr)
1396                         iterator->level = 0;
1397         }
1398 }
1399
1400 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1401 {
1402         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1403                 return false;
1404
1405         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1406                 if (is_large_pte(*iterator->sptep))
1407                         return false;
1408
1409         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1410         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1411         return true;
1412 }
1413
1414 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1415 {
1416         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1417         --iterator->level;
1418 }
1419
1420 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1421                                          struct kvm_mmu_page *sp)
1422 {
1423         unsigned i;
1424         u64 *pt;
1425         u64 ent;
1426
1427         pt = sp->spt;
1428
1429         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1430                 ent = pt[i];
1431
1432                 if (is_shadow_present_pte(ent)) {
1433                         if (!is_last_spte(ent, sp->role.level)) {
1434                                 ent &= PT64_BASE_ADDR_MASK;
1435                                 mmu_page_remove_parent_pte(page_header(ent),
1436                                                            &pt[i]);
1437                         } else {
1438                                 if (is_large_pte(ent))
1439                                         --kvm->stat.lpages;
1440                                 rmap_remove(kvm, &pt[i]);
1441                         }
1442                 }
1443                 pt[i] = shadow_trap_nonpresent_pte;
1444         }
1445 }
1446
1447 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1448 {
1449         mmu_page_remove_parent_pte(sp, parent_pte);
1450 }
1451
1452 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1453 {
1454         int i;
1455         struct kvm_vcpu *vcpu;
1456
1457         kvm_for_each_vcpu(i, vcpu, kvm)
1458                 vcpu->arch.last_pte_updated = NULL;
1459 }
1460
1461 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1462 {
1463         u64 *parent_pte;
1464
1465         while (sp->multimapped || sp->parent_pte) {
1466                 if (!sp->multimapped)
1467                         parent_pte = sp->parent_pte;
1468                 else {
1469                         struct kvm_pte_chain *chain;
1470
1471                         chain = container_of(sp->parent_ptes.first,
1472                                              struct kvm_pte_chain, link);
1473                         parent_pte = chain->parent_ptes[0];
1474                 }
1475                 BUG_ON(!parent_pte);
1476                 kvm_mmu_put_page(sp, parent_pte);
1477                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1478         }
1479 }
1480
1481 static int mmu_zap_unsync_children(struct kvm *kvm,
1482                                    struct kvm_mmu_page *parent)
1483 {
1484         int i, zapped = 0;
1485         struct mmu_page_path parents;
1486         struct kvm_mmu_pages pages;
1487
1488         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1489                 return 0;
1490
1491         kvm_mmu_pages_init(parent, &parents, &pages);
1492         while (mmu_unsync_walk(parent, &pages)) {
1493                 struct kvm_mmu_page *sp;
1494
1495                 for_each_sp(pages, sp, parents, i) {
1496                         kvm_mmu_zap_page(kvm, sp);
1497                         mmu_pages_clear_parents(&parents);
1498                 }
1499                 zapped += pages.nr;
1500                 kvm_mmu_pages_init(parent, &parents, &pages);
1501         }
1502
1503         return zapped;
1504 }
1505
1506 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1507 {
1508         int ret;
1509
1510         trace_kvm_mmu_zap_page(sp);
1511         ++kvm->stat.mmu_shadow_zapped;
1512         ret = mmu_zap_unsync_children(kvm, sp);
1513         kvm_mmu_page_unlink_children(kvm, sp);
1514         kvm_mmu_unlink_parents(kvm, sp);
1515         kvm_flush_remote_tlbs(kvm);
1516         if (!sp->role.invalid && !sp->role.direct)
1517                 unaccount_shadowed(kvm, sp->gfn);
1518         if (sp->unsync)
1519                 kvm_unlink_unsync_page(kvm, sp);
1520         if (!sp->root_count) {
1521                 hlist_del(&sp->hash_link);
1522                 kvm_mmu_free_page(kvm, sp);
1523         } else {
1524                 sp->role.invalid = 1;
1525                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1526                 kvm_reload_remote_mmus(kvm);
1527         }
1528         kvm_mmu_reset_last_pte_updated(kvm);
1529         return ret;
1530 }
1531
1532 /*
1533  * Changing the number of mmu pages allocated to the vm
1534  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1535  */
1536 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1537 {
1538         int used_pages;
1539
1540         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1541         used_pages = max(0, used_pages);
1542
1543         /*
1544          * If we set the number of mmu pages to be smaller be than the
1545          * number of actived pages , we must to free some mmu pages before we
1546          * change the value
1547          */
1548
1549         if (used_pages > kvm_nr_mmu_pages) {
1550                 while (used_pages > kvm_nr_mmu_pages) {
1551                         struct kvm_mmu_page *page;
1552
1553                         page = container_of(kvm->arch.active_mmu_pages.prev,
1554                                             struct kvm_mmu_page, link);
1555                         kvm_mmu_zap_page(kvm, page);
1556                         used_pages--;
1557                 }
1558                 kvm->arch.n_free_mmu_pages = 0;
1559         }
1560         else
1561                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1562                                          - kvm->arch.n_alloc_mmu_pages;
1563
1564         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1565 }
1566
1567 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1568 {
1569         unsigned index;
1570         struct hlist_head *bucket;
1571         struct kvm_mmu_page *sp;
1572         struct hlist_node *node, *n;
1573         int r;
1574
1575         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1576         r = 0;
1577         index = kvm_page_table_hashfn(gfn);
1578         bucket = &kvm->arch.mmu_page_hash[index];
1579         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1580                 if (sp->gfn == gfn && !sp->role.direct) {
1581                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1582                                  sp->role.word);
1583                         r = 1;
1584                         if (kvm_mmu_zap_page(kvm, sp))
1585                                 n = bucket->first;
1586                 }
1587         return r;
1588 }
1589
1590 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1591 {
1592         unsigned index;
1593         struct hlist_head *bucket;
1594         struct kvm_mmu_page *sp;
1595         struct hlist_node *node, *nn;
1596
1597         index = kvm_page_table_hashfn(gfn);
1598         bucket = &kvm->arch.mmu_page_hash[index];
1599         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1600                 if (sp->gfn == gfn && !sp->role.direct
1601                     && !sp->role.invalid) {
1602                         pgprintk("%s: zap %lx %x\n",
1603                                  __func__, gfn, sp->role.word);
1604                         kvm_mmu_zap_page(kvm, sp);
1605                 }
1606         }
1607 }
1608
1609 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1610 {
1611         int slot = memslot_id(kvm, gfn);
1612         struct kvm_mmu_page *sp = page_header(__pa(pte));
1613
1614         __set_bit(slot, sp->slot_bitmap);
1615 }
1616
1617 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1618 {
1619         int i;
1620         u64 *pt = sp->spt;
1621
1622         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1623                 return;
1624
1625         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1626                 if (pt[i] == shadow_notrap_nonpresent_pte)
1627                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1628         }
1629 }
1630
1631 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1632 {
1633         struct page *page;
1634
1635         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1636
1637         if (gpa == UNMAPPED_GVA)
1638                 return NULL;
1639
1640         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1641
1642         return page;
1643 }
1644
1645 /*
1646  * The function is based on mtrr_type_lookup() in
1647  * arch/x86/kernel/cpu/mtrr/generic.c
1648  */
1649 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1650                          u64 start, u64 end)
1651 {
1652         int i;
1653         u64 base, mask;
1654         u8 prev_match, curr_match;
1655         int num_var_ranges = KVM_NR_VAR_MTRR;
1656
1657         if (!mtrr_state->enabled)
1658                 return 0xFF;
1659
1660         /* Make end inclusive end, instead of exclusive */
1661         end--;
1662
1663         /* Look in fixed ranges. Just return the type as per start */
1664         if (mtrr_state->have_fixed && (start < 0x100000)) {
1665                 int idx;
1666
1667                 if (start < 0x80000) {
1668                         idx = 0;
1669                         idx += (start >> 16);
1670                         return mtrr_state->fixed_ranges[idx];
1671                 } else if (start < 0xC0000) {
1672                         idx = 1 * 8;
1673                         idx += ((start - 0x80000) >> 14);
1674                         return mtrr_state->fixed_ranges[idx];
1675                 } else if (start < 0x1000000) {
1676                         idx = 3 * 8;
1677                         idx += ((start - 0xC0000) >> 12);
1678                         return mtrr_state->fixed_ranges[idx];
1679                 }
1680         }
1681
1682         /*
1683          * Look in variable ranges
1684          * Look of multiple ranges matching this address and pick type
1685          * as per MTRR precedence
1686          */
1687         if (!(mtrr_state->enabled & 2))
1688                 return mtrr_state->def_type;
1689
1690         prev_match = 0xFF;
1691         for (i = 0; i < num_var_ranges; ++i) {
1692                 unsigned short start_state, end_state;
1693
1694                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1695                         continue;
1696
1697                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1698                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1699                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1700                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1701
1702                 start_state = ((start & mask) == (base & mask));
1703                 end_state = ((end & mask) == (base & mask));
1704                 if (start_state != end_state)
1705                         return 0xFE;
1706
1707                 if ((start & mask) != (base & mask))
1708                         continue;
1709
1710                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1711                 if (prev_match == 0xFF) {
1712                         prev_match = curr_match;
1713                         continue;
1714                 }
1715
1716                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1717                     curr_match == MTRR_TYPE_UNCACHABLE)
1718                         return MTRR_TYPE_UNCACHABLE;
1719
1720                 if ((prev_match == MTRR_TYPE_WRBACK &&
1721                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1722                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1723                      curr_match == MTRR_TYPE_WRBACK)) {
1724                         prev_match = MTRR_TYPE_WRTHROUGH;
1725                         curr_match = MTRR_TYPE_WRTHROUGH;
1726                 }
1727
1728                 if (prev_match != curr_match)
1729                         return MTRR_TYPE_UNCACHABLE;
1730         }
1731
1732         if (prev_match != 0xFF)
1733                 return prev_match;
1734
1735         return mtrr_state->def_type;
1736 }
1737
1738 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1739 {
1740         u8 mtrr;
1741
1742         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1743                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1744         if (mtrr == 0xfe || mtrr == 0xff)
1745                 mtrr = MTRR_TYPE_WRBACK;
1746         return mtrr;
1747 }
1748 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1749
1750 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1751 {
1752         unsigned index;
1753         struct hlist_head *bucket;
1754         struct kvm_mmu_page *s;
1755         struct hlist_node *node, *n;
1756
1757         trace_kvm_mmu_unsync_page(sp);
1758         index = kvm_page_table_hashfn(sp->gfn);
1759         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1760         /* don't unsync if pagetable is shadowed with multiple roles */
1761         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1762                 if (s->gfn != sp->gfn || s->role.direct)
1763                         continue;
1764                 if (s->role.word != sp->role.word)
1765                         return 1;
1766         }
1767         ++vcpu->kvm->stat.mmu_unsync;
1768         sp->unsync = 1;
1769
1770         kvm_mmu_mark_parents_unsync(vcpu, sp);
1771
1772         mmu_convert_notrap(sp);
1773         return 0;
1774 }
1775
1776 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1777                                   bool can_unsync)
1778 {
1779         struct kvm_mmu_page *shadow;
1780
1781         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1782         if (shadow) {
1783                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1784                         return 1;
1785                 if (shadow->unsync)
1786                         return 0;
1787                 if (can_unsync && oos_shadow)
1788                         return kvm_unsync_page(vcpu, shadow);
1789                 return 1;
1790         }
1791         return 0;
1792 }
1793
1794 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1795                     unsigned pte_access, int user_fault,
1796                     int write_fault, int dirty, int level,
1797                     gfn_t gfn, pfn_t pfn, bool speculative,
1798                     bool can_unsync, bool reset_host_protection)
1799 {
1800         u64 spte;
1801         int ret = 0;
1802
1803         /*
1804          * We don't set the accessed bit, since we sometimes want to see
1805          * whether the guest actually used the pte (in order to detect
1806          * demand paging).
1807          */
1808         spte = shadow_base_present_pte | shadow_dirty_mask;
1809         if (!speculative)
1810                 spte |= shadow_accessed_mask;
1811         if (!dirty)
1812                 pte_access &= ~ACC_WRITE_MASK;
1813         if (pte_access & ACC_EXEC_MASK)
1814                 spte |= shadow_x_mask;
1815         else
1816                 spte |= shadow_nx_mask;
1817         if (pte_access & ACC_USER_MASK)
1818                 spte |= shadow_user_mask;
1819         if (level > PT_PAGE_TABLE_LEVEL)
1820                 spte |= PT_PAGE_SIZE_MASK;
1821         if (tdp_enabled)
1822                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1823                         kvm_is_mmio_pfn(pfn));
1824
1825         if (reset_host_protection)
1826                 spte |= SPTE_HOST_WRITEABLE;
1827
1828         spte |= (u64)pfn << PAGE_SHIFT;
1829
1830         if ((pte_access & ACC_WRITE_MASK)
1831             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1832
1833                 if (level > PT_PAGE_TABLE_LEVEL &&
1834                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1835                         ret = 1;
1836                         spte = shadow_trap_nonpresent_pte;
1837                         goto set_pte;
1838                 }
1839
1840                 spte |= PT_WRITABLE_MASK;
1841
1842                 /*
1843                  * Optimization: for pte sync, if spte was writable the hash
1844                  * lookup is unnecessary (and expensive). Write protection
1845                  * is responsibility of mmu_get_page / kvm_sync_page.
1846                  * Same reasoning can be applied to dirty page accounting.
1847                  */
1848                 if (!can_unsync && is_writeble_pte(*sptep))
1849                         goto set_pte;
1850
1851                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1852                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1853                                  __func__, gfn);
1854                         ret = 1;
1855                         pte_access &= ~ACC_WRITE_MASK;
1856                         if (is_writeble_pte(spte))
1857                                 spte &= ~PT_WRITABLE_MASK;
1858                 }
1859         }
1860
1861         if (pte_access & ACC_WRITE_MASK)
1862                 mark_page_dirty(vcpu->kvm, gfn);
1863
1864 set_pte:
1865         __set_spte(sptep, spte);
1866         return ret;
1867 }
1868
1869 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1870                          unsigned pt_access, unsigned pte_access,
1871                          int user_fault, int write_fault, int dirty,
1872                          int *ptwrite, int level, gfn_t gfn,
1873                          pfn_t pfn, bool speculative,
1874                          bool reset_host_protection)
1875 {
1876         int was_rmapped = 0;
1877         int was_writeble = is_writeble_pte(*sptep);
1878         int rmap_count;
1879
1880         pgprintk("%s: spte %llx access %x write_fault %d"
1881                  " user_fault %d gfn %lx\n",
1882                  __func__, *sptep, pt_access,
1883                  write_fault, user_fault, gfn);
1884
1885         if (is_rmap_spte(*sptep)) {
1886                 /*
1887                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1888                  * the parent of the now unreachable PTE.
1889                  */
1890                 if (level > PT_PAGE_TABLE_LEVEL &&
1891                     !is_large_pte(*sptep)) {
1892                         struct kvm_mmu_page *child;
1893                         u64 pte = *sptep;
1894
1895                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1896                         mmu_page_remove_parent_pte(child, sptep);
1897                 } else if (pfn != spte_to_pfn(*sptep)) {
1898                         pgprintk("hfn old %lx new %lx\n",
1899                                  spte_to_pfn(*sptep), pfn);
1900                         rmap_remove(vcpu->kvm, sptep);
1901                 } else
1902                         was_rmapped = 1;
1903         }
1904
1905         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1906                       dirty, level, gfn, pfn, speculative, true,
1907                       reset_host_protection)) {
1908                 if (write_fault)
1909                         *ptwrite = 1;
1910                 kvm_x86_ops->tlb_flush(vcpu);
1911         }
1912
1913         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1914         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1915                  is_large_pte(*sptep)? "2MB" : "4kB",
1916                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1917                  *sptep, sptep);
1918         if (!was_rmapped && is_large_pte(*sptep))
1919                 ++vcpu->kvm->stat.lpages;
1920
1921         page_header_update_slot(vcpu->kvm, sptep, gfn);
1922         if (!was_rmapped) {
1923                 rmap_count = rmap_add(vcpu, sptep, gfn);
1924                 kvm_release_pfn_clean(pfn);
1925                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1926                         rmap_recycle(vcpu, sptep, gfn);
1927         } else {
1928                 if (was_writeble)
1929                         kvm_release_pfn_dirty(pfn);
1930                 else
1931                         kvm_release_pfn_clean(pfn);
1932         }
1933         if (speculative) {
1934                 vcpu->arch.last_pte_updated = sptep;
1935                 vcpu->arch.last_pte_gfn = gfn;
1936         }
1937 }
1938
1939 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1940 {
1941 }
1942
1943 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1944                         int level, gfn_t gfn, pfn_t pfn)
1945 {
1946         struct kvm_shadow_walk_iterator iterator;
1947         struct kvm_mmu_page *sp;
1948         int pt_write = 0;
1949         gfn_t pseudo_gfn;
1950
1951         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1952                 if (iterator.level == level) {
1953                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1954                                      0, write, 1, &pt_write,
1955                                      level, gfn, pfn, false, true);
1956                         ++vcpu->stat.pf_fixed;
1957                         break;
1958                 }
1959
1960                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1961                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1962                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1963                                               iterator.level - 1,
1964                                               1, ACC_ALL, iterator.sptep);
1965                         if (!sp) {
1966                                 pgprintk("nonpaging_map: ENOMEM\n");
1967                                 kvm_release_pfn_clean(pfn);
1968                                 return -ENOMEM;
1969                         }
1970
1971                         __set_spte(iterator.sptep,
1972                                    __pa(sp->spt)
1973                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1974                                    | shadow_user_mask | shadow_x_mask);
1975                 }
1976         }
1977         return pt_write;
1978 }
1979
1980 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1981 {
1982         int r;
1983         int level;
1984         pfn_t pfn;
1985         unsigned long mmu_seq;
1986
1987         level = mapping_level(vcpu, gfn);
1988
1989         /*
1990          * This path builds a PAE pagetable - so we can map 2mb pages at
1991          * maximum. Therefore check if the level is larger than that.
1992          */
1993         if (level > PT_DIRECTORY_LEVEL)
1994                 level = PT_DIRECTORY_LEVEL;
1995
1996         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
1997
1998         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1999         smp_rmb();
2000         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2001
2002         /* mmio */
2003         if (is_error_pfn(pfn)) {
2004                 kvm_release_pfn_clean(pfn);
2005                 return 1;
2006         }
2007
2008         spin_lock(&vcpu->kvm->mmu_lock);
2009         if (mmu_notifier_retry(vcpu, mmu_seq))
2010                 goto out_unlock;
2011         kvm_mmu_free_some_pages(vcpu);
2012         r = __direct_map(vcpu, v, write, level, gfn, pfn);
2013         spin_unlock(&vcpu->kvm->mmu_lock);
2014
2015
2016         return r;
2017
2018 out_unlock:
2019         spin_unlock(&vcpu->kvm->mmu_lock);
2020         kvm_release_pfn_clean(pfn);
2021         return 0;
2022 }
2023
2024
2025 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2026 {
2027         int i;
2028         struct kvm_mmu_page *sp;
2029
2030         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2031                 return;
2032         spin_lock(&vcpu->kvm->mmu_lock);
2033         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2034                 hpa_t root = vcpu->arch.mmu.root_hpa;
2035
2036                 sp = page_header(root);
2037                 --sp->root_count;
2038                 if (!sp->root_count && sp->role.invalid)
2039                         kvm_mmu_zap_page(vcpu->kvm, sp);
2040                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2041                 spin_unlock(&vcpu->kvm->mmu_lock);
2042                 return;
2043         }
2044         for (i = 0; i < 4; ++i) {
2045                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2046
2047                 if (root) {
2048                         root &= PT64_BASE_ADDR_MASK;
2049                         sp = page_header(root);
2050                         --sp->root_count;
2051                         if (!sp->root_count && sp->role.invalid)
2052                                 kvm_mmu_zap_page(vcpu->kvm, sp);
2053                 }
2054                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2055         }
2056         spin_unlock(&vcpu->kvm->mmu_lock);
2057         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2058 }
2059
2060 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2061 {
2062         int ret = 0;
2063
2064         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2065                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2066                 ret = 1;
2067         }
2068
2069         return ret;
2070 }
2071
2072 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2073 {
2074         int i;
2075         gfn_t root_gfn;
2076         struct kvm_mmu_page *sp;
2077         int direct = 0;
2078         u64 pdptr;
2079
2080         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2081
2082         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2083                 hpa_t root = vcpu->arch.mmu.root_hpa;
2084
2085                 ASSERT(!VALID_PAGE(root));
2086                 if (tdp_enabled)
2087                         direct = 1;
2088                 if (mmu_check_root(vcpu, root_gfn))
2089                         return 1;
2090                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2091                                       PT64_ROOT_LEVEL, direct,
2092                                       ACC_ALL, NULL);
2093                 root = __pa(sp->spt);
2094                 ++sp->root_count;
2095                 vcpu->arch.mmu.root_hpa = root;
2096                 return 0;
2097         }
2098         direct = !is_paging(vcpu);
2099         if (tdp_enabled)
2100                 direct = 1;
2101         for (i = 0; i < 4; ++i) {
2102                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2103
2104                 ASSERT(!VALID_PAGE(root));
2105                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2106                         pdptr = kvm_pdptr_read(vcpu, i);
2107                         if (!is_present_gpte(pdptr)) {
2108                                 vcpu->arch.mmu.pae_root[i] = 0;
2109                                 continue;
2110                         }
2111                         root_gfn = pdptr >> PAGE_SHIFT;
2112                 } else if (vcpu->arch.mmu.root_level == 0)
2113                         root_gfn = 0;
2114                 if (mmu_check_root(vcpu, root_gfn))
2115                         return 1;
2116                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2117                                       PT32_ROOT_LEVEL, direct,
2118                                       ACC_ALL, NULL);
2119                 root = __pa(sp->spt);
2120                 ++sp->root_count;
2121                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2122         }
2123         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2124         return 0;
2125 }
2126
2127 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2128 {
2129         int i;
2130         struct kvm_mmu_page *sp;
2131
2132         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2133                 return;
2134         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2135                 hpa_t root = vcpu->arch.mmu.root_hpa;
2136                 sp = page_header(root);
2137                 mmu_sync_children(vcpu, sp);
2138                 return;
2139         }
2140         for (i = 0; i < 4; ++i) {
2141                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2142
2143                 if (root && VALID_PAGE(root)) {
2144                         root &= PT64_BASE_ADDR_MASK;
2145                         sp = page_header(root);
2146                         mmu_sync_children(vcpu, sp);
2147                 }
2148         }
2149 }
2150
2151 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2152 {
2153         spin_lock(&vcpu->kvm->mmu_lock);
2154         mmu_sync_roots(vcpu);
2155         spin_unlock(&vcpu->kvm->mmu_lock);
2156 }
2157
2158 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2159 {
2160         return vaddr;
2161 }
2162
2163 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2164                                 u32 error_code)
2165 {
2166         gfn_t gfn;
2167         int r;
2168
2169         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2170         r = mmu_topup_memory_caches(vcpu);
2171         if (r)
2172                 return r;
2173
2174         ASSERT(vcpu);
2175         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2176
2177         gfn = gva >> PAGE_SHIFT;
2178
2179         return nonpaging_map(vcpu, gva & PAGE_MASK,
2180                              error_code & PFERR_WRITE_MASK, gfn);
2181 }
2182
2183 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2184                                 u32 error_code)
2185 {
2186         pfn_t pfn;
2187         int r;
2188         int level;
2189         gfn_t gfn = gpa >> PAGE_SHIFT;
2190         unsigned long mmu_seq;
2191
2192         ASSERT(vcpu);
2193         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2194
2195         r = mmu_topup_memory_caches(vcpu);
2196         if (r)
2197                 return r;
2198
2199         level = mapping_level(vcpu, gfn);
2200
2201         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2202
2203         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2204         smp_rmb();
2205         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2206         if (is_error_pfn(pfn)) {
2207                 kvm_release_pfn_clean(pfn);
2208                 return 1;
2209         }
2210         spin_lock(&vcpu->kvm->mmu_lock);
2211         if (mmu_notifier_retry(vcpu, mmu_seq))
2212                 goto out_unlock;
2213         kvm_mmu_free_some_pages(vcpu);
2214         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2215                          level, gfn, pfn);
2216         spin_unlock(&vcpu->kvm->mmu_lock);
2217
2218         return r;
2219
2220 out_unlock:
2221         spin_unlock(&vcpu->kvm->mmu_lock);
2222         kvm_release_pfn_clean(pfn);
2223         return 0;
2224 }
2225
2226 static void nonpaging_free(struct kvm_vcpu *vcpu)
2227 {
2228         mmu_free_roots(vcpu);
2229 }
2230
2231 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2232 {
2233         struct kvm_mmu *context = &vcpu->arch.mmu;
2234
2235         context->new_cr3 = nonpaging_new_cr3;
2236         context->page_fault = nonpaging_page_fault;
2237         context->gva_to_gpa = nonpaging_gva_to_gpa;
2238         context->free = nonpaging_free;
2239         context->prefetch_page = nonpaging_prefetch_page;
2240         context->sync_page = nonpaging_sync_page;
2241         context->invlpg = nonpaging_invlpg;
2242         context->root_level = 0;
2243         context->shadow_root_level = PT32E_ROOT_LEVEL;
2244         context->root_hpa = INVALID_PAGE;
2245         return 0;
2246 }
2247
2248 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2249 {
2250         ++vcpu->stat.tlb_flush;
2251         kvm_x86_ops->tlb_flush(vcpu);
2252 }
2253
2254 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2255 {
2256         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2257         mmu_free_roots(vcpu);
2258 }
2259
2260 static void inject_page_fault(struct kvm_vcpu *vcpu,
2261                               u64 addr,
2262                               u32 err_code)
2263 {
2264         kvm_inject_page_fault(vcpu, addr, err_code);
2265 }
2266
2267 static void paging_free(struct kvm_vcpu *vcpu)
2268 {
2269         nonpaging_free(vcpu);
2270 }
2271
2272 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2273 {
2274         int bit7;
2275
2276         bit7 = (gpte >> 7) & 1;
2277         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2278 }
2279
2280 #define PTTYPE 64
2281 #include "paging_tmpl.h"
2282 #undef PTTYPE
2283
2284 #define PTTYPE 32
2285 #include "paging_tmpl.h"
2286 #undef PTTYPE
2287
2288 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2289 {
2290         struct kvm_mmu *context = &vcpu->arch.mmu;
2291         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2292         u64 exb_bit_rsvd = 0;
2293
2294         if (!is_nx(vcpu))
2295                 exb_bit_rsvd = rsvd_bits(63, 63);
2296         switch (level) {
2297         case PT32_ROOT_LEVEL:
2298                 /* no rsvd bits for 2 level 4K page table entries */
2299                 context->rsvd_bits_mask[0][1] = 0;
2300                 context->rsvd_bits_mask[0][0] = 0;
2301                 if (is_cpuid_PSE36())
2302                         /* 36bits PSE 4MB page */
2303                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2304                 else
2305                         /* 32 bits PSE 4MB page */
2306                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2307                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2308                 break;
2309         case PT32E_ROOT_LEVEL:
2310                 context->rsvd_bits_mask[0][2] =
2311                         rsvd_bits(maxphyaddr, 63) |
2312                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2313                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2314                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2315                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2316                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2317                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2318                         rsvd_bits(maxphyaddr, 62) |
2319                         rsvd_bits(13, 20);              /* large page */
2320                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2321                 break;
2322         case PT64_ROOT_LEVEL:
2323                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2324                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2325                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2326                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2327                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2328                         rsvd_bits(maxphyaddr, 51);
2329                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2330                         rsvd_bits(maxphyaddr, 51);
2331                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2332                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2333                         rsvd_bits(maxphyaddr, 51) |
2334                         rsvd_bits(13, 29);
2335                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2336                         rsvd_bits(maxphyaddr, 51) |
2337                         rsvd_bits(13, 20);              /* large page */
2338                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2339                 break;
2340         }
2341 }
2342
2343 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2344 {
2345         struct kvm_mmu *context = &vcpu->arch.mmu;
2346
2347         ASSERT(is_pae(vcpu));
2348         context->new_cr3 = paging_new_cr3;
2349         context->page_fault = paging64_page_fault;
2350         context->gva_to_gpa = paging64_gva_to_gpa;
2351         context->prefetch_page = paging64_prefetch_page;
2352         context->sync_page = paging64_sync_page;
2353         context->invlpg = paging64_invlpg;
2354         context->free = paging_free;
2355         context->root_level = level;
2356         context->shadow_root_level = level;
2357         context->root_hpa = INVALID_PAGE;
2358         return 0;
2359 }
2360
2361 static int paging64_init_context(struct kvm_vcpu *vcpu)
2362 {
2363         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2364         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2365 }
2366
2367 static int paging32_init_context(struct kvm_vcpu *vcpu)
2368 {
2369         struct kvm_mmu *context = &vcpu->arch.mmu;
2370
2371         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2372         context->new_cr3 = paging_new_cr3;
2373         context->page_fault = paging32_page_fault;
2374         context->gva_to_gpa = paging32_gva_to_gpa;
2375         context->free = paging_free;
2376         context->prefetch_page = paging32_prefetch_page;
2377         context->sync_page = paging32_sync_page;
2378         context->invlpg = paging32_invlpg;
2379         context->root_level = PT32_ROOT_LEVEL;
2380         context->shadow_root_level = PT32E_ROOT_LEVEL;
2381         context->root_hpa = INVALID_PAGE;
2382         return 0;
2383 }
2384
2385 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2386 {
2387         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2388         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2389 }
2390
2391 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2392 {
2393         struct kvm_mmu *context = &vcpu->arch.mmu;
2394
2395         context->new_cr3 = nonpaging_new_cr3;
2396         context->page_fault = tdp_page_fault;
2397         context->free = nonpaging_free;
2398         context->prefetch_page = nonpaging_prefetch_page;
2399         context->sync_page = nonpaging_sync_page;
2400         context->invlpg = nonpaging_invlpg;
2401         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2402         context->root_hpa = INVALID_PAGE;
2403
2404         if (!is_paging(vcpu)) {
2405                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2406                 context->root_level = 0;
2407         } else if (is_long_mode(vcpu)) {
2408                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2409                 context->gva_to_gpa = paging64_gva_to_gpa;
2410                 context->root_level = PT64_ROOT_LEVEL;
2411         } else if (is_pae(vcpu)) {
2412                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2413                 context->gva_to_gpa = paging64_gva_to_gpa;
2414                 context->root_level = PT32E_ROOT_LEVEL;
2415         } else {
2416                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2417                 context->gva_to_gpa = paging32_gva_to_gpa;
2418                 context->root_level = PT32_ROOT_LEVEL;
2419         }
2420
2421         return 0;
2422 }
2423
2424 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2425 {
2426         int r;
2427
2428         ASSERT(vcpu);
2429         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2430
2431         if (!is_paging(vcpu))
2432                 r = nonpaging_init_context(vcpu);
2433         else if (is_long_mode(vcpu))
2434                 r = paging64_init_context(vcpu);
2435         else if (is_pae(vcpu))
2436                 r = paging32E_init_context(vcpu);
2437         else
2438                 r = paging32_init_context(vcpu);
2439
2440         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2441
2442         return r;
2443 }
2444
2445 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2446 {
2447         vcpu->arch.update_pte.pfn = bad_pfn;
2448
2449         if (tdp_enabled)
2450                 return init_kvm_tdp_mmu(vcpu);
2451         else
2452                 return init_kvm_softmmu(vcpu);
2453 }
2454
2455 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2456 {
2457         ASSERT(vcpu);
2458         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2459                 vcpu->arch.mmu.free(vcpu);
2460                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2461         }
2462 }
2463
2464 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2465 {
2466         destroy_kvm_mmu(vcpu);
2467         return init_kvm_mmu(vcpu);
2468 }
2469 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2470
2471 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2472 {
2473         int r;
2474
2475         r = mmu_topup_memory_caches(vcpu);
2476         if (r)
2477                 goto out;
2478         spin_lock(&vcpu->kvm->mmu_lock);
2479         kvm_mmu_free_some_pages(vcpu);
2480         r = mmu_alloc_roots(vcpu);
2481         mmu_sync_roots(vcpu);
2482         spin_unlock(&vcpu->kvm->mmu_lock);
2483         if (r)
2484                 goto out;
2485         /* set_cr3() should ensure TLB has been flushed */
2486         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2487 out:
2488         return r;
2489 }
2490 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2491
2492 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2493 {
2494         mmu_free_roots(vcpu);
2495 }
2496
2497 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2498                                   struct kvm_mmu_page *sp,
2499                                   u64 *spte)
2500 {
2501         u64 pte;
2502         struct kvm_mmu_page *child;
2503
2504         pte = *spte;
2505         if (is_shadow_present_pte(pte)) {
2506                 if (is_last_spte(pte, sp->role.level))
2507                         rmap_remove(vcpu->kvm, spte);
2508                 else {
2509                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2510                         mmu_page_remove_parent_pte(child, spte);
2511                 }
2512         }
2513         __set_spte(spte, shadow_trap_nonpresent_pte);
2514         if (is_large_pte(pte))
2515                 --vcpu->kvm->stat.lpages;
2516 }
2517
2518 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2519                                   struct kvm_mmu_page *sp,
2520                                   u64 *spte,
2521                                   const void *new)
2522 {
2523         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2524                 ++vcpu->kvm->stat.mmu_pde_zapped;
2525                 return;
2526         }
2527
2528         ++vcpu->kvm->stat.mmu_pte_updated;
2529         if (sp->role.glevels == PT32_ROOT_LEVEL)
2530                 paging32_update_pte(vcpu, sp, spte, new);
2531         else
2532                 paging64_update_pte(vcpu, sp, spte, new);
2533 }
2534
2535 static bool need_remote_flush(u64 old, u64 new)
2536 {
2537         if (!is_shadow_present_pte(old))
2538                 return false;
2539         if (!is_shadow_present_pte(new))
2540                 return true;
2541         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2542                 return true;
2543         old ^= PT64_NX_MASK;
2544         new ^= PT64_NX_MASK;
2545         return (old & ~new & PT64_PERM_MASK) != 0;
2546 }
2547
2548 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2549 {
2550         if (need_remote_flush(old, new))
2551                 kvm_flush_remote_tlbs(vcpu->kvm);
2552         else
2553                 kvm_mmu_flush_tlb(vcpu);
2554 }
2555
2556 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2557 {
2558         u64 *spte = vcpu->arch.last_pte_updated;
2559
2560         return !!(spte && (*spte & shadow_accessed_mask));
2561 }
2562
2563 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2564                                           const u8 *new, int bytes)
2565 {
2566         gfn_t gfn;
2567         int r;
2568         u64 gpte = 0;
2569         pfn_t pfn;
2570
2571         if (bytes != 4 && bytes != 8)
2572                 return;
2573
2574         /*
2575          * Assume that the pte write on a page table of the same type
2576          * as the current vcpu paging mode.  This is nearly always true
2577          * (might be false while changing modes).  Note it is verified later
2578          * by update_pte().
2579          */
2580         if (is_pae(vcpu)) {
2581                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2582                 if ((bytes == 4) && (gpa % 4 == 0)) {
2583                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2584                         if (r)
2585                                 return;
2586                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2587                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2588                         memcpy((void *)&gpte, new, 8);
2589                 }
2590         } else {
2591                 if ((bytes == 4) && (gpa % 4 == 0))
2592                         memcpy((void *)&gpte, new, 4);
2593         }
2594         if (!is_present_gpte(gpte))
2595                 return;
2596         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2597
2598         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2599         smp_rmb();
2600         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2601
2602         if (is_error_pfn(pfn)) {
2603                 kvm_release_pfn_clean(pfn);
2604                 return;
2605         }
2606         vcpu->arch.update_pte.gfn = gfn;
2607         vcpu->arch.update_pte.pfn = pfn;
2608 }
2609
2610 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2611 {
2612         u64 *spte = vcpu->arch.last_pte_updated;
2613
2614         if (spte
2615             && vcpu->arch.last_pte_gfn == gfn
2616             && shadow_accessed_mask
2617             && !(*spte & shadow_accessed_mask)
2618             && is_shadow_present_pte(*spte))
2619                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2620 }
2621
2622 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2623                        const u8 *new, int bytes,
2624                        bool guest_initiated)
2625 {
2626         gfn_t gfn = gpa >> PAGE_SHIFT;
2627         struct kvm_mmu_page *sp;
2628         struct hlist_node *node, *n;
2629         struct hlist_head *bucket;
2630         unsigned index;
2631         u64 entry, gentry;
2632         u64 *spte;
2633         unsigned offset = offset_in_page(gpa);
2634         unsigned pte_size;
2635         unsigned page_offset;
2636         unsigned misaligned;
2637         unsigned quadrant;
2638         int level;
2639         int flooded = 0;
2640         int npte;
2641         int r;
2642
2643         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2644         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2645         spin_lock(&vcpu->kvm->mmu_lock);
2646         kvm_mmu_access_page(vcpu, gfn);
2647         kvm_mmu_free_some_pages(vcpu);
2648         ++vcpu->kvm->stat.mmu_pte_write;
2649         kvm_mmu_audit(vcpu, "pre pte write");
2650         if (guest_initiated) {
2651                 if (gfn == vcpu->arch.last_pt_write_gfn
2652                     && !last_updated_pte_accessed(vcpu)) {
2653                         ++vcpu->arch.last_pt_write_count;
2654                         if (vcpu->arch.last_pt_write_count >= 3)
2655                                 flooded = 1;
2656                 } else {
2657                         vcpu->arch.last_pt_write_gfn = gfn;
2658                         vcpu->arch.last_pt_write_count = 1;
2659                         vcpu->arch.last_pte_updated = NULL;
2660                 }
2661         }
2662         index = kvm_page_table_hashfn(gfn);
2663         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2664         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2665                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2666                         continue;
2667                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2668                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2669                 misaligned |= bytes < 4;
2670                 if (misaligned || flooded) {
2671                         /*
2672                          * Misaligned accesses are too much trouble to fix
2673                          * up; also, they usually indicate a page is not used
2674                          * as a page table.
2675                          *
2676                          * If we're seeing too many writes to a page,
2677                          * it may no longer be a page table, or we may be
2678                          * forking, in which case it is better to unmap the
2679                          * page.
2680                          */
2681                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2682                                  gpa, bytes, sp->role.word);
2683                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2684                                 n = bucket->first;
2685                         ++vcpu->kvm->stat.mmu_flooded;
2686                         continue;
2687                 }
2688                 page_offset = offset;
2689                 level = sp->role.level;
2690                 npte = 1;
2691                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2692                         page_offset <<= 1;      /* 32->64 */
2693                         /*
2694                          * A 32-bit pde maps 4MB while the shadow pdes map
2695                          * only 2MB.  So we need to double the offset again
2696                          * and zap two pdes instead of one.
2697                          */
2698                         if (level == PT32_ROOT_LEVEL) {
2699                                 page_offset &= ~7; /* kill rounding error */
2700                                 page_offset <<= 1;
2701                                 npte = 2;
2702                         }
2703                         quadrant = page_offset >> PAGE_SHIFT;
2704                         page_offset &= ~PAGE_MASK;
2705                         if (quadrant != sp->role.quadrant)
2706                                 continue;
2707                 }
2708                 spte = &sp->spt[page_offset / sizeof(*spte)];
2709                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2710                         gentry = 0;
2711                         r = kvm_read_guest_atomic(vcpu->kvm,
2712                                                   gpa & ~(u64)(pte_size - 1),
2713                                                   &gentry, pte_size);
2714                         new = (const void *)&gentry;
2715                         if (r < 0)
2716                                 new = NULL;
2717                 }
2718                 while (npte--) {
2719                         entry = *spte;
2720                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2721                         if (new)
2722                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2723                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2724                         ++spte;
2725                 }
2726         }
2727         kvm_mmu_audit(vcpu, "post pte write");
2728         spin_unlock(&vcpu->kvm->mmu_lock);
2729         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2730                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2731                 vcpu->arch.update_pte.pfn = bad_pfn;
2732         }
2733 }
2734
2735 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2736 {
2737         gpa_t gpa;
2738         int r;
2739
2740         if (tdp_enabled)
2741                 return 0;
2742
2743         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2744
2745         spin_lock(&vcpu->kvm->mmu_lock);
2746         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2747         spin_unlock(&vcpu->kvm->mmu_lock);
2748         return r;
2749 }
2750 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2751
2752 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2753 {
2754         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2755                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2756                 struct kvm_mmu_page *sp;
2757
2758                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2759                                   struct kvm_mmu_page, link);
2760                 kvm_mmu_zap_page(vcpu->kvm, sp);
2761                 ++vcpu->kvm->stat.mmu_recycled;
2762         }
2763 }
2764
2765 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2766 {
2767         int r;
2768         enum emulation_result er;
2769
2770         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2771         if (r < 0)
2772                 goto out;
2773
2774         if (!r) {
2775                 r = 1;
2776                 goto out;
2777         }
2778
2779         r = mmu_topup_memory_caches(vcpu);
2780         if (r)
2781                 goto out;
2782
2783         er = emulate_instruction(vcpu, cr2, error_code, 0);
2784
2785         switch (er) {
2786         case EMULATE_DONE:
2787                 return 1;
2788         case EMULATE_DO_MMIO:
2789                 ++vcpu->stat.mmio_exits;
2790                 return 0;
2791         case EMULATE_FAIL:
2792                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2793                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2794                 vcpu->run->internal.ndata = 0;
2795                 return 0;
2796         default:
2797                 BUG();
2798         }
2799 out:
2800         return r;
2801 }
2802 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2803
2804 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2805 {
2806         vcpu->arch.mmu.invlpg(vcpu, gva);
2807         kvm_mmu_flush_tlb(vcpu);
2808         ++vcpu->stat.invlpg;
2809 }
2810 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2811
2812 void kvm_enable_tdp(void)
2813 {
2814         tdp_enabled = true;
2815 }
2816 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2817
2818 void kvm_disable_tdp(void)
2819 {
2820         tdp_enabled = false;
2821 }
2822 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2823
2824 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2825 {
2826         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2827 }
2828
2829 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2830 {
2831         struct page *page;
2832         int i;
2833
2834         ASSERT(vcpu);
2835
2836         /*
2837          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2838          * Therefore we need to allocate shadow page tables in the first
2839          * 4GB of memory, which happens to fit the DMA32 zone.
2840          */
2841         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2842         if (!page)
2843                 goto error_1;
2844         vcpu->arch.mmu.pae_root = page_address(page);
2845         for (i = 0; i < 4; ++i)
2846                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2847
2848         return 0;
2849
2850 error_1:
2851         free_mmu_pages(vcpu);
2852         return -ENOMEM;
2853 }
2854
2855 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2856 {
2857         ASSERT(vcpu);
2858         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2859
2860         return alloc_mmu_pages(vcpu);
2861 }
2862
2863 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2864 {
2865         ASSERT(vcpu);
2866         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2867
2868         return init_kvm_mmu(vcpu);
2869 }
2870
2871 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2872 {
2873         ASSERT(vcpu);
2874
2875         destroy_kvm_mmu(vcpu);
2876         free_mmu_pages(vcpu);
2877         mmu_free_memory_caches(vcpu);
2878 }
2879
2880 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2881 {
2882         struct kvm_mmu_page *sp;
2883
2884         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2885                 int i;
2886                 u64 *pt;
2887
2888                 if (!test_bit(slot, sp->slot_bitmap))
2889                         continue;
2890
2891                 pt = sp->spt;
2892                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2893                         /* avoid RMW */
2894                         if (pt[i] & PT_WRITABLE_MASK)
2895                                 pt[i] &= ~PT_WRITABLE_MASK;
2896         }
2897         kvm_flush_remote_tlbs(kvm);
2898 }
2899
2900 void kvm_mmu_zap_all(struct kvm *kvm)
2901 {
2902         struct kvm_mmu_page *sp, *node;
2903
2904         spin_lock(&kvm->mmu_lock);
2905         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2906                 if (kvm_mmu_zap_page(kvm, sp))
2907                         node = container_of(kvm->arch.active_mmu_pages.next,
2908                                             struct kvm_mmu_page, link);
2909         spin_unlock(&kvm->mmu_lock);
2910
2911         kvm_flush_remote_tlbs(kvm);
2912 }
2913
2914 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2915 {
2916         struct kvm_mmu_page *page;
2917
2918         page = container_of(kvm->arch.active_mmu_pages.prev,
2919                             struct kvm_mmu_page, link);
2920         kvm_mmu_zap_page(kvm, page);
2921 }
2922
2923 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2924 {
2925         struct kvm *kvm;
2926         struct kvm *kvm_freed = NULL;
2927         int cache_count = 0;
2928
2929         spin_lock(&kvm_lock);
2930
2931         list_for_each_entry(kvm, &vm_list, vm_list) {
2932                 int npages, idx;
2933
2934                 idx = srcu_read_lock(&kvm->srcu);
2935                 spin_lock(&kvm->mmu_lock);
2936                 npages = kvm->arch.n_alloc_mmu_pages -
2937                          kvm->arch.n_free_mmu_pages;
2938                 cache_count += npages;
2939                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2940                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2941                         cache_count--;
2942                         kvm_freed = kvm;
2943                 }
2944                 nr_to_scan--;
2945
2946                 spin_unlock(&kvm->mmu_lock);
2947                 srcu_read_unlock(&kvm->srcu, idx);
2948         }
2949         if (kvm_freed)
2950                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2951
2952         spin_unlock(&kvm_lock);
2953
2954         return cache_count;
2955 }
2956
2957 static struct shrinker mmu_shrinker = {
2958         .shrink = mmu_shrink,
2959         .seeks = DEFAULT_SEEKS * 10,
2960 };
2961
2962 static void mmu_destroy_caches(void)
2963 {
2964         if (pte_chain_cache)
2965                 kmem_cache_destroy(pte_chain_cache);
2966         if (rmap_desc_cache)
2967                 kmem_cache_destroy(rmap_desc_cache);
2968         if (mmu_page_header_cache)
2969                 kmem_cache_destroy(mmu_page_header_cache);
2970 }
2971
2972 void kvm_mmu_module_exit(void)
2973 {
2974         mmu_destroy_caches();
2975         unregister_shrinker(&mmu_shrinker);
2976 }
2977
2978 int kvm_mmu_module_init(void)
2979 {
2980         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2981                                             sizeof(struct kvm_pte_chain),
2982                                             0, 0, NULL);
2983         if (!pte_chain_cache)
2984                 goto nomem;
2985         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2986                                             sizeof(struct kvm_rmap_desc),
2987                                             0, 0, NULL);
2988         if (!rmap_desc_cache)
2989                 goto nomem;
2990
2991         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2992                                                   sizeof(struct kvm_mmu_page),
2993                                                   0, 0, NULL);
2994         if (!mmu_page_header_cache)
2995                 goto nomem;
2996
2997         register_shrinker(&mmu_shrinker);
2998
2999         return 0;
3000
3001 nomem:
3002         mmu_destroy_caches();
3003         return -ENOMEM;
3004 }
3005
3006 /*
3007  * Caculate mmu pages needed for kvm.
3008  */
3009 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3010 {
3011         int i;
3012         unsigned int nr_mmu_pages;
3013         unsigned int  nr_pages = 0;
3014         struct kvm_memslots *slots;
3015
3016         slots = rcu_dereference(kvm->memslots);
3017         for (i = 0; i < slots->nmemslots; i++)
3018                 nr_pages += slots->memslots[i].npages;
3019
3020         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3021         nr_mmu_pages = max(nr_mmu_pages,
3022                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3023
3024         return nr_mmu_pages;
3025 }
3026
3027 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3028                                 unsigned len)
3029 {
3030         if (len > buffer->len)
3031                 return NULL;
3032         return buffer->ptr;
3033 }
3034
3035 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3036                                 unsigned len)
3037 {
3038         void *ret;
3039
3040         ret = pv_mmu_peek_buffer(buffer, len);
3041         if (!ret)
3042                 return ret;
3043         buffer->ptr += len;
3044         buffer->len -= len;
3045         buffer->processed += len;
3046         return ret;
3047 }
3048
3049 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3050                              gpa_t addr, gpa_t value)
3051 {
3052         int bytes = 8;
3053         int r;
3054
3055         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3056                 bytes = 4;
3057
3058         r = mmu_topup_memory_caches(vcpu);
3059         if (r)
3060                 return r;
3061
3062         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3063                 return -EFAULT;
3064
3065         return 1;
3066 }
3067
3068 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3069 {
3070         kvm_set_cr3(vcpu, vcpu->arch.cr3);
3071         return 1;
3072 }
3073
3074 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3075 {
3076         spin_lock(&vcpu->kvm->mmu_lock);
3077         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3078         spin_unlock(&vcpu->kvm->mmu_lock);
3079         return 1;
3080 }
3081
3082 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3083                              struct kvm_pv_mmu_op_buffer *buffer)
3084 {
3085         struct kvm_mmu_op_header *header;
3086
3087         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3088         if (!header)
3089                 return 0;
3090         switch (header->op) {
3091         case KVM_MMU_OP_WRITE_PTE: {
3092                 struct kvm_mmu_op_write_pte *wpte;
3093
3094                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3095                 if (!wpte)
3096                         return 0;
3097                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3098                                         wpte->pte_val);
3099         }
3100         case KVM_MMU_OP_FLUSH_TLB: {
3101                 struct kvm_mmu_op_flush_tlb *ftlb;
3102
3103                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3104                 if (!ftlb)
3105                         return 0;
3106                 return kvm_pv_mmu_flush_tlb(vcpu);
3107         }
3108         case KVM_MMU_OP_RELEASE_PT: {
3109                 struct kvm_mmu_op_release_pt *rpt;
3110
3111                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3112                 if (!rpt)
3113                         return 0;
3114                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3115         }
3116         default: return 0;
3117         }
3118 }
3119
3120 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3121                   gpa_t addr, unsigned long *ret)
3122 {
3123         int r;
3124         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3125
3126         buffer->ptr = buffer->buf;
3127         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3128         buffer->processed = 0;
3129
3130         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3131         if (r)
3132                 goto out;
3133
3134         while (buffer->len) {
3135                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3136                 if (r < 0)
3137                         goto out;
3138                 if (r == 0)
3139                         break;
3140         }
3141
3142         r = 1;
3143 out:
3144         *ret = buffer->processed;
3145         return r;
3146 }
3147
3148 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3149 {
3150         struct kvm_shadow_walk_iterator iterator;
3151         int nr_sptes = 0;
3152
3153         spin_lock(&vcpu->kvm->mmu_lock);
3154         for_each_shadow_entry(vcpu, addr, iterator) {
3155                 sptes[iterator.level-1] = *iterator.sptep;
3156                 nr_sptes++;
3157                 if (!is_shadow_present_pte(*iterator.sptep))
3158                         break;
3159         }
3160         spin_unlock(&vcpu->kvm->mmu_lock);
3161
3162         return nr_sptes;
3163 }
3164 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3165
3166 #ifdef AUDIT
3167
3168 static const char *audit_msg;
3169
3170 static gva_t canonicalize(gva_t gva)
3171 {
3172 #ifdef CONFIG_X86_64
3173         gva = (long long)(gva << 16) >> 16;
3174 #endif
3175         return gva;
3176 }
3177
3178
3179 typedef void (*inspect_spte_fn) (struct kvm *kvm, struct kvm_mmu_page *sp,
3180                                  u64 *sptep);
3181
3182 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3183                             inspect_spte_fn fn)
3184 {
3185         int i;
3186
3187         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3188                 u64 ent = sp->spt[i];
3189
3190                 if (is_shadow_present_pte(ent)) {
3191                         if (!is_last_spte(ent, sp->role.level)) {
3192                                 struct kvm_mmu_page *child;
3193                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3194                                 __mmu_spte_walk(kvm, child, fn);
3195                         } else
3196                                 fn(kvm, sp, &sp->spt[i]);
3197                 }
3198         }
3199 }
3200
3201 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3202 {
3203         int i;
3204         struct kvm_mmu_page *sp;
3205
3206         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3207                 return;
3208         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3209                 hpa_t root = vcpu->arch.mmu.root_hpa;
3210                 sp = page_header(root);
3211                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3212                 return;
3213         }
3214         for (i = 0; i < 4; ++i) {
3215                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3216
3217                 if (root && VALID_PAGE(root)) {
3218                         root &= PT64_BASE_ADDR_MASK;
3219                         sp = page_header(root);
3220                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3221                 }
3222         }
3223         return;
3224 }
3225
3226 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3227                                 gva_t va, int level)
3228 {
3229         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3230         int i;
3231         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3232
3233         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3234                 u64 ent = pt[i];
3235
3236                 if (ent == shadow_trap_nonpresent_pte)
3237                         continue;
3238
3239                 va = canonicalize(va);
3240                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3241                         audit_mappings_page(vcpu, ent, va, level - 1);
3242                 else {
3243                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3244                         gfn_t gfn = gpa >> PAGE_SHIFT;
3245                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3246                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3247
3248                         if (is_error_pfn(pfn)) {
3249                                 kvm_release_pfn_clean(pfn);
3250                                 continue;
3251                         }
3252
3253                         if (is_shadow_present_pte(ent)
3254                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3255                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3256                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3257                                        audit_msg, vcpu->arch.mmu.root_level,
3258                                        va, gpa, hpa, ent,
3259                                        is_shadow_present_pte(ent));
3260                         else if (ent == shadow_notrap_nonpresent_pte
3261                                  && !is_error_hpa(hpa))
3262                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3263                                        " valid guest gva %lx\n", audit_msg, va);
3264                         kvm_release_pfn_clean(pfn);
3265
3266                 }
3267         }
3268 }
3269
3270 static void audit_mappings(struct kvm_vcpu *vcpu)
3271 {
3272         unsigned i;
3273
3274         if (vcpu->arch.mmu.root_level == 4)
3275                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3276         else
3277                 for (i = 0; i < 4; ++i)
3278                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3279                                 audit_mappings_page(vcpu,
3280                                                     vcpu->arch.mmu.pae_root[i],
3281                                                     i << 30,
3282                                                     2);
3283 }
3284
3285 static int count_rmaps(struct kvm_vcpu *vcpu)
3286 {
3287         int nmaps = 0;
3288         int i, j, k, idx;
3289
3290         idx = srcu_read_lock(&kvm->srcu);
3291         slots = rcu_dereference(kvm->memslots);
3292         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3293                 struct kvm_memory_slot *m = &slots->memslots[i];
3294                 struct kvm_rmap_desc *d;
3295
3296                 for (j = 0; j < m->npages; ++j) {
3297                         unsigned long *rmapp = &m->rmap[j];
3298
3299                         if (!*rmapp)
3300                                 continue;
3301                         if (!(*rmapp & 1)) {
3302                                 ++nmaps;
3303                                 continue;
3304                         }
3305                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3306                         while (d) {
3307                                 for (k = 0; k < RMAP_EXT; ++k)
3308                                         if (d->sptes[k])
3309                                                 ++nmaps;
3310                                         else
3311                                                 break;
3312                                 d = d->more;
3313                         }
3314                 }
3315         }
3316         srcu_read_unlock(&kvm->srcu, idx);
3317         return nmaps;
3318 }
3319
3320 void inspect_spte_has_rmap(struct kvm *kvm, struct kvm_mmu_page *sp, u64 *sptep)
3321 {
3322         unsigned long *rmapp;
3323         struct kvm_mmu_page *rev_sp;
3324         gfn_t gfn;
3325
3326         if (*sptep & PT_WRITABLE_MASK) {
3327                 rev_sp = page_header(__pa(sptep));
3328                 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3329
3330                 if (!gfn_to_memslot(kvm, gfn)) {
3331                         if (!printk_ratelimit())
3332                                 return;
3333                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3334                                          audit_msg, gfn);
3335                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3336                                         audit_msg, sptep - rev_sp->spt,
3337                                         rev_sp->gfn);
3338                         dump_stack();
3339                         return;
3340                 }
3341
3342                 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3343                                     is_large_pte(*sptep));
3344                 if (!*rmapp) {
3345                         if (!printk_ratelimit())
3346                                 return;
3347                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3348                                          audit_msg, *sptep);
3349                         dump_stack();
3350                 }
3351         }
3352
3353 }
3354
3355 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3356 {
3357         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3358 }
3359
3360 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3361 {
3362         struct kvm_mmu_page *sp;
3363         int i;
3364
3365         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3366                 u64 *pt = sp->spt;
3367
3368                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3369                         continue;
3370
3371                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3372                         u64 ent = pt[i];
3373
3374                         if (!(ent & PT_PRESENT_MASK))
3375                                 continue;
3376                         if (!(ent & PT_WRITABLE_MASK))
3377                                 continue;
3378                         inspect_spte_has_rmap(vcpu->kvm, sp, &pt[i]);
3379                 }
3380         }
3381         return;
3382 }
3383
3384 static void audit_rmap(struct kvm_vcpu *vcpu)
3385 {
3386         check_writable_mappings_rmap(vcpu);
3387         count_rmaps(vcpu);
3388 }
3389
3390 static void audit_write_protection(struct kvm_vcpu *vcpu)
3391 {
3392         struct kvm_mmu_page *sp;
3393         struct kvm_memory_slot *slot;
3394         unsigned long *rmapp;
3395         u64 *spte;
3396         gfn_t gfn;
3397
3398         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3399                 if (sp->role.direct)
3400                         continue;
3401                 if (sp->unsync)
3402                         continue;
3403
3404                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3405                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3406                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3407
3408                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3409                 while (spte) {
3410                         if (*spte & PT_WRITABLE_MASK)
3411                                 printk(KERN_ERR "%s: (%s) shadow page has "
3412                                 "writable mappings: gfn %lx role %x\n",
3413                                __func__, audit_msg, sp->gfn,
3414                                sp->role.word);
3415                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3416                 }
3417         }
3418 }
3419
3420 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3421 {
3422         int olddbg = dbg;
3423
3424         dbg = 0;
3425         audit_msg = msg;
3426         audit_rmap(vcpu);
3427         audit_write_protection(vcpu);
3428         if (strcmp("pre pte write", audit_msg) != 0)
3429                 audit_mappings(vcpu);
3430         audit_writable_sptes_have_rmaps(vcpu);
3431         dbg = olddbg;
3432 }
3433
3434 #endif