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