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
  22 #include <linux/kvm_host.h>
  23 #include <linux/types.h>
  24 #include <linux/string.h>
  25 #include <linux/mm.h>
  26 #include <linux/highmem.h>
  27 #include <linux/module.h>
  28 #include <linux/swap.h>
  29 #include <linux/hugetlb.h>
  30 #include <linux/compiler.h>
  31
  32 #include <asm/page.h>
  33 #include <asm/cmpxchg.h>
  34 #include <asm/io.h>
  35 #include <asm/vmx.h>
  36
  37 /*
  38  * When setting this variable to true it enables Two-Dimensional-Paging
  39  * where the hardware walks 2 page tables:
  40  * 1. the guest-virtual to guest-physical
  41  * 2. while doing 1. it walks guest-physical to host-physical
  42  * If the hardware supports that we don't need to do shadow paging.
  43  */
  44 bool tdp_enabled = false;
  45
  46 #undef MMU_DEBUG
  47
  48 #undef AUDIT
  49
  50 #ifdef AUDIT
  51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  52 #else
  53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  54 #endif
  55
  56 #ifdef MMU_DEBUG
  57
  58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  60
  61 #else
  62
  63 #define pgprintk(x...) do { } while (0)
  64 #define rmap_printk(x...) do { } while (0)
  65
  66 #endif
  67
  68 #if defined(MMU_DEBUG) || defined(AUDIT)
  69 static int dbg = 0;
  70 module_param(dbg, bool, 0644);
  71 #endif
  72
  73 static int oos_shadow = 1;
  74 module_param(oos_shadow, bool, 0644);
  75
  76 #ifndef MMU_DEBUG
  77 #define ASSERT(x) do { } while (0)
  78 #else
  79 #define ASSERT(x)                                                       \
  80         if (!(x)) {                                                     \
  81                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  82                        __FILE__, __LINE__, #x);                         \
  83         }
  84 #endif
  85
  86 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  87 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  88
  89 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  90
  91 #define PT64_LEVEL_BITS 9
  92
  93 #define PT64_LEVEL_SHIFT(level) \
  94                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  95
  96 #define PT64_LEVEL_MASK(level) \
  97                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  98
  99 #define PT64_INDEX(address, level)\
 100         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 101
 102
 103 #define PT32_LEVEL_BITS 10
 104
 105 #define PT32_LEVEL_SHIFT(level) \
 106                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 107
 108 #define PT32_LEVEL_MASK(level) \
 109                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 110
 111 #define PT32_INDEX(address, level)\
 112         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 113
 114
 115 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 116 #define PT64_DIR_BASE_ADDR_MASK \
 117         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 118
 119 #define PT32_BASE_ADDR_MASK PAGE_MASK
 120 #define PT32_DIR_BASE_ADDR_MASK \
 121         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 122
 123 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 124                         | PT64_NX_MASK)
 125
 126 #define PFERR_PRESENT_MASK (1U << 0)
 127 #define PFERR_WRITE_MASK (1U << 1)
 128 #define PFERR_USER_MASK (1U << 2)
 129 #define PFERR_RSVD_MASK (1U << 3)
 130 #define PFERR_FETCH_MASK (1U << 4)
 131
 132 #define PT_DIRECTORY_LEVEL 2
 133 #define PT_PAGE_TABLE_LEVEL 1
 134
 135 #define RMAP_EXT 4
 136
 137 #define ACC_EXEC_MASK    1
 138 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 139 #define ACC_USER_MASK    PT_USER_MASK
 140 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 141
 142 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 143
 144 struct kvm_rmap_desc {
 145         u64 *shadow_ptes[RMAP_EXT];
 146         struct kvm_rmap_desc *more;
 147 };
 148
 149 struct kvm_shadow_walk_iterator {
 150         u64 addr;
 151         hpa_t shadow_addr;
 152         int level;
 153         u64 *sptep;
 154         unsigned index;
 155 };
 156
 157 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
 158         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
 159              shadow_walk_okay(&(_walker));                      \
 160              shadow_walk_next(&(_walker)))
 161
 162
 163 struct kvm_unsync_walk {
 164         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 165 };
 166
 167 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 168
 169 static struct kmem_cache *pte_chain_cache;
 170 static struct kmem_cache *rmap_desc_cache;
 171 static struct kmem_cache *mmu_page_header_cache;
 172
 173 static u64 __read_mostly shadow_trap_nonpresent_pte;
 174 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 175 static u64 __read_mostly shadow_base_present_pte;
 176 static u64 __read_mostly shadow_nx_mask;
 177 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 178 static u64 __read_mostly shadow_user_mask;
 179 static u64 __read_mostly shadow_accessed_mask;
 180 static u64 __read_mostly shadow_dirty_mask;
 181 static u64 __read_mostly shadow_mt_mask;
 182
 183 static inline u64 rsvd_bits(int s, int e)
 184 {
 185         return ((1ULL << (e - s + 1)) - 1) << s;
 186 }
 187
 188 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 189 {
 190         shadow_trap_nonpresent_pte = trap_pte;
 191         shadow_notrap_nonpresent_pte = notrap_pte;
 192 }
 193 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 194
 195 void kvm_mmu_set_base_ptes(u64 base_pte)
 196 {
 197         shadow_base_present_pte = base_pte;
 198 }
 199 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 200
 201 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 202                 u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 mt_mask)
 203 {
 204         shadow_user_mask = user_mask;
 205         shadow_accessed_mask = accessed_mask;
 206         shadow_dirty_mask = dirty_mask;
 207         shadow_nx_mask = nx_mask;
 208         shadow_x_mask = x_mask;
 209         shadow_mt_mask = mt_mask;
 210 }
 211 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 212
 213 static int is_write_protection(struct kvm_vcpu *vcpu)
 214 {
 215         return vcpu->arch.cr0 & X86_CR0_WP;
 216 }
 217
 218 static int is_cpuid_PSE36(void)
 219 {
 220         return 1;
 221 }
 222
 223 static int is_nx(struct kvm_vcpu *vcpu)
 224 {
 225         return vcpu->arch.shadow_efer & EFER_NX;
 226 }
 227
 228 static int is_present_pte(unsigned long pte)
 229 {
 230         return pte & PT_PRESENT_MASK;
 231 }
 232
 233 static int is_shadow_present_pte(u64 pte)
 234 {
 235         return pte != shadow_trap_nonpresent_pte
 236                 && pte != shadow_notrap_nonpresent_pte;
 237 }
 238
 239 static int is_large_pte(u64 pte)
 240 {
 241         return pte & PT_PAGE_SIZE_MASK;
 242 }
 243
 244 static int is_writeble_pte(unsigned long pte)
 245 {
 246         return pte & PT_WRITABLE_MASK;
 247 }
 248
 249 static int is_dirty_pte(unsigned long pte)
 250 {
 251         return pte & shadow_dirty_mask;
 252 }
 253
 254 static int is_rmap_pte(u64 pte)
 255 {
 256         return is_shadow_present_pte(pte);
 257 }
 258
 259 static pfn_t spte_to_pfn(u64 pte)
 260 {
 261         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 262 }
 263
 264 static gfn_t pse36_gfn_delta(u32 gpte)
 265 {
 266         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 267
 268         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 269 }
 270
 271 static void set_shadow_pte(u64 *sptep, u64 spte)
 272 {
 273 #ifdef CONFIG_X86_64
 274         set_64bit((unsigned long *)sptep, spte);
 275 #else
 276         set_64bit((unsigned long long *)sptep, spte);
 277 #endif
 278 }
 279
 280 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 281                                   struct kmem_cache *base_cache, int min)
 282 {
 283         void *obj;
 284
 285         if (cache->nobjs >= min)
 286                 return 0;
 287         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 288                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 289                 if (!obj)
 290                         return -ENOMEM;
 291                 cache->objects[cache->nobjs++] = obj;
 292         }
 293         return 0;
 294 }
 295
 296 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 297 {
 298         while (mc->nobjs)
 299                 kfree(mc->objects[--mc->nobjs]);
 300 }
 301
 302 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 303                                        int min)
 304 {
 305         struct page *page;
 306
 307         if (cache->nobjs >= min)
 308                 return 0;
 309         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 310                 page = alloc_page(GFP_KERNEL);
 311                 if (!page)
 312                         return -ENOMEM;
 313                 set_page_private(page, 0);
 314                 cache->objects[cache->nobjs++] = page_address(page);
 315         }
 316         return 0;
 317 }
 318
 319 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 320 {
 321         while (mc->nobjs)
 322                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 323 }
 324
 325 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 326 {
 327         int r;
 328
 329         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 330                                    pte_chain_cache, 4);
 331         if (r)
 332                 goto out;
 333         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 334                                    rmap_desc_cache, 4);
 335         if (r)
 336                 goto out;
 337         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 338         if (r)
 339                 goto out;
 340         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 341                                    mmu_page_header_cache, 4);
 342 out:
 343         return r;
 344 }
 345
 346 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 347 {
 348         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 349         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 350         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 351         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 352 }
 353
 354 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 355                                     size_t size)
 356 {
 357         void *p;
 358
 359         BUG_ON(!mc->nobjs);
 360         p = mc->objects[--mc->nobjs];
 361         return p;
 362 }
 363
 364 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 365 {
 366         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 367                                       sizeof(struct kvm_pte_chain));
 368 }
 369
 370 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 371 {
 372         kfree(pc);
 373 }
 374
 375 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 376 {
 377         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 378                                       sizeof(struct kvm_rmap_desc));
 379 }
 380
 381 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 382 {
 383         kfree(rd);
 384 }
 385
 386 /*
 387  * Return the pointer to the largepage write count for a given
 388  * gfn, handling slots that are not large page aligned.
 389  */
 390 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 391 {
 392         unsigned long idx;
 393
 394         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 395               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 396         return &slot->lpage_info[idx].write_count;
 397 }
 398
 399 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 400 {
 401         int *write_count;
 402
 403         gfn = unalias_gfn(kvm, gfn);
 404         write_count = slot_largepage_idx(gfn,
 405                                          gfn_to_memslot_unaliased(kvm, gfn));
 406         *write_count += 1;
 407 }
 408
 409 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 410 {
 411         int *write_count;
 412
 413         gfn = unalias_gfn(kvm, gfn);
 414         write_count = slot_largepage_idx(gfn,
 415                                          gfn_to_memslot_unaliased(kvm, gfn));
 416         *write_count -= 1;
 417         WARN_ON(*write_count < 0);
 418 }
 419
 420 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 421 {
 422         struct kvm_memory_slot *slot;
 423         int *largepage_idx;
 424
 425         gfn = unalias_gfn(kvm, gfn);
 426         slot = gfn_to_memslot_unaliased(kvm, gfn);
 427         if (slot) {
 428                 largepage_idx = slot_largepage_idx(gfn, slot);
 429                 return *largepage_idx;
 430         }
 431
 432         return 1;
 433 }
 434
 435 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 436 {
 437         struct vm_area_struct *vma;
 438         unsigned long addr;
 439         int ret = 0;
 440
 441         addr = gfn_to_hva(kvm, gfn);
 442         if (kvm_is_error_hva(addr))
 443                 return ret;
 444
 445         down_read(&current->mm->mmap_sem);
 446         vma = find_vma(current->mm, addr);
 447         if (vma && is_vm_hugetlb_page(vma))
 448                 ret = 1;
 449         up_read(&current->mm->mmap_sem);
 450
 451         return ret;
 452 }
 453
 454 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 455 {
 456         struct kvm_memory_slot *slot;
 457
 458         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 459                 return 0;
 460
 461         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 462                 return 0;
 463
 464         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 465         if (slot && slot->dirty_bitmap)
 466                 return 0;
 467
 468         return 1;
 469 }
 470
 471 /*
 472  * Take gfn and return the reverse mapping to it.
 473  * Note: gfn must be unaliased before this function get called
 474  */
 475
 476 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 477 {
 478         struct kvm_memory_slot *slot;
 479         unsigned long idx;
 480
 481         slot = gfn_to_memslot(kvm, gfn);
 482         if (!lpage)
 483                 return &slot->rmap[gfn - slot->base_gfn];
 484
 485         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 486               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 487
 488         return &slot->lpage_info[idx].rmap_pde;
 489 }
 490
 491 /*
 492  * Reverse mapping data structures:
 493  *
 494  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 495  * that points to page_address(page).
 496  *
 497  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 498  * containing more mappings.
 499  */
 500 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 501 {
 502         struct kvm_mmu_page *sp;
 503         struct kvm_rmap_desc *desc;
 504         unsigned long *rmapp;
 505         int i;
 506
 507         if (!is_rmap_pte(*spte))
 508                 return;
 509         gfn = unalias_gfn(vcpu->kvm, gfn);
 510         sp = page_header(__pa(spte));
 511         sp->gfns[spte - sp->spt] = gfn;
 512         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 513         if (!*rmapp) {
 514                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 515                 *rmapp = (unsigned long)spte;
 516         } else if (!(*rmapp & 1)) {
 517                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 518                 desc = mmu_alloc_rmap_desc(vcpu);
 519                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 520                 desc->shadow_ptes[1] = spte;
 521                 *rmapp = (unsigned long)desc | 1;
 522         } else {
 523                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 524                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 525                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 526                         desc = desc->more;
 527                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 528                         desc->more = mmu_alloc_rmap_desc(vcpu);
 529                         desc = desc->more;
 530                 }
 531                 for (i = 0; desc->shadow_ptes[i]; ++i)
 532                         ;
 533                 desc->shadow_ptes[i] = spte;
 534         }
 535 }
 536
 537 static void rmap_desc_remove_entry(unsigned long *rmapp,
 538                                    struct kvm_rmap_desc *desc,
 539                                    int i,
 540                                    struct kvm_rmap_desc *prev_desc)
 541 {
 542         int j;
 543
 544         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 545                 ;
 546         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 547         desc->shadow_ptes[j] = NULL;
 548         if (j != 0)
 549                 return;
 550         if (!prev_desc && !desc->more)
 551                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 552         else
 553                 if (prev_desc)
 554                         prev_desc->more = desc->more;
 555                 else
 556                         *rmapp = (unsigned long)desc->more | 1;
 557         mmu_free_rmap_desc(desc);
 558 }
 559
 560 static void rmap_remove(struct kvm *kvm, u64 *spte)
 561 {
 562         struct kvm_rmap_desc *desc;
 563         struct kvm_rmap_desc *prev_desc;
 564         struct kvm_mmu_page *sp;
 565         pfn_t pfn;
 566         unsigned long *rmapp;
 567         int i;
 568
 569         if (!is_rmap_pte(*spte))
 570                 return;
 571         sp = page_header(__pa(spte));
 572         pfn = spte_to_pfn(*spte);
 573         if (*spte & shadow_accessed_mask)
 574                 kvm_set_pfn_accessed(pfn);
 575         if (is_writeble_pte(*spte))
 576                 kvm_release_pfn_dirty(pfn);
 577         else
 578                 kvm_release_pfn_clean(pfn);
 579         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 580         if (!*rmapp) {
 581                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 582                 BUG();
 583         } else if (!(*rmapp & 1)) {
 584                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 585                 if ((u64 *)*rmapp != spte) {
 586                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 587                                spte, *spte);
 588                         BUG();
 589                 }
 590                 *rmapp = 0;
 591         } else {
 592                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 593                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 594                 prev_desc = NULL;
 595                 while (desc) {
 596                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 597                                 if (desc->shadow_ptes[i] == spte) {
 598                                         rmap_desc_remove_entry(rmapp,
 599                                                                desc, i,
 600                                                                prev_desc);
 601                                         return;
 602                                 }
 603                         prev_desc = desc;
 604                         desc = desc->more;
 605                 }
 606                 BUG();
 607         }
 608 }
 609
 610 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 611 {
 612         struct kvm_rmap_desc *desc;
 613         struct kvm_rmap_desc *prev_desc;
 614         u64 *prev_spte;
 615         int i;
 616
 617         if (!*rmapp)
 618                 return NULL;
 619         else if (!(*rmapp & 1)) {
 620                 if (!spte)
 621                         return (u64 *)*rmapp;
 622                 return NULL;
 623         }
 624         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 625         prev_desc = NULL;
 626         prev_spte = NULL;
 627         while (desc) {
 628                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 629                         if (prev_spte == spte)
 630                                 return desc->shadow_ptes[i];
 631                         prev_spte = desc->shadow_ptes[i];
 632                 }
 633                 desc = desc->more;
 634         }
 635         return NULL;
 636 }
 637
 638 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 639 {
 640         unsigned long *rmapp;
 641         u64 *spte;
 642         int write_protected = 0;
 643
 644         gfn = unalias_gfn(kvm, gfn);
 645         rmapp = gfn_to_rmap(kvm, gfn, 0);
 646
 647         spte = rmap_next(kvm, rmapp, NULL);
 648         while (spte) {
 649                 BUG_ON(!spte);
 650                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 651                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 652                 if (is_writeble_pte(*spte)) {
 653                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 654                         write_protected = 1;
 655                 }
 656                 spte = rmap_next(kvm, rmapp, spte);
 657         }
 658         if (write_protected) {
 659                 pfn_t pfn;
 660
 661                 spte = rmap_next(kvm, rmapp, NULL);
 662                 pfn = spte_to_pfn(*spte);
 663                 kvm_set_pfn_dirty(pfn);
 664         }
 665
 666         /* check for huge page mappings */
 667         rmapp = gfn_to_rmap(kvm, gfn, 1);
 668         spte = rmap_next(kvm, rmapp, NULL);
 669         while (spte) {
 670                 BUG_ON(!spte);
 671                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 672                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 673                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 674                 if (is_writeble_pte(*spte)) {
 675                         rmap_remove(kvm, spte);
 676                         --kvm->stat.lpages;
 677                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 678                         spte = NULL;
 679                         write_protected = 1;
 680                 }
 681                 spte = rmap_next(kvm, rmapp, spte);
 682         }
 683
 684         return write_protected;
 685 }
 686
 687 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 688 {
 689         u64 *spte;
 690         int need_tlb_flush = 0;
 691
 692         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 693                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 694                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 695                 rmap_remove(kvm, spte);
 696                 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 697                 need_tlb_flush = 1;
 698         }
 699         return need_tlb_flush;
 700 }
 701
 702 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 703                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 704 {
 705         int i;
 706         int retval = 0;
 707
 708         /*
 709          * If mmap_sem isn't taken, we can look the memslots with only
 710          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 711          */
 712         for (i = 0; i < kvm->nmemslots; i++) {
 713                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 714                 unsigned long start = memslot->userspace_addr;
 715                 unsigned long end;
 716
 717                 /* mmu_lock protects userspace_addr */
 718                 if (!start)
 719                         continue;
 720
 721                 end = start + (memslot->npages << PAGE_SHIFT);
 722                 if (hva >= start && hva < end) {
 723                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 724                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 725                         retval |= handler(kvm,
 726                                           &memslot->lpage_info[
 727                                                   gfn_offset /
 728                                                   KVM_PAGES_PER_HPAGE].rmap_pde);
 729                 }
 730         }
 731
 732         return retval;
 733 }
 734
 735 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 736 {
 737         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 738 }
 739
 740 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 741 {
 742         u64 *spte;
 743         int young = 0;
 744
 745         /* always return old for EPT */
 746         if (!shadow_accessed_mask)
 747                 return 0;
 748
 749         spte = rmap_next(kvm, rmapp, NULL);
 750         while (spte) {
 751                 int _young;
 752                 u64 _spte = *spte;
 753                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 754                 _young = _spte & PT_ACCESSED_MASK;
 755                 if (_young) {
 756                         young = 1;
 757                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 758                 }
 759                 spte = rmap_next(kvm, rmapp, spte);
 760         }
 761         return young;
 762 }
 763
 764 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 765 {
 766         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 767 }
 768
 769 #ifdef MMU_DEBUG
 770 static int is_empty_shadow_page(u64 *spt)
 771 {
 772         u64 *pos;
 773         u64 *end;
 774
 775         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 776                 if (is_shadow_present_pte(*pos)) {
 777                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 778                                pos, *pos);
 779                         return 0;
 780                 }
 781         return 1;
 782 }
 783 #endif
 784
 785 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 786 {
 787         ASSERT(is_empty_shadow_page(sp->spt));
 788         list_del(&sp->link);
 789         __free_page(virt_to_page(sp->spt));
 790         __free_page(virt_to_page(sp->gfns));
 791         kfree(sp);
 792         ++kvm->arch.n_free_mmu_pages;
 793 }
 794
 795 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 796 {
 797         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 798 }
 799
 800 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 801                                                u64 *parent_pte)
 802 {
 803         struct kvm_mmu_page *sp;
 804
 805         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 806         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 807         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 808         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 809         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 810         INIT_LIST_HEAD(&sp->oos_link);
 811         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 812         sp->multimapped = 0;
 813         sp->parent_pte = parent_pte;
 814         --vcpu->kvm->arch.n_free_mmu_pages;
 815         return sp;
 816 }
 817
 818 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 819                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 820 {
 821         struct kvm_pte_chain *pte_chain;
 822         struct hlist_node *node;
 823         int i;
 824
 825         if (!parent_pte)
 826                 return;
 827         if (!sp->multimapped) {
 828                 u64 *old = sp->parent_pte;
 829
 830                 if (!old) {
 831                         sp->parent_pte = parent_pte;
 832                         return;
 833                 }
 834                 sp->multimapped = 1;
 835                 pte_chain = mmu_alloc_pte_chain(vcpu);
 836                 INIT_HLIST_HEAD(&sp->parent_ptes);
 837                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 838                 pte_chain->parent_ptes[0] = old;
 839         }
 840         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 841                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 842                         continue;
 843                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 844                         if (!pte_chain->parent_ptes[i]) {
 845                                 pte_chain->parent_ptes[i] = parent_pte;
 846                                 return;
 847                         }
 848         }
 849         pte_chain = mmu_alloc_pte_chain(vcpu);
 850         BUG_ON(!pte_chain);
 851         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 852         pte_chain->parent_ptes[0] = parent_pte;
 853 }
 854
 855 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 856                                        u64 *parent_pte)
 857 {
 858         struct kvm_pte_chain *pte_chain;
 859         struct hlist_node *node;
 860         int i;
 861
 862         if (!sp->multimapped) {
 863                 BUG_ON(sp->parent_pte != parent_pte);
 864                 sp->parent_pte = NULL;
 865                 return;
 866         }
 867         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 868                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 869                         if (!pte_chain->parent_ptes[i])
 870                                 break;
 871                         if (pte_chain->parent_ptes[i] != parent_pte)
 872                                 continue;
 873                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 874                                 && pte_chain->parent_ptes[i + 1]) {
 875                                 pte_chain->parent_ptes[i]
 876                                         = pte_chain->parent_ptes[i + 1];
 877                                 ++i;
 878                         }
 879                         pte_chain->parent_ptes[i] = NULL;
 880                         if (i == 0) {
 881                                 hlist_del(&pte_chain->link);
 882                                 mmu_free_pte_chain(pte_chain);
 883                                 if (hlist_empty(&sp->parent_ptes)) {
 884                                         sp->multimapped = 0;
 885                                         sp->parent_pte = NULL;
 886                                 }
 887                         }
 888                         return;
 889                 }
 890         BUG();
 891 }
 892
 893
 894 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 895                             mmu_parent_walk_fn fn)
 896 {
 897         struct kvm_pte_chain *pte_chain;
 898         struct hlist_node *node;
 899         struct kvm_mmu_page *parent_sp;
 900         int i;
 901
 902         if (!sp->multimapped && sp->parent_pte) {
 903                 parent_sp = page_header(__pa(sp->parent_pte));
 904                 fn(vcpu, parent_sp);
 905                 mmu_parent_walk(vcpu, parent_sp, fn);
 906                 return;
 907         }
 908         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 909                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 910                         if (!pte_chain->parent_ptes[i])
 911                                 break;
 912                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 913                         fn(vcpu, parent_sp);
 914                         mmu_parent_walk(vcpu, parent_sp, fn);
 915                 }
 916 }
 917
 918 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 919 {
 920         unsigned int index;
 921         struct kvm_mmu_page *sp = page_header(__pa(spte));
 922
 923         index = spte - sp->spt;
 924         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
 925                 sp->unsync_children++;
 926         WARN_ON(!sp->unsync_children);
 927 }
 928
 929 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 930 {
 931         struct kvm_pte_chain *pte_chain;
 932         struct hlist_node *node;
 933         int i;
 934
 935         if (!sp->parent_pte)
 936                 return;
 937
 938         if (!sp->multimapped) {
 939                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 940                 return;
 941         }
 942
 943         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 944                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 945                         if (!pte_chain->parent_ptes[i])
 946                                 break;
 947                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 948                 }
 949 }
 950
 951 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 952 {
 953         kvm_mmu_update_parents_unsync(sp);
 954         return 1;
 955 }
 956
 957 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 958                                         struct kvm_mmu_page *sp)
 959 {
 960         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 961         kvm_mmu_update_parents_unsync(sp);
 962 }
 963
 964 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 965                                     struct kvm_mmu_page *sp)
 966 {
 967         int i;
 968
 969         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 970                 sp->spt[i] = shadow_trap_nonpresent_pte;
 971 }
 972
 973 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
 974                                struct kvm_mmu_page *sp)
 975 {
 976         return 1;
 977 }
 978
 979 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
 980 {
 981 }
 982
 983 #define KVM_PAGE_ARRAY_NR 16
 984
 985 struct kvm_mmu_pages {
 986         struct mmu_page_and_offset {
 987                 struct kvm_mmu_page *sp;
 988                 unsigned int idx;
 989         } page[KVM_PAGE_ARRAY_NR];
 990         unsigned int nr;
 991 };
 992
 993 #define for_each_unsync_children(bitmap, idx)           \
 994         for (idx = find_first_bit(bitmap, 512);         \
 995              idx < 512;                                 \
 996              idx = find_next_bit(bitmap, 512, idx+1))
 997
 998 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
 999                          int idx)
1000 {
1001         int i;
1002
1003         if (sp->unsync)
1004                 for (i=0; i < pvec->nr; i++)
1005                         if (pvec->page[i].sp == sp)
1006                                 return 0;
1007
1008         pvec->page[pvec->nr].sp = sp;
1009         pvec->page[pvec->nr].idx = idx;
1010         pvec->nr++;
1011         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1012 }
1013
1014 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1015                            struct kvm_mmu_pages *pvec)
1016 {
1017         int i, ret, nr_unsync_leaf = 0;
1018
1019         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1020                 u64 ent = sp->spt[i];
1021
1022                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1023                         struct kvm_mmu_page *child;
1024                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1025
1026                         if (child->unsync_children) {
1027                                 if (mmu_pages_add(pvec, child, i))
1028                                         return -ENOSPC;
1029
1030                                 ret = __mmu_unsync_walk(child, pvec);
1031                                 if (!ret)
1032                                         __clear_bit(i, sp->unsync_child_bitmap);
1033                                 else if (ret > 0)
1034                                         nr_unsync_leaf += ret;
1035                                 else
1036                                         return ret;
1037                         }
1038
1039                         if (child->unsync) {
1040                                 nr_unsync_leaf++;
1041                                 if (mmu_pages_add(pvec, child, i))
1042                                         return -ENOSPC;
1043                         }
1044                 }
1045         }
1046
1047         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1048                 sp->unsync_children = 0;
1049
1050         return nr_unsync_leaf;
1051 }
1052
1053 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1054                            struct kvm_mmu_pages *pvec)
1055 {
1056         if (!sp->unsync_children)
1057                 return 0;
1058
1059         mmu_pages_add(pvec, sp, 0);
1060         return __mmu_unsync_walk(sp, pvec);
1061 }
1062
1063 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1064 {
1065         unsigned index;
1066         struct hlist_head *bucket;
1067         struct kvm_mmu_page *sp;
1068         struct hlist_node *node;
1069
1070         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1071         index = kvm_page_table_hashfn(gfn);
1072         bucket = &kvm->arch.mmu_page_hash[index];
1073         hlist_for_each_entry(sp, node, bucket, hash_link)
1074                 if (sp->gfn == gfn && !sp->role.direct
1075                     && !sp->role.invalid) {
1076                         pgprintk("%s: found role %x\n",
1077                                  __func__, sp->role.word);
1078                         return sp;
1079                 }
1080         return NULL;
1081 }
1082
1083 static void kvm_unlink_unsync_global(struct kvm *kvm, struct kvm_mmu_page *sp)
1084 {
1085         list_del(&sp->oos_link);
1086         --kvm->stat.mmu_unsync_global;
1087 }
1088
1089 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1090 {
1091         WARN_ON(!sp->unsync);
1092         sp->unsync = 0;
1093         if (sp->global)
1094                 kvm_unlink_unsync_global(kvm, sp);
1095         --kvm->stat.mmu_unsync;
1096 }
1097
1098 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1099
1100 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1101 {
1102         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1103                 kvm_mmu_zap_page(vcpu->kvm, sp);
1104                 return 1;
1105         }
1106
1107         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1108                 kvm_flush_remote_tlbs(vcpu->kvm);
1109         kvm_unlink_unsync_page(vcpu->kvm, sp);
1110         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1111                 kvm_mmu_zap_page(vcpu->kvm, sp);
1112                 return 1;
1113         }
1114
1115         kvm_mmu_flush_tlb(vcpu);
1116         return 0;
1117 }
1118
1119 struct mmu_page_path {
1120         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1121         unsigned int idx[PT64_ROOT_LEVEL-1];
1122 };
1123
1124 #define for_each_sp(pvec, sp, parents, i)                       \
1125                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1126                         sp = pvec.page[i].sp;                   \
1127                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1128                         i = mmu_pages_next(&pvec, &parents, i))
1129
1130 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1131                           struct mmu_page_path *parents,
1132                           int i)
1133 {
1134         int n;
1135
1136         for (n = i+1; n < pvec->nr; n++) {
1137                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1138
1139                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1140                         parents->idx[0] = pvec->page[n].idx;
1141                         return n;
1142                 }
1143
1144                 parents->parent[sp->role.level-2] = sp;
1145                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1146         }
1147
1148         return n;
1149 }
1150
1151 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1152 {
1153         struct kvm_mmu_page *sp;
1154         unsigned int level = 0;
1155
1156         do {
1157                 unsigned int idx = parents->idx[level];
1158
1159                 sp = parents->parent[level];
1160                 if (!sp)
1161                         return;
1162
1163                 --sp->unsync_children;
1164                 WARN_ON((int)sp->unsync_children < 0);
1165                 __clear_bit(idx, sp->unsync_child_bitmap);
1166                 level++;
1167         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1168 }
1169
1170 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1171                                struct mmu_page_path *parents,
1172                                struct kvm_mmu_pages *pvec)
1173 {
1174         parents->parent[parent->role.level-1] = NULL;
1175         pvec->nr = 0;
1176 }
1177
1178 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1179                               struct kvm_mmu_page *parent)
1180 {
1181         int i;
1182         struct kvm_mmu_page *sp;
1183         struct mmu_page_path parents;
1184         struct kvm_mmu_pages pages;
1185
1186         kvm_mmu_pages_init(parent, &parents, &pages);
1187         while (mmu_unsync_walk(parent, &pages)) {
1188                 int protected = 0;
1189
1190                 for_each_sp(pages, sp, parents, i)
1191                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1192
1193                 if (protected)
1194                         kvm_flush_remote_tlbs(vcpu->kvm);
1195
1196                 for_each_sp(pages, sp, parents, i) {
1197                         kvm_sync_page(vcpu, sp);
1198                         mmu_pages_clear_parents(&parents);
1199                 }
1200                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1201                 kvm_mmu_pages_init(parent, &parents, &pages);
1202         }
1203 }
1204
1205 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1206                                              gfn_t gfn,
1207                                              gva_t gaddr,
1208                                              unsigned level,
1209                                              int direct,
1210                                              unsigned access,
1211                                              u64 *parent_pte)
1212 {
1213         union kvm_mmu_page_role role;
1214         unsigned index;
1215         unsigned quadrant;
1216         struct hlist_head *bucket;
1217         struct kvm_mmu_page *sp;
1218         struct hlist_node *node, *tmp;
1219
1220         role = vcpu->arch.mmu.base_role;
1221         role.level = level;
1222         role.direct = direct;
1223         role.access = access;
1224         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1225                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1226                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1227                 role.quadrant = quadrant;
1228         }
1229         pgprintk("%s: looking gfn %lx role %x\n", __func__,
1230                  gfn, role.word);
1231         index = kvm_page_table_hashfn(gfn);
1232         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1233         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1234                 if (sp->gfn == gfn) {
1235                         if (sp->unsync)
1236                                 if (kvm_sync_page(vcpu, sp))
1237                                         continue;
1238
1239                         if (sp->role.word != role.word)
1240                                 continue;
1241
1242                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1243                         if (sp->unsync_children) {
1244                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1245                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1246                         }
1247                         pgprintk("%s: found\n", __func__);
1248                         return sp;
1249                 }
1250         ++vcpu->kvm->stat.mmu_cache_miss;
1251         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1252         if (!sp)
1253                 return sp;
1254         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1255         sp->gfn = gfn;
1256         sp->role = role;
1257         sp->global = 0;
1258         hlist_add_head(&sp->hash_link, bucket);
1259         if (!direct) {
1260                 if (rmap_write_protect(vcpu->kvm, gfn))
1261                         kvm_flush_remote_tlbs(vcpu->kvm);
1262                 account_shadowed(vcpu->kvm, gfn);
1263         }
1264         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1265                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1266         else
1267                 nonpaging_prefetch_page(vcpu, sp);
1268         return sp;
1269 }
1270
1271 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1272                              struct kvm_vcpu *vcpu, u64 addr)
1273 {
1274         iterator->addr = addr;
1275         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1276         iterator->level = vcpu->arch.mmu.shadow_root_level;
1277         if (iterator->level == PT32E_ROOT_LEVEL) {
1278                 iterator->shadow_addr
1279                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1280                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1281                 --iterator->level;
1282                 if (!iterator->shadow_addr)
1283                         iterator->level = 0;
1284         }
1285 }
1286
1287 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1288 {
1289         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1290                 return false;
1291         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1292         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1293         return true;
1294 }
1295
1296 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1297 {
1298         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1299         --iterator->level;
1300 }
1301
1302 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1303                                          struct kvm_mmu_page *sp)
1304 {
1305         unsigned i;
1306         u64 *pt;
1307         u64 ent;
1308
1309         pt = sp->spt;
1310
1311         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1312                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1313                         if (is_shadow_present_pte(pt[i]))
1314                                 rmap_remove(kvm, &pt[i]);
1315                         pt[i] = shadow_trap_nonpresent_pte;
1316                 }
1317                 return;
1318         }
1319
1320         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1321                 ent = pt[i];
1322
1323                 if (is_shadow_present_pte(ent)) {
1324                         if (!is_large_pte(ent)) {
1325                                 ent &= PT64_BASE_ADDR_MASK;
1326                                 mmu_page_remove_parent_pte(page_header(ent),
1327                                                            &pt[i]);
1328                         } else {
1329                                 --kvm->stat.lpages;
1330                                 rmap_remove(kvm, &pt[i]);
1331                         }
1332                 }
1333                 pt[i] = shadow_trap_nonpresent_pte;
1334         }
1335 }
1336
1337 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1338 {
1339         mmu_page_remove_parent_pte(sp, parent_pte);
1340 }
1341
1342 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1343 {
1344         int i;
1345
1346         for (i = 0; i < KVM_MAX_VCPUS; ++i)
1347                 if (kvm->vcpus[i])
1348                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
1349 }
1350
1351 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1352 {
1353         u64 *parent_pte;
1354
1355         while (sp->multimapped || sp->parent_pte) {
1356                 if (!sp->multimapped)
1357                         parent_pte = sp->parent_pte;
1358                 else {
1359                         struct kvm_pte_chain *chain;
1360
1361                         chain = container_of(sp->parent_ptes.first,
1362                                              struct kvm_pte_chain, link);
1363                         parent_pte = chain->parent_ptes[0];
1364                 }
1365                 BUG_ON(!parent_pte);
1366                 kvm_mmu_put_page(sp, parent_pte);
1367                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1368         }
1369 }
1370
1371 static int mmu_zap_unsync_children(struct kvm *kvm,
1372                                    struct kvm_mmu_page *parent)
1373 {
1374         int i, zapped = 0;
1375         struct mmu_page_path parents;
1376         struct kvm_mmu_pages pages;
1377
1378         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1379                 return 0;
1380
1381         kvm_mmu_pages_init(parent, &parents, &pages);
1382         while (mmu_unsync_walk(parent, &pages)) {
1383                 struct kvm_mmu_page *sp;
1384
1385                 for_each_sp(pages, sp, parents, i) {
1386                         kvm_mmu_zap_page(kvm, sp);
1387                         mmu_pages_clear_parents(&parents);
1388                 }
1389                 zapped += pages.nr;
1390                 kvm_mmu_pages_init(parent, &parents, &pages);
1391         }
1392
1393         return zapped;
1394 }
1395
1396 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1397 {
1398         int ret;
1399         ++kvm->stat.mmu_shadow_zapped;
1400         ret = mmu_zap_unsync_children(kvm, sp);
1401         kvm_mmu_page_unlink_children(kvm, sp);
1402         kvm_mmu_unlink_parents(kvm, sp);
1403         kvm_flush_remote_tlbs(kvm);
1404         if (!sp->role.invalid && !sp->role.direct)
1405                 unaccount_shadowed(kvm, sp->gfn);
1406         if (sp->unsync)
1407                 kvm_unlink_unsync_page(kvm, sp);
1408         if (!sp->root_count) {
1409                 hlist_del(&sp->hash_link);
1410                 kvm_mmu_free_page(kvm, sp);
1411         } else {
1412                 sp->role.invalid = 1;
1413                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1414                 kvm_reload_remote_mmus(kvm);
1415         }
1416         kvm_mmu_reset_last_pte_updated(kvm);
1417         return ret;
1418 }
1419
1420 /*
1421  * Changing the number of mmu pages allocated to the vm
1422  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1423  */
1424 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1425 {
1426         /*
1427          * If we set the number of mmu pages to be smaller be than the
1428          * number of actived pages , we must to free some mmu pages before we
1429          * change the value
1430          */
1431
1432         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1433             kvm_nr_mmu_pages) {
1434                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1435                                        - kvm->arch.n_free_mmu_pages;
1436
1437                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1438                         struct kvm_mmu_page *page;
1439
1440                         page = container_of(kvm->arch.active_mmu_pages.prev,
1441                                             struct kvm_mmu_page, link);
1442                         kvm_mmu_zap_page(kvm, page);
1443                         n_used_mmu_pages--;
1444                 }
1445                 kvm->arch.n_free_mmu_pages = 0;
1446         }
1447         else
1448                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1449                                          - kvm->arch.n_alloc_mmu_pages;
1450
1451         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1452 }
1453
1454 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1455 {
1456         unsigned index;
1457         struct hlist_head *bucket;
1458         struct kvm_mmu_page *sp;
1459         struct hlist_node *node, *n;
1460         int r;
1461
1462         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1463         r = 0;
1464         index = kvm_page_table_hashfn(gfn);
1465         bucket = &kvm->arch.mmu_page_hash[index];
1466         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1467                 if (sp->gfn == gfn && !sp->role.direct) {
1468                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1469                                  sp->role.word);
1470                         r = 1;
1471                         if (kvm_mmu_zap_page(kvm, sp))
1472                                 n = bucket->first;
1473                 }
1474         return r;
1475 }
1476
1477 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1478 {
1479         unsigned index;
1480         struct hlist_head *bucket;
1481         struct kvm_mmu_page *sp;
1482         struct hlist_node *node, *nn;
1483
1484         index = kvm_page_table_hashfn(gfn);
1485         bucket = &kvm->arch.mmu_page_hash[index];
1486         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1487                 if (sp->gfn == gfn && !sp->role.direct
1488                     && !sp->role.invalid) {
1489                         pgprintk("%s: zap %lx %x\n",
1490                                  __func__, gfn, sp->role.word);
1491                         kvm_mmu_zap_page(kvm, sp);
1492                 }
1493         }
1494 }
1495
1496 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1497 {
1498         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1499         struct kvm_mmu_page *sp = page_header(__pa(pte));
1500
1501         __set_bit(slot, sp->slot_bitmap);
1502 }
1503
1504 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1505 {
1506         int i;
1507         u64 *pt = sp->spt;
1508
1509         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1510                 return;
1511
1512         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1513                 if (pt[i] == shadow_notrap_nonpresent_pte)
1514                         set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1515         }
1516 }
1517
1518 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1519 {
1520         struct page *page;
1521
1522         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1523
1524         if (gpa == UNMAPPED_GVA)
1525                 return NULL;
1526
1527         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1528
1529         return page;
1530 }
1531
1532 /*
1533  * The function is based on mtrr_type_lookup() in
1534  * arch/x86/kernel/cpu/mtrr/generic.c
1535  */
1536 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1537                          u64 start, u64 end)
1538 {
1539         int i;
1540         u64 base, mask;
1541         u8 prev_match, curr_match;
1542         int num_var_ranges = KVM_NR_VAR_MTRR;
1543
1544         if (!mtrr_state->enabled)
1545                 return 0xFF;
1546
1547         /* Make end inclusive end, instead of exclusive */
1548         end--;
1549
1550         /* Look in fixed ranges. Just return the type as per start */
1551         if (mtrr_state->have_fixed && (start < 0x100000)) {
1552                 int idx;
1553
1554                 if (start < 0x80000) {
1555                         idx = 0;
1556                         idx += (start >> 16);
1557                         return mtrr_state->fixed_ranges[idx];
1558                 } else if (start < 0xC0000) {
1559                         idx = 1 * 8;
1560                         idx += ((start - 0x80000) >> 14);
1561                         return mtrr_state->fixed_ranges[idx];
1562                 } else if (start < 0x1000000) {
1563                         idx = 3 * 8;
1564                         idx += ((start - 0xC0000) >> 12);
1565                         return mtrr_state->fixed_ranges[idx];
1566                 }
1567         }
1568
1569         /*
1570          * Look in variable ranges
1571          * Look of multiple ranges matching this address and pick type
1572          * as per MTRR precedence
1573          */
1574         if (!(mtrr_state->enabled & 2))
1575                 return mtrr_state->def_type;
1576
1577         prev_match = 0xFF;
1578         for (i = 0; i < num_var_ranges; ++i) {
1579                 unsigned short start_state, end_state;
1580
1581                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1582                         continue;
1583
1584                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1585                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1586                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1587                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1588
1589                 start_state = ((start & mask) == (base & mask));
1590                 end_state = ((end & mask) == (base & mask));
1591                 if (start_state != end_state)
1592                         return 0xFE;
1593
1594                 if ((start & mask) != (base & mask))
1595                         continue;
1596
1597                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1598                 if (prev_match == 0xFF) {
1599                         prev_match = curr_match;
1600                         continue;
1601                 }
1602
1603                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1604                     curr_match == MTRR_TYPE_UNCACHABLE)
1605                         return MTRR_TYPE_UNCACHABLE;
1606
1607                 if ((prev_match == MTRR_TYPE_WRBACK &&
1608                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1609                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1610                      curr_match == MTRR_TYPE_WRBACK)) {
1611                         prev_match = MTRR_TYPE_WRTHROUGH;
1612                         curr_match = MTRR_TYPE_WRTHROUGH;
1613                 }
1614
1615                 if (prev_match != curr_match)
1616                         return MTRR_TYPE_UNCACHABLE;
1617         }
1618
1619         if (prev_match != 0xFF)
1620                 return prev_match;
1621
1622         return mtrr_state->def_type;
1623 }
1624
1625 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1626 {
1627         u8 mtrr;
1628
1629         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1630                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1631         if (mtrr == 0xfe || mtrr == 0xff)
1632                 mtrr = MTRR_TYPE_WRBACK;
1633         return mtrr;
1634 }
1635
1636 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1637 {
1638         unsigned index;
1639         struct hlist_head *bucket;
1640         struct kvm_mmu_page *s;
1641         struct hlist_node *node, *n;
1642
1643         index = kvm_page_table_hashfn(sp->gfn);
1644         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1645         /* don't unsync if pagetable is shadowed with multiple roles */
1646         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1647                 if (s->gfn != sp->gfn || s->role.direct)
1648                         continue;
1649                 if (s->role.word != sp->role.word)
1650                         return 1;
1651         }
1652         ++vcpu->kvm->stat.mmu_unsync;
1653         sp->unsync = 1;
1654
1655         if (sp->global) {
1656                 list_add(&sp->oos_link, &vcpu->kvm->arch.oos_global_pages);
1657                 ++vcpu->kvm->stat.mmu_unsync_global;
1658         } else
1659                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1660
1661         mmu_convert_notrap(sp);
1662         return 0;
1663 }
1664
1665 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1666                                   bool can_unsync)
1667 {
1668         struct kvm_mmu_page *shadow;
1669
1670         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1671         if (shadow) {
1672                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1673                         return 1;
1674                 if (shadow->unsync)
1675                         return 0;
1676                 if (can_unsync && oos_shadow)
1677                         return kvm_unsync_page(vcpu, shadow);
1678                 return 1;
1679         }
1680         return 0;
1681 }
1682
1683 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1684                     unsigned pte_access, int user_fault,
1685                     int write_fault, int dirty, int largepage,
1686                     int global, gfn_t gfn, pfn_t pfn, bool speculative,
1687                     bool can_unsync)
1688 {
1689         u64 spte;
1690         int ret = 0;
1691         u64 mt_mask = shadow_mt_mask;
1692         struct kvm_mmu_page *sp = page_header(__pa(shadow_pte));
1693
1694         if (!global && sp->global) {
1695                 sp->global = 0;
1696                 if (sp->unsync) {
1697                         kvm_unlink_unsync_global(vcpu->kvm, sp);
1698                         kvm_mmu_mark_parents_unsync(vcpu, sp);
1699                 }
1700         }
1701
1702         /*
1703          * We don't set the accessed bit, since we sometimes want to see
1704          * whether the guest actually used the pte (in order to detect
1705          * demand paging).
1706          */
1707         spte = shadow_base_present_pte | shadow_dirty_mask;
1708         if (!speculative)
1709                 spte |= shadow_accessed_mask;
1710         if (!dirty)
1711                 pte_access &= ~ACC_WRITE_MASK;
1712         if (pte_access & ACC_EXEC_MASK)
1713                 spte |= shadow_x_mask;
1714         else
1715                 spte |= shadow_nx_mask;
1716         if (pte_access & ACC_USER_MASK)
1717                 spte |= shadow_user_mask;
1718         if (largepage)
1719                 spte |= PT_PAGE_SIZE_MASK;
1720         if (mt_mask) {
1721                 if (!kvm_is_mmio_pfn(pfn)) {
1722                         mt_mask = get_memory_type(vcpu, gfn) <<
1723                                 kvm_x86_ops->get_mt_mask_shift();
1724                         mt_mask |= VMX_EPT_IGMT_BIT;
1725                 } else
1726                         mt_mask = MTRR_TYPE_UNCACHABLE <<
1727                                 kvm_x86_ops->get_mt_mask_shift();
1728                 spte |= mt_mask;
1729         }
1730
1731         spte |= (u64)pfn << PAGE_SHIFT;
1732
1733         if ((pte_access & ACC_WRITE_MASK)
1734             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1735
1736                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1737                         ret = 1;
1738                         spte = shadow_trap_nonpresent_pte;
1739                         goto set_pte;
1740                 }
1741
1742                 spte |= PT_WRITABLE_MASK;
1743
1744                 /*
1745                  * Optimization: for pte sync, if spte was writable the hash
1746                  * lookup is unnecessary (and expensive). Write protection
1747                  * is responsibility of mmu_get_page / kvm_sync_page.
1748                  * Same reasoning can be applied to dirty page accounting.
1749                  */
1750                 if (!can_unsync && is_writeble_pte(*shadow_pte))
1751                         goto set_pte;
1752
1753                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1754                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1755                                  __func__, gfn);
1756                         ret = 1;
1757                         pte_access &= ~ACC_WRITE_MASK;
1758                         if (is_writeble_pte(spte))
1759                                 spte &= ~PT_WRITABLE_MASK;
1760                 }
1761         }
1762
1763         if (pte_access & ACC_WRITE_MASK)
1764                 mark_page_dirty(vcpu->kvm, gfn);
1765
1766 set_pte:
1767         set_shadow_pte(shadow_pte, spte);
1768         return ret;
1769 }
1770
1771 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1772                          unsigned pt_access, unsigned pte_access,
1773                          int user_fault, int write_fault, int dirty,
1774                          int *ptwrite, int largepage, int global,
1775                          gfn_t gfn, pfn_t pfn, bool speculative)
1776 {
1777         int was_rmapped = 0;
1778         int was_writeble = is_writeble_pte(*shadow_pte);
1779
1780         pgprintk("%s: spte %llx access %x write_fault %d"
1781                  " user_fault %d gfn %lx\n",
1782                  __func__, *shadow_pte, pt_access,
1783                  write_fault, user_fault, gfn);
1784
1785         if (is_rmap_pte(*shadow_pte)) {
1786                 /*
1787                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1788                  * the parent of the now unreachable PTE.
1789                  */
1790                 if (largepage && !is_large_pte(*shadow_pte)) {
1791                         struct kvm_mmu_page *child;
1792                         u64 pte = *shadow_pte;
1793
1794                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1795                         mmu_page_remove_parent_pte(child, shadow_pte);
1796                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1797                         pgprintk("hfn old %lx new %lx\n",
1798                                  spte_to_pfn(*shadow_pte), pfn);
1799                         rmap_remove(vcpu->kvm, shadow_pte);
1800                 } else
1801                         was_rmapped = 1;
1802         }
1803         if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1804                       dirty, largepage, global, gfn, pfn, speculative, true)) {
1805                 if (write_fault)
1806                         *ptwrite = 1;
1807                 kvm_x86_ops->tlb_flush(vcpu);
1808         }
1809
1810         pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1811         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1812                  is_large_pte(*shadow_pte)? "2MB" : "4kB",
1813                  is_present_pte(*shadow_pte)?"RW":"R", gfn,
1814                  *shadow_pte, shadow_pte);
1815         if (!was_rmapped && is_large_pte(*shadow_pte))
1816                 ++vcpu->kvm->stat.lpages;
1817
1818         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1819         if (!was_rmapped) {
1820                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1821                 if (!is_rmap_pte(*shadow_pte))
1822                         kvm_release_pfn_clean(pfn);
1823         } else {
1824                 if (was_writeble)
1825                         kvm_release_pfn_dirty(pfn);
1826                 else
1827                         kvm_release_pfn_clean(pfn);
1828         }
1829         if (speculative) {
1830                 vcpu->arch.last_pte_updated = shadow_pte;
1831                 vcpu->arch.last_pte_gfn = gfn;
1832         }
1833 }
1834
1835 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1836 {
1837 }
1838
1839 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1840                         int largepage, gfn_t gfn, pfn_t pfn)
1841 {
1842         struct kvm_shadow_walk_iterator iterator;
1843         struct kvm_mmu_page *sp;
1844         int pt_write = 0;
1845         gfn_t pseudo_gfn;
1846
1847         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1848                 if (iterator.level == PT_PAGE_TABLE_LEVEL
1849                     || (largepage && iterator.level == PT_DIRECTORY_LEVEL)) {
1850                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1851                                      0, write, 1, &pt_write,
1852                                      largepage, 0, gfn, pfn, false);
1853                         ++vcpu->stat.pf_fixed;
1854                         break;
1855                 }
1856
1857                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1858                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1859                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1860                                               iterator.level - 1,
1861                                               1, ACC_ALL, iterator.sptep);
1862                         if (!sp) {
1863                                 pgprintk("nonpaging_map: ENOMEM\n");
1864                                 kvm_release_pfn_clean(pfn);
1865                                 return -ENOMEM;
1866                         }
1867
1868                         set_shadow_pte(iterator.sptep,
1869                                        __pa(sp->spt)
1870                                        | PT_PRESENT_MASK | PT_WRITABLE_MASK
1871                                        | shadow_user_mask | shadow_x_mask);
1872                 }
1873         }
1874         return pt_write;
1875 }
1876
1877 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1878 {
1879         int r;
1880         int largepage = 0;
1881         pfn_t pfn;
1882         unsigned long mmu_seq;
1883
1884         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1885                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1886                 largepage = 1;
1887         }
1888
1889         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1890         smp_rmb();
1891         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1892
1893         /* mmio */
1894         if (is_error_pfn(pfn)) {
1895                 kvm_release_pfn_clean(pfn);
1896                 return 1;
1897         }
1898
1899         spin_lock(&vcpu->kvm->mmu_lock);
1900         if (mmu_notifier_retry(vcpu, mmu_seq))
1901                 goto out_unlock;
1902         kvm_mmu_free_some_pages(vcpu);
1903         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1904         spin_unlock(&vcpu->kvm->mmu_lock);
1905
1906
1907         return r;
1908
1909 out_unlock:
1910         spin_unlock(&vcpu->kvm->mmu_lock);
1911         kvm_release_pfn_clean(pfn);
1912         return 0;
1913 }
1914
1915
1916 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1917 {
1918         int i;
1919         struct kvm_mmu_page *sp;
1920
1921         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1922                 return;
1923         spin_lock(&vcpu->kvm->mmu_lock);
1924         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1925                 hpa_t root = vcpu->arch.mmu.root_hpa;
1926
1927                 sp = page_header(root);
1928                 --sp->root_count;
1929                 if (!sp->root_count && sp->role.invalid)
1930                         kvm_mmu_zap_page(vcpu->kvm, sp);
1931                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1932                 spin_unlock(&vcpu->kvm->mmu_lock);
1933                 return;
1934         }
1935         for (i = 0; i < 4; ++i) {
1936                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1937
1938                 if (root) {
1939                         root &= PT64_BASE_ADDR_MASK;
1940                         sp = page_header(root);
1941                         --sp->root_count;
1942                         if (!sp->root_count && sp->role.invalid)
1943                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1944                 }
1945                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1946         }
1947         spin_unlock(&vcpu->kvm->mmu_lock);
1948         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1949 }
1950
1951 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1952 {
1953         int i;
1954         gfn_t root_gfn;
1955         struct kvm_mmu_page *sp;
1956         int direct = 0;
1957
1958         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1959
1960         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1961                 hpa_t root = vcpu->arch.mmu.root_hpa;
1962
1963                 ASSERT(!VALID_PAGE(root));
1964                 if (tdp_enabled)
1965                         direct = 1;
1966                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1967                                       PT64_ROOT_LEVEL, direct,
1968                                       ACC_ALL, NULL);
1969                 root = __pa(sp->spt);
1970                 ++sp->root_count;
1971                 vcpu->arch.mmu.root_hpa = root;
1972                 return;
1973         }
1974         direct = !is_paging(vcpu);
1975         if (tdp_enabled)
1976                 direct = 1;
1977         for (i = 0; i < 4; ++i) {
1978                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1979
1980                 ASSERT(!VALID_PAGE(root));
1981                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1982                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1983                                 vcpu->arch.mmu.pae_root[i] = 0;
1984                                 continue;
1985                         }
1986                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1987                 } else if (vcpu->arch.mmu.root_level == 0)
1988                         root_gfn = 0;
1989                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1990                                       PT32_ROOT_LEVEL, direct,
1991                                       ACC_ALL, NULL);
1992                 root = __pa(sp->spt);
1993                 ++sp->root_count;
1994                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1995         }
1996         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1997 }
1998
1999 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2000 {
2001         int i;
2002         struct kvm_mmu_page *sp;
2003
2004         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2005                 return;
2006         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2007                 hpa_t root = vcpu->arch.mmu.root_hpa;
2008                 sp = page_header(root);
2009                 mmu_sync_children(vcpu, sp);
2010                 return;
2011         }
2012         for (i = 0; i < 4; ++i) {
2013                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2014
2015                 if (root) {
2016                         root &= PT64_BASE_ADDR_MASK;
2017                         sp = page_header(root);
2018                         mmu_sync_children(vcpu, sp);
2019                 }
2020         }
2021 }
2022
2023 static void mmu_sync_global(struct kvm_vcpu *vcpu)
2024 {
2025         struct kvm *kvm = vcpu->kvm;
2026         struct kvm_mmu_page *sp, *n;
2027
2028         list_for_each_entry_safe(sp, n, &kvm->arch.oos_global_pages, oos_link)
2029                 kvm_sync_page(vcpu, sp);
2030 }
2031
2032 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2033 {
2034         spin_lock(&vcpu->kvm->mmu_lock);
2035         mmu_sync_roots(vcpu);
2036         spin_unlock(&vcpu->kvm->mmu_lock);
2037 }
2038
2039 void kvm_mmu_sync_global(struct kvm_vcpu *vcpu)
2040 {
2041         spin_lock(&vcpu->kvm->mmu_lock);
2042         mmu_sync_global(vcpu);
2043         spin_unlock(&vcpu->kvm->mmu_lock);
2044 }
2045
2046 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2047 {
2048         return vaddr;
2049 }
2050
2051 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2052                                 u32 error_code)
2053 {
2054         gfn_t gfn;
2055         int r;
2056
2057         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2058         r = mmu_topup_memory_caches(vcpu);
2059         if (r)
2060                 return r;
2061
2062         ASSERT(vcpu);
2063         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2064
2065         gfn = gva >> PAGE_SHIFT;
2066
2067         return nonpaging_map(vcpu, gva & PAGE_MASK,
2068                              error_code & PFERR_WRITE_MASK, gfn);
2069 }
2070
2071 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2072                                 u32 error_code)
2073 {
2074         pfn_t pfn;
2075         int r;
2076         int largepage = 0;
2077         gfn_t gfn = gpa >> PAGE_SHIFT;
2078         unsigned long mmu_seq;
2079
2080         ASSERT(vcpu);
2081         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2082
2083         r = mmu_topup_memory_caches(vcpu);
2084         if (r)
2085                 return r;
2086
2087         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
2088                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2089                 largepage = 1;
2090         }
2091         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2092         smp_rmb();
2093         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2094         if (is_error_pfn(pfn)) {
2095                 kvm_release_pfn_clean(pfn);
2096                 return 1;
2097         }
2098         spin_lock(&vcpu->kvm->mmu_lock);
2099         if (mmu_notifier_retry(vcpu, mmu_seq))
2100                 goto out_unlock;
2101         kvm_mmu_free_some_pages(vcpu);
2102         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2103                          largepage, gfn, pfn);
2104         spin_unlock(&vcpu->kvm->mmu_lock);
2105
2106         return r;
2107
2108 out_unlock:
2109         spin_unlock(&vcpu->kvm->mmu_lock);
2110         kvm_release_pfn_clean(pfn);
2111         return 0;
2112 }
2113
2114 static void nonpaging_free(struct kvm_vcpu *vcpu)
2115 {
2116         mmu_free_roots(vcpu);
2117 }
2118
2119 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2120 {
2121         struct kvm_mmu *context = &vcpu->arch.mmu;
2122
2123         context->new_cr3 = nonpaging_new_cr3;
2124         context->page_fault = nonpaging_page_fault;
2125         context->gva_to_gpa = nonpaging_gva_to_gpa;
2126         context->free = nonpaging_free;
2127         context->prefetch_page = nonpaging_prefetch_page;
2128         context->sync_page = nonpaging_sync_page;
2129         context->invlpg = nonpaging_invlpg;
2130         context->root_level = 0;
2131         context->shadow_root_level = PT32E_ROOT_LEVEL;
2132         context->root_hpa = INVALID_PAGE;
2133         return 0;
2134 }
2135
2136 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2137 {
2138         ++vcpu->stat.tlb_flush;
2139         kvm_x86_ops->tlb_flush(vcpu);
2140 }
2141
2142 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2143 {
2144         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2145         mmu_free_roots(vcpu);
2146 }
2147
2148 static void inject_page_fault(struct kvm_vcpu *vcpu,
2149                               u64 addr,
2150                               u32 err_code)
2151 {
2152         kvm_inject_page_fault(vcpu, addr, err_code);
2153 }
2154
2155 static void paging_free(struct kvm_vcpu *vcpu)
2156 {
2157         nonpaging_free(vcpu);
2158 }
2159
2160 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2161 {
2162         int bit7;
2163
2164         bit7 = (gpte >> 7) & 1;
2165         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2166 }
2167
2168 #define PTTYPE 64
2169 #include "paging_tmpl.h"
2170 #undef PTTYPE
2171
2172 #define PTTYPE 32
2173 #include "paging_tmpl.h"
2174 #undef PTTYPE
2175
2176 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2177 {
2178         struct kvm_mmu *context = &vcpu->arch.mmu;
2179         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2180         u64 exb_bit_rsvd = 0;
2181
2182         if (!is_nx(vcpu))
2183                 exb_bit_rsvd = rsvd_bits(63, 63);
2184         switch (level) {
2185         case PT32_ROOT_LEVEL:
2186                 /* no rsvd bits for 2 level 4K page table entries */
2187                 context->rsvd_bits_mask[0][1] = 0;
2188                 context->rsvd_bits_mask[0][0] = 0;
2189                 if (is_cpuid_PSE36())
2190                         /* 36bits PSE 4MB page */
2191                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2192                 else
2193                         /* 32 bits PSE 4MB page */
2194                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2195                 context->rsvd_bits_mask[1][0] = ~0ull;
2196                 break;
2197         case PT32E_ROOT_LEVEL:
2198                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2199                         rsvd_bits(maxphyaddr, 62);              /* PDE */
2200                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2201                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2202                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2203                         rsvd_bits(maxphyaddr, 62) |
2204                         rsvd_bits(13, 20);              /* large page */
2205                 context->rsvd_bits_mask[1][0] = ~0ull;
2206                 break;
2207         case PT64_ROOT_LEVEL:
2208                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2209                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2210                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2211                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2212                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2213                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2214                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2215                         rsvd_bits(maxphyaddr, 51);
2216                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2217                 context->rsvd_bits_mask[1][2] = context->rsvd_bits_mask[0][2];
2218                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2219                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(13, 20);
2220                 context->rsvd_bits_mask[1][0] = ~0ull;
2221                 break;
2222         }
2223 }
2224
2225 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2226 {
2227         struct kvm_mmu *context = &vcpu->arch.mmu;
2228
2229         ASSERT(is_pae(vcpu));
2230         context->new_cr3 = paging_new_cr3;
2231         context->page_fault = paging64_page_fault;
2232         context->gva_to_gpa = paging64_gva_to_gpa;
2233         context->prefetch_page = paging64_prefetch_page;
2234         context->sync_page = paging64_sync_page;
2235         context->invlpg = paging64_invlpg;
2236         context->free = paging_free;
2237         context->root_level = level;
2238         context->shadow_root_level = level;
2239         context->root_hpa = INVALID_PAGE;
2240         return 0;
2241 }
2242
2243 static int paging64_init_context(struct kvm_vcpu *vcpu)
2244 {
2245         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2246         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2247 }
2248
2249 static int paging32_init_context(struct kvm_vcpu *vcpu)
2250 {
2251         struct kvm_mmu *context = &vcpu->arch.mmu;
2252
2253         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2254         context->new_cr3 = paging_new_cr3;
2255         context->page_fault = paging32_page_fault;
2256         context->gva_to_gpa = paging32_gva_to_gpa;
2257         context->free = paging_free;
2258         context->prefetch_page = paging32_prefetch_page;
2259         context->sync_page = paging32_sync_page;
2260         context->invlpg = paging32_invlpg;
2261         context->root_level = PT32_ROOT_LEVEL;
2262         context->shadow_root_level = PT32E_ROOT_LEVEL;
2263         context->root_hpa = INVALID_PAGE;
2264         return 0;
2265 }
2266
2267 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2268 {
2269         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2270         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2271 }
2272
2273 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2274 {
2275         struct kvm_mmu *context = &vcpu->arch.mmu;
2276
2277         context->new_cr3 = nonpaging_new_cr3;
2278         context->page_fault = tdp_page_fault;
2279         context->free = nonpaging_free;
2280         context->prefetch_page = nonpaging_prefetch_page;
2281         context->sync_page = nonpaging_sync_page;
2282         context->invlpg = nonpaging_invlpg;
2283         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2284         context->root_hpa = INVALID_PAGE;
2285
2286         if (!is_paging(vcpu)) {
2287                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2288                 context->root_level = 0;
2289         } else if (is_long_mode(vcpu)) {
2290                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2291                 context->gva_to_gpa = paging64_gva_to_gpa;
2292                 context->root_level = PT64_ROOT_LEVEL;
2293         } else if (is_pae(vcpu)) {
2294                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2295                 context->gva_to_gpa = paging64_gva_to_gpa;
2296                 context->root_level = PT32E_ROOT_LEVEL;
2297         } else {
2298                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2299                 context->gva_to_gpa = paging32_gva_to_gpa;
2300                 context->root_level = PT32_ROOT_LEVEL;
2301         }
2302
2303         return 0;
2304 }
2305
2306 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2307 {
2308         int r;
2309
2310         ASSERT(vcpu);
2311         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2312
2313         if (!is_paging(vcpu))
2314                 r = nonpaging_init_context(vcpu);
2315         else if (is_long_mode(vcpu))
2316                 r = paging64_init_context(vcpu);
2317         else if (is_pae(vcpu))
2318                 r = paging32E_init_context(vcpu);
2319         else
2320                 r = paging32_init_context(vcpu);
2321
2322         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2323
2324         return r;
2325 }
2326
2327 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2328 {
2329         vcpu->arch.update_pte.pfn = bad_pfn;
2330
2331         if (tdp_enabled)
2332                 return init_kvm_tdp_mmu(vcpu);
2333         else
2334                 return init_kvm_softmmu(vcpu);
2335 }
2336
2337 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2338 {
2339         ASSERT(vcpu);
2340         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2341                 vcpu->arch.mmu.free(vcpu);
2342                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2343         }
2344 }
2345
2346 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2347 {
2348         destroy_kvm_mmu(vcpu);
2349         return init_kvm_mmu(vcpu);
2350 }
2351 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2352
2353 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2354 {
2355         int r;
2356
2357         r = mmu_topup_memory_caches(vcpu);
2358         if (r)
2359                 goto out;
2360         spin_lock(&vcpu->kvm->mmu_lock);
2361         kvm_mmu_free_some_pages(vcpu);
2362         mmu_alloc_roots(vcpu);
2363         mmu_sync_roots(vcpu);
2364         spin_unlock(&vcpu->kvm->mmu_lock);
2365         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2366         kvm_mmu_flush_tlb(vcpu);
2367 out:
2368         return r;
2369 }
2370 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2371
2372 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2373 {
2374         mmu_free_roots(vcpu);
2375 }
2376
2377 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2378                                   struct kvm_mmu_page *sp,
2379                                   u64 *spte)
2380 {
2381         u64 pte;
2382         struct kvm_mmu_page *child;
2383
2384         pte = *spte;
2385         if (is_shadow_present_pte(pte)) {
2386                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2387                     is_large_pte(pte))
2388                         rmap_remove(vcpu->kvm, spte);
2389                 else {
2390                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2391                         mmu_page_remove_parent_pte(child, spte);
2392                 }
2393         }
2394         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2395         if (is_large_pte(pte))
2396                 --vcpu->kvm->stat.lpages;
2397 }
2398
2399 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2400                                   struct kvm_mmu_page *sp,
2401                                   u64 *spte,
2402                                   const void *new)
2403 {
2404         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2405                 if (!vcpu->arch.update_pte.largepage ||
2406                     sp->role.glevels == PT32_ROOT_LEVEL) {
2407                         ++vcpu->kvm->stat.mmu_pde_zapped;
2408                         return;
2409                 }
2410         }
2411
2412         ++vcpu->kvm->stat.mmu_pte_updated;
2413         if (sp->role.glevels == PT32_ROOT_LEVEL)
2414                 paging32_update_pte(vcpu, sp, spte, new);
2415         else
2416                 paging64_update_pte(vcpu, sp, spte, new);
2417 }
2418
2419 static bool need_remote_flush(u64 old, u64 new)
2420 {
2421         if (!is_shadow_present_pte(old))
2422                 return false;
2423         if (!is_shadow_present_pte(new))
2424                 return true;
2425         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2426                 return true;
2427         old ^= PT64_NX_MASK;
2428         new ^= PT64_NX_MASK;
2429         return (old & ~new & PT64_PERM_MASK) != 0;
2430 }
2431
2432 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2433 {
2434         if (need_remote_flush(old, new))
2435                 kvm_flush_remote_tlbs(vcpu->kvm);
2436         else
2437                 kvm_mmu_flush_tlb(vcpu);
2438 }
2439
2440 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2441 {
2442         u64 *spte = vcpu->arch.last_pte_updated;
2443
2444         return !!(spte && (*spte & shadow_accessed_mask));
2445 }
2446
2447 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2448                                           const u8 *new, int bytes)
2449 {
2450         gfn_t gfn;
2451         int r;
2452         u64 gpte = 0;
2453         pfn_t pfn;
2454
2455         vcpu->arch.update_pte.largepage = 0;
2456
2457         if (bytes != 4 && bytes != 8)
2458                 return;
2459
2460         /*
2461          * Assume that the pte write on a page table of the same type
2462          * as the current vcpu paging mode.  This is nearly always true
2463          * (might be false while changing modes).  Note it is verified later
2464          * by update_pte().
2465          */
2466         if (is_pae(vcpu)) {
2467                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2468                 if ((bytes == 4) && (gpa % 4 == 0)) {
2469                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2470                         if (r)
2471                                 return;
2472                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2473                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2474                         memcpy((void *)&gpte, new, 8);
2475                 }
2476         } else {
2477                 if ((bytes == 4) && (gpa % 4 == 0))
2478                         memcpy((void *)&gpte, new, 4);
2479         }
2480         if (!is_present_pte(gpte))
2481                 return;
2482         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2483
2484         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2485                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2486                 vcpu->arch.update_pte.largepage = 1;
2487         }
2488         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2489         smp_rmb();
2490         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2491
2492         if (is_error_pfn(pfn)) {
2493                 kvm_release_pfn_clean(pfn);
2494                 return;
2495         }
2496         vcpu->arch.update_pte.gfn = gfn;
2497         vcpu->arch.update_pte.pfn = pfn;
2498 }
2499
2500 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2501 {
2502         u64 *spte = vcpu->arch.last_pte_updated;
2503
2504         if (spte
2505             && vcpu->arch.last_pte_gfn == gfn
2506             && shadow_accessed_mask
2507             && !(*spte & shadow_accessed_mask)
2508             && is_shadow_present_pte(*spte))
2509                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2510 }
2511
2512 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2513                        const u8 *new, int bytes,
2514                        bool guest_initiated)
2515 {
2516         gfn_t gfn = gpa >> PAGE_SHIFT;
2517         struct kvm_mmu_page *sp;
2518         struct hlist_node *node, *n;
2519         struct hlist_head *bucket;
2520         unsigned index;
2521         u64 entry, gentry;
2522         u64 *spte;
2523         unsigned offset = offset_in_page(gpa);
2524         unsigned pte_size;
2525         unsigned page_offset;
2526         unsigned misaligned;
2527         unsigned quadrant;
2528         int level;
2529         int flooded = 0;
2530         int npte;
2531         int r;
2532
2533         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2534         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2535         spin_lock(&vcpu->kvm->mmu_lock);
2536         kvm_mmu_access_page(vcpu, gfn);
2537         kvm_mmu_free_some_pages(vcpu);
2538         ++vcpu->kvm->stat.mmu_pte_write;
2539         kvm_mmu_audit(vcpu, "pre pte write");
2540         if (guest_initiated) {
2541                 if (gfn == vcpu->arch.last_pt_write_gfn
2542                     && !last_updated_pte_accessed(vcpu)) {
2543                         ++vcpu->arch.last_pt_write_count;
2544                         if (vcpu->arch.last_pt_write_count >= 3)
2545                                 flooded = 1;
2546                 } else {
2547                         vcpu->arch.last_pt_write_gfn = gfn;
2548                         vcpu->arch.last_pt_write_count = 1;
2549                         vcpu->arch.last_pte_updated = NULL;
2550                 }
2551         }
2552         index = kvm_page_table_hashfn(gfn);
2553         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2554         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2555                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2556                         continue;
2557                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2558                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2559                 misaligned |= bytes < 4;
2560                 if (misaligned || flooded) {
2561                         /*
2562                          * Misaligned accesses are too much trouble to fix
2563                          * up; also, they usually indicate a page is not used
2564                          * as a page table.
2565                          *
2566                          * If we're seeing too many writes to a page,
2567                          * it may no longer be a page table, or we may be
2568                          * forking, in which case it is better to unmap the
2569                          * page.
2570                          */
2571                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2572                                  gpa, bytes, sp->role.word);
2573                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2574                                 n = bucket->first;
2575                         ++vcpu->kvm->stat.mmu_flooded;
2576                         continue;
2577                 }
2578                 page_offset = offset;
2579                 level = sp->role.level;
2580                 npte = 1;
2581                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2582                         page_offset <<= 1;      /* 32->64 */
2583                         /*
2584                          * A 32-bit pde maps 4MB while the shadow pdes map
2585                          * only 2MB.  So we need to double the offset again
2586                          * and zap two pdes instead of one.
2587                          */
2588                         if (level == PT32_ROOT_LEVEL) {
2589                                 page_offset &= ~7; /* kill rounding error */
2590                                 page_offset <<= 1;
2591                                 npte = 2;
2592                         }
2593                         quadrant = page_offset >> PAGE_SHIFT;
2594                         page_offset &= ~PAGE_MASK;
2595                         if (quadrant != sp->role.quadrant)
2596                                 continue;
2597                 }
2598                 spte = &sp->spt[page_offset / sizeof(*spte)];
2599                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2600                         gentry = 0;
2601                         r = kvm_read_guest_atomic(vcpu->kvm,
2602                                                   gpa & ~(u64)(pte_size - 1),
2603                                                   &gentry, pte_size);
2604                         new = (const void *)&gentry;
2605                         if (r < 0)
2606                                 new = NULL;
2607                 }
2608                 while (npte--) {
2609                         entry = *spte;
2610                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2611                         if (new)
2612                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2613                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2614                         ++spte;
2615                 }
2616         }
2617         kvm_mmu_audit(vcpu, "post pte write");
2618         spin_unlock(&vcpu->kvm->mmu_lock);
2619         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2620                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2621                 vcpu->arch.update_pte.pfn = bad_pfn;
2622         }
2623 }
2624
2625 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2626 {
2627         gpa_t gpa;
2628         int r;
2629
2630         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2631
2632         spin_lock(&vcpu->kvm->mmu_lock);
2633         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2634         spin_unlock(&vcpu->kvm->mmu_lock);
2635         return r;
2636 }
2637 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2638
2639 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2640 {
2641         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2642                 struct kvm_mmu_page *sp;
2643
2644                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2645                                   struct kvm_mmu_page, link);
2646                 kvm_mmu_zap_page(vcpu->kvm, sp);
2647                 ++vcpu->kvm->stat.mmu_recycled;
2648         }
2649 }
2650
2651 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2652 {
2653         int r;
2654         enum emulation_result er;
2655
2656         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2657         if (r < 0)
2658                 goto out;
2659
2660         if (!r) {
2661                 r = 1;
2662                 goto out;
2663         }
2664
2665         r = mmu_topup_memory_caches(vcpu);
2666         if (r)
2667                 goto out;
2668
2669         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2670
2671         switch (er) {
2672         case EMULATE_DONE:
2673                 return 1;
2674         case EMULATE_DO_MMIO:
2675                 ++vcpu->stat.mmio_exits;
2676                 return 0;
2677         case EMULATE_FAIL:
2678                 kvm_report_emulation_failure(vcpu, "pagetable");
2679                 return 1;
2680         default:
2681                 BUG();
2682         }
2683 out:
2684         return r;
2685 }
2686 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2687
2688 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2689 {
2690         vcpu->arch.mmu.invlpg(vcpu, gva);
2691         kvm_mmu_flush_tlb(vcpu);
2692         ++vcpu->stat.invlpg;
2693 }
2694 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2695
2696 void kvm_enable_tdp(void)
2697 {
2698         tdp_enabled = true;
2699 }
2700 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2701
2702 void kvm_disable_tdp(void)
2703 {
2704         tdp_enabled = false;
2705 }
2706 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2707
2708 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2709 {
2710         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2711 }
2712
2713 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2714 {
2715         struct page *page;
2716         int i;
2717
2718         ASSERT(vcpu);
2719
2720         if (vcpu->kvm->arch.n_requested_mmu_pages)
2721                 vcpu->kvm->arch.n_free_mmu_pages =
2722                                         vcpu->kvm->arch.n_requested_mmu_pages;
2723         else
2724                 vcpu->kvm->arch.n_free_mmu_pages =
2725                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2726         /*
2727          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2728          * Therefore we need to allocate shadow page tables in the first
2729          * 4GB of memory, which happens to fit the DMA32 zone.
2730          */
2731         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2732         if (!page)
2733                 goto error_1;
2734         vcpu->arch.mmu.pae_root = page_address(page);
2735         for (i = 0; i < 4; ++i)
2736                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2737
2738         return 0;
2739
2740 error_1:
2741         free_mmu_pages(vcpu);
2742         return -ENOMEM;
2743 }
2744
2745 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2746 {
2747         ASSERT(vcpu);
2748         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2749
2750         return alloc_mmu_pages(vcpu);
2751 }
2752
2753 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2754 {
2755         ASSERT(vcpu);
2756         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2757
2758         return init_kvm_mmu(vcpu);
2759 }
2760
2761 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2762 {
2763         ASSERT(vcpu);
2764
2765         destroy_kvm_mmu(vcpu);
2766         free_mmu_pages(vcpu);
2767         mmu_free_memory_caches(vcpu);
2768 }
2769
2770 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2771 {
2772         struct kvm_mmu_page *sp;
2773
2774         spin_lock(&kvm->mmu_lock);
2775         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2776                 int i;
2777                 u64 *pt;
2778
2779                 if (!test_bit(slot, sp->slot_bitmap))
2780                         continue;
2781
2782                 pt = sp->spt;
2783                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2784                         /* avoid RMW */
2785                         if (pt[i] & PT_WRITABLE_MASK)
2786                                 pt[i] &= ~PT_WRITABLE_MASK;
2787         }
2788         kvm_flush_remote_tlbs(kvm);
2789         spin_unlock(&kvm->mmu_lock);
2790 }
2791
2792 void kvm_mmu_zap_all(struct kvm *kvm)
2793 {
2794         struct kvm_mmu_page *sp, *node;
2795
2796         spin_lock(&kvm->mmu_lock);
2797         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2798                 if (kvm_mmu_zap_page(kvm, sp))
2799                         node = container_of(kvm->arch.active_mmu_pages.next,
2800                                             struct kvm_mmu_page, link);
2801         spin_unlock(&kvm->mmu_lock);
2802
2803         kvm_flush_remote_tlbs(kvm);
2804 }
2805
2806 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2807 {
2808         struct kvm_mmu_page *page;
2809
2810         page = container_of(kvm->arch.active_mmu_pages.prev,
2811                             struct kvm_mmu_page, link);
2812         kvm_mmu_zap_page(kvm, page);
2813 }
2814
2815 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2816 {
2817         struct kvm *kvm;
2818         struct kvm *kvm_freed = NULL;
2819         int cache_count = 0;
2820
2821         spin_lock(&kvm_lock);
2822
2823         list_for_each_entry(kvm, &vm_list, vm_list) {
2824                 int npages;
2825
2826                 if (!down_read_trylock(&kvm->slots_lock))
2827                         continue;
2828                 spin_lock(&kvm->mmu_lock);
2829                 npages = kvm->arch.n_alloc_mmu_pages -
2830                          kvm->arch.n_free_mmu_pages;
2831                 cache_count += npages;
2832                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2833                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2834                         cache_count--;
2835                         kvm_freed = kvm;
2836                 }
2837                 nr_to_scan--;
2838
2839                 spin_unlock(&kvm->mmu_lock);
2840                 up_read(&kvm->slots_lock);
2841         }
2842         if (kvm_freed)
2843                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2844
2845         spin_unlock(&kvm_lock);
2846
2847         return cache_count;
2848 }
2849
2850 static struct shrinker mmu_shrinker = {
2851         .shrink = mmu_shrink,
2852         .seeks = DEFAULT_SEEKS * 10,
2853 };
2854
2855 static void mmu_destroy_caches(void)
2856 {
2857         if (pte_chain_cache)
2858                 kmem_cache_destroy(pte_chain_cache);
2859         if (rmap_desc_cache)
2860                 kmem_cache_destroy(rmap_desc_cache);
2861         if (mmu_page_header_cache)
2862                 kmem_cache_destroy(mmu_page_header_cache);
2863 }
2864
2865 void kvm_mmu_module_exit(void)
2866 {
2867         mmu_destroy_caches();
2868         unregister_shrinker(&mmu_shrinker);
2869 }
2870
2871 int kvm_mmu_module_init(void)
2872 {
2873         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2874                                             sizeof(struct kvm_pte_chain),
2875                                             0, 0, NULL);
2876         if (!pte_chain_cache)
2877                 goto nomem;
2878         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2879                                             sizeof(struct kvm_rmap_desc),
2880                                             0, 0, NULL);
2881         if (!rmap_desc_cache)
2882                 goto nomem;
2883
2884         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2885                                                   sizeof(struct kvm_mmu_page),
2886                                                   0, 0, NULL);
2887         if (!mmu_page_header_cache)
2888                 goto nomem;
2889
2890         register_shrinker(&mmu_shrinker);
2891
2892         return 0;
2893
2894 nomem:
2895         mmu_destroy_caches();
2896         return -ENOMEM;
2897 }
2898
2899 /*
2900  * Caculate mmu pages needed for kvm.
2901  */
2902 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2903 {
2904         int i;
2905         unsigned int nr_mmu_pages;
2906         unsigned int  nr_pages = 0;
2907
2908         for (i = 0; i < kvm->nmemslots; i++)
2909                 nr_pages += kvm->memslots[i].npages;
2910
2911         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2912         nr_mmu_pages = max(nr_mmu_pages,
2913                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2914
2915         return nr_mmu_pages;
2916 }
2917
2918 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2919                                 unsigned len)
2920 {
2921         if (len > buffer->len)
2922                 return NULL;
2923         return buffer->ptr;
2924 }
2925
2926 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2927                                 unsigned len)
2928 {
2929         void *ret;
2930
2931         ret = pv_mmu_peek_buffer(buffer, len);
2932         if (!ret)
2933                 return ret;
2934         buffer->ptr += len;
2935         buffer->len -= len;
2936         buffer->processed += len;
2937         return ret;
2938 }
2939
2940 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2941                              gpa_t addr, gpa_t value)
2942 {
2943         int bytes = 8;
2944         int r;
2945
2946         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2947                 bytes = 4;
2948
2949         r = mmu_topup_memory_caches(vcpu);
2950         if (r)
2951                 return r;
2952
2953         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2954                 return -EFAULT;
2955
2956         return 1;
2957 }
2958
2959 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2960 {
2961         kvm_set_cr3(vcpu, vcpu->arch.cr3);
2962         return 1;
2963 }
2964
2965 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2966 {
2967         spin_lock(&vcpu->kvm->mmu_lock);
2968         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2969         spin_unlock(&vcpu->kvm->mmu_lock);
2970         return 1;
2971 }
2972
2973 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2974                              struct kvm_pv_mmu_op_buffer *buffer)
2975 {
2976         struct kvm_mmu_op_header *header;
2977
2978         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2979         if (!header)
2980                 return 0;
2981         switch (header->op) {
2982         case KVM_MMU_OP_WRITE_PTE: {
2983                 struct kvm_mmu_op_write_pte *wpte;
2984
2985                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2986                 if (!wpte)
2987                         return 0;
2988                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2989                                         wpte->pte_val);
2990         }
2991         case KVM_MMU_OP_FLUSH_TLB: {
2992                 struct kvm_mmu_op_flush_tlb *ftlb;
2993
2994                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2995                 if (!ftlb)
2996                         return 0;
2997                 return kvm_pv_mmu_flush_tlb(vcpu);
2998         }
2999         case KVM_MMU_OP_RELEASE_PT: {
3000                 struct kvm_mmu_op_release_pt *rpt;
3001
3002                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3003                 if (!rpt)
3004                         return 0;
3005                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3006         }
3007         default: return 0;
3008         }
3009 }
3010
3011 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3012                   gpa_t addr, unsigned long *ret)
3013 {
3014         int r;
3015         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3016
3017         buffer->ptr = buffer->buf;
3018         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3019         buffer->processed = 0;
3020
3021         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3022         if (r)
3023                 goto out;
3024
3025         while (buffer->len) {
3026                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3027                 if (r < 0)
3028                         goto out;
3029                 if (r == 0)
3030                         break;
3031         }
3032
3033         r = 1;
3034 out:
3035         *ret = buffer->processed;
3036         return r;
3037 }
3038
3039 #ifdef AUDIT
3040
3041 static const char *audit_msg;
3042
3043 static gva_t canonicalize(gva_t gva)
3044 {
3045 #ifdef CONFIG_X86_64
3046         gva = (long long)(gva << 16) >> 16;
3047 #endif
3048         return gva;
3049 }
3050
3051 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3052                                 gva_t va, int level)
3053 {
3054         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3055         int i;
3056         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3057
3058         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3059                 u64 ent = pt[i];
3060
3061                 if (ent == shadow_trap_nonpresent_pte)
3062                         continue;
3063
3064                 va = canonicalize(va);
3065                 if (level > 1) {
3066                         if (ent == shadow_notrap_nonpresent_pte)
3067                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
3068                                        " in nonleaf level: levels %d gva %lx"
3069                                        " level %d pte %llx\n", audit_msg,
3070                                        vcpu->arch.mmu.root_level, va, level, ent);
3071
3072                         audit_mappings_page(vcpu, ent, va, level - 1);
3073                 } else {
3074                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3075                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
3076
3077                         if (is_shadow_present_pte(ent)
3078                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3079                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3080                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3081                                        audit_msg, vcpu->arch.mmu.root_level,
3082                                        va, gpa, hpa, ent,
3083                                        is_shadow_present_pte(ent));
3084                         else if (ent == shadow_notrap_nonpresent_pte
3085                                  && !is_error_hpa(hpa))
3086                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3087                                        " valid guest gva %lx\n", audit_msg, va);
3088                         kvm_release_pfn_clean(pfn);
3089
3090                 }
3091         }
3092 }
3093
3094 static void audit_mappings(struct kvm_vcpu *vcpu)
3095 {
3096         unsigned i;
3097
3098         if (vcpu->arch.mmu.root_level == 4)
3099                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3100         else
3101                 for (i = 0; i < 4; ++i)
3102                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3103                                 audit_mappings_page(vcpu,
3104                                                     vcpu->arch.mmu.pae_root[i],
3105                                                     i << 30,
3106                                                     2);
3107 }
3108
3109 static int count_rmaps(struct kvm_vcpu *vcpu)
3110 {
3111         int nmaps = 0;
3112         int i, j, k;
3113
3114         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3115                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3116                 struct kvm_rmap_desc *d;
3117
3118                 for (j = 0; j < m->npages; ++j) {
3119                         unsigned long *rmapp = &m->rmap[j];
3120
3121                         if (!*rmapp)
3122                                 continue;
3123                         if (!(*rmapp & 1)) {
3124                                 ++nmaps;
3125                                 continue;
3126                         }
3127                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3128                         while (d) {
3129                                 for (k = 0; k < RMAP_EXT; ++k)
3130                                         if (d->shadow_ptes[k])
3131                                                 ++nmaps;
3132                                         else
3133                                                 break;
3134                                 d = d->more;
3135                         }
3136                 }
3137         }
3138         return nmaps;
3139 }
3140
3141 static int count_writable_mappings(struct kvm_vcpu *vcpu)
3142 {
3143         int nmaps = 0;
3144         struct kvm_mmu_page *sp;
3145         int i;
3146
3147         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3148                 u64 *pt = sp->spt;
3149
3150                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3151                         continue;
3152
3153                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3154                         u64 ent = pt[i];
3155
3156                         if (!(ent & PT_PRESENT_MASK))
3157                                 continue;
3158                         if (!(ent & PT_WRITABLE_MASK))
3159                                 continue;
3160                         ++nmaps;
3161                 }
3162         }
3163         return nmaps;
3164 }
3165
3166 static void audit_rmap(struct kvm_vcpu *vcpu)
3167 {
3168         int n_rmap = count_rmaps(vcpu);
3169         int n_actual = count_writable_mappings(vcpu);
3170
3171         if (n_rmap != n_actual)
3172                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
3173                        __func__, audit_msg, n_rmap, n_actual);
3174 }
3175
3176 static void audit_write_protection(struct kvm_vcpu *vcpu)
3177 {
3178         struct kvm_mmu_page *sp;
3179         struct kvm_memory_slot *slot;
3180         unsigned long *rmapp;
3181         gfn_t gfn;
3182
3183         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3184                 if (sp->role.direct)
3185                         continue;
3186
3187                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3188                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3189                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3190                 if (*rmapp)
3191                         printk(KERN_ERR "%s: (%s) shadow page has writable"
3192                                " mappings: gfn %lx role %x\n",
3193                                __func__, audit_msg, sp->gfn,
3194                                sp->role.word);
3195         }
3196 }
3197
3198 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3199 {
3200         int olddbg = dbg;
3201
3202         dbg = 0;
3203         audit_msg = msg;
3204         audit_rmap(vcpu);
3205         audit_write_protection(vcpu);
3206         audit_mappings(vcpu);
3207         dbg = olddbg;
3208 }
3209
3210 #endif