err.no Git - 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 "vmx.h"
  21 #include "mmu.h"
  22
  23 #include <linux/kvm_host.h>
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29 #include <linux/swap.h>
  30
  31 #include <asm/page.h>
  32 #include <asm/cmpxchg.h>
  33 #include <asm/io.h>
  34
  35 /*
  36  * When setting this variable to true it enables Two-Dimensional-Paging
  37  * where the hardware walks 2 page tables:
  38  * 1. the guest-virtual to guest-physical
  39  * 2. while doing 1. it walks guest-physical to host-physical
  40  * If the hardware supports that we don't need to do shadow paging.
  41  */
  42 static bool tdp_enabled = false;
  43
  44 #undef MMU_DEBUG
  45
  46 #undef AUDIT
  47
  48 #ifdef AUDIT
  49 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  50 #else
  51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  52 #endif
  53
  54 #ifdef MMU_DEBUG
  55
  56 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  57 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  58
  59 #else
  60
  61 #define pgprintk(x...) do { } while (0)
  62 #define rmap_printk(x...) do { } while (0)
  63
  64 #endif
  65
  66 #if defined(MMU_DEBUG) || defined(AUDIT)
  67 static int dbg = 1;
  68 #endif
  69
  70 #ifndef MMU_DEBUG
  71 #define ASSERT(x) do { } while (0)
  72 #else
  73 #define ASSERT(x)                                                       \
  74         if (!(x)) {                                                     \
  75                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  76                        __FILE__, __LINE__, #x);                         \
  77         }
  78 #endif
  79
  80 #define PT64_PT_BITS 9
  81 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  82 #define PT32_PT_BITS 10
  83 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  84
  85 #define PT_WRITABLE_SHIFT 1
  86
  87 #define PT_PRESENT_MASK (1ULL << 0)
  88 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  89 #define PT_USER_MASK (1ULL << 2)
  90 #define PT_PWT_MASK (1ULL << 3)
  91 #define PT_PCD_MASK (1ULL << 4)
  92 #define PT_ACCESSED_MASK (1ULL << 5)
  93 #define PT_DIRTY_MASK (1ULL << 6)
  94 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  95 #define PT_PAT_MASK (1ULL << 7)
  96 #define PT_GLOBAL_MASK (1ULL << 8)
  97 #define PT64_NX_SHIFT 63
  98 #define PT64_NX_MASK (1ULL << PT64_NX_SHIFT)
  99
 100 #define PT_PAT_SHIFT 7
 101 #define PT_DIR_PAT_SHIFT 12
 102 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
 103
 104 #define PT32_DIR_PSE36_SIZE 4
 105 #define PT32_DIR_PSE36_SHIFT 13
 106 #define PT32_DIR_PSE36_MASK \
 107         (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
 108
 109
 110 #define PT_FIRST_AVAIL_BITS_SHIFT 9
 111 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
 112
 113 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 114
 115 #define PT64_LEVEL_BITS 9
 116
 117 #define PT64_LEVEL_SHIFT(level) \
 118                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
 119
 120 #define PT64_LEVEL_MASK(level) \
 121                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 122
 123 #define PT64_INDEX(address, level)\
 124         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 125
 126
 127 #define PT32_LEVEL_BITS 10
 128
 129 #define PT32_LEVEL_SHIFT(level) \
 130                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 131
 132 #define PT32_LEVEL_MASK(level) \
 133                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 134
 135 #define PT32_INDEX(address, level)\
 136         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 137
 138
 139 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 140 #define PT64_DIR_BASE_ADDR_MASK \
 141         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 142
 143 #define PT32_BASE_ADDR_MASK PAGE_MASK
 144 #define PT32_DIR_BASE_ADDR_MASK \
 145         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 146
 147 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 148                         | PT64_NX_MASK)
 149
 150 #define PFERR_PRESENT_MASK (1U << 0)
 151 #define PFERR_WRITE_MASK (1U << 1)
 152 #define PFERR_USER_MASK (1U << 2)
 153 #define PFERR_FETCH_MASK (1U << 4)
 154
 155 #define PT64_ROOT_LEVEL 4
 156 #define PT32_ROOT_LEVEL 2
 157 #define PT32E_ROOT_LEVEL 3
 158
 159 #define PT_DIRECTORY_LEVEL 2
 160 #define PT_PAGE_TABLE_LEVEL 1
 161
 162 #define RMAP_EXT 4
 163
 164 #define ACC_EXEC_MASK    1
 165 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 166 #define ACC_USER_MASK    PT_USER_MASK
 167 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 168
 169 struct kvm_rmap_desc {
 170         u64 *shadow_ptes[RMAP_EXT];
 171         struct kvm_rmap_desc *more;
 172 };
 173
 174 static struct kmem_cache *pte_chain_cache;
 175 static struct kmem_cache *rmap_desc_cache;
 176 static struct kmem_cache *mmu_page_header_cache;
 177
 178 static u64 __read_mostly shadow_trap_nonpresent_pte;
 179 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 180
 181 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 182 {
 183         shadow_trap_nonpresent_pte = trap_pte;
 184         shadow_notrap_nonpresent_pte = notrap_pte;
 185 }
 186 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 187
 188 static int is_write_protection(struct kvm_vcpu *vcpu)
 189 {
 190         return vcpu->arch.cr0 & X86_CR0_WP;
 191 }
 192
 193 static int is_cpuid_PSE36(void)
 194 {
 195         return 1;
 196 }
 197
 198 static int is_nx(struct kvm_vcpu *vcpu)
 199 {
 200         return vcpu->arch.shadow_efer & EFER_NX;
 201 }
 202
 203 static int is_present_pte(unsigned long pte)
 204 {
 205         return pte & PT_PRESENT_MASK;
 206 }
 207
 208 static int is_shadow_present_pte(u64 pte)
 209 {
 210         return pte != shadow_trap_nonpresent_pte
 211                 && pte != shadow_notrap_nonpresent_pte;
 212 }
 213
 214 static int is_writeble_pte(unsigned long pte)
 215 {
 216         return pte & PT_WRITABLE_MASK;
 217 }
 218
 219 static int is_dirty_pte(unsigned long pte)
 220 {
 221         return pte & PT_DIRTY_MASK;
 222 }
 223
 224 static int is_rmap_pte(u64 pte)
 225 {
 226         return is_shadow_present_pte(pte);
 227 }
 228
 229 static gfn_t pse36_gfn_delta(u32 gpte)
 230 {
 231         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 232
 233         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 234 }
 235
 236 static void set_shadow_pte(u64 *sptep, u64 spte)
 237 {
 238 #ifdef CONFIG_X86_64
 239         set_64bit((unsigned long *)sptep, spte);
 240 #else
 241         set_64bit((unsigned long long *)sptep, spte);
 242 #endif
 243 }
 244
 245 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 246                                   struct kmem_cache *base_cache, int min)
 247 {
 248         void *obj;
 249
 250         if (cache->nobjs >= min)
 251                 return 0;
 252         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 253                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 254                 if (!obj)
 255                         return -ENOMEM;
 256                 cache->objects[cache->nobjs++] = obj;
 257         }
 258         return 0;
 259 }
 260
 261 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 262 {
 263         while (mc->nobjs)
 264                 kfree(mc->objects[--mc->nobjs]);
 265 }
 266
 267 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 268                                        int min)
 269 {
 270         struct page *page;
 271
 272         if (cache->nobjs >= min)
 273                 return 0;
 274         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 275                 page = alloc_page(GFP_KERNEL);
 276                 if (!page)
 277                         return -ENOMEM;
 278                 set_page_private(page, 0);
 279                 cache->objects[cache->nobjs++] = page_address(page);
 280         }
 281         return 0;
 282 }
 283
 284 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 285 {
 286         while (mc->nobjs)
 287                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 288 }
 289
 290 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 291 {
 292         int r;
 293
 294         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 295                                    pte_chain_cache, 4);
 296         if (r)
 297                 goto out;
 298         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 299                                    rmap_desc_cache, 1);
 300         if (r)
 301                 goto out;
 302         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 303         if (r)
 304                 goto out;
 305         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 306                                    mmu_page_header_cache, 4);
 307 out:
 308         return r;
 309 }
 310
 311 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 312 {
 313         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 314         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 315         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 316         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 317 }
 318
 319 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 320                                     size_t size)
 321 {
 322         void *p;
 323
 324         BUG_ON(!mc->nobjs);
 325         p = mc->objects[--mc->nobjs];
 326         memset(p, 0, size);
 327         return p;
 328 }
 329
 330 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 331 {
 332         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 333                                       sizeof(struct kvm_pte_chain));
 334 }
 335
 336 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 337 {
 338         kfree(pc);
 339 }
 340
 341 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 342 {
 343         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 344                                       sizeof(struct kvm_rmap_desc));
 345 }
 346
 347 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 348 {
 349         kfree(rd);
 350 }
 351
 352 /*
 353  * Take gfn and return the reverse mapping to it.
 354  * Note: gfn must be unaliased before this function get called
 355  */
 356
 357 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
 358 {
 359         struct kvm_memory_slot *slot;
 360
 361         slot = gfn_to_memslot(kvm, gfn);
 362         return &slot->rmap[gfn - slot->base_gfn];
 363 }
 364
 365 /*
 366  * Reverse mapping data structures:
 367  *
 368  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 369  * that points to page_address(page).
 370  *
 371  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 372  * containing more mappings.
 373  */
 374 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 375 {
 376         struct kvm_mmu_page *sp;
 377         struct kvm_rmap_desc *desc;
 378         unsigned long *rmapp;
 379         int i;
 380
 381         if (!is_rmap_pte(*spte))
 382                 return;
 383         gfn = unalias_gfn(vcpu->kvm, gfn);
 384         sp = page_header(__pa(spte));
 385         sp->gfns[spte - sp->spt] = gfn;
 386         rmapp = gfn_to_rmap(vcpu->kvm, gfn);
 387         if (!*rmapp) {
 388                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 389                 *rmapp = (unsigned long)spte;
 390         } else if (!(*rmapp & 1)) {
 391                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 392                 desc = mmu_alloc_rmap_desc(vcpu);
 393                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 394                 desc->shadow_ptes[1] = spte;
 395                 *rmapp = (unsigned long)desc | 1;
 396         } else {
 397                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 398                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 399                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 400                         desc = desc->more;
 401                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 402                         desc->more = mmu_alloc_rmap_desc(vcpu);
 403                         desc = desc->more;
 404                 }
 405                 for (i = 0; desc->shadow_ptes[i]; ++i)
 406                         ;
 407                 desc->shadow_ptes[i] = spte;
 408         }
 409 }
 410
 411 static void rmap_desc_remove_entry(unsigned long *rmapp,
 412                                    struct kvm_rmap_desc *desc,
 413                                    int i,
 414                                    struct kvm_rmap_desc *prev_desc)
 415 {
 416         int j;
 417
 418         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 419                 ;
 420         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 421         desc->shadow_ptes[j] = NULL;
 422         if (j != 0)
 423                 return;
 424         if (!prev_desc && !desc->more)
 425                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 426         else
 427                 if (prev_desc)
 428                         prev_desc->more = desc->more;
 429                 else
 430                         *rmapp = (unsigned long)desc->more | 1;
 431         mmu_free_rmap_desc(desc);
 432 }
 433
 434 static void rmap_remove(struct kvm *kvm, u64 *spte)
 435 {
 436         struct kvm_rmap_desc *desc;
 437         struct kvm_rmap_desc *prev_desc;
 438         struct kvm_mmu_page *sp;
 439         struct page *page;
 440         unsigned long *rmapp;
 441         int i;
 442
 443         if (!is_rmap_pte(*spte))
 444                 return;
 445         sp = page_header(__pa(spte));
 446         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 447         mark_page_accessed(page);
 448         if (is_writeble_pte(*spte))
 449                 kvm_release_page_dirty(page);
 450         else
 451                 kvm_release_page_clean(page);
 452         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt]);
 453         if (!*rmapp) {
 454                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 455                 BUG();
 456         } else if (!(*rmapp & 1)) {
 457                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 458                 if ((u64 *)*rmapp != spte) {
 459                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 460                                spte, *spte);
 461                         BUG();
 462                 }
 463                 *rmapp = 0;
 464         } else {
 465                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 466                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 467                 prev_desc = NULL;
 468                 while (desc) {
 469                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 470                                 if (desc->shadow_ptes[i] == spte) {
 471                                         rmap_desc_remove_entry(rmapp,
 472                                                                desc, i,
 473                                                                prev_desc);
 474                                         return;
 475                                 }
 476                         prev_desc = desc;
 477                         desc = desc->more;
 478                 }
 479                 BUG();
 480         }
 481 }
 482
 483 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 484 {
 485         struct kvm_rmap_desc *desc;
 486         struct kvm_rmap_desc *prev_desc;
 487         u64 *prev_spte;
 488         int i;
 489
 490         if (!*rmapp)
 491                 return NULL;
 492         else if (!(*rmapp & 1)) {
 493                 if (!spte)
 494                         return (u64 *)*rmapp;
 495                 return NULL;
 496         }
 497         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 498         prev_desc = NULL;
 499         prev_spte = NULL;
 500         while (desc) {
 501                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 502                         if (prev_spte == spte)
 503                                 return desc->shadow_ptes[i];
 504                         prev_spte = desc->shadow_ptes[i];
 505                 }
 506                 desc = desc->more;
 507         }
 508         return NULL;
 509 }
 510
 511 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 512 {
 513         unsigned long *rmapp;
 514         u64 *spte;
 515         int write_protected = 0;
 516
 517         gfn = unalias_gfn(kvm, gfn);
 518         rmapp = gfn_to_rmap(kvm, gfn);
 519
 520         spte = rmap_next(kvm, rmapp, NULL);
 521         while (spte) {
 522                 BUG_ON(!spte);
 523                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 524                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 525                 if (is_writeble_pte(*spte)) {
 526                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 527                         write_protected = 1;
 528                 }
 529                 spte = rmap_next(kvm, rmapp, spte);
 530         }
 531         if (write_protected)
 532                 kvm_flush_remote_tlbs(kvm);
 533 }
 534
 535 #ifdef MMU_DEBUG
 536 static int is_empty_shadow_page(u64 *spt)
 537 {
 538         u64 *pos;
 539         u64 *end;
 540
 541         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 542                 if (*pos != shadow_trap_nonpresent_pte) {
 543                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 544                                pos, *pos);
 545                         return 0;
 546                 }
 547         return 1;
 548 }
 549 #endif
 550
 551 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 552 {
 553         ASSERT(is_empty_shadow_page(sp->spt));
 554         list_del(&sp->link);
 555         __free_page(virt_to_page(sp->spt));
 556         __free_page(virt_to_page(sp->gfns));
 557         kfree(sp);
 558         ++kvm->arch.n_free_mmu_pages;
 559 }
 560
 561 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 562 {
 563         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 564 }
 565
 566 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 567                                                u64 *parent_pte)
 568 {
 569         struct kvm_mmu_page *sp;
 570
 571         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 572         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 573         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 574         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 575         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 576         ASSERT(is_empty_shadow_page(sp->spt));
 577         sp->slot_bitmap = 0;
 578         sp->multimapped = 0;
 579         sp->parent_pte = parent_pte;
 580         --vcpu->kvm->arch.n_free_mmu_pages;
 581         return sp;
 582 }
 583
 584 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 585                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 586 {
 587         struct kvm_pte_chain *pte_chain;
 588         struct hlist_node *node;
 589         int i;
 590
 591         if (!parent_pte)
 592                 return;
 593         if (!sp->multimapped) {
 594                 u64 *old = sp->parent_pte;
 595
 596                 if (!old) {
 597                         sp->parent_pte = parent_pte;
 598                         return;
 599                 }
 600                 sp->multimapped = 1;
 601                 pte_chain = mmu_alloc_pte_chain(vcpu);
 602                 INIT_HLIST_HEAD(&sp->parent_ptes);
 603                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 604                 pte_chain->parent_ptes[0] = old;
 605         }
 606         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 607                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 608                         continue;
 609                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 610                         if (!pte_chain->parent_ptes[i]) {
 611                                 pte_chain->parent_ptes[i] = parent_pte;
 612                                 return;
 613                         }
 614         }
 615         pte_chain = mmu_alloc_pte_chain(vcpu);
 616         BUG_ON(!pte_chain);
 617         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 618         pte_chain->parent_ptes[0] = parent_pte;
 619 }
 620
 621 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 622                                        u64 *parent_pte)
 623 {
 624         struct kvm_pte_chain *pte_chain;
 625         struct hlist_node *node;
 626         int i;
 627
 628         if (!sp->multimapped) {
 629                 BUG_ON(sp->parent_pte != parent_pte);
 630                 sp->parent_pte = NULL;
 631                 return;
 632         }
 633         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 634                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 635                         if (!pte_chain->parent_ptes[i])
 636                                 break;
 637                         if (pte_chain->parent_ptes[i] != parent_pte)
 638                                 continue;
 639                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 640                                 && pte_chain->parent_ptes[i + 1]) {
 641                                 pte_chain->parent_ptes[i]
 642                                         = pte_chain->parent_ptes[i + 1];
 643                                 ++i;
 644                         }
 645                         pte_chain->parent_ptes[i] = NULL;
 646                         if (i == 0) {
 647                                 hlist_del(&pte_chain->link);
 648                                 mmu_free_pte_chain(pte_chain);
 649                                 if (hlist_empty(&sp->parent_ptes)) {
 650                                         sp->multimapped = 0;
 651                                         sp->parent_pte = NULL;
 652                                 }
 653                         }
 654                         return;
 655                 }
 656         BUG();
 657 }
 658
 659 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
 660 {
 661         unsigned index;
 662         struct hlist_head *bucket;
 663         struct kvm_mmu_page *sp;
 664         struct hlist_node *node;
 665
 666         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 667         index = kvm_page_table_hashfn(gfn);
 668         bucket = &kvm->arch.mmu_page_hash[index];
 669         hlist_for_each_entry(sp, node, bucket, hash_link)
 670                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 671                         pgprintk("%s: found role %x\n",
 672                                  __FUNCTION__, sp->role.word);
 673                         return sp;
 674                 }
 675         return NULL;
 676 }
 677
 678 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 679                                              gfn_t gfn,
 680                                              gva_t gaddr,
 681                                              unsigned level,
 682                                              int metaphysical,
 683                                              unsigned access,
 684                                              u64 *parent_pte)
 685 {
 686         union kvm_mmu_page_role role;
 687         unsigned index;
 688         unsigned quadrant;
 689         struct hlist_head *bucket;
 690         struct kvm_mmu_page *sp;
 691         struct hlist_node *node;
 692
 693         role.word = 0;
 694         role.glevels = vcpu->arch.mmu.root_level;
 695         role.level = level;
 696         role.metaphysical = metaphysical;
 697         role.access = access;
 698         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
 699                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 700                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 701                 role.quadrant = quadrant;
 702         }
 703         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 704                  gfn, role.word);
 705         index = kvm_page_table_hashfn(gfn);
 706         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
 707         hlist_for_each_entry(sp, node, bucket, hash_link)
 708                 if (sp->gfn == gfn && sp->role.word == role.word) {
 709                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
 710                         pgprintk("%s: found\n", __FUNCTION__);
 711                         return sp;
 712                 }
 713         ++vcpu->kvm->stat.mmu_cache_miss;
 714         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
 715         if (!sp)
 716                 return sp;
 717         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 718         sp->gfn = gfn;
 719         sp->role = role;
 720         hlist_add_head(&sp->hash_link, bucket);
 721         vcpu->arch.mmu.prefetch_page(vcpu, sp);
 722         if (!metaphysical)
 723                 rmap_write_protect(vcpu->kvm, gfn);
 724         return sp;
 725 }
 726
 727 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 728                                          struct kvm_mmu_page *sp)
 729 {
 730         unsigned i;
 731         u64 *pt;
 732         u64 ent;
 733
 734         pt = sp->spt;
 735
 736         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
 737                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 738                         if (is_shadow_present_pte(pt[i]))
 739                                 rmap_remove(kvm, &pt[i]);
 740                         pt[i] = shadow_trap_nonpresent_pte;
 741                 }
 742                 kvm_flush_remote_tlbs(kvm);
 743                 return;
 744         }
 745
 746         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 747                 ent = pt[i];
 748
 749                 pt[i] = shadow_trap_nonpresent_pte;
 750                 if (!is_shadow_present_pte(ent))
 751                         continue;
 752                 ent &= PT64_BASE_ADDR_MASK;
 753                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 754         }
 755         kvm_flush_remote_tlbs(kvm);
 756 }
 757
 758 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
 759 {
 760         mmu_page_remove_parent_pte(sp, parent_pte);
 761 }
 762
 763 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 764 {
 765         int i;
 766
 767         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 768                 if (kvm->vcpus[i])
 769                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
 770 }
 771
 772 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 773 {
 774         u64 *parent_pte;
 775
 776         ++kvm->stat.mmu_shadow_zapped;
 777         while (sp->multimapped || sp->parent_pte) {
 778                 if (!sp->multimapped)
 779                         parent_pte = sp->parent_pte;
 780                 else {
 781                         struct kvm_pte_chain *chain;
 782
 783                         chain = container_of(sp->parent_ptes.first,
 784                                              struct kvm_pte_chain, link);
 785                         parent_pte = chain->parent_ptes[0];
 786                 }
 787                 BUG_ON(!parent_pte);
 788                 kvm_mmu_put_page(sp, parent_pte);
 789                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 790         }
 791         kvm_mmu_page_unlink_children(kvm, sp);
 792         if (!sp->root_count) {
 793                 hlist_del(&sp->hash_link);
 794                 kvm_mmu_free_page(kvm, sp);
 795         } else
 796                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
 797         kvm_mmu_reset_last_pte_updated(kvm);
 798 }
 799
 800 /*
 801  * Changing the number of mmu pages allocated to the vm
 802  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 803  */
 804 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 805 {
 806         /*
 807          * If we set the number of mmu pages to be smaller be than the
 808          * number of actived pages , we must to free some mmu pages before we
 809          * change the value
 810          */
 811
 812         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
 813             kvm_nr_mmu_pages) {
 814                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
 815                                        - kvm->arch.n_free_mmu_pages;
 816
 817                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 818                         struct kvm_mmu_page *page;
 819
 820                         page = container_of(kvm->arch.active_mmu_pages.prev,
 821                                             struct kvm_mmu_page, link);
 822                         kvm_mmu_zap_page(kvm, page);
 823                         n_used_mmu_pages--;
 824                 }
 825                 kvm->arch.n_free_mmu_pages = 0;
 826         }
 827         else
 828                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
 829                                          - kvm->arch.n_alloc_mmu_pages;
 830
 831         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
 832 }
 833
 834 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 835 {
 836         unsigned index;
 837         struct hlist_head *bucket;
 838         struct kvm_mmu_page *sp;
 839         struct hlist_node *node, *n;
 840         int r;
 841
 842         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 843         r = 0;
 844         index = kvm_page_table_hashfn(gfn);
 845         bucket = &kvm->arch.mmu_page_hash[index];
 846         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
 847                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 848                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 849                                  sp->role.word);
 850                         kvm_mmu_zap_page(kvm, sp);
 851                         r = 1;
 852                 }
 853         return r;
 854 }
 855
 856 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
 857 {
 858         struct kvm_mmu_page *sp;
 859
 860         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
 861                 pgprintk("%s: zap %lx %x\n", __FUNCTION__, gfn, sp->role.word);
 862                 kvm_mmu_zap_page(kvm, sp);
 863         }
 864 }
 865
 866 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
 867 {
 868         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
 869         struct kvm_mmu_page *sp = page_header(__pa(pte));
 870
 871         __set_bit(slot, &sp->slot_bitmap);
 872 }
 873
 874 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 875 {
 876         struct page *page;
 877
 878         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
 879
 880         if (gpa == UNMAPPED_GVA)
 881                 return NULL;
 882
 883         down_read(&current->mm->mmap_sem);
 884         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 885         up_read(&current->mm->mmap_sem);
 886
 887         return page;
 888 }
 889
 890 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
 891                          unsigned pt_access, unsigned pte_access,
 892                          int user_fault, int write_fault, int dirty,
 893                          int *ptwrite, gfn_t gfn, struct page *page)
 894 {
 895         u64 spte;
 896         int was_rmapped = 0;
 897         int was_writeble = is_writeble_pte(*shadow_pte);
 898         hfn_t host_pfn = (*shadow_pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 899
 900         pgprintk("%s: spte %llx access %x write_fault %d"
 901                  " user_fault %d gfn %lx\n",
 902                  __FUNCTION__, *shadow_pte, pt_access,
 903                  write_fault, user_fault, gfn);
 904
 905         if (is_rmap_pte(*shadow_pte)) {
 906                 if (host_pfn != page_to_pfn(page)) {
 907                         pgprintk("hfn old %lx new %lx\n",
 908                                  host_pfn, page_to_pfn(page));
 909                         rmap_remove(vcpu->kvm, shadow_pte);
 910                 }
 911                 else
 912                         was_rmapped = 1;
 913         }
 914
 915         /*
 916          * We don't set the accessed bit, since we sometimes want to see
 917          * whether the guest actually used the pte (in order to detect
 918          * demand paging).
 919          */
 920         spte = PT_PRESENT_MASK | PT_DIRTY_MASK;
 921         if (!dirty)
 922                 pte_access &= ~ACC_WRITE_MASK;
 923         if (!(pte_access & ACC_EXEC_MASK))
 924                 spte |= PT64_NX_MASK;
 925
 926         spte |= PT_PRESENT_MASK;
 927         if (pte_access & ACC_USER_MASK)
 928                 spte |= PT_USER_MASK;
 929
 930         spte |= page_to_phys(page);
 931
 932         if ((pte_access & ACC_WRITE_MASK)
 933             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
 934                 struct kvm_mmu_page *shadow;
 935
 936                 spte |= PT_WRITABLE_MASK;
 937                 if (user_fault) {
 938                         mmu_unshadow(vcpu->kvm, gfn);
 939                         goto unshadowed;
 940                 }
 941
 942                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
 943                 if (shadow) {
 944                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 945                                  __FUNCTION__, gfn);
 946                         pte_access &= ~ACC_WRITE_MASK;
 947                         if (is_writeble_pte(spte)) {
 948                                 spte &= ~PT_WRITABLE_MASK;
 949                                 kvm_x86_ops->tlb_flush(vcpu);
 950                         }
 951                         if (write_fault)
 952                                 *ptwrite = 1;
 953                 }
 954         }
 955
 956 unshadowed:
 957
 958         if (pte_access & ACC_WRITE_MASK)
 959                 mark_page_dirty(vcpu->kvm, gfn);
 960
 961         pgprintk("%s: setting spte %llx\n", __FUNCTION__, spte);
 962         set_shadow_pte(shadow_pte, spte);
 963         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
 964         if (!was_rmapped) {
 965                 rmap_add(vcpu, shadow_pte, gfn);
 966                 if (!is_rmap_pte(*shadow_pte))
 967                         kvm_release_page_clean(page);
 968         } else {
 969                 if (was_writeble)
 970                         kvm_release_page_dirty(page);
 971                 else
 972                         kvm_release_page_clean(page);
 973         }
 974         if (!ptwrite || !*ptwrite)
 975                 vcpu->arch.last_pte_updated = shadow_pte;
 976 }
 977
 978 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 979 {
 980 }
 981
 982 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
 983                            gfn_t gfn, struct page *page, int level)
 984 {
 985         hpa_t table_addr = vcpu->arch.mmu.root_hpa;
 986         int pt_write = 0;
 987
 988         for (; ; level--) {
 989                 u32 index = PT64_INDEX(v, level);
 990                 u64 *table;
 991
 992                 ASSERT(VALID_PAGE(table_addr));
 993                 table = __va(table_addr);
 994
 995                 if (level == 1) {
 996                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
 997                                      0, write, 1, &pt_write, gfn, page);
 998                         return pt_write;
 999                 }
1000
1001                 if (table[index] == shadow_trap_nonpresent_pte) {
1002                         struct kvm_mmu_page *new_table;
1003                         gfn_t pseudo_gfn;
1004
1005                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1006                                 >> PAGE_SHIFT;
1007                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1008                                                      v, level - 1,
1009                                                      1, ACC_ALL, &table[index]);
1010                         if (!new_table) {
1011                                 pgprintk("nonpaging_map: ENOMEM\n");
1012                                 kvm_release_page_clean(page);
1013                                 return -ENOMEM;
1014                         }
1015
1016                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
1017                                 | PT_WRITABLE_MASK | PT_USER_MASK;
1018                 }
1019                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1020         }
1021 }
1022
1023 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1024 {
1025         int r;
1026
1027         struct page *page;
1028
1029         down_read(&vcpu->kvm->slots_lock);
1030
1031         down_read(&current->mm->mmap_sem);
1032         page = gfn_to_page(vcpu->kvm, gfn);
1033         up_read(&current->mm->mmap_sem);
1034
1035         /* mmio */
1036         if (is_error_page(page)) {
1037                 kvm_release_page_clean(page);
1038                 up_read(&vcpu->kvm->slots_lock);
1039                 return 1;
1040         }
1041
1042         spin_lock(&vcpu->kvm->mmu_lock);
1043         kvm_mmu_free_some_pages(vcpu);
1044         r = __direct_map(vcpu, v, write, gfn, page, PT32E_ROOT_LEVEL);
1045         spin_unlock(&vcpu->kvm->mmu_lock);
1046
1047         up_read(&vcpu->kvm->slots_lock);
1048
1049         return r;
1050 }
1051
1052
1053 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1054                                     struct kvm_mmu_page *sp)
1055 {
1056         int i;
1057
1058         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1059                 sp->spt[i] = shadow_trap_nonpresent_pte;
1060 }
1061
1062 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1063 {
1064         int i;
1065         struct kvm_mmu_page *sp;
1066
1067         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1068                 return;
1069         spin_lock(&vcpu->kvm->mmu_lock);
1070 #ifdef CONFIG_X86_64
1071         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1072                 hpa_t root = vcpu->arch.mmu.root_hpa;
1073
1074                 sp = page_header(root);
1075                 --sp->root_count;
1076                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1077                 spin_unlock(&vcpu->kvm->mmu_lock);
1078                 return;
1079         }
1080 #endif
1081         for (i = 0; i < 4; ++i) {
1082                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1083
1084                 if (root) {
1085                         root &= PT64_BASE_ADDR_MASK;
1086                         sp = page_header(root);
1087                         --sp->root_count;
1088                 }
1089                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1090         }
1091         spin_unlock(&vcpu->kvm->mmu_lock);
1092         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1093 }
1094
1095 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1096 {
1097         int i;
1098         gfn_t root_gfn;
1099         struct kvm_mmu_page *sp;
1100         int metaphysical = 0;
1101
1102         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1103
1104 #ifdef CONFIG_X86_64
1105         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1106                 hpa_t root = vcpu->arch.mmu.root_hpa;
1107
1108                 ASSERT(!VALID_PAGE(root));
1109                 if (tdp_enabled)
1110                         metaphysical = 1;
1111                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1112                                       PT64_ROOT_LEVEL, metaphysical,
1113                                       ACC_ALL, NULL);
1114                 root = __pa(sp->spt);
1115                 ++sp->root_count;
1116                 vcpu->arch.mmu.root_hpa = root;
1117                 return;
1118         }
1119 #endif
1120         metaphysical = !is_paging(vcpu);
1121         if (tdp_enabled)
1122                 metaphysical = 1;
1123         for (i = 0; i < 4; ++i) {
1124                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1125
1126                 ASSERT(!VALID_PAGE(root));
1127                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1128                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1129                                 vcpu->arch.mmu.pae_root[i] = 0;
1130                                 continue;
1131                         }
1132                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1133                 } else if (vcpu->arch.mmu.root_level == 0)
1134                         root_gfn = 0;
1135                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1136                                       PT32_ROOT_LEVEL, metaphysical,
1137                                       ACC_ALL, NULL);
1138                 root = __pa(sp->spt);
1139                 ++sp->root_count;
1140                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1141         }
1142         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1143 }
1144
1145 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1146 {
1147         return vaddr;
1148 }
1149
1150 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1151                                 u32 error_code)
1152 {
1153         gfn_t gfn;
1154         int r;
1155
1156         pgprintk("%s: gva %lx error %x\n", __FUNCTION__, gva, error_code);
1157         r = mmu_topup_memory_caches(vcpu);
1158         if (r)
1159                 return r;
1160
1161         ASSERT(vcpu);
1162         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1163
1164         gfn = gva >> PAGE_SHIFT;
1165
1166         return nonpaging_map(vcpu, gva & PAGE_MASK,
1167                              error_code & PFERR_WRITE_MASK, gfn);
1168 }
1169
1170 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1171                                 u32 error_code)
1172 {
1173         struct page *page;
1174         int r;
1175
1176         ASSERT(vcpu);
1177         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1178
1179         r = mmu_topup_memory_caches(vcpu);
1180         if (r)
1181                 return r;
1182
1183         down_read(&current->mm->mmap_sem);
1184         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1185         if (is_error_page(page)) {
1186                 kvm_release_page_clean(page);
1187                 up_read(&current->mm->mmap_sem);
1188                 return 1;
1189         }
1190         spin_lock(&vcpu->kvm->mmu_lock);
1191         kvm_mmu_free_some_pages(vcpu);
1192         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1193                          gpa >> PAGE_SHIFT, page, TDP_ROOT_LEVEL);
1194         spin_unlock(&vcpu->kvm->mmu_lock);
1195         up_read(&current->mm->mmap_sem);
1196
1197         return r;
1198 }
1199
1200 static void nonpaging_free(struct kvm_vcpu *vcpu)
1201 {
1202         mmu_free_roots(vcpu);
1203 }
1204
1205 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1206 {
1207         struct kvm_mmu *context = &vcpu->arch.mmu;
1208
1209         context->new_cr3 = nonpaging_new_cr3;
1210         context->page_fault = nonpaging_page_fault;
1211         context->gva_to_gpa = nonpaging_gva_to_gpa;
1212         context->free = nonpaging_free;
1213         context->prefetch_page = nonpaging_prefetch_page;
1214         context->root_level = 0;
1215         context->shadow_root_level = PT32E_ROOT_LEVEL;
1216         context->root_hpa = INVALID_PAGE;
1217         return 0;
1218 }
1219
1220 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1221 {
1222         ++vcpu->stat.tlb_flush;
1223         kvm_x86_ops->tlb_flush(vcpu);
1224 }
1225
1226 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1227 {
1228         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->arch.cr3);
1229         mmu_free_roots(vcpu);
1230 }
1231
1232 static void inject_page_fault(struct kvm_vcpu *vcpu,
1233                               u64 addr,
1234                               u32 err_code)
1235 {
1236         kvm_inject_page_fault(vcpu, addr, err_code);
1237 }
1238
1239 static void paging_free(struct kvm_vcpu *vcpu)
1240 {
1241         nonpaging_free(vcpu);
1242 }
1243
1244 #define PTTYPE 64
1245 #include "paging_tmpl.h"
1246 #undef PTTYPE
1247
1248 #define PTTYPE 32
1249 #include "paging_tmpl.h"
1250 #undef PTTYPE
1251
1252 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1253 {
1254         struct kvm_mmu *context = &vcpu->arch.mmu;
1255
1256         ASSERT(is_pae(vcpu));
1257         context->new_cr3 = paging_new_cr3;
1258         context->page_fault = paging64_page_fault;
1259         context->gva_to_gpa = paging64_gva_to_gpa;
1260         context->prefetch_page = paging64_prefetch_page;
1261         context->free = paging_free;
1262         context->root_level = level;
1263         context->shadow_root_level = level;
1264         context->root_hpa = INVALID_PAGE;
1265         return 0;
1266 }
1267
1268 static int paging64_init_context(struct kvm_vcpu *vcpu)
1269 {
1270         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1271 }
1272
1273 static int paging32_init_context(struct kvm_vcpu *vcpu)
1274 {
1275         struct kvm_mmu *context = &vcpu->arch.mmu;
1276
1277         context->new_cr3 = paging_new_cr3;
1278         context->page_fault = paging32_page_fault;
1279         context->gva_to_gpa = paging32_gva_to_gpa;
1280         context->free = paging_free;
1281         context->prefetch_page = paging32_prefetch_page;
1282         context->root_level = PT32_ROOT_LEVEL;
1283         context->shadow_root_level = PT32E_ROOT_LEVEL;
1284         context->root_hpa = INVALID_PAGE;
1285         return 0;
1286 }
1287
1288 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1289 {
1290         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1291 }
1292
1293 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1294 {
1295         struct kvm_mmu *context = &vcpu->arch.mmu;
1296
1297         context->new_cr3 = nonpaging_new_cr3;
1298         context->page_fault = tdp_page_fault;
1299         context->free = nonpaging_free;
1300         context->prefetch_page = nonpaging_prefetch_page;
1301         context->shadow_root_level = TDP_ROOT_LEVEL;
1302         context->root_hpa = INVALID_PAGE;
1303
1304         if (!is_paging(vcpu)) {
1305                 context->gva_to_gpa = nonpaging_gva_to_gpa;
1306                 context->root_level = 0;
1307         } else if (is_long_mode(vcpu)) {
1308                 context->gva_to_gpa = paging64_gva_to_gpa;
1309                 context->root_level = PT64_ROOT_LEVEL;
1310         } else if (is_pae(vcpu)) {
1311                 context->gva_to_gpa = paging64_gva_to_gpa;
1312                 context->root_level = PT32E_ROOT_LEVEL;
1313         } else {
1314                 context->gva_to_gpa = paging32_gva_to_gpa;
1315                 context->root_level = PT32_ROOT_LEVEL;
1316         }
1317
1318         return 0;
1319 }
1320
1321 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1322 {
1323         ASSERT(vcpu);
1324         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1325
1326         if (!is_paging(vcpu))
1327                 return nonpaging_init_context(vcpu);
1328         else if (is_long_mode(vcpu))
1329                 return paging64_init_context(vcpu);
1330         else if (is_pae(vcpu))
1331                 return paging32E_init_context(vcpu);
1332         else
1333                 return paging32_init_context(vcpu);
1334 }
1335
1336 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1337 {
1338         if (tdp_enabled)
1339                 return init_kvm_tdp_mmu(vcpu);
1340         else
1341                 return init_kvm_softmmu(vcpu);
1342 }
1343
1344 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1345 {
1346         ASSERT(vcpu);
1347         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1348                 vcpu->arch.mmu.free(vcpu);
1349                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1350         }
1351 }
1352
1353 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1354 {
1355         destroy_kvm_mmu(vcpu);
1356         return init_kvm_mmu(vcpu);
1357 }
1358 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1359
1360 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1361 {
1362         int r;
1363
1364         r = mmu_topup_memory_caches(vcpu);
1365         if (r)
1366                 goto out;
1367         spin_lock(&vcpu->kvm->mmu_lock);
1368         kvm_mmu_free_some_pages(vcpu);
1369         mmu_alloc_roots(vcpu);
1370         spin_unlock(&vcpu->kvm->mmu_lock);
1371         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1372         kvm_mmu_flush_tlb(vcpu);
1373 out:
1374         return r;
1375 }
1376 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1377
1378 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1379 {
1380         mmu_free_roots(vcpu);
1381 }
1382
1383 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1384                                   struct kvm_mmu_page *sp,
1385                                   u64 *spte)
1386 {
1387         u64 pte;
1388         struct kvm_mmu_page *child;
1389
1390         pte = *spte;
1391         if (is_shadow_present_pte(pte)) {
1392                 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1393                         rmap_remove(vcpu->kvm, spte);
1394                 else {
1395                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1396                         mmu_page_remove_parent_pte(child, spte);
1397                 }
1398         }
1399         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1400 }
1401
1402 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1403                                   struct kvm_mmu_page *sp,
1404                                   u64 *spte,
1405                                   const void *new)
1406 {
1407         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1408                 ++vcpu->kvm->stat.mmu_pde_zapped;
1409                 return;
1410         }
1411
1412         ++vcpu->kvm->stat.mmu_pte_updated;
1413         if (sp->role.glevels == PT32_ROOT_LEVEL)
1414                 paging32_update_pte(vcpu, sp, spte, new);
1415         else
1416                 paging64_update_pte(vcpu, sp, spte, new);
1417 }
1418
1419 static bool need_remote_flush(u64 old, u64 new)
1420 {
1421         if (!is_shadow_present_pte(old))
1422                 return false;
1423         if (!is_shadow_present_pte(new))
1424                 return true;
1425         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1426                 return true;
1427         old ^= PT64_NX_MASK;
1428         new ^= PT64_NX_MASK;
1429         return (old & ~new & PT64_PERM_MASK) != 0;
1430 }
1431
1432 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1433 {
1434         if (need_remote_flush(old, new))
1435                 kvm_flush_remote_tlbs(vcpu->kvm);
1436         else
1437                 kvm_mmu_flush_tlb(vcpu);
1438 }
1439
1440 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1441 {
1442         u64 *spte = vcpu->arch.last_pte_updated;
1443
1444         return !!(spte && (*spte & PT_ACCESSED_MASK));
1445 }
1446
1447 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1448                                           const u8 *new, int bytes)
1449 {
1450         gfn_t gfn;
1451         int r;
1452         u64 gpte = 0;
1453         struct page *page;
1454
1455         if (bytes != 4 && bytes != 8)
1456                 return;
1457
1458         /*
1459          * Assume that the pte write on a page table of the same type
1460          * as the current vcpu paging mode.  This is nearly always true
1461          * (might be false while changing modes).  Note it is verified later
1462          * by update_pte().
1463          */
1464         if (is_pae(vcpu)) {
1465                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1466                 if ((bytes == 4) && (gpa % 4 == 0)) {
1467                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1468                         if (r)
1469                                 return;
1470                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1471                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1472                         memcpy((void *)&gpte, new, 8);
1473                 }
1474         } else {
1475                 if ((bytes == 4) && (gpa % 4 == 0))
1476                         memcpy((void *)&gpte, new, 4);
1477         }
1478         if (!is_present_pte(gpte))
1479                 return;
1480         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1481
1482         down_read(&vcpu->kvm->slots_lock);
1483         page = gfn_to_page(vcpu->kvm, gfn);
1484         up_read(&vcpu->kvm->slots_lock);
1485
1486         if (is_error_page(page)) {
1487                 kvm_release_page_clean(page);
1488                 return;
1489         }
1490         vcpu->arch.update_pte.gfn = gfn;
1491         vcpu->arch.update_pte.page = page;
1492 }
1493
1494 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1495                        const u8 *new, int bytes)
1496 {
1497         gfn_t gfn = gpa >> PAGE_SHIFT;
1498         struct kvm_mmu_page *sp;
1499         struct hlist_node *node, *n;
1500         struct hlist_head *bucket;
1501         unsigned index;
1502         u64 entry, gentry;
1503         u64 *spte;
1504         unsigned offset = offset_in_page(gpa);
1505         unsigned pte_size;
1506         unsigned page_offset;
1507         unsigned misaligned;
1508         unsigned quadrant;
1509         int level;
1510         int flooded = 0;
1511         int npte;
1512         int r;
1513
1514         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1515         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1516         spin_lock(&vcpu->kvm->mmu_lock);
1517         kvm_mmu_free_some_pages(vcpu);
1518         ++vcpu->kvm->stat.mmu_pte_write;
1519         kvm_mmu_audit(vcpu, "pre pte write");
1520         if (gfn == vcpu->arch.last_pt_write_gfn
1521             && !last_updated_pte_accessed(vcpu)) {
1522                 ++vcpu->arch.last_pt_write_count;
1523                 if (vcpu->arch.last_pt_write_count >= 3)
1524                         flooded = 1;
1525         } else {
1526                 vcpu->arch.last_pt_write_gfn = gfn;
1527                 vcpu->arch.last_pt_write_count = 1;
1528                 vcpu->arch.last_pte_updated = NULL;
1529         }
1530         index = kvm_page_table_hashfn(gfn);
1531         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1532         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1533                 if (sp->gfn != gfn || sp->role.metaphysical)
1534                         continue;
1535                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1536                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1537                 misaligned |= bytes < 4;
1538                 if (misaligned || flooded) {
1539                         /*
1540                          * Misaligned accesses are too much trouble to fix
1541                          * up; also, they usually indicate a page is not used
1542                          * as a page table.
1543                          *
1544                          * If we're seeing too many writes to a page,
1545                          * it may no longer be a page table, or we may be
1546                          * forking, in which case it is better to unmap the
1547                          * page.
1548                          */
1549                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1550                                  gpa, bytes, sp->role.word);
1551                         kvm_mmu_zap_page(vcpu->kvm, sp);
1552                         ++vcpu->kvm->stat.mmu_flooded;
1553                         continue;
1554                 }
1555                 page_offset = offset;
1556                 level = sp->role.level;
1557                 npte = 1;
1558                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1559                         page_offset <<= 1;      /* 32->64 */
1560                         /*
1561                          * A 32-bit pde maps 4MB while the shadow pdes map
1562                          * only 2MB.  So we need to double the offset again
1563                          * and zap two pdes instead of one.
1564                          */
1565                         if (level == PT32_ROOT_LEVEL) {
1566                                 page_offset &= ~7; /* kill rounding error */
1567                                 page_offset <<= 1;
1568                                 npte = 2;
1569                         }
1570                         quadrant = page_offset >> PAGE_SHIFT;
1571                         page_offset &= ~PAGE_MASK;
1572                         if (quadrant != sp->role.quadrant)
1573                                 continue;
1574                 }
1575                 spte = &sp->spt[page_offset / sizeof(*spte)];
1576                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1577                         gentry = 0;
1578                         r = kvm_read_guest_atomic(vcpu->kvm,
1579                                                   gpa & ~(u64)(pte_size - 1),
1580                                                   &gentry, pte_size);
1581                         new = (const void *)&gentry;
1582                         if (r < 0)
1583                                 new = NULL;
1584                 }
1585                 while (npte--) {
1586                         entry = *spte;
1587                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1588                         if (new)
1589                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1590                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1591                         ++spte;
1592                 }
1593         }
1594         kvm_mmu_audit(vcpu, "post pte write");
1595         spin_unlock(&vcpu->kvm->mmu_lock);
1596         if (vcpu->arch.update_pte.page) {
1597                 kvm_release_page_clean(vcpu->arch.update_pte.page);
1598                 vcpu->arch.update_pte.page = NULL;
1599         }
1600 }
1601
1602 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1603 {
1604         gpa_t gpa;
1605         int r;
1606
1607         down_read(&vcpu->kvm->slots_lock);
1608         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1609         up_read(&vcpu->kvm->slots_lock);
1610
1611         spin_lock(&vcpu->kvm->mmu_lock);
1612         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1613         spin_unlock(&vcpu->kvm->mmu_lock);
1614         return r;
1615 }
1616
1617 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1618 {
1619         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1620                 struct kvm_mmu_page *sp;
1621
1622                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1623                                   struct kvm_mmu_page, link);
1624                 kvm_mmu_zap_page(vcpu->kvm, sp);
1625                 ++vcpu->kvm->stat.mmu_recycled;
1626         }
1627 }
1628
1629 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1630 {
1631         int r;
1632         enum emulation_result er;
1633
1634         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1635         if (r < 0)
1636                 goto out;
1637
1638         if (!r) {
1639                 r = 1;
1640                 goto out;
1641         }
1642
1643         r = mmu_topup_memory_caches(vcpu);
1644         if (r)
1645                 goto out;
1646
1647         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1648
1649         switch (er) {
1650         case EMULATE_DONE:
1651                 return 1;
1652         case EMULATE_DO_MMIO:
1653                 ++vcpu->stat.mmio_exits;
1654                 return 0;
1655         case EMULATE_FAIL:
1656                 kvm_report_emulation_failure(vcpu, "pagetable");
1657                 return 1;
1658         default:
1659                 BUG();
1660         }
1661 out:
1662         return r;
1663 }
1664 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1665
1666 void kvm_enable_tdp(void)
1667 {
1668         tdp_enabled = true;
1669 }
1670 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1671
1672 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1673 {
1674         struct kvm_mmu_page *sp;
1675
1676         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1677                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1678                                   struct kvm_mmu_page, link);
1679                 kvm_mmu_zap_page(vcpu->kvm, sp);
1680         }
1681         free_page((unsigned long)vcpu->arch.mmu.pae_root);
1682 }
1683
1684 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1685 {
1686         struct page *page;
1687         int i;
1688
1689         ASSERT(vcpu);
1690
1691         if (vcpu->kvm->arch.n_requested_mmu_pages)
1692                 vcpu->kvm->arch.n_free_mmu_pages =
1693                                         vcpu->kvm->arch.n_requested_mmu_pages;
1694         else
1695                 vcpu->kvm->arch.n_free_mmu_pages =
1696                                         vcpu->kvm->arch.n_alloc_mmu_pages;
1697         /*
1698          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1699          * Therefore we need to allocate shadow page tables in the first
1700          * 4GB of memory, which happens to fit the DMA32 zone.
1701          */
1702         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1703         if (!page)
1704                 goto error_1;
1705         vcpu->arch.mmu.pae_root = page_address(page);
1706         for (i = 0; i < 4; ++i)
1707                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1708
1709         return 0;
1710
1711 error_1:
1712         free_mmu_pages(vcpu);
1713         return -ENOMEM;
1714 }
1715
1716 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1717 {
1718         ASSERT(vcpu);
1719         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1720
1721         return alloc_mmu_pages(vcpu);
1722 }
1723
1724 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1725 {
1726         ASSERT(vcpu);
1727         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1728
1729         return init_kvm_mmu(vcpu);
1730 }
1731
1732 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1733 {
1734         ASSERT(vcpu);
1735
1736         destroy_kvm_mmu(vcpu);
1737         free_mmu_pages(vcpu);
1738         mmu_free_memory_caches(vcpu);
1739 }
1740
1741 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1742 {
1743         struct kvm_mmu_page *sp;
1744
1745         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1746                 int i;
1747                 u64 *pt;
1748
1749                 if (!test_bit(slot, &sp->slot_bitmap))
1750                         continue;
1751
1752                 pt = sp->spt;
1753                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1754                         /* avoid RMW */
1755                         if (pt[i] & PT_WRITABLE_MASK)
1756                                 pt[i] &= ~PT_WRITABLE_MASK;
1757         }
1758 }
1759
1760 void kvm_mmu_zap_all(struct kvm *kvm)
1761 {
1762         struct kvm_mmu_page *sp, *node;
1763
1764         spin_lock(&kvm->mmu_lock);
1765         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1766                 kvm_mmu_zap_page(kvm, sp);
1767         spin_unlock(&kvm->mmu_lock);
1768
1769         kvm_flush_remote_tlbs(kvm);
1770 }
1771
1772 void kvm_mmu_module_exit(void)
1773 {
1774         if (pte_chain_cache)
1775                 kmem_cache_destroy(pte_chain_cache);
1776         if (rmap_desc_cache)
1777                 kmem_cache_destroy(rmap_desc_cache);
1778         if (mmu_page_header_cache)
1779                 kmem_cache_destroy(mmu_page_header_cache);
1780 }
1781
1782 int kvm_mmu_module_init(void)
1783 {
1784         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1785                                             sizeof(struct kvm_pte_chain),
1786                                             0, 0, NULL);
1787         if (!pte_chain_cache)
1788                 goto nomem;
1789         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1790                                             sizeof(struct kvm_rmap_desc),
1791                                             0, 0, NULL);
1792         if (!rmap_desc_cache)
1793                 goto nomem;
1794
1795         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1796                                                   sizeof(struct kvm_mmu_page),
1797                                                   0, 0, NULL);
1798         if (!mmu_page_header_cache)
1799                 goto nomem;
1800
1801         return 0;
1802
1803 nomem:
1804         kvm_mmu_module_exit();
1805         return -ENOMEM;
1806 }
1807
1808 /*
1809  * Caculate mmu pages needed for kvm.
1810  */
1811 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
1812 {
1813         int i;
1814         unsigned int nr_mmu_pages;
1815         unsigned int  nr_pages = 0;
1816
1817         for (i = 0; i < kvm->nmemslots; i++)
1818                 nr_pages += kvm->memslots[i].npages;
1819
1820         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
1821         nr_mmu_pages = max(nr_mmu_pages,
1822                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
1823
1824         return nr_mmu_pages;
1825 }
1826
1827 #ifdef AUDIT
1828
1829 static const char *audit_msg;
1830
1831 static gva_t canonicalize(gva_t gva)
1832 {
1833 #ifdef CONFIG_X86_64
1834         gva = (long long)(gva << 16) >> 16;
1835 #endif
1836         return gva;
1837 }
1838
1839 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1840                                 gva_t va, int level)
1841 {
1842         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1843         int i;
1844         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1845
1846         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1847                 u64 ent = pt[i];
1848
1849                 if (ent == shadow_trap_nonpresent_pte)
1850                         continue;
1851
1852                 va = canonicalize(va);
1853                 if (level > 1) {
1854                         if (ent == shadow_notrap_nonpresent_pte)
1855                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1856                                        " in nonleaf level: levels %d gva %lx"
1857                                        " level %d pte %llx\n", audit_msg,
1858                                        vcpu->arch.mmu.root_level, va, level, ent);
1859
1860                         audit_mappings_page(vcpu, ent, va, level - 1);
1861                 } else {
1862                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
1863                         struct page *page = gpa_to_page(vcpu, gpa);
1864                         hpa_t hpa = page_to_phys(page);
1865
1866                         if (is_shadow_present_pte(ent)
1867                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1868                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1869                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1870                                        audit_msg, vcpu->arch.mmu.root_level,
1871                                        va, gpa, hpa, ent,
1872                                        is_shadow_present_pte(ent));
1873                         else if (ent == shadow_notrap_nonpresent_pte
1874                                  && !is_error_hpa(hpa))
1875                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1876                                        " valid guest gva %lx\n", audit_msg, va);
1877                         kvm_release_page_clean(page);
1878
1879                 }
1880         }
1881 }
1882
1883 static void audit_mappings(struct kvm_vcpu *vcpu)
1884 {
1885         unsigned i;
1886
1887         if (vcpu->arch.mmu.root_level == 4)
1888                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
1889         else
1890                 for (i = 0; i < 4; ++i)
1891                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
1892                                 audit_mappings_page(vcpu,
1893                                                     vcpu->arch.mmu.pae_root[i],
1894                                                     i << 30,
1895                                                     2);
1896 }
1897
1898 static int count_rmaps(struct kvm_vcpu *vcpu)
1899 {
1900         int nmaps = 0;
1901         int i, j, k;
1902
1903         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1904                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1905                 struct kvm_rmap_desc *d;
1906
1907                 for (j = 0; j < m->npages; ++j) {
1908                         unsigned long *rmapp = &m->rmap[j];
1909
1910                         if (!*rmapp)
1911                                 continue;
1912                         if (!(*rmapp & 1)) {
1913                                 ++nmaps;
1914                                 continue;
1915                         }
1916                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1917                         while (d) {
1918                                 for (k = 0; k < RMAP_EXT; ++k)
1919                                         if (d->shadow_ptes[k])
1920                                                 ++nmaps;
1921                                         else
1922                                                 break;
1923                                 d = d->more;
1924                         }
1925                 }
1926         }
1927         return nmaps;
1928 }
1929
1930 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1931 {
1932         int nmaps = 0;
1933         struct kvm_mmu_page *sp;
1934         int i;
1935
1936         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1937                 u64 *pt = sp->spt;
1938
1939                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
1940                         continue;
1941
1942                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1943                         u64 ent = pt[i];
1944
1945                         if (!(ent & PT_PRESENT_MASK))
1946                                 continue;
1947                         if (!(ent & PT_WRITABLE_MASK))
1948                                 continue;
1949                         ++nmaps;
1950                 }
1951         }
1952         return nmaps;
1953 }
1954
1955 static void audit_rmap(struct kvm_vcpu *vcpu)
1956 {
1957         int n_rmap = count_rmaps(vcpu);
1958         int n_actual = count_writable_mappings(vcpu);
1959
1960         if (n_rmap != n_actual)
1961                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1962                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1963 }
1964
1965 static void audit_write_protection(struct kvm_vcpu *vcpu)
1966 {
1967         struct kvm_mmu_page *sp;
1968         struct kvm_memory_slot *slot;
1969         unsigned long *rmapp;
1970         gfn_t gfn;
1971
1972         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1973                 if (sp->role.metaphysical)
1974                         continue;
1975
1976                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
1977                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
1978                 rmapp = &slot->rmap[gfn - slot->base_gfn];
1979                 if (*rmapp)
1980                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1981                                " mappings: gfn %lx role %x\n",
1982                                __FUNCTION__, audit_msg, sp->gfn,
1983                                sp->role.word);
1984         }
1985 }
1986
1987 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1988 {
1989         int olddbg = dbg;
1990
1991         dbg = 0;
1992         audit_msg = msg;
1993         audit_rmap(vcpu);
1994         audit_write_protection(vcpu);
1995         audit_mappings(vcpu);
1996         dbg = olddbg;
1997 }
1998
1999 #endif