drivers/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 #include <linux/types.h>
  20 #include <linux/string.h>
  21 #include <asm/page.h>
  22 #include <linux/mm.h>
  23 #include <linux/highmem.h>
  24 #include <linux/module.h>
  25
  26 #include "vmx.h"
  27 #include "kvm.h"
  28
  29 #undef MMU_DEBUG
  30
  31 #undef AUDIT
  32
  33 #ifdef AUDIT
  34 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  35 #else
  36 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  37 #endif
  38
  39 #ifdef MMU_DEBUG
  40
  41 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  42 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  43
  44 #else
  45
  46 #define pgprintk(x...) do { } while (0)
  47 #define rmap_printk(x...) do { } while (0)
  48
  49 #endif
  50
  51 #if defined(MMU_DEBUG) || defined(AUDIT)
  52 static int dbg = 1;
  53 #endif
  54
  55 #ifndef MMU_DEBUG
  56 #define ASSERT(x) do { } while (0)
  57 #else
  58 #define ASSERT(x)                                                       \
  59         if (!(x)) {                                                     \
  60                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  61                        __FILE__, __LINE__, #x);                         \
  62         }
  63 #endif
  64
  65 #define PT64_PT_BITS 9
  66 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  67 #define PT32_PT_BITS 10
  68 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  69
  70 #define PT_WRITABLE_SHIFT 1
  71
  72 #define PT_PRESENT_MASK (1ULL << 0)
  73 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  74 #define PT_USER_MASK (1ULL << 2)
  75 #define PT_PWT_MASK (1ULL << 3)
  76 #define PT_PCD_MASK (1ULL << 4)
  77 #define PT_ACCESSED_MASK (1ULL << 5)
  78 #define PT_DIRTY_MASK (1ULL << 6)
  79 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  80 #define PT_PAT_MASK (1ULL << 7)
  81 #define PT_GLOBAL_MASK (1ULL << 8)
  82 #define PT64_NX_MASK (1ULL << 63)
  83
  84 #define PT_PAT_SHIFT 7
  85 #define PT_DIR_PAT_SHIFT 12
  86 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  87
  88 #define PT32_DIR_PSE36_SIZE 4
  89 #define PT32_DIR_PSE36_SHIFT 13
  90 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  91
  92
  93 #define PT32_PTE_COPY_MASK \
  94         (PT_PRESENT_MASK | PT_ACCESSED_MASK | PT_DIRTY_MASK | PT_GLOBAL_MASK)
  95
  96 #define PT64_PTE_COPY_MASK (PT64_NX_MASK | PT32_PTE_COPY_MASK)
  97
  98 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  99 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
 100
 101 #define PT_SHADOW_PS_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 102 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 103
 104 #define PT_SHADOW_WRITABLE_SHIFT (PT_FIRST_AVAIL_BITS_SHIFT + 1)
 105 #define PT_SHADOW_WRITABLE_MASK (1ULL << PT_SHADOW_WRITABLE_SHIFT)
 106
 107 #define PT_SHADOW_USER_SHIFT (PT_SHADOW_WRITABLE_SHIFT + 1)
 108 #define PT_SHADOW_USER_MASK (1ULL << (PT_SHADOW_USER_SHIFT))
 109
 110 #define PT_SHADOW_BITS_OFFSET (PT_SHADOW_WRITABLE_SHIFT - PT_WRITABLE_SHIFT)
 111
 112 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 113
 114 #define PT64_LEVEL_BITS 9
 115
 116 #define PT64_LEVEL_SHIFT(level) \
 117                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 118
 119 #define PT64_LEVEL_MASK(level) \
 120                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 121
 122 #define PT64_INDEX(address, level)\
 123         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 124
 125
 126 #define PT32_LEVEL_BITS 10
 127
 128 #define PT32_LEVEL_SHIFT(level) \
 129                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 130
 131 #define PT32_LEVEL_MASK(level) \
 132                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 133
 134 #define PT32_INDEX(address, level)\
 135         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 136
 137
 138 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 139 #define PT64_DIR_BASE_ADDR_MASK \
 140         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 141
 142 #define PT32_BASE_ADDR_MASK PAGE_MASK
 143 #define PT32_DIR_BASE_ADDR_MASK \
 144         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 145
 146
 147 #define PFERR_PRESENT_MASK (1U << 0)
 148 #define PFERR_WRITE_MASK (1U << 1)
 149 #define PFERR_USER_MASK (1U << 2)
 150 #define PFERR_FETCH_MASK (1U << 4)
 151
 152 #define PT64_ROOT_LEVEL 4
 153 #define PT32_ROOT_LEVEL 2
 154 #define PT32E_ROOT_LEVEL 3
 155
 156 #define PT_DIRECTORY_LEVEL 2
 157 #define PT_PAGE_TABLE_LEVEL 1
 158
 159 #define RMAP_EXT 4
 160
 161 struct kvm_rmap_desc {
 162         u64 *shadow_ptes[RMAP_EXT];
 163         struct kvm_rmap_desc *more;
 164 };
 165
 166 static struct kmem_cache *pte_chain_cache;
 167 static struct kmem_cache *rmap_desc_cache;
 168
 169 static int is_write_protection(struct kvm_vcpu *vcpu)
 170 {
 171         return vcpu->cr0 & CR0_WP_MASK;
 172 }
 173
 174 static int is_cpuid_PSE36(void)
 175 {
 176         return 1;
 177 }
 178
 179 static int is_nx(struct kvm_vcpu *vcpu)
 180 {
 181         return vcpu->shadow_efer & EFER_NX;
 182 }
 183
 184 static int is_present_pte(unsigned long pte)
 185 {
 186         return pte & PT_PRESENT_MASK;
 187 }
 188
 189 static int is_writeble_pte(unsigned long pte)
 190 {
 191         return pte & PT_WRITABLE_MASK;
 192 }
 193
 194 static int is_io_pte(unsigned long pte)
 195 {
 196         return pte & PT_SHADOW_IO_MARK;
 197 }
 198
 199 static int is_rmap_pte(u64 pte)
 200 {
 201         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 202                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 203 }
 204
 205 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 206                                   struct kmem_cache *base_cache, int min,
 207                                   gfp_t gfp_flags)
 208 {
 209         void *obj;
 210
 211         if (cache->nobjs >= min)
 212                 return 0;
 213         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 214                 obj = kmem_cache_zalloc(base_cache, gfp_flags);
 215                 if (!obj)
 216                         return -ENOMEM;
 217                 cache->objects[cache->nobjs++] = obj;
 218         }
 219         return 0;
 220 }
 221
 222 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 223 {
 224         while (mc->nobjs)
 225                 kfree(mc->objects[--mc->nobjs]);
 226 }
 227
 228 static int __mmu_topup_memory_caches(struct kvm_vcpu *vcpu, gfp_t gfp_flags)
 229 {
 230         int r;
 231
 232         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 233                                    pte_chain_cache, 4, gfp_flags);
 234         if (r)
 235                 goto out;
 236         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 237                                    rmap_desc_cache, 1, gfp_flags);
 238 out:
 239         return r;
 240 }
 241
 242 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 243 {
 244         int r;
 245
 246         r = __mmu_topup_memory_caches(vcpu, GFP_NOWAIT);
 247         if (r < 0) {
 248                 spin_unlock(&vcpu->kvm->lock);
 249                 kvm_arch_ops->vcpu_put(vcpu);
 250                 r = __mmu_topup_memory_caches(vcpu, GFP_KERNEL);
 251                 kvm_arch_ops->vcpu_load(vcpu);
 252                 spin_lock(&vcpu->kvm->lock);
 253         }
 254         return r;
 255 }
 256
 257 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 258 {
 259         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 260         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 261 }
 262
 263 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 264                                     size_t size)
 265 {
 266         void *p;
 267
 268         BUG_ON(!mc->nobjs);
 269         p = mc->objects[--mc->nobjs];
 270         memset(p, 0, size);
 271         return p;
 272 }
 273
 274 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
 275 {
 276         if (mc->nobjs < KVM_NR_MEM_OBJS)
 277                 mc->objects[mc->nobjs++] = obj;
 278         else
 279                 kfree(obj);
 280 }
 281
 282 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 283 {
 284         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 285                                       sizeof(struct kvm_pte_chain));
 286 }
 287
 288 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
 289                                struct kvm_pte_chain *pc)
 290 {
 291         mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
 292 }
 293
 294 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 295 {
 296         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 297                                       sizeof(struct kvm_rmap_desc));
 298 }
 299
 300 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
 301                                struct kvm_rmap_desc *rd)
 302 {
 303         mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
 304 }
 305
 306 /*
 307  * Reverse mapping data structures:
 308  *
 309  * If page->private bit zero is zero, then page->private points to the
 310  * shadow page table entry that points to page_address(page).
 311  *
 312  * If page->private bit zero is one, (then page->private & ~1) points
 313  * to a struct kvm_rmap_desc containing more mappings.
 314  */
 315 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 316 {
 317         struct page *page;
 318         struct kvm_rmap_desc *desc;
 319         int i;
 320
 321         if (!is_rmap_pte(*spte))
 322                 return;
 323         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 324         if (!page_private(page)) {
 325                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 326                 set_page_private(page,(unsigned long)spte);
 327         } else if (!(page_private(page) & 1)) {
 328                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 329                 desc = mmu_alloc_rmap_desc(vcpu);
 330                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 331                 desc->shadow_ptes[1] = spte;
 332                 set_page_private(page,(unsigned long)desc | 1);
 333         } else {
 334                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 335                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 336                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 337                         desc = desc->more;
 338                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 339                         desc->more = mmu_alloc_rmap_desc(vcpu);
 340                         desc = desc->more;
 341                 }
 342                 for (i = 0; desc->shadow_ptes[i]; ++i)
 343                         ;
 344                 desc->shadow_ptes[i] = spte;
 345         }
 346 }
 347
 348 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
 349                                    struct page *page,
 350                                    struct kvm_rmap_desc *desc,
 351                                    int i,
 352                                    struct kvm_rmap_desc *prev_desc)
 353 {
 354         int j;
 355
 356         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 357                 ;
 358         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 359         desc->shadow_ptes[j] = NULL;
 360         if (j != 0)
 361                 return;
 362         if (!prev_desc && !desc->more)
 363                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 364         else
 365                 if (prev_desc)
 366                         prev_desc->more = desc->more;
 367                 else
 368                         set_page_private(page,(unsigned long)desc->more | 1);
 369         mmu_free_rmap_desc(vcpu, desc);
 370 }
 371
 372 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
 373 {
 374         struct page *page;
 375         struct kvm_rmap_desc *desc;
 376         struct kvm_rmap_desc *prev_desc;
 377         int i;
 378
 379         if (!is_rmap_pte(*spte))
 380                 return;
 381         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 382         if (!page_private(page)) {
 383                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 384                 BUG();
 385         } else if (!(page_private(page) & 1)) {
 386                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 387                 if ((u64 *)page_private(page) != spte) {
 388                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 389                                spte, *spte);
 390                         BUG();
 391                 }
 392                 set_page_private(page,0);
 393         } else {
 394                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 395                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 396                 prev_desc = NULL;
 397                 while (desc) {
 398                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 399                                 if (desc->shadow_ptes[i] == spte) {
 400                                         rmap_desc_remove_entry(vcpu, page,
 401                                                                desc, i,
 402                                                                prev_desc);
 403                                         return;
 404                                 }
 405                         prev_desc = desc;
 406                         desc = desc->more;
 407                 }
 408                 BUG();
 409         }
 410 }
 411
 412 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 413 {
 414         struct kvm *kvm = vcpu->kvm;
 415         struct page *page;
 416         struct kvm_rmap_desc *desc;
 417         u64 *spte;
 418
 419         page = gfn_to_page(kvm, gfn);
 420         BUG_ON(!page);
 421
 422         while (page_private(page)) {
 423                 if (!(page_private(page) & 1))
 424                         spte = (u64 *)page_private(page);
 425                 else {
 426                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 427                         spte = desc->shadow_ptes[0];
 428                 }
 429                 BUG_ON(!spte);
 430                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 431                        != page_to_pfn(page));
 432                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 433                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 434                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 435                 rmap_remove(vcpu, spte);
 436                 kvm_arch_ops->tlb_flush(vcpu);
 437                 *spte &= ~(u64)PT_WRITABLE_MASK;
 438         }
 439 }
 440
 441 #ifdef MMU_DEBUG
 442 static int is_empty_shadow_page(hpa_t page_hpa)
 443 {
 444         u64 *pos;
 445         u64 *end;
 446
 447         for (pos = __va(page_hpa), end = pos + PAGE_SIZE / sizeof(u64);
 448                       pos != end; pos++)
 449                 if (*pos != 0) {
 450                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 451                                pos, *pos);
 452                         return 0;
 453                 }
 454         return 1;
 455 }
 456 #endif
 457
 458 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu, hpa_t page_hpa)
 459 {
 460         struct kvm_mmu_page *page_head = page_header(page_hpa);
 461
 462         ASSERT(is_empty_shadow_page(page_hpa));
 463         page_head->page_hpa = page_hpa;
 464         list_move(&page_head->link, &vcpu->free_pages);
 465         ++vcpu->kvm->n_free_mmu_pages;
 466 }
 467
 468 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 469 {
 470         return gfn;
 471 }
 472
 473 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 474                                                u64 *parent_pte)
 475 {
 476         struct kvm_mmu_page *page;
 477
 478         if (list_empty(&vcpu->free_pages))
 479                 return NULL;
 480
 481         page = list_entry(vcpu->free_pages.next, struct kvm_mmu_page, link);
 482         list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 483         ASSERT(is_empty_shadow_page(page->page_hpa));
 484         page->slot_bitmap = 0;
 485         page->multimapped = 0;
 486         page->parent_pte = parent_pte;
 487         --vcpu->kvm->n_free_mmu_pages;
 488         return page;
 489 }
 490
 491 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 492                                     struct kvm_mmu_page *page, u64 *parent_pte)
 493 {
 494         struct kvm_pte_chain *pte_chain;
 495         struct hlist_node *node;
 496         int i;
 497
 498         if (!parent_pte)
 499                 return;
 500         if (!page->multimapped) {
 501                 u64 *old = page->parent_pte;
 502
 503                 if (!old) {
 504                         page->parent_pte = parent_pte;
 505                         return;
 506                 }
 507                 page->multimapped = 1;
 508                 pte_chain = mmu_alloc_pte_chain(vcpu);
 509                 INIT_HLIST_HEAD(&page->parent_ptes);
 510                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 511                 pte_chain->parent_ptes[0] = old;
 512         }
 513         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 514                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 515                         continue;
 516                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 517                         if (!pte_chain->parent_ptes[i]) {
 518                                 pte_chain->parent_ptes[i] = parent_pte;
 519                                 return;
 520                         }
 521         }
 522         pte_chain = mmu_alloc_pte_chain(vcpu);
 523         BUG_ON(!pte_chain);
 524         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 525         pte_chain->parent_ptes[0] = parent_pte;
 526 }
 527
 528 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
 529                                        struct kvm_mmu_page *page,
 530                                        u64 *parent_pte)
 531 {
 532         struct kvm_pte_chain *pte_chain;
 533         struct hlist_node *node;
 534         int i;
 535
 536         if (!page->multimapped) {
 537                 BUG_ON(page->parent_pte != parent_pte);
 538                 page->parent_pte = NULL;
 539                 return;
 540         }
 541         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 542                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 543                         if (!pte_chain->parent_ptes[i])
 544                                 break;
 545                         if (pte_chain->parent_ptes[i] != parent_pte)
 546                                 continue;
 547                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 548                                 && pte_chain->parent_ptes[i + 1]) {
 549                                 pte_chain->parent_ptes[i]
 550                                         = pte_chain->parent_ptes[i + 1];
 551                                 ++i;
 552                         }
 553                         pte_chain->parent_ptes[i] = NULL;
 554                         if (i == 0) {
 555                                 hlist_del(&pte_chain->link);
 556                                 mmu_free_pte_chain(vcpu, pte_chain);
 557                                 if (hlist_empty(&page->parent_ptes)) {
 558                                         page->multimapped = 0;
 559                                         page->parent_pte = NULL;
 560                                 }
 561                         }
 562                         return;
 563                 }
 564         BUG();
 565 }
 566
 567 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 568                                                 gfn_t gfn)
 569 {
 570         unsigned index;
 571         struct hlist_head *bucket;
 572         struct kvm_mmu_page *page;
 573         struct hlist_node *node;
 574
 575         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 576         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 577         bucket = &vcpu->kvm->mmu_page_hash[index];
 578         hlist_for_each_entry(page, node, bucket, hash_link)
 579                 if (page->gfn == gfn && !page->role.metaphysical) {
 580                         pgprintk("%s: found role %x\n",
 581                                  __FUNCTION__, page->role.word);
 582                         return page;
 583                 }
 584         return NULL;
 585 }
 586
 587 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 588                                              gfn_t gfn,
 589                                              gva_t gaddr,
 590                                              unsigned level,
 591                                              int metaphysical,
 592                                              unsigned hugepage_access,
 593                                              u64 *parent_pte)
 594 {
 595         union kvm_mmu_page_role role;
 596         unsigned index;
 597         unsigned quadrant;
 598         struct hlist_head *bucket;
 599         struct kvm_mmu_page *page;
 600         struct hlist_node *node;
 601
 602         role.word = 0;
 603         role.glevels = vcpu->mmu.root_level;
 604         role.level = level;
 605         role.metaphysical = metaphysical;
 606         role.hugepage_access = hugepage_access;
 607         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 608                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 609                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 610                 role.quadrant = quadrant;
 611         }
 612         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 613                  gfn, role.word);
 614         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 615         bucket = &vcpu->kvm->mmu_page_hash[index];
 616         hlist_for_each_entry(page, node, bucket, hash_link)
 617                 if (page->gfn == gfn && page->role.word == role.word) {
 618                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 619                         pgprintk("%s: found\n", __FUNCTION__);
 620                         return page;
 621                 }
 622         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 623         if (!page)
 624                 return page;
 625         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 626         page->gfn = gfn;
 627         page->role = role;
 628         hlist_add_head(&page->hash_link, bucket);
 629         if (!metaphysical)
 630                 rmap_write_protect(vcpu, gfn);
 631         return page;
 632 }
 633
 634 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 635                                          struct kvm_mmu_page *page)
 636 {
 637         unsigned i;
 638         u64 *pt;
 639         u64 ent;
 640
 641         pt = __va(page->page_hpa);
 642
 643         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 644                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 645                         if (pt[i] & PT_PRESENT_MASK)
 646                                 rmap_remove(vcpu, &pt[i]);
 647                         pt[i] = 0;
 648                 }
 649                 kvm_arch_ops->tlb_flush(vcpu);
 650                 return;
 651         }
 652
 653         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 654                 ent = pt[i];
 655
 656                 pt[i] = 0;
 657                 if (!(ent & PT_PRESENT_MASK))
 658                         continue;
 659                 ent &= PT64_BASE_ADDR_MASK;
 660                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
 661         }
 662 }
 663
 664 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 665                              struct kvm_mmu_page *page,
 666                              u64 *parent_pte)
 667 {
 668         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
 669 }
 670
 671 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 672                              struct kvm_mmu_page *page)
 673 {
 674         u64 *parent_pte;
 675
 676         while (page->multimapped || page->parent_pte) {
 677                 if (!page->multimapped)
 678                         parent_pte = page->parent_pte;
 679                 else {
 680                         struct kvm_pte_chain *chain;
 681
 682                         chain = container_of(page->parent_ptes.first,
 683                                              struct kvm_pte_chain, link);
 684                         parent_pte = chain->parent_ptes[0];
 685                 }
 686                 BUG_ON(!parent_pte);
 687                 kvm_mmu_put_page(vcpu, page, parent_pte);
 688                 *parent_pte = 0;
 689         }
 690         kvm_mmu_page_unlink_children(vcpu, page);
 691         if (!page->root_count) {
 692                 hlist_del(&page->hash_link);
 693                 kvm_mmu_free_page(vcpu, page->page_hpa);
 694         } else
 695                 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 696 }
 697
 698 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 699 {
 700         unsigned index;
 701         struct hlist_head *bucket;
 702         struct kvm_mmu_page *page;
 703         struct hlist_node *node, *n;
 704         int r;
 705
 706         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 707         r = 0;
 708         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 709         bucket = &vcpu->kvm->mmu_page_hash[index];
 710         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 711                 if (page->gfn == gfn && !page->role.metaphysical) {
 712                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 713                                  page->role.word);
 714                         kvm_mmu_zap_page(vcpu, page);
 715                         r = 1;
 716                 }
 717         return r;
 718 }
 719
 720 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 721 {
 722         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 723         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 724
 725         __set_bit(slot, &page_head->slot_bitmap);
 726 }
 727
 728 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 729 {
 730         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 731
 732         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 733 }
 734
 735 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 736 {
 737         struct page *page;
 738
 739         ASSERT((gpa & HPA_ERR_MASK) == 0);
 740         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 741         if (!page)
 742                 return gpa | HPA_ERR_MASK;
 743         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 744                 | (gpa & (PAGE_SIZE-1));
 745 }
 746
 747 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 748 {
 749         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 750
 751         if (gpa == UNMAPPED_GVA)
 752                 return UNMAPPED_GVA;
 753         return gpa_to_hpa(vcpu, gpa);
 754 }
 755
 756 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 757 {
 758         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 759
 760         if (gpa == UNMAPPED_GVA)
 761                 return NULL;
 762         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 763 }
 764
 765 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 766 {
 767 }
 768
 769 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 770 {
 771         int level = PT32E_ROOT_LEVEL;
 772         hpa_t table_addr = vcpu->mmu.root_hpa;
 773
 774         for (; ; level--) {
 775                 u32 index = PT64_INDEX(v, level);
 776                 u64 *table;
 777                 u64 pte;
 778
 779                 ASSERT(VALID_PAGE(table_addr));
 780                 table = __va(table_addr);
 781
 782                 if (level == 1) {
 783                         pte = table[index];
 784                         if (is_present_pte(pte) && is_writeble_pte(pte))
 785                                 return 0;
 786                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 787                         page_header_update_slot(vcpu->kvm, table, v);
 788                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 789                                                                 PT_USER_MASK;
 790                         rmap_add(vcpu, &table[index]);
 791                         return 0;
 792                 }
 793
 794                 if (table[index] == 0) {
 795                         struct kvm_mmu_page *new_table;
 796                         gfn_t pseudo_gfn;
 797
 798                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 799                                 >> PAGE_SHIFT;
 800                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 801                                                      v, level - 1,
 802                                                      1, 0, &table[index]);
 803                         if (!new_table) {
 804                                 pgprintk("nonpaging_map: ENOMEM\n");
 805                                 return -ENOMEM;
 806                         }
 807
 808                         table[index] = new_table->page_hpa | PT_PRESENT_MASK
 809                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 810                 }
 811                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 812         }
 813 }
 814
 815 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 816 {
 817         int i;
 818         struct kvm_mmu_page *page;
 819
 820 #ifdef CONFIG_X86_64
 821         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 822                 hpa_t root = vcpu->mmu.root_hpa;
 823
 824                 ASSERT(VALID_PAGE(root));
 825                 page = page_header(root);
 826                 --page->root_count;
 827                 vcpu->mmu.root_hpa = INVALID_PAGE;
 828                 return;
 829         }
 830 #endif
 831         for (i = 0; i < 4; ++i) {
 832                 hpa_t root = vcpu->mmu.pae_root[i];
 833
 834                 if (root) {
 835                         ASSERT(VALID_PAGE(root));
 836                         root &= PT64_BASE_ADDR_MASK;
 837                         page = page_header(root);
 838                         --page->root_count;
 839                 }
 840                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 841         }
 842         vcpu->mmu.root_hpa = INVALID_PAGE;
 843 }
 844
 845 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 846 {
 847         int i;
 848         gfn_t root_gfn;
 849         struct kvm_mmu_page *page;
 850
 851         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 852
 853 #ifdef CONFIG_X86_64
 854         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 855                 hpa_t root = vcpu->mmu.root_hpa;
 856
 857                 ASSERT(!VALID_PAGE(root));
 858                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 859                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 860                 root = page->page_hpa;
 861                 ++page->root_count;
 862                 vcpu->mmu.root_hpa = root;
 863                 return;
 864         }
 865 #endif
 866         for (i = 0; i < 4; ++i) {
 867                 hpa_t root = vcpu->mmu.pae_root[i];
 868
 869                 ASSERT(!VALID_PAGE(root));
 870                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 871                         if (!is_present_pte(vcpu->pdptrs[i])) {
 872                                 vcpu->mmu.pae_root[i] = 0;
 873                                 continue;
 874                         }
 875                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 876                 } else if (vcpu->mmu.root_level == 0)
 877                         root_gfn = 0;
 878                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 879                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 880                                         0, NULL);
 881                 root = page->page_hpa;
 882                 ++page->root_count;
 883                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 884         }
 885         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 886 }
 887
 888 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 889 {
 890         return vaddr;
 891 }
 892
 893 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 894                                u32 error_code)
 895 {
 896         gpa_t addr = gva;
 897         hpa_t paddr;
 898         int r;
 899
 900         r = mmu_topup_memory_caches(vcpu);
 901         if (r)
 902                 return r;
 903
 904         ASSERT(vcpu);
 905         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 906
 907
 908         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 909
 910         if (is_error_hpa(paddr))
 911                 return 1;
 912
 913         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 914 }
 915
 916 static void nonpaging_free(struct kvm_vcpu *vcpu)
 917 {
 918         mmu_free_roots(vcpu);
 919 }
 920
 921 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 922 {
 923         struct kvm_mmu *context = &vcpu->mmu;
 924
 925         context->new_cr3 = nonpaging_new_cr3;
 926         context->page_fault = nonpaging_page_fault;
 927         context->gva_to_gpa = nonpaging_gva_to_gpa;
 928         context->free = nonpaging_free;
 929         context->root_level = 0;
 930         context->shadow_root_level = PT32E_ROOT_LEVEL;
 931         mmu_alloc_roots(vcpu);
 932         ASSERT(VALID_PAGE(context->root_hpa));
 933         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 934         return 0;
 935 }
 936
 937 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 938 {
 939         ++vcpu->stat.tlb_flush;
 940         kvm_arch_ops->tlb_flush(vcpu);
 941 }
 942
 943 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 944 {
 945         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 946         mmu_free_roots(vcpu);
 947         if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
 948                 kvm_mmu_free_some_pages(vcpu);
 949         mmu_alloc_roots(vcpu);
 950         kvm_mmu_flush_tlb(vcpu);
 951         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 952 }
 953
 954 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 955                              u64 *shadow_pte,
 956                              gpa_t gaddr,
 957                              int dirty,
 958                              u64 access_bits,
 959                              gfn_t gfn)
 960 {
 961         hpa_t paddr;
 962
 963         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 964         if (!dirty)
 965                 access_bits &= ~PT_WRITABLE_MASK;
 966
 967         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 968
 969         *shadow_pte |= access_bits;
 970
 971         if (is_error_hpa(paddr)) {
 972                 *shadow_pte |= gaddr;
 973                 *shadow_pte |= PT_SHADOW_IO_MARK;
 974                 *shadow_pte &= ~PT_PRESENT_MASK;
 975                 return;
 976         }
 977
 978         *shadow_pte |= paddr;
 979
 980         if (access_bits & PT_WRITABLE_MASK) {
 981                 struct kvm_mmu_page *shadow;
 982
 983                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 984                 if (shadow) {
 985                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 986                                  __FUNCTION__, gfn);
 987                         access_bits &= ~PT_WRITABLE_MASK;
 988                         if (is_writeble_pte(*shadow_pte)) {
 989                                     *shadow_pte &= ~PT_WRITABLE_MASK;
 990                                     kvm_arch_ops->tlb_flush(vcpu);
 991                         }
 992                 }
 993         }
 994
 995         if (access_bits & PT_WRITABLE_MASK)
 996                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 997
 998         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 999         rmap_add(vcpu, shadow_pte);
1000 }
1001
1002 static void inject_page_fault(struct kvm_vcpu *vcpu,
1003                               u64 addr,
1004                               u32 err_code)
1005 {
1006         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
1007 }
1008
1009 static inline int fix_read_pf(u64 *shadow_ent)
1010 {
1011         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
1012             !(*shadow_ent & PT_USER_MASK)) {
1013                 /*
1014                  * If supervisor write protect is disabled, we shadow kernel
1015                  * pages as user pages so we can trap the write access.
1016                  */
1017                 *shadow_ent |= PT_USER_MASK;
1018                 *shadow_ent &= ~PT_WRITABLE_MASK;
1019
1020                 return 1;
1021
1022         }
1023         return 0;
1024 }
1025
1026 static void paging_free(struct kvm_vcpu *vcpu)
1027 {
1028         nonpaging_free(vcpu);
1029 }
1030
1031 #define PTTYPE 64
1032 #include "paging_tmpl.h"
1033 #undef PTTYPE
1034
1035 #define PTTYPE 32
1036 #include "paging_tmpl.h"
1037 #undef PTTYPE
1038
1039 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1040 {
1041         struct kvm_mmu *context = &vcpu->mmu;
1042
1043         ASSERT(is_pae(vcpu));
1044         context->new_cr3 = paging_new_cr3;
1045         context->page_fault = paging64_page_fault;
1046         context->gva_to_gpa = paging64_gva_to_gpa;
1047         context->free = paging_free;
1048         context->root_level = level;
1049         context->shadow_root_level = level;
1050         mmu_alloc_roots(vcpu);
1051         ASSERT(VALID_PAGE(context->root_hpa));
1052         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1053                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1054         return 0;
1055 }
1056
1057 static int paging64_init_context(struct kvm_vcpu *vcpu)
1058 {
1059         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1060 }
1061
1062 static int paging32_init_context(struct kvm_vcpu *vcpu)
1063 {
1064         struct kvm_mmu *context = &vcpu->mmu;
1065
1066         context->new_cr3 = paging_new_cr3;
1067         context->page_fault = paging32_page_fault;
1068         context->gva_to_gpa = paging32_gva_to_gpa;
1069         context->free = paging_free;
1070         context->root_level = PT32_ROOT_LEVEL;
1071         context->shadow_root_level = PT32E_ROOT_LEVEL;
1072         mmu_alloc_roots(vcpu);
1073         ASSERT(VALID_PAGE(context->root_hpa));
1074         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1075                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1076         return 0;
1077 }
1078
1079 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1080 {
1081         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1082 }
1083
1084 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1085 {
1086         ASSERT(vcpu);
1087         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1088
1089         if (!is_paging(vcpu))
1090                 return nonpaging_init_context(vcpu);
1091         else if (is_long_mode(vcpu))
1092                 return paging64_init_context(vcpu);
1093         else if (is_pae(vcpu))
1094                 return paging32E_init_context(vcpu);
1095         else
1096                 return paging32_init_context(vcpu);
1097 }
1098
1099 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1100 {
1101         ASSERT(vcpu);
1102         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1103                 vcpu->mmu.free(vcpu);
1104                 vcpu->mmu.root_hpa = INVALID_PAGE;
1105         }
1106 }
1107
1108 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1109 {
1110         int r;
1111
1112         destroy_kvm_mmu(vcpu);
1113         r = init_kvm_mmu(vcpu);
1114         if (r < 0)
1115                 goto out;
1116         r = mmu_topup_memory_caches(vcpu);
1117 out:
1118         return r;
1119 }
1120
1121 static void mmu_pre_write_zap_pte(struct kvm_vcpu *vcpu,
1122                                   struct kvm_mmu_page *page,
1123                                   u64 *spte)
1124 {
1125         u64 pte;
1126         struct kvm_mmu_page *child;
1127
1128         pte = *spte;
1129         if (is_present_pte(pte)) {
1130                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1131                         rmap_remove(vcpu, spte);
1132                 else {
1133                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1134                         mmu_page_remove_parent_pte(vcpu, child, spte);
1135                 }
1136         }
1137         *spte = 0;
1138 }
1139
1140 void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1141 {
1142         gfn_t gfn = gpa >> PAGE_SHIFT;
1143         struct kvm_mmu_page *page;
1144         struct hlist_node *node, *n;
1145         struct hlist_head *bucket;
1146         unsigned index;
1147         u64 *spte;
1148         unsigned offset = offset_in_page(gpa);
1149         unsigned pte_size;
1150         unsigned page_offset;
1151         unsigned misaligned;
1152         int level;
1153         int flooded = 0;
1154         int npte;
1155
1156         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1157         if (gfn == vcpu->last_pt_write_gfn) {
1158                 ++vcpu->last_pt_write_count;
1159                 if (vcpu->last_pt_write_count >= 3)
1160                         flooded = 1;
1161         } else {
1162                 vcpu->last_pt_write_gfn = gfn;
1163                 vcpu->last_pt_write_count = 1;
1164         }
1165         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1166         bucket = &vcpu->kvm->mmu_page_hash[index];
1167         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1168                 if (page->gfn != gfn || page->role.metaphysical)
1169                         continue;
1170                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1171                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1172                 if (misaligned || flooded) {
1173                         /*
1174                          * Misaligned accesses are too much trouble to fix
1175                          * up; also, they usually indicate a page is not used
1176                          * as a page table.
1177                          *
1178                          * If we're seeing too many writes to a page,
1179                          * it may no longer be a page table, or we may be
1180                          * forking, in which case it is better to unmap the
1181                          * page.
1182                          */
1183                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1184                                  gpa, bytes, page->role.word);
1185                         kvm_mmu_zap_page(vcpu, page);
1186                         continue;
1187                 }
1188                 page_offset = offset;
1189                 level = page->role.level;
1190                 npte = 1;
1191                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1192                         page_offset <<= 1;      /* 32->64 */
1193                         /*
1194                          * A 32-bit pde maps 4MB while the shadow pdes map
1195                          * only 2MB.  So we need to double the offset again
1196                          * and zap two pdes instead of one.
1197                          */
1198                         if (level == PT32_ROOT_LEVEL) {
1199                                 page_offset &= ~7; /* kill rounding error */
1200                                 page_offset <<= 1;
1201                                 npte = 2;
1202                         }
1203                         page_offset &= ~PAGE_MASK;
1204                 }
1205                 spte = __va(page->page_hpa);
1206                 spte += page_offset / sizeof(*spte);
1207                 while (npte--) {
1208                         mmu_pre_write_zap_pte(vcpu, page, spte);
1209                         ++spte;
1210                 }
1211         }
1212 }
1213
1214 void kvm_mmu_post_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1215 {
1216 }
1217
1218 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1219 {
1220         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1221
1222         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1223 }
1224
1225 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1226 {
1227         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1228                 struct kvm_mmu_page *page;
1229
1230                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1231                                     struct kvm_mmu_page, link);
1232                 kvm_mmu_zap_page(vcpu, page);
1233         }
1234 }
1235 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1236
1237 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1238 {
1239         struct kvm_mmu_page *page;
1240
1241         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1242                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1243                                     struct kvm_mmu_page, link);
1244                 kvm_mmu_zap_page(vcpu, page);
1245         }
1246         while (!list_empty(&vcpu->free_pages)) {
1247                 page = list_entry(vcpu->free_pages.next,
1248                                   struct kvm_mmu_page, link);
1249                 list_del(&page->link);
1250                 __free_page(pfn_to_page(page->page_hpa >> PAGE_SHIFT));
1251                 page->page_hpa = INVALID_PAGE;
1252         }
1253         free_page((unsigned long)vcpu->mmu.pae_root);
1254 }
1255
1256 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1257 {
1258         struct page *page;
1259         int i;
1260
1261         ASSERT(vcpu);
1262
1263         for (i = 0; i < KVM_NUM_MMU_PAGES; i++) {
1264                 struct kvm_mmu_page *page_header = &vcpu->page_header_buf[i];
1265
1266                 INIT_LIST_HEAD(&page_header->link);
1267                 if ((page = alloc_page(GFP_KERNEL)) == NULL)
1268                         goto error_1;
1269                 set_page_private(page, (unsigned long)page_header);
1270                 page_header->page_hpa = (hpa_t)page_to_pfn(page) << PAGE_SHIFT;
1271                 memset(__va(page_header->page_hpa), 0, PAGE_SIZE);
1272                 list_add(&page_header->link, &vcpu->free_pages);
1273                 ++vcpu->kvm->n_free_mmu_pages;
1274         }
1275
1276         /*
1277          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1278          * Therefore we need to allocate shadow page tables in the first
1279          * 4GB of memory, which happens to fit the DMA32 zone.
1280          */
1281         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1282         if (!page)
1283                 goto error_1;
1284         vcpu->mmu.pae_root = page_address(page);
1285         for (i = 0; i < 4; ++i)
1286                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1287
1288         return 0;
1289
1290 error_1:
1291         free_mmu_pages(vcpu);
1292         return -ENOMEM;
1293 }
1294
1295 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1296 {
1297         ASSERT(vcpu);
1298         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1299         ASSERT(list_empty(&vcpu->free_pages));
1300
1301         return alloc_mmu_pages(vcpu);
1302 }
1303
1304 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1305 {
1306         ASSERT(vcpu);
1307         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1308         ASSERT(!list_empty(&vcpu->free_pages));
1309
1310         return init_kvm_mmu(vcpu);
1311 }
1312
1313 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1314 {
1315         ASSERT(vcpu);
1316
1317         destroy_kvm_mmu(vcpu);
1318         free_mmu_pages(vcpu);
1319         mmu_free_memory_caches(vcpu);
1320 }
1321
1322 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1323 {
1324         struct kvm *kvm = vcpu->kvm;
1325         struct kvm_mmu_page *page;
1326
1327         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1328                 int i;
1329                 u64 *pt;
1330
1331                 if (!test_bit(slot, &page->slot_bitmap))
1332                         continue;
1333
1334                 pt = __va(page->page_hpa);
1335                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1336                         /* avoid RMW */
1337                         if (pt[i] & PT_WRITABLE_MASK) {
1338                                 rmap_remove(vcpu, &pt[i]);
1339                                 pt[i] &= ~PT_WRITABLE_MASK;
1340                         }
1341         }
1342 }
1343
1344 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1345 {
1346         destroy_kvm_mmu(vcpu);
1347
1348         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1349                 struct kvm_mmu_page *page;
1350
1351                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1352                                     struct kvm_mmu_page, link);
1353                 kvm_mmu_zap_page(vcpu, page);
1354         }
1355
1356         mmu_free_memory_caches(vcpu);
1357         kvm_arch_ops->tlb_flush(vcpu);
1358         init_kvm_mmu(vcpu);
1359 }
1360
1361 void kvm_mmu_module_exit(void)
1362 {
1363         if (pte_chain_cache)
1364                 kmem_cache_destroy(pte_chain_cache);
1365         if (rmap_desc_cache)
1366                 kmem_cache_destroy(rmap_desc_cache);
1367 }
1368
1369 int kvm_mmu_module_init(void)
1370 {
1371         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1372                                             sizeof(struct kvm_pte_chain),
1373                                             0, 0, NULL, NULL);
1374         if (!pte_chain_cache)
1375                 goto nomem;
1376         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1377                                             sizeof(struct kvm_rmap_desc),
1378                                             0, 0, NULL, NULL);
1379         if (!rmap_desc_cache)
1380                 goto nomem;
1381
1382         return 0;
1383
1384 nomem:
1385         kvm_mmu_module_exit();
1386         return -ENOMEM;
1387 }
1388
1389 #ifdef AUDIT
1390
1391 static const char *audit_msg;
1392
1393 static gva_t canonicalize(gva_t gva)
1394 {
1395 #ifdef CONFIG_X86_64
1396         gva = (long long)(gva << 16) >> 16;
1397 #endif
1398         return gva;
1399 }
1400
1401 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1402                                 gva_t va, int level)
1403 {
1404         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1405         int i;
1406         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1407
1408         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1409                 u64 ent = pt[i];
1410
1411                 if (!(ent & PT_PRESENT_MASK))
1412                         continue;
1413
1414                 va = canonicalize(va);
1415                 if (level > 1)
1416                         audit_mappings_page(vcpu, ent, va, level - 1);
1417                 else {
1418                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1419                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1420
1421                         if ((ent & PT_PRESENT_MASK)
1422                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1423                                 printk(KERN_ERR "audit error: (%s) levels %d"
1424                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1425                                        audit_msg, vcpu->mmu.root_level,
1426                                        va, gpa, hpa, ent);
1427                 }
1428         }
1429 }
1430
1431 static void audit_mappings(struct kvm_vcpu *vcpu)
1432 {
1433         unsigned i;
1434
1435         if (vcpu->mmu.root_level == 4)
1436                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1437         else
1438                 for (i = 0; i < 4; ++i)
1439                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1440                                 audit_mappings_page(vcpu,
1441                                                     vcpu->mmu.pae_root[i],
1442                                                     i << 30,
1443                                                     2);
1444 }
1445
1446 static int count_rmaps(struct kvm_vcpu *vcpu)
1447 {
1448         int nmaps = 0;
1449         int i, j, k;
1450
1451         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1452                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1453                 struct kvm_rmap_desc *d;
1454
1455                 for (j = 0; j < m->npages; ++j) {
1456                         struct page *page = m->phys_mem[j];
1457
1458                         if (!page->private)
1459                                 continue;
1460                         if (!(page->private & 1)) {
1461                                 ++nmaps;
1462                                 continue;
1463                         }
1464                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1465                         while (d) {
1466                                 for (k = 0; k < RMAP_EXT; ++k)
1467                                         if (d->shadow_ptes[k])
1468                                                 ++nmaps;
1469                                         else
1470                                                 break;
1471                                 d = d->more;
1472                         }
1473                 }
1474         }
1475         return nmaps;
1476 }
1477
1478 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1479 {
1480         int nmaps = 0;
1481         struct kvm_mmu_page *page;
1482         int i;
1483
1484         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1485                 u64 *pt = __va(page->page_hpa);
1486
1487                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1488                         continue;
1489
1490                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1491                         u64 ent = pt[i];
1492
1493                         if (!(ent & PT_PRESENT_MASK))
1494                                 continue;
1495                         if (!(ent & PT_WRITABLE_MASK))
1496                                 continue;
1497                         ++nmaps;
1498                 }
1499         }
1500         return nmaps;
1501 }
1502
1503 static void audit_rmap(struct kvm_vcpu *vcpu)
1504 {
1505         int n_rmap = count_rmaps(vcpu);
1506         int n_actual = count_writable_mappings(vcpu);
1507
1508         if (n_rmap != n_actual)
1509                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1510                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1511 }
1512
1513 static void audit_write_protection(struct kvm_vcpu *vcpu)
1514 {
1515         struct kvm_mmu_page *page;
1516
1517         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1518                 hfn_t hfn;
1519                 struct page *pg;
1520
1521                 if (page->role.metaphysical)
1522                         continue;
1523
1524                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1525                         >> PAGE_SHIFT;
1526                 pg = pfn_to_page(hfn);
1527                 if (pg->private)
1528                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1529                                " mappings: gfn %lx role %x\n",
1530                                __FUNCTION__, audit_msg, page->gfn,
1531                                page->role.word);
1532         }
1533 }
1534
1535 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1536 {
1537         int olddbg = dbg;
1538
1539         dbg = 0;
1540         audit_msg = msg;
1541         audit_rmap(vcpu);
1542         audit_write_protection(vcpu);
1543         audit_mappings(vcpu);
1544         dbg = olddbg;
1545 }
1546
1547 #endif