err.no Git - linux-2.6/blob - 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 #define ASSERT(x)                                                       \
  56         if (!(x)) {                                                     \
  57                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  58                        __FILE__, __LINE__, #x);                         \
  59         }
  60
  61 #define PT64_PT_BITS 9
  62 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  63 #define PT32_PT_BITS 10
  64 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  65
  66 #define PT_WRITABLE_SHIFT 1
  67
  68 #define PT_PRESENT_MASK (1ULL << 0)
  69 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  70 #define PT_USER_MASK (1ULL << 2)
  71 #define PT_PWT_MASK (1ULL << 3)
  72 #define PT_PCD_MASK (1ULL << 4)
  73 #define PT_ACCESSED_MASK (1ULL << 5)
  74 #define PT_DIRTY_MASK (1ULL << 6)
  75 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  76 #define PT_PAT_MASK (1ULL << 7)
  77 #define PT_GLOBAL_MASK (1ULL << 8)
  78 #define PT64_NX_MASK (1ULL << 63)
  79
  80 #define PT_PAT_SHIFT 7
  81 #define PT_DIR_PAT_SHIFT 12
  82 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  83
  84 #define PT32_DIR_PSE36_SIZE 4
  85 #define PT32_DIR_PSE36_SHIFT 13
  86 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  87
  88
  89 #define PT32_PTE_COPY_MASK \
  90         (PT_PRESENT_MASK | PT_ACCESSED_MASK | PT_DIRTY_MASK | PT_GLOBAL_MASK)
  91
  92 #define PT64_PTE_COPY_MASK (PT64_NX_MASK | PT32_PTE_COPY_MASK)
  93
  94 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  95 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  96
  97 #define PT_SHADOW_PS_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
  98 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
  99
 100 #define PT_SHADOW_WRITABLE_SHIFT (PT_FIRST_AVAIL_BITS_SHIFT + 1)
 101 #define PT_SHADOW_WRITABLE_MASK (1ULL << PT_SHADOW_WRITABLE_SHIFT)
 102
 103 #define PT_SHADOW_USER_SHIFT (PT_SHADOW_WRITABLE_SHIFT + 1)
 104 #define PT_SHADOW_USER_MASK (1ULL << (PT_SHADOW_USER_SHIFT))
 105
 106 #define PT_SHADOW_BITS_OFFSET (PT_SHADOW_WRITABLE_SHIFT - PT_WRITABLE_SHIFT)
 107
 108 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 109
 110 #define PT64_LEVEL_BITS 9
 111
 112 #define PT64_LEVEL_SHIFT(level) \
 113                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 114
 115 #define PT64_LEVEL_MASK(level) \
 116                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 117
 118 #define PT64_INDEX(address, level)\
 119         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 120
 121
 122 #define PT32_LEVEL_BITS 10
 123
 124 #define PT32_LEVEL_SHIFT(level) \
 125                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 126
 127 #define PT32_LEVEL_MASK(level) \
 128                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 129
 130 #define PT32_INDEX(address, level)\
 131         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 132
 133
 134 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 135 #define PT64_DIR_BASE_ADDR_MASK \
 136         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 137
 138 #define PT32_BASE_ADDR_MASK PAGE_MASK
 139 #define PT32_DIR_BASE_ADDR_MASK \
 140         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 141
 142
 143 #define PFERR_PRESENT_MASK (1U << 0)
 144 #define PFERR_WRITE_MASK (1U << 1)
 145 #define PFERR_USER_MASK (1U << 2)
 146 #define PFERR_FETCH_MASK (1U << 4)
 147
 148 #define PT64_ROOT_LEVEL 4
 149 #define PT32_ROOT_LEVEL 2
 150 #define PT32E_ROOT_LEVEL 3
 151
 152 #define PT_DIRECTORY_LEVEL 2
 153 #define PT_PAGE_TABLE_LEVEL 1
 154
 155 #define RMAP_EXT 4
 156
 157 struct kvm_rmap_desc {
 158         u64 *shadow_ptes[RMAP_EXT];
 159         struct kvm_rmap_desc *more;
 160 };
 161
 162 static struct kmem_cache *pte_chain_cache;
 163 static struct kmem_cache *rmap_desc_cache;
 164
 165 static int is_write_protection(struct kvm_vcpu *vcpu)
 166 {
 167         return vcpu->cr0 & CR0_WP_MASK;
 168 }
 169
 170 static int is_cpuid_PSE36(void)
 171 {
 172         return 1;
 173 }
 174
 175 static int is_nx(struct kvm_vcpu *vcpu)
 176 {
 177         return vcpu->shadow_efer & EFER_NX;
 178 }
 179
 180 static int is_present_pte(unsigned long pte)
 181 {
 182         return pte & PT_PRESENT_MASK;
 183 }
 184
 185 static int is_writeble_pte(unsigned long pte)
 186 {
 187         return pte & PT_WRITABLE_MASK;
 188 }
 189
 190 static int is_io_pte(unsigned long pte)
 191 {
 192         return pte & PT_SHADOW_IO_MARK;
 193 }
 194
 195 static int is_rmap_pte(u64 pte)
 196 {
 197         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 198                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 199 }
 200
 201 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 202                                   struct kmem_cache *base_cache, int min,
 203                                   gfp_t gfp_flags)
 204 {
 205         void *obj;
 206
 207         if (cache->nobjs >= min)
 208                 return 0;
 209         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 210                 obj = kmem_cache_zalloc(base_cache, gfp_flags);
 211                 if (!obj)
 212                         return -ENOMEM;
 213                 cache->objects[cache->nobjs++] = obj;
 214         }
 215         return 0;
 216 }
 217
 218 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 219 {
 220         while (mc->nobjs)
 221                 kfree(mc->objects[--mc->nobjs]);
 222 }
 223
 224 static int __mmu_topup_memory_caches(struct kvm_vcpu *vcpu, gfp_t gfp_flags)
 225 {
 226         int r;
 227
 228         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 229                                    pte_chain_cache, 4, gfp_flags);
 230         if (r)
 231                 goto out;
 232         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 233                                    rmap_desc_cache, 1, gfp_flags);
 234 out:
 235         return r;
 236 }
 237
 238 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 239 {
 240         int r;
 241
 242         r = __mmu_topup_memory_caches(vcpu, GFP_NOWAIT);
 243         if (r < 0) {
 244                 spin_unlock(&vcpu->kvm->lock);
 245                 kvm_arch_ops->vcpu_put(vcpu);
 246                 r = __mmu_topup_memory_caches(vcpu, GFP_KERNEL);
 247                 kvm_arch_ops->vcpu_load(vcpu);
 248                 spin_lock(&vcpu->kvm->lock);
 249         }
 250         return r;
 251 }
 252
 253 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 254 {
 255         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 256         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 257 }
 258
 259 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 260                                     size_t size)
 261 {
 262         void *p;
 263
 264         BUG_ON(!mc->nobjs);
 265         p = mc->objects[--mc->nobjs];
 266         memset(p, 0, size);
 267         return p;
 268 }
 269
 270 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
 271 {
 272         if (mc->nobjs < KVM_NR_MEM_OBJS)
 273                 mc->objects[mc->nobjs++] = obj;
 274         else
 275                 kfree(obj);
 276 }
 277
 278 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 279 {
 280         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 281                                       sizeof(struct kvm_pte_chain));
 282 }
 283
 284 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
 285                                struct kvm_pte_chain *pc)
 286 {
 287         mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
 288 }
 289
 290 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 291 {
 292         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 293                                       sizeof(struct kvm_rmap_desc));
 294 }
 295
 296 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
 297                                struct kvm_rmap_desc *rd)
 298 {
 299         mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
 300 }
 301
 302 /*
 303  * Reverse mapping data structures:
 304  *
 305  * If page->private bit zero is zero, then page->private points to the
 306  * shadow page table entry that points to page_address(page).
 307  *
 308  * If page->private bit zero is one, (then page->private & ~1) points
 309  * to a struct kvm_rmap_desc containing more mappings.
 310  */
 311 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 312 {
 313         struct page *page;
 314         struct kvm_rmap_desc *desc;
 315         int i;
 316
 317         if (!is_rmap_pte(*spte))
 318                 return;
 319         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 320         if (!page_private(page)) {
 321                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 322                 set_page_private(page,(unsigned long)spte);
 323         } else if (!(page_private(page) & 1)) {
 324                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 325                 desc = mmu_alloc_rmap_desc(vcpu);
 326                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 327                 desc->shadow_ptes[1] = spte;
 328                 set_page_private(page,(unsigned long)desc | 1);
 329         } else {
 330                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 331                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 332                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 333                         desc = desc->more;
 334                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 335                         desc->more = mmu_alloc_rmap_desc(vcpu);
 336                         desc = desc->more;
 337                 }
 338                 for (i = 0; desc->shadow_ptes[i]; ++i)
 339                         ;
 340                 desc->shadow_ptes[i] = spte;
 341         }
 342 }
 343
 344 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
 345                                    struct page *page,
 346                                    struct kvm_rmap_desc *desc,
 347                                    int i,
 348                                    struct kvm_rmap_desc *prev_desc)
 349 {
 350         int j;
 351
 352         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 353                 ;
 354         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 355         desc->shadow_ptes[j] = NULL;
 356         if (j != 0)
 357                 return;
 358         if (!prev_desc && !desc->more)
 359                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 360         else
 361                 if (prev_desc)
 362                         prev_desc->more = desc->more;
 363                 else
 364                         set_page_private(page,(unsigned long)desc->more | 1);
 365         mmu_free_rmap_desc(vcpu, desc);
 366 }
 367
 368 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
 369 {
 370         struct page *page;
 371         struct kvm_rmap_desc *desc;
 372         struct kvm_rmap_desc *prev_desc;
 373         int i;
 374
 375         if (!is_rmap_pte(*spte))
 376                 return;
 377         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 378         if (!page_private(page)) {
 379                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 380                 BUG();
 381         } else if (!(page_private(page) & 1)) {
 382                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 383                 if ((u64 *)page_private(page) != spte) {
 384                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 385                                spte, *spte);
 386                         BUG();
 387                 }
 388                 set_page_private(page,0);
 389         } else {
 390                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 391                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 392                 prev_desc = NULL;
 393                 while (desc) {
 394                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 395                                 if (desc->shadow_ptes[i] == spte) {
 396                                         rmap_desc_remove_entry(vcpu, page,
 397                                                                desc, i,
 398                                                                prev_desc);
 399                                         return;
 400                                 }
 401                         prev_desc = desc;
 402                         desc = desc->more;
 403                 }
 404                 BUG();
 405         }
 406 }
 407
 408 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 409 {
 410         struct kvm *kvm = vcpu->kvm;
 411         struct page *page;
 412         struct kvm_rmap_desc *desc;
 413         u64 *spte;
 414
 415         page = gfn_to_page(kvm, gfn);
 416         BUG_ON(!page);
 417
 418         while (page_private(page)) {
 419                 if (!(page_private(page) & 1))
 420                         spte = (u64 *)page_private(page);
 421                 else {
 422                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 423                         spte = desc->shadow_ptes[0];
 424                 }
 425                 BUG_ON(!spte);
 426                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 427                        != page_to_pfn(page));
 428                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 429                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 430                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 431                 rmap_remove(vcpu, spte);
 432                 kvm_arch_ops->tlb_flush(vcpu);
 433                 *spte &= ~(u64)PT_WRITABLE_MASK;
 434         }
 435 }
 436
 437 static int is_empty_shadow_page(hpa_t page_hpa)
 438 {
 439         u64 *pos;
 440         u64 *end;
 441
 442         for (pos = __va(page_hpa), end = pos + PAGE_SIZE / sizeof(u64);
 443                       pos != end; pos++)
 444                 if (*pos != 0) {
 445                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 446                                pos, *pos);
 447                         return 0;
 448                 }
 449         return 1;
 450 }
 451
 452 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu, hpa_t page_hpa)
 453 {
 454         struct kvm_mmu_page *page_head = page_header(page_hpa);
 455
 456         ASSERT(is_empty_shadow_page(page_hpa));
 457         page_head->page_hpa = page_hpa;
 458         list_move(&page_head->link, &vcpu->free_pages);
 459         ++vcpu->kvm->n_free_mmu_pages;
 460 }
 461
 462 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 463 {
 464         return gfn;
 465 }
 466
 467 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 468                                                u64 *parent_pte)
 469 {
 470         struct kvm_mmu_page *page;
 471
 472         if (list_empty(&vcpu->free_pages))
 473                 return NULL;
 474
 475         page = list_entry(vcpu->free_pages.next, struct kvm_mmu_page, link);
 476         list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 477         ASSERT(is_empty_shadow_page(page->page_hpa));
 478         page->slot_bitmap = 0;
 479         page->multimapped = 0;
 480         page->parent_pte = parent_pte;
 481         --vcpu->kvm->n_free_mmu_pages;
 482         return page;
 483 }
 484
 485 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 486                                     struct kvm_mmu_page *page, u64 *parent_pte)
 487 {
 488         struct kvm_pte_chain *pte_chain;
 489         struct hlist_node *node;
 490         int i;
 491
 492         if (!parent_pte)
 493                 return;
 494         if (!page->multimapped) {
 495                 u64 *old = page->parent_pte;
 496
 497                 if (!old) {
 498                         page->parent_pte = parent_pte;
 499                         return;
 500                 }
 501                 page->multimapped = 1;
 502                 pte_chain = mmu_alloc_pte_chain(vcpu);
 503                 INIT_HLIST_HEAD(&page->parent_ptes);
 504                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 505                 pte_chain->parent_ptes[0] = old;
 506         }
 507         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 508                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 509                         continue;
 510                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 511                         if (!pte_chain->parent_ptes[i]) {
 512                                 pte_chain->parent_ptes[i] = parent_pte;
 513                                 return;
 514                         }
 515         }
 516         pte_chain = mmu_alloc_pte_chain(vcpu);
 517         BUG_ON(!pte_chain);
 518         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 519         pte_chain->parent_ptes[0] = parent_pte;
 520 }
 521
 522 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
 523                                        struct kvm_mmu_page *page,
 524                                        u64 *parent_pte)
 525 {
 526         struct kvm_pte_chain *pte_chain;
 527         struct hlist_node *node;
 528         int i;
 529
 530         if (!page->multimapped) {
 531                 BUG_ON(page->parent_pte != parent_pte);
 532                 page->parent_pte = NULL;
 533                 return;
 534         }
 535         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 536                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 537                         if (!pte_chain->parent_ptes[i])
 538                                 break;
 539                         if (pte_chain->parent_ptes[i] != parent_pte)
 540                                 continue;
 541                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 542                                 && pte_chain->parent_ptes[i + 1]) {
 543                                 pte_chain->parent_ptes[i]
 544                                         = pte_chain->parent_ptes[i + 1];
 545                                 ++i;
 546                         }
 547                         pte_chain->parent_ptes[i] = NULL;
 548                         if (i == 0) {
 549                                 hlist_del(&pte_chain->link);
 550                                 mmu_free_pte_chain(vcpu, pte_chain);
 551                                 if (hlist_empty(&page->parent_ptes)) {
 552                                         page->multimapped = 0;
 553                                         page->parent_pte = NULL;
 554                                 }
 555                         }
 556                         return;
 557                 }
 558         BUG();
 559 }
 560
 561 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 562                                                 gfn_t gfn)
 563 {
 564         unsigned index;
 565         struct hlist_head *bucket;
 566         struct kvm_mmu_page *page;
 567         struct hlist_node *node;
 568
 569         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 570         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 571         bucket = &vcpu->kvm->mmu_page_hash[index];
 572         hlist_for_each_entry(page, node, bucket, hash_link)
 573                 if (page->gfn == gfn && !page->role.metaphysical) {
 574                         pgprintk("%s: found role %x\n",
 575                                  __FUNCTION__, page->role.word);
 576                         return page;
 577                 }
 578         return NULL;
 579 }
 580
 581 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 582                                              gfn_t gfn,
 583                                              gva_t gaddr,
 584                                              unsigned level,
 585                                              int metaphysical,
 586                                              unsigned hugepage_access,
 587                                              u64 *parent_pte)
 588 {
 589         union kvm_mmu_page_role role;
 590         unsigned index;
 591         unsigned quadrant;
 592         struct hlist_head *bucket;
 593         struct kvm_mmu_page *page;
 594         struct hlist_node *node;
 595
 596         role.word = 0;
 597         role.glevels = vcpu->mmu.root_level;
 598         role.level = level;
 599         role.metaphysical = metaphysical;
 600         role.hugepage_access = hugepage_access;
 601         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 602                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 603                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 604                 role.quadrant = quadrant;
 605         }
 606         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 607                  gfn, role.word);
 608         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 609         bucket = &vcpu->kvm->mmu_page_hash[index];
 610         hlist_for_each_entry(page, node, bucket, hash_link)
 611                 if (page->gfn == gfn && page->role.word == role.word) {
 612                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 613                         pgprintk("%s: found\n", __FUNCTION__);
 614                         return page;
 615                 }
 616         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 617         if (!page)
 618                 return page;
 619         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 620         page->gfn = gfn;
 621         page->role = role;
 622         hlist_add_head(&page->hash_link, bucket);
 623         if (!metaphysical)
 624                 rmap_write_protect(vcpu, gfn);
 625         return page;
 626 }
 627
 628 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 629                                          struct kvm_mmu_page *page)
 630 {
 631         unsigned i;
 632         u64 *pt;
 633         u64 ent;
 634
 635         pt = __va(page->page_hpa);
 636
 637         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 638                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 639                         if (pt[i] & PT_PRESENT_MASK)
 640                                 rmap_remove(vcpu, &pt[i]);
 641                         pt[i] = 0;
 642                 }
 643                 kvm_arch_ops->tlb_flush(vcpu);
 644                 return;
 645         }
 646
 647         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 648                 ent = pt[i];
 649
 650                 pt[i] = 0;
 651                 if (!(ent & PT_PRESENT_MASK))
 652                         continue;
 653                 ent &= PT64_BASE_ADDR_MASK;
 654                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
 655         }
 656 }
 657
 658 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 659                              struct kvm_mmu_page *page,
 660                              u64 *parent_pte)
 661 {
 662         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
 663 }
 664
 665 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 666                              struct kvm_mmu_page *page)
 667 {
 668         u64 *parent_pte;
 669
 670         while (page->multimapped || page->parent_pte) {
 671                 if (!page->multimapped)
 672                         parent_pte = page->parent_pte;
 673                 else {
 674                         struct kvm_pte_chain *chain;
 675
 676                         chain = container_of(page->parent_ptes.first,
 677                                              struct kvm_pte_chain, link);
 678                         parent_pte = chain->parent_ptes[0];
 679                 }
 680                 BUG_ON(!parent_pte);
 681                 kvm_mmu_put_page(vcpu, page, parent_pte);
 682                 *parent_pte = 0;
 683         }
 684         kvm_mmu_page_unlink_children(vcpu, page);
 685         if (!page->root_count) {
 686                 hlist_del(&page->hash_link);
 687                 kvm_mmu_free_page(vcpu, page->page_hpa);
 688         } else
 689                 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 690 }
 691
 692 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 693 {
 694         unsigned index;
 695         struct hlist_head *bucket;
 696         struct kvm_mmu_page *page;
 697         struct hlist_node *node, *n;
 698         int r;
 699
 700         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 701         r = 0;
 702         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 703         bucket = &vcpu->kvm->mmu_page_hash[index];
 704         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 705                 if (page->gfn == gfn && !page->role.metaphysical) {
 706                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 707                                  page->role.word);
 708                         kvm_mmu_zap_page(vcpu, page);
 709                         r = 1;
 710                 }
 711         return r;
 712 }
 713
 714 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 715 {
 716         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 717         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 718
 719         __set_bit(slot, &page_head->slot_bitmap);
 720 }
 721
 722 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 723 {
 724         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 725
 726         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 727 }
 728
 729 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 730 {
 731         struct page *page;
 732
 733         ASSERT((gpa & HPA_ERR_MASK) == 0);
 734         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 735         if (!page)
 736                 return gpa | HPA_ERR_MASK;
 737         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 738                 | (gpa & (PAGE_SIZE-1));
 739 }
 740
 741 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 742 {
 743         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 744
 745         if (gpa == UNMAPPED_GVA)
 746                 return UNMAPPED_GVA;
 747         return gpa_to_hpa(vcpu, gpa);
 748 }
 749
 750 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 751 {
 752         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 753
 754         if (gpa == UNMAPPED_GVA)
 755                 return NULL;
 756         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 757 }
 758
 759 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 760 {
 761 }
 762
 763 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 764 {
 765         int level = PT32E_ROOT_LEVEL;
 766         hpa_t table_addr = vcpu->mmu.root_hpa;
 767
 768         for (; ; level--) {
 769                 u32 index = PT64_INDEX(v, level);
 770                 u64 *table;
 771                 u64 pte;
 772
 773                 ASSERT(VALID_PAGE(table_addr));
 774                 table = __va(table_addr);
 775
 776                 if (level == 1) {
 777                         pte = table[index];
 778                         if (is_present_pte(pte) && is_writeble_pte(pte))
 779                                 return 0;
 780                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 781                         page_header_update_slot(vcpu->kvm, table, v);
 782                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 783                                                                 PT_USER_MASK;
 784                         rmap_add(vcpu, &table[index]);
 785                         return 0;
 786                 }
 787
 788                 if (table[index] == 0) {
 789                         struct kvm_mmu_page *new_table;
 790                         gfn_t pseudo_gfn;
 791
 792                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 793                                 >> PAGE_SHIFT;
 794                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 795                                                      v, level - 1,
 796                                                      1, 0, &table[index]);
 797                         if (!new_table) {
 798                                 pgprintk("nonpaging_map: ENOMEM\n");
 799                                 return -ENOMEM;
 800                         }
 801
 802                         table[index] = new_table->page_hpa | PT_PRESENT_MASK
 803                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 804                 }
 805                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 806         }
 807 }
 808
 809 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 810 {
 811         int i;
 812         struct kvm_mmu_page *page;
 813
 814 #ifdef CONFIG_X86_64
 815         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 816                 hpa_t root = vcpu->mmu.root_hpa;
 817
 818                 ASSERT(VALID_PAGE(root));
 819                 page = page_header(root);
 820                 --page->root_count;
 821                 vcpu->mmu.root_hpa = INVALID_PAGE;
 822                 return;
 823         }
 824 #endif
 825         for (i = 0; i < 4; ++i) {
 826                 hpa_t root = vcpu->mmu.pae_root[i];
 827
 828                 if (root) {
 829                         ASSERT(VALID_PAGE(root));
 830                         root &= PT64_BASE_ADDR_MASK;
 831                         page = page_header(root);
 832                         --page->root_count;
 833                 }
 834                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 835         }
 836         vcpu->mmu.root_hpa = INVALID_PAGE;
 837 }
 838
 839 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 840 {
 841         int i;
 842         gfn_t root_gfn;
 843         struct kvm_mmu_page *page;
 844
 845         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 846
 847 #ifdef CONFIG_X86_64
 848         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 849                 hpa_t root = vcpu->mmu.root_hpa;
 850
 851                 ASSERT(!VALID_PAGE(root));
 852                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 853                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 854                 root = page->page_hpa;
 855                 ++page->root_count;
 856                 vcpu->mmu.root_hpa = root;
 857                 return;
 858         }
 859 #endif
 860         for (i = 0; i < 4; ++i) {
 861                 hpa_t root = vcpu->mmu.pae_root[i];
 862
 863                 ASSERT(!VALID_PAGE(root));
 864                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 865                         if (!is_present_pte(vcpu->pdptrs[i])) {
 866                                 vcpu->mmu.pae_root[i] = 0;
 867                                 continue;
 868                         }
 869                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 870                 } else if (vcpu->mmu.root_level == 0)
 871                         root_gfn = 0;
 872                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 873                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 874                                         0, NULL);
 875                 root = page->page_hpa;
 876                 ++page->root_count;
 877                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 878         }
 879         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 880 }
 881
 882 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 883 {
 884         return vaddr;
 885 }
 886
 887 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 888                                u32 error_code)
 889 {
 890         gpa_t addr = gva;
 891         hpa_t paddr;
 892         int r;
 893
 894         r = mmu_topup_memory_caches(vcpu);
 895         if (r)
 896                 return r;
 897
 898         ASSERT(vcpu);
 899         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 900
 901
 902         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 903
 904         if (is_error_hpa(paddr))
 905                 return 1;
 906
 907         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 908 }
 909
 910 static void nonpaging_free(struct kvm_vcpu *vcpu)
 911 {
 912         mmu_free_roots(vcpu);
 913 }
 914
 915 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 916 {
 917         struct kvm_mmu *context = &vcpu->mmu;
 918
 919         context->new_cr3 = nonpaging_new_cr3;
 920         context->page_fault = nonpaging_page_fault;
 921         context->gva_to_gpa = nonpaging_gva_to_gpa;
 922         context->free = nonpaging_free;
 923         context->root_level = 0;
 924         context->shadow_root_level = PT32E_ROOT_LEVEL;
 925         mmu_alloc_roots(vcpu);
 926         ASSERT(VALID_PAGE(context->root_hpa));
 927         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 928         return 0;
 929 }
 930
 931 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 932 {
 933         ++kvm_stat.tlb_flush;
 934         kvm_arch_ops->tlb_flush(vcpu);
 935 }
 936
 937 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 938 {
 939         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 940         mmu_free_roots(vcpu);
 941         if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
 942                 kvm_mmu_free_some_pages(vcpu);
 943         mmu_alloc_roots(vcpu);
 944         kvm_mmu_flush_tlb(vcpu);
 945         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 946 }
 947
 948 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 949                              u64 *shadow_pte,
 950                              gpa_t gaddr,
 951                              int dirty,
 952                              u64 access_bits,
 953                              gfn_t gfn)
 954 {
 955         hpa_t paddr;
 956
 957         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 958         if (!dirty)
 959                 access_bits &= ~PT_WRITABLE_MASK;
 960
 961         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 962
 963         *shadow_pte |= access_bits;
 964
 965         if (is_error_hpa(paddr)) {
 966                 *shadow_pte |= gaddr;
 967                 *shadow_pte |= PT_SHADOW_IO_MARK;
 968                 *shadow_pte &= ~PT_PRESENT_MASK;
 969                 return;
 970         }
 971
 972         *shadow_pte |= paddr;
 973
 974         if (access_bits & PT_WRITABLE_MASK) {
 975                 struct kvm_mmu_page *shadow;
 976
 977                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 978                 if (shadow) {
 979                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 980                                  __FUNCTION__, gfn);
 981                         access_bits &= ~PT_WRITABLE_MASK;
 982                         if (is_writeble_pte(*shadow_pte)) {
 983                                     *shadow_pte &= ~PT_WRITABLE_MASK;
 984                                     kvm_arch_ops->tlb_flush(vcpu);
 985                         }
 986                 }
 987         }
 988
 989         if (access_bits & PT_WRITABLE_MASK)
 990                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 991
 992         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 993         rmap_add(vcpu, shadow_pte);
 994 }
 995
 996 static void inject_page_fault(struct kvm_vcpu *vcpu,
 997                               u64 addr,
 998                               u32 err_code)
 999 {
1000         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
1001 }
1002
1003 static inline int fix_read_pf(u64 *shadow_ent)
1004 {
1005         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
1006             !(*shadow_ent & PT_USER_MASK)) {
1007                 /*
1008                  * If supervisor write protect is disabled, we shadow kernel
1009                  * pages as user pages so we can trap the write access.
1010                  */
1011                 *shadow_ent |= PT_USER_MASK;
1012                 *shadow_ent &= ~PT_WRITABLE_MASK;
1013
1014                 return 1;
1015
1016         }
1017         return 0;
1018 }
1019
1020 static void paging_free(struct kvm_vcpu *vcpu)
1021 {
1022         nonpaging_free(vcpu);
1023 }
1024
1025 #define PTTYPE 64
1026 #include "paging_tmpl.h"
1027 #undef PTTYPE
1028
1029 #define PTTYPE 32
1030 #include "paging_tmpl.h"
1031 #undef PTTYPE
1032
1033 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1034 {
1035         struct kvm_mmu *context = &vcpu->mmu;
1036
1037         ASSERT(is_pae(vcpu));
1038         context->new_cr3 = paging_new_cr3;
1039         context->page_fault = paging64_page_fault;
1040         context->gva_to_gpa = paging64_gva_to_gpa;
1041         context->free = paging_free;
1042         context->root_level = level;
1043         context->shadow_root_level = level;
1044         mmu_alloc_roots(vcpu);
1045         ASSERT(VALID_PAGE(context->root_hpa));
1046         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1047                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1048         return 0;
1049 }
1050
1051 static int paging64_init_context(struct kvm_vcpu *vcpu)
1052 {
1053         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1054 }
1055
1056 static int paging32_init_context(struct kvm_vcpu *vcpu)
1057 {
1058         struct kvm_mmu *context = &vcpu->mmu;
1059
1060         context->new_cr3 = paging_new_cr3;
1061         context->page_fault = paging32_page_fault;
1062         context->gva_to_gpa = paging32_gva_to_gpa;
1063         context->free = paging_free;
1064         context->root_level = PT32_ROOT_LEVEL;
1065         context->shadow_root_level = PT32E_ROOT_LEVEL;
1066         mmu_alloc_roots(vcpu);
1067         ASSERT(VALID_PAGE(context->root_hpa));
1068         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1069                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1070         return 0;
1071 }
1072
1073 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1074 {
1075         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1076 }
1077
1078 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1079 {
1080         ASSERT(vcpu);
1081         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1082
1083         if (!is_paging(vcpu))
1084                 return nonpaging_init_context(vcpu);
1085         else if (is_long_mode(vcpu))
1086                 return paging64_init_context(vcpu);
1087         else if (is_pae(vcpu))
1088                 return paging32E_init_context(vcpu);
1089         else
1090                 return paging32_init_context(vcpu);
1091 }
1092
1093 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1094 {
1095         ASSERT(vcpu);
1096         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1097                 vcpu->mmu.free(vcpu);
1098                 vcpu->mmu.root_hpa = INVALID_PAGE;
1099         }
1100 }
1101
1102 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1103 {
1104         int r;
1105
1106         destroy_kvm_mmu(vcpu);
1107         r = init_kvm_mmu(vcpu);
1108         if (r < 0)
1109                 goto out;
1110         r = mmu_topup_memory_caches(vcpu);
1111 out:
1112         return r;
1113 }
1114
1115 static void mmu_pre_write_zap_pte(struct kvm_vcpu *vcpu,
1116                                   struct kvm_mmu_page *page,
1117                                   u64 *spte)
1118 {
1119         u64 pte;
1120         struct kvm_mmu_page *child;
1121
1122         pte = *spte;
1123         if (is_present_pte(pte)) {
1124                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1125                         rmap_remove(vcpu, spte);
1126                 else {
1127                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1128                         mmu_page_remove_parent_pte(vcpu, child, spte);
1129                 }
1130         }
1131         *spte = 0;
1132 }
1133
1134 void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1135 {
1136         gfn_t gfn = gpa >> PAGE_SHIFT;
1137         struct kvm_mmu_page *page;
1138         struct hlist_node *node, *n;
1139         struct hlist_head *bucket;
1140         unsigned index;
1141         u64 *spte;
1142         unsigned offset = offset_in_page(gpa);
1143         unsigned pte_size;
1144         unsigned page_offset;
1145         unsigned misaligned;
1146         int level;
1147         int flooded = 0;
1148         int npte;
1149
1150         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1151         if (gfn == vcpu->last_pt_write_gfn) {
1152                 ++vcpu->last_pt_write_count;
1153                 if (vcpu->last_pt_write_count >= 3)
1154                         flooded = 1;
1155         } else {
1156                 vcpu->last_pt_write_gfn = gfn;
1157                 vcpu->last_pt_write_count = 1;
1158         }
1159         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1160         bucket = &vcpu->kvm->mmu_page_hash[index];
1161         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1162                 if (page->gfn != gfn || page->role.metaphysical)
1163                         continue;
1164                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1165                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1166                 if (misaligned || flooded) {
1167                         /*
1168                          * Misaligned accesses are too much trouble to fix
1169                          * up; also, they usually indicate a page is not used
1170                          * as a page table.
1171                          *
1172                          * If we're seeing too many writes to a page,
1173                          * it may no longer be a page table, or we may be
1174                          * forking, in which case it is better to unmap the
1175                          * page.
1176                          */
1177                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1178                                  gpa, bytes, page->role.word);
1179                         kvm_mmu_zap_page(vcpu, page);
1180                         continue;
1181                 }
1182                 page_offset = offset;
1183                 level = page->role.level;
1184                 npte = 1;
1185                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1186                         page_offset <<= 1;      /* 32->64 */
1187                         /*
1188                          * A 32-bit pde maps 4MB while the shadow pdes map
1189                          * only 2MB.  So we need to double the offset again
1190                          * and zap two pdes instead of one.
1191                          */
1192                         if (level == PT32_ROOT_LEVEL) {
1193                                 page_offset &= ~7; /* kill rounding error */
1194                                 page_offset <<= 1;
1195                                 npte = 2;
1196                         }
1197                         page_offset &= ~PAGE_MASK;
1198                 }
1199                 spte = __va(page->page_hpa);
1200                 spte += page_offset / sizeof(*spte);
1201                 while (npte--) {
1202                         mmu_pre_write_zap_pte(vcpu, page, spte);
1203                         ++spte;
1204                 }
1205         }
1206 }
1207
1208 void kvm_mmu_post_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1209 {
1210 }
1211
1212 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1213 {
1214         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1215
1216         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1217 }
1218
1219 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1220 {
1221         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1222                 struct kvm_mmu_page *page;
1223
1224                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1225                                     struct kvm_mmu_page, link);
1226                 kvm_mmu_zap_page(vcpu, page);
1227         }
1228 }
1229 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1230
1231 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1232 {
1233         struct kvm_mmu_page *page;
1234
1235         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1236                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1237                                     struct kvm_mmu_page, link);
1238                 kvm_mmu_zap_page(vcpu, page);
1239         }
1240         while (!list_empty(&vcpu->free_pages)) {
1241                 page = list_entry(vcpu->free_pages.next,
1242                                   struct kvm_mmu_page, link);
1243                 list_del(&page->link);
1244                 __free_page(pfn_to_page(page->page_hpa >> PAGE_SHIFT));
1245                 page->page_hpa = INVALID_PAGE;
1246         }
1247         free_page((unsigned long)vcpu->mmu.pae_root);
1248 }
1249
1250 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1251 {
1252         struct page *page;
1253         int i;
1254
1255         ASSERT(vcpu);
1256
1257         for (i = 0; i < KVM_NUM_MMU_PAGES; i++) {
1258                 struct kvm_mmu_page *page_header = &vcpu->page_header_buf[i];
1259
1260                 INIT_LIST_HEAD(&page_header->link);
1261                 if ((page = alloc_page(GFP_KERNEL)) == NULL)
1262                         goto error_1;
1263                 set_page_private(page, (unsigned long)page_header);
1264                 page_header->page_hpa = (hpa_t)page_to_pfn(page) << PAGE_SHIFT;
1265                 memset(__va(page_header->page_hpa), 0, PAGE_SIZE);
1266                 list_add(&page_header->link, &vcpu->free_pages);
1267                 ++vcpu->kvm->n_free_mmu_pages;
1268         }
1269
1270         /*
1271          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1272          * Therefore we need to allocate shadow page tables in the first
1273          * 4GB of memory, which happens to fit the DMA32 zone.
1274          */
1275         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1276         if (!page)
1277                 goto error_1;
1278         vcpu->mmu.pae_root = page_address(page);
1279         for (i = 0; i < 4; ++i)
1280                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1281
1282         return 0;
1283
1284 error_1:
1285         free_mmu_pages(vcpu);
1286         return -ENOMEM;
1287 }
1288
1289 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1290 {
1291         ASSERT(vcpu);
1292         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1293         ASSERT(list_empty(&vcpu->free_pages));
1294
1295         return alloc_mmu_pages(vcpu);
1296 }
1297
1298 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1299 {
1300         ASSERT(vcpu);
1301         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1302         ASSERT(!list_empty(&vcpu->free_pages));
1303
1304         return init_kvm_mmu(vcpu);
1305 }
1306
1307 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1308 {
1309         ASSERT(vcpu);
1310
1311         destroy_kvm_mmu(vcpu);
1312         free_mmu_pages(vcpu);
1313         mmu_free_memory_caches(vcpu);
1314 }
1315
1316 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1317 {
1318         struct kvm *kvm = vcpu->kvm;
1319         struct kvm_mmu_page *page;
1320
1321         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1322                 int i;
1323                 u64 *pt;
1324
1325                 if (!test_bit(slot, &page->slot_bitmap))
1326                         continue;
1327
1328                 pt = __va(page->page_hpa);
1329                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1330                         /* avoid RMW */
1331                         if (pt[i] & PT_WRITABLE_MASK) {
1332                                 rmap_remove(vcpu, &pt[i]);
1333                                 pt[i] &= ~PT_WRITABLE_MASK;
1334                         }
1335         }
1336 }
1337
1338 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1339 {
1340         destroy_kvm_mmu(vcpu);
1341
1342         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1343                 struct kvm_mmu_page *page;
1344
1345                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1346                                     struct kvm_mmu_page, link);
1347                 kvm_mmu_zap_page(vcpu, page);
1348         }
1349
1350         mmu_free_memory_caches(vcpu);
1351         kvm_arch_ops->tlb_flush(vcpu);
1352         init_kvm_mmu(vcpu);
1353 }
1354
1355 void kvm_mmu_module_exit(void)
1356 {
1357         if (pte_chain_cache)
1358                 kmem_cache_destroy(pte_chain_cache);
1359         if (rmap_desc_cache)
1360                 kmem_cache_destroy(rmap_desc_cache);
1361 }
1362
1363 int kvm_mmu_module_init(void)
1364 {
1365         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1366                                             sizeof(struct kvm_pte_chain),
1367                                             0, 0, NULL, NULL);
1368         if (!pte_chain_cache)
1369                 goto nomem;
1370         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1371                                             sizeof(struct kvm_rmap_desc),
1372                                             0, 0, NULL, NULL);
1373         if (!rmap_desc_cache)
1374                 goto nomem;
1375
1376         return 0;
1377
1378 nomem:
1379         kvm_mmu_module_exit();
1380         return -ENOMEM;
1381 }
1382
1383 #ifdef AUDIT
1384
1385 static const char *audit_msg;
1386
1387 static gva_t canonicalize(gva_t gva)
1388 {
1389 #ifdef CONFIG_X86_64
1390         gva = (long long)(gva << 16) >> 16;
1391 #endif
1392         return gva;
1393 }
1394
1395 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1396                                 gva_t va, int level)
1397 {
1398         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1399         int i;
1400         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1401
1402         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1403                 u64 ent = pt[i];
1404
1405                 if (!ent & PT_PRESENT_MASK)
1406                         continue;
1407
1408                 va = canonicalize(va);
1409                 if (level > 1)
1410                         audit_mappings_page(vcpu, ent, va, level - 1);
1411                 else {
1412                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1413                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1414
1415                         if ((ent & PT_PRESENT_MASK)
1416                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1417                                 printk(KERN_ERR "audit error: (%s) levels %d"
1418                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1419                                        audit_msg, vcpu->mmu.root_level,
1420                                        va, gpa, hpa, ent);
1421                 }
1422         }
1423 }
1424
1425 static void audit_mappings(struct kvm_vcpu *vcpu)
1426 {
1427         unsigned i;
1428
1429         if (vcpu->mmu.root_level == 4)
1430                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1431         else
1432                 for (i = 0; i < 4; ++i)
1433                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1434                                 audit_mappings_page(vcpu,
1435                                                     vcpu->mmu.pae_root[i],
1436                                                     i << 30,
1437                                                     2);
1438 }
1439
1440 static int count_rmaps(struct kvm_vcpu *vcpu)
1441 {
1442         int nmaps = 0;
1443         int i, j, k;
1444
1445         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1446                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1447                 struct kvm_rmap_desc *d;
1448
1449                 for (j = 0; j < m->npages; ++j) {
1450                         struct page *page = m->phys_mem[j];
1451
1452                         if (!page->private)
1453                                 continue;
1454                         if (!(page->private & 1)) {
1455                                 ++nmaps;
1456                                 continue;
1457                         }
1458                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1459                         while (d) {
1460                                 for (k = 0; k < RMAP_EXT; ++k)
1461                                         if (d->shadow_ptes[k])
1462                                                 ++nmaps;
1463                                         else
1464                                                 break;
1465                                 d = d->more;
1466                         }
1467                 }
1468         }
1469         return nmaps;
1470 }
1471
1472 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1473 {
1474         int nmaps = 0;
1475         struct kvm_mmu_page *page;
1476         int i;
1477
1478         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1479                 u64 *pt = __va(page->page_hpa);
1480
1481                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1482                         continue;
1483
1484                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1485                         u64 ent = pt[i];
1486
1487                         if (!(ent & PT_PRESENT_MASK))
1488                                 continue;
1489                         if (!(ent & PT_WRITABLE_MASK))
1490                                 continue;
1491                         ++nmaps;
1492                 }
1493         }
1494         return nmaps;
1495 }
1496
1497 static void audit_rmap(struct kvm_vcpu *vcpu)
1498 {
1499         int n_rmap = count_rmaps(vcpu);
1500         int n_actual = count_writable_mappings(vcpu);
1501
1502         if (n_rmap != n_actual)
1503                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1504                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1505 }
1506
1507 static void audit_write_protection(struct kvm_vcpu *vcpu)
1508 {
1509         struct kvm_mmu_page *page;
1510
1511         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1512                 hfn_t hfn;
1513                 struct page *pg;
1514
1515                 if (page->role.metaphysical)
1516                         continue;
1517
1518                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1519                         >> PAGE_SHIFT;
1520                 pg = pfn_to_page(hfn);
1521                 if (pg->private)
1522                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1523                                " mappings: gfn %lx role %x\n",
1524                                __FUNCTION__, audit_msg, page->gfn,
1525                                page->role.word);
1526         }
1527 }
1528
1529 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1530 {
1531         int olddbg = dbg;
1532
1533         dbg = 0;
1534         audit_msg = msg;
1535         audit_rmap(vcpu);
1536         audit_write_protection(vcpu);
1537         audit_mappings(vcpu);
1538         dbg = olddbg;
1539 }
1540
1541 #endif