Pileus Git - ~andy/linux/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 #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,
 459                               struct kvm_mmu_page *page_head)
 460 {
 461         ASSERT(is_empty_shadow_page(page_head->page_hpa));
 462         list_move(&page_head->link, &vcpu->free_pages);
 463         ++vcpu->kvm->n_free_mmu_pages;
 464 }
 465
 466 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 467 {
 468         return gfn;
 469 }
 470
 471 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 472                                                u64 *parent_pte)
 473 {
 474         struct kvm_mmu_page *page;
 475
 476         if (list_empty(&vcpu->free_pages))
 477                 return NULL;
 478
 479         page = list_entry(vcpu->free_pages.next, struct kvm_mmu_page, link);
 480         list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 481         ASSERT(is_empty_shadow_page(page->page_hpa));
 482         page->slot_bitmap = 0;
 483         page->multimapped = 0;
 484         page->parent_pte = parent_pte;
 485         --vcpu->kvm->n_free_mmu_pages;
 486         return page;
 487 }
 488
 489 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 490                                     struct kvm_mmu_page *page, u64 *parent_pte)
 491 {
 492         struct kvm_pte_chain *pte_chain;
 493         struct hlist_node *node;
 494         int i;
 495
 496         if (!parent_pte)
 497                 return;
 498         if (!page->multimapped) {
 499                 u64 *old = page->parent_pte;
 500
 501                 if (!old) {
 502                         page->parent_pte = parent_pte;
 503                         return;
 504                 }
 505                 page->multimapped = 1;
 506                 pte_chain = mmu_alloc_pte_chain(vcpu);
 507                 INIT_HLIST_HEAD(&page->parent_ptes);
 508                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 509                 pte_chain->parent_ptes[0] = old;
 510         }
 511         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 512                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 513                         continue;
 514                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 515                         if (!pte_chain->parent_ptes[i]) {
 516                                 pte_chain->parent_ptes[i] = parent_pte;
 517                                 return;
 518                         }
 519         }
 520         pte_chain = mmu_alloc_pte_chain(vcpu);
 521         BUG_ON(!pte_chain);
 522         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 523         pte_chain->parent_ptes[0] = parent_pte;
 524 }
 525
 526 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
 527                                        struct kvm_mmu_page *page,
 528                                        u64 *parent_pte)
 529 {
 530         struct kvm_pte_chain *pte_chain;
 531         struct hlist_node *node;
 532         int i;
 533
 534         if (!page->multimapped) {
 535                 BUG_ON(page->parent_pte != parent_pte);
 536                 page->parent_pte = NULL;
 537                 return;
 538         }
 539         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 540                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 541                         if (!pte_chain->parent_ptes[i])
 542                                 break;
 543                         if (pte_chain->parent_ptes[i] != parent_pte)
 544                                 continue;
 545                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 546                                 && pte_chain->parent_ptes[i + 1]) {
 547                                 pte_chain->parent_ptes[i]
 548                                         = pte_chain->parent_ptes[i + 1];
 549                                 ++i;
 550                         }
 551                         pte_chain->parent_ptes[i] = NULL;
 552                         if (i == 0) {
 553                                 hlist_del(&pte_chain->link);
 554                                 mmu_free_pte_chain(vcpu, pte_chain);
 555                                 if (hlist_empty(&page->parent_ptes)) {
 556                                         page->multimapped = 0;
 557                                         page->parent_pte = NULL;
 558                                 }
 559                         }
 560                         return;
 561                 }
 562         BUG();
 563 }
 564
 565 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 566                                                 gfn_t gfn)
 567 {
 568         unsigned index;
 569         struct hlist_head *bucket;
 570         struct kvm_mmu_page *page;
 571         struct hlist_node *node;
 572
 573         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 574         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 575         bucket = &vcpu->kvm->mmu_page_hash[index];
 576         hlist_for_each_entry(page, node, bucket, hash_link)
 577                 if (page->gfn == gfn && !page->role.metaphysical) {
 578                         pgprintk("%s: found role %x\n",
 579                                  __FUNCTION__, page->role.word);
 580                         return page;
 581                 }
 582         return NULL;
 583 }
 584
 585 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 586                                              gfn_t gfn,
 587                                              gva_t gaddr,
 588                                              unsigned level,
 589                                              int metaphysical,
 590                                              unsigned hugepage_access,
 591                                              u64 *parent_pte)
 592 {
 593         union kvm_mmu_page_role role;
 594         unsigned index;
 595         unsigned quadrant;
 596         struct hlist_head *bucket;
 597         struct kvm_mmu_page *page;
 598         struct hlist_node *node;
 599
 600         role.word = 0;
 601         role.glevels = vcpu->mmu.root_level;
 602         role.level = level;
 603         role.metaphysical = metaphysical;
 604         role.hugepage_access = hugepage_access;
 605         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 606                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 607                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 608                 role.quadrant = quadrant;
 609         }
 610         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 611                  gfn, role.word);
 612         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 613         bucket = &vcpu->kvm->mmu_page_hash[index];
 614         hlist_for_each_entry(page, node, bucket, hash_link)
 615                 if (page->gfn == gfn && page->role.word == role.word) {
 616                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 617                         pgprintk("%s: found\n", __FUNCTION__);
 618                         return page;
 619                 }
 620         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 621         if (!page)
 622                 return page;
 623         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 624         page->gfn = gfn;
 625         page->role = role;
 626         hlist_add_head(&page->hash_link, bucket);
 627         if (!metaphysical)
 628                 rmap_write_protect(vcpu, gfn);
 629         return page;
 630 }
 631
 632 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 633                                          struct kvm_mmu_page *page)
 634 {
 635         unsigned i;
 636         u64 *pt;
 637         u64 ent;
 638
 639         pt = __va(page->page_hpa);
 640
 641         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 642                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 643                         if (pt[i] & PT_PRESENT_MASK)
 644                                 rmap_remove(vcpu, &pt[i]);
 645                         pt[i] = 0;
 646                 }
 647                 kvm_arch_ops->tlb_flush(vcpu);
 648                 return;
 649         }
 650
 651         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 652                 ent = pt[i];
 653
 654                 pt[i] = 0;
 655                 if (!(ent & PT_PRESENT_MASK))
 656                         continue;
 657                 ent &= PT64_BASE_ADDR_MASK;
 658                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
 659         }
 660 }
 661
 662 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 663                              struct kvm_mmu_page *page,
 664                              u64 *parent_pte)
 665 {
 666         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
 667 }
 668
 669 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 670                              struct kvm_mmu_page *page)
 671 {
 672         u64 *parent_pte;
 673
 674         while (page->multimapped || page->parent_pte) {
 675                 if (!page->multimapped)
 676                         parent_pte = page->parent_pte;
 677                 else {
 678                         struct kvm_pte_chain *chain;
 679
 680                         chain = container_of(page->parent_ptes.first,
 681                                              struct kvm_pte_chain, link);
 682                         parent_pte = chain->parent_ptes[0];
 683                 }
 684                 BUG_ON(!parent_pte);
 685                 kvm_mmu_put_page(vcpu, page, parent_pte);
 686                 *parent_pte = 0;
 687         }
 688         kvm_mmu_page_unlink_children(vcpu, page);
 689         if (!page->root_count) {
 690                 hlist_del(&page->hash_link);
 691                 kvm_mmu_free_page(vcpu, page);
 692         } else
 693                 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 694 }
 695
 696 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 697 {
 698         unsigned index;
 699         struct hlist_head *bucket;
 700         struct kvm_mmu_page *page;
 701         struct hlist_node *node, *n;
 702         int r;
 703
 704         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 705         r = 0;
 706         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 707         bucket = &vcpu->kvm->mmu_page_hash[index];
 708         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 709                 if (page->gfn == gfn && !page->role.metaphysical) {
 710                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 711                                  page->role.word);
 712                         kvm_mmu_zap_page(vcpu, page);
 713                         r = 1;
 714                 }
 715         return r;
 716 }
 717
 718 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 719 {
 720         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 721         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 722
 723         __set_bit(slot, &page_head->slot_bitmap);
 724 }
 725
 726 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 727 {
 728         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 729
 730         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 731 }
 732
 733 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 734 {
 735         struct page *page;
 736
 737         ASSERT((gpa & HPA_ERR_MASK) == 0);
 738         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 739         if (!page)
 740                 return gpa | HPA_ERR_MASK;
 741         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 742                 | (gpa & (PAGE_SIZE-1));
 743 }
 744
 745 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 746 {
 747         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 748
 749         if (gpa == UNMAPPED_GVA)
 750                 return UNMAPPED_GVA;
 751         return gpa_to_hpa(vcpu, gpa);
 752 }
 753
 754 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 755 {
 756         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 757
 758         if (gpa == UNMAPPED_GVA)
 759                 return NULL;
 760         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 761 }
 762
 763 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 764 {
 765 }
 766
 767 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 768 {
 769         int level = PT32E_ROOT_LEVEL;
 770         hpa_t table_addr = vcpu->mmu.root_hpa;
 771
 772         for (; ; level--) {
 773                 u32 index = PT64_INDEX(v, level);
 774                 u64 *table;
 775                 u64 pte;
 776
 777                 ASSERT(VALID_PAGE(table_addr));
 778                 table = __va(table_addr);
 779
 780                 if (level == 1) {
 781                         pte = table[index];
 782                         if (is_present_pte(pte) && is_writeble_pte(pte))
 783                                 return 0;
 784                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 785                         page_header_update_slot(vcpu->kvm, table, v);
 786                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 787                                                                 PT_USER_MASK;
 788                         rmap_add(vcpu, &table[index]);
 789                         return 0;
 790                 }
 791
 792                 if (table[index] == 0) {
 793                         struct kvm_mmu_page *new_table;
 794                         gfn_t pseudo_gfn;
 795
 796                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 797                                 >> PAGE_SHIFT;
 798                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 799                                                      v, level - 1,
 800                                                      1, 0, &table[index]);
 801                         if (!new_table) {
 802                                 pgprintk("nonpaging_map: ENOMEM\n");
 803                                 return -ENOMEM;
 804                         }
 805
 806                         table[index] = new_table->page_hpa | PT_PRESENT_MASK
 807                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 808                 }
 809                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 810         }
 811 }
 812
 813 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 814 {
 815         int i;
 816         struct kvm_mmu_page *page;
 817
 818 #ifdef CONFIG_X86_64
 819         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 820                 hpa_t root = vcpu->mmu.root_hpa;
 821
 822                 ASSERT(VALID_PAGE(root));
 823                 page = page_header(root);
 824                 --page->root_count;
 825                 vcpu->mmu.root_hpa = INVALID_PAGE;
 826                 return;
 827         }
 828 #endif
 829         for (i = 0; i < 4; ++i) {
 830                 hpa_t root = vcpu->mmu.pae_root[i];
 831
 832                 if (root) {
 833                         ASSERT(VALID_PAGE(root));
 834                         root &= PT64_BASE_ADDR_MASK;
 835                         page = page_header(root);
 836                         --page->root_count;
 837                 }
 838                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 839         }
 840         vcpu->mmu.root_hpa = INVALID_PAGE;
 841 }
 842
 843 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 844 {
 845         int i;
 846         gfn_t root_gfn;
 847         struct kvm_mmu_page *page;
 848
 849         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 850
 851 #ifdef CONFIG_X86_64
 852         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 853                 hpa_t root = vcpu->mmu.root_hpa;
 854
 855                 ASSERT(!VALID_PAGE(root));
 856                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 857                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 858                 root = page->page_hpa;
 859                 ++page->root_count;
 860                 vcpu->mmu.root_hpa = root;
 861                 return;
 862         }
 863 #endif
 864         for (i = 0; i < 4; ++i) {
 865                 hpa_t root = vcpu->mmu.pae_root[i];
 866
 867                 ASSERT(!VALID_PAGE(root));
 868                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 869                         if (!is_present_pte(vcpu->pdptrs[i])) {
 870                                 vcpu->mmu.pae_root[i] = 0;
 871                                 continue;
 872                         }
 873                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 874                 } else if (vcpu->mmu.root_level == 0)
 875                         root_gfn = 0;
 876                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 877                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 878                                         0, NULL);
 879                 root = page->page_hpa;
 880                 ++page->root_count;
 881                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 882         }
 883         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 884 }
 885
 886 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 887 {
 888         return vaddr;
 889 }
 890
 891 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 892                                u32 error_code)
 893 {
 894         gpa_t addr = gva;
 895         hpa_t paddr;
 896         int r;
 897
 898         r = mmu_topup_memory_caches(vcpu);
 899         if (r)
 900                 return r;
 901
 902         ASSERT(vcpu);
 903         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 904
 905
 906         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 907
 908         if (is_error_hpa(paddr))
 909                 return 1;
 910
 911         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 912 }
 913
 914 static void nonpaging_free(struct kvm_vcpu *vcpu)
 915 {
 916         mmu_free_roots(vcpu);
 917 }
 918
 919 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 920 {
 921         struct kvm_mmu *context = &vcpu->mmu;
 922
 923         context->new_cr3 = nonpaging_new_cr3;
 924         context->page_fault = nonpaging_page_fault;
 925         context->gva_to_gpa = nonpaging_gva_to_gpa;
 926         context->free = nonpaging_free;
 927         context->root_level = 0;
 928         context->shadow_root_level = PT32E_ROOT_LEVEL;
 929         mmu_alloc_roots(vcpu);
 930         ASSERT(VALID_PAGE(context->root_hpa));
 931         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 932         return 0;
 933 }
 934
 935 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 936 {
 937         ++vcpu->stat.tlb_flush;
 938         kvm_arch_ops->tlb_flush(vcpu);
 939 }
 940
 941 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 942 {
 943         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 944         mmu_free_roots(vcpu);
 945         if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
 946                 kvm_mmu_free_some_pages(vcpu);
 947         mmu_alloc_roots(vcpu);
 948         kvm_mmu_flush_tlb(vcpu);
 949         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 950 }
 951
 952 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 953                              u64 *shadow_pte,
 954                              gpa_t gaddr,
 955                              int dirty,
 956                              u64 access_bits,
 957                              gfn_t gfn)
 958 {
 959         hpa_t paddr;
 960
 961         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 962         if (!dirty)
 963                 access_bits &= ~PT_WRITABLE_MASK;
 964
 965         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 966
 967         *shadow_pte |= access_bits;
 968
 969         if (is_error_hpa(paddr)) {
 970                 *shadow_pte |= gaddr;
 971                 *shadow_pte |= PT_SHADOW_IO_MARK;
 972                 *shadow_pte &= ~PT_PRESENT_MASK;
 973                 return;
 974         }
 975
 976         *shadow_pte |= paddr;
 977
 978         if (access_bits & PT_WRITABLE_MASK) {
 979                 struct kvm_mmu_page *shadow;
 980
 981                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 982                 if (shadow) {
 983                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 984                                  __FUNCTION__, gfn);
 985                         access_bits &= ~PT_WRITABLE_MASK;
 986                         if (is_writeble_pte(*shadow_pte)) {
 987                                     *shadow_pte &= ~PT_WRITABLE_MASK;
 988                                     kvm_arch_ops->tlb_flush(vcpu);
 989                         }
 990                 }
 991         }
 992
 993         if (access_bits & PT_WRITABLE_MASK)
 994                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 995
 996         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 997         rmap_add(vcpu, shadow_pte);
 998 }
 999
1000 static void inject_page_fault(struct kvm_vcpu *vcpu,
1001                               u64 addr,
1002                               u32 err_code)
1003 {
1004         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
1005 }
1006
1007 static inline int fix_read_pf(u64 *shadow_ent)
1008 {
1009         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
1010             !(*shadow_ent & PT_USER_MASK)) {
1011                 /*
1012                  * If supervisor write protect is disabled, we shadow kernel
1013                  * pages as user pages so we can trap the write access.
1014                  */
1015                 *shadow_ent |= PT_USER_MASK;
1016                 *shadow_ent &= ~PT_WRITABLE_MASK;
1017
1018                 return 1;
1019
1020         }
1021         return 0;
1022 }
1023
1024 static void paging_free(struct kvm_vcpu *vcpu)
1025 {
1026         nonpaging_free(vcpu);
1027 }
1028
1029 #define PTTYPE 64
1030 #include "paging_tmpl.h"
1031 #undef PTTYPE
1032
1033 #define PTTYPE 32
1034 #include "paging_tmpl.h"
1035 #undef PTTYPE
1036
1037 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1038 {
1039         struct kvm_mmu *context = &vcpu->mmu;
1040
1041         ASSERT(is_pae(vcpu));
1042         context->new_cr3 = paging_new_cr3;
1043         context->page_fault = paging64_page_fault;
1044         context->gva_to_gpa = paging64_gva_to_gpa;
1045         context->free = paging_free;
1046         context->root_level = level;
1047         context->shadow_root_level = level;
1048         mmu_alloc_roots(vcpu);
1049         ASSERT(VALID_PAGE(context->root_hpa));
1050         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1051                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1052         return 0;
1053 }
1054
1055 static int paging64_init_context(struct kvm_vcpu *vcpu)
1056 {
1057         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1058 }
1059
1060 static int paging32_init_context(struct kvm_vcpu *vcpu)
1061 {
1062         struct kvm_mmu *context = &vcpu->mmu;
1063
1064         context->new_cr3 = paging_new_cr3;
1065         context->page_fault = paging32_page_fault;
1066         context->gva_to_gpa = paging32_gva_to_gpa;
1067         context->free = paging_free;
1068         context->root_level = PT32_ROOT_LEVEL;
1069         context->shadow_root_level = PT32E_ROOT_LEVEL;
1070         mmu_alloc_roots(vcpu);
1071         ASSERT(VALID_PAGE(context->root_hpa));
1072         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1073                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1074         return 0;
1075 }
1076
1077 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1078 {
1079         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1080 }
1081
1082 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1083 {
1084         ASSERT(vcpu);
1085         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1086
1087         if (!is_paging(vcpu))
1088                 return nonpaging_init_context(vcpu);
1089         else if (is_long_mode(vcpu))
1090                 return paging64_init_context(vcpu);
1091         else if (is_pae(vcpu))
1092                 return paging32E_init_context(vcpu);
1093         else
1094                 return paging32_init_context(vcpu);
1095 }
1096
1097 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1098 {
1099         ASSERT(vcpu);
1100         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1101                 vcpu->mmu.free(vcpu);
1102                 vcpu->mmu.root_hpa = INVALID_PAGE;
1103         }
1104 }
1105
1106 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1107 {
1108         int r;
1109
1110         destroy_kvm_mmu(vcpu);
1111         r = init_kvm_mmu(vcpu);
1112         if (r < 0)
1113                 goto out;
1114         r = mmu_topup_memory_caches(vcpu);
1115 out:
1116         return r;
1117 }
1118
1119 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1120                                   struct kvm_mmu_page *page,
1121                                   u64 *spte)
1122 {
1123         u64 pte;
1124         struct kvm_mmu_page *child;
1125
1126         pte = *spte;
1127         if (is_present_pte(pte)) {
1128                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1129                         rmap_remove(vcpu, spte);
1130                 else {
1131                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1132                         mmu_page_remove_parent_pte(vcpu, child, spte);
1133                 }
1134         }
1135         *spte = 0;
1136 }
1137
1138 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1139                                   struct kvm_mmu_page *page,
1140                                   u64 *spte,
1141                                   const void *new, int bytes)
1142 {
1143         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1144                 return;
1145
1146         if (page->role.glevels == PT32_ROOT_LEVEL)
1147                 paging32_update_pte(vcpu, page, spte, new, bytes);
1148         else
1149                 paging64_update_pte(vcpu, page, spte, new, bytes);
1150 }
1151
1152 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1153                        const u8 *old, const u8 *new, int bytes)
1154 {
1155         gfn_t gfn = gpa >> PAGE_SHIFT;
1156         struct kvm_mmu_page *page;
1157         struct hlist_node *node, *n;
1158         struct hlist_head *bucket;
1159         unsigned index;
1160         u64 *spte;
1161         unsigned offset = offset_in_page(gpa);
1162         unsigned pte_size;
1163         unsigned page_offset;
1164         unsigned misaligned;
1165         unsigned quadrant;
1166         int level;
1167         int flooded = 0;
1168         int npte;
1169
1170         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1171         if (gfn == vcpu->last_pt_write_gfn) {
1172                 ++vcpu->last_pt_write_count;
1173                 if (vcpu->last_pt_write_count >= 3)
1174                         flooded = 1;
1175         } else {
1176                 vcpu->last_pt_write_gfn = gfn;
1177                 vcpu->last_pt_write_count = 1;
1178         }
1179         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1180         bucket = &vcpu->kvm->mmu_page_hash[index];
1181         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1182                 if (page->gfn != gfn || page->role.metaphysical)
1183                         continue;
1184                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1185                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1186                 misaligned |= bytes < 4;
1187                 if (misaligned || flooded) {
1188                         /*
1189                          * Misaligned accesses are too much trouble to fix
1190                          * up; also, they usually indicate a page is not used
1191                          * as a page table.
1192                          *
1193                          * If we're seeing too many writes to a page,
1194                          * it may no longer be a page table, or we may be
1195                          * forking, in which case it is better to unmap the
1196                          * page.
1197                          */
1198                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1199                                  gpa, bytes, page->role.word);
1200                         kvm_mmu_zap_page(vcpu, page);
1201                         continue;
1202                 }
1203                 page_offset = offset;
1204                 level = page->role.level;
1205                 npte = 1;
1206                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1207                         page_offset <<= 1;      /* 32->64 */
1208                         /*
1209                          * A 32-bit pde maps 4MB while the shadow pdes map
1210                          * only 2MB.  So we need to double the offset again
1211                          * and zap two pdes instead of one.
1212                          */
1213                         if (level == PT32_ROOT_LEVEL) {
1214                                 page_offset &= ~7; /* kill rounding error */
1215                                 page_offset <<= 1;
1216                                 npte = 2;
1217                         }
1218                         quadrant = page_offset >> PAGE_SHIFT;
1219                         page_offset &= ~PAGE_MASK;
1220                         if (quadrant != page->role.quadrant)
1221                                 continue;
1222                 }
1223                 spte = __va(page->page_hpa);
1224                 spte += page_offset / sizeof(*spte);
1225                 while (npte--) {
1226                         mmu_pte_write_zap_pte(vcpu, page, spte);
1227                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes);
1228                         ++spte;
1229                 }
1230         }
1231 }
1232
1233 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1234 {
1235         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1236
1237         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1238 }
1239
1240 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1241 {
1242         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1243                 struct kvm_mmu_page *page;
1244
1245                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1246                                     struct kvm_mmu_page, link);
1247                 kvm_mmu_zap_page(vcpu, page);
1248         }
1249 }
1250 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1251
1252 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1253 {
1254         struct kvm_mmu_page *page;
1255
1256         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1257                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1258                                     struct kvm_mmu_page, link);
1259                 kvm_mmu_zap_page(vcpu, page);
1260         }
1261         while (!list_empty(&vcpu->free_pages)) {
1262                 page = list_entry(vcpu->free_pages.next,
1263                                   struct kvm_mmu_page, link);
1264                 list_del(&page->link);
1265                 __free_page(pfn_to_page(page->page_hpa >> PAGE_SHIFT));
1266                 page->page_hpa = INVALID_PAGE;
1267         }
1268         free_page((unsigned long)vcpu->mmu.pae_root);
1269 }
1270
1271 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1272 {
1273         struct page *page;
1274         int i;
1275
1276         ASSERT(vcpu);
1277
1278         for (i = 0; i < KVM_NUM_MMU_PAGES; i++) {
1279                 struct kvm_mmu_page *page_header = &vcpu->page_header_buf[i];
1280
1281                 INIT_LIST_HEAD(&page_header->link);
1282                 if ((page = alloc_page(GFP_KERNEL)) == NULL)
1283                         goto error_1;
1284                 set_page_private(page, (unsigned long)page_header);
1285                 page_header->page_hpa = (hpa_t)page_to_pfn(page) << PAGE_SHIFT;
1286                 memset(__va(page_header->page_hpa), 0, PAGE_SIZE);
1287                 list_add(&page_header->link, &vcpu->free_pages);
1288                 ++vcpu->kvm->n_free_mmu_pages;
1289         }
1290
1291         /*
1292          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1293          * Therefore we need to allocate shadow page tables in the first
1294          * 4GB of memory, which happens to fit the DMA32 zone.
1295          */
1296         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1297         if (!page)
1298                 goto error_1;
1299         vcpu->mmu.pae_root = page_address(page);
1300         for (i = 0; i < 4; ++i)
1301                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1302
1303         return 0;
1304
1305 error_1:
1306         free_mmu_pages(vcpu);
1307         return -ENOMEM;
1308 }
1309
1310 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1311 {
1312         ASSERT(vcpu);
1313         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1314         ASSERT(list_empty(&vcpu->free_pages));
1315
1316         return alloc_mmu_pages(vcpu);
1317 }
1318
1319 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1320 {
1321         ASSERT(vcpu);
1322         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1323         ASSERT(!list_empty(&vcpu->free_pages));
1324
1325         return init_kvm_mmu(vcpu);
1326 }
1327
1328 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1329 {
1330         ASSERT(vcpu);
1331
1332         destroy_kvm_mmu(vcpu);
1333         free_mmu_pages(vcpu);
1334         mmu_free_memory_caches(vcpu);
1335 }
1336
1337 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1338 {
1339         struct kvm *kvm = vcpu->kvm;
1340         struct kvm_mmu_page *page;
1341
1342         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1343                 int i;
1344                 u64 *pt;
1345
1346                 if (!test_bit(slot, &page->slot_bitmap))
1347                         continue;
1348
1349                 pt = __va(page->page_hpa);
1350                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1351                         /* avoid RMW */
1352                         if (pt[i] & PT_WRITABLE_MASK) {
1353                                 rmap_remove(vcpu, &pt[i]);
1354                                 pt[i] &= ~PT_WRITABLE_MASK;
1355                         }
1356         }
1357 }
1358
1359 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1360 {
1361         destroy_kvm_mmu(vcpu);
1362
1363         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1364                 struct kvm_mmu_page *page;
1365
1366                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1367                                     struct kvm_mmu_page, link);
1368                 kvm_mmu_zap_page(vcpu, page);
1369         }
1370
1371         mmu_free_memory_caches(vcpu);
1372         kvm_arch_ops->tlb_flush(vcpu);
1373         init_kvm_mmu(vcpu);
1374 }
1375
1376 void kvm_mmu_module_exit(void)
1377 {
1378         if (pte_chain_cache)
1379                 kmem_cache_destroy(pte_chain_cache);
1380         if (rmap_desc_cache)
1381                 kmem_cache_destroy(rmap_desc_cache);
1382 }
1383
1384 int kvm_mmu_module_init(void)
1385 {
1386         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1387                                             sizeof(struct kvm_pte_chain),
1388                                             0, 0, NULL, NULL);
1389         if (!pte_chain_cache)
1390                 goto nomem;
1391         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1392                                             sizeof(struct kvm_rmap_desc),
1393                                             0, 0, NULL, NULL);
1394         if (!rmap_desc_cache)
1395                 goto nomem;
1396
1397         return 0;
1398
1399 nomem:
1400         kvm_mmu_module_exit();
1401         return -ENOMEM;
1402 }
1403
1404 #ifdef AUDIT
1405
1406 static const char *audit_msg;
1407
1408 static gva_t canonicalize(gva_t gva)
1409 {
1410 #ifdef CONFIG_X86_64
1411         gva = (long long)(gva << 16) >> 16;
1412 #endif
1413         return gva;
1414 }
1415
1416 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1417                                 gva_t va, int level)
1418 {
1419         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1420         int i;
1421         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1422
1423         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1424                 u64 ent = pt[i];
1425
1426                 if (!(ent & PT_PRESENT_MASK))
1427                         continue;
1428
1429                 va = canonicalize(va);
1430                 if (level > 1)
1431                         audit_mappings_page(vcpu, ent, va, level - 1);
1432                 else {
1433                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1434                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1435
1436                         if ((ent & PT_PRESENT_MASK)
1437                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1438                                 printk(KERN_ERR "audit error: (%s) levels %d"
1439                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1440                                        audit_msg, vcpu->mmu.root_level,
1441                                        va, gpa, hpa, ent);
1442                 }
1443         }
1444 }
1445
1446 static void audit_mappings(struct kvm_vcpu *vcpu)
1447 {
1448         unsigned i;
1449
1450         if (vcpu->mmu.root_level == 4)
1451                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1452         else
1453                 for (i = 0; i < 4; ++i)
1454                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1455                                 audit_mappings_page(vcpu,
1456                                                     vcpu->mmu.pae_root[i],
1457                                                     i << 30,
1458                                                     2);
1459 }
1460
1461 static int count_rmaps(struct kvm_vcpu *vcpu)
1462 {
1463         int nmaps = 0;
1464         int i, j, k;
1465
1466         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1467                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1468                 struct kvm_rmap_desc *d;
1469
1470                 for (j = 0; j < m->npages; ++j) {
1471                         struct page *page = m->phys_mem[j];
1472
1473                         if (!page->private)
1474                                 continue;
1475                         if (!(page->private & 1)) {
1476                                 ++nmaps;
1477                                 continue;
1478                         }
1479                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1480                         while (d) {
1481                                 for (k = 0; k < RMAP_EXT; ++k)
1482                                         if (d->shadow_ptes[k])
1483                                                 ++nmaps;
1484                                         else
1485                                                 break;
1486                                 d = d->more;
1487                         }
1488                 }
1489         }
1490         return nmaps;
1491 }
1492
1493 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1494 {
1495         int nmaps = 0;
1496         struct kvm_mmu_page *page;
1497         int i;
1498
1499         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1500                 u64 *pt = __va(page->page_hpa);
1501
1502                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1503                         continue;
1504
1505                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1506                         u64 ent = pt[i];
1507
1508                         if (!(ent & PT_PRESENT_MASK))
1509                                 continue;
1510                         if (!(ent & PT_WRITABLE_MASK))
1511                                 continue;
1512                         ++nmaps;
1513                 }
1514         }
1515         return nmaps;
1516 }
1517
1518 static void audit_rmap(struct kvm_vcpu *vcpu)
1519 {
1520         int n_rmap = count_rmaps(vcpu);
1521         int n_actual = count_writable_mappings(vcpu);
1522
1523         if (n_rmap != n_actual)
1524                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1525                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1526 }
1527
1528 static void audit_write_protection(struct kvm_vcpu *vcpu)
1529 {
1530         struct kvm_mmu_page *page;
1531
1532         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1533                 hfn_t hfn;
1534                 struct page *pg;
1535
1536                 if (page->role.metaphysical)
1537                         continue;
1538
1539                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1540                         >> PAGE_SHIFT;
1541                 pg = pfn_to_page(hfn);
1542                 if (pg->private)
1543                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1544                                " mappings: gfn %lx role %x\n",
1545                                __FUNCTION__, audit_msg, page->gfn,
1546                                page->role.word);
1547         }
1548 }
1549
1550 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1551 {
1552         int olddbg = dbg;
1553
1554         dbg = 0;
1555         audit_msg = msg;
1556         audit_rmap(vcpu);
1557         audit_write_protection(vcpu);
1558         audit_mappings(vcpu);
1559         dbg = olddbg;
1560 }
1561
1562 #endif