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