err.no Git - linux-2.6/blob - drivers/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19 #include <linux/types.h>
  20 #include <linux/string.h>
  21 #include <asm/page.h>
  22 #include <linux/mm.h>
  23 #include <linux/highmem.h>
  24 #include <linux/module.h>
  25
  26 #include "vmx.h"
  27 #include "kvm.h"
  28
  29 #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 /*
 170  * Reverse mapping data structures:
 171  *
 172  * If page->private bit zero is zero, then page->private points to the
 173  * shadow page table entry that points to page_address(page).
 174  *
 175  * If page->private bit zero is one, (then page->private & ~1) points
 176  * to a struct kvm_rmap_desc containing more mappings.
 177  */
 178 static void rmap_add(struct kvm *kvm, u64 *spte)
 179 {
 180         struct page *page;
 181         struct kvm_rmap_desc *desc;
 182         int i;
 183
 184         if (!is_rmap_pte(*spte))
 185                 return;
 186         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 187         if (!page->private) {
 188                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 189                 page->private = (unsigned long)spte;
 190         } else if (!(page->private & 1)) {
 191                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 192                 desc = kzalloc(sizeof *desc, GFP_NOWAIT);
 193                 if (!desc)
 194                         BUG(); /* FIXME: return error */
 195                 desc->shadow_ptes[0] = (u64 *)page->private;
 196                 desc->shadow_ptes[1] = spte;
 197                 page->private = (unsigned long)desc | 1;
 198         } else {
 199                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 200                 desc = (struct kvm_rmap_desc *)(page->private & ~1ul);
 201                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 202                         desc = desc->more;
 203                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 204                         desc->more = kzalloc(sizeof *desc->more, GFP_NOWAIT);
 205                         if (!desc->more)
 206                                 BUG(); /* FIXME: return error */
 207                         desc = desc->more;
 208                 }
 209                 for (i = 0; desc->shadow_ptes[i]; ++i)
 210                         ;
 211                 desc->shadow_ptes[i] = spte;
 212         }
 213 }
 214
 215 static void rmap_desc_remove_entry(struct page *page,
 216                                    struct kvm_rmap_desc *desc,
 217                                    int i,
 218                                    struct kvm_rmap_desc *prev_desc)
 219 {
 220         int j;
 221
 222         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 223                 ;
 224         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 225         desc->shadow_ptes[j] = 0;
 226         if (j != 0)
 227                 return;
 228         if (!prev_desc && !desc->more)
 229                 page->private = (unsigned long)desc->shadow_ptes[0];
 230         else
 231                 if (prev_desc)
 232                         prev_desc->more = desc->more;
 233                 else
 234                         page->private = (unsigned long)desc->more | 1;
 235         kfree(desc);
 236 }
 237
 238 static void rmap_remove(struct kvm *kvm, u64 *spte)
 239 {
 240         struct page *page;
 241         struct kvm_rmap_desc *desc;
 242         struct kvm_rmap_desc *prev_desc;
 243         int i;
 244
 245         if (!is_rmap_pte(*spte))
 246                 return;
 247         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 248         if (!page->private) {
 249                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 250                 BUG();
 251         } else if (!(page->private & 1)) {
 252                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 253                 if ((u64 *)page->private != spte) {
 254                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 255                                spte, *spte);
 256                         BUG();
 257                 }
 258                 page->private = 0;
 259         } else {
 260                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 261                 desc = (struct kvm_rmap_desc *)(page->private & ~1ul);
 262                 prev_desc = NULL;
 263                 while (desc) {
 264                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 265                                 if (desc->shadow_ptes[i] == spte) {
 266                                         rmap_desc_remove_entry(page, desc, i,
 267                                                                prev_desc);
 268                                         return;
 269                                 }
 270                         prev_desc = desc;
 271                         desc = desc->more;
 272                 }
 273                 BUG();
 274         }
 275 }
 276
 277 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 278 {
 279         struct page *page;
 280         struct kvm_memory_slot *slot;
 281         struct kvm_rmap_desc *desc;
 282         u64 *spte;
 283
 284         slot = gfn_to_memslot(kvm, gfn);
 285         BUG_ON(!slot);
 286         page = gfn_to_page(slot, gfn);
 287
 288         while (page->private) {
 289                 if (!(page->private & 1))
 290                         spte = (u64 *)page->private;
 291                 else {
 292                         desc = (struct kvm_rmap_desc *)(page->private & ~1ul);
 293                         spte = desc->shadow_ptes[0];
 294                 }
 295                 BUG_ON(!spte);
 296                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) !=
 297                        page_to_pfn(page) << PAGE_SHIFT);
 298                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 299                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 300                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 301                 rmap_remove(kvm, spte);
 302                 *spte &= ~(u64)PT_WRITABLE_MASK;
 303         }
 304 }
 305
 306 static int is_empty_shadow_page(hpa_t page_hpa)
 307 {
 308         u32 *pos;
 309         u32 *end;
 310         for (pos = __va(page_hpa), end = pos + PAGE_SIZE / sizeof(u32);
 311                       pos != end; pos++)
 312                 if (*pos != 0)
 313                         return 0;
 314         return 1;
 315 }
 316
 317 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu, hpa_t page_hpa)
 318 {
 319         struct kvm_mmu_page *page_head = page_header(page_hpa);
 320
 321         ASSERT(is_empty_shadow_page(page_hpa));
 322         list_del(&page_head->link);
 323         page_head->page_hpa = page_hpa;
 324         list_add(&page_head->link, &vcpu->free_pages);
 325         ++vcpu->kvm->n_free_mmu_pages;
 326 }
 327
 328 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 329 {
 330         return gfn;
 331 }
 332
 333 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 334                                                u64 *parent_pte)
 335 {
 336         struct kvm_mmu_page *page;
 337
 338         if (list_empty(&vcpu->free_pages))
 339                 return NULL;
 340
 341         page = list_entry(vcpu->free_pages.next, struct kvm_mmu_page, link);
 342         list_del(&page->link);
 343         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 344         ASSERT(is_empty_shadow_page(page->page_hpa));
 345         page->slot_bitmap = 0;
 346         page->global = 1;
 347         page->multimapped = 0;
 348         page->parent_pte = parent_pte;
 349         --vcpu->kvm->n_free_mmu_pages;
 350         return page;
 351 }
 352
 353 static void mmu_page_add_parent_pte(struct kvm_mmu_page *page, u64 *parent_pte)
 354 {
 355         struct kvm_pte_chain *pte_chain;
 356         struct hlist_node *node;
 357         int i;
 358
 359         if (!parent_pte)
 360                 return;
 361         if (!page->multimapped) {
 362                 u64 *old = page->parent_pte;
 363
 364                 if (!old) {
 365                         page->parent_pte = parent_pte;
 366                         return;
 367                 }
 368                 page->multimapped = 1;
 369                 pte_chain = kzalloc(sizeof(struct kvm_pte_chain), GFP_NOWAIT);
 370                 BUG_ON(!pte_chain);
 371                 INIT_HLIST_HEAD(&page->parent_ptes);
 372                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 373                 pte_chain->parent_ptes[0] = old;
 374         }
 375         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 376                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 377                         continue;
 378                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 379                         if (!pte_chain->parent_ptes[i]) {
 380                                 pte_chain->parent_ptes[i] = parent_pte;
 381                                 return;
 382                         }
 383         }
 384         pte_chain = kzalloc(sizeof(struct kvm_pte_chain), GFP_NOWAIT);
 385         BUG_ON(!pte_chain);
 386         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 387         pte_chain->parent_ptes[0] = parent_pte;
 388 }
 389
 390 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 391                                        u64 *parent_pte)
 392 {
 393         struct kvm_pte_chain *pte_chain;
 394         struct hlist_node *node;
 395         int i;
 396
 397         if (!page->multimapped) {
 398                 BUG_ON(page->parent_pte != parent_pte);
 399                 page->parent_pte = NULL;
 400                 return;
 401         }
 402         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 403                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 404                         if (!pte_chain->parent_ptes[i])
 405                                 break;
 406                         if (pte_chain->parent_ptes[i] != parent_pte)
 407                                 continue;
 408                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 409                                 && pte_chain->parent_ptes[i + 1]) {
 410                                 pte_chain->parent_ptes[i]
 411                                         = pte_chain->parent_ptes[i + 1];
 412                                 ++i;
 413                         }
 414                         pte_chain->parent_ptes[i] = NULL;
 415                         if (i == 0) {
 416                                 hlist_del(&pte_chain->link);
 417                                 kfree(pte_chain);
 418                                 if (hlist_empty(&page->parent_ptes)) {
 419                                         page->multimapped = 0;
 420                                         page->parent_pte = NULL;
 421                                 }
 422                         }
 423                         return;
 424                 }
 425         BUG();
 426 }
 427
 428 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 429                                                 gfn_t gfn)
 430 {
 431         unsigned index;
 432         struct hlist_head *bucket;
 433         struct kvm_mmu_page *page;
 434         struct hlist_node *node;
 435
 436         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 437         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 438         bucket = &vcpu->kvm->mmu_page_hash[index];
 439         hlist_for_each_entry(page, node, bucket, hash_link)
 440                 if (page->gfn == gfn && !page->role.metaphysical) {
 441                         pgprintk("%s: found role %x\n",
 442                                  __FUNCTION__, page->role.word);
 443                         return page;
 444                 }
 445         return NULL;
 446 }
 447
 448 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 449                                              gfn_t gfn,
 450                                              gva_t gaddr,
 451                                              unsigned level,
 452                                              int metaphysical,
 453                                              u64 *parent_pte)
 454 {
 455         union kvm_mmu_page_role role;
 456         unsigned index;
 457         unsigned quadrant;
 458         struct hlist_head *bucket;
 459         struct kvm_mmu_page *page;
 460         struct hlist_node *node;
 461
 462         role.word = 0;
 463         role.glevels = vcpu->mmu.root_level;
 464         role.level = level;
 465         role.metaphysical = metaphysical;
 466         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 467                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 468                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 469                 role.quadrant = quadrant;
 470         }
 471         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 472                  gfn, role.word);
 473         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 474         bucket = &vcpu->kvm->mmu_page_hash[index];
 475         hlist_for_each_entry(page, node, bucket, hash_link)
 476                 if (page->gfn == gfn && page->role.word == role.word) {
 477                         mmu_page_add_parent_pte(page, parent_pte);
 478                         pgprintk("%s: found\n", __FUNCTION__);
 479                         return page;
 480                 }
 481         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 482         if (!page)
 483                 return page;
 484         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 485         page->gfn = gfn;
 486         page->role = role;
 487         hlist_add_head(&page->hash_link, bucket);
 488         if (!metaphysical)
 489                 rmap_write_protect(vcpu->kvm, gfn);
 490         return page;
 491 }
 492
 493 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 494                                          struct kvm_mmu_page *page)
 495 {
 496         unsigned i;
 497         u64 *pt;
 498         u64 ent;
 499
 500         pt = __va(page->page_hpa);
 501
 502         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 503                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 504                         if (pt[i] & PT_PRESENT_MASK)
 505                                 rmap_remove(vcpu->kvm, &pt[i]);
 506                         pt[i] = 0;
 507                 }
 508                 return;
 509         }
 510
 511         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 512                 ent = pt[i];
 513
 514                 pt[i] = 0;
 515                 if (!(ent & PT_PRESENT_MASK))
 516                         continue;
 517                 ent &= PT64_BASE_ADDR_MASK;
 518                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 519         }
 520 }
 521
 522 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 523                              struct kvm_mmu_page *page,
 524                              u64 *parent_pte)
 525 {
 526         mmu_page_remove_parent_pte(page, parent_pte);
 527 }
 528
 529 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 530                              struct kvm_mmu_page *page)
 531 {
 532         u64 *parent_pte;
 533
 534         while (page->multimapped || page->parent_pte) {
 535                 if (!page->multimapped)
 536                         parent_pte = page->parent_pte;
 537                 else {
 538                         struct kvm_pte_chain *chain;
 539
 540                         chain = container_of(page->parent_ptes.first,
 541                                              struct kvm_pte_chain, link);
 542                         parent_pte = chain->parent_ptes[0];
 543                 }
 544                 BUG_ON(!parent_pte);
 545                 kvm_mmu_put_page(vcpu, page, parent_pte);
 546                 *parent_pte = 0;
 547         }
 548         kvm_mmu_page_unlink_children(vcpu, page);
 549         hlist_del(&page->hash_link);
 550         kvm_mmu_free_page(vcpu, page->page_hpa);
 551 }
 552
 553 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 554 {
 555         unsigned index;
 556         struct hlist_head *bucket;
 557         struct kvm_mmu_page *page;
 558         struct hlist_node *node, *n;
 559         int r;
 560
 561         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 562         r = 0;
 563         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 564         bucket = &vcpu->kvm->mmu_page_hash[index];
 565         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 566                 if (page->gfn == gfn && !page->role.metaphysical) {
 567                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 568                                  page->role.word);
 569                         kvm_mmu_zap_page(vcpu, page);
 570                         r = 1;
 571                 }
 572         return r;
 573 }
 574
 575 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 576 {
 577         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 578         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 579
 580         __set_bit(slot, &page_head->slot_bitmap);
 581 }
 582
 583 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 584 {
 585         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 586
 587         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 588 }
 589
 590 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 591 {
 592         struct kvm_memory_slot *slot;
 593         struct page *page;
 594
 595         ASSERT((gpa & HPA_ERR_MASK) == 0);
 596         slot = gfn_to_memslot(vcpu->kvm, gpa >> PAGE_SHIFT);
 597         if (!slot)
 598                 return gpa | HPA_ERR_MASK;
 599         page = gfn_to_page(slot, gpa >> PAGE_SHIFT);
 600         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 601                 | (gpa & (PAGE_SIZE-1));
 602 }
 603
 604 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 605 {
 606         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 607
 608         if (gpa == UNMAPPED_GVA)
 609                 return UNMAPPED_GVA;
 610         return gpa_to_hpa(vcpu, gpa);
 611 }
 612
 613 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 614 {
 615 }
 616
 617 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 618 {
 619         int level = PT32E_ROOT_LEVEL;
 620         hpa_t table_addr = vcpu->mmu.root_hpa;
 621
 622         for (; ; level--) {
 623                 u32 index = PT64_INDEX(v, level);
 624                 u64 *table;
 625                 u64 pte;
 626
 627                 ASSERT(VALID_PAGE(table_addr));
 628                 table = __va(table_addr);
 629
 630                 if (level == 1) {
 631                         pte = table[index];
 632                         if (is_present_pte(pte) && is_writeble_pte(pte))
 633                                 return 0;
 634                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 635                         page_header_update_slot(vcpu->kvm, table, v);
 636                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 637                                                                 PT_USER_MASK;
 638                         rmap_add(vcpu->kvm, &table[index]);
 639                         return 0;
 640                 }
 641
 642                 if (table[index] == 0) {
 643                         struct kvm_mmu_page *new_table;
 644                         gfn_t pseudo_gfn;
 645
 646                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 647                                 >> PAGE_SHIFT;
 648                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 649                                                      v, level - 1,
 650                                                      1, &table[index]);
 651                         if (!new_table) {
 652                                 pgprintk("nonpaging_map: ENOMEM\n");
 653                                 return -ENOMEM;
 654                         }
 655
 656                         table[index] = new_table->page_hpa | PT_PRESENT_MASK
 657                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 658                 }
 659                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 660         }
 661 }
 662
 663 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 664 {
 665         int i;
 666
 667 #ifdef CONFIG_X86_64
 668         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 669                 hpa_t root = vcpu->mmu.root_hpa;
 670
 671                 ASSERT(VALID_PAGE(root));
 672                 vcpu->mmu.root_hpa = INVALID_PAGE;
 673                 return;
 674         }
 675 #endif
 676         for (i = 0; i < 4; ++i) {
 677                 hpa_t root = vcpu->mmu.pae_root[i];
 678
 679                 ASSERT(VALID_PAGE(root));
 680                 root &= PT64_BASE_ADDR_MASK;
 681                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 682         }
 683         vcpu->mmu.root_hpa = INVALID_PAGE;
 684 }
 685
 686 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 687 {
 688         int i;
 689         gfn_t root_gfn;
 690         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 691
 692 #ifdef CONFIG_X86_64
 693         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 694                 hpa_t root = vcpu->mmu.root_hpa;
 695
 696                 ASSERT(!VALID_PAGE(root));
 697                 root = kvm_mmu_get_page(vcpu, root_gfn, 0,
 698                                         PT64_ROOT_LEVEL, 0, NULL)->page_hpa;
 699                 vcpu->mmu.root_hpa = root;
 700                 return;
 701         }
 702 #endif
 703         for (i = 0; i < 4; ++i) {
 704                 hpa_t root = vcpu->mmu.pae_root[i];
 705
 706                 ASSERT(!VALID_PAGE(root));
 707                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL)
 708                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 709                 else if (vcpu->mmu.root_level == 0)
 710                         root_gfn = 0;
 711                 root = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 712                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 713                                         NULL)->page_hpa;
 714                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 715         }
 716         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 717 }
 718
 719 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 720 {
 721         return vaddr;
 722 }
 723
 724 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 725                                u32 error_code)
 726 {
 727         gpa_t addr = gva;
 728         hpa_t paddr;
 729
 730         ASSERT(vcpu);
 731         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 732
 733
 734         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 735
 736         if (is_error_hpa(paddr))
 737                 return 1;
 738
 739         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 740 }
 741
 742 static void nonpaging_free(struct kvm_vcpu *vcpu)
 743 {
 744         mmu_free_roots(vcpu);
 745 }
 746
 747 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 748 {
 749         struct kvm_mmu *context = &vcpu->mmu;
 750
 751         context->new_cr3 = nonpaging_new_cr3;
 752         context->page_fault = nonpaging_page_fault;
 753         context->gva_to_gpa = nonpaging_gva_to_gpa;
 754         context->free = nonpaging_free;
 755         context->root_level = 0;
 756         context->shadow_root_level = PT32E_ROOT_LEVEL;
 757         mmu_alloc_roots(vcpu);
 758         ASSERT(VALID_PAGE(context->root_hpa));
 759         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 760         return 0;
 761 }
 762
 763 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 764 {
 765         ++kvm_stat.tlb_flush;
 766         kvm_arch_ops->tlb_flush(vcpu);
 767 }
 768
 769 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 770 {
 771         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 772         mmu_free_roots(vcpu);
 773         mmu_alloc_roots(vcpu);
 774         kvm_mmu_flush_tlb(vcpu);
 775         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 776 }
 777
 778 static void mark_pagetable_nonglobal(void *shadow_pte)
 779 {
 780         page_header(__pa(shadow_pte))->global = 0;
 781 }
 782
 783 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 784                              u64 *shadow_pte,
 785                              gpa_t gaddr,
 786                              int dirty,
 787                              u64 access_bits,
 788                              gfn_t gfn)
 789 {
 790         hpa_t paddr;
 791
 792         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 793         if (!dirty)
 794                 access_bits &= ~PT_WRITABLE_MASK;
 795
 796         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 797
 798         *shadow_pte |= access_bits;
 799
 800         if (!(*shadow_pte & PT_GLOBAL_MASK))
 801                 mark_pagetable_nonglobal(shadow_pte);
 802
 803         if (is_error_hpa(paddr)) {
 804                 *shadow_pte |= gaddr;
 805                 *shadow_pte |= PT_SHADOW_IO_MARK;
 806                 *shadow_pte &= ~PT_PRESENT_MASK;
 807                 return;
 808         }
 809
 810         *shadow_pte |= paddr;
 811
 812         if (access_bits & PT_WRITABLE_MASK) {
 813                 struct kvm_mmu_page *shadow;
 814
 815                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 816                 if (shadow) {
 817                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 818                                  __FUNCTION__, gfn);
 819                         access_bits &= ~PT_WRITABLE_MASK;
 820                         *shadow_pte &= ~PT_WRITABLE_MASK;
 821                 }
 822         }
 823
 824         if (access_bits & PT_WRITABLE_MASK)
 825                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 826
 827         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 828         rmap_add(vcpu->kvm, shadow_pte);
 829 }
 830
 831 static void inject_page_fault(struct kvm_vcpu *vcpu,
 832                               u64 addr,
 833                               u32 err_code)
 834 {
 835         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
 836 }
 837
 838 static inline int fix_read_pf(u64 *shadow_ent)
 839 {
 840         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
 841             !(*shadow_ent & PT_USER_MASK)) {
 842                 /*
 843                  * If supervisor write protect is disabled, we shadow kernel
 844                  * pages as user pages so we can trap the write access.
 845                  */
 846                 *shadow_ent |= PT_USER_MASK;
 847                 *shadow_ent &= ~PT_WRITABLE_MASK;
 848
 849                 return 1;
 850
 851         }
 852         return 0;
 853 }
 854
 855 static int may_access(u64 pte, int write, int user)
 856 {
 857
 858         if (user && !(pte & PT_USER_MASK))
 859                 return 0;
 860         if (write && !(pte & PT_WRITABLE_MASK))
 861                 return 0;
 862         return 1;
 863 }
 864
 865 static void paging_free(struct kvm_vcpu *vcpu)
 866 {
 867         nonpaging_free(vcpu);
 868 }
 869
 870 #define PTTYPE 64
 871 #include "paging_tmpl.h"
 872 #undef PTTYPE
 873
 874 #define PTTYPE 32
 875 #include "paging_tmpl.h"
 876 #undef PTTYPE
 877
 878 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
 879 {
 880         struct kvm_mmu *context = &vcpu->mmu;
 881
 882         ASSERT(is_pae(vcpu));
 883         context->new_cr3 = paging_new_cr3;
 884         context->page_fault = paging64_page_fault;
 885         context->gva_to_gpa = paging64_gva_to_gpa;
 886         context->free = paging_free;
 887         context->root_level = level;
 888         context->shadow_root_level = level;
 889         mmu_alloc_roots(vcpu);
 890         ASSERT(VALID_PAGE(context->root_hpa));
 891         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
 892                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
 893         return 0;
 894 }
 895
 896 static int paging64_init_context(struct kvm_vcpu *vcpu)
 897 {
 898         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
 899 }
 900
 901 static int paging32_init_context(struct kvm_vcpu *vcpu)
 902 {
 903         struct kvm_mmu *context = &vcpu->mmu;
 904
 905         context->new_cr3 = paging_new_cr3;
 906         context->page_fault = paging32_page_fault;
 907         context->gva_to_gpa = paging32_gva_to_gpa;
 908         context->free = paging_free;
 909         context->root_level = PT32_ROOT_LEVEL;
 910         context->shadow_root_level = PT32E_ROOT_LEVEL;
 911         mmu_alloc_roots(vcpu);
 912         ASSERT(VALID_PAGE(context->root_hpa));
 913         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
 914                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
 915         return 0;
 916 }
 917
 918 static int paging32E_init_context(struct kvm_vcpu *vcpu)
 919 {
 920         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
 921 }
 922
 923 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
 924 {
 925         ASSERT(vcpu);
 926         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
 927
 928         if (!is_paging(vcpu))
 929                 return nonpaging_init_context(vcpu);
 930         else if (is_long_mode(vcpu))
 931                 return paging64_init_context(vcpu);
 932         else if (is_pae(vcpu))
 933                 return paging32E_init_context(vcpu);
 934         else
 935                 return paging32_init_context(vcpu);
 936 }
 937
 938 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
 939 {
 940         ASSERT(vcpu);
 941         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
 942                 vcpu->mmu.free(vcpu);
 943                 vcpu->mmu.root_hpa = INVALID_PAGE;
 944         }
 945 }
 946
 947 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
 948 {
 949         destroy_kvm_mmu(vcpu);
 950         return init_kvm_mmu(vcpu);
 951 }
 952
 953 void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
 954 {
 955         gfn_t gfn = gpa >> PAGE_SHIFT;
 956         struct kvm_mmu_page *page;
 957         struct kvm_mmu_page *child;
 958         struct hlist_node *node, *n;
 959         struct hlist_head *bucket;
 960         unsigned index;
 961         u64 *spte;
 962         u64 pte;
 963         unsigned offset = offset_in_page(gpa);
 964         unsigned pte_size;
 965         unsigned page_offset;
 966         unsigned misaligned;
 967         int level;
 968
 969         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
 970         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 971         bucket = &vcpu->kvm->mmu_page_hash[index];
 972         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
 973                 if (page->gfn != gfn || page->role.metaphysical)
 974                         continue;
 975                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
 976                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
 977                 if (misaligned) {
 978                         /*
 979                          * Misaligned accesses are too much trouble to fix
 980                          * up; also, they usually indicate a page is not used
 981                          * as a page table.
 982                          */
 983                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
 984                                  gpa, bytes, page->role.word);
 985                         kvm_mmu_zap_page(vcpu, page);
 986                         continue;
 987                 }
 988                 page_offset = offset;
 989                 level = page->role.level;
 990                 if (page->role.glevels == PT32_ROOT_LEVEL) {
 991                         page_offset <<= 1;          /* 32->64 */
 992                         page_offset &= ~PAGE_MASK;
 993                 }
 994                 spte = __va(page->page_hpa);
 995                 spte += page_offset / sizeof(*spte);
 996                 pte = *spte;
 997                 if (is_present_pte(pte)) {
 998                         if (level == PT_PAGE_TABLE_LEVEL)
 999                                 rmap_remove(vcpu->kvm, spte);
1000                         else {
1001                                 child = page_header(pte & PT64_BASE_ADDR_MASK);
1002                                 mmu_page_remove_parent_pte(child, spte);
1003                         }
1004                 }
1005                 *spte = 0;
1006         }
1007 }
1008
1009 void kvm_mmu_post_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1010 {
1011 }
1012
1013 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1014 {
1015         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1016
1017         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1018 }
1019
1020 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1021 {
1022         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1023                 struct kvm_mmu_page *page;
1024
1025                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1026                                     struct kvm_mmu_page, link);
1027                 kvm_mmu_zap_page(vcpu, page);
1028         }
1029 }
1030 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1031
1032 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1033 {
1034         while (!list_empty(&vcpu->free_pages)) {
1035                 struct kvm_mmu_page *page;
1036
1037                 page = list_entry(vcpu->free_pages.next,
1038                                   struct kvm_mmu_page, link);
1039                 list_del(&page->link);
1040                 __free_page(pfn_to_page(page->page_hpa >> PAGE_SHIFT));
1041                 page->page_hpa = INVALID_PAGE;
1042         }
1043         free_page((unsigned long)vcpu->mmu.pae_root);
1044 }
1045
1046 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1047 {
1048         struct page *page;
1049         int i;
1050
1051         ASSERT(vcpu);
1052
1053         for (i = 0; i < KVM_NUM_MMU_PAGES; i++) {
1054                 struct kvm_mmu_page *page_header = &vcpu->page_header_buf[i];
1055
1056                 INIT_LIST_HEAD(&page_header->link);
1057                 if ((page = alloc_page(GFP_KERNEL)) == NULL)
1058                         goto error_1;
1059                 page->private = (unsigned long)page_header;
1060                 page_header->page_hpa = (hpa_t)page_to_pfn(page) << PAGE_SHIFT;
1061                 memset(__va(page_header->page_hpa), 0, PAGE_SIZE);
1062                 list_add(&page_header->link, &vcpu->free_pages);
1063                 ++vcpu->kvm->n_free_mmu_pages;
1064         }
1065
1066         /*
1067          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1068          * Therefore we need to allocate shadow page tables in the first
1069          * 4GB of memory, which happens to fit the DMA32 zone.
1070          */
1071         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1072         if (!page)
1073                 goto error_1;
1074         vcpu->mmu.pae_root = page_address(page);
1075         for (i = 0; i < 4; ++i)
1076                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1077
1078         return 0;
1079
1080 error_1:
1081         free_mmu_pages(vcpu);
1082         return -ENOMEM;
1083 }
1084
1085 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1086 {
1087         ASSERT(vcpu);
1088         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1089         ASSERT(list_empty(&vcpu->free_pages));
1090
1091         return alloc_mmu_pages(vcpu);
1092 }
1093
1094 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1095 {
1096         ASSERT(vcpu);
1097         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1098         ASSERT(!list_empty(&vcpu->free_pages));
1099
1100         return init_kvm_mmu(vcpu);
1101 }
1102
1103 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1104 {
1105         ASSERT(vcpu);
1106
1107         destroy_kvm_mmu(vcpu);
1108         free_mmu_pages(vcpu);
1109 }
1110
1111 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1112 {
1113         struct kvm_mmu_page *page;
1114
1115         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1116                 int i;
1117                 u64 *pt;
1118
1119                 if (!test_bit(slot, &page->slot_bitmap))
1120                         continue;
1121
1122                 pt = __va(page->page_hpa);
1123                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1124                         /* avoid RMW */
1125                         if (pt[i] & PT_WRITABLE_MASK) {
1126                                 rmap_remove(kvm, &pt[i]);
1127                                 pt[i] &= ~PT_WRITABLE_MASK;
1128                         }
1129         }
1130 }