SAFE public projects git trees. - safe/jmp/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 void kvm_mmu_free_page(struct kvm_vcpu *vcpu, hpa_t page_hpa)
 307 {
 308         struct kvm_mmu_page *page_head = page_header(page_hpa);
 309
 310         list_del(&page_head->link);
 311         page_head->page_hpa = page_hpa;
 312         list_add(&page_head->link, &vcpu->free_pages);
 313         ++vcpu->kvm->n_free_mmu_pages;
 314 }
 315
 316 static int is_empty_shadow_page(hpa_t page_hpa)
 317 {
 318         u32 *pos;
 319         u32 *end;
 320         for (pos = __va(page_hpa), end = pos + PAGE_SIZE / sizeof(u32);
 321                       pos != end; pos++)
 322                 if (*pos != 0)
 323                         return 0;
 324         return 1;
 325 }
 326
 327 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 328 {
 329         return gfn;
 330 }
 331
 332 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 333                                                u64 *parent_pte)
 334 {
 335         struct kvm_mmu_page *page;
 336
 337         if (list_empty(&vcpu->free_pages))
 338                 return NULL;
 339
 340         page = list_entry(vcpu->free_pages.next, struct kvm_mmu_page, link);
 341         list_del(&page->link);
 342         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 343         ASSERT(is_empty_shadow_page(page->page_hpa));
 344         page->slot_bitmap = 0;
 345         page->global = 1;
 346         page->multimapped = 0;
 347         page->parent_pte = parent_pte;
 348         --vcpu->kvm->n_free_mmu_pages;
 349         return page;
 350 }
 351
 352 static void mmu_page_add_parent_pte(struct kvm_mmu_page *page, u64 *parent_pte)
 353 {
 354         struct kvm_pte_chain *pte_chain;
 355         struct hlist_node *node;
 356         int i;
 357
 358         if (!parent_pte)
 359                 return;
 360         if (!page->multimapped) {
 361                 u64 *old = page->parent_pte;
 362
 363                 if (!old) {
 364                         page->parent_pte = parent_pte;
 365                         return;
 366                 }
 367                 page->multimapped = 1;
 368                 pte_chain = kzalloc(sizeof(struct kvm_pte_chain), GFP_NOWAIT);
 369                 BUG_ON(!pte_chain);
 370                 INIT_HLIST_HEAD(&page->parent_ptes);
 371                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 372                 pte_chain->parent_ptes[0] = old;
 373         }
 374         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 375                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 376                         continue;
 377                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 378                         if (!pte_chain->parent_ptes[i]) {
 379                                 pte_chain->parent_ptes[i] = parent_pte;
 380                                 return;
 381                         }
 382         }
 383         pte_chain = kzalloc(sizeof(struct kvm_pte_chain), GFP_NOWAIT);
 384         BUG_ON(!pte_chain);
 385         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 386         pte_chain->parent_ptes[0] = parent_pte;
 387 }
 388
 389 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 390                                        u64 *parent_pte)
 391 {
 392         struct kvm_pte_chain *pte_chain;
 393         struct hlist_node *node;
 394         int i;
 395
 396         if (!page->multimapped) {
 397                 BUG_ON(page->parent_pte != parent_pte);
 398                 page->parent_pte = NULL;
 399                 return;
 400         }
 401         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 402                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 403                         if (!pte_chain->parent_ptes[i])
 404                                 break;
 405                         if (pte_chain->parent_ptes[i] != parent_pte)
 406                                 continue;
 407                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 408                                 && pte_chain->parent_ptes[i + 1]) {
 409                                 pte_chain->parent_ptes[i]
 410                                         = pte_chain->parent_ptes[i + 1];
 411                                 ++i;
 412                         }
 413                         pte_chain->parent_ptes[i] = NULL;
 414                         if (i == 0) {
 415                                 hlist_del(&pte_chain->link);
 416                                 kfree(pte_chain);
 417                                 if (hlist_empty(&page->parent_ptes)) {
 418                                         page->multimapped = 0;
 419                                         page->parent_pte = NULL;
 420                                 }
 421                         }
 422                         return;
 423                 }
 424         BUG();
 425 }
 426
 427 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 428                                                 gfn_t gfn)
 429 {
 430         unsigned index;
 431         struct hlist_head *bucket;
 432         struct kvm_mmu_page *page;
 433         struct hlist_node *node;
 434
 435         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 436         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 437         bucket = &vcpu->kvm->mmu_page_hash[index];
 438         hlist_for_each_entry(page, node, bucket, hash_link)
 439                 if (page->gfn == gfn && !page->role.metaphysical) {
 440                         pgprintk("%s: found role %x\n",
 441                                  __FUNCTION__, page->role.word);
 442                         return page;
 443                 }
 444         return NULL;
 445 }
 446
 447 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 448                                              gfn_t gfn,
 449                                              gva_t gaddr,
 450                                              unsigned level,
 451                                              int metaphysical,
 452                                              u64 *parent_pte)
 453 {
 454         union kvm_mmu_page_role role;
 455         unsigned index;
 456         unsigned quadrant;
 457         struct hlist_head *bucket;
 458         struct kvm_mmu_page *page;
 459         struct hlist_node *node;
 460
 461         role.word = 0;
 462         role.glevels = vcpu->mmu.root_level;
 463         role.level = level;
 464         role.metaphysical = metaphysical;
 465         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 466                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 467                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 468                 role.quadrant = quadrant;
 469         }
 470         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 471                  gfn, role.word);
 472         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 473         bucket = &vcpu->kvm->mmu_page_hash[index];
 474         hlist_for_each_entry(page, node, bucket, hash_link)
 475                 if (page->gfn == gfn && page->role.word == role.word) {
 476                         mmu_page_add_parent_pte(page, parent_pte);
 477                         pgprintk("%s: found\n", __FUNCTION__);
 478                         return page;
 479                 }
 480         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 481         if (!page)
 482                 return page;
 483         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 484         page->gfn = gfn;
 485         page->role = role;
 486         hlist_add_head(&page->hash_link, bucket);
 487         if (!metaphysical)
 488                 rmap_write_protect(vcpu->kvm, gfn);
 489         return page;
 490 }
 491
 492 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 493                                          struct kvm_mmu_page *page)
 494 {
 495         unsigned i;
 496         u64 *pt;
 497         u64 ent;
 498
 499         pt = __va(page->page_hpa);
 500
 501         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 502                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 503                         if (pt[i] & PT_PRESENT_MASK)
 504                                 rmap_remove(vcpu->kvm, &pt[i]);
 505                         pt[i] = 0;
 506                 }
 507                 return;
 508         }
 509
 510         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 511                 ent = pt[i];
 512
 513                 pt[i] = 0;
 514                 if (!(ent & PT_PRESENT_MASK))
 515                         continue;
 516                 ent &= PT64_BASE_ADDR_MASK;
 517                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 518         }
 519 }
 520
 521 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 522                              struct kvm_mmu_page *page,
 523                              u64 *parent_pte)
 524 {
 525         mmu_page_remove_parent_pte(page, parent_pte);
 526 }
 527
 528 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 529                              struct kvm_mmu_page *page)
 530 {
 531         u64 *parent_pte;
 532
 533         while (page->multimapped || page->parent_pte) {
 534                 if (!page->multimapped)
 535                         parent_pte = page->parent_pte;
 536                 else {
 537                         struct kvm_pte_chain *chain;
 538
 539                         chain = container_of(page->parent_ptes.first,
 540                                              struct kvm_pte_chain, link);
 541                         parent_pte = chain->parent_ptes[0];
 542                 }
 543                 BUG_ON(!parent_pte);
 544                 kvm_mmu_put_page(vcpu, page, parent_pte);
 545                 *parent_pte = 0;
 546         }
 547         kvm_mmu_page_unlink_children(vcpu, page);
 548         hlist_del(&page->hash_link);
 549         kvm_mmu_free_page(vcpu, page->page_hpa);
 550 }
 551
 552 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 553 {
 554         unsigned index;
 555         struct hlist_head *bucket;
 556         struct kvm_mmu_page *page;
 557         struct hlist_node *node, *n;
 558         int r;
 559
 560         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 561         r = 0;
 562         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 563         bucket = &vcpu->kvm->mmu_page_hash[index];
 564         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 565                 if (page->gfn == gfn && !page->role.metaphysical) {
 566                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 567                                  page->role.word);
 568                         kvm_mmu_zap_page(vcpu, page);
 569                         r = 1;
 570                 }
 571         return r;
 572 }
 573
 574 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 575 {
 576         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 577         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 578
 579         __set_bit(slot, &page_head->slot_bitmap);
 580 }
 581
 582 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 583 {
 584         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 585
 586         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 587 }
 588
 589 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 590 {
 591         struct kvm_memory_slot *slot;
 592         struct page *page;
 593
 594         ASSERT((gpa & HPA_ERR_MASK) == 0);
 595         slot = gfn_to_memslot(vcpu->kvm, gpa >> PAGE_SHIFT);
 596         if (!slot)
 597                 return gpa | HPA_ERR_MASK;
 598         page = gfn_to_page(slot, gpa >> PAGE_SHIFT);
 599         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 600                 | (gpa & (PAGE_SIZE-1));
 601 }
 602
 603 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 604 {
 605         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 606
 607         if (gpa == UNMAPPED_GVA)
 608                 return UNMAPPED_GVA;
 609         return gpa_to_hpa(vcpu, gpa);
 610 }
 611
 612 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 613 {
 614 }
 615
 616 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 617 {
 618         int level = PT32E_ROOT_LEVEL;
 619         hpa_t table_addr = vcpu->mmu.root_hpa;
 620
 621         for (; ; level--) {
 622                 u32 index = PT64_INDEX(v, level);
 623                 u64 *table;
 624                 u64 pte;
 625
 626                 ASSERT(VALID_PAGE(table_addr));
 627                 table = __va(table_addr);
 628
 629                 if (level == 1) {
 630                         pte = table[index];
 631                         if (is_present_pte(pte) && is_writeble_pte(pte))
 632                                 return 0;
 633                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 634                         page_header_update_slot(vcpu->kvm, table, v);
 635                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 636                                                                 PT_USER_MASK;
 637                         rmap_add(vcpu->kvm, &table[index]);
 638                         return 0;
 639                 }
 640
 641                 if (table[index] == 0) {
 642                         struct kvm_mmu_page *new_table;
 643                         gfn_t pseudo_gfn;
 644
 645                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 646                                 >> PAGE_SHIFT;
 647                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 648                                                      v, level - 1,
 649                                                      1, &table[index]);
 650                         if (!new_table) {
 651                                 pgprintk("nonpaging_map: ENOMEM\n");
 652                                 return -ENOMEM;
 653                         }
 654
 655                         table[index] = new_table->page_hpa | PT_PRESENT_MASK
 656                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 657                 }
 658                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 659         }
 660 }
 661
 662 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 663 {
 664         int i;
 665
 666 #ifdef CONFIG_X86_64
 667         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 668                 hpa_t root = vcpu->mmu.root_hpa;
 669
 670                 ASSERT(VALID_PAGE(root));
 671                 vcpu->mmu.root_hpa = INVALID_PAGE;
 672                 return;
 673         }
 674 #endif
 675         for (i = 0; i < 4; ++i) {
 676                 hpa_t root = vcpu->mmu.pae_root[i];
 677
 678                 ASSERT(VALID_PAGE(root));
 679                 root &= PT64_BASE_ADDR_MASK;
 680                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 681         }
 682         vcpu->mmu.root_hpa = INVALID_PAGE;
 683 }
 684
 685 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 686 {
 687         int i;
 688         gfn_t root_gfn;
 689         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 690
 691 #ifdef CONFIG_X86_64
 692         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 693                 hpa_t root = vcpu->mmu.root_hpa;
 694
 695                 ASSERT(!VALID_PAGE(root));
 696                 root = kvm_mmu_get_page(vcpu, root_gfn, 0,
 697                                         PT64_ROOT_LEVEL, 0, NULL)->page_hpa;
 698                 vcpu->mmu.root_hpa = root;
 699                 return;
 700         }
 701 #endif
 702         for (i = 0; i < 4; ++i) {
 703                 hpa_t root = vcpu->mmu.pae_root[i];
 704
 705                 ASSERT(!VALID_PAGE(root));
 706                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL)
 707                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 708                 else if (vcpu->mmu.root_level == 0)
 709                         root_gfn = 0;
 710                 root = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 711                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 712                                         NULL)->page_hpa;
 713                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 714         }
 715         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 716 }
 717
 718 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 719 {
 720         return vaddr;
 721 }
 722
 723 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 724                                u32 error_code)
 725 {
 726         gpa_t addr = gva;
 727         hpa_t paddr;
 728
 729         ASSERT(vcpu);
 730         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 731
 732
 733         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 734
 735         if (is_error_hpa(paddr))
 736                 return 1;
 737
 738         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 739 }
 740
 741 static void nonpaging_free(struct kvm_vcpu *vcpu)
 742 {
 743         mmu_free_roots(vcpu);
 744 }
 745
 746 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 747 {
 748         struct kvm_mmu *context = &vcpu->mmu;
 749
 750         context->new_cr3 = nonpaging_new_cr3;
 751         context->page_fault = nonpaging_page_fault;
 752         context->gva_to_gpa = nonpaging_gva_to_gpa;
 753         context->free = nonpaging_free;
 754         context->root_level = 0;
 755         context->shadow_root_level = PT32E_ROOT_LEVEL;
 756         mmu_alloc_roots(vcpu);
 757         ASSERT(VALID_PAGE(context->root_hpa));
 758         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 759         return 0;
 760 }
 761
 762 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 763 {
 764         ++kvm_stat.tlb_flush;
 765         kvm_arch_ops->tlb_flush(vcpu);
 766 }
 767
 768 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 769 {
 770         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 771         mmu_free_roots(vcpu);
 772         mmu_alloc_roots(vcpu);
 773         kvm_mmu_flush_tlb(vcpu);
 774         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 775 }
 776
 777 static void mark_pagetable_nonglobal(void *shadow_pte)
 778 {
 779         page_header(__pa(shadow_pte))->global = 0;
 780 }
 781
 782 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 783                              u64 *shadow_pte,
 784                              gpa_t gaddr,
 785                              int dirty,
 786                              u64 access_bits,
 787                              gfn_t gfn)
 788 {
 789         hpa_t paddr;
 790
 791         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 792         if (!dirty)
 793                 access_bits &= ~PT_WRITABLE_MASK;
 794
 795         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 796
 797         *shadow_pte |= access_bits;
 798
 799         if (!(*shadow_pte & PT_GLOBAL_MASK))
 800                 mark_pagetable_nonglobal(shadow_pte);
 801
 802         if (is_error_hpa(paddr)) {
 803                 *shadow_pte |= gaddr;
 804                 *shadow_pte |= PT_SHADOW_IO_MARK;
 805                 *shadow_pte &= ~PT_PRESENT_MASK;
 806                 return;
 807         }
 808
 809         *shadow_pte |= paddr;
 810
 811         if (access_bits & PT_WRITABLE_MASK) {
 812                 struct kvm_mmu_page *shadow;
 813
 814                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 815                 if (shadow) {
 816                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 817                                  __FUNCTION__, gfn);
 818                         access_bits &= ~PT_WRITABLE_MASK;
 819                         *shadow_pte &= ~PT_WRITABLE_MASK;
 820                 }
 821         }
 822
 823         if (access_bits & PT_WRITABLE_MASK)
 824                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 825
 826         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 827         rmap_add(vcpu->kvm, shadow_pte);
 828 }
 829
 830 static void inject_page_fault(struct kvm_vcpu *vcpu,
 831                               u64 addr,
 832                               u32 err_code)
 833 {
 834         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
 835 }
 836
 837 static inline int fix_read_pf(u64 *shadow_ent)
 838 {
 839         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
 840             !(*shadow_ent & PT_USER_MASK)) {
 841                 /*
 842                  * If supervisor write protect is disabled, we shadow kernel
 843                  * pages as user pages so we can trap the write access.
 844                  */
 845                 *shadow_ent |= PT_USER_MASK;
 846                 *shadow_ent &= ~PT_WRITABLE_MASK;
 847
 848                 return 1;
 849
 850         }
 851         return 0;
 852 }
 853
 854 static int may_access(u64 pte, int write, int user)
 855 {
 856
 857         if (user && !(pte & PT_USER_MASK))
 858                 return 0;
 859         if (write && !(pte & PT_WRITABLE_MASK))
 860                 return 0;
 861         return 1;
 862 }
 863
 864 static void paging_free(struct kvm_vcpu *vcpu)
 865 {
 866         nonpaging_free(vcpu);
 867 }
 868
 869 #define PTTYPE 64
 870 #include "paging_tmpl.h"
 871 #undef PTTYPE
 872
 873 #define PTTYPE 32
 874 #include "paging_tmpl.h"
 875 #undef PTTYPE
 876
 877 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
 878 {
 879         struct kvm_mmu *context = &vcpu->mmu;
 880
 881         ASSERT(is_pae(vcpu));
 882         context->new_cr3 = paging_new_cr3;
 883         context->page_fault = paging64_page_fault;
 884         context->gva_to_gpa = paging64_gva_to_gpa;
 885         context->free = paging_free;
 886         context->root_level = level;
 887         context->shadow_root_level = level;
 888         mmu_alloc_roots(vcpu);
 889         ASSERT(VALID_PAGE(context->root_hpa));
 890         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
 891                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
 892         return 0;
 893 }
 894
 895 static int paging64_init_context(struct kvm_vcpu *vcpu)
 896 {
 897         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
 898 }
 899
 900 static int paging32_init_context(struct kvm_vcpu *vcpu)
 901 {
 902         struct kvm_mmu *context = &vcpu->mmu;
 903
 904         context->new_cr3 = paging_new_cr3;
 905         context->page_fault = paging32_page_fault;
 906         context->gva_to_gpa = paging32_gva_to_gpa;
 907         context->free = paging_free;
 908         context->root_level = PT32_ROOT_LEVEL;
 909         context->shadow_root_level = PT32E_ROOT_LEVEL;
 910         mmu_alloc_roots(vcpu);
 911         ASSERT(VALID_PAGE(context->root_hpa));
 912         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
 913                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
 914         return 0;
 915 }
 916
 917 static int paging32E_init_context(struct kvm_vcpu *vcpu)
 918 {
 919         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
 920 }
 921
 922 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
 923 {
 924         ASSERT(vcpu);
 925         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
 926
 927         if (!is_paging(vcpu))
 928                 return nonpaging_init_context(vcpu);
 929         else if (is_long_mode(vcpu))
 930                 return paging64_init_context(vcpu);
 931         else if (is_pae(vcpu))
 932                 return paging32E_init_context(vcpu);
 933         else
 934                 return paging32_init_context(vcpu);
 935 }
 936
 937 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
 938 {
 939         ASSERT(vcpu);
 940         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
 941                 vcpu->mmu.free(vcpu);
 942                 vcpu->mmu.root_hpa = INVALID_PAGE;
 943         }
 944 }
 945
 946 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
 947 {
 948         destroy_kvm_mmu(vcpu);
 949         return init_kvm_mmu(vcpu);
 950 }
 951
 952 void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
 953 {
 954         gfn_t gfn = gpa >> PAGE_SHIFT;
 955         struct kvm_mmu_page *page;
 956         struct kvm_mmu_page *child;
 957         struct hlist_node *node;
 958         struct hlist_head *bucket;
 959         unsigned index;
 960         u64 *spte;
 961         u64 pte;
 962         unsigned offset = offset_in_page(gpa);
 963         unsigned page_offset;
 964         int level;
 965
 966         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
 967         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 968         bucket = &vcpu->kvm->mmu_page_hash[index];
 969         hlist_for_each_entry(page, node, bucket, hash_link) {
 970                 if (page->gfn != gfn || page->role.metaphysical)
 971                         continue;
 972                 page_offset = offset;
 973                 level = page->role.level;
 974                 if (page->role.glevels == PT32_ROOT_LEVEL) {
 975                         page_offset <<= 1;          /* 32->64 */
 976                         page_offset &= ~PAGE_MASK;
 977                 }
 978                 spte = __va(page->page_hpa);
 979                 spte += page_offset / sizeof(*spte);
 980                 pte = *spte;
 981                 if (is_present_pte(pte)) {
 982                         if (level == PT_PAGE_TABLE_LEVEL)
 983                                 rmap_remove(vcpu->kvm, spte);
 984                         else {
 985                                 child = page_header(pte & PT64_BASE_ADDR_MASK);
 986                                 mmu_page_remove_parent_pte(child, spte);
 987                         }
 988                 }
 989                 *spte = 0;
 990         }
 991 }
 992
 993 void kvm_mmu_post_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
 994 {
 995 }
 996
 997 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
 998 {
 999         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1000
1001         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1002 }
1003
1004 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1005 {
1006         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1007                 struct kvm_mmu_page *page;
1008
1009                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1010                                     struct kvm_mmu_page, link);
1011                 kvm_mmu_zap_page(vcpu, page);
1012         }
1013 }
1014 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1015
1016 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1017 {
1018         while (!list_empty(&vcpu->free_pages)) {
1019                 struct kvm_mmu_page *page;
1020
1021                 page = list_entry(vcpu->free_pages.next,
1022                                   struct kvm_mmu_page, link);
1023                 list_del(&page->link);
1024                 __free_page(pfn_to_page(page->page_hpa >> PAGE_SHIFT));
1025                 page->page_hpa = INVALID_PAGE;
1026         }
1027         free_page((unsigned long)vcpu->mmu.pae_root);
1028 }
1029
1030 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1031 {
1032         struct page *page;
1033         int i;
1034
1035         ASSERT(vcpu);
1036
1037         for (i = 0; i < KVM_NUM_MMU_PAGES; i++) {
1038                 struct kvm_mmu_page *page_header = &vcpu->page_header_buf[i];
1039
1040                 INIT_LIST_HEAD(&page_header->link);
1041                 if ((page = alloc_page(GFP_KERNEL)) == NULL)
1042                         goto error_1;
1043                 page->private = (unsigned long)page_header;
1044                 page_header->page_hpa = (hpa_t)page_to_pfn(page) << PAGE_SHIFT;
1045                 memset(__va(page_header->page_hpa), 0, PAGE_SIZE);
1046                 list_add(&page_header->link, &vcpu->free_pages);
1047                 ++vcpu->kvm->n_free_mmu_pages;
1048         }
1049
1050         /*
1051          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1052          * Therefore we need to allocate shadow page tables in the first
1053          * 4GB of memory, which happens to fit the DMA32 zone.
1054          */
1055         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1056         if (!page)
1057                 goto error_1;
1058         vcpu->mmu.pae_root = page_address(page);
1059         for (i = 0; i < 4; ++i)
1060                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1061
1062         return 0;
1063
1064 error_1:
1065         free_mmu_pages(vcpu);
1066         return -ENOMEM;
1067 }
1068
1069 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1070 {
1071         ASSERT(vcpu);
1072         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1073         ASSERT(list_empty(&vcpu->free_pages));
1074
1075         return alloc_mmu_pages(vcpu);
1076 }
1077
1078 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1079 {
1080         ASSERT(vcpu);
1081         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1082         ASSERT(!list_empty(&vcpu->free_pages));
1083
1084         return init_kvm_mmu(vcpu);
1085 }
1086
1087 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1088 {
1089         ASSERT(vcpu);
1090
1091         destroy_kvm_mmu(vcpu);
1092         free_mmu_pages(vcpu);
1093 }
1094
1095 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1096 {
1097         struct kvm_mmu_page *page;
1098
1099         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1100                 int i;
1101                 u64 *pt;
1102
1103                 if (!test_bit(slot, &page->slot_bitmap))
1104                         continue;
1105
1106                 pt = __va(page->page_hpa);
1107                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1108                         /* avoid RMW */
1109                         if (pt[i] & PT_WRITABLE_MASK) {
1110                                 rmap_remove(kvm, &pt[i]);
1111                                 pt[i] &= ~PT_WRITABLE_MASK;
1112                         }
1113         }
1114 }