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         u64 *pos;
 309         u64 *end;
 310
 311         for (pos = __va(page_hpa), end = pos + PAGE_SIZE / sizeof(u64);
 312                       pos != end; pos++)
 313                 if (*pos != 0) {
 314                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 315                                pos, *pos);
 316                         return 0;
 317                 }
 318         return 1;
 319 }
 320
 321 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu, hpa_t page_hpa)
 322 {
 323         struct kvm_mmu_page *page_head = page_header(page_hpa);
 324
 325         ASSERT(is_empty_shadow_page(page_hpa));
 326         list_del(&page_head->link);
 327         page_head->page_hpa = page_hpa;
 328         list_add(&page_head->link, &vcpu->free_pages);
 329         ++vcpu->kvm->n_free_mmu_pages;
 330 }
 331
 332 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 333 {
 334         return gfn;
 335 }
 336
 337 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 338                                                u64 *parent_pte)
 339 {
 340         struct kvm_mmu_page *page;
 341
 342         if (list_empty(&vcpu->free_pages))
 343                 return NULL;
 344
 345         page = list_entry(vcpu->free_pages.next, struct kvm_mmu_page, link);
 346         list_del(&page->link);
 347         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 348         ASSERT(is_empty_shadow_page(page->page_hpa));
 349         page->slot_bitmap = 0;
 350         page->global = 1;
 351         page->multimapped = 0;
 352         page->parent_pte = parent_pte;
 353         --vcpu->kvm->n_free_mmu_pages;
 354         return page;
 355 }
 356
 357 static void mmu_page_add_parent_pte(struct kvm_mmu_page *page, u64 *parent_pte)
 358 {
 359         struct kvm_pte_chain *pte_chain;
 360         struct hlist_node *node;
 361         int i;
 362
 363         if (!parent_pte)
 364                 return;
 365         if (!page->multimapped) {
 366                 u64 *old = page->parent_pte;
 367
 368                 if (!old) {
 369                         page->parent_pte = parent_pte;
 370                         return;
 371                 }
 372                 page->multimapped = 1;
 373                 pte_chain = kzalloc(sizeof(struct kvm_pte_chain), GFP_NOWAIT);
 374                 BUG_ON(!pte_chain);
 375                 INIT_HLIST_HEAD(&page->parent_ptes);
 376                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 377                 pte_chain->parent_ptes[0] = old;
 378         }
 379         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 380                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 381                         continue;
 382                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 383                         if (!pte_chain->parent_ptes[i]) {
 384                                 pte_chain->parent_ptes[i] = parent_pte;
 385                                 return;
 386                         }
 387         }
 388         pte_chain = kzalloc(sizeof(struct kvm_pte_chain), GFP_NOWAIT);
 389         BUG_ON(!pte_chain);
 390         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 391         pte_chain->parent_ptes[0] = parent_pte;
 392 }
 393
 394 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 395                                        u64 *parent_pte)
 396 {
 397         struct kvm_pte_chain *pte_chain;
 398         struct hlist_node *node;
 399         int i;
 400
 401         if (!page->multimapped) {
 402                 BUG_ON(page->parent_pte != parent_pte);
 403                 page->parent_pte = NULL;
 404                 return;
 405         }
 406         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 407                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 408                         if (!pte_chain->parent_ptes[i])
 409                                 break;
 410                         if (pte_chain->parent_ptes[i] != parent_pte)
 411                                 continue;
 412                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 413                                 && pte_chain->parent_ptes[i + 1]) {
 414                                 pte_chain->parent_ptes[i]
 415                                         = pte_chain->parent_ptes[i + 1];
 416                                 ++i;
 417                         }
 418                         pte_chain->parent_ptes[i] = NULL;
 419                         if (i == 0) {
 420                                 hlist_del(&pte_chain->link);
 421                                 kfree(pte_chain);
 422                                 if (hlist_empty(&page->parent_ptes)) {
 423                                         page->multimapped = 0;
 424                                         page->parent_pte = NULL;
 425                                 }
 426                         }
 427                         return;
 428                 }
 429         BUG();
 430 }
 431
 432 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 433                                                 gfn_t gfn)
 434 {
 435         unsigned index;
 436         struct hlist_head *bucket;
 437         struct kvm_mmu_page *page;
 438         struct hlist_node *node;
 439
 440         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 441         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 442         bucket = &vcpu->kvm->mmu_page_hash[index];
 443         hlist_for_each_entry(page, node, bucket, hash_link)
 444                 if (page->gfn == gfn && !page->role.metaphysical) {
 445                         pgprintk("%s: found role %x\n",
 446                                  __FUNCTION__, page->role.word);
 447                         return page;
 448                 }
 449         return NULL;
 450 }
 451
 452 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 453                                              gfn_t gfn,
 454                                              gva_t gaddr,
 455                                              unsigned level,
 456                                              int metaphysical,
 457                                              u64 *parent_pte)
 458 {
 459         union kvm_mmu_page_role role;
 460         unsigned index;
 461         unsigned quadrant;
 462         struct hlist_head *bucket;
 463         struct kvm_mmu_page *page;
 464         struct hlist_node *node;
 465
 466         role.word = 0;
 467         role.glevels = vcpu->mmu.root_level;
 468         role.level = level;
 469         role.metaphysical = metaphysical;
 470         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 471                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 472                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 473                 role.quadrant = quadrant;
 474         }
 475         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 476                  gfn, role.word);
 477         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 478         bucket = &vcpu->kvm->mmu_page_hash[index];
 479         hlist_for_each_entry(page, node, bucket, hash_link)
 480                 if (page->gfn == gfn && page->role.word == role.word) {
 481                         mmu_page_add_parent_pte(page, parent_pte);
 482                         pgprintk("%s: found\n", __FUNCTION__);
 483                         return page;
 484                 }
 485         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 486         if (!page)
 487                 return page;
 488         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 489         page->gfn = gfn;
 490         page->role = role;
 491         hlist_add_head(&page->hash_link, bucket);
 492         if (!metaphysical)
 493                 rmap_write_protect(vcpu->kvm, gfn);
 494         return page;
 495 }
 496
 497 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 498                                          struct kvm_mmu_page *page)
 499 {
 500         unsigned i;
 501         u64 *pt;
 502         u64 ent;
 503
 504         pt = __va(page->page_hpa);
 505
 506         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 507                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 508                         if (pt[i] & PT_PRESENT_MASK)
 509                                 rmap_remove(vcpu->kvm, &pt[i]);
 510                         pt[i] = 0;
 511                 }
 512                 return;
 513         }
 514
 515         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 516                 ent = pt[i];
 517
 518                 pt[i] = 0;
 519                 if (!(ent & PT_PRESENT_MASK))
 520                         continue;
 521                 ent &= PT64_BASE_ADDR_MASK;
 522                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 523         }
 524 }
 525
 526 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 527                              struct kvm_mmu_page *page,
 528                              u64 *parent_pte)
 529 {
 530         mmu_page_remove_parent_pte(page, parent_pte);
 531 }
 532
 533 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 534                              struct kvm_mmu_page *page)
 535 {
 536         u64 *parent_pte;
 537
 538         while (page->multimapped || page->parent_pte) {
 539                 if (!page->multimapped)
 540                         parent_pte = page->parent_pte;
 541                 else {
 542                         struct kvm_pte_chain *chain;
 543
 544                         chain = container_of(page->parent_ptes.first,
 545                                              struct kvm_pte_chain, link);
 546                         parent_pte = chain->parent_ptes[0];
 547                 }
 548                 BUG_ON(!parent_pte);
 549                 kvm_mmu_put_page(vcpu, page, parent_pte);
 550                 *parent_pte = 0;
 551         }
 552         kvm_mmu_page_unlink_children(vcpu, page);
 553         hlist_del(&page->hash_link);
 554         kvm_mmu_free_page(vcpu, page->page_hpa);
 555 }
 556
 557 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 558 {
 559         unsigned index;
 560         struct hlist_head *bucket;
 561         struct kvm_mmu_page *page;
 562         struct hlist_node *node, *n;
 563         int r;
 564
 565         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 566         r = 0;
 567         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 568         bucket = &vcpu->kvm->mmu_page_hash[index];
 569         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 570                 if (page->gfn == gfn && !page->role.metaphysical) {
 571                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 572                                  page->role.word);
 573                         kvm_mmu_zap_page(vcpu, page);
 574                         r = 1;
 575                 }
 576         return r;
 577 }
 578
 579 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 580 {
 581         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 582         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 583
 584         __set_bit(slot, &page_head->slot_bitmap);
 585 }
 586
 587 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 588 {
 589         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 590
 591         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 592 }
 593
 594 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 595 {
 596         struct kvm_memory_slot *slot;
 597         struct page *page;
 598
 599         ASSERT((gpa & HPA_ERR_MASK) == 0);
 600         slot = gfn_to_memslot(vcpu->kvm, gpa >> PAGE_SHIFT);
 601         if (!slot)
 602                 return gpa | HPA_ERR_MASK;
 603         page = gfn_to_page(slot, gpa >> PAGE_SHIFT);
 604         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 605                 | (gpa & (PAGE_SIZE-1));
 606 }
 607
 608 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 609 {
 610         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 611
 612         if (gpa == UNMAPPED_GVA)
 613                 return UNMAPPED_GVA;
 614         return gpa_to_hpa(vcpu, gpa);
 615 }
 616
 617 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 618 {
 619 }
 620
 621 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 622 {
 623         int level = PT32E_ROOT_LEVEL;
 624         hpa_t table_addr = vcpu->mmu.root_hpa;
 625
 626         for (; ; level--) {
 627                 u32 index = PT64_INDEX(v, level);
 628                 u64 *table;
 629                 u64 pte;
 630
 631                 ASSERT(VALID_PAGE(table_addr));
 632                 table = __va(table_addr);
 633
 634                 if (level == 1) {
 635                         pte = table[index];
 636                         if (is_present_pte(pte) && is_writeble_pte(pte))
 637                                 return 0;
 638                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 639                         page_header_update_slot(vcpu->kvm, table, v);
 640                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 641                                                                 PT_USER_MASK;
 642                         rmap_add(vcpu->kvm, &table[index]);
 643                         return 0;
 644                 }
 645
 646                 if (table[index] == 0) {
 647                         struct kvm_mmu_page *new_table;
 648                         gfn_t pseudo_gfn;
 649
 650                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 651                                 >> PAGE_SHIFT;
 652                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 653                                                      v, level - 1,
 654                                                      1, &table[index]);
 655                         if (!new_table) {
 656                                 pgprintk("nonpaging_map: ENOMEM\n");
 657                                 return -ENOMEM;
 658                         }
 659
 660                         table[index] = new_table->page_hpa | PT_PRESENT_MASK
 661                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 662                 }
 663                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 664         }
 665 }
 666
 667 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 668 {
 669         int i;
 670
 671 #ifdef CONFIG_X86_64
 672         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 673                 hpa_t root = vcpu->mmu.root_hpa;
 674
 675                 ASSERT(VALID_PAGE(root));
 676                 vcpu->mmu.root_hpa = INVALID_PAGE;
 677                 return;
 678         }
 679 #endif
 680         for (i = 0; i < 4; ++i) {
 681                 hpa_t root = vcpu->mmu.pae_root[i];
 682
 683                 ASSERT(VALID_PAGE(root));
 684                 root &= PT64_BASE_ADDR_MASK;
 685                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 686         }
 687         vcpu->mmu.root_hpa = INVALID_PAGE;
 688 }
 689
 690 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 691 {
 692         int i;
 693         gfn_t root_gfn;
 694         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 695
 696 #ifdef CONFIG_X86_64
 697         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 698                 hpa_t root = vcpu->mmu.root_hpa;
 699
 700                 ASSERT(!VALID_PAGE(root));
 701                 root = kvm_mmu_get_page(vcpu, root_gfn, 0,
 702                                         PT64_ROOT_LEVEL, 0, NULL)->page_hpa;
 703                 vcpu->mmu.root_hpa = root;
 704                 return;
 705         }
 706 #endif
 707         for (i = 0; i < 4; ++i) {
 708                 hpa_t root = vcpu->mmu.pae_root[i];
 709
 710                 ASSERT(!VALID_PAGE(root));
 711                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL)
 712                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 713                 else if (vcpu->mmu.root_level == 0)
 714                         root_gfn = 0;
 715                 root = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 716                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 717                                         NULL)->page_hpa;
 718                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 719         }
 720         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 721 }
 722
 723 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 724 {
 725         return vaddr;
 726 }
 727
 728 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 729                                u32 error_code)
 730 {
 731         gpa_t addr = gva;
 732         hpa_t paddr;
 733
 734         ASSERT(vcpu);
 735         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 736
 737
 738         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 739
 740         if (is_error_hpa(paddr))
 741                 return 1;
 742
 743         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 744 }
 745
 746 static void nonpaging_free(struct kvm_vcpu *vcpu)
 747 {
 748         mmu_free_roots(vcpu);
 749 }
 750
 751 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 752 {
 753         struct kvm_mmu *context = &vcpu->mmu;
 754
 755         context->new_cr3 = nonpaging_new_cr3;
 756         context->page_fault = nonpaging_page_fault;
 757         context->gva_to_gpa = nonpaging_gva_to_gpa;
 758         context->free = nonpaging_free;
 759         context->root_level = 0;
 760         context->shadow_root_level = PT32E_ROOT_LEVEL;
 761         mmu_alloc_roots(vcpu);
 762         ASSERT(VALID_PAGE(context->root_hpa));
 763         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 764         return 0;
 765 }
 766
 767 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 768 {
 769         ++kvm_stat.tlb_flush;
 770         kvm_arch_ops->tlb_flush(vcpu);
 771 }
 772
 773 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 774 {
 775         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 776         mmu_free_roots(vcpu);
 777         mmu_alloc_roots(vcpu);
 778         kvm_mmu_flush_tlb(vcpu);
 779         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 780 }
 781
 782 static void mark_pagetable_nonglobal(void *shadow_pte)
 783 {
 784         page_header(__pa(shadow_pte))->global = 0;
 785 }
 786
 787 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 788                              u64 *shadow_pte,
 789                              gpa_t gaddr,
 790                              int dirty,
 791                              u64 access_bits,
 792                              gfn_t gfn)
 793 {
 794         hpa_t paddr;
 795
 796         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 797         if (!dirty)
 798                 access_bits &= ~PT_WRITABLE_MASK;
 799
 800         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 801
 802         *shadow_pte |= access_bits;
 803
 804         if (!(*shadow_pte & PT_GLOBAL_MASK))
 805                 mark_pagetable_nonglobal(shadow_pte);
 806
 807         if (is_error_hpa(paddr)) {
 808                 *shadow_pte |= gaddr;
 809                 *shadow_pte |= PT_SHADOW_IO_MARK;
 810                 *shadow_pte &= ~PT_PRESENT_MASK;
 811                 return;
 812         }
 813
 814         *shadow_pte |= paddr;
 815
 816         if (access_bits & PT_WRITABLE_MASK) {
 817                 struct kvm_mmu_page *shadow;
 818
 819                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 820                 if (shadow) {
 821                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 822                                  __FUNCTION__, gfn);
 823                         access_bits &= ~PT_WRITABLE_MASK;
 824                         *shadow_pte &= ~PT_WRITABLE_MASK;
 825                 }
 826         }
 827
 828         if (access_bits & PT_WRITABLE_MASK)
 829                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 830
 831         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 832         rmap_add(vcpu->kvm, shadow_pte);
 833 }
 834
 835 static void inject_page_fault(struct kvm_vcpu *vcpu,
 836                               u64 addr,
 837                               u32 err_code)
 838 {
 839         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
 840 }
 841
 842 static inline int fix_read_pf(u64 *shadow_ent)
 843 {
 844         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
 845             !(*shadow_ent & PT_USER_MASK)) {
 846                 /*
 847                  * If supervisor write protect is disabled, we shadow kernel
 848                  * pages as user pages so we can trap the write access.
 849                  */
 850                 *shadow_ent |= PT_USER_MASK;
 851                 *shadow_ent &= ~PT_WRITABLE_MASK;
 852
 853                 return 1;
 854
 855         }
 856         return 0;
 857 }
 858
 859 static int may_access(u64 pte, int write, int user)
 860 {
 861
 862         if (user && !(pte & PT_USER_MASK))
 863                 return 0;
 864         if (write && !(pte & PT_WRITABLE_MASK))
 865                 return 0;
 866         return 1;
 867 }
 868
 869 static void paging_free(struct kvm_vcpu *vcpu)
 870 {
 871         nonpaging_free(vcpu);
 872 }
 873
 874 #define PTTYPE 64
 875 #include "paging_tmpl.h"
 876 #undef PTTYPE
 877
 878 #define PTTYPE 32
 879 #include "paging_tmpl.h"
 880 #undef PTTYPE
 881
 882 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
 883 {
 884         struct kvm_mmu *context = &vcpu->mmu;
 885
 886         ASSERT(is_pae(vcpu));
 887         context->new_cr3 = paging_new_cr3;
 888         context->page_fault = paging64_page_fault;
 889         context->gva_to_gpa = paging64_gva_to_gpa;
 890         context->free = paging_free;
 891         context->root_level = level;
 892         context->shadow_root_level = level;
 893         mmu_alloc_roots(vcpu);
 894         ASSERT(VALID_PAGE(context->root_hpa));
 895         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
 896                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
 897         return 0;
 898 }
 899
 900 static int paging64_init_context(struct kvm_vcpu *vcpu)
 901 {
 902         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
 903 }
 904
 905 static int paging32_init_context(struct kvm_vcpu *vcpu)
 906 {
 907         struct kvm_mmu *context = &vcpu->mmu;
 908
 909         context->new_cr3 = paging_new_cr3;
 910         context->page_fault = paging32_page_fault;
 911         context->gva_to_gpa = paging32_gva_to_gpa;
 912         context->free = paging_free;
 913         context->root_level = PT32_ROOT_LEVEL;
 914         context->shadow_root_level = PT32E_ROOT_LEVEL;
 915         mmu_alloc_roots(vcpu);
 916         ASSERT(VALID_PAGE(context->root_hpa));
 917         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
 918                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
 919         return 0;
 920 }
 921
 922 static int paging32E_init_context(struct kvm_vcpu *vcpu)
 923 {
 924         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
 925 }
 926
 927 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
 928 {
 929         ASSERT(vcpu);
 930         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
 931
 932         if (!is_paging(vcpu))
 933                 return nonpaging_init_context(vcpu);
 934         else if (is_long_mode(vcpu))
 935                 return paging64_init_context(vcpu);
 936         else if (is_pae(vcpu))
 937                 return paging32E_init_context(vcpu);
 938         else
 939                 return paging32_init_context(vcpu);
 940 }
 941
 942 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
 943 {
 944         ASSERT(vcpu);
 945         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
 946                 vcpu->mmu.free(vcpu);
 947                 vcpu->mmu.root_hpa = INVALID_PAGE;
 948         }
 949 }
 950
 951 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
 952 {
 953         destroy_kvm_mmu(vcpu);
 954         return init_kvm_mmu(vcpu);
 955 }
 956
 957 void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
 958 {
 959         gfn_t gfn = gpa >> PAGE_SHIFT;
 960         struct kvm_mmu_page *page;
 961         struct kvm_mmu_page *child;
 962         struct hlist_node *node, *n;
 963         struct hlist_head *bucket;
 964         unsigned index;
 965         u64 *spte;
 966         u64 pte;
 967         unsigned offset = offset_in_page(gpa);
 968         unsigned pte_size;
 969         unsigned page_offset;
 970         unsigned misaligned;
 971         int level;
 972         int flooded = 0;
 973
 974         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
 975         if (gfn == vcpu->last_pt_write_gfn) {
 976                 ++vcpu->last_pt_write_count;
 977                 if (vcpu->last_pt_write_count >= 3)
 978                         flooded = 1;
 979         } else {
 980                 vcpu->last_pt_write_gfn = gfn;
 981                 vcpu->last_pt_write_count = 1;
 982         }
 983         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 984         bucket = &vcpu->kvm->mmu_page_hash[index];
 985         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
 986                 if (page->gfn != gfn || page->role.metaphysical)
 987                         continue;
 988                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
 989                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
 990                 if (misaligned || flooded) {
 991                         /*
 992                          * Misaligned accesses are too much trouble to fix
 993                          * up; also, they usually indicate a page is not used
 994                          * as a page table.
 995                          *
 996                          * If we're seeing too many writes to a page,
 997                          * it may no longer be a page table, or we may be
 998                          * forking, in which case it is better to unmap the
 999                          * page.
1000                          */
1001                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1002                                  gpa, bytes, page->role.word);
1003                         kvm_mmu_zap_page(vcpu, page);
1004                         continue;
1005                 }
1006                 page_offset = offset;
1007                 level = page->role.level;
1008                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1009                         page_offset <<= 1;          /* 32->64 */
1010                         page_offset &= ~PAGE_MASK;
1011                 }
1012                 spte = __va(page->page_hpa);
1013                 spte += page_offset / sizeof(*spte);
1014                 pte = *spte;
1015                 if (is_present_pte(pte)) {
1016                         if (level == PT_PAGE_TABLE_LEVEL)
1017                                 rmap_remove(vcpu->kvm, spte);
1018                         else {
1019                                 child = page_header(pte & PT64_BASE_ADDR_MASK);
1020                                 mmu_page_remove_parent_pte(child, spte);
1021                         }
1022                 }
1023                 *spte = 0;
1024         }
1025 }
1026
1027 void kvm_mmu_post_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1028 {
1029 }
1030
1031 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1032 {
1033         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1034
1035         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1036 }
1037
1038 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1039 {
1040         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1041                 struct kvm_mmu_page *page;
1042
1043                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1044                                     struct kvm_mmu_page, link);
1045                 kvm_mmu_zap_page(vcpu, page);
1046         }
1047 }
1048 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1049
1050 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1051 {
1052         while (!list_empty(&vcpu->free_pages)) {
1053                 struct kvm_mmu_page *page;
1054
1055                 page = list_entry(vcpu->free_pages.next,
1056                                   struct kvm_mmu_page, link);
1057                 list_del(&page->link);
1058                 __free_page(pfn_to_page(page->page_hpa >> PAGE_SHIFT));
1059                 page->page_hpa = INVALID_PAGE;
1060         }
1061         free_page((unsigned long)vcpu->mmu.pae_root);
1062 }
1063
1064 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1065 {
1066         struct page *page;
1067         int i;
1068
1069         ASSERT(vcpu);
1070
1071         for (i = 0; i < KVM_NUM_MMU_PAGES; i++) {
1072                 struct kvm_mmu_page *page_header = &vcpu->page_header_buf[i];
1073
1074                 INIT_LIST_HEAD(&page_header->link);
1075                 if ((page = alloc_page(GFP_KERNEL)) == NULL)
1076                         goto error_1;
1077                 page->private = (unsigned long)page_header;
1078                 page_header->page_hpa = (hpa_t)page_to_pfn(page) << PAGE_SHIFT;
1079                 memset(__va(page_header->page_hpa), 0, PAGE_SIZE);
1080                 list_add(&page_header->link, &vcpu->free_pages);
1081                 ++vcpu->kvm->n_free_mmu_pages;
1082         }
1083
1084         /*
1085          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1086          * Therefore we need to allocate shadow page tables in the first
1087          * 4GB of memory, which happens to fit the DMA32 zone.
1088          */
1089         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1090         if (!page)
1091                 goto error_1;
1092         vcpu->mmu.pae_root = page_address(page);
1093         for (i = 0; i < 4; ++i)
1094                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1095
1096         return 0;
1097
1098 error_1:
1099         free_mmu_pages(vcpu);
1100         return -ENOMEM;
1101 }
1102
1103 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1104 {
1105         ASSERT(vcpu);
1106         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1107         ASSERT(list_empty(&vcpu->free_pages));
1108
1109         return alloc_mmu_pages(vcpu);
1110 }
1111
1112 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1113 {
1114         ASSERT(vcpu);
1115         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1116         ASSERT(!list_empty(&vcpu->free_pages));
1117
1118         return init_kvm_mmu(vcpu);
1119 }
1120
1121 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1122 {
1123         ASSERT(vcpu);
1124
1125         destroy_kvm_mmu(vcpu);
1126         free_mmu_pages(vcpu);
1127 }
1128
1129 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1130 {
1131         struct kvm_mmu_page *page;
1132
1133         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1134                 int i;
1135                 u64 *pt;
1136
1137                 if (!test_bit(slot, &page->slot_bitmap))
1138                         continue;
1139
1140                 pt = __va(page->page_hpa);
1141                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1142                         /* avoid RMW */
1143                         if (pt[i] & PT_WRITABLE_MASK) {
1144                                 rmap_remove(kvm, &pt[i]);
1145                                 pt[i] &= ~PT_WRITABLE_MASK;
1146                         }
1147         }
1148 }