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         if (!page->root_count) {
 554                 hlist_del(&page->hash_link);
 555                 kvm_mmu_free_page(vcpu, page->page_hpa);
 556         } else {
 557                 list_del(&page->link);
 558                 list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 559         }
 560 }
 561
 562 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 563 {
 564         unsigned index;
 565         struct hlist_head *bucket;
 566         struct kvm_mmu_page *page;
 567         struct hlist_node *node, *n;
 568         int r;
 569
 570         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 571         r = 0;
 572         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 573         bucket = &vcpu->kvm->mmu_page_hash[index];
 574         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 575                 if (page->gfn == gfn && !page->role.metaphysical) {
 576                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 577                                  page->role.word);
 578                         kvm_mmu_zap_page(vcpu, page);
 579                         r = 1;
 580                 }
 581         return r;
 582 }
 583
 584 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 585 {
 586         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 587         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 588
 589         __set_bit(slot, &page_head->slot_bitmap);
 590 }
 591
 592 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 593 {
 594         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 595
 596         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 597 }
 598
 599 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 600 {
 601         struct kvm_memory_slot *slot;
 602         struct page *page;
 603
 604         ASSERT((gpa & HPA_ERR_MASK) == 0);
 605         slot = gfn_to_memslot(vcpu->kvm, gpa >> PAGE_SHIFT);
 606         if (!slot)
 607                 return gpa | HPA_ERR_MASK;
 608         page = gfn_to_page(slot, gpa >> PAGE_SHIFT);
 609         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 610                 | (gpa & (PAGE_SIZE-1));
 611 }
 612
 613 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 614 {
 615         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 616
 617         if (gpa == UNMAPPED_GVA)
 618                 return UNMAPPED_GVA;
 619         return gpa_to_hpa(vcpu, gpa);
 620 }
 621
 622 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 623 {
 624 }
 625
 626 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 627 {
 628         int level = PT32E_ROOT_LEVEL;
 629         hpa_t table_addr = vcpu->mmu.root_hpa;
 630
 631         for (; ; level--) {
 632                 u32 index = PT64_INDEX(v, level);
 633                 u64 *table;
 634                 u64 pte;
 635
 636                 ASSERT(VALID_PAGE(table_addr));
 637                 table = __va(table_addr);
 638
 639                 if (level == 1) {
 640                         pte = table[index];
 641                         if (is_present_pte(pte) && is_writeble_pte(pte))
 642                                 return 0;
 643                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 644                         page_header_update_slot(vcpu->kvm, table, v);
 645                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 646                                                                 PT_USER_MASK;
 647                         rmap_add(vcpu->kvm, &table[index]);
 648                         return 0;
 649                 }
 650
 651                 if (table[index] == 0) {
 652                         struct kvm_mmu_page *new_table;
 653                         gfn_t pseudo_gfn;
 654
 655                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 656                                 >> PAGE_SHIFT;
 657                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 658                                                      v, level - 1,
 659                                                      1, &table[index]);
 660                         if (!new_table) {
 661                                 pgprintk("nonpaging_map: ENOMEM\n");
 662                                 return -ENOMEM;
 663                         }
 664
 665                         table[index] = new_table->page_hpa | PT_PRESENT_MASK
 666                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 667                 }
 668                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 669         }
 670 }
 671
 672 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 673 {
 674         int i;
 675         struct kvm_mmu_page *page;
 676
 677 #ifdef CONFIG_X86_64
 678         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 679                 hpa_t root = vcpu->mmu.root_hpa;
 680
 681                 ASSERT(VALID_PAGE(root));
 682                 page = page_header(root);
 683                 --page->root_count;
 684                 vcpu->mmu.root_hpa = INVALID_PAGE;
 685                 return;
 686         }
 687 #endif
 688         for (i = 0; i < 4; ++i) {
 689                 hpa_t root = vcpu->mmu.pae_root[i];
 690
 691                 ASSERT(VALID_PAGE(root));
 692                 root &= PT64_BASE_ADDR_MASK;
 693                 page = page_header(root);
 694                 --page->root_count;
 695                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 696         }
 697         vcpu->mmu.root_hpa = INVALID_PAGE;
 698 }
 699
 700 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 701 {
 702         int i;
 703         gfn_t root_gfn;
 704         struct kvm_mmu_page *page;
 705
 706         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 707
 708 #ifdef CONFIG_X86_64
 709         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 710                 hpa_t root = vcpu->mmu.root_hpa;
 711
 712                 ASSERT(!VALID_PAGE(root));
 713                 root = kvm_mmu_get_page(vcpu, root_gfn, 0,
 714                                         PT64_ROOT_LEVEL, 0, NULL)->page_hpa;
 715                 page = page_header(root);
 716                 ++page->root_count;
 717                 vcpu->mmu.root_hpa = root;
 718                 return;
 719         }
 720 #endif
 721         for (i = 0; i < 4; ++i) {
 722                 hpa_t root = vcpu->mmu.pae_root[i];
 723
 724                 ASSERT(!VALID_PAGE(root));
 725                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL)
 726                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 727                 else if (vcpu->mmu.root_level == 0)
 728                         root_gfn = 0;
 729                 root = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 730                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 731                                         NULL)->page_hpa;
 732                 page = page_header(root);
 733                 ++page->root_count;
 734                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 735         }
 736         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 737 }
 738
 739 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 740 {
 741         return vaddr;
 742 }
 743
 744 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 745                                u32 error_code)
 746 {
 747         gpa_t addr = gva;
 748         hpa_t paddr;
 749
 750         ASSERT(vcpu);
 751         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 752
 753
 754         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 755
 756         if (is_error_hpa(paddr))
 757                 return 1;
 758
 759         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 760 }
 761
 762 static void nonpaging_free(struct kvm_vcpu *vcpu)
 763 {
 764         mmu_free_roots(vcpu);
 765 }
 766
 767 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 768 {
 769         struct kvm_mmu *context = &vcpu->mmu;
 770
 771         context->new_cr3 = nonpaging_new_cr3;
 772         context->page_fault = nonpaging_page_fault;
 773         context->gva_to_gpa = nonpaging_gva_to_gpa;
 774         context->free = nonpaging_free;
 775         context->root_level = 0;
 776         context->shadow_root_level = PT32E_ROOT_LEVEL;
 777         mmu_alloc_roots(vcpu);
 778         ASSERT(VALID_PAGE(context->root_hpa));
 779         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 780         return 0;
 781 }
 782
 783 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 784 {
 785         ++kvm_stat.tlb_flush;
 786         kvm_arch_ops->tlb_flush(vcpu);
 787 }
 788
 789 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 790 {
 791         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 792         mmu_free_roots(vcpu);
 793         mmu_alloc_roots(vcpu);
 794         kvm_mmu_flush_tlb(vcpu);
 795         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 796 }
 797
 798 static void mark_pagetable_nonglobal(void *shadow_pte)
 799 {
 800         page_header(__pa(shadow_pte))->global = 0;
 801 }
 802
 803 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 804                              u64 *shadow_pte,
 805                              gpa_t gaddr,
 806                              int dirty,
 807                              u64 access_bits,
 808                              gfn_t gfn)
 809 {
 810         hpa_t paddr;
 811
 812         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 813         if (!dirty)
 814                 access_bits &= ~PT_WRITABLE_MASK;
 815
 816         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 817
 818         *shadow_pte |= access_bits;
 819
 820         if (!(*shadow_pte & PT_GLOBAL_MASK))
 821                 mark_pagetable_nonglobal(shadow_pte);
 822
 823         if (is_error_hpa(paddr)) {
 824                 *shadow_pte |= gaddr;
 825                 *shadow_pte |= PT_SHADOW_IO_MARK;
 826                 *shadow_pte &= ~PT_PRESENT_MASK;
 827                 return;
 828         }
 829
 830         *shadow_pte |= paddr;
 831
 832         if (access_bits & PT_WRITABLE_MASK) {
 833                 struct kvm_mmu_page *shadow;
 834
 835                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 836                 if (shadow) {
 837                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 838                                  __FUNCTION__, gfn);
 839                         access_bits &= ~PT_WRITABLE_MASK;
 840                         *shadow_pte &= ~PT_WRITABLE_MASK;
 841                 }
 842         }
 843
 844         if (access_bits & PT_WRITABLE_MASK)
 845                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 846
 847         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 848         rmap_add(vcpu->kvm, shadow_pte);
 849 }
 850
 851 static void inject_page_fault(struct kvm_vcpu *vcpu,
 852                               u64 addr,
 853                               u32 err_code)
 854 {
 855         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
 856 }
 857
 858 static inline int fix_read_pf(u64 *shadow_ent)
 859 {
 860         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
 861             !(*shadow_ent & PT_USER_MASK)) {
 862                 /*
 863                  * If supervisor write protect is disabled, we shadow kernel
 864                  * pages as user pages so we can trap the write access.
 865                  */
 866                 *shadow_ent |= PT_USER_MASK;
 867                 *shadow_ent &= ~PT_WRITABLE_MASK;
 868
 869                 return 1;
 870
 871         }
 872         return 0;
 873 }
 874
 875 static int may_access(u64 pte, int write, int user)
 876 {
 877
 878         if (user && !(pte & PT_USER_MASK))
 879                 return 0;
 880         if (write && !(pte & PT_WRITABLE_MASK))
 881                 return 0;
 882         return 1;
 883 }
 884
 885 static void paging_free(struct kvm_vcpu *vcpu)
 886 {
 887         nonpaging_free(vcpu);
 888 }
 889
 890 #define PTTYPE 64
 891 #include "paging_tmpl.h"
 892 #undef PTTYPE
 893
 894 #define PTTYPE 32
 895 #include "paging_tmpl.h"
 896 #undef PTTYPE
 897
 898 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
 899 {
 900         struct kvm_mmu *context = &vcpu->mmu;
 901
 902         ASSERT(is_pae(vcpu));
 903         context->new_cr3 = paging_new_cr3;
 904         context->page_fault = paging64_page_fault;
 905         context->gva_to_gpa = paging64_gva_to_gpa;
 906         context->free = paging_free;
 907         context->root_level = level;
 908         context->shadow_root_level = level;
 909         mmu_alloc_roots(vcpu);
 910         ASSERT(VALID_PAGE(context->root_hpa));
 911         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
 912                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
 913         return 0;
 914 }
 915
 916 static int paging64_init_context(struct kvm_vcpu *vcpu)
 917 {
 918         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
 919 }
 920
 921 static int paging32_init_context(struct kvm_vcpu *vcpu)
 922 {
 923         struct kvm_mmu *context = &vcpu->mmu;
 924
 925         context->new_cr3 = paging_new_cr3;
 926         context->page_fault = paging32_page_fault;
 927         context->gva_to_gpa = paging32_gva_to_gpa;
 928         context->free = paging_free;
 929         context->root_level = PT32_ROOT_LEVEL;
 930         context->shadow_root_level = PT32E_ROOT_LEVEL;
 931         mmu_alloc_roots(vcpu);
 932         ASSERT(VALID_PAGE(context->root_hpa));
 933         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
 934                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
 935         return 0;
 936 }
 937
 938 static int paging32E_init_context(struct kvm_vcpu *vcpu)
 939 {
 940         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
 941 }
 942
 943 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
 944 {
 945         ASSERT(vcpu);
 946         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
 947
 948         if (!is_paging(vcpu))
 949                 return nonpaging_init_context(vcpu);
 950         else if (is_long_mode(vcpu))
 951                 return paging64_init_context(vcpu);
 952         else if (is_pae(vcpu))
 953                 return paging32E_init_context(vcpu);
 954         else
 955                 return paging32_init_context(vcpu);
 956 }
 957
 958 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
 959 {
 960         ASSERT(vcpu);
 961         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
 962                 vcpu->mmu.free(vcpu);
 963                 vcpu->mmu.root_hpa = INVALID_PAGE;
 964         }
 965 }
 966
 967 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
 968 {
 969         destroy_kvm_mmu(vcpu);
 970         return init_kvm_mmu(vcpu);
 971 }
 972
 973 void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
 974 {
 975         gfn_t gfn = gpa >> PAGE_SHIFT;
 976         struct kvm_mmu_page *page;
 977         struct kvm_mmu_page *child;
 978         struct hlist_node *node, *n;
 979         struct hlist_head *bucket;
 980         unsigned index;
 981         u64 *spte;
 982         u64 pte;
 983         unsigned offset = offset_in_page(gpa);
 984         unsigned pte_size;
 985         unsigned page_offset;
 986         unsigned misaligned;
 987         int level;
 988         int flooded = 0;
 989
 990         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
 991         if (gfn == vcpu->last_pt_write_gfn) {
 992                 ++vcpu->last_pt_write_count;
 993                 if (vcpu->last_pt_write_count >= 3)
 994                         flooded = 1;
 995         } else {
 996                 vcpu->last_pt_write_gfn = gfn;
 997                 vcpu->last_pt_write_count = 1;
 998         }
 999         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1000         bucket = &vcpu->kvm->mmu_page_hash[index];
1001         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1002                 if (page->gfn != gfn || page->role.metaphysical)
1003                         continue;
1004                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1005                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1006                 if (misaligned || flooded) {
1007                         /*
1008                          * Misaligned accesses are too much trouble to fix
1009                          * up; also, they usually indicate a page is not used
1010                          * as a page table.
1011                          *
1012                          * If we're seeing too many writes to a page,
1013                          * it may no longer be a page table, or we may be
1014                          * forking, in which case it is better to unmap the
1015                          * page.
1016                          */
1017                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1018                                  gpa, bytes, page->role.word);
1019                         kvm_mmu_zap_page(vcpu, page);
1020                         continue;
1021                 }
1022                 page_offset = offset;
1023                 level = page->role.level;
1024                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1025                         page_offset <<= 1;          /* 32->64 */
1026                         page_offset &= ~PAGE_MASK;
1027                 }
1028                 spte = __va(page->page_hpa);
1029                 spte += page_offset / sizeof(*spte);
1030                 pte = *spte;
1031                 if (is_present_pte(pte)) {
1032                         if (level == PT_PAGE_TABLE_LEVEL)
1033                                 rmap_remove(vcpu->kvm, spte);
1034                         else {
1035                                 child = page_header(pte & PT64_BASE_ADDR_MASK);
1036                                 mmu_page_remove_parent_pte(child, spte);
1037                         }
1038                 }
1039                 *spte = 0;
1040         }
1041 }
1042
1043 void kvm_mmu_post_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1044 {
1045 }
1046
1047 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1048 {
1049         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1050
1051         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1052 }
1053
1054 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1055 {
1056         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1057                 struct kvm_mmu_page *page;
1058
1059                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1060                                     struct kvm_mmu_page, link);
1061                 kvm_mmu_zap_page(vcpu, page);
1062         }
1063 }
1064 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1065
1066 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1067 {
1068         struct kvm_mmu_page *page;
1069
1070         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1071                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1072                                     struct kvm_mmu_page, link);
1073                 kvm_mmu_zap_page(vcpu, page);
1074         }
1075         while (!list_empty(&vcpu->free_pages)) {
1076                 page = list_entry(vcpu->free_pages.next,
1077                                   struct kvm_mmu_page, link);
1078                 list_del(&page->link);
1079                 __free_page(pfn_to_page(page->page_hpa >> PAGE_SHIFT));
1080                 page->page_hpa = INVALID_PAGE;
1081         }
1082         free_page((unsigned long)vcpu->mmu.pae_root);
1083 }
1084
1085 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1086 {
1087         struct page *page;
1088         int i;
1089
1090         ASSERT(vcpu);
1091
1092         for (i = 0; i < KVM_NUM_MMU_PAGES; i++) {
1093                 struct kvm_mmu_page *page_header = &vcpu->page_header_buf[i];
1094
1095                 INIT_LIST_HEAD(&page_header->link);
1096                 if ((page = alloc_page(GFP_KERNEL)) == NULL)
1097                         goto error_1;
1098                 page->private = (unsigned long)page_header;
1099                 page_header->page_hpa = (hpa_t)page_to_pfn(page) << PAGE_SHIFT;
1100                 memset(__va(page_header->page_hpa), 0, PAGE_SIZE);
1101                 list_add(&page_header->link, &vcpu->free_pages);
1102                 ++vcpu->kvm->n_free_mmu_pages;
1103         }
1104
1105         /*
1106          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1107          * Therefore we need to allocate shadow page tables in the first
1108          * 4GB of memory, which happens to fit the DMA32 zone.
1109          */
1110         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1111         if (!page)
1112                 goto error_1;
1113         vcpu->mmu.pae_root = page_address(page);
1114         for (i = 0; i < 4; ++i)
1115                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1116
1117         return 0;
1118
1119 error_1:
1120         free_mmu_pages(vcpu);
1121         return -ENOMEM;
1122 }
1123
1124 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1125 {
1126         ASSERT(vcpu);
1127         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1128         ASSERT(list_empty(&vcpu->free_pages));
1129
1130         return alloc_mmu_pages(vcpu);
1131 }
1132
1133 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1134 {
1135         ASSERT(vcpu);
1136         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1137         ASSERT(!list_empty(&vcpu->free_pages));
1138
1139         return init_kvm_mmu(vcpu);
1140 }
1141
1142 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1143 {
1144         ASSERT(vcpu);
1145
1146         destroy_kvm_mmu(vcpu);
1147         free_mmu_pages(vcpu);
1148 }
1149
1150 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1151 {
1152         struct kvm_mmu_page *page;
1153
1154         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1155                 int i;
1156                 u64 *pt;
1157
1158                 if (!test_bit(slot, &page->slot_bitmap))
1159                         continue;
1160
1161                 pt = __va(page->page_hpa);
1162                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1163                         /* avoid RMW */
1164                         if (pt[i] & PT_WRITABLE_MASK) {
1165                                 rmap_remove(kvm, &pt[i]);
1166                                 pt[i] &= ~PT_WRITABLE_MASK;
1167                         }
1168         }
1169 }