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
  20 #include "vmx.h"
  21 #include "kvm.h"
  22 #include "x86.h"
  23
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29
  30 #include <asm/page.h>
  31 #include <asm/cmpxchg.h>
  32 #include <asm/io.h>
  33
  34 #undef MMU_DEBUG
  35
  36 #undef AUDIT
  37
  38 #ifdef AUDIT
  39 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  40 #else
  41 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  42 #endif
  43
  44 #ifdef MMU_DEBUG
  45
  46 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  47 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  48
  49 #else
  50
  51 #define pgprintk(x...) do { } while (0)
  52 #define rmap_printk(x...) do { } while (0)
  53
  54 #endif
  55
  56 #if defined(MMU_DEBUG) || defined(AUDIT)
  57 static int dbg = 1;
  58 #endif
  59
  60 #ifndef MMU_DEBUG
  61 #define ASSERT(x) do { } while (0)
  62 #else
  63 #define ASSERT(x)                                                       \
  64         if (!(x)) {                                                     \
  65                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  66                        __FILE__, __LINE__, #x);                         \
  67         }
  68 #endif
  69
  70 #define PT64_PT_BITS 9
  71 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  72 #define PT32_PT_BITS 10
  73 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  74
  75 #define PT_WRITABLE_SHIFT 1
  76
  77 #define PT_PRESENT_MASK (1ULL << 0)
  78 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  79 #define PT_USER_MASK (1ULL << 2)
  80 #define PT_PWT_MASK (1ULL << 3)
  81 #define PT_PCD_MASK (1ULL << 4)
  82 #define PT_ACCESSED_MASK (1ULL << 5)
  83 #define PT_DIRTY_MASK (1ULL << 6)
  84 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  85 #define PT_PAT_MASK (1ULL << 7)
  86 #define PT_GLOBAL_MASK (1ULL << 8)
  87 #define PT64_NX_MASK (1ULL << 63)
  88
  89 #define PT_PAT_SHIFT 7
  90 #define PT_DIR_PAT_SHIFT 12
  91 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  92
  93 #define PT32_DIR_PSE36_SIZE 4
  94 #define PT32_DIR_PSE36_SHIFT 13
  95 #define PT32_DIR_PSE36_MASK \
  96         (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  97
  98
  99 #define PT_FIRST_AVAIL_BITS_SHIFT 9
 100 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
 101
 102 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 103
 104 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 105
 106 #define PT64_LEVEL_BITS 9
 107
 108 #define PT64_LEVEL_SHIFT(level) \
 109                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
 110
 111 #define PT64_LEVEL_MASK(level) \
 112                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 113
 114 #define PT64_INDEX(address, level)\
 115         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 116
 117
 118 #define PT32_LEVEL_BITS 10
 119
 120 #define PT32_LEVEL_SHIFT(level) \
 121                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 122
 123 #define PT32_LEVEL_MASK(level) \
 124                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 125
 126 #define PT32_INDEX(address, level)\
 127         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 128
 129
 130 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 131 #define PT64_DIR_BASE_ADDR_MASK \
 132         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 133
 134 #define PT32_BASE_ADDR_MASK PAGE_MASK
 135 #define PT32_DIR_BASE_ADDR_MASK \
 136         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 137
 138 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 139                         | PT64_NX_MASK)
 140
 141 #define PFERR_PRESENT_MASK (1U << 0)
 142 #define PFERR_WRITE_MASK (1U << 1)
 143 #define PFERR_USER_MASK (1U << 2)
 144 #define PFERR_FETCH_MASK (1U << 4)
 145
 146 #define PT64_ROOT_LEVEL 4
 147 #define PT32_ROOT_LEVEL 2
 148 #define PT32E_ROOT_LEVEL 3
 149
 150 #define PT_DIRECTORY_LEVEL 2
 151 #define PT_PAGE_TABLE_LEVEL 1
 152
 153 #define RMAP_EXT 4
 154
 155 struct kvm_rmap_desc {
 156         u64 *shadow_ptes[RMAP_EXT];
 157         struct kvm_rmap_desc *more;
 158 };
 159
 160 static struct kmem_cache *pte_chain_cache;
 161 static struct kmem_cache *rmap_desc_cache;
 162 static struct kmem_cache *mmu_page_header_cache;
 163
 164 static u64 __read_mostly shadow_trap_nonpresent_pte;
 165 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 166
 167 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 168 {
 169         shadow_trap_nonpresent_pte = trap_pte;
 170         shadow_notrap_nonpresent_pte = notrap_pte;
 171 }
 172 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 173
 174 static int is_write_protection(struct kvm_vcpu *vcpu)
 175 {
 176         return vcpu->cr0 & X86_CR0_WP;
 177 }
 178
 179 static int is_cpuid_PSE36(void)
 180 {
 181         return 1;
 182 }
 183
 184 static int is_nx(struct kvm_vcpu *vcpu)
 185 {
 186         return vcpu->shadow_efer & EFER_NX;
 187 }
 188
 189 static int is_present_pte(unsigned long pte)
 190 {
 191         return pte & PT_PRESENT_MASK;
 192 }
 193
 194 static int is_shadow_present_pte(u64 pte)
 195 {
 196         pte &= ~PT_SHADOW_IO_MARK;
 197         return pte != shadow_trap_nonpresent_pte
 198                 && pte != shadow_notrap_nonpresent_pte;
 199 }
 200
 201 static int is_writeble_pte(unsigned long pte)
 202 {
 203         return pte & PT_WRITABLE_MASK;
 204 }
 205
 206 static int is_dirty_pte(unsigned long pte)
 207 {
 208         return pte & PT_DIRTY_MASK;
 209 }
 210
 211 static int is_io_pte(unsigned long pte)
 212 {
 213         return pte & PT_SHADOW_IO_MARK;
 214 }
 215
 216 static int is_rmap_pte(u64 pte)
 217 {
 218         return pte != shadow_trap_nonpresent_pte
 219                 && pte != shadow_notrap_nonpresent_pte;
 220 }
 221
 222 static gfn_t pse36_gfn_delta(u32 gpte)
 223 {
 224         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 225
 226         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 227 }
 228
 229 static void set_shadow_pte(u64 *sptep, u64 spte)
 230 {
 231 #ifdef CONFIG_X86_64
 232         set_64bit((unsigned long *)sptep, spte);
 233 #else
 234         set_64bit((unsigned long long *)sptep, spte);
 235 #endif
 236 }
 237
 238 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 239                                   struct kmem_cache *base_cache, int min)
 240 {
 241         void *obj;
 242
 243         if (cache->nobjs >= min)
 244                 return 0;
 245         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 246                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 247                 if (!obj)
 248                         return -ENOMEM;
 249                 cache->objects[cache->nobjs++] = obj;
 250         }
 251         return 0;
 252 }
 253
 254 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 255 {
 256         while (mc->nobjs)
 257                 kfree(mc->objects[--mc->nobjs]);
 258 }
 259
 260 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 261                                        int min)
 262 {
 263         struct page *page;
 264
 265         if (cache->nobjs >= min)
 266                 return 0;
 267         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 268                 page = alloc_page(GFP_KERNEL);
 269                 if (!page)
 270                         return -ENOMEM;
 271                 set_page_private(page, 0);
 272                 cache->objects[cache->nobjs++] = page_address(page);
 273         }
 274         return 0;
 275 }
 276
 277 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 278 {
 279         while (mc->nobjs)
 280                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 281 }
 282
 283 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 284 {
 285         int r;
 286
 287         kvm_mmu_free_some_pages(vcpu);
 288         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 289                                    pte_chain_cache, 4);
 290         if (r)
 291                 goto out;
 292         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 293                                    rmap_desc_cache, 1);
 294         if (r)
 295                 goto out;
 296         r = mmu_topup_memory_cache_page(&vcpu->mmu_page_cache, 8);
 297         if (r)
 298                 goto out;
 299         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 300                                    mmu_page_header_cache, 4);
 301 out:
 302         return r;
 303 }
 304
 305 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 306 {
 307         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 308         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 309         mmu_free_memory_cache_page(&vcpu->mmu_page_cache);
 310         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 311 }
 312
 313 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 314                                     size_t size)
 315 {
 316         void *p;
 317
 318         BUG_ON(!mc->nobjs);
 319         p = mc->objects[--mc->nobjs];
 320         memset(p, 0, size);
 321         return p;
 322 }
 323
 324 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 325 {
 326         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 327                                       sizeof(struct kvm_pte_chain));
 328 }
 329
 330 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 331 {
 332         kfree(pc);
 333 }
 334
 335 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 336 {
 337         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 338                                       sizeof(struct kvm_rmap_desc));
 339 }
 340
 341 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 342 {
 343         kfree(rd);
 344 }
 345
 346 /*
 347  * Take gfn and return the reverse mapping to it.
 348  * Note: gfn must be unaliased before this function get called
 349  */
 350
 351 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
 352 {
 353         struct kvm_memory_slot *slot;
 354
 355         slot = gfn_to_memslot(kvm, gfn);
 356         return &slot->rmap[gfn - slot->base_gfn];
 357 }
 358
 359 /*
 360  * Reverse mapping data structures:
 361  *
 362  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 363  * that points to page_address(page).
 364  *
 365  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 366  * containing more mappings.
 367  */
 368 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 369 {
 370         struct kvm_mmu_page *page;
 371         struct kvm_rmap_desc *desc;
 372         unsigned long *rmapp;
 373         int i;
 374
 375         if (!is_rmap_pte(*spte))
 376                 return;
 377         gfn = unalias_gfn(vcpu->kvm, gfn);
 378         page = page_header(__pa(spte));
 379         page->gfns[spte - page->spt] = gfn;
 380         rmapp = gfn_to_rmap(vcpu->kvm, gfn);
 381         if (!*rmapp) {
 382                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 383                 *rmapp = (unsigned long)spte;
 384         } else if (!(*rmapp & 1)) {
 385                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 386                 desc = mmu_alloc_rmap_desc(vcpu);
 387                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 388                 desc->shadow_ptes[1] = spte;
 389                 *rmapp = (unsigned long)desc | 1;
 390         } else {
 391                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 392                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 393                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 394                         desc = desc->more;
 395                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 396                         desc->more = mmu_alloc_rmap_desc(vcpu);
 397                         desc = desc->more;
 398                 }
 399                 for (i = 0; desc->shadow_ptes[i]; ++i)
 400                         ;
 401                 desc->shadow_ptes[i] = spte;
 402         }
 403 }
 404
 405 static void rmap_desc_remove_entry(unsigned long *rmapp,
 406                                    struct kvm_rmap_desc *desc,
 407                                    int i,
 408                                    struct kvm_rmap_desc *prev_desc)
 409 {
 410         int j;
 411
 412         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 413                 ;
 414         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 415         desc->shadow_ptes[j] = NULL;
 416         if (j != 0)
 417                 return;
 418         if (!prev_desc && !desc->more)
 419                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 420         else
 421                 if (prev_desc)
 422                         prev_desc->more = desc->more;
 423                 else
 424                         *rmapp = (unsigned long)desc->more | 1;
 425         mmu_free_rmap_desc(desc);
 426 }
 427
 428 static void rmap_remove(struct kvm *kvm, u64 *spte)
 429 {
 430         struct kvm_rmap_desc *desc;
 431         struct kvm_rmap_desc *prev_desc;
 432         struct kvm_mmu_page *page;
 433         struct page *release_page;
 434         unsigned long *rmapp;
 435         int i;
 436
 437         if (!is_rmap_pte(*spte))
 438                 return;
 439         page = page_header(__pa(spte));
 440         release_page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 441         if (is_writeble_pte(*spte))
 442                 kvm_release_page_dirty(release_page);
 443         else
 444                 kvm_release_page_clean(release_page);
 445         rmapp = gfn_to_rmap(kvm, page->gfns[spte - page->spt]);
 446         if (!*rmapp) {
 447                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 448                 BUG();
 449         } else if (!(*rmapp & 1)) {
 450                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 451                 if ((u64 *)*rmapp != spte) {
 452                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 453                                spte, *spte);
 454                         BUG();
 455                 }
 456                 *rmapp = 0;
 457         } else {
 458                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 459                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 460                 prev_desc = NULL;
 461                 while (desc) {
 462                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 463                                 if (desc->shadow_ptes[i] == spte) {
 464                                         rmap_desc_remove_entry(rmapp,
 465                                                                desc, i,
 466                                                                prev_desc);
 467                                         return;
 468                                 }
 469                         prev_desc = desc;
 470                         desc = desc->more;
 471                 }
 472                 BUG();
 473         }
 474 }
 475
 476 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 477 {
 478         struct kvm_rmap_desc *desc;
 479         struct kvm_rmap_desc *prev_desc;
 480         u64 *prev_spte;
 481         int i;
 482
 483         if (!*rmapp)
 484                 return NULL;
 485         else if (!(*rmapp & 1)) {
 486                 if (!spte)
 487                         return (u64 *)*rmapp;
 488                 return NULL;
 489         }
 490         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 491         prev_desc = NULL;
 492         prev_spte = NULL;
 493         while (desc) {
 494                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 495                         if (prev_spte == spte)
 496                                 return desc->shadow_ptes[i];
 497                         prev_spte = desc->shadow_ptes[i];
 498                 }
 499                 desc = desc->more;
 500         }
 501         return NULL;
 502 }
 503
 504 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 505 {
 506         unsigned long *rmapp;
 507         u64 *spte;
 508
 509         gfn = unalias_gfn(kvm, gfn);
 510         rmapp = gfn_to_rmap(kvm, gfn);
 511
 512         spte = rmap_next(kvm, rmapp, NULL);
 513         while (spte) {
 514                 BUG_ON(!spte);
 515                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 516                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 517                 if (is_writeble_pte(*spte))
 518                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 519                 kvm_flush_remote_tlbs(kvm);
 520                 spte = rmap_next(kvm, rmapp, spte);
 521         }
 522 }
 523
 524 #ifdef MMU_DEBUG
 525 static int is_empty_shadow_page(u64 *spt)
 526 {
 527         u64 *pos;
 528         u64 *end;
 529
 530         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 531                 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
 532                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 533                                pos, *pos);
 534                         return 0;
 535                 }
 536         return 1;
 537 }
 538 #endif
 539
 540 static void kvm_mmu_free_page(struct kvm *kvm,
 541                               struct kvm_mmu_page *page_head)
 542 {
 543         ASSERT(is_empty_shadow_page(page_head->spt));
 544         list_del(&page_head->link);
 545         __free_page(virt_to_page(page_head->spt));
 546         __free_page(virt_to_page(page_head->gfns));
 547         kfree(page_head);
 548         ++kvm->n_free_mmu_pages;
 549 }
 550
 551 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 552 {
 553         return gfn;
 554 }
 555
 556 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 557                                                u64 *parent_pte)
 558 {
 559         struct kvm_mmu_page *page;
 560
 561         if (!vcpu->kvm->n_free_mmu_pages)
 562                 return NULL;
 563
 564         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 565                                       sizeof *page);
 566         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 567         page->gfns = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 568         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 569         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 570         ASSERT(is_empty_shadow_page(page->spt));
 571         page->slot_bitmap = 0;
 572         page->multimapped = 0;
 573         page->parent_pte = parent_pte;
 574         --vcpu->kvm->n_free_mmu_pages;
 575         return page;
 576 }
 577
 578 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 579                                     struct kvm_mmu_page *page, u64 *parent_pte)
 580 {
 581         struct kvm_pte_chain *pte_chain;
 582         struct hlist_node *node;
 583         int i;
 584
 585         if (!parent_pte)
 586                 return;
 587         if (!page->multimapped) {
 588                 u64 *old = page->parent_pte;
 589
 590                 if (!old) {
 591                         page->parent_pte = parent_pte;
 592                         return;
 593                 }
 594                 page->multimapped = 1;
 595                 pte_chain = mmu_alloc_pte_chain(vcpu);
 596                 INIT_HLIST_HEAD(&page->parent_ptes);
 597                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 598                 pte_chain->parent_ptes[0] = old;
 599         }
 600         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 601                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 602                         continue;
 603                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 604                         if (!pte_chain->parent_ptes[i]) {
 605                                 pte_chain->parent_ptes[i] = parent_pte;
 606                                 return;
 607                         }
 608         }
 609         pte_chain = mmu_alloc_pte_chain(vcpu);
 610         BUG_ON(!pte_chain);
 611         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 612         pte_chain->parent_ptes[0] = parent_pte;
 613 }
 614
 615 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 616                                        u64 *parent_pte)
 617 {
 618         struct kvm_pte_chain *pte_chain;
 619         struct hlist_node *node;
 620         int i;
 621
 622         if (!page->multimapped) {
 623                 BUG_ON(page->parent_pte != parent_pte);
 624                 page->parent_pte = NULL;
 625                 return;
 626         }
 627         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 628                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 629                         if (!pte_chain->parent_ptes[i])
 630                                 break;
 631                         if (pte_chain->parent_ptes[i] != parent_pte)
 632                                 continue;
 633                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 634                                 && pte_chain->parent_ptes[i + 1]) {
 635                                 pte_chain->parent_ptes[i]
 636                                         = pte_chain->parent_ptes[i + 1];
 637                                 ++i;
 638                         }
 639                         pte_chain->parent_ptes[i] = NULL;
 640                         if (i == 0) {
 641                                 hlist_del(&pte_chain->link);
 642                                 mmu_free_pte_chain(pte_chain);
 643                                 if (hlist_empty(&page->parent_ptes)) {
 644                                         page->multimapped = 0;
 645                                         page->parent_pte = NULL;
 646                                 }
 647                         }
 648                         return;
 649                 }
 650         BUG();
 651 }
 652
 653 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm,
 654                                                 gfn_t gfn)
 655 {
 656         unsigned index;
 657         struct hlist_head *bucket;
 658         struct kvm_mmu_page *page;
 659         struct hlist_node *node;
 660
 661         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 662         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 663         bucket = &kvm->mmu_page_hash[index];
 664         hlist_for_each_entry(page, node, bucket, hash_link)
 665                 if (page->gfn == gfn && !page->role.metaphysical) {
 666                         pgprintk("%s: found role %x\n",
 667                                  __FUNCTION__, page->role.word);
 668                         return page;
 669                 }
 670         return NULL;
 671 }
 672
 673 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 674                                              gfn_t gfn,
 675                                              gva_t gaddr,
 676                                              unsigned level,
 677                                              int metaphysical,
 678                                              unsigned hugepage_access,
 679                                              u64 *parent_pte)
 680 {
 681         union kvm_mmu_page_role role;
 682         unsigned index;
 683         unsigned quadrant;
 684         struct hlist_head *bucket;
 685         struct kvm_mmu_page *page;
 686         struct hlist_node *node;
 687
 688         role.word = 0;
 689         role.glevels = vcpu->mmu.root_level;
 690         role.level = level;
 691         role.metaphysical = metaphysical;
 692         role.hugepage_access = hugepage_access;
 693         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 694                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 695                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 696                 role.quadrant = quadrant;
 697         }
 698         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 699                  gfn, role.word);
 700         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 701         bucket = &vcpu->kvm->mmu_page_hash[index];
 702         hlist_for_each_entry(page, node, bucket, hash_link)
 703                 if (page->gfn == gfn && page->role.word == role.word) {
 704                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 705                         pgprintk("%s: found\n", __FUNCTION__);
 706                         return page;
 707                 }
 708         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 709         if (!page)
 710                 return page;
 711         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 712         page->gfn = gfn;
 713         page->role = role;
 714         hlist_add_head(&page->hash_link, bucket);
 715         vcpu->mmu.prefetch_page(vcpu, page);
 716         if (!metaphysical)
 717                 rmap_write_protect(vcpu->kvm, gfn);
 718         return page;
 719 }
 720
 721 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 722                                          struct kvm_mmu_page *page)
 723 {
 724         unsigned i;
 725         u64 *pt;
 726         u64 ent;
 727
 728         pt = page->spt;
 729
 730         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 731                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 732                         if (is_shadow_present_pte(pt[i]))
 733                                 rmap_remove(kvm, &pt[i]);
 734                         pt[i] = shadow_trap_nonpresent_pte;
 735                 }
 736                 kvm_flush_remote_tlbs(kvm);
 737                 return;
 738         }
 739
 740         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 741                 ent = pt[i];
 742
 743                 pt[i] = shadow_trap_nonpresent_pte;
 744                 if (!is_shadow_present_pte(ent))
 745                         continue;
 746                 ent &= PT64_BASE_ADDR_MASK;
 747                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 748         }
 749         kvm_flush_remote_tlbs(kvm);
 750 }
 751
 752 static void kvm_mmu_put_page(struct kvm_mmu_page *page,
 753                              u64 *parent_pte)
 754 {
 755         mmu_page_remove_parent_pte(page, parent_pte);
 756 }
 757
 758 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 759 {
 760         int i;
 761
 762         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 763                 if (kvm->vcpus[i])
 764                         kvm->vcpus[i]->last_pte_updated = NULL;
 765 }
 766
 767 static void kvm_mmu_zap_page(struct kvm *kvm,
 768                              struct kvm_mmu_page *page)
 769 {
 770         u64 *parent_pte;
 771
 772         ++kvm->stat.mmu_shadow_zapped;
 773         while (page->multimapped || page->parent_pte) {
 774                 if (!page->multimapped)
 775                         parent_pte = page->parent_pte;
 776                 else {
 777                         struct kvm_pte_chain *chain;
 778
 779                         chain = container_of(page->parent_ptes.first,
 780                                              struct kvm_pte_chain, link);
 781                         parent_pte = chain->parent_ptes[0];
 782                 }
 783                 BUG_ON(!parent_pte);
 784                 kvm_mmu_put_page(page, parent_pte);
 785                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 786         }
 787         kvm_mmu_page_unlink_children(kvm, page);
 788         if (!page->root_count) {
 789                 hlist_del(&page->hash_link);
 790                 kvm_mmu_free_page(kvm, page);
 791         } else
 792                 list_move(&page->link, &kvm->active_mmu_pages);
 793         kvm_mmu_reset_last_pte_updated(kvm);
 794 }
 795
 796 /*
 797  * Changing the number of mmu pages allocated to the vm
 798  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 799  */
 800 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 801 {
 802         /*
 803          * If we set the number of mmu pages to be smaller be than the
 804          * number of actived pages , we must to free some mmu pages before we
 805          * change the value
 806          */
 807
 808         if ((kvm->n_alloc_mmu_pages - kvm->n_free_mmu_pages) >
 809             kvm_nr_mmu_pages) {
 810                 int n_used_mmu_pages = kvm->n_alloc_mmu_pages
 811                                        - kvm->n_free_mmu_pages;
 812
 813                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 814                         struct kvm_mmu_page *page;
 815
 816                         page = container_of(kvm->active_mmu_pages.prev,
 817                                             struct kvm_mmu_page, link);
 818                         kvm_mmu_zap_page(kvm, page);
 819                         n_used_mmu_pages--;
 820                 }
 821                 kvm->n_free_mmu_pages = 0;
 822         }
 823         else
 824                 kvm->n_free_mmu_pages += kvm_nr_mmu_pages
 825                                          - kvm->n_alloc_mmu_pages;
 826
 827         kvm->n_alloc_mmu_pages = kvm_nr_mmu_pages;
 828 }
 829
 830 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 831 {
 832         unsigned index;
 833         struct hlist_head *bucket;
 834         struct kvm_mmu_page *page;
 835         struct hlist_node *node, *n;
 836         int r;
 837
 838         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 839         r = 0;
 840         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 841         bucket = &kvm->mmu_page_hash[index];
 842         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 843                 if (page->gfn == gfn && !page->role.metaphysical) {
 844                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 845                                  page->role.word);
 846                         kvm_mmu_zap_page(kvm, page);
 847                         r = 1;
 848                 }
 849         return r;
 850 }
 851
 852 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
 853 {
 854         struct kvm_mmu_page *page;
 855
 856         while ((page = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
 857                 pgprintk("%s: zap %lx %x\n",
 858                          __FUNCTION__, gfn, page->role.word);
 859                 kvm_mmu_zap_page(kvm, page);
 860         }
 861 }
 862
 863 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
 864 {
 865         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
 866         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 867
 868         __set_bit(slot, &page_head->slot_bitmap);
 869 }
 870
 871 hpa_t gpa_to_hpa(struct kvm *kvm, gpa_t gpa)
 872 {
 873         struct page *page;
 874         hpa_t hpa;
 875
 876         ASSERT((gpa & HPA_ERR_MASK) == 0);
 877         page = gfn_to_page(kvm, gpa >> PAGE_SHIFT);
 878         hpa = ((hpa_t)page_to_pfn(page) << PAGE_SHIFT) | (gpa & (PAGE_SIZE-1));
 879         if (is_error_page(page))
 880                 return hpa | HPA_ERR_MASK;
 881         return hpa;
 882 }
 883
 884 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 885 {
 886         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 887
 888         if (gpa == UNMAPPED_GVA)
 889                 return NULL;
 890         return pfn_to_page(gpa_to_hpa(vcpu->kvm, gpa) >> PAGE_SHIFT);
 891 }
 892
 893 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 894 {
 895 }
 896
 897 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, struct page *page)
 898 {
 899         int level = PT32E_ROOT_LEVEL;
 900         hpa_t table_addr = vcpu->mmu.root_hpa;
 901
 902         for (; ; level--) {
 903                 u32 index = PT64_INDEX(v, level);
 904                 u64 *table;
 905                 u64 pte;
 906
 907                 ASSERT(VALID_PAGE(table_addr));
 908                 table = __va(table_addr);
 909
 910                 if (level == 1) {
 911                         int was_rmapped;
 912
 913                         pte = table[index];
 914                         was_rmapped = is_rmap_pte(pte);
 915                         if (is_shadow_present_pte(pte) && is_writeble_pte(pte)) {
 916                                 kvm_release_page_clean(page);
 917                                 return 0;
 918                         }
 919                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 920                         page_header_update_slot(vcpu->kvm, table,
 921                                                 v >> PAGE_SHIFT);
 922                         table[index] = page_to_phys(page)
 923                                 | PT_PRESENT_MASK | PT_WRITABLE_MASK
 924                                 | PT_USER_MASK;
 925                         if (!was_rmapped)
 926                                 rmap_add(vcpu, &table[index], v >> PAGE_SHIFT);
 927                         else
 928                                 kvm_release_page_clean(page);
 929
 930                         return 0;
 931                 }
 932
 933                 if (table[index] == shadow_trap_nonpresent_pte) {
 934                         struct kvm_mmu_page *new_table;
 935                         gfn_t pseudo_gfn;
 936
 937                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 938                                 >> PAGE_SHIFT;
 939                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 940                                                      v, level - 1,
 941                                                      1, 3, &table[index]);
 942                         if (!new_table) {
 943                                 pgprintk("nonpaging_map: ENOMEM\n");
 944                                 kvm_release_page_clean(page);
 945                                 return -ENOMEM;
 946                         }
 947
 948                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 949                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 950                 }
 951                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 952         }
 953 }
 954
 955 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 956                                     struct kvm_mmu_page *sp)
 957 {
 958         int i;
 959
 960         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 961                 sp->spt[i] = shadow_trap_nonpresent_pte;
 962 }
 963
 964 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 965 {
 966         int i;
 967         struct kvm_mmu_page *page;
 968
 969         if (!VALID_PAGE(vcpu->mmu.root_hpa))
 970                 return;
 971 #ifdef CONFIG_X86_64
 972         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 973                 hpa_t root = vcpu->mmu.root_hpa;
 974
 975                 page = page_header(root);
 976                 --page->root_count;
 977                 vcpu->mmu.root_hpa = INVALID_PAGE;
 978                 return;
 979         }
 980 #endif
 981         for (i = 0; i < 4; ++i) {
 982                 hpa_t root = vcpu->mmu.pae_root[i];
 983
 984                 if (root) {
 985                         root &= PT64_BASE_ADDR_MASK;
 986                         page = page_header(root);
 987                         --page->root_count;
 988                 }
 989                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 990         }
 991         vcpu->mmu.root_hpa = INVALID_PAGE;
 992 }
 993
 994 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 995 {
 996         int i;
 997         gfn_t root_gfn;
 998         struct kvm_mmu_page *page;
 999
1000         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
1001
1002 #ifdef CONFIG_X86_64
1003         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1004                 hpa_t root = vcpu->mmu.root_hpa;
1005
1006                 ASSERT(!VALID_PAGE(root));
1007                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
1008                                         PT64_ROOT_LEVEL, 0, 0, NULL);
1009                 root = __pa(page->spt);
1010                 ++page->root_count;
1011                 vcpu->mmu.root_hpa = root;
1012                 return;
1013         }
1014 #endif
1015         for (i = 0; i < 4; ++i) {
1016                 hpa_t root = vcpu->mmu.pae_root[i];
1017
1018                 ASSERT(!VALID_PAGE(root));
1019                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
1020                         if (!is_present_pte(vcpu->pdptrs[i])) {
1021                                 vcpu->mmu.pae_root[i] = 0;
1022                                 continue;
1023                         }
1024                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
1025                 } else if (vcpu->mmu.root_level == 0)
1026                         root_gfn = 0;
1027                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1028                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
1029                                         0, NULL);
1030                 root = __pa(page->spt);
1031                 ++page->root_count;
1032                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
1033         }
1034         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
1035 }
1036
1037 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1038 {
1039         return vaddr;
1040 }
1041
1042 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1043                                 u32 error_code)
1044 {
1045         struct page *page;
1046         int r;
1047
1048         r = mmu_topup_memory_caches(vcpu);
1049         if (r)
1050                 return r;
1051
1052         ASSERT(vcpu);
1053         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
1054
1055         page = gfn_to_page(vcpu->kvm, gva >> PAGE_SHIFT);
1056
1057         if (is_error_page(page)) {
1058                 kvm_release_page_clean(page);
1059                 return 1;
1060         }
1061
1062         return nonpaging_map(vcpu, gva & PAGE_MASK, page);
1063 }
1064
1065 static void nonpaging_free(struct kvm_vcpu *vcpu)
1066 {
1067         mmu_free_roots(vcpu);
1068 }
1069
1070 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1071 {
1072         struct kvm_mmu *context = &vcpu->mmu;
1073
1074         context->new_cr3 = nonpaging_new_cr3;
1075         context->page_fault = nonpaging_page_fault;
1076         context->gva_to_gpa = nonpaging_gva_to_gpa;
1077         context->free = nonpaging_free;
1078         context->prefetch_page = nonpaging_prefetch_page;
1079         context->root_level = 0;
1080         context->shadow_root_level = PT32E_ROOT_LEVEL;
1081         context->root_hpa = INVALID_PAGE;
1082         return 0;
1083 }
1084
1085 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1086 {
1087         ++vcpu->stat.tlb_flush;
1088         kvm_x86_ops->tlb_flush(vcpu);
1089 }
1090
1091 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1092 {
1093         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
1094         mmu_free_roots(vcpu);
1095 }
1096
1097 static void inject_page_fault(struct kvm_vcpu *vcpu,
1098                               u64 addr,
1099                               u32 err_code)
1100 {
1101         kvm_x86_ops->inject_page_fault(vcpu, addr, err_code);
1102 }
1103
1104 static void paging_free(struct kvm_vcpu *vcpu)
1105 {
1106         nonpaging_free(vcpu);
1107 }
1108
1109 #define PTTYPE 64
1110 #include "paging_tmpl.h"
1111 #undef PTTYPE
1112
1113 #define PTTYPE 32
1114 #include "paging_tmpl.h"
1115 #undef PTTYPE
1116
1117 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1118 {
1119         struct kvm_mmu *context = &vcpu->mmu;
1120
1121         ASSERT(is_pae(vcpu));
1122         context->new_cr3 = paging_new_cr3;
1123         context->page_fault = paging64_page_fault;
1124         context->gva_to_gpa = paging64_gva_to_gpa;
1125         context->prefetch_page = paging64_prefetch_page;
1126         context->free = paging_free;
1127         context->root_level = level;
1128         context->shadow_root_level = level;
1129         context->root_hpa = INVALID_PAGE;
1130         return 0;
1131 }
1132
1133 static int paging64_init_context(struct kvm_vcpu *vcpu)
1134 {
1135         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1136 }
1137
1138 static int paging32_init_context(struct kvm_vcpu *vcpu)
1139 {
1140         struct kvm_mmu *context = &vcpu->mmu;
1141
1142         context->new_cr3 = paging_new_cr3;
1143         context->page_fault = paging32_page_fault;
1144         context->gva_to_gpa = paging32_gva_to_gpa;
1145         context->free = paging_free;
1146         context->prefetch_page = paging32_prefetch_page;
1147         context->root_level = PT32_ROOT_LEVEL;
1148         context->shadow_root_level = PT32E_ROOT_LEVEL;
1149         context->root_hpa = INVALID_PAGE;
1150         return 0;
1151 }
1152
1153 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1154 {
1155         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1156 }
1157
1158 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1159 {
1160         ASSERT(vcpu);
1161         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1162
1163         if (!is_paging(vcpu))
1164                 return nonpaging_init_context(vcpu);
1165         else if (is_long_mode(vcpu))
1166                 return paging64_init_context(vcpu);
1167         else if (is_pae(vcpu))
1168                 return paging32E_init_context(vcpu);
1169         else
1170                 return paging32_init_context(vcpu);
1171 }
1172
1173 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1174 {
1175         ASSERT(vcpu);
1176         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1177                 vcpu->mmu.free(vcpu);
1178                 vcpu->mmu.root_hpa = INVALID_PAGE;
1179         }
1180 }
1181
1182 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1183 {
1184         destroy_kvm_mmu(vcpu);
1185         return init_kvm_mmu(vcpu);
1186 }
1187 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1188
1189 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1190 {
1191         int r;
1192
1193         mutex_lock(&vcpu->kvm->lock);
1194         r = mmu_topup_memory_caches(vcpu);
1195         if (r)
1196                 goto out;
1197         mmu_alloc_roots(vcpu);
1198         kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1199         kvm_mmu_flush_tlb(vcpu);
1200 out:
1201         mutex_unlock(&vcpu->kvm->lock);
1202         return r;
1203 }
1204 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1205
1206 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1207 {
1208         mmu_free_roots(vcpu);
1209 }
1210
1211 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1212                                   struct kvm_mmu_page *page,
1213                                   u64 *spte)
1214 {
1215         u64 pte;
1216         struct kvm_mmu_page *child;
1217
1218         pte = *spte;
1219         if (is_shadow_present_pte(pte)) {
1220                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1221                         rmap_remove(vcpu->kvm, spte);
1222                 else {
1223                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1224                         mmu_page_remove_parent_pte(child, spte);
1225                 }
1226         }
1227         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1228 }
1229
1230 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1231                                   struct kvm_mmu_page *page,
1232                                   u64 *spte,
1233                                   const void *new, int bytes,
1234                                   int offset_in_pte)
1235 {
1236         if (page->role.level != PT_PAGE_TABLE_LEVEL) {
1237                 ++vcpu->kvm->stat.mmu_pde_zapped;
1238                 return;
1239         }
1240
1241         ++vcpu->kvm->stat.mmu_pte_updated;
1242         if (page->role.glevels == PT32_ROOT_LEVEL)
1243                 paging32_update_pte(vcpu, page, spte, new, bytes,
1244                                     offset_in_pte);
1245         else
1246                 paging64_update_pte(vcpu, page, spte, new, bytes,
1247                                     offset_in_pte);
1248 }
1249
1250 static bool need_remote_flush(u64 old, u64 new)
1251 {
1252         if (!is_shadow_present_pte(old))
1253                 return false;
1254         if (!is_shadow_present_pte(new))
1255                 return true;
1256         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1257                 return true;
1258         old ^= PT64_NX_MASK;
1259         new ^= PT64_NX_MASK;
1260         return (old & ~new & PT64_PERM_MASK) != 0;
1261 }
1262
1263 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1264 {
1265         if (need_remote_flush(old, new))
1266                 kvm_flush_remote_tlbs(vcpu->kvm);
1267         else
1268                 kvm_mmu_flush_tlb(vcpu);
1269 }
1270
1271 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1272 {
1273         u64 *spte = vcpu->last_pte_updated;
1274
1275         return !!(spte && (*spte & PT_ACCESSED_MASK));
1276 }
1277
1278 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1279                        const u8 *new, int bytes)
1280 {
1281         gfn_t gfn = gpa >> PAGE_SHIFT;
1282         struct kvm_mmu_page *page;
1283         struct hlist_node *node, *n;
1284         struct hlist_head *bucket;
1285         unsigned index;
1286         u64 entry;
1287         u64 *spte;
1288         unsigned offset = offset_in_page(gpa);
1289         unsigned pte_size;
1290         unsigned page_offset;
1291         unsigned misaligned;
1292         unsigned quadrant;
1293         int level;
1294         int flooded = 0;
1295         int npte;
1296
1297         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1298         ++vcpu->kvm->stat.mmu_pte_write;
1299         kvm_mmu_audit(vcpu, "pre pte write");
1300         if (gfn == vcpu->last_pt_write_gfn
1301             && !last_updated_pte_accessed(vcpu)) {
1302                 ++vcpu->last_pt_write_count;
1303                 if (vcpu->last_pt_write_count >= 3)
1304                         flooded = 1;
1305         } else {
1306                 vcpu->last_pt_write_gfn = gfn;
1307                 vcpu->last_pt_write_count = 1;
1308                 vcpu->last_pte_updated = NULL;
1309         }
1310         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1311         bucket = &vcpu->kvm->mmu_page_hash[index];
1312         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1313                 if (page->gfn != gfn || page->role.metaphysical)
1314                         continue;
1315                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1316                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1317                 misaligned |= bytes < 4;
1318                 if (misaligned || flooded) {
1319                         /*
1320                          * Misaligned accesses are too much trouble to fix
1321                          * up; also, they usually indicate a page is not used
1322                          * as a page table.
1323                          *
1324                          * If we're seeing too many writes to a page,
1325                          * it may no longer be a page table, or we may be
1326                          * forking, in which case it is better to unmap the
1327                          * page.
1328                          */
1329                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1330                                  gpa, bytes, page->role.word);
1331                         kvm_mmu_zap_page(vcpu->kvm, page);
1332                         ++vcpu->kvm->stat.mmu_flooded;
1333                         continue;
1334                 }
1335                 page_offset = offset;
1336                 level = page->role.level;
1337                 npte = 1;
1338                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1339                         page_offset <<= 1;      /* 32->64 */
1340                         /*
1341                          * A 32-bit pde maps 4MB while the shadow pdes map
1342                          * only 2MB.  So we need to double the offset again
1343                          * and zap two pdes instead of one.
1344                          */
1345                         if (level == PT32_ROOT_LEVEL) {
1346                                 page_offset &= ~7; /* kill rounding error */
1347                                 page_offset <<= 1;
1348                                 npte = 2;
1349                         }
1350                         quadrant = page_offset >> PAGE_SHIFT;
1351                         page_offset &= ~PAGE_MASK;
1352                         if (quadrant != page->role.quadrant)
1353                                 continue;
1354                 }
1355                 spte = &page->spt[page_offset / sizeof(*spte)];
1356                 while (npte--) {
1357                         entry = *spte;
1358                         mmu_pte_write_zap_pte(vcpu, page, spte);
1359                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes,
1360                                               page_offset & (pte_size - 1));
1361                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1362                         ++spte;
1363                 }
1364         }
1365         kvm_mmu_audit(vcpu, "post pte write");
1366 }
1367
1368 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1369 {
1370         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1371
1372         return kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1373 }
1374
1375 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1376 {
1377         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1378                 struct kvm_mmu_page *page;
1379
1380                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1381                                     struct kvm_mmu_page, link);
1382                 kvm_mmu_zap_page(vcpu->kvm, page);
1383                 ++vcpu->kvm->stat.mmu_recycled;
1384         }
1385 }
1386
1387 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1388 {
1389         int r;
1390         enum emulation_result er;
1391
1392         mutex_lock(&vcpu->kvm->lock);
1393         r = vcpu->mmu.page_fault(vcpu, cr2, error_code);
1394         if (r < 0)
1395                 goto out;
1396
1397         if (!r) {
1398                 r = 1;
1399                 goto out;
1400         }
1401
1402         r = mmu_topup_memory_caches(vcpu);
1403         if (r)
1404                 goto out;
1405
1406         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1407         mutex_unlock(&vcpu->kvm->lock);
1408
1409         switch (er) {
1410         case EMULATE_DONE:
1411                 return 1;
1412         case EMULATE_DO_MMIO:
1413                 ++vcpu->stat.mmio_exits;
1414                 return 0;
1415         case EMULATE_FAIL:
1416                 kvm_report_emulation_failure(vcpu, "pagetable");
1417                 return 1;
1418         default:
1419                 BUG();
1420         }
1421 out:
1422         mutex_unlock(&vcpu->kvm->lock);
1423         return r;
1424 }
1425 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1426
1427 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1428 {
1429         struct kvm_mmu_page *page;
1430
1431         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1432                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1433                                     struct kvm_mmu_page, link);
1434                 kvm_mmu_zap_page(vcpu->kvm, page);
1435         }
1436         free_page((unsigned long)vcpu->mmu.pae_root);
1437 }
1438
1439 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1440 {
1441         struct page *page;
1442         int i;
1443
1444         ASSERT(vcpu);
1445
1446         if (vcpu->kvm->n_requested_mmu_pages)
1447                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_requested_mmu_pages;
1448         else
1449                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_alloc_mmu_pages;
1450         /*
1451          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1452          * Therefore we need to allocate shadow page tables in the first
1453          * 4GB of memory, which happens to fit the DMA32 zone.
1454          */
1455         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1456         if (!page)
1457                 goto error_1;
1458         vcpu->mmu.pae_root = page_address(page);
1459         for (i = 0; i < 4; ++i)
1460                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1461
1462         return 0;
1463
1464 error_1:
1465         free_mmu_pages(vcpu);
1466         return -ENOMEM;
1467 }
1468
1469 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1470 {
1471         ASSERT(vcpu);
1472         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1473
1474         return alloc_mmu_pages(vcpu);
1475 }
1476
1477 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1478 {
1479         ASSERT(vcpu);
1480         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1481
1482         return init_kvm_mmu(vcpu);
1483 }
1484
1485 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1486 {
1487         ASSERT(vcpu);
1488
1489         destroy_kvm_mmu(vcpu);
1490         free_mmu_pages(vcpu);
1491         mmu_free_memory_caches(vcpu);
1492 }
1493
1494 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1495 {
1496         struct kvm_mmu_page *page;
1497
1498         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1499                 int i;
1500                 u64 *pt;
1501
1502                 if (!test_bit(slot, &page->slot_bitmap))
1503                         continue;
1504
1505                 pt = page->spt;
1506                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1507                         /* avoid RMW */
1508                         if (pt[i] & PT_WRITABLE_MASK)
1509                                 pt[i] &= ~PT_WRITABLE_MASK;
1510         }
1511 }
1512
1513 void kvm_mmu_zap_all(struct kvm *kvm)
1514 {
1515         struct kvm_mmu_page *page, *node;
1516
1517         list_for_each_entry_safe(page, node, &kvm->active_mmu_pages, link)
1518                 kvm_mmu_zap_page(kvm, page);
1519
1520         kvm_flush_remote_tlbs(kvm);
1521 }
1522
1523 void kvm_mmu_module_exit(void)
1524 {
1525         if (pte_chain_cache)
1526                 kmem_cache_destroy(pte_chain_cache);
1527         if (rmap_desc_cache)
1528                 kmem_cache_destroy(rmap_desc_cache);
1529         if (mmu_page_header_cache)
1530                 kmem_cache_destroy(mmu_page_header_cache);
1531 }
1532
1533 int kvm_mmu_module_init(void)
1534 {
1535         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1536                                             sizeof(struct kvm_pte_chain),
1537                                             0, 0, NULL);
1538         if (!pte_chain_cache)
1539                 goto nomem;
1540         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1541                                             sizeof(struct kvm_rmap_desc),
1542                                             0, 0, NULL);
1543         if (!rmap_desc_cache)
1544                 goto nomem;
1545
1546         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1547                                                   sizeof(struct kvm_mmu_page),
1548                                                   0, 0, NULL);
1549         if (!mmu_page_header_cache)
1550                 goto nomem;
1551
1552         return 0;
1553
1554 nomem:
1555         kvm_mmu_module_exit();
1556         return -ENOMEM;
1557 }
1558
1559 /*
1560  * Caculate mmu pages needed for kvm.
1561  */
1562 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
1563 {
1564         int i;
1565         unsigned int nr_mmu_pages;
1566         unsigned int  nr_pages = 0;
1567
1568         for (i = 0; i < kvm->nmemslots; i++)
1569                 nr_pages += kvm->memslots[i].npages;
1570
1571         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
1572         nr_mmu_pages = max(nr_mmu_pages,
1573                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
1574
1575         return nr_mmu_pages;
1576 }
1577
1578 #ifdef AUDIT
1579
1580 static const char *audit_msg;
1581
1582 static gva_t canonicalize(gva_t gva)
1583 {
1584 #ifdef CONFIG_X86_64
1585         gva = (long long)(gva << 16) >> 16;
1586 #endif
1587         return gva;
1588 }
1589
1590 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1591                                 gva_t va, int level)
1592 {
1593         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1594         int i;
1595         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1596
1597         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1598                 u64 ent = pt[i];
1599
1600                 if (ent == shadow_trap_nonpresent_pte)
1601                         continue;
1602
1603                 va = canonicalize(va);
1604                 if (level > 1) {
1605                         if (ent == shadow_notrap_nonpresent_pte)
1606                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1607                                        " in nonleaf level: levels %d gva %lx"
1608                                        " level %d pte %llx\n", audit_msg,
1609                                        vcpu->mmu.root_level, va, level, ent);
1610
1611                         audit_mappings_page(vcpu, ent, va, level - 1);
1612                 } else {
1613                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1614                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1615                         struct page *page;
1616
1617                         if (is_shadow_present_pte(ent)
1618                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1619                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1620                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1621                                        audit_msg, vcpu->mmu.root_level,
1622                                        va, gpa, hpa, ent,
1623                                        is_shadow_present_pte(ent));
1624                         else if (ent == shadow_notrap_nonpresent_pte
1625                                  && !is_error_hpa(hpa))
1626                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1627                                        " valid guest gva %lx\n", audit_msg, va);
1628                         page = pfn_to_page((gpa & PT64_BASE_ADDR_MASK)
1629                                            >> PAGE_SHIFT);
1630                         kvm_release_page_clean(page);
1631
1632                 }
1633         }
1634 }
1635
1636 static void audit_mappings(struct kvm_vcpu *vcpu)
1637 {
1638         unsigned i;
1639
1640         if (vcpu->mmu.root_level == 4)
1641                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1642         else
1643                 for (i = 0; i < 4; ++i)
1644                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1645                                 audit_mappings_page(vcpu,
1646                                                     vcpu->mmu.pae_root[i],
1647                                                     i << 30,
1648                                                     2);
1649 }
1650
1651 static int count_rmaps(struct kvm_vcpu *vcpu)
1652 {
1653         int nmaps = 0;
1654         int i, j, k;
1655
1656         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1657                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1658                 struct kvm_rmap_desc *d;
1659
1660                 for (j = 0; j < m->npages; ++j) {
1661                         unsigned long *rmapp = &m->rmap[j];
1662
1663                         if (!*rmapp)
1664                                 continue;
1665                         if (!(*rmapp & 1)) {
1666                                 ++nmaps;
1667                                 continue;
1668                         }
1669                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1670                         while (d) {
1671                                 for (k = 0; k < RMAP_EXT; ++k)
1672                                         if (d->shadow_ptes[k])
1673                                                 ++nmaps;
1674                                         else
1675                                                 break;
1676                                 d = d->more;
1677                         }
1678                 }
1679         }
1680         return nmaps;
1681 }
1682
1683 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1684 {
1685         int nmaps = 0;
1686         struct kvm_mmu_page *page;
1687         int i;
1688
1689         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1690                 u64 *pt = page->spt;
1691
1692                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1693                         continue;
1694
1695                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1696                         u64 ent = pt[i];
1697
1698                         if (!(ent & PT_PRESENT_MASK))
1699                                 continue;
1700                         if (!(ent & PT_WRITABLE_MASK))
1701                                 continue;
1702                         ++nmaps;
1703                 }
1704         }
1705         return nmaps;
1706 }
1707
1708 static void audit_rmap(struct kvm_vcpu *vcpu)
1709 {
1710         int n_rmap = count_rmaps(vcpu);
1711         int n_actual = count_writable_mappings(vcpu);
1712
1713         if (n_rmap != n_actual)
1714                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1715                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1716 }
1717
1718 static void audit_write_protection(struct kvm_vcpu *vcpu)
1719 {
1720         struct kvm_mmu_page *page;
1721         struct kvm_memory_slot *slot;
1722         unsigned long *rmapp;
1723         gfn_t gfn;
1724
1725         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1726                 if (page->role.metaphysical)
1727                         continue;
1728
1729                 slot = gfn_to_memslot(vcpu->kvm, page->gfn);
1730                 gfn = unalias_gfn(vcpu->kvm, page->gfn);
1731                 rmapp = &slot->rmap[gfn - slot->base_gfn];
1732                 if (*rmapp)
1733                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1734                                " mappings: gfn %lx role %x\n",
1735                                __FUNCTION__, audit_msg, page->gfn,
1736                                page->role.word);
1737         }
1738 }
1739
1740 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1741 {
1742         int olddbg = dbg;
1743
1744         dbg = 0;
1745         audit_msg = msg;
1746         audit_rmap(vcpu);
1747         audit_write_protection(vcpu);
1748         audit_mappings(vcpu);
1749         dbg = olddbg;
1750 }
1751
1752 #endif