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 hpa_t gva_to_hpa(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 UNMAPPED_GVA;
 890         return gpa_to_hpa(vcpu->kvm, gpa);
 891 }
 892
 893 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 894 {
 895         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 896
 897         if (gpa == UNMAPPED_GVA)
 898                 return NULL;
 899         return pfn_to_page(gpa_to_hpa(vcpu->kvm, gpa) >> PAGE_SHIFT);
 900 }
 901
 902 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 903 {
 904 }
 905
 906 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 907 {
 908         int level = PT32E_ROOT_LEVEL;
 909         hpa_t table_addr = vcpu->mmu.root_hpa;
 910         struct page *page;
 911
 912         page = pfn_to_page(p >> PAGE_SHIFT);
 913         for (; ; level--) {
 914                 u32 index = PT64_INDEX(v, level);
 915                 u64 *table;
 916                 u64 pte;
 917
 918                 ASSERT(VALID_PAGE(table_addr));
 919                 table = __va(table_addr);
 920
 921                 if (level == 1) {
 922                         int was_rmapped;
 923
 924                         pte = table[index];
 925                         was_rmapped = is_rmap_pte(pte);
 926                         if (is_shadow_present_pte(pte) && is_writeble_pte(pte)) {
 927                                 kvm_release_page_clean(page);
 928                                 return 0;
 929                         }
 930                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 931                         page_header_update_slot(vcpu->kvm, table,
 932                                                 v >> PAGE_SHIFT);
 933                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 934                                                                 PT_USER_MASK;
 935                         if (!was_rmapped)
 936                                 rmap_add(vcpu, &table[index], v >> PAGE_SHIFT);
 937                         else
 938                                 kvm_release_page_clean(page);
 939
 940                         return 0;
 941                 }
 942
 943                 if (table[index] == shadow_trap_nonpresent_pte) {
 944                         struct kvm_mmu_page *new_table;
 945                         gfn_t pseudo_gfn;
 946
 947                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 948                                 >> PAGE_SHIFT;
 949                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 950                                                      v, level - 1,
 951                                                      1, 3, &table[index]);
 952                         if (!new_table) {
 953                                 pgprintk("nonpaging_map: ENOMEM\n");
 954                                 kvm_release_page_clean(page);
 955                                 return -ENOMEM;
 956                         }
 957
 958                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 959                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 960                 }
 961                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 962         }
 963 }
 964
 965 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 966                                     struct kvm_mmu_page *sp)
 967 {
 968         int i;
 969
 970         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 971                 sp->spt[i] = shadow_trap_nonpresent_pte;
 972 }
 973
 974 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 975 {
 976         int i;
 977         struct kvm_mmu_page *page;
 978
 979         if (!VALID_PAGE(vcpu->mmu.root_hpa))
 980                 return;
 981 #ifdef CONFIG_X86_64
 982         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 983                 hpa_t root = vcpu->mmu.root_hpa;
 984
 985                 page = page_header(root);
 986                 --page->root_count;
 987                 vcpu->mmu.root_hpa = INVALID_PAGE;
 988                 return;
 989         }
 990 #endif
 991         for (i = 0; i < 4; ++i) {
 992                 hpa_t root = vcpu->mmu.pae_root[i];
 993
 994                 if (root) {
 995                         root &= PT64_BASE_ADDR_MASK;
 996                         page = page_header(root);
 997                         --page->root_count;
 998                 }
 999                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1000         }
1001         vcpu->mmu.root_hpa = INVALID_PAGE;
1002 }
1003
1004 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1005 {
1006         int i;
1007         gfn_t root_gfn;
1008         struct kvm_mmu_page *page;
1009
1010         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
1011
1012 #ifdef CONFIG_X86_64
1013         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1014                 hpa_t root = vcpu->mmu.root_hpa;
1015
1016                 ASSERT(!VALID_PAGE(root));
1017                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
1018                                         PT64_ROOT_LEVEL, 0, 0, NULL);
1019                 root = __pa(page->spt);
1020                 ++page->root_count;
1021                 vcpu->mmu.root_hpa = root;
1022                 return;
1023         }
1024 #endif
1025         for (i = 0; i < 4; ++i) {
1026                 hpa_t root = vcpu->mmu.pae_root[i];
1027
1028                 ASSERT(!VALID_PAGE(root));
1029                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
1030                         if (!is_present_pte(vcpu->pdptrs[i])) {
1031                                 vcpu->mmu.pae_root[i] = 0;
1032                                 continue;
1033                         }
1034                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
1035                 } else if (vcpu->mmu.root_level == 0)
1036                         root_gfn = 0;
1037                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1038                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
1039                                         0, NULL);
1040                 root = __pa(page->spt);
1041                 ++page->root_count;
1042                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
1043         }
1044         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
1045 }
1046
1047 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1048 {
1049         return vaddr;
1050 }
1051
1052 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1053                                u32 error_code)
1054 {
1055         gpa_t addr = gva;
1056         hpa_t paddr;
1057         int r;
1058
1059         r = mmu_topup_memory_caches(vcpu);
1060         if (r)
1061                 return r;
1062
1063         ASSERT(vcpu);
1064         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
1065
1066
1067         paddr = gpa_to_hpa(vcpu->kvm, addr & PT64_BASE_ADDR_MASK);
1068
1069         if (is_error_hpa(paddr)) {
1070                 kvm_release_page_clean(pfn_to_page((paddr & PT64_BASE_ADDR_MASK)
1071                                        >> PAGE_SHIFT));
1072                 return 1;
1073         }
1074
1075         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
1076 }
1077
1078 static void nonpaging_free(struct kvm_vcpu *vcpu)
1079 {
1080         mmu_free_roots(vcpu);
1081 }
1082
1083 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1084 {
1085         struct kvm_mmu *context = &vcpu->mmu;
1086
1087         context->new_cr3 = nonpaging_new_cr3;
1088         context->page_fault = nonpaging_page_fault;
1089         context->gva_to_gpa = nonpaging_gva_to_gpa;
1090         context->free = nonpaging_free;
1091         context->prefetch_page = nonpaging_prefetch_page;
1092         context->root_level = 0;
1093         context->shadow_root_level = PT32E_ROOT_LEVEL;
1094         context->root_hpa = INVALID_PAGE;
1095         return 0;
1096 }
1097
1098 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1099 {
1100         ++vcpu->stat.tlb_flush;
1101         kvm_x86_ops->tlb_flush(vcpu);
1102 }
1103
1104 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1105 {
1106         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
1107         mmu_free_roots(vcpu);
1108 }
1109
1110 static void inject_page_fault(struct kvm_vcpu *vcpu,
1111                               u64 addr,
1112                               u32 err_code)
1113 {
1114         kvm_x86_ops->inject_page_fault(vcpu, addr, err_code);
1115 }
1116
1117 static void paging_free(struct kvm_vcpu *vcpu)
1118 {
1119         nonpaging_free(vcpu);
1120 }
1121
1122 #define PTTYPE 64
1123 #include "paging_tmpl.h"
1124 #undef PTTYPE
1125
1126 #define PTTYPE 32
1127 #include "paging_tmpl.h"
1128 #undef PTTYPE
1129
1130 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1131 {
1132         struct kvm_mmu *context = &vcpu->mmu;
1133
1134         ASSERT(is_pae(vcpu));
1135         context->new_cr3 = paging_new_cr3;
1136         context->page_fault = paging64_page_fault;
1137         context->gva_to_gpa = paging64_gva_to_gpa;
1138         context->prefetch_page = paging64_prefetch_page;
1139         context->free = paging_free;
1140         context->root_level = level;
1141         context->shadow_root_level = level;
1142         context->root_hpa = INVALID_PAGE;
1143         return 0;
1144 }
1145
1146 static int paging64_init_context(struct kvm_vcpu *vcpu)
1147 {
1148         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1149 }
1150
1151 static int paging32_init_context(struct kvm_vcpu *vcpu)
1152 {
1153         struct kvm_mmu *context = &vcpu->mmu;
1154
1155         context->new_cr3 = paging_new_cr3;
1156         context->page_fault = paging32_page_fault;
1157         context->gva_to_gpa = paging32_gva_to_gpa;
1158         context->free = paging_free;
1159         context->prefetch_page = paging32_prefetch_page;
1160         context->root_level = PT32_ROOT_LEVEL;
1161         context->shadow_root_level = PT32E_ROOT_LEVEL;
1162         context->root_hpa = INVALID_PAGE;
1163         return 0;
1164 }
1165
1166 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1167 {
1168         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1169 }
1170
1171 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1172 {
1173         ASSERT(vcpu);
1174         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1175
1176         if (!is_paging(vcpu))
1177                 return nonpaging_init_context(vcpu);
1178         else if (is_long_mode(vcpu))
1179                 return paging64_init_context(vcpu);
1180         else if (is_pae(vcpu))
1181                 return paging32E_init_context(vcpu);
1182         else
1183                 return paging32_init_context(vcpu);
1184 }
1185
1186 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1187 {
1188         ASSERT(vcpu);
1189         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1190                 vcpu->mmu.free(vcpu);
1191                 vcpu->mmu.root_hpa = INVALID_PAGE;
1192         }
1193 }
1194
1195 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1196 {
1197         destroy_kvm_mmu(vcpu);
1198         return init_kvm_mmu(vcpu);
1199 }
1200 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1201
1202 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1203 {
1204         int r;
1205
1206         mutex_lock(&vcpu->kvm->lock);
1207         r = mmu_topup_memory_caches(vcpu);
1208         if (r)
1209                 goto out;
1210         mmu_alloc_roots(vcpu);
1211         kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1212         kvm_mmu_flush_tlb(vcpu);
1213 out:
1214         mutex_unlock(&vcpu->kvm->lock);
1215         return r;
1216 }
1217 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1218
1219 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1220 {
1221         mmu_free_roots(vcpu);
1222 }
1223
1224 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1225                                   struct kvm_mmu_page *page,
1226                                   u64 *spte)
1227 {
1228         u64 pte;
1229         struct kvm_mmu_page *child;
1230
1231         pte = *spte;
1232         if (is_shadow_present_pte(pte)) {
1233                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1234                         rmap_remove(vcpu->kvm, spte);
1235                 else {
1236                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1237                         mmu_page_remove_parent_pte(child, spte);
1238                 }
1239         }
1240         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1241 }
1242
1243 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1244                                   struct kvm_mmu_page *page,
1245                                   u64 *spte,
1246                                   const void *new, int bytes,
1247                                   int offset_in_pte)
1248 {
1249         if (page->role.level != PT_PAGE_TABLE_LEVEL) {
1250                 ++vcpu->kvm->stat.mmu_pde_zapped;
1251                 return;
1252         }
1253
1254         ++vcpu->kvm->stat.mmu_pte_updated;
1255         if (page->role.glevels == PT32_ROOT_LEVEL)
1256                 paging32_update_pte(vcpu, page, spte, new, bytes,
1257                                     offset_in_pte);
1258         else
1259                 paging64_update_pte(vcpu, page, spte, new, bytes,
1260                                     offset_in_pte);
1261 }
1262
1263 static bool need_remote_flush(u64 old, u64 new)
1264 {
1265         if (!is_shadow_present_pte(old))
1266                 return false;
1267         if (!is_shadow_present_pte(new))
1268                 return true;
1269         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1270                 return true;
1271         old ^= PT64_NX_MASK;
1272         new ^= PT64_NX_MASK;
1273         return (old & ~new & PT64_PERM_MASK) != 0;
1274 }
1275
1276 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1277 {
1278         if (need_remote_flush(old, new))
1279                 kvm_flush_remote_tlbs(vcpu->kvm);
1280         else
1281                 kvm_mmu_flush_tlb(vcpu);
1282 }
1283
1284 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1285 {
1286         u64 *spte = vcpu->last_pte_updated;
1287
1288         return !!(spte && (*spte & PT_ACCESSED_MASK));
1289 }
1290
1291 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1292                        const u8 *new, int bytes)
1293 {
1294         gfn_t gfn = gpa >> PAGE_SHIFT;
1295         struct kvm_mmu_page *page;
1296         struct hlist_node *node, *n;
1297         struct hlist_head *bucket;
1298         unsigned index;
1299         u64 entry;
1300         u64 *spte;
1301         unsigned offset = offset_in_page(gpa);
1302         unsigned pte_size;
1303         unsigned page_offset;
1304         unsigned misaligned;
1305         unsigned quadrant;
1306         int level;
1307         int flooded = 0;
1308         int npte;
1309
1310         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1311         ++vcpu->kvm->stat.mmu_pte_write;
1312         kvm_mmu_audit(vcpu, "pre pte write");
1313         if (gfn == vcpu->last_pt_write_gfn
1314             && !last_updated_pte_accessed(vcpu)) {
1315                 ++vcpu->last_pt_write_count;
1316                 if (vcpu->last_pt_write_count >= 3)
1317                         flooded = 1;
1318         } else {
1319                 vcpu->last_pt_write_gfn = gfn;
1320                 vcpu->last_pt_write_count = 1;
1321                 vcpu->last_pte_updated = NULL;
1322         }
1323         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1324         bucket = &vcpu->kvm->mmu_page_hash[index];
1325         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1326                 if (page->gfn != gfn || page->role.metaphysical)
1327                         continue;
1328                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1329                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1330                 misaligned |= bytes < 4;
1331                 if (misaligned || flooded) {
1332                         /*
1333                          * Misaligned accesses are too much trouble to fix
1334                          * up; also, they usually indicate a page is not used
1335                          * as a page table.
1336                          *
1337                          * If we're seeing too many writes to a page,
1338                          * it may no longer be a page table, or we may be
1339                          * forking, in which case it is better to unmap the
1340                          * page.
1341                          */
1342                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1343                                  gpa, bytes, page->role.word);
1344                         kvm_mmu_zap_page(vcpu->kvm, page);
1345                         ++vcpu->kvm->stat.mmu_flooded;
1346                         continue;
1347                 }
1348                 page_offset = offset;
1349                 level = page->role.level;
1350                 npte = 1;
1351                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1352                         page_offset <<= 1;      /* 32->64 */
1353                         /*
1354                          * A 32-bit pde maps 4MB while the shadow pdes map
1355                          * only 2MB.  So we need to double the offset again
1356                          * and zap two pdes instead of one.
1357                          */
1358                         if (level == PT32_ROOT_LEVEL) {
1359                                 page_offset &= ~7; /* kill rounding error */
1360                                 page_offset <<= 1;
1361                                 npte = 2;
1362                         }
1363                         quadrant = page_offset >> PAGE_SHIFT;
1364                         page_offset &= ~PAGE_MASK;
1365                         if (quadrant != page->role.quadrant)
1366                                 continue;
1367                 }
1368                 spte = &page->spt[page_offset / sizeof(*spte)];
1369                 while (npte--) {
1370                         entry = *spte;
1371                         mmu_pte_write_zap_pte(vcpu, page, spte);
1372                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes,
1373                                               page_offset & (pte_size - 1));
1374                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1375                         ++spte;
1376                 }
1377         }
1378         kvm_mmu_audit(vcpu, "post pte write");
1379 }
1380
1381 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1382 {
1383         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1384
1385         return kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1386 }
1387
1388 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1389 {
1390         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1391                 struct kvm_mmu_page *page;
1392
1393                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1394                                     struct kvm_mmu_page, link);
1395                 kvm_mmu_zap_page(vcpu->kvm, page);
1396                 ++vcpu->kvm->stat.mmu_recycled;
1397         }
1398 }
1399
1400 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1401 {
1402         int r;
1403         enum emulation_result er;
1404
1405         mutex_lock(&vcpu->kvm->lock);
1406         r = vcpu->mmu.page_fault(vcpu, cr2, error_code);
1407         if (r < 0)
1408                 goto out;
1409
1410         if (!r) {
1411                 r = 1;
1412                 goto out;
1413         }
1414
1415         r = mmu_topup_memory_caches(vcpu);
1416         if (r)
1417                 goto out;
1418
1419         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1420         mutex_unlock(&vcpu->kvm->lock);
1421
1422         switch (er) {
1423         case EMULATE_DONE:
1424                 return 1;
1425         case EMULATE_DO_MMIO:
1426                 ++vcpu->stat.mmio_exits;
1427                 return 0;
1428         case EMULATE_FAIL:
1429                 kvm_report_emulation_failure(vcpu, "pagetable");
1430                 return 1;
1431         default:
1432                 BUG();
1433         }
1434 out:
1435         mutex_unlock(&vcpu->kvm->lock);
1436         return r;
1437 }
1438 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1439
1440 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1441 {
1442         struct kvm_mmu_page *page;
1443
1444         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1445                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1446                                     struct kvm_mmu_page, link);
1447                 kvm_mmu_zap_page(vcpu->kvm, page);
1448         }
1449         free_page((unsigned long)vcpu->mmu.pae_root);
1450 }
1451
1452 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1453 {
1454         struct page *page;
1455         int i;
1456
1457         ASSERT(vcpu);
1458
1459         if (vcpu->kvm->n_requested_mmu_pages)
1460                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_requested_mmu_pages;
1461         else
1462                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_alloc_mmu_pages;
1463         /*
1464          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1465          * Therefore we need to allocate shadow page tables in the first
1466          * 4GB of memory, which happens to fit the DMA32 zone.
1467          */
1468         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1469         if (!page)
1470                 goto error_1;
1471         vcpu->mmu.pae_root = page_address(page);
1472         for (i = 0; i < 4; ++i)
1473                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1474
1475         return 0;
1476
1477 error_1:
1478         free_mmu_pages(vcpu);
1479         return -ENOMEM;
1480 }
1481
1482 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1483 {
1484         ASSERT(vcpu);
1485         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1486
1487         return alloc_mmu_pages(vcpu);
1488 }
1489
1490 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1491 {
1492         ASSERT(vcpu);
1493         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1494
1495         return init_kvm_mmu(vcpu);
1496 }
1497
1498 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1499 {
1500         ASSERT(vcpu);
1501
1502         destroy_kvm_mmu(vcpu);
1503         free_mmu_pages(vcpu);
1504         mmu_free_memory_caches(vcpu);
1505 }
1506
1507 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1508 {
1509         struct kvm_mmu_page *page;
1510
1511         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1512                 int i;
1513                 u64 *pt;
1514
1515                 if (!test_bit(slot, &page->slot_bitmap))
1516                         continue;
1517
1518                 pt = page->spt;
1519                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1520                         /* avoid RMW */
1521                         if (pt[i] & PT_WRITABLE_MASK)
1522                                 pt[i] &= ~PT_WRITABLE_MASK;
1523         }
1524 }
1525
1526 void kvm_mmu_zap_all(struct kvm *kvm)
1527 {
1528         struct kvm_mmu_page *page, *node;
1529
1530         list_for_each_entry_safe(page, node, &kvm->active_mmu_pages, link)
1531                 kvm_mmu_zap_page(kvm, page);
1532
1533         kvm_flush_remote_tlbs(kvm);
1534 }
1535
1536 void kvm_mmu_module_exit(void)
1537 {
1538         if (pte_chain_cache)
1539                 kmem_cache_destroy(pte_chain_cache);
1540         if (rmap_desc_cache)
1541                 kmem_cache_destroy(rmap_desc_cache);
1542         if (mmu_page_header_cache)
1543                 kmem_cache_destroy(mmu_page_header_cache);
1544 }
1545
1546 int kvm_mmu_module_init(void)
1547 {
1548         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1549                                             sizeof(struct kvm_pte_chain),
1550                                             0, 0, NULL);
1551         if (!pte_chain_cache)
1552                 goto nomem;
1553         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1554                                             sizeof(struct kvm_rmap_desc),
1555                                             0, 0, NULL);
1556         if (!rmap_desc_cache)
1557                 goto nomem;
1558
1559         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1560                                                   sizeof(struct kvm_mmu_page),
1561                                                   0, 0, NULL);
1562         if (!mmu_page_header_cache)
1563                 goto nomem;
1564
1565         return 0;
1566
1567 nomem:
1568         kvm_mmu_module_exit();
1569         return -ENOMEM;
1570 }
1571
1572 /*
1573  * Caculate mmu pages needed for kvm.
1574  */
1575 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
1576 {
1577         int i;
1578         unsigned int nr_mmu_pages;
1579         unsigned int  nr_pages = 0;
1580
1581         for (i = 0; i < kvm->nmemslots; i++)
1582                 nr_pages += kvm->memslots[i].npages;
1583
1584         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
1585         nr_mmu_pages = max(nr_mmu_pages,
1586                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
1587
1588         return nr_mmu_pages;
1589 }
1590
1591 #ifdef AUDIT
1592
1593 static const char *audit_msg;
1594
1595 static gva_t canonicalize(gva_t gva)
1596 {
1597 #ifdef CONFIG_X86_64
1598         gva = (long long)(gva << 16) >> 16;
1599 #endif
1600         return gva;
1601 }
1602
1603 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1604                                 gva_t va, int level)
1605 {
1606         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1607         int i;
1608         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1609
1610         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1611                 u64 ent = pt[i];
1612
1613                 if (ent == shadow_trap_nonpresent_pte)
1614                         continue;
1615
1616                 va = canonicalize(va);
1617                 if (level > 1) {
1618                         if (ent == shadow_notrap_nonpresent_pte)
1619                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1620                                        " in nonleaf level: levels %d gva %lx"
1621                                        " level %d pte %llx\n", audit_msg,
1622                                        vcpu->mmu.root_level, va, level, ent);
1623
1624                         audit_mappings_page(vcpu, ent, va, level - 1);
1625                 } else {
1626                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1627                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1628                         struct page *page;
1629
1630                         if (is_shadow_present_pte(ent)
1631                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1632                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1633                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1634                                        audit_msg, vcpu->mmu.root_level,
1635                                        va, gpa, hpa, ent,
1636                                        is_shadow_present_pte(ent));
1637                         else if (ent == shadow_notrap_nonpresent_pte
1638                                  && !is_error_hpa(hpa))
1639                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1640                                        " valid guest gva %lx\n", audit_msg, va);
1641                         page = pfn_to_page((gpa & PT64_BASE_ADDR_MASK)
1642                                            >> PAGE_SHIFT);
1643                         kvm_release_page_clean(page);
1644
1645                 }
1646         }
1647 }
1648
1649 static void audit_mappings(struct kvm_vcpu *vcpu)
1650 {
1651         unsigned i;
1652
1653         if (vcpu->mmu.root_level == 4)
1654                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1655         else
1656                 for (i = 0; i < 4; ++i)
1657                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1658                                 audit_mappings_page(vcpu,
1659                                                     vcpu->mmu.pae_root[i],
1660                                                     i << 30,
1661                                                     2);
1662 }
1663
1664 static int count_rmaps(struct kvm_vcpu *vcpu)
1665 {
1666         int nmaps = 0;
1667         int i, j, k;
1668
1669         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1670                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1671                 struct kvm_rmap_desc *d;
1672
1673                 for (j = 0; j < m->npages; ++j) {
1674                         unsigned long *rmapp = &m->rmap[j];
1675
1676                         if (!*rmapp)
1677                                 continue;
1678                         if (!(*rmapp & 1)) {
1679                                 ++nmaps;
1680                                 continue;
1681                         }
1682                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1683                         while (d) {
1684                                 for (k = 0; k < RMAP_EXT; ++k)
1685                                         if (d->shadow_ptes[k])
1686                                                 ++nmaps;
1687                                         else
1688                                                 break;
1689                                 d = d->more;
1690                         }
1691                 }
1692         }
1693         return nmaps;
1694 }
1695
1696 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1697 {
1698         int nmaps = 0;
1699         struct kvm_mmu_page *page;
1700         int i;
1701
1702         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1703                 u64 *pt = page->spt;
1704
1705                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1706                         continue;
1707
1708                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1709                         u64 ent = pt[i];
1710
1711                         if (!(ent & PT_PRESENT_MASK))
1712                                 continue;
1713                         if (!(ent & PT_WRITABLE_MASK))
1714                                 continue;
1715                         ++nmaps;
1716                 }
1717         }
1718         return nmaps;
1719 }
1720
1721 static void audit_rmap(struct kvm_vcpu *vcpu)
1722 {
1723         int n_rmap = count_rmaps(vcpu);
1724         int n_actual = count_writable_mappings(vcpu);
1725
1726         if (n_rmap != n_actual)
1727                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1728                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1729 }
1730
1731 static void audit_write_protection(struct kvm_vcpu *vcpu)
1732 {
1733         struct kvm_mmu_page *page;
1734         struct kvm_memory_slot *slot;
1735         unsigned long *rmapp;
1736         gfn_t gfn;
1737
1738         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1739                 if (page->role.metaphysical)
1740                         continue;
1741
1742                 slot = gfn_to_memslot(vcpu->kvm, page->gfn);
1743                 gfn = unalias_gfn(vcpu->kvm, page->gfn);
1744                 rmapp = &slot->rmap[gfn - slot->base_gfn];
1745                 if (*rmapp)
1746                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1747                                " mappings: gfn %lx role %x\n",
1748                                __FUNCTION__, audit_msg, page->gfn,
1749                                page->role.word);
1750         }
1751 }
1752
1753 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1754 {
1755         int olddbg = dbg;
1756
1757         dbg = 0;
1758         audit_msg = msg;
1759         audit_rmap(vcpu);
1760         audit_write_protection(vcpu);
1761         audit_mappings(vcpu);
1762         dbg = olddbg;
1763 }
1764
1765 #endif