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 #include <linux/swap.h>
  30
  31 #include <asm/page.h>
  32 #include <asm/cmpxchg.h>
  33 #include <asm/io.h>
  34
  35 #undef MMU_DEBUG
  36
  37 #undef AUDIT
  38
  39 #ifdef AUDIT
  40 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  41 #else
  42 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  43 #endif
  44
  45 #ifdef MMU_DEBUG
  46
  47 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  48 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  49
  50 #else
  51
  52 #define pgprintk(x...) do { } while (0)
  53 #define rmap_printk(x...) do { } while (0)
  54
  55 #endif
  56
  57 #if defined(MMU_DEBUG) || defined(AUDIT)
  58 static int dbg = 1;
  59 #endif
  60
  61 #ifndef MMU_DEBUG
  62 #define ASSERT(x) do { } while (0)
  63 #else
  64 #define ASSERT(x)                                                       \
  65         if (!(x)) {                                                     \
  66                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  67                        __FILE__, __LINE__, #x);                         \
  68         }
  69 #endif
  70
  71 #define PT64_PT_BITS 9
  72 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  73 #define PT32_PT_BITS 10
  74 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  75
  76 #define PT_WRITABLE_SHIFT 1
  77
  78 #define PT_PRESENT_MASK (1ULL << 0)
  79 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  80 #define PT_USER_MASK (1ULL << 2)
  81 #define PT_PWT_MASK (1ULL << 3)
  82 #define PT_PCD_MASK (1ULL << 4)
  83 #define PT_ACCESSED_MASK (1ULL << 5)
  84 #define PT_DIRTY_MASK (1ULL << 6)
  85 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  86 #define PT_PAT_MASK (1ULL << 7)
  87 #define PT_GLOBAL_MASK (1ULL << 8)
  88 #define PT64_NX_SHIFT 63
  89 #define PT64_NX_MASK (1ULL << PT64_NX_SHIFT)
  90
  91 #define PT_PAT_SHIFT 7
  92 #define PT_DIR_PAT_SHIFT 12
  93 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  94
  95 #define PT32_DIR_PSE36_SIZE 4
  96 #define PT32_DIR_PSE36_SHIFT 13
  97 #define PT32_DIR_PSE36_MASK \
  98         (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  99
 100
 101 #define PT_FIRST_AVAIL_BITS_SHIFT 9
 102 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
 103
 104 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 105
 106 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 107
 108 #define PT64_LEVEL_BITS 9
 109
 110 #define PT64_LEVEL_SHIFT(level) \
 111                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
 112
 113 #define PT64_LEVEL_MASK(level) \
 114                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 115
 116 #define PT64_INDEX(address, level)\
 117         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 118
 119
 120 #define PT32_LEVEL_BITS 10
 121
 122 #define PT32_LEVEL_SHIFT(level) \
 123                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 124
 125 #define PT32_LEVEL_MASK(level) \
 126                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 127
 128 #define PT32_INDEX(address, level)\
 129         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 130
 131
 132 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 133 #define PT64_DIR_BASE_ADDR_MASK \
 134         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 135
 136 #define PT32_BASE_ADDR_MASK PAGE_MASK
 137 #define PT32_DIR_BASE_ADDR_MASK \
 138         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 139
 140 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 141                         | PT64_NX_MASK)
 142
 143 #define PFERR_PRESENT_MASK (1U << 0)
 144 #define PFERR_WRITE_MASK (1U << 1)
 145 #define PFERR_USER_MASK (1U << 2)
 146 #define PFERR_FETCH_MASK (1U << 4)
 147
 148 #define PT64_ROOT_LEVEL 4
 149 #define PT32_ROOT_LEVEL 2
 150 #define PT32E_ROOT_LEVEL 3
 151
 152 #define PT_DIRECTORY_LEVEL 2
 153 #define PT_PAGE_TABLE_LEVEL 1
 154
 155 #define RMAP_EXT 4
 156
 157 #define ACC_EXEC_MASK    1
 158 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 159 #define ACC_USER_MASK    PT_USER_MASK
 160 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 161
 162 struct kvm_rmap_desc {
 163         u64 *shadow_ptes[RMAP_EXT];
 164         struct kvm_rmap_desc *more;
 165 };
 166
 167 static struct kmem_cache *pte_chain_cache;
 168 static struct kmem_cache *rmap_desc_cache;
 169 static struct kmem_cache *mmu_page_header_cache;
 170
 171 static u64 __read_mostly shadow_trap_nonpresent_pte;
 172 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 173
 174 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 175 {
 176         shadow_trap_nonpresent_pte = trap_pte;
 177         shadow_notrap_nonpresent_pte = notrap_pte;
 178 }
 179 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 180
 181 static int is_write_protection(struct kvm_vcpu *vcpu)
 182 {
 183         return vcpu->cr0 & X86_CR0_WP;
 184 }
 185
 186 static int is_cpuid_PSE36(void)
 187 {
 188         return 1;
 189 }
 190
 191 static int is_nx(struct kvm_vcpu *vcpu)
 192 {
 193         return vcpu->shadow_efer & EFER_NX;
 194 }
 195
 196 static int is_present_pte(unsigned long pte)
 197 {
 198         return pte & PT_PRESENT_MASK;
 199 }
 200
 201 static int is_shadow_present_pte(u64 pte)
 202 {
 203         pte &= ~PT_SHADOW_IO_MARK;
 204         return pte != shadow_trap_nonpresent_pte
 205                 && pte != shadow_notrap_nonpresent_pte;
 206 }
 207
 208 static int is_writeble_pte(unsigned long pte)
 209 {
 210         return pte & PT_WRITABLE_MASK;
 211 }
 212
 213 static int is_dirty_pte(unsigned long pte)
 214 {
 215         return pte & PT_DIRTY_MASK;
 216 }
 217
 218 static int is_io_pte(unsigned long pte)
 219 {
 220         return pte & PT_SHADOW_IO_MARK;
 221 }
 222
 223 static int is_rmap_pte(u64 pte)
 224 {
 225         return pte != shadow_trap_nonpresent_pte
 226                 && pte != shadow_notrap_nonpresent_pte;
 227 }
 228
 229 static gfn_t pse36_gfn_delta(u32 gpte)
 230 {
 231         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 232
 233         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 234 }
 235
 236 static void set_shadow_pte(u64 *sptep, u64 spte)
 237 {
 238 #ifdef CONFIG_X86_64
 239         set_64bit((unsigned long *)sptep, spte);
 240 #else
 241         set_64bit((unsigned long long *)sptep, spte);
 242 #endif
 243 }
 244
 245 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 246                                   struct kmem_cache *base_cache, int min)
 247 {
 248         void *obj;
 249
 250         if (cache->nobjs >= min)
 251                 return 0;
 252         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 253                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 254                 if (!obj)
 255                         return -ENOMEM;
 256                 cache->objects[cache->nobjs++] = obj;
 257         }
 258         return 0;
 259 }
 260
 261 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 262 {
 263         while (mc->nobjs)
 264                 kfree(mc->objects[--mc->nobjs]);
 265 }
 266
 267 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 268                                        int min)
 269 {
 270         struct page *page;
 271
 272         if (cache->nobjs >= min)
 273                 return 0;
 274         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 275                 page = alloc_page(GFP_KERNEL);
 276                 if (!page)
 277                         return -ENOMEM;
 278                 set_page_private(page, 0);
 279                 cache->objects[cache->nobjs++] = page_address(page);
 280         }
 281         return 0;
 282 }
 283
 284 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 285 {
 286         while (mc->nobjs)
 287                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 288 }
 289
 290 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 291 {
 292         int r;
 293
 294         kvm_mmu_free_some_pages(vcpu);
 295         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 296                                    pte_chain_cache, 4);
 297         if (r)
 298                 goto out;
 299         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 300                                    rmap_desc_cache, 1);
 301         if (r)
 302                 goto out;
 303         r = mmu_topup_memory_cache_page(&vcpu->mmu_page_cache, 8);
 304         if (r)
 305                 goto out;
 306         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 307                                    mmu_page_header_cache, 4);
 308 out:
 309         return r;
 310 }
 311
 312 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 313 {
 314         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 315         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 316         mmu_free_memory_cache_page(&vcpu->mmu_page_cache);
 317         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 318 }
 319
 320 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 321                                     size_t size)
 322 {
 323         void *p;
 324
 325         BUG_ON(!mc->nobjs);
 326         p = mc->objects[--mc->nobjs];
 327         memset(p, 0, size);
 328         return p;
 329 }
 330
 331 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 332 {
 333         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 334                                       sizeof(struct kvm_pte_chain));
 335 }
 336
 337 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 338 {
 339         kfree(pc);
 340 }
 341
 342 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 343 {
 344         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 345                                       sizeof(struct kvm_rmap_desc));
 346 }
 347
 348 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 349 {
 350         kfree(rd);
 351 }
 352
 353 /*
 354  * Take gfn and return the reverse mapping to it.
 355  * Note: gfn must be unaliased before this function get called
 356  */
 357
 358 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
 359 {
 360         struct kvm_memory_slot *slot;
 361
 362         slot = gfn_to_memslot(kvm, gfn);
 363         return &slot->rmap[gfn - slot->base_gfn];
 364 }
 365
 366 /*
 367  * Reverse mapping data structures:
 368  *
 369  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 370  * that points to page_address(page).
 371  *
 372  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 373  * containing more mappings.
 374  */
 375 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 376 {
 377         struct kvm_mmu_page *sp;
 378         struct kvm_rmap_desc *desc;
 379         unsigned long *rmapp;
 380         int i;
 381
 382         if (!is_rmap_pte(*spte))
 383                 return;
 384         gfn = unalias_gfn(vcpu->kvm, gfn);
 385         sp = page_header(__pa(spte));
 386         sp->gfns[spte - sp->spt] = gfn;
 387         rmapp = gfn_to_rmap(vcpu->kvm, gfn);
 388         if (!*rmapp) {
 389                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 390                 *rmapp = (unsigned long)spte;
 391         } else if (!(*rmapp & 1)) {
 392                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 393                 desc = mmu_alloc_rmap_desc(vcpu);
 394                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 395                 desc->shadow_ptes[1] = spte;
 396                 *rmapp = (unsigned long)desc | 1;
 397         } else {
 398                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 399                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 400                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 401                         desc = desc->more;
 402                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 403                         desc->more = mmu_alloc_rmap_desc(vcpu);
 404                         desc = desc->more;
 405                 }
 406                 for (i = 0; desc->shadow_ptes[i]; ++i)
 407                         ;
 408                 desc->shadow_ptes[i] = spte;
 409         }
 410 }
 411
 412 static void rmap_desc_remove_entry(unsigned long *rmapp,
 413                                    struct kvm_rmap_desc *desc,
 414                                    int i,
 415                                    struct kvm_rmap_desc *prev_desc)
 416 {
 417         int j;
 418
 419         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 420                 ;
 421         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 422         desc->shadow_ptes[j] = NULL;
 423         if (j != 0)
 424                 return;
 425         if (!prev_desc && !desc->more)
 426                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 427         else
 428                 if (prev_desc)
 429                         prev_desc->more = desc->more;
 430                 else
 431                         *rmapp = (unsigned long)desc->more | 1;
 432         mmu_free_rmap_desc(desc);
 433 }
 434
 435 static void rmap_remove(struct kvm *kvm, u64 *spte)
 436 {
 437         struct kvm_rmap_desc *desc;
 438         struct kvm_rmap_desc *prev_desc;
 439         struct kvm_mmu_page *sp;
 440         struct page *page;
 441         unsigned long *rmapp;
 442         int i;
 443
 444         if (!is_rmap_pte(*spte))
 445                 return;
 446         sp = page_header(__pa(spte));
 447         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 448         mark_page_accessed(page);
 449         if (is_writeble_pte(*spte))
 450                 kvm_release_page_dirty(page);
 451         else
 452                 kvm_release_page_clean(page);
 453         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt]);
 454         if (!*rmapp) {
 455                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 456                 BUG();
 457         } else if (!(*rmapp & 1)) {
 458                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 459                 if ((u64 *)*rmapp != spte) {
 460                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 461                                spte, *spte);
 462                         BUG();
 463                 }
 464                 *rmapp = 0;
 465         } else {
 466                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 467                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 468                 prev_desc = NULL;
 469                 while (desc) {
 470                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 471                                 if (desc->shadow_ptes[i] == spte) {
 472                                         rmap_desc_remove_entry(rmapp,
 473                                                                desc, i,
 474                                                                prev_desc);
 475                                         return;
 476                                 }
 477                         prev_desc = desc;
 478                         desc = desc->more;
 479                 }
 480                 BUG();
 481         }
 482 }
 483
 484 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 485 {
 486         struct kvm_rmap_desc *desc;
 487         struct kvm_rmap_desc *prev_desc;
 488         u64 *prev_spte;
 489         int i;
 490
 491         if (!*rmapp)
 492                 return NULL;
 493         else if (!(*rmapp & 1)) {
 494                 if (!spte)
 495                         return (u64 *)*rmapp;
 496                 return NULL;
 497         }
 498         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 499         prev_desc = NULL;
 500         prev_spte = NULL;
 501         while (desc) {
 502                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 503                         if (prev_spte == spte)
 504                                 return desc->shadow_ptes[i];
 505                         prev_spte = desc->shadow_ptes[i];
 506                 }
 507                 desc = desc->more;
 508         }
 509         return NULL;
 510 }
 511
 512 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 513 {
 514         unsigned long *rmapp;
 515         u64 *spte;
 516
 517         gfn = unalias_gfn(kvm, gfn);
 518         rmapp = gfn_to_rmap(kvm, gfn);
 519
 520         spte = rmap_next(kvm, rmapp, NULL);
 521         while (spte) {
 522                 BUG_ON(!spte);
 523                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 524                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 525                 if (is_writeble_pte(*spte))
 526                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 527                 kvm_flush_remote_tlbs(kvm);
 528                 spte = rmap_next(kvm, rmapp, spte);
 529         }
 530 }
 531
 532 #ifdef MMU_DEBUG
 533 static int is_empty_shadow_page(u64 *spt)
 534 {
 535         u64 *pos;
 536         u64 *end;
 537
 538         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 539                 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
 540                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 541                                pos, *pos);
 542                         return 0;
 543                 }
 544         return 1;
 545 }
 546 #endif
 547
 548 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 549 {
 550         ASSERT(is_empty_shadow_page(sp->spt));
 551         list_del(&sp->link);
 552         __free_page(virt_to_page(sp->spt));
 553         __free_page(virt_to_page(sp->gfns));
 554         kfree(sp);
 555         ++kvm->n_free_mmu_pages;
 556 }
 557
 558 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 559 {
 560         return gfn;
 561 }
 562
 563 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 564                                                u64 *parent_pte)
 565 {
 566         struct kvm_mmu_page *sp;
 567
 568         if (!vcpu->kvm->n_free_mmu_pages)
 569                 return NULL;
 570
 571         sp = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache, sizeof *sp);
 572         sp->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 573         sp->gfns = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 574         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 575         list_add(&sp->link, &vcpu->kvm->active_mmu_pages);
 576         ASSERT(is_empty_shadow_page(sp->spt));
 577         sp->slot_bitmap = 0;
 578         sp->multimapped = 0;
 579         sp->parent_pte = parent_pte;
 580         --vcpu->kvm->n_free_mmu_pages;
 581         return sp;
 582 }
 583
 584 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 585                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 586 {
 587         struct kvm_pte_chain *pte_chain;
 588         struct hlist_node *node;
 589         int i;
 590
 591         if (!parent_pte)
 592                 return;
 593         if (!sp->multimapped) {
 594                 u64 *old = sp->parent_pte;
 595
 596                 if (!old) {
 597                         sp->parent_pte = parent_pte;
 598                         return;
 599                 }
 600                 sp->multimapped = 1;
 601                 pte_chain = mmu_alloc_pte_chain(vcpu);
 602                 INIT_HLIST_HEAD(&sp->parent_ptes);
 603                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 604                 pte_chain->parent_ptes[0] = old;
 605         }
 606         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 607                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 608                         continue;
 609                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 610                         if (!pte_chain->parent_ptes[i]) {
 611                                 pte_chain->parent_ptes[i] = parent_pte;
 612                                 return;
 613                         }
 614         }
 615         pte_chain = mmu_alloc_pte_chain(vcpu);
 616         BUG_ON(!pte_chain);
 617         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 618         pte_chain->parent_ptes[0] = parent_pte;
 619 }
 620
 621 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 622                                        u64 *parent_pte)
 623 {
 624         struct kvm_pte_chain *pte_chain;
 625         struct hlist_node *node;
 626         int i;
 627
 628         if (!sp->multimapped) {
 629                 BUG_ON(sp->parent_pte != parent_pte);
 630                 sp->parent_pte = NULL;
 631                 return;
 632         }
 633         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 634                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 635                         if (!pte_chain->parent_ptes[i])
 636                                 break;
 637                         if (pte_chain->parent_ptes[i] != parent_pte)
 638                                 continue;
 639                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 640                                 && pte_chain->parent_ptes[i + 1]) {
 641                                 pte_chain->parent_ptes[i]
 642                                         = pte_chain->parent_ptes[i + 1];
 643                                 ++i;
 644                         }
 645                         pte_chain->parent_ptes[i] = NULL;
 646                         if (i == 0) {
 647                                 hlist_del(&pte_chain->link);
 648                                 mmu_free_pte_chain(pte_chain);
 649                                 if (hlist_empty(&sp->parent_ptes)) {
 650                                         sp->multimapped = 0;
 651                                         sp->parent_pte = NULL;
 652                                 }
 653                         }
 654                         return;
 655                 }
 656         BUG();
 657 }
 658
 659 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
 660 {
 661         unsigned index;
 662         struct hlist_head *bucket;
 663         struct kvm_mmu_page *sp;
 664         struct hlist_node *node;
 665
 666         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 667         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 668         bucket = &kvm->mmu_page_hash[index];
 669         hlist_for_each_entry(sp, node, bucket, hash_link)
 670                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 671                         pgprintk("%s: found role %x\n",
 672                                  __FUNCTION__, sp->role.word);
 673                         return sp;
 674                 }
 675         return NULL;
 676 }
 677
 678 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 679                                              gfn_t gfn,
 680                                              gva_t gaddr,
 681                                              unsigned level,
 682                                              int metaphysical,
 683                                              unsigned access,
 684                                              u64 *parent_pte)
 685 {
 686         union kvm_mmu_page_role role;
 687         unsigned index;
 688         unsigned quadrant;
 689         struct hlist_head *bucket;
 690         struct kvm_mmu_page *sp;
 691         struct hlist_node *node;
 692
 693         role.word = 0;
 694         role.glevels = vcpu->mmu.root_level;
 695         role.level = level;
 696         role.metaphysical = metaphysical;
 697         role.access = access;
 698         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 699                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 700                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 701                 role.quadrant = quadrant;
 702         }
 703         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 704                  gfn, role.word);
 705         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 706         bucket = &vcpu->kvm->mmu_page_hash[index];
 707         hlist_for_each_entry(sp, node, bucket, hash_link)
 708                 if (sp->gfn == gfn && sp->role.word == role.word) {
 709                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
 710                         pgprintk("%s: found\n", __FUNCTION__);
 711                         return sp;
 712                 }
 713         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
 714         if (!sp)
 715                 return sp;
 716         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 717         sp->gfn = gfn;
 718         sp->role = role;
 719         hlist_add_head(&sp->hash_link, bucket);
 720         vcpu->mmu.prefetch_page(vcpu, sp);
 721         if (!metaphysical)
 722                 rmap_write_protect(vcpu->kvm, gfn);
 723         return sp;
 724 }
 725
 726 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 727                                          struct kvm_mmu_page *sp)
 728 {
 729         unsigned i;
 730         u64 *pt;
 731         u64 ent;
 732
 733         pt = sp->spt;
 734
 735         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
 736                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 737                         if (is_shadow_present_pte(pt[i]))
 738                                 rmap_remove(kvm, &pt[i]);
 739                         pt[i] = shadow_trap_nonpresent_pte;
 740                 }
 741                 kvm_flush_remote_tlbs(kvm);
 742                 return;
 743         }
 744
 745         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 746                 ent = pt[i];
 747
 748                 pt[i] = shadow_trap_nonpresent_pte;
 749                 if (!is_shadow_present_pte(ent))
 750                         continue;
 751                 ent &= PT64_BASE_ADDR_MASK;
 752                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 753         }
 754         kvm_flush_remote_tlbs(kvm);
 755 }
 756
 757 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
 758 {
 759         mmu_page_remove_parent_pte(sp, parent_pte);
 760 }
 761
 762 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 763 {
 764         int i;
 765
 766         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 767                 if (kvm->vcpus[i])
 768                         kvm->vcpus[i]->last_pte_updated = NULL;
 769 }
 770
 771 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 772 {
 773         u64 *parent_pte;
 774
 775         ++kvm->stat.mmu_shadow_zapped;
 776         while (sp->multimapped || sp->parent_pte) {
 777                 if (!sp->multimapped)
 778                         parent_pte = sp->parent_pte;
 779                 else {
 780                         struct kvm_pte_chain *chain;
 781
 782                         chain = container_of(sp->parent_ptes.first,
 783                                              struct kvm_pte_chain, link);
 784                         parent_pte = chain->parent_ptes[0];
 785                 }
 786                 BUG_ON(!parent_pte);
 787                 kvm_mmu_put_page(sp, parent_pte);
 788                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 789         }
 790         kvm_mmu_page_unlink_children(kvm, sp);
 791         if (!sp->root_count) {
 792                 hlist_del(&sp->hash_link);
 793                 kvm_mmu_free_page(kvm, sp);
 794         } else
 795                 list_move(&sp->link, &kvm->active_mmu_pages);
 796         kvm_mmu_reset_last_pte_updated(kvm);
 797 }
 798
 799 /*
 800  * Changing the number of mmu pages allocated to the vm
 801  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 802  */
 803 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 804 {
 805         /*
 806          * If we set the number of mmu pages to be smaller be than the
 807          * number of actived pages , we must to free some mmu pages before we
 808          * change the value
 809          */
 810
 811         if ((kvm->n_alloc_mmu_pages - kvm->n_free_mmu_pages) >
 812             kvm_nr_mmu_pages) {
 813                 int n_used_mmu_pages = kvm->n_alloc_mmu_pages
 814                                        - kvm->n_free_mmu_pages;
 815
 816                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 817                         struct kvm_mmu_page *page;
 818
 819                         page = container_of(kvm->active_mmu_pages.prev,
 820                                             struct kvm_mmu_page, link);
 821                         kvm_mmu_zap_page(kvm, page);
 822                         n_used_mmu_pages--;
 823                 }
 824                 kvm->n_free_mmu_pages = 0;
 825         }
 826         else
 827                 kvm->n_free_mmu_pages += kvm_nr_mmu_pages
 828                                          - kvm->n_alloc_mmu_pages;
 829
 830         kvm->n_alloc_mmu_pages = kvm_nr_mmu_pages;
 831 }
 832
 833 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 834 {
 835         unsigned index;
 836         struct hlist_head *bucket;
 837         struct kvm_mmu_page *sp;
 838         struct hlist_node *node, *n;
 839         int r;
 840
 841         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 842         r = 0;
 843         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 844         bucket = &kvm->mmu_page_hash[index];
 845         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
 846                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 847                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 848                                  sp->role.word);
 849                         kvm_mmu_zap_page(kvm, sp);
 850                         r = 1;
 851                 }
 852         return r;
 853 }
 854
 855 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
 856 {
 857         struct kvm_mmu_page *sp;
 858
 859         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
 860                 pgprintk("%s: zap %lx %x\n", __FUNCTION__, gfn, sp->role.word);
 861                 kvm_mmu_zap_page(kvm, sp);
 862         }
 863 }
 864
 865 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
 866 {
 867         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
 868         struct kvm_mmu_page *sp = page_header(__pa(pte));
 869
 870         __set_bit(slot, &sp->slot_bitmap);
 871 }
 872
 873 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 874 {
 875         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 876
 877         if (gpa == UNMAPPED_GVA)
 878                 return NULL;
 879         return gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 880 }
 881
 882 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
 883                          unsigned pt_access, unsigned pte_access,
 884                          int user_fault, int write_fault, int dirty,
 885                          int *ptwrite, gfn_t gfn)
 886 {
 887         u64 spte;
 888         int was_rmapped = is_rmap_pte(*shadow_pte);
 889         struct page *page;
 890
 891         pgprintk("%s: spte %llx gpte %llx access %x write_fault %d"
 892                  " user_fault %d gfn %lx\n",
 893                  __FUNCTION__, *shadow_pte, (u64)gpte, pt_access,
 894                  write_fault, user_fault, gfn);
 895
 896         /*
 897          * We don't set the accessed bit, since we sometimes want to see
 898          * whether the guest actually used the pte (in order to detect
 899          * demand paging).
 900          */
 901         spte = PT_PRESENT_MASK | PT_DIRTY_MASK;
 902         if (!dirty)
 903                 pte_access &= ~ACC_WRITE_MASK;
 904         if (!(pte_access & ACC_EXEC_MASK))
 905                 spte |= PT64_NX_MASK;
 906
 907         page = gfn_to_page(vcpu->kvm, gfn);
 908
 909         spte |= PT_PRESENT_MASK;
 910         if (pte_access & ACC_USER_MASK)
 911                 spte |= PT_USER_MASK;
 912
 913         if (is_error_page(page)) {
 914                 set_shadow_pte(shadow_pte,
 915                                shadow_trap_nonpresent_pte | PT_SHADOW_IO_MARK);
 916                 kvm_release_page_clean(page);
 917                 return;
 918         }
 919
 920         spte |= page_to_phys(page);
 921
 922         if ((pte_access & ACC_WRITE_MASK)
 923             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
 924                 struct kvm_mmu_page *shadow;
 925
 926                 spte |= PT_WRITABLE_MASK;
 927                 if (user_fault) {
 928                         mmu_unshadow(vcpu->kvm, gfn);
 929                         goto unshadowed;
 930                 }
 931
 932                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
 933                 if (shadow) {
 934                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 935                                  __FUNCTION__, gfn);
 936                         pte_access &= ~ACC_WRITE_MASK;
 937                         if (is_writeble_pte(spte)) {
 938                                 spte &= ~PT_WRITABLE_MASK;
 939                                 kvm_x86_ops->tlb_flush(vcpu);
 940                         }
 941                         if (write_fault)
 942                                 *ptwrite = 1;
 943                 }
 944         }
 945
 946 unshadowed:
 947
 948         if (pte_access & ACC_WRITE_MASK)
 949                 mark_page_dirty(vcpu->kvm, gfn);
 950
 951         pgprintk("%s: setting spte %llx\n", __FUNCTION__, spte);
 952         set_shadow_pte(shadow_pte, spte);
 953         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
 954         if (!was_rmapped) {
 955                 rmap_add(vcpu, shadow_pte, gfn);
 956                 if (!is_rmap_pte(*shadow_pte))
 957                         kvm_release_page_clean(page);
 958         }
 959         else
 960                 kvm_release_page_clean(page);
 961         if (!ptwrite || !*ptwrite)
 962                 vcpu->last_pte_updated = shadow_pte;
 963 }
 964
 965 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 966 {
 967 }
 968
 969 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, struct page *page)
 970 {
 971         int level = PT32E_ROOT_LEVEL;
 972         hpa_t table_addr = vcpu->mmu.root_hpa;
 973
 974         for (; ; level--) {
 975                 u32 index = PT64_INDEX(v, level);
 976                 u64 *table;
 977                 u64 pte;
 978
 979                 ASSERT(VALID_PAGE(table_addr));
 980                 table = __va(table_addr);
 981
 982                 if (level == 1) {
 983                         int was_rmapped;
 984
 985                         pte = table[index];
 986                         was_rmapped = is_rmap_pte(pte);
 987                         if (is_shadow_present_pte(pte) && is_writeble_pte(pte)) {
 988                                 kvm_release_page_clean(page);
 989                                 return 0;
 990                         }
 991                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 992                         page_header_update_slot(vcpu->kvm, table,
 993                                                 v >> PAGE_SHIFT);
 994                         table[index] = page_to_phys(page)
 995                                 | PT_PRESENT_MASK | PT_WRITABLE_MASK
 996                                 | PT_USER_MASK;
 997                         if (!was_rmapped)
 998                                 rmap_add(vcpu, &table[index], v >> PAGE_SHIFT);
 999                         else
1000                                 kvm_release_page_clean(page);
1001
1002                         return 0;
1003                 }
1004
1005                 if (table[index] == shadow_trap_nonpresent_pte) {
1006                         struct kvm_mmu_page *new_table;
1007                         gfn_t pseudo_gfn;
1008
1009                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1010                                 >> PAGE_SHIFT;
1011                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1012                                                      v, level - 1,
1013                                                      1, ACC_ALL, &table[index]);
1014                         if (!new_table) {
1015                                 pgprintk("nonpaging_map: ENOMEM\n");
1016                                 kvm_release_page_clean(page);
1017                                 return -ENOMEM;
1018                         }
1019
1020                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
1021                                 | PT_WRITABLE_MASK | PT_USER_MASK;
1022                 }
1023                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1024         }
1025 }
1026
1027 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1028                                     struct kvm_mmu_page *sp)
1029 {
1030         int i;
1031
1032         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1033                 sp->spt[i] = shadow_trap_nonpresent_pte;
1034 }
1035
1036 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1037 {
1038         int i;
1039         struct kvm_mmu_page *sp;
1040
1041         if (!VALID_PAGE(vcpu->mmu.root_hpa))
1042                 return;
1043 #ifdef CONFIG_X86_64
1044         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1045                 hpa_t root = vcpu->mmu.root_hpa;
1046
1047                 sp = page_header(root);
1048                 --sp->root_count;
1049                 vcpu->mmu.root_hpa = INVALID_PAGE;
1050                 return;
1051         }
1052 #endif
1053         for (i = 0; i < 4; ++i) {
1054                 hpa_t root = vcpu->mmu.pae_root[i];
1055
1056                 if (root) {
1057                         root &= PT64_BASE_ADDR_MASK;
1058                         sp = page_header(root);
1059                         --sp->root_count;
1060                 }
1061                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1062         }
1063         vcpu->mmu.root_hpa = INVALID_PAGE;
1064 }
1065
1066 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1067 {
1068         int i;
1069         gfn_t root_gfn;
1070         struct kvm_mmu_page *sp;
1071
1072         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
1073
1074 #ifdef CONFIG_X86_64
1075         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1076                 hpa_t root = vcpu->mmu.root_hpa;
1077
1078                 ASSERT(!VALID_PAGE(root));
1079                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1080                                       PT64_ROOT_LEVEL, 0, ACC_ALL, NULL);
1081                 root = __pa(sp->spt);
1082                 ++sp->root_count;
1083                 vcpu->mmu.root_hpa = root;
1084                 return;
1085         }
1086 #endif
1087         for (i = 0; i < 4; ++i) {
1088                 hpa_t root = vcpu->mmu.pae_root[i];
1089
1090                 ASSERT(!VALID_PAGE(root));
1091                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
1092                         if (!is_present_pte(vcpu->pdptrs[i])) {
1093                                 vcpu->mmu.pae_root[i] = 0;
1094                                 continue;
1095                         }
1096                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
1097                 } else if (vcpu->mmu.root_level == 0)
1098                         root_gfn = 0;
1099                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1100                                       PT32_ROOT_LEVEL, !is_paging(vcpu),
1101                                       ACC_ALL, NULL);
1102                 root = __pa(sp->spt);
1103                 ++sp->root_count;
1104                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
1105         }
1106         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
1107 }
1108
1109 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1110 {
1111         return vaddr;
1112 }
1113
1114 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1115                                 u32 error_code)
1116 {
1117         struct page *page;
1118         int r;
1119
1120         r = mmu_topup_memory_caches(vcpu);
1121         if (r)
1122                 return r;
1123
1124         ASSERT(vcpu);
1125         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
1126
1127         page = gfn_to_page(vcpu->kvm, gva >> PAGE_SHIFT);
1128
1129         if (is_error_page(page)) {
1130                 kvm_release_page_clean(page);
1131                 return 1;
1132         }
1133
1134         return nonpaging_map(vcpu, gva & PAGE_MASK, page);
1135 }
1136
1137 static void nonpaging_free(struct kvm_vcpu *vcpu)
1138 {
1139         mmu_free_roots(vcpu);
1140 }
1141
1142 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1143 {
1144         struct kvm_mmu *context = &vcpu->mmu;
1145
1146         context->new_cr3 = nonpaging_new_cr3;
1147         context->page_fault = nonpaging_page_fault;
1148         context->gva_to_gpa = nonpaging_gva_to_gpa;
1149         context->free = nonpaging_free;
1150         context->prefetch_page = nonpaging_prefetch_page;
1151         context->root_level = 0;
1152         context->shadow_root_level = PT32E_ROOT_LEVEL;
1153         context->root_hpa = INVALID_PAGE;
1154         return 0;
1155 }
1156
1157 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1158 {
1159         ++vcpu->stat.tlb_flush;
1160         kvm_x86_ops->tlb_flush(vcpu);
1161 }
1162
1163 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1164 {
1165         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
1166         mmu_free_roots(vcpu);
1167 }
1168
1169 static void inject_page_fault(struct kvm_vcpu *vcpu,
1170                               u64 addr,
1171                               u32 err_code)
1172 {
1173         kvm_inject_page_fault(vcpu, addr, err_code);
1174 }
1175
1176 static void paging_free(struct kvm_vcpu *vcpu)
1177 {
1178         nonpaging_free(vcpu);
1179 }
1180
1181 #define PTTYPE 64
1182 #include "paging_tmpl.h"
1183 #undef PTTYPE
1184
1185 #define PTTYPE 32
1186 #include "paging_tmpl.h"
1187 #undef PTTYPE
1188
1189 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1190 {
1191         struct kvm_mmu *context = &vcpu->mmu;
1192
1193         ASSERT(is_pae(vcpu));
1194         context->new_cr3 = paging_new_cr3;
1195         context->page_fault = paging64_page_fault;
1196         context->gva_to_gpa = paging64_gva_to_gpa;
1197         context->prefetch_page = paging64_prefetch_page;
1198         context->free = paging_free;
1199         context->root_level = level;
1200         context->shadow_root_level = level;
1201         context->root_hpa = INVALID_PAGE;
1202         return 0;
1203 }
1204
1205 static int paging64_init_context(struct kvm_vcpu *vcpu)
1206 {
1207         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1208 }
1209
1210 static int paging32_init_context(struct kvm_vcpu *vcpu)
1211 {
1212         struct kvm_mmu *context = &vcpu->mmu;
1213
1214         context->new_cr3 = paging_new_cr3;
1215         context->page_fault = paging32_page_fault;
1216         context->gva_to_gpa = paging32_gva_to_gpa;
1217         context->free = paging_free;
1218         context->prefetch_page = paging32_prefetch_page;
1219         context->root_level = PT32_ROOT_LEVEL;
1220         context->shadow_root_level = PT32E_ROOT_LEVEL;
1221         context->root_hpa = INVALID_PAGE;
1222         return 0;
1223 }
1224
1225 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1226 {
1227         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1228 }
1229
1230 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1231 {
1232         ASSERT(vcpu);
1233         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1234
1235         if (!is_paging(vcpu))
1236                 return nonpaging_init_context(vcpu);
1237         else if (is_long_mode(vcpu))
1238                 return paging64_init_context(vcpu);
1239         else if (is_pae(vcpu))
1240                 return paging32E_init_context(vcpu);
1241         else
1242                 return paging32_init_context(vcpu);
1243 }
1244
1245 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1246 {
1247         ASSERT(vcpu);
1248         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1249                 vcpu->mmu.free(vcpu);
1250                 vcpu->mmu.root_hpa = INVALID_PAGE;
1251         }
1252 }
1253
1254 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1255 {
1256         destroy_kvm_mmu(vcpu);
1257         return init_kvm_mmu(vcpu);
1258 }
1259 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1260
1261 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1262 {
1263         int r;
1264
1265         mutex_lock(&vcpu->kvm->lock);
1266         r = mmu_topup_memory_caches(vcpu);
1267         if (r)
1268                 goto out;
1269         mmu_alloc_roots(vcpu);
1270         kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1271         kvm_mmu_flush_tlb(vcpu);
1272 out:
1273         mutex_unlock(&vcpu->kvm->lock);
1274         return r;
1275 }
1276 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1277
1278 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1279 {
1280         mmu_free_roots(vcpu);
1281 }
1282
1283 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1284                                   struct kvm_mmu_page *sp,
1285                                   u64 *spte)
1286 {
1287         u64 pte;
1288         struct kvm_mmu_page *child;
1289
1290         pte = *spte;
1291         if (is_shadow_present_pte(pte)) {
1292                 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1293                         rmap_remove(vcpu->kvm, spte);
1294                 else {
1295                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1296                         mmu_page_remove_parent_pte(child, spte);
1297                 }
1298         }
1299         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1300 }
1301
1302 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1303                                   struct kvm_mmu_page *sp,
1304                                   u64 *spte,
1305                                   const void *new, int bytes,
1306                                   int offset_in_pte)
1307 {
1308         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1309                 ++vcpu->kvm->stat.mmu_pde_zapped;
1310                 return;
1311         }
1312
1313         ++vcpu->kvm->stat.mmu_pte_updated;
1314         if (sp->role.glevels == PT32_ROOT_LEVEL)
1315                 paging32_update_pte(vcpu, sp, spte, new, bytes, offset_in_pte);
1316         else
1317                 paging64_update_pte(vcpu, sp, spte, new, bytes, offset_in_pte);
1318 }
1319
1320 static bool need_remote_flush(u64 old, u64 new)
1321 {
1322         if (!is_shadow_present_pte(old))
1323                 return false;
1324         if (!is_shadow_present_pte(new))
1325                 return true;
1326         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1327                 return true;
1328         old ^= PT64_NX_MASK;
1329         new ^= PT64_NX_MASK;
1330         return (old & ~new & PT64_PERM_MASK) != 0;
1331 }
1332
1333 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1334 {
1335         if (need_remote_flush(old, new))
1336                 kvm_flush_remote_tlbs(vcpu->kvm);
1337         else
1338                 kvm_mmu_flush_tlb(vcpu);
1339 }
1340
1341 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1342 {
1343         u64 *spte = vcpu->last_pte_updated;
1344
1345         return !!(spte && (*spte & PT_ACCESSED_MASK));
1346 }
1347
1348 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1349                        const u8 *new, int bytes)
1350 {
1351         gfn_t gfn = gpa >> PAGE_SHIFT;
1352         struct kvm_mmu_page *sp;
1353         struct hlist_node *node, *n;
1354         struct hlist_head *bucket;
1355         unsigned index;
1356         u64 entry;
1357         u64 *spte;
1358         unsigned offset = offset_in_page(gpa);
1359         unsigned pte_size;
1360         unsigned page_offset;
1361         unsigned misaligned;
1362         unsigned quadrant;
1363         int level;
1364         int flooded = 0;
1365         int npte;
1366
1367         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1368         ++vcpu->kvm->stat.mmu_pte_write;
1369         kvm_mmu_audit(vcpu, "pre pte write");
1370         if (gfn == vcpu->last_pt_write_gfn
1371             && !last_updated_pte_accessed(vcpu)) {
1372                 ++vcpu->last_pt_write_count;
1373                 if (vcpu->last_pt_write_count >= 3)
1374                         flooded = 1;
1375         } else {
1376                 vcpu->last_pt_write_gfn = gfn;
1377                 vcpu->last_pt_write_count = 1;
1378                 vcpu->last_pte_updated = NULL;
1379         }
1380         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1381         bucket = &vcpu->kvm->mmu_page_hash[index];
1382         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1383                 if (sp->gfn != gfn || sp->role.metaphysical)
1384                         continue;
1385                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1386                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1387                 misaligned |= bytes < 4;
1388                 if (misaligned || flooded) {
1389                         /*
1390                          * Misaligned accesses are too much trouble to fix
1391                          * up; also, they usually indicate a page is not used
1392                          * as a page table.
1393                          *
1394                          * If we're seeing too many writes to a page,
1395                          * it may no longer be a page table, or we may be
1396                          * forking, in which case it is better to unmap the
1397                          * page.
1398                          */
1399                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1400                                  gpa, bytes, sp->role.word);
1401                         kvm_mmu_zap_page(vcpu->kvm, sp);
1402                         ++vcpu->kvm->stat.mmu_flooded;
1403                         continue;
1404                 }
1405                 page_offset = offset;
1406                 level = sp->role.level;
1407                 npte = 1;
1408                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1409                         page_offset <<= 1;      /* 32->64 */
1410                         /*
1411                          * A 32-bit pde maps 4MB while the shadow pdes map
1412                          * only 2MB.  So we need to double the offset again
1413                          * and zap two pdes instead of one.
1414                          */
1415                         if (level == PT32_ROOT_LEVEL) {
1416                                 page_offset &= ~7; /* kill rounding error */
1417                                 page_offset <<= 1;
1418                                 npte = 2;
1419                         }
1420                         quadrant = page_offset >> PAGE_SHIFT;
1421                         page_offset &= ~PAGE_MASK;
1422                         if (quadrant != sp->role.quadrant)
1423                                 continue;
1424                 }
1425                 spte = &sp->spt[page_offset / sizeof(*spte)];
1426                 while (npte--) {
1427                         entry = *spte;
1428                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1429                         mmu_pte_write_new_pte(vcpu, sp, spte, new, bytes,
1430                                               page_offset & (pte_size - 1));
1431                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1432                         ++spte;
1433                 }
1434         }
1435         kvm_mmu_audit(vcpu, "post pte write");
1436 }
1437
1438 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1439 {
1440         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1441
1442         return kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1443 }
1444
1445 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1446 {
1447         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1448                 struct kvm_mmu_page *sp;
1449
1450                 sp = container_of(vcpu->kvm->active_mmu_pages.prev,
1451                                   struct kvm_mmu_page, link);
1452                 kvm_mmu_zap_page(vcpu->kvm, sp);
1453                 ++vcpu->kvm->stat.mmu_recycled;
1454         }
1455 }
1456
1457 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1458 {
1459         int r;
1460         enum emulation_result er;
1461
1462         mutex_lock(&vcpu->kvm->lock);
1463         r = vcpu->mmu.page_fault(vcpu, cr2, error_code);
1464         if (r < 0)
1465                 goto out;
1466
1467         if (!r) {
1468                 r = 1;
1469                 goto out;
1470         }
1471
1472         r = mmu_topup_memory_caches(vcpu);
1473         if (r)
1474                 goto out;
1475
1476         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1477         mutex_unlock(&vcpu->kvm->lock);
1478
1479         switch (er) {
1480         case EMULATE_DONE:
1481                 return 1;
1482         case EMULATE_DO_MMIO:
1483                 ++vcpu->stat.mmio_exits;
1484                 return 0;
1485         case EMULATE_FAIL:
1486                 kvm_report_emulation_failure(vcpu, "pagetable");
1487                 return 1;
1488         default:
1489                 BUG();
1490         }
1491 out:
1492         mutex_unlock(&vcpu->kvm->lock);
1493         return r;
1494 }
1495 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1496
1497 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1498 {
1499         struct kvm_mmu_page *sp;
1500
1501         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1502                 sp = container_of(vcpu->kvm->active_mmu_pages.next,
1503                                   struct kvm_mmu_page, link);
1504                 kvm_mmu_zap_page(vcpu->kvm, sp);
1505         }
1506         free_page((unsigned long)vcpu->mmu.pae_root);
1507 }
1508
1509 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1510 {
1511         struct page *page;
1512         int i;
1513
1514         ASSERT(vcpu);
1515
1516         if (vcpu->kvm->n_requested_mmu_pages)
1517                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_requested_mmu_pages;
1518         else
1519                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_alloc_mmu_pages;
1520         /*
1521          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1522          * Therefore we need to allocate shadow page tables in the first
1523          * 4GB of memory, which happens to fit the DMA32 zone.
1524          */
1525         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1526         if (!page)
1527                 goto error_1;
1528         vcpu->mmu.pae_root = page_address(page);
1529         for (i = 0; i < 4; ++i)
1530                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1531
1532         return 0;
1533
1534 error_1:
1535         free_mmu_pages(vcpu);
1536         return -ENOMEM;
1537 }
1538
1539 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1540 {
1541         ASSERT(vcpu);
1542         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1543
1544         return alloc_mmu_pages(vcpu);
1545 }
1546
1547 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1548 {
1549         ASSERT(vcpu);
1550         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1551
1552         return init_kvm_mmu(vcpu);
1553 }
1554
1555 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1556 {
1557         ASSERT(vcpu);
1558
1559         destroy_kvm_mmu(vcpu);
1560         free_mmu_pages(vcpu);
1561         mmu_free_memory_caches(vcpu);
1562 }
1563
1564 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1565 {
1566         struct kvm_mmu_page *sp;
1567
1568         list_for_each_entry(sp, &kvm->active_mmu_pages, link) {
1569                 int i;
1570                 u64 *pt;
1571
1572                 if (!test_bit(slot, &sp->slot_bitmap))
1573                         continue;
1574
1575                 pt = sp->spt;
1576                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1577                         /* avoid RMW */
1578                         if (pt[i] & PT_WRITABLE_MASK)
1579                                 pt[i] &= ~PT_WRITABLE_MASK;
1580         }
1581 }
1582
1583 void kvm_mmu_zap_all(struct kvm *kvm)
1584 {
1585         struct kvm_mmu_page *sp, *node;
1586
1587         list_for_each_entry_safe(sp, node, &kvm->active_mmu_pages, link)
1588                 kvm_mmu_zap_page(kvm, sp);
1589
1590         kvm_flush_remote_tlbs(kvm);
1591 }
1592
1593 void kvm_mmu_module_exit(void)
1594 {
1595         if (pte_chain_cache)
1596                 kmem_cache_destroy(pte_chain_cache);
1597         if (rmap_desc_cache)
1598                 kmem_cache_destroy(rmap_desc_cache);
1599         if (mmu_page_header_cache)
1600                 kmem_cache_destroy(mmu_page_header_cache);
1601 }
1602
1603 int kvm_mmu_module_init(void)
1604 {
1605         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1606                                             sizeof(struct kvm_pte_chain),
1607                                             0, 0, NULL);
1608         if (!pte_chain_cache)
1609                 goto nomem;
1610         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1611                                             sizeof(struct kvm_rmap_desc),
1612                                             0, 0, NULL);
1613         if (!rmap_desc_cache)
1614                 goto nomem;
1615
1616         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1617                                                   sizeof(struct kvm_mmu_page),
1618                                                   0, 0, NULL);
1619         if (!mmu_page_header_cache)
1620                 goto nomem;
1621
1622         return 0;
1623
1624 nomem:
1625         kvm_mmu_module_exit();
1626         return -ENOMEM;
1627 }
1628
1629 /*
1630  * Caculate mmu pages needed for kvm.
1631  */
1632 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
1633 {
1634         int i;
1635         unsigned int nr_mmu_pages;
1636         unsigned int  nr_pages = 0;
1637
1638         for (i = 0; i < kvm->nmemslots; i++)
1639                 nr_pages += kvm->memslots[i].npages;
1640
1641         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
1642         nr_mmu_pages = max(nr_mmu_pages,
1643                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
1644
1645         return nr_mmu_pages;
1646 }
1647
1648 #ifdef AUDIT
1649
1650 static const char *audit_msg;
1651
1652 static gva_t canonicalize(gva_t gva)
1653 {
1654 #ifdef CONFIG_X86_64
1655         gva = (long long)(gva << 16) >> 16;
1656 #endif
1657         return gva;
1658 }
1659
1660 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1661                                 gva_t va, int level)
1662 {
1663         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1664         int i;
1665         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1666
1667         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1668                 u64 ent = pt[i];
1669
1670                 if (ent == shadow_trap_nonpresent_pte)
1671                         continue;
1672
1673                 va = canonicalize(va);
1674                 if (level > 1) {
1675                         if (ent == shadow_notrap_nonpresent_pte)
1676                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1677                                        " in nonleaf level: levels %d gva %lx"
1678                                        " level %d pte %llx\n", audit_msg,
1679                                        vcpu->mmu.root_level, va, level, ent);
1680
1681                         audit_mappings_page(vcpu, ent, va, level - 1);
1682                 } else {
1683                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1684                         struct page *page = gpa_to_page(vcpu, gpa);
1685                         hpa_t hpa = page_to_phys(page);
1686
1687                         if (is_shadow_present_pte(ent)
1688                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1689                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1690                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1691                                        audit_msg, vcpu->mmu.root_level,
1692                                        va, gpa, hpa, ent,
1693                                        is_shadow_present_pte(ent));
1694                         else if (ent == shadow_notrap_nonpresent_pte
1695                                  && !is_error_hpa(hpa))
1696                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1697                                        " valid guest gva %lx\n", audit_msg, va);
1698                         kvm_release_page_clean(page);
1699
1700                 }
1701         }
1702 }
1703
1704 static void audit_mappings(struct kvm_vcpu *vcpu)
1705 {
1706         unsigned i;
1707
1708         if (vcpu->mmu.root_level == 4)
1709                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1710         else
1711                 for (i = 0; i < 4; ++i)
1712                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1713                                 audit_mappings_page(vcpu,
1714                                                     vcpu->mmu.pae_root[i],
1715                                                     i << 30,
1716                                                     2);
1717 }
1718
1719 static int count_rmaps(struct kvm_vcpu *vcpu)
1720 {
1721         int nmaps = 0;
1722         int i, j, k;
1723
1724         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1725                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1726                 struct kvm_rmap_desc *d;
1727
1728                 for (j = 0; j < m->npages; ++j) {
1729                         unsigned long *rmapp = &m->rmap[j];
1730
1731                         if (!*rmapp)
1732                                 continue;
1733                         if (!(*rmapp & 1)) {
1734                                 ++nmaps;
1735                                 continue;
1736                         }
1737                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1738                         while (d) {
1739                                 for (k = 0; k < RMAP_EXT; ++k)
1740                                         if (d->shadow_ptes[k])
1741                                                 ++nmaps;
1742                                         else
1743                                                 break;
1744                                 d = d->more;
1745                         }
1746                 }
1747         }
1748         return nmaps;
1749 }
1750
1751 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1752 {
1753         int nmaps = 0;
1754         struct kvm_mmu_page *sp;
1755         int i;
1756
1757         list_for_each_entry(sp, &vcpu->kvm->active_mmu_pages, link) {
1758                 u64 *pt = sp->spt;
1759
1760                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
1761                         continue;
1762
1763                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1764                         u64 ent = pt[i];
1765
1766                         if (!(ent & PT_PRESENT_MASK))
1767                                 continue;
1768                         if (!(ent & PT_WRITABLE_MASK))
1769                                 continue;
1770                         ++nmaps;
1771                 }
1772         }
1773         return nmaps;
1774 }
1775
1776 static void audit_rmap(struct kvm_vcpu *vcpu)
1777 {
1778         int n_rmap = count_rmaps(vcpu);
1779         int n_actual = count_writable_mappings(vcpu);
1780
1781         if (n_rmap != n_actual)
1782                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1783                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1784 }
1785
1786 static void audit_write_protection(struct kvm_vcpu *vcpu)
1787 {
1788         struct kvm_mmu_page *sp;
1789         struct kvm_memory_slot *slot;
1790         unsigned long *rmapp;
1791         gfn_t gfn;
1792
1793         list_for_each_entry(sp, &vcpu->kvm->active_mmu_pages, link) {
1794                 if (sp->role.metaphysical)
1795                         continue;
1796
1797                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
1798                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
1799                 rmapp = &slot->rmap[gfn - slot->base_gfn];
1800                 if (*rmapp)
1801                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1802                                " mappings: gfn %lx role %x\n",
1803                                __FUNCTION__, audit_msg, sp->gfn,
1804                                sp->role.word);
1805         }
1806 }
1807
1808 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1809 {
1810         int olddbg = dbg;
1811
1812         dbg = 0;
1813         audit_msg = msg;
1814         audit_rmap(vcpu);
1815         audit_write_protection(vcpu);
1816         audit_mappings(vcpu);
1817         dbg = olddbg;
1818 }
1819
1820 #endif