SAFE public projects git trees. - safe/jmp/linux-2.6/blob - arch/x86/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 "mmu.h"
  21
  22 #include <linux/kvm_host.h>
  23 #include <linux/types.h>
  24 #include <linux/string.h>
  25 #include <linux/mm.h>
  26 #include <linux/highmem.h>
  27 #include <linux/module.h>
  28 #include <linux/swap.h>
  29 #include <linux/hugetlb.h>
  30 #include <linux/compiler.h>
  31
  32 #include <asm/page.h>
  33 #include <asm/cmpxchg.h>
  34 #include <asm/io.h>
  35 #include <asm/vmx.h>
  36
  37 /*
  38  * When setting this variable to true it enables Two-Dimensional-Paging
  39  * where the hardware walks 2 page tables:
  40  * 1. the guest-virtual to guest-physical
  41  * 2. while doing 1. it walks guest-physical to host-physical
  42  * If the hardware supports that we don't need to do shadow paging.
  43  */
  44 bool tdp_enabled = false;
  45
  46 #undef MMU_DEBUG
  47
  48 #undef AUDIT
  49
  50 #ifdef AUDIT
  51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  52 #else
  53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  54 #endif
  55
  56 #ifdef MMU_DEBUG
  57
  58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  60
  61 #else
  62
  63 #define pgprintk(x...) do { } while (0)
  64 #define rmap_printk(x...) do { } while (0)
  65
  66 #endif
  67
  68 #if defined(MMU_DEBUG) || defined(AUDIT)
  69 static int dbg = 0;
  70 module_param(dbg, bool, 0644);
  71 #endif
  72
  73 static int oos_shadow = 1;
  74 module_param(oos_shadow, bool, 0644);
  75
  76 #ifndef MMU_DEBUG
  77 #define ASSERT(x) do { } while (0)
  78 #else
  79 #define ASSERT(x)                                                       \
  80         if (!(x)) {                                                     \
  81                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  82                        __FILE__, __LINE__, #x);                         \
  83         }
  84 #endif
  85
  86 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  87 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  88
  89 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  90
  91 #define PT64_LEVEL_BITS 9
  92
  93 #define PT64_LEVEL_SHIFT(level) \
  94                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  95
  96 #define PT64_LEVEL_MASK(level) \
  97                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  98
  99 #define PT64_INDEX(address, level)\
 100         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 101
 102
 103 #define PT32_LEVEL_BITS 10
 104
 105 #define PT32_LEVEL_SHIFT(level) \
 106                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 107
 108 #define PT32_LEVEL_MASK(level) \
 109                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 110
 111 #define PT32_INDEX(address, level)\
 112         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 113
 114
 115 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 116 #define PT64_DIR_BASE_ADDR_MASK \
 117         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 118
 119 #define PT32_BASE_ADDR_MASK PAGE_MASK
 120 #define PT32_DIR_BASE_ADDR_MASK \
 121         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 122
 123 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 124                         | PT64_NX_MASK)
 125
 126 #define PFERR_PRESENT_MASK (1U << 0)
 127 #define PFERR_WRITE_MASK (1U << 1)
 128 #define PFERR_USER_MASK (1U << 2)
 129 #define PFERR_FETCH_MASK (1U << 4)
 130
 131 #define PT_DIRECTORY_LEVEL 2
 132 #define PT_PAGE_TABLE_LEVEL 1
 133
 134 #define RMAP_EXT 4
 135
 136 #define ACC_EXEC_MASK    1
 137 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 138 #define ACC_USER_MASK    PT_USER_MASK
 139 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 140
 141 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 142
 143 struct kvm_rmap_desc {
 144         u64 *shadow_ptes[RMAP_EXT];
 145         struct kvm_rmap_desc *more;
 146 };
 147
 148 struct kvm_shadow_walk {
 149         int (*entry)(struct kvm_shadow_walk *walk, struct kvm_vcpu *vcpu,
 150                      u64 addr, u64 *spte, int level);
 151 };
 152
 153 struct kvm_unsync_walk {
 154         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 155 };
 156
 157 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 158
 159 static struct kmem_cache *pte_chain_cache;
 160 static struct kmem_cache *rmap_desc_cache;
 161 static struct kmem_cache *mmu_page_header_cache;
 162
 163 static u64 __read_mostly shadow_trap_nonpresent_pte;
 164 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 165 static u64 __read_mostly shadow_base_present_pte;
 166 static u64 __read_mostly shadow_nx_mask;
 167 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 168 static u64 __read_mostly shadow_user_mask;
 169 static u64 __read_mostly shadow_accessed_mask;
 170 static u64 __read_mostly shadow_dirty_mask;
 171 static u64 __read_mostly shadow_mt_mask;
 172
 173 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 174 {
 175         shadow_trap_nonpresent_pte = trap_pte;
 176         shadow_notrap_nonpresent_pte = notrap_pte;
 177 }
 178 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 179
 180 void kvm_mmu_set_base_ptes(u64 base_pte)
 181 {
 182         shadow_base_present_pte = base_pte;
 183 }
 184 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 185
 186 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 187                 u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 mt_mask)
 188 {
 189         shadow_user_mask = user_mask;
 190         shadow_accessed_mask = accessed_mask;
 191         shadow_dirty_mask = dirty_mask;
 192         shadow_nx_mask = nx_mask;
 193         shadow_x_mask = x_mask;
 194         shadow_mt_mask = mt_mask;
 195 }
 196 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 197
 198 static int is_write_protection(struct kvm_vcpu *vcpu)
 199 {
 200         return vcpu->arch.cr0 & X86_CR0_WP;
 201 }
 202
 203 static int is_cpuid_PSE36(void)
 204 {
 205         return 1;
 206 }
 207
 208 static int is_nx(struct kvm_vcpu *vcpu)
 209 {
 210         return vcpu->arch.shadow_efer & EFER_NX;
 211 }
 212
 213 static int is_present_pte(unsigned long pte)
 214 {
 215         return pte & PT_PRESENT_MASK;
 216 }
 217
 218 static int is_shadow_present_pte(u64 pte)
 219 {
 220         return pte != shadow_trap_nonpresent_pte
 221                 && pte != shadow_notrap_nonpresent_pte;
 222 }
 223
 224 static int is_large_pte(u64 pte)
 225 {
 226         return pte & PT_PAGE_SIZE_MASK;
 227 }
 228
 229 static int is_writeble_pte(unsigned long pte)
 230 {
 231         return pte & PT_WRITABLE_MASK;
 232 }
 233
 234 static int is_dirty_pte(unsigned long pte)
 235 {
 236         return pte & shadow_dirty_mask;
 237 }
 238
 239 static int is_rmap_pte(u64 pte)
 240 {
 241         return is_shadow_present_pte(pte);
 242 }
 243
 244 static pfn_t spte_to_pfn(u64 pte)
 245 {
 246         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 247 }
 248
 249 static gfn_t pse36_gfn_delta(u32 gpte)
 250 {
 251         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 252
 253         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 254 }
 255
 256 static void set_shadow_pte(u64 *sptep, u64 spte)
 257 {
 258 #ifdef CONFIG_X86_64
 259         set_64bit((unsigned long *)sptep, spte);
 260 #else
 261         set_64bit((unsigned long long *)sptep, spte);
 262 #endif
 263 }
 264
 265 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 266                                   struct kmem_cache *base_cache, int min)
 267 {
 268         void *obj;
 269
 270         if (cache->nobjs >= min)
 271                 return 0;
 272         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 273                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 274                 if (!obj)
 275                         return -ENOMEM;
 276                 cache->objects[cache->nobjs++] = obj;
 277         }
 278         return 0;
 279 }
 280
 281 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 282 {
 283         while (mc->nobjs)
 284                 kfree(mc->objects[--mc->nobjs]);
 285 }
 286
 287 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 288                                        int min)
 289 {
 290         struct page *page;
 291
 292         if (cache->nobjs >= min)
 293                 return 0;
 294         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 295                 page = alloc_page(GFP_KERNEL);
 296                 if (!page)
 297                         return -ENOMEM;
 298                 set_page_private(page, 0);
 299                 cache->objects[cache->nobjs++] = page_address(page);
 300         }
 301         return 0;
 302 }
 303
 304 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 305 {
 306         while (mc->nobjs)
 307                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 308 }
 309
 310 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 311 {
 312         int r;
 313
 314         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 315                                    pte_chain_cache, 4);
 316         if (r)
 317                 goto out;
 318         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 319                                    rmap_desc_cache, 4);
 320         if (r)
 321                 goto out;
 322         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 323         if (r)
 324                 goto out;
 325         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 326                                    mmu_page_header_cache, 4);
 327 out:
 328         return r;
 329 }
 330
 331 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 332 {
 333         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 334         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 335         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 336         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 337 }
 338
 339 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 340                                     size_t size)
 341 {
 342         void *p;
 343
 344         BUG_ON(!mc->nobjs);
 345         p = mc->objects[--mc->nobjs];
 346         memset(p, 0, size);
 347         return p;
 348 }
 349
 350 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 351 {
 352         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 353                                       sizeof(struct kvm_pte_chain));
 354 }
 355
 356 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 357 {
 358         kfree(pc);
 359 }
 360
 361 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 362 {
 363         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 364                                       sizeof(struct kvm_rmap_desc));
 365 }
 366
 367 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 368 {
 369         kfree(rd);
 370 }
 371
 372 /*
 373  * Return the pointer to the largepage write count for a given
 374  * gfn, handling slots that are not large page aligned.
 375  */
 376 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 377 {
 378         unsigned long idx;
 379
 380         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 381               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 382         return &slot->lpage_info[idx].write_count;
 383 }
 384
 385 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 386 {
 387         int *write_count;
 388
 389         gfn = unalias_gfn(kvm, gfn);
 390         write_count = slot_largepage_idx(gfn,
 391                                          gfn_to_memslot_unaliased(kvm, gfn));
 392         *write_count += 1;
 393 }
 394
 395 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 396 {
 397         int *write_count;
 398
 399         gfn = unalias_gfn(kvm, gfn);
 400         write_count = slot_largepage_idx(gfn,
 401                                          gfn_to_memslot_unaliased(kvm, gfn));
 402         *write_count -= 1;
 403         WARN_ON(*write_count < 0);
 404 }
 405
 406 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 407 {
 408         struct kvm_memory_slot *slot;
 409         int *largepage_idx;
 410
 411         gfn = unalias_gfn(kvm, gfn);
 412         slot = gfn_to_memslot_unaliased(kvm, gfn);
 413         if (slot) {
 414                 largepage_idx = slot_largepage_idx(gfn, slot);
 415                 return *largepage_idx;
 416         }
 417
 418         return 1;
 419 }
 420
 421 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 422 {
 423         struct vm_area_struct *vma;
 424         unsigned long addr;
 425         int ret = 0;
 426
 427         addr = gfn_to_hva(kvm, gfn);
 428         if (kvm_is_error_hva(addr))
 429                 return ret;
 430
 431         down_read(&current->mm->mmap_sem);
 432         vma = find_vma(current->mm, addr);
 433         if (vma && is_vm_hugetlb_page(vma))
 434                 ret = 1;
 435         up_read(&current->mm->mmap_sem);
 436
 437         return ret;
 438 }
 439
 440 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 441 {
 442         struct kvm_memory_slot *slot;
 443
 444         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 445                 return 0;
 446
 447         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 448                 return 0;
 449
 450         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 451         if (slot && slot->dirty_bitmap)
 452                 return 0;
 453
 454         return 1;
 455 }
 456
 457 /*
 458  * Take gfn and return the reverse mapping to it.
 459  * Note: gfn must be unaliased before this function get called
 460  */
 461
 462 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 463 {
 464         struct kvm_memory_slot *slot;
 465         unsigned long idx;
 466
 467         slot = gfn_to_memslot(kvm, gfn);
 468         if (!lpage)
 469                 return &slot->rmap[gfn - slot->base_gfn];
 470
 471         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 472               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 473
 474         return &slot->lpage_info[idx].rmap_pde;
 475 }
 476
 477 /*
 478  * Reverse mapping data structures:
 479  *
 480  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 481  * that points to page_address(page).
 482  *
 483  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 484  * containing more mappings.
 485  */
 486 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 487 {
 488         struct kvm_mmu_page *sp;
 489         struct kvm_rmap_desc *desc;
 490         unsigned long *rmapp;
 491         int i;
 492
 493         if (!is_rmap_pte(*spte))
 494                 return;
 495         gfn = unalias_gfn(vcpu->kvm, gfn);
 496         sp = page_header(__pa(spte));
 497         sp->gfns[spte - sp->spt] = gfn;
 498         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 499         if (!*rmapp) {
 500                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 501                 *rmapp = (unsigned long)spte;
 502         } else if (!(*rmapp & 1)) {
 503                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 504                 desc = mmu_alloc_rmap_desc(vcpu);
 505                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 506                 desc->shadow_ptes[1] = spte;
 507                 *rmapp = (unsigned long)desc | 1;
 508         } else {
 509                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 510                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 511                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 512                         desc = desc->more;
 513                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 514                         desc->more = mmu_alloc_rmap_desc(vcpu);
 515                         desc = desc->more;
 516                 }
 517                 for (i = 0; desc->shadow_ptes[i]; ++i)
 518                         ;
 519                 desc->shadow_ptes[i] = spte;
 520         }
 521 }
 522
 523 static void rmap_desc_remove_entry(unsigned long *rmapp,
 524                                    struct kvm_rmap_desc *desc,
 525                                    int i,
 526                                    struct kvm_rmap_desc *prev_desc)
 527 {
 528         int j;
 529
 530         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 531                 ;
 532         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 533         desc->shadow_ptes[j] = NULL;
 534         if (j != 0)
 535                 return;
 536         if (!prev_desc && !desc->more)
 537                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 538         else
 539                 if (prev_desc)
 540                         prev_desc->more = desc->more;
 541                 else
 542                         *rmapp = (unsigned long)desc->more | 1;
 543         mmu_free_rmap_desc(desc);
 544 }
 545
 546 static void rmap_remove(struct kvm *kvm, u64 *spte)
 547 {
 548         struct kvm_rmap_desc *desc;
 549         struct kvm_rmap_desc *prev_desc;
 550         struct kvm_mmu_page *sp;
 551         pfn_t pfn;
 552         unsigned long *rmapp;
 553         int i;
 554
 555         if (!is_rmap_pte(*spte))
 556                 return;
 557         sp = page_header(__pa(spte));
 558         pfn = spte_to_pfn(*spte);
 559         if (*spte & shadow_accessed_mask)
 560                 kvm_set_pfn_accessed(pfn);
 561         if (is_writeble_pte(*spte))
 562                 kvm_release_pfn_dirty(pfn);
 563         else
 564                 kvm_release_pfn_clean(pfn);
 565         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 566         if (!*rmapp) {
 567                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 568                 BUG();
 569         } else if (!(*rmapp & 1)) {
 570                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 571                 if ((u64 *)*rmapp != spte) {
 572                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 573                                spte, *spte);
 574                         BUG();
 575                 }
 576                 *rmapp = 0;
 577         } else {
 578                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 579                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 580                 prev_desc = NULL;
 581                 while (desc) {
 582                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 583                                 if (desc->shadow_ptes[i] == spte) {
 584                                         rmap_desc_remove_entry(rmapp,
 585                                                                desc, i,
 586                                                                prev_desc);
 587                                         return;
 588                                 }
 589                         prev_desc = desc;
 590                         desc = desc->more;
 591                 }
 592                 BUG();
 593         }
 594 }
 595
 596 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 597 {
 598         struct kvm_rmap_desc *desc;
 599         struct kvm_rmap_desc *prev_desc;
 600         u64 *prev_spte;
 601         int i;
 602
 603         if (!*rmapp)
 604                 return NULL;
 605         else if (!(*rmapp & 1)) {
 606                 if (!spte)
 607                         return (u64 *)*rmapp;
 608                 return NULL;
 609         }
 610         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 611         prev_desc = NULL;
 612         prev_spte = NULL;
 613         while (desc) {
 614                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 615                         if (prev_spte == spte)
 616                                 return desc->shadow_ptes[i];
 617                         prev_spte = desc->shadow_ptes[i];
 618                 }
 619                 desc = desc->more;
 620         }
 621         return NULL;
 622 }
 623
 624 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 625 {
 626         unsigned long *rmapp;
 627         u64 *spte;
 628         int write_protected = 0;
 629
 630         gfn = unalias_gfn(kvm, gfn);
 631         rmapp = gfn_to_rmap(kvm, gfn, 0);
 632
 633         spte = rmap_next(kvm, rmapp, NULL);
 634         while (spte) {
 635                 BUG_ON(!spte);
 636                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 637                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 638                 if (is_writeble_pte(*spte)) {
 639                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 640                         write_protected = 1;
 641                 }
 642                 spte = rmap_next(kvm, rmapp, spte);
 643         }
 644         if (write_protected) {
 645                 pfn_t pfn;
 646
 647                 spte = rmap_next(kvm, rmapp, NULL);
 648                 pfn = spte_to_pfn(*spte);
 649                 kvm_set_pfn_dirty(pfn);
 650         }
 651
 652         /* check for huge page mappings */
 653         rmapp = gfn_to_rmap(kvm, gfn, 1);
 654         spte = rmap_next(kvm, rmapp, NULL);
 655         while (spte) {
 656                 BUG_ON(!spte);
 657                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 658                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 659                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 660                 if (is_writeble_pte(*spte)) {
 661                         rmap_remove(kvm, spte);
 662                         --kvm->stat.lpages;
 663                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 664                         spte = NULL;
 665                         write_protected = 1;
 666                 }
 667                 spte = rmap_next(kvm, rmapp, spte);
 668         }
 669
 670         if (write_protected)
 671                 kvm_flush_remote_tlbs(kvm);
 672 }
 673
 674 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 675 {
 676         u64 *spte;
 677         int need_tlb_flush = 0;
 678
 679         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 680                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 681                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 682                 rmap_remove(kvm, spte);
 683                 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 684                 need_tlb_flush = 1;
 685         }
 686         return need_tlb_flush;
 687 }
 688
 689 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 690                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 691 {
 692         int i;
 693         int retval = 0;
 694
 695         /*
 696          * If mmap_sem isn't taken, we can look the memslots with only
 697          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 698          */
 699         for (i = 0; i < kvm->nmemslots; i++) {
 700                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 701                 unsigned long start = memslot->userspace_addr;
 702                 unsigned long end;
 703
 704                 /* mmu_lock protects userspace_addr */
 705                 if (!start)
 706                         continue;
 707
 708                 end = start + (memslot->npages << PAGE_SHIFT);
 709                 if (hva >= start && hva < end) {
 710                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 711                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 712                         retval |= handler(kvm,
 713                                           &memslot->lpage_info[
 714                                                   gfn_offset /
 715                                                   KVM_PAGES_PER_HPAGE].rmap_pde);
 716                 }
 717         }
 718
 719         return retval;
 720 }
 721
 722 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 723 {
 724         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 725 }
 726
 727 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 728 {
 729         u64 *spte;
 730         int young = 0;
 731
 732         /* always return old for EPT */
 733         if (!shadow_accessed_mask)
 734                 return 0;
 735
 736         spte = rmap_next(kvm, rmapp, NULL);
 737         while (spte) {
 738                 int _young;
 739                 u64 _spte = *spte;
 740                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 741                 _young = _spte & PT_ACCESSED_MASK;
 742                 if (_young) {
 743                         young = 1;
 744                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 745                 }
 746                 spte = rmap_next(kvm, rmapp, spte);
 747         }
 748         return young;
 749 }
 750
 751 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 752 {
 753         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 754 }
 755
 756 #ifdef MMU_DEBUG
 757 static int is_empty_shadow_page(u64 *spt)
 758 {
 759         u64 *pos;
 760         u64 *end;
 761
 762         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 763                 if (is_shadow_present_pte(*pos)) {
 764                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 765                                pos, *pos);
 766                         return 0;
 767                 }
 768         return 1;
 769 }
 770 #endif
 771
 772 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 773 {
 774         ASSERT(is_empty_shadow_page(sp->spt));
 775         list_del(&sp->link);
 776         __free_page(virt_to_page(sp->spt));
 777         __free_page(virt_to_page(sp->gfns));
 778         kfree(sp);
 779         ++kvm->arch.n_free_mmu_pages;
 780 }
 781
 782 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 783 {
 784         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 785 }
 786
 787 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 788                                                u64 *parent_pte)
 789 {
 790         struct kvm_mmu_page *sp;
 791
 792         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 793         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 794         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 795         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 796         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 797         ASSERT(is_empty_shadow_page(sp->spt));
 798         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 799         sp->multimapped = 0;
 800         sp->parent_pte = parent_pte;
 801         --vcpu->kvm->arch.n_free_mmu_pages;
 802         return sp;
 803 }
 804
 805 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 806                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 807 {
 808         struct kvm_pte_chain *pte_chain;
 809         struct hlist_node *node;
 810         int i;
 811
 812         if (!parent_pte)
 813                 return;
 814         if (!sp->multimapped) {
 815                 u64 *old = sp->parent_pte;
 816
 817                 if (!old) {
 818                         sp->parent_pte = parent_pte;
 819                         return;
 820                 }
 821                 sp->multimapped = 1;
 822                 pte_chain = mmu_alloc_pte_chain(vcpu);
 823                 INIT_HLIST_HEAD(&sp->parent_ptes);
 824                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 825                 pte_chain->parent_ptes[0] = old;
 826         }
 827         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 828                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 829                         continue;
 830                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 831                         if (!pte_chain->parent_ptes[i]) {
 832                                 pte_chain->parent_ptes[i] = parent_pte;
 833                                 return;
 834                         }
 835         }
 836         pte_chain = mmu_alloc_pte_chain(vcpu);
 837         BUG_ON(!pte_chain);
 838         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 839         pte_chain->parent_ptes[0] = parent_pte;
 840 }
 841
 842 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 843                                        u64 *parent_pte)
 844 {
 845         struct kvm_pte_chain *pte_chain;
 846         struct hlist_node *node;
 847         int i;
 848
 849         if (!sp->multimapped) {
 850                 BUG_ON(sp->parent_pte != parent_pte);
 851                 sp->parent_pte = NULL;
 852                 return;
 853         }
 854         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 855                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 856                         if (!pte_chain->parent_ptes[i])
 857                                 break;
 858                         if (pte_chain->parent_ptes[i] != parent_pte)
 859                                 continue;
 860                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 861                                 && pte_chain->parent_ptes[i + 1]) {
 862                                 pte_chain->parent_ptes[i]
 863                                         = pte_chain->parent_ptes[i + 1];
 864                                 ++i;
 865                         }
 866                         pte_chain->parent_ptes[i] = NULL;
 867                         if (i == 0) {
 868                                 hlist_del(&pte_chain->link);
 869                                 mmu_free_pte_chain(pte_chain);
 870                                 if (hlist_empty(&sp->parent_ptes)) {
 871                                         sp->multimapped = 0;
 872                                         sp->parent_pte = NULL;
 873                                 }
 874                         }
 875                         return;
 876                 }
 877         BUG();
 878 }
 879
 880
 881 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 882                             mmu_parent_walk_fn fn)
 883 {
 884         struct kvm_pte_chain *pte_chain;
 885         struct hlist_node *node;
 886         struct kvm_mmu_page *parent_sp;
 887         int i;
 888
 889         if (!sp->multimapped && sp->parent_pte) {
 890                 parent_sp = page_header(__pa(sp->parent_pte));
 891                 fn(vcpu, parent_sp);
 892                 mmu_parent_walk(vcpu, parent_sp, fn);
 893                 return;
 894         }
 895         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 896                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 897                         if (!pte_chain->parent_ptes[i])
 898                                 break;
 899                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 900                         fn(vcpu, parent_sp);
 901                         mmu_parent_walk(vcpu, parent_sp, fn);
 902                 }
 903 }
 904
 905 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 906 {
 907         unsigned int index;
 908         struct kvm_mmu_page *sp = page_header(__pa(spte));
 909
 910         index = spte - sp->spt;
 911         __set_bit(index, sp->unsync_child_bitmap);
 912         sp->unsync_children = 1;
 913 }
 914
 915 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 916 {
 917         struct kvm_pte_chain *pte_chain;
 918         struct hlist_node *node;
 919         int i;
 920
 921         if (!sp->parent_pte)
 922                 return;
 923
 924         if (!sp->multimapped) {
 925                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 926                 return;
 927         }
 928
 929         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 930                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 931                         if (!pte_chain->parent_ptes[i])
 932                                 break;
 933                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 934                 }
 935 }
 936
 937 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 938 {
 939         sp->unsync_children = 1;
 940         kvm_mmu_update_parents_unsync(sp);
 941         return 1;
 942 }
 943
 944 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 945                                         struct kvm_mmu_page *sp)
 946 {
 947         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 948         kvm_mmu_update_parents_unsync(sp);
 949 }
 950
 951 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 952                                     struct kvm_mmu_page *sp)
 953 {
 954         int i;
 955
 956         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 957                 sp->spt[i] = shadow_trap_nonpresent_pte;
 958 }
 959
 960 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
 961                                struct kvm_mmu_page *sp)
 962 {
 963         return 1;
 964 }
 965
 966 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
 967 {
 968 }
 969
 970 #define for_each_unsync_children(bitmap, idx)           \
 971         for (idx = find_first_bit(bitmap, 512);         \
 972              idx < 512;                                 \
 973              idx = find_next_bit(bitmap, 512, idx+1))
 974
 975 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
 976                            struct kvm_unsync_walk *walker)
 977 {
 978         int i, ret;
 979
 980         if (!sp->unsync_children)
 981                 return 0;
 982
 983         for_each_unsync_children(sp->unsync_child_bitmap, i) {
 984                 u64 ent = sp->spt[i];
 985
 986                 if (is_shadow_present_pte(ent)) {
 987                         struct kvm_mmu_page *child;
 988                         child = page_header(ent & PT64_BASE_ADDR_MASK);
 989
 990                         if (child->unsync_children) {
 991                                 ret = mmu_unsync_walk(child, walker);
 992                                 if (ret)
 993                                         return ret;
 994                                 __clear_bit(i, sp->unsync_child_bitmap);
 995                         }
 996
 997                         if (child->unsync) {
 998                                 ret = walker->entry(child, walker);
 999                                 __clear_bit(i, sp->unsync_child_bitmap);
1000                                 if (ret)
1001                                         return ret;
1002                         }
1003                 }
1004         }
1005
1006         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1007                 sp->unsync_children = 0;
1008
1009         return 0;
1010 }
1011
1012 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1013 {
1014         unsigned index;
1015         struct hlist_head *bucket;
1016         struct kvm_mmu_page *sp;
1017         struct hlist_node *node;
1018
1019         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1020         index = kvm_page_table_hashfn(gfn);
1021         bucket = &kvm->arch.mmu_page_hash[index];
1022         hlist_for_each_entry(sp, node, bucket, hash_link)
1023                 if (sp->gfn == gfn && !sp->role.metaphysical
1024                     && !sp->role.invalid) {
1025                         pgprintk("%s: found role %x\n",
1026                                  __func__, sp->role.word);
1027                         return sp;
1028                 }
1029         return NULL;
1030 }
1031
1032 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1033 {
1034         WARN_ON(!sp->unsync);
1035         sp->unsync = 0;
1036         --kvm->stat.mmu_unsync;
1037 }
1038
1039 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1040
1041 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1042 {
1043         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1044                 kvm_mmu_zap_page(vcpu->kvm, sp);
1045                 return 1;
1046         }
1047
1048         rmap_write_protect(vcpu->kvm, sp->gfn);
1049         kvm_unlink_unsync_page(vcpu->kvm, sp);
1050         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1051                 kvm_mmu_zap_page(vcpu->kvm, sp);
1052                 return 1;
1053         }
1054
1055         kvm_mmu_flush_tlb(vcpu);
1056         return 0;
1057 }
1058
1059 struct sync_walker {
1060         struct kvm_vcpu *vcpu;
1061         struct kvm_unsync_walk walker;
1062 };
1063
1064 static int mmu_sync_fn(struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk)
1065 {
1066         struct sync_walker *sync_walk = container_of(walk, struct sync_walker,
1067                                                      walker);
1068         struct kvm_vcpu *vcpu = sync_walk->vcpu;
1069
1070         kvm_sync_page(vcpu, sp);
1071         return (need_resched() || spin_needbreak(&vcpu->kvm->mmu_lock));
1072 }
1073
1074 static void mmu_sync_children(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1075 {
1076         struct sync_walker walker = {
1077                 .walker = { .entry = mmu_sync_fn, },
1078                 .vcpu = vcpu,
1079         };
1080
1081         while (mmu_unsync_walk(sp, &walker.walker))
1082                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1083 }
1084
1085 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1086                                              gfn_t gfn,
1087                                              gva_t gaddr,
1088                                              unsigned level,
1089                                              int metaphysical,
1090                                              unsigned access,
1091                                              u64 *parent_pte)
1092 {
1093         union kvm_mmu_page_role role;
1094         unsigned index;
1095         unsigned quadrant;
1096         struct hlist_head *bucket;
1097         struct kvm_mmu_page *sp;
1098         struct hlist_node *node, *tmp;
1099
1100         role.word = 0;
1101         role.glevels = vcpu->arch.mmu.root_level;
1102         role.level = level;
1103         role.metaphysical = metaphysical;
1104         role.access = access;
1105         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1106                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1107                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1108                 role.quadrant = quadrant;
1109         }
1110         pgprintk("%s: looking gfn %lx role %x\n", __func__,
1111                  gfn, role.word);
1112         index = kvm_page_table_hashfn(gfn);
1113         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1114         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1115                 if (sp->gfn == gfn) {
1116                         if (sp->unsync)
1117                                 if (kvm_sync_page(vcpu, sp))
1118                                         continue;
1119
1120                         if (sp->role.word != role.word)
1121                                 continue;
1122
1123                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1124                         if (sp->unsync_children) {
1125                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1126                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1127                         }
1128                         pgprintk("%s: found\n", __func__);
1129                         return sp;
1130                 }
1131         ++vcpu->kvm->stat.mmu_cache_miss;
1132         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1133         if (!sp)
1134                 return sp;
1135         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1136         sp->gfn = gfn;
1137         sp->role = role;
1138         hlist_add_head(&sp->hash_link, bucket);
1139         if (!metaphysical) {
1140                 rmap_write_protect(vcpu->kvm, gfn);
1141                 account_shadowed(vcpu->kvm, gfn);
1142         }
1143         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1144                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1145         else
1146                 nonpaging_prefetch_page(vcpu, sp);
1147         return sp;
1148 }
1149
1150 static int walk_shadow(struct kvm_shadow_walk *walker,
1151                        struct kvm_vcpu *vcpu, u64 addr)
1152 {
1153         hpa_t shadow_addr;
1154         int level;
1155         int r;
1156         u64 *sptep;
1157         unsigned index;
1158
1159         shadow_addr = vcpu->arch.mmu.root_hpa;
1160         level = vcpu->arch.mmu.shadow_root_level;
1161         if (level == PT32E_ROOT_LEVEL) {
1162                 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1163                 shadow_addr &= PT64_BASE_ADDR_MASK;
1164                 --level;
1165         }
1166
1167         while (level >= PT_PAGE_TABLE_LEVEL) {
1168                 index = SHADOW_PT_INDEX(addr, level);
1169                 sptep = ((u64 *)__va(shadow_addr)) + index;
1170                 r = walker->entry(walker, vcpu, addr, sptep, level);
1171                 if (r)
1172                         return r;
1173                 shadow_addr = *sptep & PT64_BASE_ADDR_MASK;
1174                 --level;
1175         }
1176         return 0;
1177 }
1178
1179 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1180                                          struct kvm_mmu_page *sp)
1181 {
1182         unsigned i;
1183         u64 *pt;
1184         u64 ent;
1185
1186         pt = sp->spt;
1187
1188         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1189                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1190                         if (is_shadow_present_pte(pt[i]))
1191                                 rmap_remove(kvm, &pt[i]);
1192                         pt[i] = shadow_trap_nonpresent_pte;
1193                 }
1194                 return;
1195         }
1196
1197         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1198                 ent = pt[i];
1199
1200                 if (is_shadow_present_pte(ent)) {
1201                         if (!is_large_pte(ent)) {
1202                                 ent &= PT64_BASE_ADDR_MASK;
1203                                 mmu_page_remove_parent_pte(page_header(ent),
1204                                                            &pt[i]);
1205                         } else {
1206                                 --kvm->stat.lpages;
1207                                 rmap_remove(kvm, &pt[i]);
1208                         }
1209                 }
1210                 pt[i] = shadow_trap_nonpresent_pte;
1211         }
1212 }
1213
1214 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1215 {
1216         mmu_page_remove_parent_pte(sp, parent_pte);
1217 }
1218
1219 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1220 {
1221         int i;
1222
1223         for (i = 0; i < KVM_MAX_VCPUS; ++i)
1224                 if (kvm->vcpus[i])
1225                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
1226 }
1227
1228 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1229 {
1230         u64 *parent_pte;
1231
1232         while (sp->multimapped || sp->parent_pte) {
1233                 if (!sp->multimapped)
1234                         parent_pte = sp->parent_pte;
1235                 else {
1236                         struct kvm_pte_chain *chain;
1237
1238                         chain = container_of(sp->parent_ptes.first,
1239                                              struct kvm_pte_chain, link);
1240                         parent_pte = chain->parent_ptes[0];
1241                 }
1242                 BUG_ON(!parent_pte);
1243                 kvm_mmu_put_page(sp, parent_pte);
1244                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1245         }
1246 }
1247
1248 struct zap_walker {
1249         struct kvm_unsync_walk walker;
1250         struct kvm *kvm;
1251         int zapped;
1252 };
1253
1254 static int mmu_zap_fn(struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk)
1255 {
1256         struct zap_walker *zap_walk = container_of(walk, struct zap_walker,
1257                                                      walker);
1258         kvm_mmu_zap_page(zap_walk->kvm, sp);
1259         zap_walk->zapped = 1;
1260         return 0;
1261 }
1262
1263 static int mmu_zap_unsync_children(struct kvm *kvm, struct kvm_mmu_page *sp)
1264 {
1265         struct zap_walker walker = {
1266                 .walker = { .entry = mmu_zap_fn, },
1267                 .kvm = kvm,
1268                 .zapped = 0,
1269         };
1270
1271         if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1272                 return 0;
1273         mmu_unsync_walk(sp, &walker.walker);
1274         return walker.zapped;
1275 }
1276
1277 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1278 {
1279         int ret;
1280         ++kvm->stat.mmu_shadow_zapped;
1281         ret = mmu_zap_unsync_children(kvm, sp);
1282         kvm_mmu_page_unlink_children(kvm, sp);
1283         kvm_mmu_unlink_parents(kvm, sp);
1284         kvm_flush_remote_tlbs(kvm);
1285         if (!sp->role.invalid && !sp->role.metaphysical)
1286                 unaccount_shadowed(kvm, sp->gfn);
1287         if (sp->unsync)
1288                 kvm_unlink_unsync_page(kvm, sp);
1289         if (!sp->root_count) {
1290                 hlist_del(&sp->hash_link);
1291                 kvm_mmu_free_page(kvm, sp);
1292         } else {
1293                 sp->role.invalid = 1;
1294                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1295                 kvm_reload_remote_mmus(kvm);
1296         }
1297         kvm_mmu_reset_last_pte_updated(kvm);
1298         return ret;
1299 }
1300
1301 /*
1302  * Changing the number of mmu pages allocated to the vm
1303  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1304  */
1305 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1306 {
1307         /*
1308          * If we set the number of mmu pages to be smaller be than the
1309          * number of actived pages , we must to free some mmu pages before we
1310          * change the value
1311          */
1312
1313         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1314             kvm_nr_mmu_pages) {
1315                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1316                                        - kvm->arch.n_free_mmu_pages;
1317
1318                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1319                         struct kvm_mmu_page *page;
1320
1321                         page = container_of(kvm->arch.active_mmu_pages.prev,
1322                                             struct kvm_mmu_page, link);
1323                         kvm_mmu_zap_page(kvm, page);
1324                         n_used_mmu_pages--;
1325                 }
1326                 kvm->arch.n_free_mmu_pages = 0;
1327         }
1328         else
1329                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1330                                          - kvm->arch.n_alloc_mmu_pages;
1331
1332         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1333 }
1334
1335 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1336 {
1337         unsigned index;
1338         struct hlist_head *bucket;
1339         struct kvm_mmu_page *sp;
1340         struct hlist_node *node, *n;
1341         int r;
1342
1343         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1344         r = 0;
1345         index = kvm_page_table_hashfn(gfn);
1346         bucket = &kvm->arch.mmu_page_hash[index];
1347         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1348                 if (sp->gfn == gfn && !sp->role.metaphysical) {
1349                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1350                                  sp->role.word);
1351                         r = 1;
1352                         if (kvm_mmu_zap_page(kvm, sp))
1353                                 n = bucket->first;
1354                 }
1355         return r;
1356 }
1357
1358 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1359 {
1360         struct kvm_mmu_page *sp;
1361
1362         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1363                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1364                 kvm_mmu_zap_page(kvm, sp);
1365         }
1366 }
1367
1368 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1369 {
1370         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1371         struct kvm_mmu_page *sp = page_header(__pa(pte));
1372
1373         __set_bit(slot, sp->slot_bitmap);
1374 }
1375
1376 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1377 {
1378         int i;
1379         u64 *pt = sp->spt;
1380
1381         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1382                 return;
1383
1384         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1385                 if (pt[i] == shadow_notrap_nonpresent_pte)
1386                         set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1387         }
1388 }
1389
1390 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1391 {
1392         struct page *page;
1393
1394         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1395
1396         if (gpa == UNMAPPED_GVA)
1397                 return NULL;
1398
1399         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1400
1401         return page;
1402 }
1403
1404 /*
1405  * The function is based on mtrr_type_lookup() in
1406  * arch/x86/kernel/cpu/mtrr/generic.c
1407  */
1408 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1409                          u64 start, u64 end)
1410 {
1411         int i;
1412         u64 base, mask;
1413         u8 prev_match, curr_match;
1414         int num_var_ranges = KVM_NR_VAR_MTRR;
1415
1416         if (!mtrr_state->enabled)
1417                 return 0xFF;
1418
1419         /* Make end inclusive end, instead of exclusive */
1420         end--;
1421
1422         /* Look in fixed ranges. Just return the type as per start */
1423         if (mtrr_state->have_fixed && (start < 0x100000)) {
1424                 int idx;
1425
1426                 if (start < 0x80000) {
1427                         idx = 0;
1428                         idx += (start >> 16);
1429                         return mtrr_state->fixed_ranges[idx];
1430                 } else if (start < 0xC0000) {
1431                         idx = 1 * 8;
1432                         idx += ((start - 0x80000) >> 14);
1433                         return mtrr_state->fixed_ranges[idx];
1434                 } else if (start < 0x1000000) {
1435                         idx = 3 * 8;
1436                         idx += ((start - 0xC0000) >> 12);
1437                         return mtrr_state->fixed_ranges[idx];
1438                 }
1439         }
1440
1441         /*
1442          * Look in variable ranges
1443          * Look of multiple ranges matching this address and pick type
1444          * as per MTRR precedence
1445          */
1446         if (!(mtrr_state->enabled & 2))
1447                 return mtrr_state->def_type;
1448
1449         prev_match = 0xFF;
1450         for (i = 0; i < num_var_ranges; ++i) {
1451                 unsigned short start_state, end_state;
1452
1453                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1454                         continue;
1455
1456                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1457                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1458                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1459                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1460
1461                 start_state = ((start & mask) == (base & mask));
1462                 end_state = ((end & mask) == (base & mask));
1463                 if (start_state != end_state)
1464                         return 0xFE;
1465
1466                 if ((start & mask) != (base & mask))
1467                         continue;
1468
1469                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1470                 if (prev_match == 0xFF) {
1471                         prev_match = curr_match;
1472                         continue;
1473                 }
1474
1475                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1476                     curr_match == MTRR_TYPE_UNCACHABLE)
1477                         return MTRR_TYPE_UNCACHABLE;
1478
1479                 if ((prev_match == MTRR_TYPE_WRBACK &&
1480                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1481                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1482                      curr_match == MTRR_TYPE_WRBACK)) {
1483                         prev_match = MTRR_TYPE_WRTHROUGH;
1484                         curr_match = MTRR_TYPE_WRTHROUGH;
1485                 }
1486
1487                 if (prev_match != curr_match)
1488                         return MTRR_TYPE_UNCACHABLE;
1489         }
1490
1491         if (prev_match != 0xFF)
1492                 return prev_match;
1493
1494         return mtrr_state->def_type;
1495 }
1496
1497 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1498 {
1499         u8 mtrr;
1500
1501         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1502                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1503         if (mtrr == 0xfe || mtrr == 0xff)
1504                 mtrr = MTRR_TYPE_WRBACK;
1505         return mtrr;
1506 }
1507
1508 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1509 {
1510         unsigned index;
1511         struct hlist_head *bucket;
1512         struct kvm_mmu_page *s;
1513         struct hlist_node *node, *n;
1514
1515         index = kvm_page_table_hashfn(sp->gfn);
1516         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1517         /* don't unsync if pagetable is shadowed with multiple roles */
1518         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1519                 if (s->gfn != sp->gfn || s->role.metaphysical)
1520                         continue;
1521                 if (s->role.word != sp->role.word)
1522                         return 1;
1523         }
1524         kvm_mmu_mark_parents_unsync(vcpu, sp);
1525         ++vcpu->kvm->stat.mmu_unsync;
1526         sp->unsync = 1;
1527         mmu_convert_notrap(sp);
1528         return 0;
1529 }
1530
1531 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1532                                   bool can_unsync)
1533 {
1534         struct kvm_mmu_page *shadow;
1535
1536         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1537         if (shadow) {
1538                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1539                         return 1;
1540                 if (shadow->unsync)
1541                         return 0;
1542                 if (can_unsync && oos_shadow)
1543                         return kvm_unsync_page(vcpu, shadow);
1544                 return 1;
1545         }
1546         return 0;
1547 }
1548
1549 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1550                     unsigned pte_access, int user_fault,
1551                     int write_fault, int dirty, int largepage,
1552                     gfn_t gfn, pfn_t pfn, bool speculative,
1553                     bool can_unsync)
1554 {
1555         u64 spte;
1556         int ret = 0;
1557         u64 mt_mask = shadow_mt_mask;
1558
1559         /*
1560          * We don't set the accessed bit, since we sometimes want to see
1561          * whether the guest actually used the pte (in order to detect
1562          * demand paging).
1563          */
1564         spte = shadow_base_present_pte | shadow_dirty_mask;
1565         if (!speculative)
1566                 spte |= shadow_accessed_mask;
1567         if (!dirty)
1568                 pte_access &= ~ACC_WRITE_MASK;
1569         if (pte_access & ACC_EXEC_MASK)
1570                 spte |= shadow_x_mask;
1571         else
1572                 spte |= shadow_nx_mask;
1573         if (pte_access & ACC_USER_MASK)
1574                 spte |= shadow_user_mask;
1575         if (largepage)
1576                 spte |= PT_PAGE_SIZE_MASK;
1577         if (mt_mask) {
1578                 mt_mask = get_memory_type(vcpu, gfn) <<
1579                           kvm_x86_ops->get_mt_mask_shift();
1580                 spte |= mt_mask;
1581         }
1582
1583         spte |= (u64)pfn << PAGE_SHIFT;
1584
1585         if ((pte_access & ACC_WRITE_MASK)
1586             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1587
1588                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1589                         ret = 1;
1590                         spte = shadow_trap_nonpresent_pte;
1591                         goto set_pte;
1592                 }
1593
1594                 spte |= PT_WRITABLE_MASK;
1595
1596                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1597                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1598                                  __func__, gfn);
1599                         ret = 1;
1600                         pte_access &= ~ACC_WRITE_MASK;
1601                         if (is_writeble_pte(spte))
1602                                 spte &= ~PT_WRITABLE_MASK;
1603                 }
1604         }
1605
1606         if (pte_access & ACC_WRITE_MASK)
1607                 mark_page_dirty(vcpu->kvm, gfn);
1608
1609 set_pte:
1610         set_shadow_pte(shadow_pte, spte);
1611         return ret;
1612 }
1613
1614 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1615                          unsigned pt_access, unsigned pte_access,
1616                          int user_fault, int write_fault, int dirty,
1617                          int *ptwrite, int largepage, gfn_t gfn,
1618                          pfn_t pfn, bool speculative)
1619 {
1620         int was_rmapped = 0;
1621         int was_writeble = is_writeble_pte(*shadow_pte);
1622
1623         pgprintk("%s: spte %llx access %x write_fault %d"
1624                  " user_fault %d gfn %lx\n",
1625                  __func__, *shadow_pte, pt_access,
1626                  write_fault, user_fault, gfn);
1627
1628         if (is_rmap_pte(*shadow_pte)) {
1629                 /*
1630                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1631                  * the parent of the now unreachable PTE.
1632                  */
1633                 if (largepage && !is_large_pte(*shadow_pte)) {
1634                         struct kvm_mmu_page *child;
1635                         u64 pte = *shadow_pte;
1636
1637                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1638                         mmu_page_remove_parent_pte(child, shadow_pte);
1639                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1640                         pgprintk("hfn old %lx new %lx\n",
1641                                  spte_to_pfn(*shadow_pte), pfn);
1642                         rmap_remove(vcpu->kvm, shadow_pte);
1643                 } else {
1644                         if (largepage)
1645                                 was_rmapped = is_large_pte(*shadow_pte);
1646                         else
1647                                 was_rmapped = 1;
1648                 }
1649         }
1650         if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1651                       dirty, largepage, gfn, pfn, speculative, true)) {
1652                 if (write_fault)
1653                         *ptwrite = 1;
1654                 kvm_x86_ops->tlb_flush(vcpu);
1655         }
1656
1657         pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1658         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1659                  is_large_pte(*shadow_pte)? "2MB" : "4kB",
1660                  is_present_pte(*shadow_pte)?"RW":"R", gfn,
1661                  *shadow_pte, shadow_pte);
1662         if (!was_rmapped && is_large_pte(*shadow_pte))
1663                 ++vcpu->kvm->stat.lpages;
1664
1665         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1666         if (!was_rmapped) {
1667                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1668                 if (!is_rmap_pte(*shadow_pte))
1669                         kvm_release_pfn_clean(pfn);
1670         } else {
1671                 if (was_writeble)
1672                         kvm_release_pfn_dirty(pfn);
1673                 else
1674                         kvm_release_pfn_clean(pfn);
1675         }
1676         if (speculative) {
1677                 vcpu->arch.last_pte_updated = shadow_pte;
1678                 vcpu->arch.last_pte_gfn = gfn;
1679         }
1680 }
1681
1682 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1683 {
1684 }
1685
1686 struct direct_shadow_walk {
1687         struct kvm_shadow_walk walker;
1688         pfn_t pfn;
1689         int write;
1690         int largepage;
1691         int pt_write;
1692 };
1693
1694 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1695                             struct kvm_vcpu *vcpu,
1696                             u64 addr, u64 *sptep, int level)
1697 {
1698         struct direct_shadow_walk *walk =
1699                 container_of(_walk, struct direct_shadow_walk, walker);
1700         struct kvm_mmu_page *sp;
1701         gfn_t pseudo_gfn;
1702         gfn_t gfn = addr >> PAGE_SHIFT;
1703
1704         if (level == PT_PAGE_TABLE_LEVEL
1705             || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1706                 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1707                              0, walk->write, 1, &walk->pt_write,
1708                              walk->largepage, gfn, walk->pfn, false);
1709                 ++vcpu->stat.pf_fixed;
1710                 return 1;
1711         }
1712
1713         if (*sptep == shadow_trap_nonpresent_pte) {
1714                 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1715                 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1716                                       1, ACC_ALL, sptep);
1717                 if (!sp) {
1718                         pgprintk("nonpaging_map: ENOMEM\n");
1719                         kvm_release_pfn_clean(walk->pfn);
1720                         return -ENOMEM;
1721                 }
1722
1723                 set_shadow_pte(sptep,
1724                                __pa(sp->spt)
1725                                | PT_PRESENT_MASK | PT_WRITABLE_MASK
1726                                | shadow_user_mask | shadow_x_mask);
1727         }
1728         return 0;
1729 }
1730
1731 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1732                         int largepage, gfn_t gfn, pfn_t pfn)
1733 {
1734         int r;
1735         struct direct_shadow_walk walker = {
1736                 .walker = { .entry = direct_map_entry, },
1737                 .pfn = pfn,
1738                 .largepage = largepage,
1739                 .write = write,
1740                 .pt_write = 0,
1741         };
1742
1743         r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1744         if (r < 0)
1745                 return r;
1746         return walker.pt_write;
1747 }
1748
1749 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1750 {
1751         int r;
1752         int largepage = 0;
1753         pfn_t pfn;
1754         unsigned long mmu_seq;
1755
1756         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1757                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1758                 largepage = 1;
1759         }
1760
1761         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1762         smp_rmb();
1763         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1764
1765         /* mmio */
1766         if (is_error_pfn(pfn)) {
1767                 kvm_release_pfn_clean(pfn);
1768                 return 1;
1769         }
1770
1771         spin_lock(&vcpu->kvm->mmu_lock);
1772         if (mmu_notifier_retry(vcpu, mmu_seq))
1773                 goto out_unlock;
1774         kvm_mmu_free_some_pages(vcpu);
1775         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1776         spin_unlock(&vcpu->kvm->mmu_lock);
1777
1778
1779         return r;
1780
1781 out_unlock:
1782         spin_unlock(&vcpu->kvm->mmu_lock);
1783         kvm_release_pfn_clean(pfn);
1784         return 0;
1785 }
1786
1787
1788 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1789 {
1790         int i;
1791         struct kvm_mmu_page *sp;
1792
1793         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1794                 return;
1795         spin_lock(&vcpu->kvm->mmu_lock);
1796         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1797                 hpa_t root = vcpu->arch.mmu.root_hpa;
1798
1799                 sp = page_header(root);
1800                 --sp->root_count;
1801                 if (!sp->root_count && sp->role.invalid)
1802                         kvm_mmu_zap_page(vcpu->kvm, sp);
1803                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1804                 spin_unlock(&vcpu->kvm->mmu_lock);
1805                 return;
1806         }
1807         for (i = 0; i < 4; ++i) {
1808                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1809
1810                 if (root) {
1811                         root &= PT64_BASE_ADDR_MASK;
1812                         sp = page_header(root);
1813                         --sp->root_count;
1814                         if (!sp->root_count && sp->role.invalid)
1815                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1816                 }
1817                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1818         }
1819         spin_unlock(&vcpu->kvm->mmu_lock);
1820         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1821 }
1822
1823 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1824 {
1825         int i;
1826         gfn_t root_gfn;
1827         struct kvm_mmu_page *sp;
1828         int metaphysical = 0;
1829
1830         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1831
1832         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1833                 hpa_t root = vcpu->arch.mmu.root_hpa;
1834
1835                 ASSERT(!VALID_PAGE(root));
1836                 if (tdp_enabled)
1837                         metaphysical = 1;
1838                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1839                                       PT64_ROOT_LEVEL, metaphysical,
1840                                       ACC_ALL, NULL);
1841                 root = __pa(sp->spt);
1842                 ++sp->root_count;
1843                 vcpu->arch.mmu.root_hpa = root;
1844                 return;
1845         }
1846         metaphysical = !is_paging(vcpu);
1847         if (tdp_enabled)
1848                 metaphysical = 1;
1849         for (i = 0; i < 4; ++i) {
1850                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1851
1852                 ASSERT(!VALID_PAGE(root));
1853                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1854                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1855                                 vcpu->arch.mmu.pae_root[i] = 0;
1856                                 continue;
1857                         }
1858                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1859                 } else if (vcpu->arch.mmu.root_level == 0)
1860                         root_gfn = 0;
1861                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1862                                       PT32_ROOT_LEVEL, metaphysical,
1863                                       ACC_ALL, NULL);
1864                 root = __pa(sp->spt);
1865                 ++sp->root_count;
1866                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1867         }
1868         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1869 }
1870
1871 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
1872 {
1873         int i;
1874         struct kvm_mmu_page *sp;
1875
1876         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1877                 return;
1878         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1879                 hpa_t root = vcpu->arch.mmu.root_hpa;
1880                 sp = page_header(root);
1881                 mmu_sync_children(vcpu, sp);
1882                 return;
1883         }
1884         for (i = 0; i < 4; ++i) {
1885                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1886
1887                 if (root) {
1888                         root &= PT64_BASE_ADDR_MASK;
1889                         sp = page_header(root);
1890                         mmu_sync_children(vcpu, sp);
1891                 }
1892         }
1893 }
1894
1895 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
1896 {
1897         spin_lock(&vcpu->kvm->mmu_lock);
1898         mmu_sync_roots(vcpu);
1899         spin_unlock(&vcpu->kvm->mmu_lock);
1900 }
1901
1902 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1903 {
1904         return vaddr;
1905 }
1906
1907 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1908                                 u32 error_code)
1909 {
1910         gfn_t gfn;
1911         int r;
1912
1913         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1914         r = mmu_topup_memory_caches(vcpu);
1915         if (r)
1916                 return r;
1917
1918         ASSERT(vcpu);
1919         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1920
1921         gfn = gva >> PAGE_SHIFT;
1922
1923         return nonpaging_map(vcpu, gva & PAGE_MASK,
1924                              error_code & PFERR_WRITE_MASK, gfn);
1925 }
1926
1927 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1928                                 u32 error_code)
1929 {
1930         pfn_t pfn;
1931         int r;
1932         int largepage = 0;
1933         gfn_t gfn = gpa >> PAGE_SHIFT;
1934         unsigned long mmu_seq;
1935
1936         ASSERT(vcpu);
1937         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1938
1939         r = mmu_topup_memory_caches(vcpu);
1940         if (r)
1941                 return r;
1942
1943         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1944                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1945                 largepage = 1;
1946         }
1947         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1948         smp_rmb();
1949         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1950         if (is_error_pfn(pfn)) {
1951                 kvm_release_pfn_clean(pfn);
1952                 return 1;
1953         }
1954         spin_lock(&vcpu->kvm->mmu_lock);
1955         if (mmu_notifier_retry(vcpu, mmu_seq))
1956                 goto out_unlock;
1957         kvm_mmu_free_some_pages(vcpu);
1958         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1959                          largepage, gfn, pfn);
1960         spin_unlock(&vcpu->kvm->mmu_lock);
1961
1962         return r;
1963
1964 out_unlock:
1965         spin_unlock(&vcpu->kvm->mmu_lock);
1966         kvm_release_pfn_clean(pfn);
1967         return 0;
1968 }
1969
1970 static void nonpaging_free(struct kvm_vcpu *vcpu)
1971 {
1972         mmu_free_roots(vcpu);
1973 }
1974
1975 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1976 {
1977         struct kvm_mmu *context = &vcpu->arch.mmu;
1978
1979         context->new_cr3 = nonpaging_new_cr3;
1980         context->page_fault = nonpaging_page_fault;
1981         context->gva_to_gpa = nonpaging_gva_to_gpa;
1982         context->free = nonpaging_free;
1983         context->prefetch_page = nonpaging_prefetch_page;
1984         context->sync_page = nonpaging_sync_page;
1985         context->invlpg = nonpaging_invlpg;
1986         context->root_level = 0;
1987         context->shadow_root_level = PT32E_ROOT_LEVEL;
1988         context->root_hpa = INVALID_PAGE;
1989         return 0;
1990 }
1991
1992 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1993 {
1994         ++vcpu->stat.tlb_flush;
1995         kvm_x86_ops->tlb_flush(vcpu);
1996 }
1997
1998 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1999 {
2000         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2001         mmu_free_roots(vcpu);
2002 }
2003
2004 static void inject_page_fault(struct kvm_vcpu *vcpu,
2005                               u64 addr,
2006                               u32 err_code)
2007 {
2008         kvm_inject_page_fault(vcpu, addr, err_code);
2009 }
2010
2011 static void paging_free(struct kvm_vcpu *vcpu)
2012 {
2013         nonpaging_free(vcpu);
2014 }
2015
2016 #define PTTYPE 64
2017 #include "paging_tmpl.h"
2018 #undef PTTYPE
2019
2020 #define PTTYPE 32
2021 #include "paging_tmpl.h"
2022 #undef PTTYPE
2023
2024 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2025 {
2026         struct kvm_mmu *context = &vcpu->arch.mmu;
2027
2028         ASSERT(is_pae(vcpu));
2029         context->new_cr3 = paging_new_cr3;
2030         context->page_fault = paging64_page_fault;
2031         context->gva_to_gpa = paging64_gva_to_gpa;
2032         context->prefetch_page = paging64_prefetch_page;
2033         context->sync_page = paging64_sync_page;
2034         context->invlpg = paging64_invlpg;
2035         context->free = paging_free;
2036         context->root_level = level;
2037         context->shadow_root_level = level;
2038         context->root_hpa = INVALID_PAGE;
2039         return 0;
2040 }
2041
2042 static int paging64_init_context(struct kvm_vcpu *vcpu)
2043 {
2044         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2045 }
2046
2047 static int paging32_init_context(struct kvm_vcpu *vcpu)
2048 {
2049         struct kvm_mmu *context = &vcpu->arch.mmu;
2050
2051         context->new_cr3 = paging_new_cr3;
2052         context->page_fault = paging32_page_fault;
2053         context->gva_to_gpa = paging32_gva_to_gpa;
2054         context->free = paging_free;
2055         context->prefetch_page = paging32_prefetch_page;
2056         context->sync_page = paging32_sync_page;
2057         context->invlpg = paging32_invlpg;
2058         context->root_level = PT32_ROOT_LEVEL;
2059         context->shadow_root_level = PT32E_ROOT_LEVEL;
2060         context->root_hpa = INVALID_PAGE;
2061         return 0;
2062 }
2063
2064 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2065 {
2066         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2067 }
2068
2069 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2070 {
2071         struct kvm_mmu *context = &vcpu->arch.mmu;
2072
2073         context->new_cr3 = nonpaging_new_cr3;
2074         context->page_fault = tdp_page_fault;
2075         context->free = nonpaging_free;
2076         context->prefetch_page = nonpaging_prefetch_page;
2077         context->sync_page = nonpaging_sync_page;
2078         context->invlpg = nonpaging_invlpg;
2079         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2080         context->root_hpa = INVALID_PAGE;
2081
2082         if (!is_paging(vcpu)) {
2083                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2084                 context->root_level = 0;
2085         } else if (is_long_mode(vcpu)) {
2086                 context->gva_to_gpa = paging64_gva_to_gpa;
2087                 context->root_level = PT64_ROOT_LEVEL;
2088         } else if (is_pae(vcpu)) {
2089                 context->gva_to_gpa = paging64_gva_to_gpa;
2090                 context->root_level = PT32E_ROOT_LEVEL;
2091         } else {
2092                 context->gva_to_gpa = paging32_gva_to_gpa;
2093                 context->root_level = PT32_ROOT_LEVEL;
2094         }
2095
2096         return 0;
2097 }
2098
2099 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2100 {
2101         ASSERT(vcpu);
2102         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2103
2104         if (!is_paging(vcpu))
2105                 return nonpaging_init_context(vcpu);
2106         else if (is_long_mode(vcpu))
2107                 return paging64_init_context(vcpu);
2108         else if (is_pae(vcpu))
2109                 return paging32E_init_context(vcpu);
2110         else
2111                 return paging32_init_context(vcpu);
2112 }
2113
2114 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2115 {
2116         vcpu->arch.update_pte.pfn = bad_pfn;
2117
2118         if (tdp_enabled)
2119                 return init_kvm_tdp_mmu(vcpu);
2120         else
2121                 return init_kvm_softmmu(vcpu);
2122 }
2123
2124 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2125 {
2126         ASSERT(vcpu);
2127         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2128                 vcpu->arch.mmu.free(vcpu);
2129                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2130         }
2131 }
2132
2133 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2134 {
2135         destroy_kvm_mmu(vcpu);
2136         return init_kvm_mmu(vcpu);
2137 }
2138 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2139
2140 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2141 {
2142         int r;
2143
2144         r = mmu_topup_memory_caches(vcpu);
2145         if (r)
2146                 goto out;
2147         spin_lock(&vcpu->kvm->mmu_lock);
2148         kvm_mmu_free_some_pages(vcpu);
2149         mmu_alloc_roots(vcpu);
2150         mmu_sync_roots(vcpu);
2151         spin_unlock(&vcpu->kvm->mmu_lock);
2152         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2153         kvm_mmu_flush_tlb(vcpu);
2154 out:
2155         return r;
2156 }
2157 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2158
2159 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2160 {
2161         mmu_free_roots(vcpu);
2162 }
2163
2164 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2165                                   struct kvm_mmu_page *sp,
2166                                   u64 *spte)
2167 {
2168         u64 pte;
2169         struct kvm_mmu_page *child;
2170
2171         pte = *spte;
2172         if (is_shadow_present_pte(pte)) {
2173                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2174                     is_large_pte(pte))
2175                         rmap_remove(vcpu->kvm, spte);
2176                 else {
2177                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2178                         mmu_page_remove_parent_pte(child, spte);
2179                 }
2180         }
2181         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2182         if (is_large_pte(pte))
2183                 --vcpu->kvm->stat.lpages;
2184 }
2185
2186 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2187                                   struct kvm_mmu_page *sp,
2188                                   u64 *spte,
2189                                   const void *new)
2190 {
2191         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2192                 if (!vcpu->arch.update_pte.largepage ||
2193                     sp->role.glevels == PT32_ROOT_LEVEL) {
2194                         ++vcpu->kvm->stat.mmu_pde_zapped;
2195                         return;
2196                 }
2197         }
2198
2199         ++vcpu->kvm->stat.mmu_pte_updated;
2200         if (sp->role.glevels == PT32_ROOT_LEVEL)
2201                 paging32_update_pte(vcpu, sp, spte, new);
2202         else
2203                 paging64_update_pte(vcpu, sp, spte, new);
2204 }
2205
2206 static bool need_remote_flush(u64 old, u64 new)
2207 {
2208         if (!is_shadow_present_pte(old))
2209                 return false;
2210         if (!is_shadow_present_pte(new))
2211                 return true;
2212         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2213                 return true;
2214         old ^= PT64_NX_MASK;
2215         new ^= PT64_NX_MASK;
2216         return (old & ~new & PT64_PERM_MASK) != 0;
2217 }
2218
2219 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2220 {
2221         if (need_remote_flush(old, new))
2222                 kvm_flush_remote_tlbs(vcpu->kvm);
2223         else
2224                 kvm_mmu_flush_tlb(vcpu);
2225 }
2226
2227 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2228 {
2229         u64 *spte = vcpu->arch.last_pte_updated;
2230
2231         return !!(spte && (*spte & shadow_accessed_mask));
2232 }
2233
2234 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2235                                           const u8 *new, int bytes)
2236 {
2237         gfn_t gfn;
2238         int r;
2239         u64 gpte = 0;
2240         pfn_t pfn;
2241
2242         vcpu->arch.update_pte.largepage = 0;
2243
2244         if (bytes != 4 && bytes != 8)
2245                 return;
2246
2247         /*
2248          * Assume that the pte write on a page table of the same type
2249          * as the current vcpu paging mode.  This is nearly always true
2250          * (might be false while changing modes).  Note it is verified later
2251          * by update_pte().
2252          */
2253         if (is_pae(vcpu)) {
2254                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2255                 if ((bytes == 4) && (gpa % 4 == 0)) {
2256                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2257                         if (r)
2258                                 return;
2259                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2260                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2261                         memcpy((void *)&gpte, new, 8);
2262                 }
2263         } else {
2264                 if ((bytes == 4) && (gpa % 4 == 0))
2265                         memcpy((void *)&gpte, new, 4);
2266         }
2267         if (!is_present_pte(gpte))
2268                 return;
2269         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2270
2271         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2272                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2273                 vcpu->arch.update_pte.largepage = 1;
2274         }
2275         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2276         smp_rmb();
2277         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2278
2279         if (is_error_pfn(pfn)) {
2280                 kvm_release_pfn_clean(pfn);
2281                 return;
2282         }
2283         vcpu->arch.update_pte.gfn = gfn;
2284         vcpu->arch.update_pte.pfn = pfn;
2285 }
2286
2287 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2288 {
2289         u64 *spte = vcpu->arch.last_pte_updated;
2290
2291         if (spte
2292             && vcpu->arch.last_pte_gfn == gfn
2293             && shadow_accessed_mask
2294             && !(*spte & shadow_accessed_mask)
2295             && is_shadow_present_pte(*spte))
2296                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2297 }
2298
2299 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2300                        const u8 *new, int bytes)
2301 {
2302         gfn_t gfn = gpa >> PAGE_SHIFT;
2303         struct kvm_mmu_page *sp;
2304         struct hlist_node *node, *n;
2305         struct hlist_head *bucket;
2306         unsigned index;
2307         u64 entry, gentry;
2308         u64 *spte;
2309         unsigned offset = offset_in_page(gpa);
2310         unsigned pte_size;
2311         unsigned page_offset;
2312         unsigned misaligned;
2313         unsigned quadrant;
2314         int level;
2315         int flooded = 0;
2316         int npte;
2317         int r;
2318
2319         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2320         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2321         spin_lock(&vcpu->kvm->mmu_lock);
2322         kvm_mmu_access_page(vcpu, gfn);
2323         kvm_mmu_free_some_pages(vcpu);
2324         ++vcpu->kvm->stat.mmu_pte_write;
2325         kvm_mmu_audit(vcpu, "pre pte write");
2326         if (gfn == vcpu->arch.last_pt_write_gfn
2327             && !last_updated_pte_accessed(vcpu)) {
2328                 ++vcpu->arch.last_pt_write_count;
2329                 if (vcpu->arch.last_pt_write_count >= 3)
2330                         flooded = 1;
2331         } else {
2332                 vcpu->arch.last_pt_write_gfn = gfn;
2333                 vcpu->arch.last_pt_write_count = 1;
2334                 vcpu->arch.last_pte_updated = NULL;
2335         }
2336         index = kvm_page_table_hashfn(gfn);
2337         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2338         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2339                 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
2340                         continue;
2341                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2342                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2343                 misaligned |= bytes < 4;
2344                 if (misaligned || flooded) {
2345                         /*
2346                          * Misaligned accesses are too much trouble to fix
2347                          * up; also, they usually indicate a page is not used
2348                          * as a page table.
2349                          *
2350                          * If we're seeing too many writes to a page,
2351                          * it may no longer be a page table, or we may be
2352                          * forking, in which case it is better to unmap the
2353                          * page.
2354                          */
2355                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2356                                  gpa, bytes, sp->role.word);
2357                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2358                                 n = bucket->first;
2359                         ++vcpu->kvm->stat.mmu_flooded;
2360                         continue;
2361                 }
2362                 page_offset = offset;
2363                 level = sp->role.level;
2364                 npte = 1;
2365                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2366                         page_offset <<= 1;      /* 32->64 */
2367                         /*
2368                          * A 32-bit pde maps 4MB while the shadow pdes map
2369                          * only 2MB.  So we need to double the offset again
2370                          * and zap two pdes instead of one.
2371                          */
2372                         if (level == PT32_ROOT_LEVEL) {
2373                                 page_offset &= ~7; /* kill rounding error */
2374                                 page_offset <<= 1;
2375                                 npte = 2;
2376                         }
2377                         quadrant = page_offset >> PAGE_SHIFT;
2378                         page_offset &= ~PAGE_MASK;
2379                         if (quadrant != sp->role.quadrant)
2380                                 continue;
2381                 }
2382                 spte = &sp->spt[page_offset / sizeof(*spte)];
2383                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2384                         gentry = 0;
2385                         r = kvm_read_guest_atomic(vcpu->kvm,
2386                                                   gpa & ~(u64)(pte_size - 1),
2387                                                   &gentry, pte_size);
2388                         new = (const void *)&gentry;
2389                         if (r < 0)
2390                                 new = NULL;
2391                 }
2392                 while (npte--) {
2393                         entry = *spte;
2394                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2395                         if (new)
2396                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2397                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2398                         ++spte;
2399                 }
2400         }
2401         kvm_mmu_audit(vcpu, "post pte write");
2402         spin_unlock(&vcpu->kvm->mmu_lock);
2403         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2404                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2405                 vcpu->arch.update_pte.pfn = bad_pfn;
2406         }
2407 }
2408
2409 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2410 {
2411         gpa_t gpa;
2412         int r;
2413
2414         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2415
2416         spin_lock(&vcpu->kvm->mmu_lock);
2417         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2418         spin_unlock(&vcpu->kvm->mmu_lock);
2419         return r;
2420 }
2421 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2422
2423 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2424 {
2425         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2426                 struct kvm_mmu_page *sp;
2427
2428                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2429                                   struct kvm_mmu_page, link);
2430                 kvm_mmu_zap_page(vcpu->kvm, sp);
2431                 ++vcpu->kvm->stat.mmu_recycled;
2432         }
2433 }
2434
2435 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2436 {
2437         int r;
2438         enum emulation_result er;
2439
2440         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2441         if (r < 0)
2442                 goto out;
2443
2444         if (!r) {
2445                 r = 1;
2446                 goto out;
2447         }
2448
2449         r = mmu_topup_memory_caches(vcpu);
2450         if (r)
2451                 goto out;
2452
2453         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2454
2455         switch (er) {
2456         case EMULATE_DONE:
2457                 return 1;
2458         case EMULATE_DO_MMIO:
2459                 ++vcpu->stat.mmio_exits;
2460                 return 0;
2461         case EMULATE_FAIL:
2462                 kvm_report_emulation_failure(vcpu, "pagetable");
2463                 return 1;
2464         default:
2465                 BUG();
2466         }
2467 out:
2468         return r;
2469 }
2470 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2471
2472 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2473 {
2474         spin_lock(&vcpu->kvm->mmu_lock);
2475         vcpu->arch.mmu.invlpg(vcpu, gva);
2476         spin_unlock(&vcpu->kvm->mmu_lock);
2477         kvm_mmu_flush_tlb(vcpu);
2478         ++vcpu->stat.invlpg;
2479 }
2480 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2481
2482 void kvm_enable_tdp(void)
2483 {
2484         tdp_enabled = true;
2485 }
2486 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2487
2488 void kvm_disable_tdp(void)
2489 {
2490         tdp_enabled = false;
2491 }
2492 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2493
2494 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2495 {
2496         struct kvm_mmu_page *sp;
2497
2498         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2499                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2500                                   struct kvm_mmu_page, link);
2501                 kvm_mmu_zap_page(vcpu->kvm, sp);
2502                 cond_resched();
2503         }
2504         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2505 }
2506
2507 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2508 {
2509         struct page *page;
2510         int i;
2511
2512         ASSERT(vcpu);
2513
2514         if (vcpu->kvm->arch.n_requested_mmu_pages)
2515                 vcpu->kvm->arch.n_free_mmu_pages =
2516                                         vcpu->kvm->arch.n_requested_mmu_pages;
2517         else
2518                 vcpu->kvm->arch.n_free_mmu_pages =
2519                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2520         /*
2521          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2522          * Therefore we need to allocate shadow page tables in the first
2523          * 4GB of memory, which happens to fit the DMA32 zone.
2524          */
2525         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2526         if (!page)
2527                 goto error_1;
2528         vcpu->arch.mmu.pae_root = page_address(page);
2529         for (i = 0; i < 4; ++i)
2530                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2531
2532         return 0;
2533
2534 error_1:
2535         free_mmu_pages(vcpu);
2536         return -ENOMEM;
2537 }
2538
2539 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2540 {
2541         ASSERT(vcpu);
2542         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2543
2544         return alloc_mmu_pages(vcpu);
2545 }
2546
2547 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2548 {
2549         ASSERT(vcpu);
2550         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2551
2552         return init_kvm_mmu(vcpu);
2553 }
2554
2555 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2556 {
2557         ASSERT(vcpu);
2558
2559         destroy_kvm_mmu(vcpu);
2560         free_mmu_pages(vcpu);
2561         mmu_free_memory_caches(vcpu);
2562 }
2563
2564 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2565 {
2566         struct kvm_mmu_page *sp;
2567
2568         spin_lock(&kvm->mmu_lock);
2569         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2570                 int i;
2571                 u64 *pt;
2572
2573                 if (!test_bit(slot, sp->slot_bitmap))
2574                         continue;
2575
2576                 pt = sp->spt;
2577                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2578                         /* avoid RMW */
2579                         if (pt[i] & PT_WRITABLE_MASK)
2580                                 pt[i] &= ~PT_WRITABLE_MASK;
2581         }
2582         kvm_flush_remote_tlbs(kvm);
2583         spin_unlock(&kvm->mmu_lock);
2584 }
2585
2586 void kvm_mmu_zap_all(struct kvm *kvm)
2587 {
2588         struct kvm_mmu_page *sp, *node;
2589
2590         spin_lock(&kvm->mmu_lock);
2591         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2592                 if (kvm_mmu_zap_page(kvm, sp))
2593                         node = container_of(kvm->arch.active_mmu_pages.next,
2594                                             struct kvm_mmu_page, link);
2595         spin_unlock(&kvm->mmu_lock);
2596
2597         kvm_flush_remote_tlbs(kvm);
2598 }
2599
2600 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2601 {
2602         struct kvm_mmu_page *page;
2603
2604         page = container_of(kvm->arch.active_mmu_pages.prev,
2605                             struct kvm_mmu_page, link);
2606         kvm_mmu_zap_page(kvm, page);
2607 }
2608
2609 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2610 {
2611         struct kvm *kvm;
2612         struct kvm *kvm_freed = NULL;
2613         int cache_count = 0;
2614
2615         spin_lock(&kvm_lock);
2616
2617         list_for_each_entry(kvm, &vm_list, vm_list) {
2618                 int npages;
2619
2620                 if (!down_read_trylock(&kvm->slots_lock))
2621                         continue;
2622                 spin_lock(&kvm->mmu_lock);
2623                 npages = kvm->arch.n_alloc_mmu_pages -
2624                          kvm->arch.n_free_mmu_pages;
2625                 cache_count += npages;
2626                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2627                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2628                         cache_count--;
2629                         kvm_freed = kvm;
2630                 }
2631                 nr_to_scan--;
2632
2633                 spin_unlock(&kvm->mmu_lock);
2634                 up_read(&kvm->slots_lock);
2635         }
2636         if (kvm_freed)
2637                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2638
2639         spin_unlock(&kvm_lock);
2640
2641         return cache_count;
2642 }
2643
2644 static struct shrinker mmu_shrinker = {
2645         .shrink = mmu_shrink,
2646         .seeks = DEFAULT_SEEKS * 10,
2647 };
2648
2649 static void mmu_destroy_caches(void)
2650 {
2651         if (pte_chain_cache)
2652                 kmem_cache_destroy(pte_chain_cache);
2653         if (rmap_desc_cache)
2654                 kmem_cache_destroy(rmap_desc_cache);
2655         if (mmu_page_header_cache)
2656                 kmem_cache_destroy(mmu_page_header_cache);
2657 }
2658
2659 void kvm_mmu_module_exit(void)
2660 {
2661         mmu_destroy_caches();
2662         unregister_shrinker(&mmu_shrinker);
2663 }
2664
2665 int kvm_mmu_module_init(void)
2666 {
2667         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2668                                             sizeof(struct kvm_pte_chain),
2669                                             0, 0, NULL);
2670         if (!pte_chain_cache)
2671                 goto nomem;
2672         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2673                                             sizeof(struct kvm_rmap_desc),
2674                                             0, 0, NULL);
2675         if (!rmap_desc_cache)
2676                 goto nomem;
2677
2678         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2679                                                   sizeof(struct kvm_mmu_page),
2680                                                   0, 0, NULL);
2681         if (!mmu_page_header_cache)
2682                 goto nomem;
2683
2684         register_shrinker(&mmu_shrinker);
2685
2686         return 0;
2687
2688 nomem:
2689         mmu_destroy_caches();
2690         return -ENOMEM;
2691 }
2692
2693 /*
2694  * Caculate mmu pages needed for kvm.
2695  */
2696 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2697 {
2698         int i;
2699         unsigned int nr_mmu_pages;
2700         unsigned int  nr_pages = 0;
2701
2702         for (i = 0; i < kvm->nmemslots; i++)
2703                 nr_pages += kvm->memslots[i].npages;
2704
2705         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2706         nr_mmu_pages = max(nr_mmu_pages,
2707                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2708
2709         return nr_mmu_pages;
2710 }
2711
2712 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2713                                 unsigned len)
2714 {
2715         if (len > buffer->len)
2716                 return NULL;
2717         return buffer->ptr;
2718 }
2719
2720 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2721                                 unsigned len)
2722 {
2723         void *ret;
2724
2725         ret = pv_mmu_peek_buffer(buffer, len);
2726         if (!ret)
2727                 return ret;
2728         buffer->ptr += len;
2729         buffer->len -= len;
2730         buffer->processed += len;
2731         return ret;
2732 }
2733
2734 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2735                              gpa_t addr, gpa_t value)
2736 {
2737         int bytes = 8;
2738         int r;
2739
2740         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2741                 bytes = 4;
2742
2743         r = mmu_topup_memory_caches(vcpu);
2744         if (r)
2745                 return r;
2746
2747         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2748                 return -EFAULT;
2749
2750         return 1;
2751 }
2752
2753 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2754 {
2755         kvm_x86_ops->tlb_flush(vcpu);
2756         set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
2757         return 1;
2758 }
2759
2760 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2761 {
2762         spin_lock(&vcpu->kvm->mmu_lock);
2763         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2764         spin_unlock(&vcpu->kvm->mmu_lock);
2765         return 1;
2766 }
2767
2768 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2769                              struct kvm_pv_mmu_op_buffer *buffer)
2770 {
2771         struct kvm_mmu_op_header *header;
2772
2773         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2774         if (!header)
2775                 return 0;
2776         switch (header->op) {
2777         case KVM_MMU_OP_WRITE_PTE: {
2778                 struct kvm_mmu_op_write_pte *wpte;
2779
2780                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2781                 if (!wpte)
2782                         return 0;
2783                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2784                                         wpte->pte_val);
2785         }
2786         case KVM_MMU_OP_FLUSH_TLB: {
2787                 struct kvm_mmu_op_flush_tlb *ftlb;
2788
2789                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2790                 if (!ftlb)
2791                         return 0;
2792                 return kvm_pv_mmu_flush_tlb(vcpu);
2793         }
2794         case KVM_MMU_OP_RELEASE_PT: {
2795                 struct kvm_mmu_op_release_pt *rpt;
2796
2797                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2798                 if (!rpt)
2799                         return 0;
2800                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2801         }
2802         default: return 0;
2803         }
2804 }
2805
2806 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2807                   gpa_t addr, unsigned long *ret)
2808 {
2809         int r;
2810         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2811
2812         buffer->ptr = buffer->buf;
2813         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2814         buffer->processed = 0;
2815
2816         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2817         if (r)
2818                 goto out;
2819
2820         while (buffer->len) {
2821                 r = kvm_pv_mmu_op_one(vcpu, buffer);
2822                 if (r < 0)
2823                         goto out;
2824                 if (r == 0)
2825                         break;
2826         }
2827
2828         r = 1;
2829 out:
2830         *ret = buffer->processed;
2831         return r;
2832 }
2833
2834 #ifdef AUDIT
2835
2836 static const char *audit_msg;
2837
2838 static gva_t canonicalize(gva_t gva)
2839 {
2840 #ifdef CONFIG_X86_64
2841         gva = (long long)(gva << 16) >> 16;
2842 #endif
2843         return gva;
2844 }
2845
2846 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2847                                 gva_t va, int level)
2848 {
2849         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2850         int i;
2851         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2852
2853         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2854                 u64 ent = pt[i];
2855
2856                 if (ent == shadow_trap_nonpresent_pte)
2857                         continue;
2858
2859                 va = canonicalize(va);
2860                 if (level > 1) {
2861                         if (ent == shadow_notrap_nonpresent_pte)
2862                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2863                                        " in nonleaf level: levels %d gva %lx"
2864                                        " level %d pte %llx\n", audit_msg,
2865                                        vcpu->arch.mmu.root_level, va, level, ent);
2866
2867                         audit_mappings_page(vcpu, ent, va, level - 1);
2868                 } else {
2869                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2870                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2871
2872                         if (is_shadow_present_pte(ent)
2873                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2874                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2875                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2876                                        audit_msg, vcpu->arch.mmu.root_level,
2877                                        va, gpa, hpa, ent,
2878                                        is_shadow_present_pte(ent));
2879                         else if (ent == shadow_notrap_nonpresent_pte
2880                                  && !is_error_hpa(hpa))
2881                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2882                                        " valid guest gva %lx\n", audit_msg, va);
2883                         kvm_release_pfn_clean(pfn);
2884
2885                 }
2886         }
2887 }
2888
2889 static void audit_mappings(struct kvm_vcpu *vcpu)
2890 {
2891         unsigned i;
2892
2893         if (vcpu->arch.mmu.root_level == 4)
2894                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2895         else
2896                 for (i = 0; i < 4; ++i)
2897                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2898                                 audit_mappings_page(vcpu,
2899                                                     vcpu->arch.mmu.pae_root[i],
2900                                                     i << 30,
2901                                                     2);
2902 }
2903
2904 static int count_rmaps(struct kvm_vcpu *vcpu)
2905 {
2906         int nmaps = 0;
2907         int i, j, k;
2908
2909         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2910                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2911                 struct kvm_rmap_desc *d;
2912
2913                 for (j = 0; j < m->npages; ++j) {
2914                         unsigned long *rmapp = &m->rmap[j];
2915
2916                         if (!*rmapp)
2917                                 continue;
2918                         if (!(*rmapp & 1)) {
2919                                 ++nmaps;
2920                                 continue;
2921                         }
2922                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2923                         while (d) {
2924                                 for (k = 0; k < RMAP_EXT; ++k)
2925                                         if (d->shadow_ptes[k])
2926                                                 ++nmaps;
2927                                         else
2928                                                 break;
2929                                 d = d->more;
2930                         }
2931                 }
2932         }
2933         return nmaps;
2934 }
2935
2936 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2937 {
2938         int nmaps = 0;
2939         struct kvm_mmu_page *sp;
2940         int i;
2941
2942         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2943                 u64 *pt = sp->spt;
2944
2945                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2946                         continue;
2947
2948                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2949                         u64 ent = pt[i];
2950
2951                         if (!(ent & PT_PRESENT_MASK))
2952                                 continue;
2953                         if (!(ent & PT_WRITABLE_MASK))
2954                                 continue;
2955                         ++nmaps;
2956                 }
2957         }
2958         return nmaps;
2959 }
2960
2961 static void audit_rmap(struct kvm_vcpu *vcpu)
2962 {
2963         int n_rmap = count_rmaps(vcpu);
2964         int n_actual = count_writable_mappings(vcpu);
2965
2966         if (n_rmap != n_actual)
2967                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2968                        __func__, audit_msg, n_rmap, n_actual);
2969 }
2970
2971 static void audit_write_protection(struct kvm_vcpu *vcpu)
2972 {
2973         struct kvm_mmu_page *sp;
2974         struct kvm_memory_slot *slot;
2975         unsigned long *rmapp;
2976         gfn_t gfn;
2977
2978         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2979                 if (sp->role.metaphysical)
2980                         continue;
2981
2982                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2983                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
2984                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2985                 if (*rmapp)
2986                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2987                                " mappings: gfn %lx role %x\n",
2988                                __func__, audit_msg, sp->gfn,
2989                                sp->role.word);
2990         }
2991 }
2992
2993 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2994 {
2995         int olddbg = dbg;
2996
2997         dbg = 0;
2998         audit_msg = msg;
2999         audit_rmap(vcpu);
3000         audit_write_protection(vcpu);
3001         audit_mappings(vcpu);
3002         dbg = olddbg;
3003 }
3004
3005 #endif