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 "vmx.h"
  21 #include "mmu.h"
  22
  23 #include <linux/kvm_host.h>
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29 #include <linux/swap.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/compiler.h>
  32
  33 #include <asm/page.h>
  34 #include <asm/cmpxchg.h>
  35 #include <asm/io.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 = 1;
  70 #endif
  71
  72 #ifndef MMU_DEBUG
  73 #define ASSERT(x) do { } while (0)
  74 #else
  75 #define ASSERT(x)                                                       \
  76         if (!(x)) {                                                     \
  77                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  78                        __FILE__, __LINE__, #x);                         \
  79         }
  80 #endif
  81
  82 #define PT64_PT_BITS 9
  83 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  84 #define PT32_PT_BITS 10
  85 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  86
  87 #define PT_WRITABLE_SHIFT 1
  88
  89 #define PT_PRESENT_MASK (1ULL << 0)
  90 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  91 #define PT_USER_MASK (1ULL << 2)
  92 #define PT_PWT_MASK (1ULL << 3)
  93 #define PT_PCD_MASK (1ULL << 4)
  94 #define PT_ACCESSED_MASK (1ULL << 5)
  95 #define PT_DIRTY_MASK (1ULL << 6)
  96 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  97 #define PT_PAT_MASK (1ULL << 7)
  98 #define PT_GLOBAL_MASK (1ULL << 8)
  99 #define PT64_NX_SHIFT 63
 100 #define PT64_NX_MASK (1ULL << PT64_NX_SHIFT)
 101
 102 #define PT_PAT_SHIFT 7
 103 #define PT_DIR_PAT_SHIFT 12
 104 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
 105
 106 #define PT32_DIR_PSE36_SIZE 4
 107 #define PT32_DIR_PSE36_SHIFT 13
 108 #define PT32_DIR_PSE36_MASK \
 109         (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
 110
 111
 112 #define PT_FIRST_AVAIL_BITS_SHIFT 9
 113 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
 114
 115 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 116
 117 #define PT64_LEVEL_BITS 9
 118
 119 #define PT64_LEVEL_SHIFT(level) \
 120                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
 121
 122 #define PT64_LEVEL_MASK(level) \
 123                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 124
 125 #define PT64_INDEX(address, level)\
 126         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 127
 128
 129 #define PT32_LEVEL_BITS 10
 130
 131 #define PT32_LEVEL_SHIFT(level) \
 132                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 133
 134 #define PT32_LEVEL_MASK(level) \
 135                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 136
 137 #define PT32_INDEX(address, level)\
 138         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 139
 140
 141 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 142 #define PT64_DIR_BASE_ADDR_MASK \
 143         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 144
 145 #define PT32_BASE_ADDR_MASK PAGE_MASK
 146 #define PT32_DIR_BASE_ADDR_MASK \
 147         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 148
 149 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 150                         | PT64_NX_MASK)
 151
 152 #define PFERR_PRESENT_MASK (1U << 0)
 153 #define PFERR_WRITE_MASK (1U << 1)
 154 #define PFERR_USER_MASK (1U << 2)
 155 #define PFERR_FETCH_MASK (1U << 4)
 156
 157 #define PT64_ROOT_LEVEL 4
 158 #define PT32_ROOT_LEVEL 2
 159 #define PT32E_ROOT_LEVEL 3
 160
 161 #define PT_DIRECTORY_LEVEL 2
 162 #define PT_PAGE_TABLE_LEVEL 1
 163
 164 #define RMAP_EXT 4
 165
 166 #define ACC_EXEC_MASK    1
 167 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 168 #define ACC_USER_MASK    PT_USER_MASK
 169 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 170
 171 struct kvm_pv_mmu_op_buffer {
 172         void *ptr;
 173         unsigned len;
 174         unsigned processed;
 175         char buf[512] __aligned(sizeof(long));
 176 };
 177
 178 struct kvm_rmap_desc {
 179         u64 *shadow_ptes[RMAP_EXT];
 180         struct kvm_rmap_desc *more;
 181 };
 182
 183 static struct kmem_cache *pte_chain_cache;
 184 static struct kmem_cache *rmap_desc_cache;
 185 static struct kmem_cache *mmu_page_header_cache;
 186
 187 static u64 __read_mostly shadow_trap_nonpresent_pte;
 188 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 189
 190 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 191 {
 192         shadow_trap_nonpresent_pte = trap_pte;
 193         shadow_notrap_nonpresent_pte = notrap_pte;
 194 }
 195 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 196
 197 static int is_write_protection(struct kvm_vcpu *vcpu)
 198 {
 199         return vcpu->arch.cr0 & X86_CR0_WP;
 200 }
 201
 202 static int is_cpuid_PSE36(void)
 203 {
 204         return 1;
 205 }
 206
 207 static int is_nx(struct kvm_vcpu *vcpu)
 208 {
 209         return vcpu->arch.shadow_efer & EFER_NX;
 210 }
 211
 212 static int is_present_pte(unsigned long pte)
 213 {
 214         return pte & PT_PRESENT_MASK;
 215 }
 216
 217 static int is_shadow_present_pte(u64 pte)
 218 {
 219         return pte != shadow_trap_nonpresent_pte
 220                 && pte != shadow_notrap_nonpresent_pte;
 221 }
 222
 223 static int is_large_pte(u64 pte)
 224 {
 225         return pte & PT_PAGE_SIZE_MASK;
 226 }
 227
 228 static int is_writeble_pte(unsigned long pte)
 229 {
 230         return pte & PT_WRITABLE_MASK;
 231 }
 232
 233 static int is_dirty_pte(unsigned long pte)
 234 {
 235         return pte & PT_DIRTY_MASK;
 236 }
 237
 238 static int is_rmap_pte(u64 pte)
 239 {
 240         return is_shadow_present_pte(pte);
 241 }
 242
 243 static struct page *spte_to_page(u64 pte)
 244 {
 245         hfn_t hfn = (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 246
 247         return pfn_to_page(hfn);
 248 }
 249
 250 static gfn_t pse36_gfn_delta(u32 gpte)
 251 {
 252         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 253
 254         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 255 }
 256
 257 static void set_shadow_pte(u64 *sptep, u64 spte)
 258 {
 259 #ifdef CONFIG_X86_64
 260         set_64bit((unsigned long *)sptep, spte);
 261 #else
 262         set_64bit((unsigned long long *)sptep, spte);
 263 #endif
 264 }
 265
 266 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 267                                   struct kmem_cache *base_cache, int min)
 268 {
 269         void *obj;
 270
 271         if (cache->nobjs >= min)
 272                 return 0;
 273         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 274                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 275                 if (!obj)
 276                         return -ENOMEM;
 277                 cache->objects[cache->nobjs++] = obj;
 278         }
 279         return 0;
 280 }
 281
 282 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 283 {
 284         while (mc->nobjs)
 285                 kfree(mc->objects[--mc->nobjs]);
 286 }
 287
 288 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 289                                        int min)
 290 {
 291         struct page *page;
 292
 293         if (cache->nobjs >= min)
 294                 return 0;
 295         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 296                 page = alloc_page(GFP_KERNEL);
 297                 if (!page)
 298                         return -ENOMEM;
 299                 set_page_private(page, 0);
 300                 cache->objects[cache->nobjs++] = page_address(page);
 301         }
 302         return 0;
 303 }
 304
 305 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 306 {
 307         while (mc->nobjs)
 308                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 309 }
 310
 311 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 312 {
 313         int r;
 314
 315         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 316                                    pte_chain_cache, 4);
 317         if (r)
 318                 goto out;
 319         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 320                                    rmap_desc_cache, 1);
 321         if (r)
 322                 goto out;
 323         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 324         if (r)
 325                 goto out;
 326         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 327                                    mmu_page_header_cache, 4);
 328 out:
 329         return r;
 330 }
 331
 332 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 333 {
 334         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 335         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 336         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 337         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 338 }
 339
 340 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 341                                     size_t size)
 342 {
 343         void *p;
 344
 345         BUG_ON(!mc->nobjs);
 346         p = mc->objects[--mc->nobjs];
 347         memset(p, 0, size);
 348         return p;
 349 }
 350
 351 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 352 {
 353         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 354                                       sizeof(struct kvm_pte_chain));
 355 }
 356
 357 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 358 {
 359         kfree(pc);
 360 }
 361
 362 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 363 {
 364         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 365                                       sizeof(struct kvm_rmap_desc));
 366 }
 367
 368 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 369 {
 370         kfree(rd);
 371 }
 372
 373 /*
 374  * Return the pointer to the largepage write count for a given
 375  * gfn, handling slots that are not large page aligned.
 376  */
 377 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 378 {
 379         unsigned long idx;
 380
 381         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 382               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 383         return &slot->lpage_info[idx].write_count;
 384 }
 385
 386 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 387 {
 388         int *write_count;
 389
 390         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 391         *write_count += 1;
 392         WARN_ON(*write_count > KVM_PAGES_PER_HPAGE);
 393 }
 394
 395 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 396 {
 397         int *write_count;
 398
 399         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 400         *write_count -= 1;
 401         WARN_ON(*write_count < 0);
 402 }
 403
 404 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 405 {
 406         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
 407         int *largepage_idx;
 408
 409         if (slot) {
 410                 largepage_idx = slot_largepage_idx(gfn, slot);
 411                 return *largepage_idx;
 412         }
 413
 414         return 1;
 415 }
 416
 417 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 418 {
 419         struct vm_area_struct *vma;
 420         unsigned long addr;
 421
 422         addr = gfn_to_hva(kvm, gfn);
 423         if (kvm_is_error_hva(addr))
 424                 return 0;
 425
 426         vma = find_vma(current->mm, addr);
 427         if (vma && is_vm_hugetlb_page(vma))
 428                 return 1;
 429
 430         return 0;
 431 }
 432
 433 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 434 {
 435         struct kvm_memory_slot *slot;
 436
 437         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 438                 return 0;
 439
 440         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 441                 return 0;
 442
 443         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 444         if (slot && slot->dirty_bitmap)
 445                 return 0;
 446
 447         return 1;
 448 }
 449
 450 /*
 451  * Take gfn and return the reverse mapping to it.
 452  * Note: gfn must be unaliased before this function get called
 453  */
 454
 455 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 456 {
 457         struct kvm_memory_slot *slot;
 458         unsigned long idx;
 459
 460         slot = gfn_to_memslot(kvm, gfn);
 461         if (!lpage)
 462                 return &slot->rmap[gfn - slot->base_gfn];
 463
 464         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 465               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 466
 467         return &slot->lpage_info[idx].rmap_pde;
 468 }
 469
 470 /*
 471  * Reverse mapping data structures:
 472  *
 473  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 474  * that points to page_address(page).
 475  *
 476  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 477  * containing more mappings.
 478  */
 479 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 480 {
 481         struct kvm_mmu_page *sp;
 482         struct kvm_rmap_desc *desc;
 483         unsigned long *rmapp;
 484         int i;
 485
 486         if (!is_rmap_pte(*spte))
 487                 return;
 488         gfn = unalias_gfn(vcpu->kvm, gfn);
 489         sp = page_header(__pa(spte));
 490         sp->gfns[spte - sp->spt] = gfn;
 491         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 492         if (!*rmapp) {
 493                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 494                 *rmapp = (unsigned long)spte;
 495         } else if (!(*rmapp & 1)) {
 496                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 497                 desc = mmu_alloc_rmap_desc(vcpu);
 498                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 499                 desc->shadow_ptes[1] = spte;
 500                 *rmapp = (unsigned long)desc | 1;
 501         } else {
 502                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 503                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 504                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 505                         desc = desc->more;
 506                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 507                         desc->more = mmu_alloc_rmap_desc(vcpu);
 508                         desc = desc->more;
 509                 }
 510                 for (i = 0; desc->shadow_ptes[i]; ++i)
 511                         ;
 512                 desc->shadow_ptes[i] = spte;
 513         }
 514 }
 515
 516 static void rmap_desc_remove_entry(unsigned long *rmapp,
 517                                    struct kvm_rmap_desc *desc,
 518                                    int i,
 519                                    struct kvm_rmap_desc *prev_desc)
 520 {
 521         int j;
 522
 523         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 524                 ;
 525         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 526         desc->shadow_ptes[j] = NULL;
 527         if (j != 0)
 528                 return;
 529         if (!prev_desc && !desc->more)
 530                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 531         else
 532                 if (prev_desc)
 533                         prev_desc->more = desc->more;
 534                 else
 535                         *rmapp = (unsigned long)desc->more | 1;
 536         mmu_free_rmap_desc(desc);
 537 }
 538
 539 static void rmap_remove(struct kvm *kvm, u64 *spte)
 540 {
 541         struct kvm_rmap_desc *desc;
 542         struct kvm_rmap_desc *prev_desc;
 543         struct kvm_mmu_page *sp;
 544         struct page *page;
 545         unsigned long *rmapp;
 546         int i;
 547
 548         if (!is_rmap_pte(*spte))
 549                 return;
 550         sp = page_header(__pa(spte));
 551         page = spte_to_page(*spte);
 552         mark_page_accessed(page);
 553         if (is_writeble_pte(*spte))
 554                 kvm_release_page_dirty(page);
 555         else
 556                 kvm_release_page_clean(page);
 557         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 558         if (!*rmapp) {
 559                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 560                 BUG();
 561         } else if (!(*rmapp & 1)) {
 562                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 563                 if ((u64 *)*rmapp != spte) {
 564                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 565                                spte, *spte);
 566                         BUG();
 567                 }
 568                 *rmapp = 0;
 569         } else {
 570                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 571                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 572                 prev_desc = NULL;
 573                 while (desc) {
 574                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 575                                 if (desc->shadow_ptes[i] == spte) {
 576                                         rmap_desc_remove_entry(rmapp,
 577                                                                desc, i,
 578                                                                prev_desc);
 579                                         return;
 580                                 }
 581                         prev_desc = desc;
 582                         desc = desc->more;
 583                 }
 584                 BUG();
 585         }
 586 }
 587
 588 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 589 {
 590         struct kvm_rmap_desc *desc;
 591         struct kvm_rmap_desc *prev_desc;
 592         u64 *prev_spte;
 593         int i;
 594
 595         if (!*rmapp)
 596                 return NULL;
 597         else if (!(*rmapp & 1)) {
 598                 if (!spte)
 599                         return (u64 *)*rmapp;
 600                 return NULL;
 601         }
 602         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 603         prev_desc = NULL;
 604         prev_spte = NULL;
 605         while (desc) {
 606                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 607                         if (prev_spte == spte)
 608                                 return desc->shadow_ptes[i];
 609                         prev_spte = desc->shadow_ptes[i];
 610                 }
 611                 desc = desc->more;
 612         }
 613         return NULL;
 614 }
 615
 616 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 617 {
 618         unsigned long *rmapp;
 619         u64 *spte;
 620         int write_protected = 0;
 621
 622         gfn = unalias_gfn(kvm, gfn);
 623         rmapp = gfn_to_rmap(kvm, gfn, 0);
 624
 625         spte = rmap_next(kvm, rmapp, NULL);
 626         while (spte) {
 627                 BUG_ON(!spte);
 628                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 629                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 630                 if (is_writeble_pte(*spte)) {
 631                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 632                         write_protected = 1;
 633                 }
 634                 spte = rmap_next(kvm, rmapp, spte);
 635         }
 636         if (write_protected) {
 637                 struct page *page;
 638
 639                 spte = rmap_next(kvm, rmapp, NULL);
 640                 page = spte_to_page(*spte);
 641                 SetPageDirty(page);
 642         }
 643
 644         /* check for huge page mappings */
 645         rmapp = gfn_to_rmap(kvm, gfn, 1);
 646         spte = rmap_next(kvm, rmapp, NULL);
 647         while (spte) {
 648                 BUG_ON(!spte);
 649                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 650                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 651                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 652                 if (is_writeble_pte(*spte)) {
 653                         rmap_remove(kvm, spte);
 654                         --kvm->stat.lpages;
 655                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 656                         write_protected = 1;
 657                 }
 658                 spte = rmap_next(kvm, rmapp, spte);
 659         }
 660
 661         if (write_protected)
 662                 kvm_flush_remote_tlbs(kvm);
 663
 664         account_shadowed(kvm, gfn);
 665 }
 666
 667 #ifdef MMU_DEBUG
 668 static int is_empty_shadow_page(u64 *spt)
 669 {
 670         u64 *pos;
 671         u64 *end;
 672
 673         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 674                 if (*pos != shadow_trap_nonpresent_pte) {
 675                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 676                                pos, *pos);
 677                         return 0;
 678                 }
 679         return 1;
 680 }
 681 #endif
 682
 683 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 684 {
 685         ASSERT(is_empty_shadow_page(sp->spt));
 686         list_del(&sp->link);
 687         __free_page(virt_to_page(sp->spt));
 688         __free_page(virt_to_page(sp->gfns));
 689         kfree(sp);
 690         ++kvm->arch.n_free_mmu_pages;
 691 }
 692
 693 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 694 {
 695         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 696 }
 697
 698 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 699                                                u64 *parent_pte)
 700 {
 701         struct kvm_mmu_page *sp;
 702
 703         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 704         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 705         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 706         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 707         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 708         ASSERT(is_empty_shadow_page(sp->spt));
 709         sp->slot_bitmap = 0;
 710         sp->multimapped = 0;
 711         sp->parent_pte = parent_pte;
 712         --vcpu->kvm->arch.n_free_mmu_pages;
 713         return sp;
 714 }
 715
 716 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 717                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 718 {
 719         struct kvm_pte_chain *pte_chain;
 720         struct hlist_node *node;
 721         int i;
 722
 723         if (!parent_pte)
 724                 return;
 725         if (!sp->multimapped) {
 726                 u64 *old = sp->parent_pte;
 727
 728                 if (!old) {
 729                         sp->parent_pte = parent_pte;
 730                         return;
 731                 }
 732                 sp->multimapped = 1;
 733                 pte_chain = mmu_alloc_pte_chain(vcpu);
 734                 INIT_HLIST_HEAD(&sp->parent_ptes);
 735                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 736                 pte_chain->parent_ptes[0] = old;
 737         }
 738         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 739                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 740                         continue;
 741                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 742                         if (!pte_chain->parent_ptes[i]) {
 743                                 pte_chain->parent_ptes[i] = parent_pte;
 744                                 return;
 745                         }
 746         }
 747         pte_chain = mmu_alloc_pte_chain(vcpu);
 748         BUG_ON(!pte_chain);
 749         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 750         pte_chain->parent_ptes[0] = parent_pte;
 751 }
 752
 753 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 754                                        u64 *parent_pte)
 755 {
 756         struct kvm_pte_chain *pte_chain;
 757         struct hlist_node *node;
 758         int i;
 759
 760         if (!sp->multimapped) {
 761                 BUG_ON(sp->parent_pte != parent_pte);
 762                 sp->parent_pte = NULL;
 763                 return;
 764         }
 765         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 766                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 767                         if (!pte_chain->parent_ptes[i])
 768                                 break;
 769                         if (pte_chain->parent_ptes[i] != parent_pte)
 770                                 continue;
 771                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 772                                 && pte_chain->parent_ptes[i + 1]) {
 773                                 pte_chain->parent_ptes[i]
 774                                         = pte_chain->parent_ptes[i + 1];
 775                                 ++i;
 776                         }
 777                         pte_chain->parent_ptes[i] = NULL;
 778                         if (i == 0) {
 779                                 hlist_del(&pte_chain->link);
 780                                 mmu_free_pte_chain(pte_chain);
 781                                 if (hlist_empty(&sp->parent_ptes)) {
 782                                         sp->multimapped = 0;
 783                                         sp->parent_pte = NULL;
 784                                 }
 785                         }
 786                         return;
 787                 }
 788         BUG();
 789 }
 790
 791 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
 792 {
 793         unsigned index;
 794         struct hlist_head *bucket;
 795         struct kvm_mmu_page *sp;
 796         struct hlist_node *node;
 797
 798         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 799         index = kvm_page_table_hashfn(gfn);
 800         bucket = &kvm->arch.mmu_page_hash[index];
 801         hlist_for_each_entry(sp, node, bucket, hash_link)
 802                 if (sp->gfn == gfn && !sp->role.metaphysical
 803                     && !sp->role.invalid) {
 804                         pgprintk("%s: found role %x\n",
 805                                  __func__, sp->role.word);
 806                         return sp;
 807                 }
 808         return NULL;
 809 }
 810
 811 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 812                                              gfn_t gfn,
 813                                              gva_t gaddr,
 814                                              unsigned level,
 815                                              int metaphysical,
 816                                              unsigned access,
 817                                              u64 *parent_pte)
 818 {
 819         union kvm_mmu_page_role role;
 820         unsigned index;
 821         unsigned quadrant;
 822         struct hlist_head *bucket;
 823         struct kvm_mmu_page *sp;
 824         struct hlist_node *node;
 825
 826         role.word = 0;
 827         role.glevels = vcpu->arch.mmu.root_level;
 828         role.level = level;
 829         role.metaphysical = metaphysical;
 830         role.access = access;
 831         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
 832                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 833                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 834                 role.quadrant = quadrant;
 835         }
 836         pgprintk("%s: looking gfn %lx role %x\n", __func__,
 837                  gfn, role.word);
 838         index = kvm_page_table_hashfn(gfn);
 839         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
 840         hlist_for_each_entry(sp, node, bucket, hash_link)
 841                 if (sp->gfn == gfn && sp->role.word == role.word) {
 842                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
 843                         pgprintk("%s: found\n", __func__);
 844                         return sp;
 845                 }
 846         ++vcpu->kvm->stat.mmu_cache_miss;
 847         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
 848         if (!sp)
 849                 return sp;
 850         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
 851         sp->gfn = gfn;
 852         sp->role = role;
 853         hlist_add_head(&sp->hash_link, bucket);
 854         vcpu->arch.mmu.prefetch_page(vcpu, sp);
 855         if (!metaphysical)
 856                 rmap_write_protect(vcpu->kvm, gfn);
 857         return sp;
 858 }
 859
 860 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 861                                          struct kvm_mmu_page *sp)
 862 {
 863         unsigned i;
 864         u64 *pt;
 865         u64 ent;
 866
 867         pt = sp->spt;
 868
 869         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
 870                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 871                         if (is_shadow_present_pte(pt[i]))
 872                                 rmap_remove(kvm, &pt[i]);
 873                         pt[i] = shadow_trap_nonpresent_pte;
 874                 }
 875                 kvm_flush_remote_tlbs(kvm);
 876                 return;
 877         }
 878
 879         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 880                 ent = pt[i];
 881
 882                 if (is_shadow_present_pte(ent)) {
 883                         if (!is_large_pte(ent)) {
 884                                 ent &= PT64_BASE_ADDR_MASK;
 885                                 mmu_page_remove_parent_pte(page_header(ent),
 886                                                            &pt[i]);
 887                         } else {
 888                                 --kvm->stat.lpages;
 889                                 rmap_remove(kvm, &pt[i]);
 890                         }
 891                 }
 892                 pt[i] = shadow_trap_nonpresent_pte;
 893         }
 894         kvm_flush_remote_tlbs(kvm);
 895 }
 896
 897 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
 898 {
 899         mmu_page_remove_parent_pte(sp, parent_pte);
 900 }
 901
 902 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 903 {
 904         int i;
 905
 906         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 907                 if (kvm->vcpus[i])
 908                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
 909 }
 910
 911 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 912 {
 913         u64 *parent_pte;
 914
 915         ++kvm->stat.mmu_shadow_zapped;
 916         while (sp->multimapped || sp->parent_pte) {
 917                 if (!sp->multimapped)
 918                         parent_pte = sp->parent_pte;
 919                 else {
 920                         struct kvm_pte_chain *chain;
 921
 922                         chain = container_of(sp->parent_ptes.first,
 923                                              struct kvm_pte_chain, link);
 924                         parent_pte = chain->parent_ptes[0];
 925                 }
 926                 BUG_ON(!parent_pte);
 927                 kvm_mmu_put_page(sp, parent_pte);
 928                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 929         }
 930         kvm_mmu_page_unlink_children(kvm, sp);
 931         if (!sp->root_count) {
 932                 if (!sp->role.metaphysical)
 933                         unaccount_shadowed(kvm, sp->gfn);
 934                 hlist_del(&sp->hash_link);
 935                 kvm_mmu_free_page(kvm, sp);
 936         } else {
 937                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
 938                 sp->role.invalid = 1;
 939                 kvm_reload_remote_mmus(kvm);
 940         }
 941         kvm_mmu_reset_last_pte_updated(kvm);
 942 }
 943
 944 /*
 945  * Changing the number of mmu pages allocated to the vm
 946  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 947  */
 948 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 949 {
 950         /*
 951          * If we set the number of mmu pages to be smaller be than the
 952          * number of actived pages , we must to free some mmu pages before we
 953          * change the value
 954          */
 955
 956         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
 957             kvm_nr_mmu_pages) {
 958                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
 959                                        - kvm->arch.n_free_mmu_pages;
 960
 961                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 962                         struct kvm_mmu_page *page;
 963
 964                         page = container_of(kvm->arch.active_mmu_pages.prev,
 965                                             struct kvm_mmu_page, link);
 966                         kvm_mmu_zap_page(kvm, page);
 967                         n_used_mmu_pages--;
 968                 }
 969                 kvm->arch.n_free_mmu_pages = 0;
 970         }
 971         else
 972                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
 973                                          - kvm->arch.n_alloc_mmu_pages;
 974
 975         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
 976 }
 977
 978 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 979 {
 980         unsigned index;
 981         struct hlist_head *bucket;
 982         struct kvm_mmu_page *sp;
 983         struct hlist_node *node, *n;
 984         int r;
 985
 986         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 987         r = 0;
 988         index = kvm_page_table_hashfn(gfn);
 989         bucket = &kvm->arch.mmu_page_hash[index];
 990         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
 991                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 992                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
 993                                  sp->role.word);
 994                         kvm_mmu_zap_page(kvm, sp);
 995                         r = 1;
 996                 }
 997         return r;
 998 }
 999
1000 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1001 {
1002         struct kvm_mmu_page *sp;
1003
1004         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1005                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1006                 kvm_mmu_zap_page(kvm, sp);
1007         }
1008 }
1009
1010 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1011 {
1012         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1013         struct kvm_mmu_page *sp = page_header(__pa(pte));
1014
1015         __set_bit(slot, &sp->slot_bitmap);
1016 }
1017
1018 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1019 {
1020         struct page *page;
1021
1022         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1023
1024         if (gpa == UNMAPPED_GVA)
1025                 return NULL;
1026
1027         down_read(&current->mm->mmap_sem);
1028         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1029         up_read(&current->mm->mmap_sem);
1030
1031         return page;
1032 }
1033
1034 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1035                          unsigned pt_access, unsigned pte_access,
1036                          int user_fault, int write_fault, int dirty,
1037                          int *ptwrite, int largepage, gfn_t gfn,
1038                          struct page *page, bool speculative)
1039 {
1040         u64 spte;
1041         int was_rmapped = 0;
1042         int was_writeble = is_writeble_pte(*shadow_pte);
1043
1044         pgprintk("%s: spte %llx access %x write_fault %d"
1045                  " user_fault %d gfn %lx\n",
1046                  __func__, *shadow_pte, pt_access,
1047                  write_fault, user_fault, gfn);
1048
1049         if (is_rmap_pte(*shadow_pte)) {
1050                 /*
1051                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1052                  * the parent of the now unreachable PTE.
1053                  */
1054                 if (largepage && !is_large_pte(*shadow_pte)) {
1055                         struct kvm_mmu_page *child;
1056                         u64 pte = *shadow_pte;
1057
1058                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1059                         mmu_page_remove_parent_pte(child, shadow_pte);
1060                 } else if (page != spte_to_page(*shadow_pte)) {
1061                         pgprintk("hfn old %lx new %lx\n",
1062                                  page_to_pfn(spte_to_page(*shadow_pte)),
1063                                  page_to_pfn(page));
1064                         rmap_remove(vcpu->kvm, shadow_pte);
1065                 } else {
1066                         if (largepage)
1067                                 was_rmapped = is_large_pte(*shadow_pte);
1068                         else
1069                                 was_rmapped = 1;
1070                 }
1071         }
1072
1073         /*
1074          * We don't set the accessed bit, since we sometimes want to see
1075          * whether the guest actually used the pte (in order to detect
1076          * demand paging).
1077          */
1078         spte = PT_PRESENT_MASK | PT_DIRTY_MASK;
1079         if (!speculative)
1080                 pte_access |= PT_ACCESSED_MASK;
1081         if (!dirty)
1082                 pte_access &= ~ACC_WRITE_MASK;
1083         if (!(pte_access & ACC_EXEC_MASK))
1084                 spte |= PT64_NX_MASK;
1085
1086         spte |= PT_PRESENT_MASK;
1087         if (pte_access & ACC_USER_MASK)
1088                 spte |= PT_USER_MASK;
1089         if (largepage)
1090                 spte |= PT_PAGE_SIZE_MASK;
1091
1092         spte |= page_to_phys(page);
1093
1094         if ((pte_access & ACC_WRITE_MASK)
1095             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1096                 struct kvm_mmu_page *shadow;
1097
1098                 spte |= PT_WRITABLE_MASK;
1099                 if (user_fault) {
1100                         mmu_unshadow(vcpu->kvm, gfn);
1101                         goto unshadowed;
1102                 }
1103
1104                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1105                 if (shadow ||
1106                    (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
1107                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1108                                  __func__, gfn);
1109                         pte_access &= ~ACC_WRITE_MASK;
1110                         if (is_writeble_pte(spte)) {
1111                                 spte &= ~PT_WRITABLE_MASK;
1112                                 kvm_x86_ops->tlb_flush(vcpu);
1113                         }
1114                         if (write_fault)
1115                                 *ptwrite = 1;
1116                 }
1117         }
1118
1119 unshadowed:
1120
1121         if (pte_access & ACC_WRITE_MASK)
1122                 mark_page_dirty(vcpu->kvm, gfn);
1123
1124         pgprintk("%s: setting spte %llx\n", __func__, spte);
1125         pgprintk("instantiating %s PTE (%s) at %d (%llx) addr %llx\n",
1126                  (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
1127                  (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
1128         set_shadow_pte(shadow_pte, spte);
1129         if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
1130             && (spte & PT_PRESENT_MASK))
1131                 ++vcpu->kvm->stat.lpages;
1132
1133         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1134         if (!was_rmapped) {
1135                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1136                 if (!is_rmap_pte(*shadow_pte))
1137                         kvm_release_page_clean(page);
1138         } else {
1139                 if (was_writeble)
1140                         kvm_release_page_dirty(page);
1141                 else
1142                         kvm_release_page_clean(page);
1143         }
1144         if (!ptwrite || !*ptwrite)
1145                 vcpu->arch.last_pte_updated = shadow_pte;
1146 }
1147
1148 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1149 {
1150 }
1151
1152 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1153                            int largepage, gfn_t gfn, struct page *page,
1154                            int level)
1155 {
1156         hpa_t table_addr = vcpu->arch.mmu.root_hpa;
1157         int pt_write = 0;
1158
1159         for (; ; level--) {
1160                 u32 index = PT64_INDEX(v, level);
1161                 u64 *table;
1162
1163                 ASSERT(VALID_PAGE(table_addr));
1164                 table = __va(table_addr);
1165
1166                 if (level == 1) {
1167                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1168                                      0, write, 1, &pt_write, 0, gfn, page, false);
1169                         return pt_write;
1170                 }
1171
1172                 if (largepage && level == 2) {
1173                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1174                                      0, write, 1, &pt_write, 1, gfn, page, false);
1175                         return pt_write;
1176                 }
1177
1178                 if (table[index] == shadow_trap_nonpresent_pte) {
1179                         struct kvm_mmu_page *new_table;
1180                         gfn_t pseudo_gfn;
1181
1182                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1183                                 >> PAGE_SHIFT;
1184                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1185                                                      v, level - 1,
1186                                                      1, ACC_ALL, &table[index]);
1187                         if (!new_table) {
1188                                 pgprintk("nonpaging_map: ENOMEM\n");
1189                                 kvm_release_page_clean(page);
1190                                 return -ENOMEM;
1191                         }
1192
1193                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
1194                                 | PT_WRITABLE_MASK | PT_USER_MASK;
1195                 }
1196                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1197         }
1198 }
1199
1200 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1201 {
1202         int r;
1203         int largepage = 0;
1204
1205         struct page *page;
1206
1207         down_read(&vcpu->kvm->slots_lock);
1208
1209         down_read(&current->mm->mmap_sem);
1210         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1211                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1212                 largepage = 1;
1213         }
1214
1215         page = gfn_to_page(vcpu->kvm, gfn);
1216         up_read(&current->mm->mmap_sem);
1217
1218         /* mmio */
1219         if (is_error_page(page)) {
1220                 kvm_release_page_clean(page);
1221                 up_read(&vcpu->kvm->slots_lock);
1222                 return 1;
1223         }
1224
1225         spin_lock(&vcpu->kvm->mmu_lock);
1226         kvm_mmu_free_some_pages(vcpu);
1227         r = __direct_map(vcpu, v, write, largepage, gfn, page,
1228                          PT32E_ROOT_LEVEL);
1229         spin_unlock(&vcpu->kvm->mmu_lock);
1230
1231         up_read(&vcpu->kvm->slots_lock);
1232
1233         return r;
1234 }
1235
1236
1237 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1238                                     struct kvm_mmu_page *sp)
1239 {
1240         int i;
1241
1242         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1243                 sp->spt[i] = shadow_trap_nonpresent_pte;
1244 }
1245
1246 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1247 {
1248         int i;
1249         struct kvm_mmu_page *sp;
1250
1251         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1252                 return;
1253         spin_lock(&vcpu->kvm->mmu_lock);
1254 #ifdef CONFIG_X86_64
1255         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1256                 hpa_t root = vcpu->arch.mmu.root_hpa;
1257
1258                 sp = page_header(root);
1259                 --sp->root_count;
1260                 if (!sp->root_count && sp->role.invalid)
1261                         kvm_mmu_zap_page(vcpu->kvm, sp);
1262                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1263                 spin_unlock(&vcpu->kvm->mmu_lock);
1264                 return;
1265         }
1266 #endif
1267         for (i = 0; i < 4; ++i) {
1268                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1269
1270                 if (root) {
1271                         root &= PT64_BASE_ADDR_MASK;
1272                         sp = page_header(root);
1273                         --sp->root_count;
1274                         if (!sp->root_count && sp->role.invalid)
1275                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1276                 }
1277                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1278         }
1279         spin_unlock(&vcpu->kvm->mmu_lock);
1280         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1281 }
1282
1283 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1284 {
1285         int i;
1286         gfn_t root_gfn;
1287         struct kvm_mmu_page *sp;
1288         int metaphysical = 0;
1289
1290         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1291
1292 #ifdef CONFIG_X86_64
1293         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1294                 hpa_t root = vcpu->arch.mmu.root_hpa;
1295
1296                 ASSERT(!VALID_PAGE(root));
1297                 if (tdp_enabled)
1298                         metaphysical = 1;
1299                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1300                                       PT64_ROOT_LEVEL, metaphysical,
1301                                       ACC_ALL, NULL);
1302                 root = __pa(sp->spt);
1303                 ++sp->root_count;
1304                 vcpu->arch.mmu.root_hpa = root;
1305                 return;
1306         }
1307 #endif
1308         metaphysical = !is_paging(vcpu);
1309         if (tdp_enabled)
1310                 metaphysical = 1;
1311         for (i = 0; i < 4; ++i) {
1312                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1313
1314                 ASSERT(!VALID_PAGE(root));
1315                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1316                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1317                                 vcpu->arch.mmu.pae_root[i] = 0;
1318                                 continue;
1319                         }
1320                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1321                 } else if (vcpu->arch.mmu.root_level == 0)
1322                         root_gfn = 0;
1323                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1324                                       PT32_ROOT_LEVEL, metaphysical,
1325                                       ACC_ALL, NULL);
1326                 root = __pa(sp->spt);
1327                 ++sp->root_count;
1328                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1329         }
1330         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1331 }
1332
1333 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1334 {
1335         return vaddr;
1336 }
1337
1338 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1339                                 u32 error_code)
1340 {
1341         gfn_t gfn;
1342         int r;
1343
1344         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1345         r = mmu_topup_memory_caches(vcpu);
1346         if (r)
1347                 return r;
1348
1349         ASSERT(vcpu);
1350         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1351
1352         gfn = gva >> PAGE_SHIFT;
1353
1354         return nonpaging_map(vcpu, gva & PAGE_MASK,
1355                              error_code & PFERR_WRITE_MASK, gfn);
1356 }
1357
1358 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1359                                 u32 error_code)
1360 {
1361         struct page *page;
1362         int r;
1363         int largepage = 0;
1364         gfn_t gfn = gpa >> PAGE_SHIFT;
1365
1366         ASSERT(vcpu);
1367         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1368
1369         r = mmu_topup_memory_caches(vcpu);
1370         if (r)
1371                 return r;
1372
1373         down_read(&current->mm->mmap_sem);
1374         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1375                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1376                 largepage = 1;
1377         }
1378         page = gfn_to_page(vcpu->kvm, gfn);
1379         if (is_error_page(page)) {
1380                 kvm_release_page_clean(page);
1381                 up_read(&current->mm->mmap_sem);
1382                 return 1;
1383         }
1384         spin_lock(&vcpu->kvm->mmu_lock);
1385         kvm_mmu_free_some_pages(vcpu);
1386         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1387                          largepage, gfn, page, TDP_ROOT_LEVEL);
1388         spin_unlock(&vcpu->kvm->mmu_lock);
1389         up_read(&current->mm->mmap_sem);
1390
1391         return r;
1392 }
1393
1394 static void nonpaging_free(struct kvm_vcpu *vcpu)
1395 {
1396         mmu_free_roots(vcpu);
1397 }
1398
1399 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1400 {
1401         struct kvm_mmu *context = &vcpu->arch.mmu;
1402
1403         context->new_cr3 = nonpaging_new_cr3;
1404         context->page_fault = nonpaging_page_fault;
1405         context->gva_to_gpa = nonpaging_gva_to_gpa;
1406         context->free = nonpaging_free;
1407         context->prefetch_page = nonpaging_prefetch_page;
1408         context->root_level = 0;
1409         context->shadow_root_level = PT32E_ROOT_LEVEL;
1410         context->root_hpa = INVALID_PAGE;
1411         return 0;
1412 }
1413
1414 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1415 {
1416         ++vcpu->stat.tlb_flush;
1417         kvm_x86_ops->tlb_flush(vcpu);
1418 }
1419
1420 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1421 {
1422         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1423         mmu_free_roots(vcpu);
1424 }
1425
1426 static void inject_page_fault(struct kvm_vcpu *vcpu,
1427                               u64 addr,
1428                               u32 err_code)
1429 {
1430         kvm_inject_page_fault(vcpu, addr, err_code);
1431 }
1432
1433 static void paging_free(struct kvm_vcpu *vcpu)
1434 {
1435         nonpaging_free(vcpu);
1436 }
1437
1438 #define PTTYPE 64
1439 #include "paging_tmpl.h"
1440 #undef PTTYPE
1441
1442 #define PTTYPE 32
1443 #include "paging_tmpl.h"
1444 #undef PTTYPE
1445
1446 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1447 {
1448         struct kvm_mmu *context = &vcpu->arch.mmu;
1449
1450         ASSERT(is_pae(vcpu));
1451         context->new_cr3 = paging_new_cr3;
1452         context->page_fault = paging64_page_fault;
1453         context->gva_to_gpa = paging64_gva_to_gpa;
1454         context->prefetch_page = paging64_prefetch_page;
1455         context->free = paging_free;
1456         context->root_level = level;
1457         context->shadow_root_level = level;
1458         context->root_hpa = INVALID_PAGE;
1459         return 0;
1460 }
1461
1462 static int paging64_init_context(struct kvm_vcpu *vcpu)
1463 {
1464         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1465 }
1466
1467 static int paging32_init_context(struct kvm_vcpu *vcpu)
1468 {
1469         struct kvm_mmu *context = &vcpu->arch.mmu;
1470
1471         context->new_cr3 = paging_new_cr3;
1472         context->page_fault = paging32_page_fault;
1473         context->gva_to_gpa = paging32_gva_to_gpa;
1474         context->free = paging_free;
1475         context->prefetch_page = paging32_prefetch_page;
1476         context->root_level = PT32_ROOT_LEVEL;
1477         context->shadow_root_level = PT32E_ROOT_LEVEL;
1478         context->root_hpa = INVALID_PAGE;
1479         return 0;
1480 }
1481
1482 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1483 {
1484         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1485 }
1486
1487 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1488 {
1489         struct kvm_mmu *context = &vcpu->arch.mmu;
1490
1491         context->new_cr3 = nonpaging_new_cr3;
1492         context->page_fault = tdp_page_fault;
1493         context->free = nonpaging_free;
1494         context->prefetch_page = nonpaging_prefetch_page;
1495         context->shadow_root_level = TDP_ROOT_LEVEL;
1496         context->root_hpa = INVALID_PAGE;
1497
1498         if (!is_paging(vcpu)) {
1499                 context->gva_to_gpa = nonpaging_gva_to_gpa;
1500                 context->root_level = 0;
1501         } else if (is_long_mode(vcpu)) {
1502                 context->gva_to_gpa = paging64_gva_to_gpa;
1503                 context->root_level = PT64_ROOT_LEVEL;
1504         } else if (is_pae(vcpu)) {
1505                 context->gva_to_gpa = paging64_gva_to_gpa;
1506                 context->root_level = PT32E_ROOT_LEVEL;
1507         } else {
1508                 context->gva_to_gpa = paging32_gva_to_gpa;
1509                 context->root_level = PT32_ROOT_LEVEL;
1510         }
1511
1512         return 0;
1513 }
1514
1515 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1516 {
1517         ASSERT(vcpu);
1518         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1519
1520         if (!is_paging(vcpu))
1521                 return nonpaging_init_context(vcpu);
1522         else if (is_long_mode(vcpu))
1523                 return paging64_init_context(vcpu);
1524         else if (is_pae(vcpu))
1525                 return paging32E_init_context(vcpu);
1526         else
1527                 return paging32_init_context(vcpu);
1528 }
1529
1530 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1531 {
1532         if (tdp_enabled)
1533                 return init_kvm_tdp_mmu(vcpu);
1534         else
1535                 return init_kvm_softmmu(vcpu);
1536 }
1537
1538 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1539 {
1540         ASSERT(vcpu);
1541         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1542                 vcpu->arch.mmu.free(vcpu);
1543                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1544         }
1545 }
1546
1547 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1548 {
1549         destroy_kvm_mmu(vcpu);
1550         return init_kvm_mmu(vcpu);
1551 }
1552 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1553
1554 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1555 {
1556         int r;
1557
1558         r = mmu_topup_memory_caches(vcpu);
1559         if (r)
1560                 goto out;
1561         spin_lock(&vcpu->kvm->mmu_lock);
1562         kvm_mmu_free_some_pages(vcpu);
1563         mmu_alloc_roots(vcpu);
1564         spin_unlock(&vcpu->kvm->mmu_lock);
1565         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1566         kvm_mmu_flush_tlb(vcpu);
1567 out:
1568         return r;
1569 }
1570 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1571
1572 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1573 {
1574         mmu_free_roots(vcpu);
1575 }
1576
1577 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1578                                   struct kvm_mmu_page *sp,
1579                                   u64 *spte)
1580 {
1581         u64 pte;
1582         struct kvm_mmu_page *child;
1583
1584         pte = *spte;
1585         if (is_shadow_present_pte(pte)) {
1586                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1587                     is_large_pte(pte))
1588                         rmap_remove(vcpu->kvm, spte);
1589                 else {
1590                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1591                         mmu_page_remove_parent_pte(child, spte);
1592                 }
1593         }
1594         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1595         if (is_large_pte(pte))
1596                 --vcpu->kvm->stat.lpages;
1597 }
1598
1599 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1600                                   struct kvm_mmu_page *sp,
1601                                   u64 *spte,
1602                                   const void *new)
1603 {
1604         if ((sp->role.level != PT_PAGE_TABLE_LEVEL)
1605             && !vcpu->arch.update_pte.largepage) {
1606                 ++vcpu->kvm->stat.mmu_pde_zapped;
1607                 return;
1608         }
1609
1610         ++vcpu->kvm->stat.mmu_pte_updated;
1611         if (sp->role.glevels == PT32_ROOT_LEVEL)
1612                 paging32_update_pte(vcpu, sp, spte, new);
1613         else
1614                 paging64_update_pte(vcpu, sp, spte, new);
1615 }
1616
1617 static bool need_remote_flush(u64 old, u64 new)
1618 {
1619         if (!is_shadow_present_pte(old))
1620                 return false;
1621         if (!is_shadow_present_pte(new))
1622                 return true;
1623         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1624                 return true;
1625         old ^= PT64_NX_MASK;
1626         new ^= PT64_NX_MASK;
1627         return (old & ~new & PT64_PERM_MASK) != 0;
1628 }
1629
1630 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1631 {
1632         if (need_remote_flush(old, new))
1633                 kvm_flush_remote_tlbs(vcpu->kvm);
1634         else
1635                 kvm_mmu_flush_tlb(vcpu);
1636 }
1637
1638 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1639 {
1640         u64 *spte = vcpu->arch.last_pte_updated;
1641
1642         return !!(spte && (*spte & PT_ACCESSED_MASK));
1643 }
1644
1645 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1646                                           const u8 *new, int bytes)
1647 {
1648         gfn_t gfn;
1649         int r;
1650         u64 gpte = 0;
1651         struct page *page;
1652
1653         vcpu->arch.update_pte.largepage = 0;
1654
1655         if (bytes != 4 && bytes != 8)
1656                 return;
1657
1658         /*
1659          * Assume that the pte write on a page table of the same type
1660          * as the current vcpu paging mode.  This is nearly always true
1661          * (might be false while changing modes).  Note it is verified later
1662          * by update_pte().
1663          */
1664         if (is_pae(vcpu)) {
1665                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1666                 if ((bytes == 4) && (gpa % 4 == 0)) {
1667                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1668                         if (r)
1669                                 return;
1670                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1671                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1672                         memcpy((void *)&gpte, new, 8);
1673                 }
1674         } else {
1675                 if ((bytes == 4) && (gpa % 4 == 0))
1676                         memcpy((void *)&gpte, new, 4);
1677         }
1678         if (!is_present_pte(gpte))
1679                 return;
1680         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1681
1682         down_read(&current->mm->mmap_sem);
1683         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1684                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1685                 vcpu->arch.update_pte.largepage = 1;
1686         }
1687         page = gfn_to_page(vcpu->kvm, gfn);
1688         up_read(&current->mm->mmap_sem);
1689
1690         if (is_error_page(page)) {
1691                 kvm_release_page_clean(page);
1692                 return;
1693         }
1694         vcpu->arch.update_pte.gfn = gfn;
1695         vcpu->arch.update_pte.page = page;
1696 }
1697
1698 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1699                        const u8 *new, int bytes)
1700 {
1701         gfn_t gfn = gpa >> PAGE_SHIFT;
1702         struct kvm_mmu_page *sp;
1703         struct hlist_node *node, *n;
1704         struct hlist_head *bucket;
1705         unsigned index;
1706         u64 entry, gentry;
1707         u64 *spte;
1708         unsigned offset = offset_in_page(gpa);
1709         unsigned pte_size;
1710         unsigned page_offset;
1711         unsigned misaligned;
1712         unsigned quadrant;
1713         int level;
1714         int flooded = 0;
1715         int npte;
1716         int r;
1717
1718         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1719         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1720         spin_lock(&vcpu->kvm->mmu_lock);
1721         kvm_mmu_free_some_pages(vcpu);
1722         ++vcpu->kvm->stat.mmu_pte_write;
1723         kvm_mmu_audit(vcpu, "pre pte write");
1724         if (gfn == vcpu->arch.last_pt_write_gfn
1725             && !last_updated_pte_accessed(vcpu)) {
1726                 ++vcpu->arch.last_pt_write_count;
1727                 if (vcpu->arch.last_pt_write_count >= 3)
1728                         flooded = 1;
1729         } else {
1730                 vcpu->arch.last_pt_write_gfn = gfn;
1731                 vcpu->arch.last_pt_write_count = 1;
1732                 vcpu->arch.last_pte_updated = NULL;
1733         }
1734         index = kvm_page_table_hashfn(gfn);
1735         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1736         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1737                 if (sp->gfn != gfn || sp->role.metaphysical)
1738                         continue;
1739                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1740                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1741                 misaligned |= bytes < 4;
1742                 if (misaligned || flooded) {
1743                         /*
1744                          * Misaligned accesses are too much trouble to fix
1745                          * up; also, they usually indicate a page is not used
1746                          * as a page table.
1747                          *
1748                          * If we're seeing too many writes to a page,
1749                          * it may no longer be a page table, or we may be
1750                          * forking, in which case it is better to unmap the
1751                          * page.
1752                          */
1753                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1754                                  gpa, bytes, sp->role.word);
1755                         kvm_mmu_zap_page(vcpu->kvm, sp);
1756                         ++vcpu->kvm->stat.mmu_flooded;
1757                         continue;
1758                 }
1759                 page_offset = offset;
1760                 level = sp->role.level;
1761                 npte = 1;
1762                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1763                         page_offset <<= 1;      /* 32->64 */
1764                         /*
1765                          * A 32-bit pde maps 4MB while the shadow pdes map
1766                          * only 2MB.  So we need to double the offset again
1767                          * and zap two pdes instead of one.
1768                          */
1769                         if (level == PT32_ROOT_LEVEL) {
1770                                 page_offset &= ~7; /* kill rounding error */
1771                                 page_offset <<= 1;
1772                                 npte = 2;
1773                         }
1774                         quadrant = page_offset >> PAGE_SHIFT;
1775                         page_offset &= ~PAGE_MASK;
1776                         if (quadrant != sp->role.quadrant)
1777                                 continue;
1778                 }
1779                 spte = &sp->spt[page_offset / sizeof(*spte)];
1780                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1781                         gentry = 0;
1782                         r = kvm_read_guest_atomic(vcpu->kvm,
1783                                                   gpa & ~(u64)(pte_size - 1),
1784                                                   &gentry, pte_size);
1785                         new = (const void *)&gentry;
1786                         if (r < 0)
1787                                 new = NULL;
1788                 }
1789                 while (npte--) {
1790                         entry = *spte;
1791                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1792                         if (new)
1793                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1794                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1795                         ++spte;
1796                 }
1797         }
1798         kvm_mmu_audit(vcpu, "post pte write");
1799         spin_unlock(&vcpu->kvm->mmu_lock);
1800         if (vcpu->arch.update_pte.page) {
1801                 kvm_release_page_clean(vcpu->arch.update_pte.page);
1802                 vcpu->arch.update_pte.page = NULL;
1803         }
1804 }
1805
1806 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1807 {
1808         gpa_t gpa;
1809         int r;
1810
1811         down_read(&vcpu->kvm->slots_lock);
1812         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1813         up_read(&vcpu->kvm->slots_lock);
1814
1815         spin_lock(&vcpu->kvm->mmu_lock);
1816         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1817         spin_unlock(&vcpu->kvm->mmu_lock);
1818         return r;
1819 }
1820
1821 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1822 {
1823         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1824                 struct kvm_mmu_page *sp;
1825
1826                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1827                                   struct kvm_mmu_page, link);
1828                 kvm_mmu_zap_page(vcpu->kvm, sp);
1829                 ++vcpu->kvm->stat.mmu_recycled;
1830         }
1831 }
1832
1833 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1834 {
1835         int r;
1836         enum emulation_result er;
1837
1838         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1839         if (r < 0)
1840                 goto out;
1841
1842         if (!r) {
1843                 r = 1;
1844                 goto out;
1845         }
1846
1847         r = mmu_topup_memory_caches(vcpu);
1848         if (r)
1849                 goto out;
1850
1851         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1852
1853         switch (er) {
1854         case EMULATE_DONE:
1855                 return 1;
1856         case EMULATE_DO_MMIO:
1857                 ++vcpu->stat.mmio_exits;
1858                 return 0;
1859         case EMULATE_FAIL:
1860                 kvm_report_emulation_failure(vcpu, "pagetable");
1861                 return 1;
1862         default:
1863                 BUG();
1864         }
1865 out:
1866         return r;
1867 }
1868 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1869
1870 void kvm_enable_tdp(void)
1871 {
1872         tdp_enabled = true;
1873 }
1874 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1875
1876 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1877 {
1878         struct kvm_mmu_page *sp;
1879
1880         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1881                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1882                                   struct kvm_mmu_page, link);
1883                 kvm_mmu_zap_page(vcpu->kvm, sp);
1884         }
1885         free_page((unsigned long)vcpu->arch.mmu.pae_root);
1886 }
1887
1888 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1889 {
1890         struct page *page;
1891         int i;
1892
1893         ASSERT(vcpu);
1894
1895         if (vcpu->kvm->arch.n_requested_mmu_pages)
1896                 vcpu->kvm->arch.n_free_mmu_pages =
1897                                         vcpu->kvm->arch.n_requested_mmu_pages;
1898         else
1899                 vcpu->kvm->arch.n_free_mmu_pages =
1900                                         vcpu->kvm->arch.n_alloc_mmu_pages;
1901         /*
1902          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1903          * Therefore we need to allocate shadow page tables in the first
1904          * 4GB of memory, which happens to fit the DMA32 zone.
1905          */
1906         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1907         if (!page)
1908                 goto error_1;
1909         vcpu->arch.mmu.pae_root = page_address(page);
1910         for (i = 0; i < 4; ++i)
1911                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1912
1913         return 0;
1914
1915 error_1:
1916         free_mmu_pages(vcpu);
1917         return -ENOMEM;
1918 }
1919
1920 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1921 {
1922         ASSERT(vcpu);
1923         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1924
1925         return alloc_mmu_pages(vcpu);
1926 }
1927
1928 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1929 {
1930         ASSERT(vcpu);
1931         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1932
1933         return init_kvm_mmu(vcpu);
1934 }
1935
1936 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1937 {
1938         ASSERT(vcpu);
1939
1940         destroy_kvm_mmu(vcpu);
1941         free_mmu_pages(vcpu);
1942         mmu_free_memory_caches(vcpu);
1943 }
1944
1945 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1946 {
1947         struct kvm_mmu_page *sp;
1948
1949         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1950                 int i;
1951                 u64 *pt;
1952
1953                 if (!test_bit(slot, &sp->slot_bitmap))
1954                         continue;
1955
1956                 pt = sp->spt;
1957                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1958                         /* avoid RMW */
1959                         if (pt[i] & PT_WRITABLE_MASK)
1960                                 pt[i] &= ~PT_WRITABLE_MASK;
1961         }
1962 }
1963
1964 void kvm_mmu_zap_all(struct kvm *kvm)
1965 {
1966         struct kvm_mmu_page *sp, *node;
1967
1968         spin_lock(&kvm->mmu_lock);
1969         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1970                 kvm_mmu_zap_page(kvm, sp);
1971         spin_unlock(&kvm->mmu_lock);
1972
1973         kvm_flush_remote_tlbs(kvm);
1974 }
1975
1976 void kvm_mmu_module_exit(void)
1977 {
1978         if (pte_chain_cache)
1979                 kmem_cache_destroy(pte_chain_cache);
1980         if (rmap_desc_cache)
1981                 kmem_cache_destroy(rmap_desc_cache);
1982         if (mmu_page_header_cache)
1983                 kmem_cache_destroy(mmu_page_header_cache);
1984 }
1985
1986 int kvm_mmu_module_init(void)
1987 {
1988         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1989                                             sizeof(struct kvm_pte_chain),
1990                                             0, 0, NULL);
1991         if (!pte_chain_cache)
1992                 goto nomem;
1993         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1994                                             sizeof(struct kvm_rmap_desc),
1995                                             0, 0, NULL);
1996         if (!rmap_desc_cache)
1997                 goto nomem;
1998
1999         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2000                                                   sizeof(struct kvm_mmu_page),
2001                                                   0, 0, NULL);
2002         if (!mmu_page_header_cache)
2003                 goto nomem;
2004
2005         return 0;
2006
2007 nomem:
2008         kvm_mmu_module_exit();
2009         return -ENOMEM;
2010 }
2011
2012 /*
2013  * Caculate mmu pages needed for kvm.
2014  */
2015 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2016 {
2017         int i;
2018         unsigned int nr_mmu_pages;
2019         unsigned int  nr_pages = 0;
2020
2021         for (i = 0; i < kvm->nmemslots; i++)
2022                 nr_pages += kvm->memslots[i].npages;
2023
2024         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2025         nr_mmu_pages = max(nr_mmu_pages,
2026                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2027
2028         return nr_mmu_pages;
2029 }
2030
2031 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2032                                 unsigned len)
2033 {
2034         if (len > buffer->len)
2035                 return NULL;
2036         return buffer->ptr;
2037 }
2038
2039 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2040                                 unsigned len)
2041 {
2042         void *ret;
2043
2044         ret = pv_mmu_peek_buffer(buffer, len);
2045         if (!ret)
2046                 return ret;
2047         buffer->ptr += len;
2048         buffer->len -= len;
2049         buffer->processed += len;
2050         return ret;
2051 }
2052
2053 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2054                              gpa_t addr, gpa_t value)
2055 {
2056         int bytes = 8;
2057         int r;
2058
2059         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2060                 bytes = 4;
2061
2062         r = mmu_topup_memory_caches(vcpu);
2063         if (r)
2064                 return r;
2065
2066         if (!__emulator_write_phys(vcpu, addr, &value, bytes))
2067                 return -EFAULT;
2068
2069         return 1;
2070 }
2071
2072 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2073 {
2074         kvm_x86_ops->tlb_flush(vcpu);
2075         return 1;
2076 }
2077
2078 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2079 {
2080         spin_lock(&vcpu->kvm->mmu_lock);
2081         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2082         spin_unlock(&vcpu->kvm->mmu_lock);
2083         return 1;
2084 }
2085
2086 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2087                              struct kvm_pv_mmu_op_buffer *buffer)
2088 {
2089         struct kvm_mmu_op_header *header;
2090
2091         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2092         if (!header)
2093                 return 0;
2094         switch (header->op) {
2095         case KVM_MMU_OP_WRITE_PTE: {
2096                 struct kvm_mmu_op_write_pte *wpte;
2097
2098                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2099                 if (!wpte)
2100                         return 0;
2101                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2102                                         wpte->pte_val);
2103         }
2104         case KVM_MMU_OP_FLUSH_TLB: {
2105                 struct kvm_mmu_op_flush_tlb *ftlb;
2106
2107                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2108                 if (!ftlb)
2109                         return 0;
2110                 return kvm_pv_mmu_flush_tlb(vcpu);
2111         }
2112         case KVM_MMU_OP_RELEASE_PT: {
2113                 struct kvm_mmu_op_release_pt *rpt;
2114
2115                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2116                 if (!rpt)
2117                         return 0;
2118                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2119         }
2120         default: return 0;
2121         }
2122 }
2123
2124 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2125                   gpa_t addr, unsigned long *ret)
2126 {
2127         int r;
2128         struct kvm_pv_mmu_op_buffer buffer;
2129
2130         down_read(&vcpu->kvm->slots_lock);
2131         down_read(&current->mm->mmap_sem);
2132
2133         buffer.ptr = buffer.buf;
2134         buffer.len = min_t(unsigned long, bytes, sizeof buffer.buf);
2135         buffer.processed = 0;
2136
2137         r = kvm_read_guest(vcpu->kvm, addr, buffer.buf, buffer.len);
2138         if (r)
2139                 goto out;
2140
2141         while (buffer.len) {
2142                 r = kvm_pv_mmu_op_one(vcpu, &buffer);
2143                 if (r < 0)
2144                         goto out;
2145                 if (r == 0)
2146                         break;
2147         }
2148
2149         r = 1;
2150 out:
2151         *ret = buffer.processed;
2152         up_read(&current->mm->mmap_sem);
2153         up_read(&vcpu->kvm->slots_lock);
2154         return r;
2155 }
2156
2157 #ifdef AUDIT
2158
2159 static const char *audit_msg;
2160
2161 static gva_t canonicalize(gva_t gva)
2162 {
2163 #ifdef CONFIG_X86_64
2164         gva = (long long)(gva << 16) >> 16;
2165 #endif
2166         return gva;
2167 }
2168
2169 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2170                                 gva_t va, int level)
2171 {
2172         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2173         int i;
2174         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2175
2176         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2177                 u64 ent = pt[i];
2178
2179                 if (ent == shadow_trap_nonpresent_pte)
2180                         continue;
2181
2182                 va = canonicalize(va);
2183                 if (level > 1) {
2184                         if (ent == shadow_notrap_nonpresent_pte)
2185                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2186                                        " in nonleaf level: levels %d gva %lx"
2187                                        " level %d pte %llx\n", audit_msg,
2188                                        vcpu->arch.mmu.root_level, va, level, ent);
2189
2190                         audit_mappings_page(vcpu, ent, va, level - 1);
2191                 } else {
2192                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2193                         struct page *page = gpa_to_page(vcpu, gpa);
2194                         hpa_t hpa = page_to_phys(page);
2195
2196                         if (is_shadow_present_pte(ent)
2197                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2198                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2199                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2200                                        audit_msg, vcpu->arch.mmu.root_level,
2201                                        va, gpa, hpa, ent,
2202                                        is_shadow_present_pte(ent));
2203                         else if (ent == shadow_notrap_nonpresent_pte
2204                                  && !is_error_hpa(hpa))
2205                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2206                                        " valid guest gva %lx\n", audit_msg, va);
2207                         kvm_release_page_clean(page);
2208
2209                 }
2210         }
2211 }
2212
2213 static void audit_mappings(struct kvm_vcpu *vcpu)
2214 {
2215         unsigned i;
2216
2217         if (vcpu->arch.mmu.root_level == 4)
2218                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2219         else
2220                 for (i = 0; i < 4; ++i)
2221                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2222                                 audit_mappings_page(vcpu,
2223                                                     vcpu->arch.mmu.pae_root[i],
2224                                                     i << 30,
2225                                                     2);
2226 }
2227
2228 static int count_rmaps(struct kvm_vcpu *vcpu)
2229 {
2230         int nmaps = 0;
2231         int i, j, k;
2232
2233         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2234                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2235                 struct kvm_rmap_desc *d;
2236
2237                 for (j = 0; j < m->npages; ++j) {
2238                         unsigned long *rmapp = &m->rmap[j];
2239
2240                         if (!*rmapp)
2241                                 continue;
2242                         if (!(*rmapp & 1)) {
2243                                 ++nmaps;
2244                                 continue;
2245                         }
2246                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2247                         while (d) {
2248                                 for (k = 0; k < RMAP_EXT; ++k)
2249                                         if (d->shadow_ptes[k])
2250                                                 ++nmaps;
2251                                         else
2252                                                 break;
2253                                 d = d->more;
2254                         }
2255                 }
2256         }
2257         return nmaps;
2258 }
2259
2260 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2261 {
2262         int nmaps = 0;
2263         struct kvm_mmu_page *sp;
2264         int i;
2265
2266         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2267                 u64 *pt = sp->spt;
2268
2269                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2270                         continue;
2271
2272                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2273                         u64 ent = pt[i];
2274
2275                         if (!(ent & PT_PRESENT_MASK))
2276                                 continue;
2277                         if (!(ent & PT_WRITABLE_MASK))
2278                                 continue;
2279                         ++nmaps;
2280                 }
2281         }
2282         return nmaps;
2283 }
2284
2285 static void audit_rmap(struct kvm_vcpu *vcpu)
2286 {
2287         int n_rmap = count_rmaps(vcpu);
2288         int n_actual = count_writable_mappings(vcpu);
2289
2290         if (n_rmap != n_actual)
2291                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2292                        __func__, audit_msg, n_rmap, n_actual);
2293 }
2294
2295 static void audit_write_protection(struct kvm_vcpu *vcpu)
2296 {
2297         struct kvm_mmu_page *sp;
2298         struct kvm_memory_slot *slot;
2299         unsigned long *rmapp;
2300         gfn_t gfn;
2301
2302         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2303                 if (sp->role.metaphysical)
2304                         continue;
2305
2306                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2307                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2308                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2309                 if (*rmapp)
2310                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2311                                " mappings: gfn %lx role %x\n",
2312                                __func__, audit_msg, sp->gfn,
2313                                sp->role.word);
2314         }
2315 }
2316
2317 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2318 {
2319         int olddbg = dbg;
2320
2321         dbg = 0;
2322         audit_msg = msg;
2323         audit_rmap(vcpu);
2324         audit_write_protection(vcpu);
2325         audit_mappings(vcpu);
2326         dbg = olddbg;
2327 }
2328
2329 #endif