git.oblomov.eu Git - 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 "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 = 0;
  70 module_param(dbg, bool, 0644);
  71 #endif
  72
  73 #ifndef MMU_DEBUG
  74 #define ASSERT(x) do { } while (0)
  75 #else
  76 #define ASSERT(x)                                                       \
  77         if (!(x)) {                                                     \
  78                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  79                        __FILE__, __LINE__, #x);                         \
  80         }
  81 #endif
  82
  83 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  84 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  85
  86 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  87
  88 #define PT64_LEVEL_BITS 9
  89
  90 #define PT64_LEVEL_SHIFT(level) \
  91                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  92
  93 #define PT64_LEVEL_MASK(level) \
  94                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  95
  96 #define PT64_INDEX(address, level)\
  97         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
  98
  99
 100 #define PT32_LEVEL_BITS 10
 101
 102 #define PT32_LEVEL_SHIFT(level) \
 103                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 104
 105 #define PT32_LEVEL_MASK(level) \
 106                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 107
 108 #define PT32_INDEX(address, level)\
 109         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 110
 111
 112 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 113 #define PT64_DIR_BASE_ADDR_MASK \
 114         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 115
 116 #define PT32_BASE_ADDR_MASK PAGE_MASK
 117 #define PT32_DIR_BASE_ADDR_MASK \
 118         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 119
 120 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 121                         | PT64_NX_MASK)
 122
 123 #define PFERR_PRESENT_MASK (1U << 0)
 124 #define PFERR_WRITE_MASK (1U << 1)
 125 #define PFERR_USER_MASK (1U << 2)
 126 #define PFERR_FETCH_MASK (1U << 4)
 127
 128 #define PT_DIRECTORY_LEVEL 2
 129 #define PT_PAGE_TABLE_LEVEL 1
 130
 131 #define RMAP_EXT 4
 132
 133 #define ACC_EXEC_MASK    1
 134 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 135 #define ACC_USER_MASK    PT_USER_MASK
 136 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 137
 138 struct kvm_pv_mmu_op_buffer {
 139         void *ptr;
 140         unsigned len;
 141         unsigned processed;
 142         char buf[512] __aligned(sizeof(long));
 143 };
 144
 145 struct kvm_rmap_desc {
 146         u64 *shadow_ptes[RMAP_EXT];
 147         struct kvm_rmap_desc *more;
 148 };
 149
 150 static struct kmem_cache *pte_chain_cache;
 151 static struct kmem_cache *rmap_desc_cache;
 152 static struct kmem_cache *mmu_page_header_cache;
 153
 154 static u64 __read_mostly shadow_trap_nonpresent_pte;
 155 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 156 static u64 __read_mostly shadow_base_present_pte;
 157 static u64 __read_mostly shadow_nx_mask;
 158 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 159 static u64 __read_mostly shadow_user_mask;
 160 static u64 __read_mostly shadow_accessed_mask;
 161 static u64 __read_mostly shadow_dirty_mask;
 162
 163 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 164 {
 165         shadow_trap_nonpresent_pte = trap_pte;
 166         shadow_notrap_nonpresent_pte = notrap_pte;
 167 }
 168 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 169
 170 void kvm_mmu_set_base_ptes(u64 base_pte)
 171 {
 172         shadow_base_present_pte = base_pte;
 173 }
 174 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 175
 176 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 177                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 178 {
 179         shadow_user_mask = user_mask;
 180         shadow_accessed_mask = accessed_mask;
 181         shadow_dirty_mask = dirty_mask;
 182         shadow_nx_mask = nx_mask;
 183         shadow_x_mask = x_mask;
 184 }
 185 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 186
 187 static int is_write_protection(struct kvm_vcpu *vcpu)
 188 {
 189         return vcpu->arch.cr0 & X86_CR0_WP;
 190 }
 191
 192 static int is_cpuid_PSE36(void)
 193 {
 194         return 1;
 195 }
 196
 197 static int is_nx(struct kvm_vcpu *vcpu)
 198 {
 199         return vcpu->arch.shadow_efer & EFER_NX;
 200 }
 201
 202 static int is_present_pte(unsigned long pte)
 203 {
 204         return pte & PT_PRESENT_MASK;
 205 }
 206
 207 static int is_shadow_present_pte(u64 pte)
 208 {
 209         return pte != shadow_trap_nonpresent_pte
 210                 && pte != shadow_notrap_nonpresent_pte;
 211 }
 212
 213 static int is_large_pte(u64 pte)
 214 {
 215         return pte & PT_PAGE_SIZE_MASK;
 216 }
 217
 218 static int is_writeble_pte(unsigned long pte)
 219 {
 220         return pte & PT_WRITABLE_MASK;
 221 }
 222
 223 static int is_dirty_pte(unsigned long pte)
 224 {
 225         return pte & shadow_dirty_mask;
 226 }
 227
 228 static int is_rmap_pte(u64 pte)
 229 {
 230         return is_shadow_present_pte(pte);
 231 }
 232
 233 static pfn_t spte_to_pfn(u64 pte)
 234 {
 235         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 236 }
 237
 238 static gfn_t pse36_gfn_delta(u32 gpte)
 239 {
 240         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 241
 242         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 243 }
 244
 245 static void set_shadow_pte(u64 *sptep, u64 spte)
 246 {
 247 #ifdef CONFIG_X86_64
 248         set_64bit((unsigned long *)sptep, spte);
 249 #else
 250         set_64bit((unsigned long long *)sptep, spte);
 251 #endif
 252 }
 253
 254 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 255                                   struct kmem_cache *base_cache, int min)
 256 {
 257         void *obj;
 258
 259         if (cache->nobjs >= min)
 260                 return 0;
 261         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 262                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 263                 if (!obj)
 264                         return -ENOMEM;
 265                 cache->objects[cache->nobjs++] = obj;
 266         }
 267         return 0;
 268 }
 269
 270 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 271 {
 272         while (mc->nobjs)
 273                 kfree(mc->objects[--mc->nobjs]);
 274 }
 275
 276 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 277                                        int min)
 278 {
 279         struct page *page;
 280
 281         if (cache->nobjs >= min)
 282                 return 0;
 283         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 284                 page = alloc_page(GFP_KERNEL);
 285                 if (!page)
 286                         return -ENOMEM;
 287                 set_page_private(page, 0);
 288                 cache->objects[cache->nobjs++] = page_address(page);
 289         }
 290         return 0;
 291 }
 292
 293 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 294 {
 295         while (mc->nobjs)
 296                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 297 }
 298
 299 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 300 {
 301         int r;
 302
 303         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 304                                    pte_chain_cache, 4);
 305         if (r)
 306                 goto out;
 307         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 308                                    rmap_desc_cache, 1);
 309         if (r)
 310                 goto out;
 311         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 312         if (r)
 313                 goto out;
 314         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 315                                    mmu_page_header_cache, 4);
 316 out:
 317         return r;
 318 }
 319
 320 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 321 {
 322         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 323         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 324         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 325         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 326 }
 327
 328 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 329                                     size_t size)
 330 {
 331         void *p;
 332
 333         BUG_ON(!mc->nobjs);
 334         p = mc->objects[--mc->nobjs];
 335         memset(p, 0, size);
 336         return p;
 337 }
 338
 339 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 340 {
 341         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 342                                       sizeof(struct kvm_pte_chain));
 343 }
 344
 345 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 346 {
 347         kfree(pc);
 348 }
 349
 350 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 351 {
 352         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 353                                       sizeof(struct kvm_rmap_desc));
 354 }
 355
 356 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 357 {
 358         kfree(rd);
 359 }
 360
 361 /*
 362  * Return the pointer to the largepage write count for a given
 363  * gfn, handling slots that are not large page aligned.
 364  */
 365 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 366 {
 367         unsigned long idx;
 368
 369         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 370               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 371         return &slot->lpage_info[idx].write_count;
 372 }
 373
 374 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 375 {
 376         int *write_count;
 377
 378         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 379         *write_count += 1;
 380 }
 381
 382 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 383 {
 384         int *write_count;
 385
 386         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 387         *write_count -= 1;
 388         WARN_ON(*write_count < 0);
 389 }
 390
 391 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 392 {
 393         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
 394         int *largepage_idx;
 395
 396         if (slot) {
 397                 largepage_idx = slot_largepage_idx(gfn, slot);
 398                 return *largepage_idx;
 399         }
 400
 401         return 1;
 402 }
 403
 404 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 405 {
 406         struct vm_area_struct *vma;
 407         unsigned long addr;
 408
 409         addr = gfn_to_hva(kvm, gfn);
 410         if (kvm_is_error_hva(addr))
 411                 return 0;
 412
 413         vma = find_vma(current->mm, addr);
 414         if (vma && is_vm_hugetlb_page(vma))
 415                 return 1;
 416
 417         return 0;
 418 }
 419
 420 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 421 {
 422         struct kvm_memory_slot *slot;
 423
 424         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 425                 return 0;
 426
 427         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 428                 return 0;
 429
 430         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 431         if (slot && slot->dirty_bitmap)
 432                 return 0;
 433
 434         return 1;
 435 }
 436
 437 /*
 438  * Take gfn and return the reverse mapping to it.
 439  * Note: gfn must be unaliased before this function get called
 440  */
 441
 442 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 443 {
 444         struct kvm_memory_slot *slot;
 445         unsigned long idx;
 446
 447         slot = gfn_to_memslot(kvm, gfn);
 448         if (!lpage)
 449                 return &slot->rmap[gfn - slot->base_gfn];
 450
 451         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 452               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 453
 454         return &slot->lpage_info[idx].rmap_pde;
 455 }
 456
 457 /*
 458  * Reverse mapping data structures:
 459  *
 460  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 461  * that points to page_address(page).
 462  *
 463  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 464  * containing more mappings.
 465  */
 466 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 467 {
 468         struct kvm_mmu_page *sp;
 469         struct kvm_rmap_desc *desc;
 470         unsigned long *rmapp;
 471         int i;
 472
 473         if (!is_rmap_pte(*spte))
 474                 return;
 475         gfn = unalias_gfn(vcpu->kvm, gfn);
 476         sp = page_header(__pa(spte));
 477         sp->gfns[spte - sp->spt] = gfn;
 478         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 479         if (!*rmapp) {
 480                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 481                 *rmapp = (unsigned long)spte;
 482         } else if (!(*rmapp & 1)) {
 483                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 484                 desc = mmu_alloc_rmap_desc(vcpu);
 485                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 486                 desc->shadow_ptes[1] = spte;
 487                 *rmapp = (unsigned long)desc | 1;
 488         } else {
 489                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 490                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 491                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 492                         desc = desc->more;
 493                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 494                         desc->more = mmu_alloc_rmap_desc(vcpu);
 495                         desc = desc->more;
 496                 }
 497                 for (i = 0; desc->shadow_ptes[i]; ++i)
 498                         ;
 499                 desc->shadow_ptes[i] = spte;
 500         }
 501 }
 502
 503 static void rmap_desc_remove_entry(unsigned long *rmapp,
 504                                    struct kvm_rmap_desc *desc,
 505                                    int i,
 506                                    struct kvm_rmap_desc *prev_desc)
 507 {
 508         int j;
 509
 510         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 511                 ;
 512         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 513         desc->shadow_ptes[j] = NULL;
 514         if (j != 0)
 515                 return;
 516         if (!prev_desc && !desc->more)
 517                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 518         else
 519                 if (prev_desc)
 520                         prev_desc->more = desc->more;
 521                 else
 522                         *rmapp = (unsigned long)desc->more | 1;
 523         mmu_free_rmap_desc(desc);
 524 }
 525
 526 static void rmap_remove(struct kvm *kvm, u64 *spte)
 527 {
 528         struct kvm_rmap_desc *desc;
 529         struct kvm_rmap_desc *prev_desc;
 530         struct kvm_mmu_page *sp;
 531         pfn_t pfn;
 532         unsigned long *rmapp;
 533         int i;
 534
 535         if (!is_rmap_pte(*spte))
 536                 return;
 537         sp = page_header(__pa(spte));
 538         pfn = spte_to_pfn(*spte);
 539         if (*spte & shadow_accessed_mask)
 540                 kvm_set_pfn_accessed(pfn);
 541         if (is_writeble_pte(*spte))
 542                 kvm_release_pfn_dirty(pfn);
 543         else
 544                 kvm_release_pfn_clean(pfn);
 545         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 546         if (!*rmapp) {
 547                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 548                 BUG();
 549         } else if (!(*rmapp & 1)) {
 550                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 551                 if ((u64 *)*rmapp != spte) {
 552                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 553                                spte, *spte);
 554                         BUG();
 555                 }
 556                 *rmapp = 0;
 557         } else {
 558                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 559                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 560                 prev_desc = NULL;
 561                 while (desc) {
 562                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 563                                 if (desc->shadow_ptes[i] == spte) {
 564                                         rmap_desc_remove_entry(rmapp,
 565                                                                desc, i,
 566                                                                prev_desc);
 567                                         return;
 568                                 }
 569                         prev_desc = desc;
 570                         desc = desc->more;
 571                 }
 572                 BUG();
 573         }
 574 }
 575
 576 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 577 {
 578         struct kvm_rmap_desc *desc;
 579         struct kvm_rmap_desc *prev_desc;
 580         u64 *prev_spte;
 581         int i;
 582
 583         if (!*rmapp)
 584                 return NULL;
 585         else if (!(*rmapp & 1)) {
 586                 if (!spte)
 587                         return (u64 *)*rmapp;
 588                 return NULL;
 589         }
 590         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 591         prev_desc = NULL;
 592         prev_spte = NULL;
 593         while (desc) {
 594                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 595                         if (prev_spte == spte)
 596                                 return desc->shadow_ptes[i];
 597                         prev_spte = desc->shadow_ptes[i];
 598                 }
 599                 desc = desc->more;
 600         }
 601         return NULL;
 602 }
 603
 604 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 605 {
 606         unsigned long *rmapp;
 607         u64 *spte;
 608         int write_protected = 0;
 609
 610         gfn = unalias_gfn(kvm, gfn);
 611         rmapp = gfn_to_rmap(kvm, gfn, 0);
 612
 613         spte = rmap_next(kvm, rmapp, NULL);
 614         while (spte) {
 615                 BUG_ON(!spte);
 616                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 617                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 618                 if (is_writeble_pte(*spte)) {
 619                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 620                         write_protected = 1;
 621                 }
 622                 spte = rmap_next(kvm, rmapp, spte);
 623         }
 624         if (write_protected) {
 625                 pfn_t pfn;
 626
 627                 spte = rmap_next(kvm, rmapp, NULL);
 628                 pfn = spte_to_pfn(*spte);
 629                 kvm_set_pfn_dirty(pfn);
 630         }
 631
 632         /* check for huge page mappings */
 633         rmapp = gfn_to_rmap(kvm, gfn, 1);
 634         spte = rmap_next(kvm, rmapp, NULL);
 635         while (spte) {
 636                 BUG_ON(!spte);
 637                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 638                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 639                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 640                 if (is_writeble_pte(*spte)) {
 641                         rmap_remove(kvm, spte);
 642                         --kvm->stat.lpages;
 643                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 644                         spte = NULL;
 645                         write_protected = 1;
 646                 }
 647                 spte = rmap_next(kvm, rmapp, spte);
 648         }
 649
 650         if (write_protected)
 651                 kvm_flush_remote_tlbs(kvm);
 652
 653         account_shadowed(kvm, gfn);
 654 }
 655
 656 #ifdef MMU_DEBUG
 657 static int is_empty_shadow_page(u64 *spt)
 658 {
 659         u64 *pos;
 660         u64 *end;
 661
 662         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 663                 if (is_shadow_present_pte(*pos)) {
 664                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 665                                pos, *pos);
 666                         return 0;
 667                 }
 668         return 1;
 669 }
 670 #endif
 671
 672 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 673 {
 674         ASSERT(is_empty_shadow_page(sp->spt));
 675         list_del(&sp->link);
 676         __free_page(virt_to_page(sp->spt));
 677         __free_page(virt_to_page(sp->gfns));
 678         kfree(sp);
 679         ++kvm->arch.n_free_mmu_pages;
 680 }
 681
 682 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 683 {
 684         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 685 }
 686
 687 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 688                                                u64 *parent_pte)
 689 {
 690         struct kvm_mmu_page *sp;
 691
 692         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 693         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 694         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 695         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 696         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 697         ASSERT(is_empty_shadow_page(sp->spt));
 698         sp->slot_bitmap = 0;
 699         sp->multimapped = 0;
 700         sp->parent_pte = parent_pte;
 701         --vcpu->kvm->arch.n_free_mmu_pages;
 702         return sp;
 703 }
 704
 705 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 706                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 707 {
 708         struct kvm_pte_chain *pte_chain;
 709         struct hlist_node *node;
 710         int i;
 711
 712         if (!parent_pte)
 713                 return;
 714         if (!sp->multimapped) {
 715                 u64 *old = sp->parent_pte;
 716
 717                 if (!old) {
 718                         sp->parent_pte = parent_pte;
 719                         return;
 720                 }
 721                 sp->multimapped = 1;
 722                 pte_chain = mmu_alloc_pte_chain(vcpu);
 723                 INIT_HLIST_HEAD(&sp->parent_ptes);
 724                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 725                 pte_chain->parent_ptes[0] = old;
 726         }
 727         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 728                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 729                         continue;
 730                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 731                         if (!pte_chain->parent_ptes[i]) {
 732                                 pte_chain->parent_ptes[i] = parent_pte;
 733                                 return;
 734                         }
 735         }
 736         pte_chain = mmu_alloc_pte_chain(vcpu);
 737         BUG_ON(!pte_chain);
 738         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 739         pte_chain->parent_ptes[0] = parent_pte;
 740 }
 741
 742 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 743                                        u64 *parent_pte)
 744 {
 745         struct kvm_pte_chain *pte_chain;
 746         struct hlist_node *node;
 747         int i;
 748
 749         if (!sp->multimapped) {
 750                 BUG_ON(sp->parent_pte != parent_pte);
 751                 sp->parent_pte = NULL;
 752                 return;
 753         }
 754         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 755                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 756                         if (!pte_chain->parent_ptes[i])
 757                                 break;
 758                         if (pte_chain->parent_ptes[i] != parent_pte)
 759                                 continue;
 760                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 761                                 && pte_chain->parent_ptes[i + 1]) {
 762                                 pte_chain->parent_ptes[i]
 763                                         = pte_chain->parent_ptes[i + 1];
 764                                 ++i;
 765                         }
 766                         pte_chain->parent_ptes[i] = NULL;
 767                         if (i == 0) {
 768                                 hlist_del(&pte_chain->link);
 769                                 mmu_free_pte_chain(pte_chain);
 770                                 if (hlist_empty(&sp->parent_ptes)) {
 771                                         sp->multimapped = 0;
 772                                         sp->parent_pte = NULL;
 773                                 }
 774                         }
 775                         return;
 776                 }
 777         BUG();
 778 }
 779
 780 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 781                                     struct kvm_mmu_page *sp)
 782 {
 783         int i;
 784
 785         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 786                 sp->spt[i] = shadow_trap_nonpresent_pte;
 787 }
 788
 789 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
 790 {
 791         unsigned index;
 792         struct hlist_head *bucket;
 793         struct kvm_mmu_page *sp;
 794         struct hlist_node *node;
 795
 796         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 797         index = kvm_page_table_hashfn(gfn);
 798         bucket = &kvm->arch.mmu_page_hash[index];
 799         hlist_for_each_entry(sp, node, bucket, hash_link)
 800                 if (sp->gfn == gfn && !sp->role.metaphysical
 801                     && !sp->role.invalid) {
 802                         pgprintk("%s: found role %x\n",
 803                                  __func__, sp->role.word);
 804                         return sp;
 805                 }
 806         return NULL;
 807 }
 808
 809 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 810                                              gfn_t gfn,
 811                                              gva_t gaddr,
 812                                              unsigned level,
 813                                              int metaphysical,
 814                                              unsigned access,
 815                                              u64 *parent_pte)
 816 {
 817         union kvm_mmu_page_role role;
 818         unsigned index;
 819         unsigned quadrant;
 820         struct hlist_head *bucket;
 821         struct kvm_mmu_page *sp;
 822         struct hlist_node *node;
 823
 824         role.word = 0;
 825         role.glevels = vcpu->arch.mmu.root_level;
 826         role.level = level;
 827         role.metaphysical = metaphysical;
 828         role.access = access;
 829         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
 830                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 831                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 832                 role.quadrant = quadrant;
 833         }
 834         pgprintk("%s: looking gfn %lx role %x\n", __func__,
 835                  gfn, role.word);
 836         index = kvm_page_table_hashfn(gfn);
 837         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
 838         hlist_for_each_entry(sp, node, bucket, hash_link)
 839                 if (sp->gfn == gfn && sp->role.word == role.word) {
 840                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
 841                         pgprintk("%s: found\n", __func__);
 842                         return sp;
 843                 }
 844         ++vcpu->kvm->stat.mmu_cache_miss;
 845         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
 846         if (!sp)
 847                 return sp;
 848         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
 849         sp->gfn = gfn;
 850         sp->role = role;
 851         hlist_add_head(&sp->hash_link, bucket);
 852         if (!metaphysical)
 853                 rmap_write_protect(vcpu->kvm, gfn);
 854         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
 855                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
 856         else
 857                 nonpaging_prefetch_page(vcpu, sp);
 858         return sp;
 859 }
 860
 861 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 862                                          struct kvm_mmu_page *sp)
 863 {
 864         unsigned i;
 865         u64 *pt;
 866         u64 ent;
 867
 868         pt = sp->spt;
 869
 870         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
 871                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 872                         if (is_shadow_present_pte(pt[i]))
 873                                 rmap_remove(kvm, &pt[i]);
 874                         pt[i] = shadow_trap_nonpresent_pte;
 875                 }
 876                 kvm_flush_remote_tlbs(kvm);
 877                 return;
 878         }
 879
 880         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 881                 ent = pt[i];
 882
 883                 if (is_shadow_present_pte(ent)) {
 884                         if (!is_large_pte(ent)) {
 885                                 ent &= PT64_BASE_ADDR_MASK;
 886                                 mmu_page_remove_parent_pte(page_header(ent),
 887                                                            &pt[i]);
 888                         } else {
 889                                 --kvm->stat.lpages;
 890                                 rmap_remove(kvm, &pt[i]);
 891                         }
 892                 }
 893                 pt[i] = shadow_trap_nonpresent_pte;
 894         }
 895         kvm_flush_remote_tlbs(kvm);
 896 }
 897
 898 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
 899 {
 900         mmu_page_remove_parent_pte(sp, parent_pte);
 901 }
 902
 903 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 904 {
 905         int i;
 906
 907         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 908                 if (kvm->vcpus[i])
 909                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
 910 }
 911
 912 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 913 {
 914         u64 *parent_pte;
 915
 916         ++kvm->stat.mmu_shadow_zapped;
 917         while (sp->multimapped || sp->parent_pte) {
 918                 if (!sp->multimapped)
 919                         parent_pte = sp->parent_pte;
 920                 else {
 921                         struct kvm_pte_chain *chain;
 922
 923                         chain = container_of(sp->parent_ptes.first,
 924                                              struct kvm_pte_chain, link);
 925                         parent_pte = chain->parent_ptes[0];
 926                 }
 927                 BUG_ON(!parent_pte);
 928                 kvm_mmu_put_page(sp, parent_pte);
 929                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 930         }
 931         kvm_mmu_page_unlink_children(kvm, sp);
 932         if (!sp->root_count) {
 933                 if (!sp->role.metaphysical && !sp->role.invalid)
 934                         unaccount_shadowed(kvm, sp->gfn);
 935                 hlist_del(&sp->hash_link);
 936                 kvm_mmu_free_page(kvm, sp);
 937         } else {
 938                 int invalid = sp->role.invalid;
 939                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
 940                 sp->role.invalid = 1;
 941                 kvm_reload_remote_mmus(kvm);
 942                 if (!sp->role.metaphysical && !invalid)
 943                         unaccount_shadowed(kvm, sp->gfn);
 944         }
 945         kvm_mmu_reset_last_pte_updated(kvm);
 946 }
 947
 948 /*
 949  * Changing the number of mmu pages allocated to the vm
 950  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 951  */
 952 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 953 {
 954         /*
 955          * If we set the number of mmu pages to be smaller be than the
 956          * number of actived pages , we must to free some mmu pages before we
 957          * change the value
 958          */
 959
 960         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
 961             kvm_nr_mmu_pages) {
 962                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
 963                                        - kvm->arch.n_free_mmu_pages;
 964
 965                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 966                         struct kvm_mmu_page *page;
 967
 968                         page = container_of(kvm->arch.active_mmu_pages.prev,
 969                                             struct kvm_mmu_page, link);
 970                         kvm_mmu_zap_page(kvm, page);
 971                         n_used_mmu_pages--;
 972                 }
 973                 kvm->arch.n_free_mmu_pages = 0;
 974         }
 975         else
 976                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
 977                                          - kvm->arch.n_alloc_mmu_pages;
 978
 979         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
 980 }
 981
 982 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 983 {
 984         unsigned index;
 985         struct hlist_head *bucket;
 986         struct kvm_mmu_page *sp;
 987         struct hlist_node *node, *n;
 988         int r;
 989
 990         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 991         r = 0;
 992         index = kvm_page_table_hashfn(gfn);
 993         bucket = &kvm->arch.mmu_page_hash[index];
 994         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
 995                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 996                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
 997                                  sp->role.word);
 998                         kvm_mmu_zap_page(kvm, sp);
 999                         r = 1;
1000                 }
1001         return r;
1002 }
1003
1004 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1005 {
1006         struct kvm_mmu_page *sp;
1007
1008         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1009                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1010                 kvm_mmu_zap_page(kvm, sp);
1011         }
1012 }
1013
1014 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1015 {
1016         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1017         struct kvm_mmu_page *sp = page_header(__pa(pte));
1018
1019         __set_bit(slot, &sp->slot_bitmap);
1020 }
1021
1022 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1023 {
1024         struct page *page;
1025
1026         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1027
1028         if (gpa == UNMAPPED_GVA)
1029                 return NULL;
1030
1031         down_read(&current->mm->mmap_sem);
1032         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1033         up_read(&current->mm->mmap_sem);
1034
1035         return page;
1036 }
1037
1038 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1039                          unsigned pt_access, unsigned pte_access,
1040                          int user_fault, int write_fault, int dirty,
1041                          int *ptwrite, int largepage, gfn_t gfn,
1042                          pfn_t pfn, bool speculative)
1043 {
1044         u64 spte;
1045         int was_rmapped = 0;
1046         int was_writeble = is_writeble_pte(*shadow_pte);
1047
1048         pgprintk("%s: spte %llx access %x write_fault %d"
1049                  " user_fault %d gfn %lx\n",
1050                  __func__, *shadow_pte, pt_access,
1051                  write_fault, user_fault, gfn);
1052
1053         if (is_rmap_pte(*shadow_pte)) {
1054                 /*
1055                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1056                  * the parent of the now unreachable PTE.
1057                  */
1058                 if (largepage && !is_large_pte(*shadow_pte)) {
1059                         struct kvm_mmu_page *child;
1060                         u64 pte = *shadow_pte;
1061
1062                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1063                         mmu_page_remove_parent_pte(child, shadow_pte);
1064                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1065                         pgprintk("hfn old %lx new %lx\n",
1066                                  spte_to_pfn(*shadow_pte), pfn);
1067                         rmap_remove(vcpu->kvm, shadow_pte);
1068                 } else {
1069                         if (largepage)
1070                                 was_rmapped = is_large_pte(*shadow_pte);
1071                         else
1072                                 was_rmapped = 1;
1073                 }
1074         }
1075
1076         /*
1077          * We don't set the accessed bit, since we sometimes want to see
1078          * whether the guest actually used the pte (in order to detect
1079          * demand paging).
1080          */
1081         spte = shadow_base_present_pte | shadow_dirty_mask;
1082         if (!speculative)
1083                 pte_access |= PT_ACCESSED_MASK;
1084         if (!dirty)
1085                 pte_access &= ~ACC_WRITE_MASK;
1086         if (pte_access & ACC_EXEC_MASK)
1087                 spte |= shadow_x_mask;
1088         else
1089                 spte |= shadow_nx_mask;
1090         if (pte_access & ACC_USER_MASK)
1091                 spte |= shadow_user_mask;
1092         if (largepage)
1093                 spte |= PT_PAGE_SIZE_MASK;
1094
1095         spte |= (u64)pfn << PAGE_SHIFT;
1096
1097         if ((pte_access & ACC_WRITE_MASK)
1098             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1099                 struct kvm_mmu_page *shadow;
1100
1101                 spte |= PT_WRITABLE_MASK;
1102
1103                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1104                 if (shadow ||
1105                    (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
1106                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1107                                  __func__, gfn);
1108                         pte_access &= ~ACC_WRITE_MASK;
1109                         if (is_writeble_pte(spte)) {
1110                                 spte &= ~PT_WRITABLE_MASK;
1111                                 kvm_x86_ops->tlb_flush(vcpu);
1112                         }
1113                         if (write_fault)
1114                                 *ptwrite = 1;
1115                 }
1116         }
1117
1118         if (pte_access & ACC_WRITE_MASK)
1119                 mark_page_dirty(vcpu->kvm, gfn);
1120
1121         pgprintk("%s: setting spte %llx\n", __func__, spte);
1122         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1123                  (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
1124                  (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
1125         set_shadow_pte(shadow_pte, spte);
1126         if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
1127             && (spte & PT_PRESENT_MASK))
1128                 ++vcpu->kvm->stat.lpages;
1129
1130         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1131         if (!was_rmapped) {
1132                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1133                 if (!is_rmap_pte(*shadow_pte))
1134                         kvm_release_pfn_clean(pfn);
1135         } else {
1136                 if (was_writeble)
1137                         kvm_release_pfn_dirty(pfn);
1138                 else
1139                         kvm_release_pfn_clean(pfn);
1140         }
1141         if (speculative) {
1142                 vcpu->arch.last_pte_updated = shadow_pte;
1143                 vcpu->arch.last_pte_gfn = gfn;
1144         }
1145 }
1146
1147 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1148 {
1149 }
1150
1151 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1152                            int largepage, gfn_t gfn, pfn_t pfn,
1153                            int level)
1154 {
1155         hpa_t table_addr = vcpu->arch.mmu.root_hpa;
1156         int pt_write = 0;
1157
1158         for (; ; level--) {
1159                 u32 index = PT64_INDEX(v, level);
1160                 u64 *table;
1161
1162                 ASSERT(VALID_PAGE(table_addr));
1163                 table = __va(table_addr);
1164
1165                 if (level == 1) {
1166                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1167                                      0, write, 1, &pt_write, 0, gfn, pfn, false);
1168                         return pt_write;
1169                 }
1170
1171                 if (largepage && level == 2) {
1172                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1173                                      0, write, 1, &pt_write, 1, gfn, pfn, false);
1174                         return pt_write;
1175                 }
1176
1177                 if (table[index] == shadow_trap_nonpresent_pte) {
1178                         struct kvm_mmu_page *new_table;
1179                         gfn_t pseudo_gfn;
1180
1181                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1182                                 >> PAGE_SHIFT;
1183                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1184                                                      v, level - 1,
1185                                                      1, ACC_ALL, &table[index]);
1186                         if (!new_table) {
1187                                 pgprintk("nonpaging_map: ENOMEM\n");
1188                                 kvm_release_pfn_clean(pfn);
1189                                 return -ENOMEM;
1190                         }
1191
1192                         set_shadow_pte(&table[index],
1193                                        __pa(new_table->spt)
1194                                        | PT_PRESENT_MASK | PT_WRITABLE_MASK
1195                                        | shadow_user_mask | shadow_x_mask);
1196                 }
1197                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1198         }
1199 }
1200
1201 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1202 {
1203         int r;
1204         int largepage = 0;
1205         pfn_t pfn;
1206
1207         down_read(&current->mm->mmap_sem);
1208         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1209                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1210                 largepage = 1;
1211         }
1212
1213         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1214         up_read(&current->mm->mmap_sem);
1215
1216         /* mmio */
1217         if (is_error_pfn(pfn)) {
1218                 kvm_release_pfn_clean(pfn);
1219                 return 1;
1220         }
1221
1222         spin_lock(&vcpu->kvm->mmu_lock);
1223         kvm_mmu_free_some_pages(vcpu);
1224         r = __direct_map(vcpu, v, write, largepage, gfn, pfn,
1225                          PT32E_ROOT_LEVEL);
1226         spin_unlock(&vcpu->kvm->mmu_lock);
1227
1228
1229         return r;
1230 }
1231
1232
1233 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1234 {
1235         int i;
1236         struct kvm_mmu_page *sp;
1237
1238         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1239                 return;
1240         spin_lock(&vcpu->kvm->mmu_lock);
1241         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1242                 hpa_t root = vcpu->arch.mmu.root_hpa;
1243
1244                 sp = page_header(root);
1245                 --sp->root_count;
1246                 if (!sp->root_count && sp->role.invalid)
1247                         kvm_mmu_zap_page(vcpu->kvm, sp);
1248                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1249                 spin_unlock(&vcpu->kvm->mmu_lock);
1250                 return;
1251         }
1252         for (i = 0; i < 4; ++i) {
1253                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1254
1255                 if (root) {
1256                         root &= PT64_BASE_ADDR_MASK;
1257                         sp = page_header(root);
1258                         --sp->root_count;
1259                         if (!sp->root_count && sp->role.invalid)
1260                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1261                 }
1262                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1263         }
1264         spin_unlock(&vcpu->kvm->mmu_lock);
1265         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1266 }
1267
1268 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1269 {
1270         int i;
1271         gfn_t root_gfn;
1272         struct kvm_mmu_page *sp;
1273         int metaphysical = 0;
1274
1275         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1276
1277         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1278                 hpa_t root = vcpu->arch.mmu.root_hpa;
1279
1280                 ASSERT(!VALID_PAGE(root));
1281                 if (tdp_enabled)
1282                         metaphysical = 1;
1283                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1284                                       PT64_ROOT_LEVEL, metaphysical,
1285                                       ACC_ALL, NULL);
1286                 root = __pa(sp->spt);
1287                 ++sp->root_count;
1288                 vcpu->arch.mmu.root_hpa = root;
1289                 return;
1290         }
1291         metaphysical = !is_paging(vcpu);
1292         if (tdp_enabled)
1293                 metaphysical = 1;
1294         for (i = 0; i < 4; ++i) {
1295                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1296
1297                 ASSERT(!VALID_PAGE(root));
1298                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1299                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1300                                 vcpu->arch.mmu.pae_root[i] = 0;
1301                                 continue;
1302                         }
1303                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1304                 } else if (vcpu->arch.mmu.root_level == 0)
1305                         root_gfn = 0;
1306                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1307                                       PT32_ROOT_LEVEL, metaphysical,
1308                                       ACC_ALL, NULL);
1309                 root = __pa(sp->spt);
1310                 ++sp->root_count;
1311                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1312         }
1313         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1314 }
1315
1316 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1317 {
1318         return vaddr;
1319 }
1320
1321 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1322                                 u32 error_code)
1323 {
1324         gfn_t gfn;
1325         int r;
1326
1327         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1328         r = mmu_topup_memory_caches(vcpu);
1329         if (r)
1330                 return r;
1331
1332         ASSERT(vcpu);
1333         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1334
1335         gfn = gva >> PAGE_SHIFT;
1336
1337         return nonpaging_map(vcpu, gva & PAGE_MASK,
1338                              error_code & PFERR_WRITE_MASK, gfn);
1339 }
1340
1341 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1342                                 u32 error_code)
1343 {
1344         pfn_t pfn;
1345         int r;
1346         int largepage = 0;
1347         gfn_t gfn = gpa >> PAGE_SHIFT;
1348
1349         ASSERT(vcpu);
1350         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1351
1352         r = mmu_topup_memory_caches(vcpu);
1353         if (r)
1354                 return r;
1355
1356         down_read(&current->mm->mmap_sem);
1357         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1358                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1359                 largepage = 1;
1360         }
1361         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1362         up_read(&current->mm->mmap_sem);
1363         if (is_error_pfn(pfn)) {
1364                 kvm_release_pfn_clean(pfn);
1365                 return 1;
1366         }
1367         spin_lock(&vcpu->kvm->mmu_lock);
1368         kvm_mmu_free_some_pages(vcpu);
1369         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1370                          largepage, gfn, pfn, kvm_x86_ops->get_tdp_level());
1371         spin_unlock(&vcpu->kvm->mmu_lock);
1372
1373         return r;
1374 }
1375
1376 static void nonpaging_free(struct kvm_vcpu *vcpu)
1377 {
1378         mmu_free_roots(vcpu);
1379 }
1380
1381 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1382 {
1383         struct kvm_mmu *context = &vcpu->arch.mmu;
1384
1385         context->new_cr3 = nonpaging_new_cr3;
1386         context->page_fault = nonpaging_page_fault;
1387         context->gva_to_gpa = nonpaging_gva_to_gpa;
1388         context->free = nonpaging_free;
1389         context->prefetch_page = nonpaging_prefetch_page;
1390         context->root_level = 0;
1391         context->shadow_root_level = PT32E_ROOT_LEVEL;
1392         context->root_hpa = INVALID_PAGE;
1393         return 0;
1394 }
1395
1396 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1397 {
1398         ++vcpu->stat.tlb_flush;
1399         kvm_x86_ops->tlb_flush(vcpu);
1400 }
1401
1402 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1403 {
1404         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1405         mmu_free_roots(vcpu);
1406 }
1407
1408 static void inject_page_fault(struct kvm_vcpu *vcpu,
1409                               u64 addr,
1410                               u32 err_code)
1411 {
1412         kvm_inject_page_fault(vcpu, addr, err_code);
1413 }
1414
1415 static void paging_free(struct kvm_vcpu *vcpu)
1416 {
1417         nonpaging_free(vcpu);
1418 }
1419
1420 #define PTTYPE 64
1421 #include "paging_tmpl.h"
1422 #undef PTTYPE
1423
1424 #define PTTYPE 32
1425 #include "paging_tmpl.h"
1426 #undef PTTYPE
1427
1428 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1429 {
1430         struct kvm_mmu *context = &vcpu->arch.mmu;
1431
1432         ASSERT(is_pae(vcpu));
1433         context->new_cr3 = paging_new_cr3;
1434         context->page_fault = paging64_page_fault;
1435         context->gva_to_gpa = paging64_gva_to_gpa;
1436         context->prefetch_page = paging64_prefetch_page;
1437         context->free = paging_free;
1438         context->root_level = level;
1439         context->shadow_root_level = level;
1440         context->root_hpa = INVALID_PAGE;
1441         return 0;
1442 }
1443
1444 static int paging64_init_context(struct kvm_vcpu *vcpu)
1445 {
1446         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1447 }
1448
1449 static int paging32_init_context(struct kvm_vcpu *vcpu)
1450 {
1451         struct kvm_mmu *context = &vcpu->arch.mmu;
1452
1453         context->new_cr3 = paging_new_cr3;
1454         context->page_fault = paging32_page_fault;
1455         context->gva_to_gpa = paging32_gva_to_gpa;
1456         context->free = paging_free;
1457         context->prefetch_page = paging32_prefetch_page;
1458         context->root_level = PT32_ROOT_LEVEL;
1459         context->shadow_root_level = PT32E_ROOT_LEVEL;
1460         context->root_hpa = INVALID_PAGE;
1461         return 0;
1462 }
1463
1464 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1465 {
1466         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1467 }
1468
1469 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1470 {
1471         struct kvm_mmu *context = &vcpu->arch.mmu;
1472
1473         context->new_cr3 = nonpaging_new_cr3;
1474         context->page_fault = tdp_page_fault;
1475         context->free = nonpaging_free;
1476         context->prefetch_page = nonpaging_prefetch_page;
1477         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
1478         context->root_hpa = INVALID_PAGE;
1479
1480         if (!is_paging(vcpu)) {
1481                 context->gva_to_gpa = nonpaging_gva_to_gpa;
1482                 context->root_level = 0;
1483         } else if (is_long_mode(vcpu)) {
1484                 context->gva_to_gpa = paging64_gva_to_gpa;
1485                 context->root_level = PT64_ROOT_LEVEL;
1486         } else if (is_pae(vcpu)) {
1487                 context->gva_to_gpa = paging64_gva_to_gpa;
1488                 context->root_level = PT32E_ROOT_LEVEL;
1489         } else {
1490                 context->gva_to_gpa = paging32_gva_to_gpa;
1491                 context->root_level = PT32_ROOT_LEVEL;
1492         }
1493
1494         return 0;
1495 }
1496
1497 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1498 {
1499         ASSERT(vcpu);
1500         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1501
1502         if (!is_paging(vcpu))
1503                 return nonpaging_init_context(vcpu);
1504         else if (is_long_mode(vcpu))
1505                 return paging64_init_context(vcpu);
1506         else if (is_pae(vcpu))
1507                 return paging32E_init_context(vcpu);
1508         else
1509                 return paging32_init_context(vcpu);
1510 }
1511
1512 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1513 {
1514         vcpu->arch.update_pte.pfn = bad_pfn;
1515
1516         if (tdp_enabled)
1517                 return init_kvm_tdp_mmu(vcpu);
1518         else
1519                 return init_kvm_softmmu(vcpu);
1520 }
1521
1522 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1523 {
1524         ASSERT(vcpu);
1525         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1526                 vcpu->arch.mmu.free(vcpu);
1527                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1528         }
1529 }
1530
1531 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1532 {
1533         destroy_kvm_mmu(vcpu);
1534         return init_kvm_mmu(vcpu);
1535 }
1536 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1537
1538 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1539 {
1540         int r;
1541
1542         r = mmu_topup_memory_caches(vcpu);
1543         if (r)
1544                 goto out;
1545         spin_lock(&vcpu->kvm->mmu_lock);
1546         kvm_mmu_free_some_pages(vcpu);
1547         mmu_alloc_roots(vcpu);
1548         spin_unlock(&vcpu->kvm->mmu_lock);
1549         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1550         kvm_mmu_flush_tlb(vcpu);
1551 out:
1552         return r;
1553 }
1554 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1555
1556 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1557 {
1558         mmu_free_roots(vcpu);
1559 }
1560
1561 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1562                                   struct kvm_mmu_page *sp,
1563                                   u64 *spte)
1564 {
1565         u64 pte;
1566         struct kvm_mmu_page *child;
1567
1568         pte = *spte;
1569         if (is_shadow_present_pte(pte)) {
1570                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1571                     is_large_pte(pte))
1572                         rmap_remove(vcpu->kvm, spte);
1573                 else {
1574                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1575                         mmu_page_remove_parent_pte(child, spte);
1576                 }
1577         }
1578         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1579         if (is_large_pte(pte))
1580                 --vcpu->kvm->stat.lpages;
1581 }
1582
1583 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1584                                   struct kvm_mmu_page *sp,
1585                                   u64 *spte,
1586                                   const void *new)
1587 {
1588         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1589                 if (!vcpu->arch.update_pte.largepage ||
1590                     sp->role.glevels == PT32_ROOT_LEVEL) {
1591                         ++vcpu->kvm->stat.mmu_pde_zapped;
1592                         return;
1593                 }
1594         }
1595
1596         ++vcpu->kvm->stat.mmu_pte_updated;
1597         if (sp->role.glevels == PT32_ROOT_LEVEL)
1598                 paging32_update_pte(vcpu, sp, spte, new);
1599         else
1600                 paging64_update_pte(vcpu, sp, spte, new);
1601 }
1602
1603 static bool need_remote_flush(u64 old, u64 new)
1604 {
1605         if (!is_shadow_present_pte(old))
1606                 return false;
1607         if (!is_shadow_present_pte(new))
1608                 return true;
1609         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1610                 return true;
1611         old ^= PT64_NX_MASK;
1612         new ^= PT64_NX_MASK;
1613         return (old & ~new & PT64_PERM_MASK) != 0;
1614 }
1615
1616 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1617 {
1618         if (need_remote_flush(old, new))
1619                 kvm_flush_remote_tlbs(vcpu->kvm);
1620         else
1621                 kvm_mmu_flush_tlb(vcpu);
1622 }
1623
1624 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1625 {
1626         u64 *spte = vcpu->arch.last_pte_updated;
1627
1628         return !!(spte && (*spte & shadow_accessed_mask));
1629 }
1630
1631 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1632                                           const u8 *new, int bytes)
1633 {
1634         gfn_t gfn;
1635         int r;
1636         u64 gpte = 0;
1637         pfn_t pfn;
1638
1639         vcpu->arch.update_pte.largepage = 0;
1640
1641         if (bytes != 4 && bytes != 8)
1642                 return;
1643
1644         /*
1645          * Assume that the pte write on a page table of the same type
1646          * as the current vcpu paging mode.  This is nearly always true
1647          * (might be false while changing modes).  Note it is verified later
1648          * by update_pte().
1649          */
1650         if (is_pae(vcpu)) {
1651                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1652                 if ((bytes == 4) && (gpa % 4 == 0)) {
1653                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1654                         if (r)
1655                                 return;
1656                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1657                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1658                         memcpy((void *)&gpte, new, 8);
1659                 }
1660         } else {
1661                 if ((bytes == 4) && (gpa % 4 == 0))
1662                         memcpy((void *)&gpte, new, 4);
1663         }
1664         if (!is_present_pte(gpte))
1665                 return;
1666         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1667
1668         down_read(&current->mm->mmap_sem);
1669         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1670                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1671                 vcpu->arch.update_pte.largepage = 1;
1672         }
1673         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1674         up_read(&current->mm->mmap_sem);
1675
1676         if (is_error_pfn(pfn)) {
1677                 kvm_release_pfn_clean(pfn);
1678                 return;
1679         }
1680         vcpu->arch.update_pte.gfn = gfn;
1681         vcpu->arch.update_pte.pfn = pfn;
1682 }
1683
1684 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1685 {
1686         u64 *spte = vcpu->arch.last_pte_updated;
1687
1688         if (spte
1689             && vcpu->arch.last_pte_gfn == gfn
1690             && shadow_accessed_mask
1691             && !(*spte & shadow_accessed_mask)
1692             && is_shadow_present_pte(*spte))
1693                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
1694 }
1695
1696 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1697                        const u8 *new, int bytes)
1698 {
1699         gfn_t gfn = gpa >> PAGE_SHIFT;
1700         struct kvm_mmu_page *sp;
1701         struct hlist_node *node, *n;
1702         struct hlist_head *bucket;
1703         unsigned index;
1704         u64 entry, gentry;
1705         u64 *spte;
1706         unsigned offset = offset_in_page(gpa);
1707         unsigned pte_size;
1708         unsigned page_offset;
1709         unsigned misaligned;
1710         unsigned quadrant;
1711         int level;
1712         int flooded = 0;
1713         int npte;
1714         int r;
1715
1716         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1717         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1718         spin_lock(&vcpu->kvm->mmu_lock);
1719         kvm_mmu_access_page(vcpu, gfn);
1720         kvm_mmu_free_some_pages(vcpu);
1721         ++vcpu->kvm->stat.mmu_pte_write;
1722         kvm_mmu_audit(vcpu, "pre pte write");
1723         if (gfn == vcpu->arch.last_pt_write_gfn
1724             && !last_updated_pte_accessed(vcpu)) {
1725                 ++vcpu->arch.last_pt_write_count;
1726                 if (vcpu->arch.last_pt_write_count >= 3)
1727                         flooded = 1;
1728         } else {
1729                 vcpu->arch.last_pt_write_gfn = gfn;
1730                 vcpu->arch.last_pt_write_count = 1;
1731                 vcpu->arch.last_pte_updated = NULL;
1732         }
1733         index = kvm_page_table_hashfn(gfn);
1734         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1735         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1736                 if (sp->gfn != gfn || sp->role.metaphysical)
1737                         continue;
1738                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1739                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1740                 misaligned |= bytes < 4;
1741                 if (misaligned || flooded) {
1742                         /*
1743                          * Misaligned accesses are too much trouble to fix
1744                          * up; also, they usually indicate a page is not used
1745                          * as a page table.
1746                          *
1747                          * If we're seeing too many writes to a page,
1748                          * it may no longer be a page table, or we may be
1749                          * forking, in which case it is better to unmap the
1750                          * page.
1751                          */
1752                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1753                                  gpa, bytes, sp->role.word);
1754                         kvm_mmu_zap_page(vcpu->kvm, sp);
1755                         ++vcpu->kvm->stat.mmu_flooded;
1756                         continue;
1757                 }
1758                 page_offset = offset;
1759                 level = sp->role.level;
1760                 npte = 1;
1761                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1762                         page_offset <<= 1;      /* 32->64 */
1763                         /*
1764                          * A 32-bit pde maps 4MB while the shadow pdes map
1765                          * only 2MB.  So we need to double the offset again
1766                          * and zap two pdes instead of one.
1767                          */
1768                         if (level == PT32_ROOT_LEVEL) {
1769                                 page_offset &= ~7; /* kill rounding error */
1770                                 page_offset <<= 1;
1771                                 npte = 2;
1772                         }
1773                         quadrant = page_offset >> PAGE_SHIFT;
1774                         page_offset &= ~PAGE_MASK;
1775                         if (quadrant != sp->role.quadrant)
1776                                 continue;
1777                 }
1778                 spte = &sp->spt[page_offset / sizeof(*spte)];
1779                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1780                         gentry = 0;
1781                         r = kvm_read_guest_atomic(vcpu->kvm,
1782                                                   gpa & ~(u64)(pte_size - 1),
1783                                                   &gentry, pte_size);
1784                         new = (const void *)&gentry;
1785                         if (r < 0)
1786                                 new = NULL;
1787                 }
1788                 while (npte--) {
1789                         entry = *spte;
1790                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1791                         if (new)
1792                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1793                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1794                         ++spte;
1795                 }
1796         }
1797         kvm_mmu_audit(vcpu, "post pte write");
1798         spin_unlock(&vcpu->kvm->mmu_lock);
1799         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
1800                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
1801                 vcpu->arch.update_pte.pfn = bad_pfn;
1802         }
1803 }
1804
1805 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1806 {
1807         gpa_t gpa;
1808         int r;
1809
1810         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1811
1812         spin_lock(&vcpu->kvm->mmu_lock);
1813         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1814         spin_unlock(&vcpu->kvm->mmu_lock);
1815         return r;
1816 }
1817
1818 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1819 {
1820         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1821                 struct kvm_mmu_page *sp;
1822
1823                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1824                                   struct kvm_mmu_page, link);
1825                 kvm_mmu_zap_page(vcpu->kvm, sp);
1826                 ++vcpu->kvm->stat.mmu_recycled;
1827         }
1828 }
1829
1830 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1831 {
1832         int r;
1833         enum emulation_result er;
1834
1835         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1836         if (r < 0)
1837                 goto out;
1838
1839         if (!r) {
1840                 r = 1;
1841                 goto out;
1842         }
1843
1844         r = mmu_topup_memory_caches(vcpu);
1845         if (r)
1846                 goto out;
1847
1848         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1849
1850         switch (er) {
1851         case EMULATE_DONE:
1852                 return 1;
1853         case EMULATE_DO_MMIO:
1854                 ++vcpu->stat.mmio_exits;
1855                 return 0;
1856         case EMULATE_FAIL:
1857                 kvm_report_emulation_failure(vcpu, "pagetable");
1858                 return 1;
1859         default:
1860                 BUG();
1861         }
1862 out:
1863         return r;
1864 }
1865 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1866
1867 void kvm_enable_tdp(void)
1868 {
1869         tdp_enabled = true;
1870 }
1871 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1872
1873 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1874 {
1875         struct kvm_mmu_page *sp;
1876
1877         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1878                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1879                                   struct kvm_mmu_page, link);
1880                 kvm_mmu_zap_page(vcpu->kvm, sp);
1881                 cond_resched();
1882         }
1883         free_page((unsigned long)vcpu->arch.mmu.pae_root);
1884 }
1885
1886 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1887 {
1888         struct page *page;
1889         int i;
1890
1891         ASSERT(vcpu);
1892
1893         if (vcpu->kvm->arch.n_requested_mmu_pages)
1894                 vcpu->kvm->arch.n_free_mmu_pages =
1895                                         vcpu->kvm->arch.n_requested_mmu_pages;
1896         else
1897                 vcpu->kvm->arch.n_free_mmu_pages =
1898                                         vcpu->kvm->arch.n_alloc_mmu_pages;
1899         /*
1900          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1901          * Therefore we need to allocate shadow page tables in the first
1902          * 4GB of memory, which happens to fit the DMA32 zone.
1903          */
1904         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1905         if (!page)
1906                 goto error_1;
1907         vcpu->arch.mmu.pae_root = page_address(page);
1908         for (i = 0; i < 4; ++i)
1909                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1910
1911         return 0;
1912
1913 error_1:
1914         free_mmu_pages(vcpu);
1915         return -ENOMEM;
1916 }
1917
1918 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1919 {
1920         ASSERT(vcpu);
1921         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1922
1923         return alloc_mmu_pages(vcpu);
1924 }
1925
1926 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1927 {
1928         ASSERT(vcpu);
1929         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1930
1931         return init_kvm_mmu(vcpu);
1932 }
1933
1934 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1935 {
1936         ASSERT(vcpu);
1937
1938         destroy_kvm_mmu(vcpu);
1939         free_mmu_pages(vcpu);
1940         mmu_free_memory_caches(vcpu);
1941 }
1942
1943 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1944 {
1945         struct kvm_mmu_page *sp;
1946
1947         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1948                 int i;
1949                 u64 *pt;
1950
1951                 if (!test_bit(slot, &sp->slot_bitmap))
1952                         continue;
1953
1954                 pt = sp->spt;
1955                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1956                         /* avoid RMW */
1957                         if (pt[i] & PT_WRITABLE_MASK)
1958                                 pt[i] &= ~PT_WRITABLE_MASK;
1959         }
1960 }
1961
1962 void kvm_mmu_zap_all(struct kvm *kvm)
1963 {
1964         struct kvm_mmu_page *sp, *node;
1965
1966         spin_lock(&kvm->mmu_lock);
1967         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1968                 kvm_mmu_zap_page(kvm, sp);
1969         spin_unlock(&kvm->mmu_lock);
1970
1971         kvm_flush_remote_tlbs(kvm);
1972 }
1973
1974 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
1975 {
1976         struct kvm_mmu_page *page;
1977
1978         page = container_of(kvm->arch.active_mmu_pages.prev,
1979                             struct kvm_mmu_page, link);
1980         kvm_mmu_zap_page(kvm, page);
1981 }
1982
1983 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
1984 {
1985         struct kvm *kvm;
1986         struct kvm *kvm_freed = NULL;
1987         int cache_count = 0;
1988
1989         spin_lock(&kvm_lock);
1990
1991         list_for_each_entry(kvm, &vm_list, vm_list) {
1992                 int npages;
1993
1994                 if (!down_read_trylock(&kvm->slots_lock))
1995                         continue;
1996                 spin_lock(&kvm->mmu_lock);
1997                 npages = kvm->arch.n_alloc_mmu_pages -
1998                          kvm->arch.n_free_mmu_pages;
1999                 cache_count += npages;
2000                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2001                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2002                         cache_count--;
2003                         kvm_freed = kvm;
2004                 }
2005                 nr_to_scan--;
2006
2007                 spin_unlock(&kvm->mmu_lock);
2008                 up_read(&kvm->slots_lock);
2009         }
2010         if (kvm_freed)
2011                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2012
2013         spin_unlock(&kvm_lock);
2014
2015         return cache_count;
2016 }
2017
2018 static struct shrinker mmu_shrinker = {
2019         .shrink = mmu_shrink,
2020         .seeks = DEFAULT_SEEKS * 10,
2021 };
2022
2023 static void mmu_destroy_caches(void)
2024 {
2025         if (pte_chain_cache)
2026                 kmem_cache_destroy(pte_chain_cache);
2027         if (rmap_desc_cache)
2028                 kmem_cache_destroy(rmap_desc_cache);
2029         if (mmu_page_header_cache)
2030                 kmem_cache_destroy(mmu_page_header_cache);
2031 }
2032
2033 void kvm_mmu_module_exit(void)
2034 {
2035         mmu_destroy_caches();
2036         unregister_shrinker(&mmu_shrinker);
2037 }
2038
2039 int kvm_mmu_module_init(void)
2040 {
2041         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2042                                             sizeof(struct kvm_pte_chain),
2043                                             0, 0, NULL);
2044         if (!pte_chain_cache)
2045                 goto nomem;
2046         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2047                                             sizeof(struct kvm_rmap_desc),
2048                                             0, 0, NULL);
2049         if (!rmap_desc_cache)
2050                 goto nomem;
2051
2052         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2053                                                   sizeof(struct kvm_mmu_page),
2054                                                   0, 0, NULL);
2055         if (!mmu_page_header_cache)
2056                 goto nomem;
2057
2058         register_shrinker(&mmu_shrinker);
2059
2060         return 0;
2061
2062 nomem:
2063         mmu_destroy_caches();
2064         return -ENOMEM;
2065 }
2066
2067 /*
2068  * Caculate mmu pages needed for kvm.
2069  */
2070 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2071 {
2072         int i;
2073         unsigned int nr_mmu_pages;
2074         unsigned int  nr_pages = 0;
2075
2076         for (i = 0; i < kvm->nmemslots; i++)
2077                 nr_pages += kvm->memslots[i].npages;
2078
2079         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2080         nr_mmu_pages = max(nr_mmu_pages,
2081                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2082
2083         return nr_mmu_pages;
2084 }
2085
2086 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2087                                 unsigned len)
2088 {
2089         if (len > buffer->len)
2090                 return NULL;
2091         return buffer->ptr;
2092 }
2093
2094 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2095                                 unsigned len)
2096 {
2097         void *ret;
2098
2099         ret = pv_mmu_peek_buffer(buffer, len);
2100         if (!ret)
2101                 return ret;
2102         buffer->ptr += len;
2103         buffer->len -= len;
2104         buffer->processed += len;
2105         return ret;
2106 }
2107
2108 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2109                              gpa_t addr, gpa_t value)
2110 {
2111         int bytes = 8;
2112         int r;
2113
2114         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2115                 bytes = 4;
2116
2117         r = mmu_topup_memory_caches(vcpu);
2118         if (r)
2119                 return r;
2120
2121         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2122                 return -EFAULT;
2123
2124         return 1;
2125 }
2126
2127 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2128 {
2129         kvm_x86_ops->tlb_flush(vcpu);
2130         return 1;
2131 }
2132
2133 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2134 {
2135         spin_lock(&vcpu->kvm->mmu_lock);
2136         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2137         spin_unlock(&vcpu->kvm->mmu_lock);
2138         return 1;
2139 }
2140
2141 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2142                              struct kvm_pv_mmu_op_buffer *buffer)
2143 {
2144         struct kvm_mmu_op_header *header;
2145
2146         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2147         if (!header)
2148                 return 0;
2149         switch (header->op) {
2150         case KVM_MMU_OP_WRITE_PTE: {
2151                 struct kvm_mmu_op_write_pte *wpte;
2152
2153                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2154                 if (!wpte)
2155                         return 0;
2156                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2157                                         wpte->pte_val);
2158         }
2159         case KVM_MMU_OP_FLUSH_TLB: {
2160                 struct kvm_mmu_op_flush_tlb *ftlb;
2161
2162                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2163                 if (!ftlb)
2164                         return 0;
2165                 return kvm_pv_mmu_flush_tlb(vcpu);
2166         }
2167         case KVM_MMU_OP_RELEASE_PT: {
2168                 struct kvm_mmu_op_release_pt *rpt;
2169
2170                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2171                 if (!rpt)
2172                         return 0;
2173                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2174         }
2175         default: return 0;
2176         }
2177 }
2178
2179 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2180                   gpa_t addr, unsigned long *ret)
2181 {
2182         int r;
2183         struct kvm_pv_mmu_op_buffer buffer;
2184
2185         buffer.ptr = buffer.buf;
2186         buffer.len = min_t(unsigned long, bytes, sizeof buffer.buf);
2187         buffer.processed = 0;
2188
2189         r = kvm_read_guest(vcpu->kvm, addr, buffer.buf, buffer.len);
2190         if (r)
2191                 goto out;
2192
2193         while (buffer.len) {
2194                 r = kvm_pv_mmu_op_one(vcpu, &buffer);
2195                 if (r < 0)
2196                         goto out;
2197                 if (r == 0)
2198                         break;
2199         }
2200
2201         r = 1;
2202 out:
2203         *ret = buffer.processed;
2204         return r;
2205 }
2206
2207 #ifdef AUDIT
2208
2209 static const char *audit_msg;
2210
2211 static gva_t canonicalize(gva_t gva)
2212 {
2213 #ifdef CONFIG_X86_64
2214         gva = (long long)(gva << 16) >> 16;
2215 #endif
2216         return gva;
2217 }
2218
2219 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2220                                 gva_t va, int level)
2221 {
2222         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2223         int i;
2224         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2225
2226         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2227                 u64 ent = pt[i];
2228
2229                 if (ent == shadow_trap_nonpresent_pte)
2230                         continue;
2231
2232                 va = canonicalize(va);
2233                 if (level > 1) {
2234                         if (ent == shadow_notrap_nonpresent_pte)
2235                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2236                                        " in nonleaf level: levels %d gva %lx"
2237                                        " level %d pte %llx\n", audit_msg,
2238                                        vcpu->arch.mmu.root_level, va, level, ent);
2239
2240                         audit_mappings_page(vcpu, ent, va, level - 1);
2241                 } else {
2242                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2243                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2244
2245                         if (is_shadow_present_pte(ent)
2246                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2247                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2248                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2249                                        audit_msg, vcpu->arch.mmu.root_level,
2250                                        va, gpa, hpa, ent,
2251                                        is_shadow_present_pte(ent));
2252                         else if (ent == shadow_notrap_nonpresent_pte
2253                                  && !is_error_hpa(hpa))
2254                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2255                                        " valid guest gva %lx\n", audit_msg, va);
2256                         kvm_release_pfn_clean(pfn);
2257
2258                 }
2259         }
2260 }
2261
2262 static void audit_mappings(struct kvm_vcpu *vcpu)
2263 {
2264         unsigned i;
2265
2266         if (vcpu->arch.mmu.root_level == 4)
2267                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2268         else
2269                 for (i = 0; i < 4; ++i)
2270                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2271                                 audit_mappings_page(vcpu,
2272                                                     vcpu->arch.mmu.pae_root[i],
2273                                                     i << 30,
2274                                                     2);
2275 }
2276
2277 static int count_rmaps(struct kvm_vcpu *vcpu)
2278 {
2279         int nmaps = 0;
2280         int i, j, k;
2281
2282         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2283                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2284                 struct kvm_rmap_desc *d;
2285
2286                 for (j = 0; j < m->npages; ++j) {
2287                         unsigned long *rmapp = &m->rmap[j];
2288
2289                         if (!*rmapp)
2290                                 continue;
2291                         if (!(*rmapp & 1)) {
2292                                 ++nmaps;
2293                                 continue;
2294                         }
2295                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2296                         while (d) {
2297                                 for (k = 0; k < RMAP_EXT; ++k)
2298                                         if (d->shadow_ptes[k])
2299                                                 ++nmaps;
2300                                         else
2301                                                 break;
2302                                 d = d->more;
2303                         }
2304                 }
2305         }
2306         return nmaps;
2307 }
2308
2309 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2310 {
2311         int nmaps = 0;
2312         struct kvm_mmu_page *sp;
2313         int i;
2314
2315         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2316                 u64 *pt = sp->spt;
2317
2318                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2319                         continue;
2320
2321                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2322                         u64 ent = pt[i];
2323
2324                         if (!(ent & PT_PRESENT_MASK))
2325                                 continue;
2326                         if (!(ent & PT_WRITABLE_MASK))
2327                                 continue;
2328                         ++nmaps;
2329                 }
2330         }
2331         return nmaps;
2332 }
2333
2334 static void audit_rmap(struct kvm_vcpu *vcpu)
2335 {
2336         int n_rmap = count_rmaps(vcpu);
2337         int n_actual = count_writable_mappings(vcpu);
2338
2339         if (n_rmap != n_actual)
2340                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2341                        __func__, audit_msg, n_rmap, n_actual);
2342 }
2343
2344 static void audit_write_protection(struct kvm_vcpu *vcpu)
2345 {
2346         struct kvm_mmu_page *sp;
2347         struct kvm_memory_slot *slot;
2348         unsigned long *rmapp;
2349         gfn_t gfn;
2350
2351         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2352                 if (sp->role.metaphysical)
2353                         continue;
2354
2355                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2356                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2357                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2358                 if (*rmapp)
2359                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2360                                " mappings: gfn %lx role %x\n",
2361                                __func__, audit_msg, sp->gfn,
2362                                sp->role.word);
2363         }
2364 }
2365
2366 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2367 {
2368         int olddbg = dbg;
2369
2370         dbg = 0;
2371         audit_msg = msg;
2372         audit_rmap(vcpu);
2373         audit_write_protection(vcpu);
2374         audit_mappings(vcpu);
2375         dbg = olddbg;
2376 }
2377
2378 #endif