SAFE public projects git trees. - safe/jmp/linux-2.6/blob - arch/x86/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "mmu.h"
  21 #include "kvm_cache_regs.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 #include <asm/vmx.h>
  37
  38 /*
  39  * When setting this variable to true it enables Two-Dimensional-Paging
  40  * where the hardware walks 2 page tables:
  41  * 1. the guest-virtual to guest-physical
  42  * 2. while doing 1. it walks guest-physical to host-physical
  43  * If the hardware supports that we don't need to do shadow paging.
  44  */
  45 bool tdp_enabled = false;
  46
  47 #undef MMU_DEBUG
  48
  49 #undef AUDIT
  50
  51 #ifdef AUDIT
  52 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  53 #else
  54 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  55 #endif
  56
  57 #ifdef MMU_DEBUG
  58
  59 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  60 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  61
  62 #else
  63
  64 #define pgprintk(x...) do { } while (0)
  65 #define rmap_printk(x...) do { } while (0)
  66
  67 #endif
  68
  69 #if defined(MMU_DEBUG) || defined(AUDIT)
  70 static int dbg = 0;
  71 module_param(dbg, bool, 0644);
  72 #endif
  73
  74 static int oos_shadow = 1;
  75 module_param(oos_shadow, bool, 0644);
  76
  77 #ifndef MMU_DEBUG
  78 #define ASSERT(x) do { } while (0)
  79 #else
  80 #define ASSERT(x)                                                       \
  81         if (!(x)) {                                                     \
  82                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  83                        __FILE__, __LINE__, #x);                         \
  84         }
  85 #endif
  86
  87 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  88 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  89
  90 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  91
  92 #define PT64_LEVEL_BITS 9
  93
  94 #define PT64_LEVEL_SHIFT(level) \
  95                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  96
  97 #define PT64_LEVEL_MASK(level) \
  98                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  99
 100 #define PT64_INDEX(address, level)\
 101         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 102
 103
 104 #define PT32_LEVEL_BITS 10
 105
 106 #define PT32_LEVEL_SHIFT(level) \
 107                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 108
 109 #define PT32_LEVEL_MASK(level) \
 110                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 111
 112 #define PT32_INDEX(address, level)\
 113         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 114
 115
 116 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 117 #define PT64_DIR_BASE_ADDR_MASK \
 118         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 119
 120 #define PT32_BASE_ADDR_MASK PAGE_MASK
 121 #define PT32_DIR_BASE_ADDR_MASK \
 122         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 123
 124 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 125                         | PT64_NX_MASK)
 126
 127 #define PFERR_PRESENT_MASK (1U << 0)
 128 #define PFERR_WRITE_MASK (1U << 1)
 129 #define PFERR_USER_MASK (1U << 2)
 130 #define PFERR_RSVD_MASK (1U << 3)
 131 #define PFERR_FETCH_MASK (1U << 4)
 132
 133 #define PT_DIRECTORY_LEVEL 2
 134 #define PT_PAGE_TABLE_LEVEL 1
 135
 136 #define RMAP_EXT 4
 137
 138 #define ACC_EXEC_MASK    1
 139 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 140 #define ACC_USER_MASK    PT_USER_MASK
 141 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 142
 143 #define CREATE_TRACE_POINTS
 144 #include "mmutrace.h"
 145
 146 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 147
 148 struct kvm_rmap_desc {
 149         u64 *sptes[RMAP_EXT];
 150         struct kvm_rmap_desc *more;
 151 };
 152
 153 struct kvm_shadow_walk_iterator {
 154         u64 addr;
 155         hpa_t shadow_addr;
 156         int level;
 157         u64 *sptep;
 158         unsigned index;
 159 };
 160
 161 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
 162         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
 163              shadow_walk_okay(&(_walker));                      \
 164              shadow_walk_next(&(_walker)))
 165
 166
 167 struct kvm_unsync_walk {
 168         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 169 };
 170
 171 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 172
 173 static struct kmem_cache *pte_chain_cache;
 174 static struct kmem_cache *rmap_desc_cache;
 175 static struct kmem_cache *mmu_page_header_cache;
 176
 177 static u64 __read_mostly shadow_trap_nonpresent_pte;
 178 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 179 static u64 __read_mostly shadow_base_present_pte;
 180 static u64 __read_mostly shadow_nx_mask;
 181 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 182 static u64 __read_mostly shadow_user_mask;
 183 static u64 __read_mostly shadow_accessed_mask;
 184 static u64 __read_mostly shadow_dirty_mask;
 185
 186 static inline u64 rsvd_bits(int s, int e)
 187 {
 188         return ((1ULL << (e - s + 1)) - 1) << s;
 189 }
 190
 191 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 192 {
 193         shadow_trap_nonpresent_pte = trap_pte;
 194         shadow_notrap_nonpresent_pte = notrap_pte;
 195 }
 196 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 197
 198 void kvm_mmu_set_base_ptes(u64 base_pte)
 199 {
 200         shadow_base_present_pte = base_pte;
 201 }
 202 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 203
 204 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 205                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 206 {
 207         shadow_user_mask = user_mask;
 208         shadow_accessed_mask = accessed_mask;
 209         shadow_dirty_mask = dirty_mask;
 210         shadow_nx_mask = nx_mask;
 211         shadow_x_mask = x_mask;
 212 }
 213 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 214
 215 static int is_write_protection(struct kvm_vcpu *vcpu)
 216 {
 217         return vcpu->arch.cr0 & X86_CR0_WP;
 218 }
 219
 220 static int is_cpuid_PSE36(void)
 221 {
 222         return 1;
 223 }
 224
 225 static int is_nx(struct kvm_vcpu *vcpu)
 226 {
 227         return vcpu->arch.shadow_efer & EFER_NX;
 228 }
 229
 230 static int is_shadow_present_pte(u64 pte)
 231 {
 232         return pte != shadow_trap_nonpresent_pte
 233                 && pte != shadow_notrap_nonpresent_pte;
 234 }
 235
 236 static int is_large_pte(u64 pte)
 237 {
 238         return pte & PT_PAGE_SIZE_MASK;
 239 }
 240
 241 static int is_writeble_pte(unsigned long pte)
 242 {
 243         return pte & PT_WRITABLE_MASK;
 244 }
 245
 246 static int is_dirty_gpte(unsigned long pte)
 247 {
 248         return pte & PT_DIRTY_MASK;
 249 }
 250
 251 static int is_rmap_spte(u64 pte)
 252 {
 253         return is_shadow_present_pte(pte);
 254 }
 255
 256 static int is_last_spte(u64 pte, int level)
 257 {
 258         if (level == PT_PAGE_TABLE_LEVEL)
 259                 return 1;
 260         if (level == PT_DIRECTORY_LEVEL && is_large_pte(pte))
 261                 return 1;
 262         return 0;
 263 }
 264
 265 static pfn_t spte_to_pfn(u64 pte)
 266 {
 267         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 268 }
 269
 270 static gfn_t pse36_gfn_delta(u32 gpte)
 271 {
 272         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 273
 274         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 275 }
 276
 277 static void __set_spte(u64 *sptep, u64 spte)
 278 {
 279 #ifdef CONFIG_X86_64
 280         set_64bit((unsigned long *)sptep, spte);
 281 #else
 282         set_64bit((unsigned long long *)sptep, spte);
 283 #endif
 284 }
 285
 286 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 287                                   struct kmem_cache *base_cache, int min)
 288 {
 289         void *obj;
 290
 291         if (cache->nobjs >= min)
 292                 return 0;
 293         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 294                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 295                 if (!obj)
 296                         return -ENOMEM;
 297                 cache->objects[cache->nobjs++] = obj;
 298         }
 299         return 0;
 300 }
 301
 302 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 303 {
 304         while (mc->nobjs)
 305                 kfree(mc->objects[--mc->nobjs]);
 306 }
 307
 308 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 309                                        int min)
 310 {
 311         struct page *page;
 312
 313         if (cache->nobjs >= min)
 314                 return 0;
 315         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 316                 page = alloc_page(GFP_KERNEL);
 317                 if (!page)
 318                         return -ENOMEM;
 319                 set_page_private(page, 0);
 320                 cache->objects[cache->nobjs++] = page_address(page);
 321         }
 322         return 0;
 323 }
 324
 325 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 326 {
 327         while (mc->nobjs)
 328                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 329 }
 330
 331 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 332 {
 333         int r;
 334
 335         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 336                                    pte_chain_cache, 4);
 337         if (r)
 338                 goto out;
 339         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 340                                    rmap_desc_cache, 4);
 341         if (r)
 342                 goto out;
 343         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 344         if (r)
 345                 goto out;
 346         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 347                                    mmu_page_header_cache, 4);
 348 out:
 349         return r;
 350 }
 351
 352 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 353 {
 354         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 355         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 356         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 357         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 358 }
 359
 360 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 361                                     size_t size)
 362 {
 363         void *p;
 364
 365         BUG_ON(!mc->nobjs);
 366         p = mc->objects[--mc->nobjs];
 367         return p;
 368 }
 369
 370 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 371 {
 372         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 373                                       sizeof(struct kvm_pte_chain));
 374 }
 375
 376 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 377 {
 378         kfree(pc);
 379 }
 380
 381 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 382 {
 383         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 384                                       sizeof(struct kvm_rmap_desc));
 385 }
 386
 387 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 388 {
 389         kfree(rd);
 390 }
 391
 392 /*
 393  * Return the pointer to the largepage write count for a given
 394  * gfn, handling slots that are not large page aligned.
 395  */
 396 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 397 {
 398         unsigned long idx;
 399
 400         idx = (gfn / KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL)) -
 401               (slot->base_gfn / KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL));
 402         return &slot->lpage_info[0][idx].write_count;
 403 }
 404
 405 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 406 {
 407         int *write_count;
 408
 409         gfn = unalias_gfn(kvm, gfn);
 410         write_count = slot_largepage_idx(gfn,
 411                                          gfn_to_memslot_unaliased(kvm, gfn));
 412         *write_count += 1;
 413 }
 414
 415 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 416 {
 417         int *write_count;
 418
 419         gfn = unalias_gfn(kvm, gfn);
 420         write_count = slot_largepage_idx(gfn,
 421                                          gfn_to_memslot_unaliased(kvm, gfn));
 422         *write_count -= 1;
 423         WARN_ON(*write_count < 0);
 424 }
 425
 426 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 427 {
 428         struct kvm_memory_slot *slot;
 429         int *largepage_idx;
 430
 431         gfn = unalias_gfn(kvm, gfn);
 432         slot = gfn_to_memslot_unaliased(kvm, gfn);
 433         if (slot) {
 434                 largepage_idx = slot_largepage_idx(gfn, slot);
 435                 return *largepage_idx;
 436         }
 437
 438         return 1;
 439 }
 440
 441 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 442 {
 443         struct vm_area_struct *vma;
 444         unsigned long addr;
 445         int ret = 0;
 446
 447         addr = gfn_to_hva(kvm, gfn);
 448         if (kvm_is_error_hva(addr))
 449                 return ret;
 450
 451         down_read(&current->mm->mmap_sem);
 452         vma = find_vma(current->mm, addr);
 453         if (vma && is_vm_hugetlb_page(vma))
 454                 ret = 1;
 455         up_read(&current->mm->mmap_sem);
 456
 457         return ret;
 458 }
 459
 460 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 461 {
 462         struct kvm_memory_slot *slot;
 463
 464         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 465                 return 0;
 466
 467         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 468                 return 0;
 469
 470         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 471         if (slot && slot->dirty_bitmap)
 472                 return 0;
 473
 474         return 1;
 475 }
 476
 477 /*
 478  * Take gfn and return the reverse mapping to it.
 479  * Note: gfn must be unaliased before this function get called
 480  */
 481
 482 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 483 {
 484         struct kvm_memory_slot *slot;
 485         unsigned long idx;
 486
 487         slot = gfn_to_memslot(kvm, gfn);
 488         if (!lpage)
 489                 return &slot->rmap[gfn - slot->base_gfn];
 490
 491         idx = (gfn / KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL)) -
 492               (slot->base_gfn / KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL));
 493
 494         return &slot->lpage_info[0][idx].rmap_pde;
 495 }
 496
 497 /*
 498  * Reverse mapping data structures:
 499  *
 500  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 501  * that points to page_address(page).
 502  *
 503  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 504  * containing more mappings.
 505  *
 506  * Returns the number of rmap entries before the spte was added or zero if
 507  * the spte was not added.
 508  *
 509  */
 510 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 511 {
 512         struct kvm_mmu_page *sp;
 513         struct kvm_rmap_desc *desc;
 514         unsigned long *rmapp;
 515         int i, count = 0;
 516
 517         if (!is_rmap_spte(*spte))
 518                 return count;
 519         gfn = unalias_gfn(vcpu->kvm, gfn);
 520         sp = page_header(__pa(spte));
 521         sp->gfns[spte - sp->spt] = gfn;
 522         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 523         if (!*rmapp) {
 524                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 525                 *rmapp = (unsigned long)spte;
 526         } else if (!(*rmapp & 1)) {
 527                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 528                 desc = mmu_alloc_rmap_desc(vcpu);
 529                 desc->sptes[0] = (u64 *)*rmapp;
 530                 desc->sptes[1] = spte;
 531                 *rmapp = (unsigned long)desc | 1;
 532         } else {
 533                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 534                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 535                 while (desc->sptes[RMAP_EXT-1] && desc->more) {
 536                         desc = desc->more;
 537                         count += RMAP_EXT;
 538                 }
 539                 if (desc->sptes[RMAP_EXT-1]) {
 540                         desc->more = mmu_alloc_rmap_desc(vcpu);
 541                         desc = desc->more;
 542                 }
 543                 for (i = 0; desc->sptes[i]; ++i)
 544                         ;
 545                 desc->sptes[i] = spte;
 546         }
 547         return count;
 548 }
 549
 550 static void rmap_desc_remove_entry(unsigned long *rmapp,
 551                                    struct kvm_rmap_desc *desc,
 552                                    int i,
 553                                    struct kvm_rmap_desc *prev_desc)
 554 {
 555         int j;
 556
 557         for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
 558                 ;
 559         desc->sptes[i] = desc->sptes[j];
 560         desc->sptes[j] = NULL;
 561         if (j != 0)
 562                 return;
 563         if (!prev_desc && !desc->more)
 564                 *rmapp = (unsigned long)desc->sptes[0];
 565         else
 566                 if (prev_desc)
 567                         prev_desc->more = desc->more;
 568                 else
 569                         *rmapp = (unsigned long)desc->more | 1;
 570         mmu_free_rmap_desc(desc);
 571 }
 572
 573 static void rmap_remove(struct kvm *kvm, u64 *spte)
 574 {
 575         struct kvm_rmap_desc *desc;
 576         struct kvm_rmap_desc *prev_desc;
 577         struct kvm_mmu_page *sp;
 578         pfn_t pfn;
 579         unsigned long *rmapp;
 580         int i;
 581
 582         if (!is_rmap_spte(*spte))
 583                 return;
 584         sp = page_header(__pa(spte));
 585         pfn = spte_to_pfn(*spte);
 586         if (*spte & shadow_accessed_mask)
 587                 kvm_set_pfn_accessed(pfn);
 588         if (is_writeble_pte(*spte))
 589                 kvm_release_pfn_dirty(pfn);
 590         else
 591                 kvm_release_pfn_clean(pfn);
 592         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 593         if (!*rmapp) {
 594                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 595                 BUG();
 596         } else if (!(*rmapp & 1)) {
 597                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 598                 if ((u64 *)*rmapp != spte) {
 599                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 600                                spte, *spte);
 601                         BUG();
 602                 }
 603                 *rmapp = 0;
 604         } else {
 605                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 606                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 607                 prev_desc = NULL;
 608                 while (desc) {
 609                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
 610                                 if (desc->sptes[i] == spte) {
 611                                         rmap_desc_remove_entry(rmapp,
 612                                                                desc, i,
 613                                                                prev_desc);
 614                                         return;
 615                                 }
 616                         prev_desc = desc;
 617                         desc = desc->more;
 618                 }
 619                 BUG();
 620         }
 621 }
 622
 623 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 624 {
 625         struct kvm_rmap_desc *desc;
 626         struct kvm_rmap_desc *prev_desc;
 627         u64 *prev_spte;
 628         int i;
 629
 630         if (!*rmapp)
 631                 return NULL;
 632         else if (!(*rmapp & 1)) {
 633                 if (!spte)
 634                         return (u64 *)*rmapp;
 635                 return NULL;
 636         }
 637         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 638         prev_desc = NULL;
 639         prev_spte = NULL;
 640         while (desc) {
 641                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
 642                         if (prev_spte == spte)
 643                                 return desc->sptes[i];
 644                         prev_spte = desc->sptes[i];
 645                 }
 646                 desc = desc->more;
 647         }
 648         return NULL;
 649 }
 650
 651 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 652 {
 653         unsigned long *rmapp;
 654         u64 *spte;
 655         int write_protected = 0;
 656
 657         gfn = unalias_gfn(kvm, gfn);
 658         rmapp = gfn_to_rmap(kvm, gfn, 0);
 659
 660         spte = rmap_next(kvm, rmapp, NULL);
 661         while (spte) {
 662                 BUG_ON(!spte);
 663                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 664                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 665                 if (is_writeble_pte(*spte)) {
 666                         __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
 667                         write_protected = 1;
 668                 }
 669                 spte = rmap_next(kvm, rmapp, spte);
 670         }
 671         if (write_protected) {
 672                 pfn_t pfn;
 673
 674                 spte = rmap_next(kvm, rmapp, NULL);
 675                 pfn = spte_to_pfn(*spte);
 676                 kvm_set_pfn_dirty(pfn);
 677         }
 678
 679         /* check for huge page mappings */
 680         rmapp = gfn_to_rmap(kvm, gfn, 1);
 681         spte = rmap_next(kvm, rmapp, NULL);
 682         while (spte) {
 683                 BUG_ON(!spte);
 684                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 685                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 686                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 687                 if (is_writeble_pte(*spte)) {
 688                         rmap_remove(kvm, spte);
 689                         --kvm->stat.lpages;
 690                         __set_spte(spte, shadow_trap_nonpresent_pte);
 691                         spte = NULL;
 692                         write_protected = 1;
 693                 }
 694                 spte = rmap_next(kvm, rmapp, spte);
 695         }
 696
 697         return write_protected;
 698 }
 699
 700 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 701 {
 702         u64 *spte;
 703         int need_tlb_flush = 0;
 704
 705         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 706                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 707                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 708                 rmap_remove(kvm, spte);
 709                 __set_spte(spte, shadow_trap_nonpresent_pte);
 710                 need_tlb_flush = 1;
 711         }
 712         return need_tlb_flush;
 713 }
 714
 715 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 716                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 717 {
 718         int i;
 719         int retval = 0;
 720
 721         /*
 722          * If mmap_sem isn't taken, we can look the memslots with only
 723          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 724          */
 725         for (i = 0; i < kvm->nmemslots; i++) {
 726                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 727                 unsigned long start = memslot->userspace_addr;
 728                 unsigned long end;
 729
 730                 /* mmu_lock protects userspace_addr */
 731                 if (!start)
 732                         continue;
 733
 734                 end = start + (memslot->npages << PAGE_SHIFT);
 735                 if (hva >= start && hva < end) {
 736                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 737                         int idx = gfn_offset /
 738                                   KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL);
 739                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 740                         retval |= handler(kvm,
 741                                         &memslot->lpage_info[0][idx].rmap_pde);
 742                 }
 743         }
 744
 745         return retval;
 746 }
 747
 748 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 749 {
 750         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 751 }
 752
 753 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 754 {
 755         u64 *spte;
 756         int young = 0;
 757
 758         /* always return old for EPT */
 759         if (!shadow_accessed_mask)
 760                 return 0;
 761
 762         spte = rmap_next(kvm, rmapp, NULL);
 763         while (spte) {
 764                 int _young;
 765                 u64 _spte = *spte;
 766                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 767                 _young = _spte & PT_ACCESSED_MASK;
 768                 if (_young) {
 769                         young = 1;
 770                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 771                 }
 772                 spte = rmap_next(kvm, rmapp, spte);
 773         }
 774         return young;
 775 }
 776
 777 #define RMAP_RECYCLE_THRESHOLD 1000
 778
 779 static void rmap_recycle(struct kvm_vcpu *vcpu, gfn_t gfn, int lpage)
 780 {
 781         unsigned long *rmapp;
 782
 783         gfn = unalias_gfn(vcpu->kvm, gfn);
 784         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 785
 786         kvm_unmap_rmapp(vcpu->kvm, rmapp);
 787         kvm_flush_remote_tlbs(vcpu->kvm);
 788 }
 789
 790 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 791 {
 792         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 793 }
 794
 795 #ifdef MMU_DEBUG
 796 static int is_empty_shadow_page(u64 *spt)
 797 {
 798         u64 *pos;
 799         u64 *end;
 800
 801         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 802                 if (is_shadow_present_pte(*pos)) {
 803                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 804                                pos, *pos);
 805                         return 0;
 806                 }
 807         return 1;
 808 }
 809 #endif
 810
 811 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 812 {
 813         ASSERT(is_empty_shadow_page(sp->spt));
 814         list_del(&sp->link);
 815         __free_page(virt_to_page(sp->spt));
 816         __free_page(virt_to_page(sp->gfns));
 817         kfree(sp);
 818         ++kvm->arch.n_free_mmu_pages;
 819 }
 820
 821 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 822 {
 823         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 824 }
 825
 826 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 827                                                u64 *parent_pte)
 828 {
 829         struct kvm_mmu_page *sp;
 830
 831         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 832         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 833         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 834         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 835         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 836         INIT_LIST_HEAD(&sp->oos_link);
 837         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 838         sp->multimapped = 0;
 839         sp->parent_pte = parent_pte;
 840         --vcpu->kvm->arch.n_free_mmu_pages;
 841         return sp;
 842 }
 843
 844 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 845                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 846 {
 847         struct kvm_pte_chain *pte_chain;
 848         struct hlist_node *node;
 849         int i;
 850
 851         if (!parent_pte)
 852                 return;
 853         if (!sp->multimapped) {
 854                 u64 *old = sp->parent_pte;
 855
 856                 if (!old) {
 857                         sp->parent_pte = parent_pte;
 858                         return;
 859                 }
 860                 sp->multimapped = 1;
 861                 pte_chain = mmu_alloc_pte_chain(vcpu);
 862                 INIT_HLIST_HEAD(&sp->parent_ptes);
 863                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 864                 pte_chain->parent_ptes[0] = old;
 865         }
 866         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 867                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 868                         continue;
 869                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 870                         if (!pte_chain->parent_ptes[i]) {
 871                                 pte_chain->parent_ptes[i] = parent_pte;
 872                                 return;
 873                         }
 874         }
 875         pte_chain = mmu_alloc_pte_chain(vcpu);
 876         BUG_ON(!pte_chain);
 877         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 878         pte_chain->parent_ptes[0] = parent_pte;
 879 }
 880
 881 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 882                                        u64 *parent_pte)
 883 {
 884         struct kvm_pte_chain *pte_chain;
 885         struct hlist_node *node;
 886         int i;
 887
 888         if (!sp->multimapped) {
 889                 BUG_ON(sp->parent_pte != parent_pte);
 890                 sp->parent_pte = NULL;
 891                 return;
 892         }
 893         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 894                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 895                         if (!pte_chain->parent_ptes[i])
 896                                 break;
 897                         if (pte_chain->parent_ptes[i] != parent_pte)
 898                                 continue;
 899                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 900                                 && pte_chain->parent_ptes[i + 1]) {
 901                                 pte_chain->parent_ptes[i]
 902                                         = pte_chain->parent_ptes[i + 1];
 903                                 ++i;
 904                         }
 905                         pte_chain->parent_ptes[i] = NULL;
 906                         if (i == 0) {
 907                                 hlist_del(&pte_chain->link);
 908                                 mmu_free_pte_chain(pte_chain);
 909                                 if (hlist_empty(&sp->parent_ptes)) {
 910                                         sp->multimapped = 0;
 911                                         sp->parent_pte = NULL;
 912                                 }
 913                         }
 914                         return;
 915                 }
 916         BUG();
 917 }
 918
 919
 920 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 921                             mmu_parent_walk_fn fn)
 922 {
 923         struct kvm_pte_chain *pte_chain;
 924         struct hlist_node *node;
 925         struct kvm_mmu_page *parent_sp;
 926         int i;
 927
 928         if (!sp->multimapped && sp->parent_pte) {
 929                 parent_sp = page_header(__pa(sp->parent_pte));
 930                 fn(vcpu, parent_sp);
 931                 mmu_parent_walk(vcpu, parent_sp, fn);
 932                 return;
 933         }
 934         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 935                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 936                         if (!pte_chain->parent_ptes[i])
 937                                 break;
 938                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 939                         fn(vcpu, parent_sp);
 940                         mmu_parent_walk(vcpu, parent_sp, fn);
 941                 }
 942 }
 943
 944 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 945 {
 946         unsigned int index;
 947         struct kvm_mmu_page *sp = page_header(__pa(spte));
 948
 949         index = spte - sp->spt;
 950         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
 951                 sp->unsync_children++;
 952         WARN_ON(!sp->unsync_children);
 953 }
 954
 955 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 956 {
 957         struct kvm_pte_chain *pte_chain;
 958         struct hlist_node *node;
 959         int i;
 960
 961         if (!sp->parent_pte)
 962                 return;
 963
 964         if (!sp->multimapped) {
 965                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 966                 return;
 967         }
 968
 969         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 970                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 971                         if (!pte_chain->parent_ptes[i])
 972                                 break;
 973                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 974                 }
 975 }
 976
 977 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 978 {
 979         kvm_mmu_update_parents_unsync(sp);
 980         return 1;
 981 }
 982
 983 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 984                                         struct kvm_mmu_page *sp)
 985 {
 986         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 987         kvm_mmu_update_parents_unsync(sp);
 988 }
 989
 990 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 991                                     struct kvm_mmu_page *sp)
 992 {
 993         int i;
 994
 995         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 996                 sp->spt[i] = shadow_trap_nonpresent_pte;
 997 }
 998
 999 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1000                                struct kvm_mmu_page *sp)
1001 {
1002         return 1;
1003 }
1004
1005 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1006 {
1007 }
1008
1009 #define KVM_PAGE_ARRAY_NR 16
1010
1011 struct kvm_mmu_pages {
1012         struct mmu_page_and_offset {
1013                 struct kvm_mmu_page *sp;
1014                 unsigned int idx;
1015         } page[KVM_PAGE_ARRAY_NR];
1016         unsigned int nr;
1017 };
1018
1019 #define for_each_unsync_children(bitmap, idx)           \
1020         for (idx = find_first_bit(bitmap, 512);         \
1021              idx < 512;                                 \
1022              idx = find_next_bit(bitmap, 512, idx+1))
1023
1024 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1025                          int idx)
1026 {
1027         int i;
1028
1029         if (sp->unsync)
1030                 for (i=0; i < pvec->nr; i++)
1031                         if (pvec->page[i].sp == sp)
1032                                 return 0;
1033
1034         pvec->page[pvec->nr].sp = sp;
1035         pvec->page[pvec->nr].idx = idx;
1036         pvec->nr++;
1037         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1038 }
1039
1040 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1041                            struct kvm_mmu_pages *pvec)
1042 {
1043         int i, ret, nr_unsync_leaf = 0;
1044
1045         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1046                 u64 ent = sp->spt[i];
1047
1048                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1049                         struct kvm_mmu_page *child;
1050                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1051
1052                         if (child->unsync_children) {
1053                                 if (mmu_pages_add(pvec, child, i))
1054                                         return -ENOSPC;
1055
1056                                 ret = __mmu_unsync_walk(child, pvec);
1057                                 if (!ret)
1058                                         __clear_bit(i, sp->unsync_child_bitmap);
1059                                 else if (ret > 0)
1060                                         nr_unsync_leaf += ret;
1061                                 else
1062                                         return ret;
1063                         }
1064
1065                         if (child->unsync) {
1066                                 nr_unsync_leaf++;
1067                                 if (mmu_pages_add(pvec, child, i))
1068                                         return -ENOSPC;
1069                         }
1070                 }
1071         }
1072
1073         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1074                 sp->unsync_children = 0;
1075
1076         return nr_unsync_leaf;
1077 }
1078
1079 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1080                            struct kvm_mmu_pages *pvec)
1081 {
1082         if (!sp->unsync_children)
1083                 return 0;
1084
1085         mmu_pages_add(pvec, sp, 0);
1086         return __mmu_unsync_walk(sp, pvec);
1087 }
1088
1089 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1090 {
1091         unsigned index;
1092         struct hlist_head *bucket;
1093         struct kvm_mmu_page *sp;
1094         struct hlist_node *node;
1095
1096         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1097         index = kvm_page_table_hashfn(gfn);
1098         bucket = &kvm->arch.mmu_page_hash[index];
1099         hlist_for_each_entry(sp, node, bucket, hash_link)
1100                 if (sp->gfn == gfn && !sp->role.direct
1101                     && !sp->role.invalid) {
1102                         pgprintk("%s: found role %x\n",
1103                                  __func__, sp->role.word);
1104                         return sp;
1105                 }
1106         return NULL;
1107 }
1108
1109 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1110 {
1111         WARN_ON(!sp->unsync);
1112         sp->unsync = 0;
1113         --kvm->stat.mmu_unsync;
1114 }
1115
1116 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1117
1118 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1119 {
1120         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1121                 kvm_mmu_zap_page(vcpu->kvm, sp);
1122                 return 1;
1123         }
1124
1125         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1126                 kvm_flush_remote_tlbs(vcpu->kvm);
1127         kvm_unlink_unsync_page(vcpu->kvm, sp);
1128         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1129                 kvm_mmu_zap_page(vcpu->kvm, sp);
1130                 return 1;
1131         }
1132
1133         kvm_mmu_flush_tlb(vcpu);
1134         return 0;
1135 }
1136
1137 struct mmu_page_path {
1138         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1139         unsigned int idx[PT64_ROOT_LEVEL-1];
1140 };
1141
1142 #define for_each_sp(pvec, sp, parents, i)                       \
1143                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1144                         sp = pvec.page[i].sp;                   \
1145                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1146                         i = mmu_pages_next(&pvec, &parents, i))
1147
1148 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1149                           struct mmu_page_path *parents,
1150                           int i)
1151 {
1152         int n;
1153
1154         for (n = i+1; n < pvec->nr; n++) {
1155                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1156
1157                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1158                         parents->idx[0] = pvec->page[n].idx;
1159                         return n;
1160                 }
1161
1162                 parents->parent[sp->role.level-2] = sp;
1163                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1164         }
1165
1166         return n;
1167 }
1168
1169 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1170 {
1171         struct kvm_mmu_page *sp;
1172         unsigned int level = 0;
1173
1174         do {
1175                 unsigned int idx = parents->idx[level];
1176
1177                 sp = parents->parent[level];
1178                 if (!sp)
1179                         return;
1180
1181                 --sp->unsync_children;
1182                 WARN_ON((int)sp->unsync_children < 0);
1183                 __clear_bit(idx, sp->unsync_child_bitmap);
1184                 level++;
1185         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1186 }
1187
1188 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1189                                struct mmu_page_path *parents,
1190                                struct kvm_mmu_pages *pvec)
1191 {
1192         parents->parent[parent->role.level-1] = NULL;
1193         pvec->nr = 0;
1194 }
1195
1196 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1197                               struct kvm_mmu_page *parent)
1198 {
1199         int i;
1200         struct kvm_mmu_page *sp;
1201         struct mmu_page_path parents;
1202         struct kvm_mmu_pages pages;
1203
1204         kvm_mmu_pages_init(parent, &parents, &pages);
1205         while (mmu_unsync_walk(parent, &pages)) {
1206                 int protected = 0;
1207
1208                 for_each_sp(pages, sp, parents, i)
1209                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1210
1211                 if (protected)
1212                         kvm_flush_remote_tlbs(vcpu->kvm);
1213
1214                 for_each_sp(pages, sp, parents, i) {
1215                         kvm_sync_page(vcpu, sp);
1216                         mmu_pages_clear_parents(&parents);
1217                 }
1218                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1219                 kvm_mmu_pages_init(parent, &parents, &pages);
1220         }
1221 }
1222
1223 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1224                                              gfn_t gfn,
1225                                              gva_t gaddr,
1226                                              unsigned level,
1227                                              int direct,
1228                                              unsigned access,
1229                                              u64 *parent_pte)
1230 {
1231         union kvm_mmu_page_role role;
1232         unsigned index;
1233         unsigned quadrant;
1234         struct hlist_head *bucket;
1235         struct kvm_mmu_page *sp;
1236         struct hlist_node *node, *tmp;
1237
1238         role = vcpu->arch.mmu.base_role;
1239         role.level = level;
1240         role.direct = direct;
1241         role.access = access;
1242         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1243                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1244                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1245                 role.quadrant = quadrant;
1246         }
1247         pgprintk("%s: looking gfn %lx role %x\n", __func__,
1248                  gfn, role.word);
1249         index = kvm_page_table_hashfn(gfn);
1250         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1251         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1252                 if (sp->gfn == gfn) {
1253                         if (sp->unsync)
1254                                 if (kvm_sync_page(vcpu, sp))
1255                                         continue;
1256
1257                         if (sp->role.word != role.word)
1258                                 continue;
1259
1260                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1261                         if (sp->unsync_children) {
1262                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1263                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1264                         }
1265                         pgprintk("%s: found\n", __func__);
1266                         return sp;
1267                 }
1268         ++vcpu->kvm->stat.mmu_cache_miss;
1269         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1270         if (!sp)
1271                 return sp;
1272         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1273         sp->gfn = gfn;
1274         sp->role = role;
1275         hlist_add_head(&sp->hash_link, bucket);
1276         if (!direct) {
1277                 if (rmap_write_protect(vcpu->kvm, gfn))
1278                         kvm_flush_remote_tlbs(vcpu->kvm);
1279                 account_shadowed(vcpu->kvm, gfn);
1280         }
1281         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1282                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1283         else
1284                 nonpaging_prefetch_page(vcpu, sp);
1285         return sp;
1286 }
1287
1288 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1289                              struct kvm_vcpu *vcpu, u64 addr)
1290 {
1291         iterator->addr = addr;
1292         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1293         iterator->level = vcpu->arch.mmu.shadow_root_level;
1294         if (iterator->level == PT32E_ROOT_LEVEL) {
1295                 iterator->shadow_addr
1296                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1297                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1298                 --iterator->level;
1299                 if (!iterator->shadow_addr)
1300                         iterator->level = 0;
1301         }
1302 }
1303
1304 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1305 {
1306         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1307                 return false;
1308
1309         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1310                 if (is_large_pte(*iterator->sptep))
1311                         return false;
1312
1313         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1314         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1315         return true;
1316 }
1317
1318 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1319 {
1320         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1321         --iterator->level;
1322 }
1323
1324 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1325                                          struct kvm_mmu_page *sp)
1326 {
1327         unsigned i;
1328         u64 *pt;
1329         u64 ent;
1330
1331         pt = sp->spt;
1332
1333         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1334                 ent = pt[i];
1335
1336                 if (is_shadow_present_pte(ent)) {
1337                         if (!is_last_spte(ent, sp->role.level)) {
1338                                 ent &= PT64_BASE_ADDR_MASK;
1339                                 mmu_page_remove_parent_pte(page_header(ent),
1340                                                            &pt[i]);
1341                         } else {
1342                                 if (is_large_pte(ent))
1343                                         --kvm->stat.lpages;
1344                                 rmap_remove(kvm, &pt[i]);
1345                         }
1346                 }
1347                 pt[i] = shadow_trap_nonpresent_pte;
1348         }
1349 }
1350
1351 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1352 {
1353         mmu_page_remove_parent_pte(sp, parent_pte);
1354 }
1355
1356 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1357 {
1358         int i;
1359         struct kvm_vcpu *vcpu;
1360
1361         kvm_for_each_vcpu(i, vcpu, kvm)
1362                 vcpu->arch.last_pte_updated = NULL;
1363 }
1364
1365 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1366 {
1367         u64 *parent_pte;
1368
1369         while (sp->multimapped || sp->parent_pte) {
1370                 if (!sp->multimapped)
1371                         parent_pte = sp->parent_pte;
1372                 else {
1373                         struct kvm_pte_chain *chain;
1374
1375                         chain = container_of(sp->parent_ptes.first,
1376                                              struct kvm_pte_chain, link);
1377                         parent_pte = chain->parent_ptes[0];
1378                 }
1379                 BUG_ON(!parent_pte);
1380                 kvm_mmu_put_page(sp, parent_pte);
1381                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1382         }
1383 }
1384
1385 static int mmu_zap_unsync_children(struct kvm *kvm,
1386                                    struct kvm_mmu_page *parent)
1387 {
1388         int i, zapped = 0;
1389         struct mmu_page_path parents;
1390         struct kvm_mmu_pages pages;
1391
1392         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1393                 return 0;
1394
1395         kvm_mmu_pages_init(parent, &parents, &pages);
1396         while (mmu_unsync_walk(parent, &pages)) {
1397                 struct kvm_mmu_page *sp;
1398
1399                 for_each_sp(pages, sp, parents, i) {
1400                         kvm_mmu_zap_page(kvm, sp);
1401                         mmu_pages_clear_parents(&parents);
1402                 }
1403                 zapped += pages.nr;
1404                 kvm_mmu_pages_init(parent, &parents, &pages);
1405         }
1406
1407         return zapped;
1408 }
1409
1410 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1411 {
1412         int ret;
1413         ++kvm->stat.mmu_shadow_zapped;
1414         ret = mmu_zap_unsync_children(kvm, sp);
1415         kvm_mmu_page_unlink_children(kvm, sp);
1416         kvm_mmu_unlink_parents(kvm, sp);
1417         kvm_flush_remote_tlbs(kvm);
1418         if (!sp->role.invalid && !sp->role.direct)
1419                 unaccount_shadowed(kvm, sp->gfn);
1420         if (sp->unsync)
1421                 kvm_unlink_unsync_page(kvm, sp);
1422         if (!sp->root_count) {
1423                 hlist_del(&sp->hash_link);
1424                 kvm_mmu_free_page(kvm, sp);
1425         } else {
1426                 sp->role.invalid = 1;
1427                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1428                 kvm_reload_remote_mmus(kvm);
1429         }
1430         kvm_mmu_reset_last_pte_updated(kvm);
1431         return ret;
1432 }
1433
1434 /*
1435  * Changing the number of mmu pages allocated to the vm
1436  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1437  */
1438 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1439 {
1440         int used_pages;
1441
1442         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1443         used_pages = max(0, used_pages);
1444
1445         /*
1446          * If we set the number of mmu pages to be smaller be than the
1447          * number of actived pages , we must to free some mmu pages before we
1448          * change the value
1449          */
1450
1451         if (used_pages > kvm_nr_mmu_pages) {
1452                 while (used_pages > kvm_nr_mmu_pages) {
1453                         struct kvm_mmu_page *page;
1454
1455                         page = container_of(kvm->arch.active_mmu_pages.prev,
1456                                             struct kvm_mmu_page, link);
1457                         kvm_mmu_zap_page(kvm, page);
1458                         used_pages--;
1459                 }
1460                 kvm->arch.n_free_mmu_pages = 0;
1461         }
1462         else
1463                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1464                                          - kvm->arch.n_alloc_mmu_pages;
1465
1466         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1467 }
1468
1469 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1470 {
1471         unsigned index;
1472         struct hlist_head *bucket;
1473         struct kvm_mmu_page *sp;
1474         struct hlist_node *node, *n;
1475         int r;
1476
1477         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1478         r = 0;
1479         index = kvm_page_table_hashfn(gfn);
1480         bucket = &kvm->arch.mmu_page_hash[index];
1481         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1482                 if (sp->gfn == gfn && !sp->role.direct) {
1483                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1484                                  sp->role.word);
1485                         r = 1;
1486                         if (kvm_mmu_zap_page(kvm, sp))
1487                                 n = bucket->first;
1488                 }
1489         return r;
1490 }
1491
1492 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1493 {
1494         unsigned index;
1495         struct hlist_head *bucket;
1496         struct kvm_mmu_page *sp;
1497         struct hlist_node *node, *nn;
1498
1499         index = kvm_page_table_hashfn(gfn);
1500         bucket = &kvm->arch.mmu_page_hash[index];
1501         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1502                 if (sp->gfn == gfn && !sp->role.direct
1503                     && !sp->role.invalid) {
1504                         pgprintk("%s: zap %lx %x\n",
1505                                  __func__, gfn, sp->role.word);
1506                         kvm_mmu_zap_page(kvm, sp);
1507                 }
1508         }
1509 }
1510
1511 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1512 {
1513         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1514         struct kvm_mmu_page *sp = page_header(__pa(pte));
1515
1516         __set_bit(slot, sp->slot_bitmap);
1517 }
1518
1519 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1520 {
1521         int i;
1522         u64 *pt = sp->spt;
1523
1524         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1525                 return;
1526
1527         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1528                 if (pt[i] == shadow_notrap_nonpresent_pte)
1529                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1530         }
1531 }
1532
1533 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1534 {
1535         struct page *page;
1536
1537         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1538
1539         if (gpa == UNMAPPED_GVA)
1540                 return NULL;
1541
1542         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1543
1544         return page;
1545 }
1546
1547 /*
1548  * The function is based on mtrr_type_lookup() in
1549  * arch/x86/kernel/cpu/mtrr/generic.c
1550  */
1551 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1552                          u64 start, u64 end)
1553 {
1554         int i;
1555         u64 base, mask;
1556         u8 prev_match, curr_match;
1557         int num_var_ranges = KVM_NR_VAR_MTRR;
1558
1559         if (!mtrr_state->enabled)
1560                 return 0xFF;
1561
1562         /* Make end inclusive end, instead of exclusive */
1563         end--;
1564
1565         /* Look in fixed ranges. Just return the type as per start */
1566         if (mtrr_state->have_fixed && (start < 0x100000)) {
1567                 int idx;
1568
1569                 if (start < 0x80000) {
1570                         idx = 0;
1571                         idx += (start >> 16);
1572                         return mtrr_state->fixed_ranges[idx];
1573                 } else if (start < 0xC0000) {
1574                         idx = 1 * 8;
1575                         idx += ((start - 0x80000) >> 14);
1576                         return mtrr_state->fixed_ranges[idx];
1577                 } else if (start < 0x1000000) {
1578                         idx = 3 * 8;
1579                         idx += ((start - 0xC0000) >> 12);
1580                         return mtrr_state->fixed_ranges[idx];
1581                 }
1582         }
1583
1584         /*
1585          * Look in variable ranges
1586          * Look of multiple ranges matching this address and pick type
1587          * as per MTRR precedence
1588          */
1589         if (!(mtrr_state->enabled & 2))
1590                 return mtrr_state->def_type;
1591
1592         prev_match = 0xFF;
1593         for (i = 0; i < num_var_ranges; ++i) {
1594                 unsigned short start_state, end_state;
1595
1596                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1597                         continue;
1598
1599                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1600                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1601                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1602                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1603
1604                 start_state = ((start & mask) == (base & mask));
1605                 end_state = ((end & mask) == (base & mask));
1606                 if (start_state != end_state)
1607                         return 0xFE;
1608
1609                 if ((start & mask) != (base & mask))
1610                         continue;
1611
1612                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1613                 if (prev_match == 0xFF) {
1614                         prev_match = curr_match;
1615                         continue;
1616                 }
1617
1618                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1619                     curr_match == MTRR_TYPE_UNCACHABLE)
1620                         return MTRR_TYPE_UNCACHABLE;
1621
1622                 if ((prev_match == MTRR_TYPE_WRBACK &&
1623                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1624                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1625                      curr_match == MTRR_TYPE_WRBACK)) {
1626                         prev_match = MTRR_TYPE_WRTHROUGH;
1627                         curr_match = MTRR_TYPE_WRTHROUGH;
1628                 }
1629
1630                 if (prev_match != curr_match)
1631                         return MTRR_TYPE_UNCACHABLE;
1632         }
1633
1634         if (prev_match != 0xFF)
1635                 return prev_match;
1636
1637         return mtrr_state->def_type;
1638 }
1639
1640 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1641 {
1642         u8 mtrr;
1643
1644         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1645                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1646         if (mtrr == 0xfe || mtrr == 0xff)
1647                 mtrr = MTRR_TYPE_WRBACK;
1648         return mtrr;
1649 }
1650 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1651
1652 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1653 {
1654         unsigned index;
1655         struct hlist_head *bucket;
1656         struct kvm_mmu_page *s;
1657         struct hlist_node *node, *n;
1658
1659         index = kvm_page_table_hashfn(sp->gfn);
1660         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1661         /* don't unsync if pagetable is shadowed with multiple roles */
1662         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1663                 if (s->gfn != sp->gfn || s->role.direct)
1664                         continue;
1665                 if (s->role.word != sp->role.word)
1666                         return 1;
1667         }
1668         ++vcpu->kvm->stat.mmu_unsync;
1669         sp->unsync = 1;
1670
1671         kvm_mmu_mark_parents_unsync(vcpu, sp);
1672
1673         mmu_convert_notrap(sp);
1674         return 0;
1675 }
1676
1677 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1678                                   bool can_unsync)
1679 {
1680         struct kvm_mmu_page *shadow;
1681
1682         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1683         if (shadow) {
1684                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1685                         return 1;
1686                 if (shadow->unsync)
1687                         return 0;
1688                 if (can_unsync && oos_shadow)
1689                         return kvm_unsync_page(vcpu, shadow);
1690                 return 1;
1691         }
1692         return 0;
1693 }
1694
1695 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1696                     unsigned pte_access, int user_fault,
1697                     int write_fault, int dirty, int largepage,
1698                     gfn_t gfn, pfn_t pfn, bool speculative,
1699                     bool can_unsync)
1700 {
1701         u64 spte;
1702         int ret = 0;
1703
1704         /*
1705          * We don't set the accessed bit, since we sometimes want to see
1706          * whether the guest actually used the pte (in order to detect
1707          * demand paging).
1708          */
1709         spte = shadow_base_present_pte | shadow_dirty_mask;
1710         if (!speculative)
1711                 spte |= shadow_accessed_mask;
1712         if (!dirty)
1713                 pte_access &= ~ACC_WRITE_MASK;
1714         if (pte_access & ACC_EXEC_MASK)
1715                 spte |= shadow_x_mask;
1716         else
1717                 spte |= shadow_nx_mask;
1718         if (pte_access & ACC_USER_MASK)
1719                 spte |= shadow_user_mask;
1720         if (largepage)
1721                 spte |= PT_PAGE_SIZE_MASK;
1722         if (tdp_enabled)
1723                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1724                         kvm_is_mmio_pfn(pfn));
1725
1726         spte |= (u64)pfn << PAGE_SHIFT;
1727
1728         if ((pte_access & ACC_WRITE_MASK)
1729             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1730
1731                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1732                         ret = 1;
1733                         spte = shadow_trap_nonpresent_pte;
1734                         goto set_pte;
1735                 }
1736
1737                 spte |= PT_WRITABLE_MASK;
1738
1739                 /*
1740                  * Optimization: for pte sync, if spte was writable the hash
1741                  * lookup is unnecessary (and expensive). Write protection
1742                  * is responsibility of mmu_get_page / kvm_sync_page.
1743                  * Same reasoning can be applied to dirty page accounting.
1744                  */
1745                 if (!can_unsync && is_writeble_pte(*sptep))
1746                         goto set_pte;
1747
1748                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1749                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1750                                  __func__, gfn);
1751                         ret = 1;
1752                         pte_access &= ~ACC_WRITE_MASK;
1753                         if (is_writeble_pte(spte))
1754                                 spte &= ~PT_WRITABLE_MASK;
1755                 }
1756         }
1757
1758         if (pte_access & ACC_WRITE_MASK)
1759                 mark_page_dirty(vcpu->kvm, gfn);
1760
1761 set_pte:
1762         __set_spte(sptep, spte);
1763         return ret;
1764 }
1765
1766 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1767                          unsigned pt_access, unsigned pte_access,
1768                          int user_fault, int write_fault, int dirty,
1769                          int *ptwrite, int largepage, gfn_t gfn,
1770                          pfn_t pfn, bool speculative)
1771 {
1772         int was_rmapped = 0;
1773         int was_writeble = is_writeble_pte(*sptep);
1774         int rmap_count;
1775
1776         pgprintk("%s: spte %llx access %x write_fault %d"
1777                  " user_fault %d gfn %lx\n",
1778                  __func__, *sptep, pt_access,
1779                  write_fault, user_fault, gfn);
1780
1781         if (is_rmap_spte(*sptep)) {
1782                 /*
1783                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1784                  * the parent of the now unreachable PTE.
1785                  */
1786                 if (largepage && !is_large_pte(*sptep)) {
1787                         struct kvm_mmu_page *child;
1788                         u64 pte = *sptep;
1789
1790                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1791                         mmu_page_remove_parent_pte(child, sptep);
1792                 } else if (pfn != spte_to_pfn(*sptep)) {
1793                         pgprintk("hfn old %lx new %lx\n",
1794                                  spte_to_pfn(*sptep), pfn);
1795                         rmap_remove(vcpu->kvm, sptep);
1796                 } else
1797                         was_rmapped = 1;
1798         }
1799         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1800                       dirty, largepage, gfn, pfn, speculative, true)) {
1801                 if (write_fault)
1802                         *ptwrite = 1;
1803                 kvm_x86_ops->tlb_flush(vcpu);
1804         }
1805
1806         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1807         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1808                  is_large_pte(*sptep)? "2MB" : "4kB",
1809                  is_present_pte(*sptep)?"RW":"R", gfn,
1810                  *shadow_pte, sptep);
1811         if (!was_rmapped && is_large_pte(*sptep))
1812                 ++vcpu->kvm->stat.lpages;
1813
1814         page_header_update_slot(vcpu->kvm, sptep, gfn);
1815         if (!was_rmapped) {
1816                 rmap_count = rmap_add(vcpu, sptep, gfn, largepage);
1817                 if (!is_rmap_spte(*sptep))
1818                         kvm_release_pfn_clean(pfn);
1819                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1820                         rmap_recycle(vcpu, gfn, largepage);
1821         } else {
1822                 if (was_writeble)
1823                         kvm_release_pfn_dirty(pfn);
1824                 else
1825                         kvm_release_pfn_clean(pfn);
1826         }
1827         if (speculative) {
1828                 vcpu->arch.last_pte_updated = sptep;
1829                 vcpu->arch.last_pte_gfn = gfn;
1830         }
1831 }
1832
1833 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1834 {
1835 }
1836
1837 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1838                         int largepage, gfn_t gfn, pfn_t pfn)
1839 {
1840         struct kvm_shadow_walk_iterator iterator;
1841         struct kvm_mmu_page *sp;
1842         int pt_write = 0;
1843         gfn_t pseudo_gfn;
1844
1845         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1846                 if (iterator.level == PT_PAGE_TABLE_LEVEL
1847                     || (largepage && iterator.level == PT_DIRECTORY_LEVEL)) {
1848                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1849                                      0, write, 1, &pt_write,
1850                                      largepage, gfn, pfn, false);
1851                         ++vcpu->stat.pf_fixed;
1852                         break;
1853                 }
1854
1855                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1856                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1857                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1858                                               iterator.level - 1,
1859                                               1, ACC_ALL, iterator.sptep);
1860                         if (!sp) {
1861                                 pgprintk("nonpaging_map: ENOMEM\n");
1862                                 kvm_release_pfn_clean(pfn);
1863                                 return -ENOMEM;
1864                         }
1865
1866                         __set_spte(iterator.sptep,
1867                                    __pa(sp->spt)
1868                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1869                                    | shadow_user_mask | shadow_x_mask);
1870                 }
1871         }
1872         return pt_write;
1873 }
1874
1875 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1876 {
1877         int r;
1878         int largepage = 0;
1879         pfn_t pfn;
1880         unsigned long mmu_seq;
1881
1882         if (is_largepage_backed(vcpu, gfn &
1883                         ~(KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL) - 1))) {
1884                 gfn &= ~(KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL) - 1);
1885                 largepage = 1;
1886         }
1887
1888         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1889         smp_rmb();
1890         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1891
1892         /* mmio */
1893         if (is_error_pfn(pfn)) {
1894                 kvm_release_pfn_clean(pfn);
1895                 return 1;
1896         }
1897
1898         spin_lock(&vcpu->kvm->mmu_lock);
1899         if (mmu_notifier_retry(vcpu, mmu_seq))
1900                 goto out_unlock;
1901         kvm_mmu_free_some_pages(vcpu);
1902         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1903         spin_unlock(&vcpu->kvm->mmu_lock);
1904
1905
1906         return r;
1907
1908 out_unlock:
1909         spin_unlock(&vcpu->kvm->mmu_lock);
1910         kvm_release_pfn_clean(pfn);
1911         return 0;
1912 }
1913
1914
1915 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1916 {
1917         int i;
1918         struct kvm_mmu_page *sp;
1919
1920         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1921                 return;
1922         spin_lock(&vcpu->kvm->mmu_lock);
1923         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1924                 hpa_t root = vcpu->arch.mmu.root_hpa;
1925
1926                 sp = page_header(root);
1927                 --sp->root_count;
1928                 if (!sp->root_count && sp->role.invalid)
1929                         kvm_mmu_zap_page(vcpu->kvm, sp);
1930                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1931                 spin_unlock(&vcpu->kvm->mmu_lock);
1932                 return;
1933         }
1934         for (i = 0; i < 4; ++i) {
1935                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1936
1937                 if (root) {
1938                         root &= PT64_BASE_ADDR_MASK;
1939                         sp = page_header(root);
1940                         --sp->root_count;
1941                         if (!sp->root_count && sp->role.invalid)
1942                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1943                 }
1944                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1945         }
1946         spin_unlock(&vcpu->kvm->mmu_lock);
1947         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1948 }
1949
1950 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
1951 {
1952         int ret = 0;
1953
1954         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
1955                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
1956                 ret = 1;
1957         }
1958
1959         return ret;
1960 }
1961
1962 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
1963 {
1964         int i;
1965         gfn_t root_gfn;
1966         struct kvm_mmu_page *sp;
1967         int direct = 0;
1968         u64 pdptr;
1969
1970         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1971
1972         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1973                 hpa_t root = vcpu->arch.mmu.root_hpa;
1974
1975                 ASSERT(!VALID_PAGE(root));
1976                 if (tdp_enabled)
1977                         direct = 1;
1978                 if (mmu_check_root(vcpu, root_gfn))
1979                         return 1;
1980                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1981                                       PT64_ROOT_LEVEL, direct,
1982                                       ACC_ALL, NULL);
1983                 root = __pa(sp->spt);
1984                 ++sp->root_count;
1985                 vcpu->arch.mmu.root_hpa = root;
1986                 return 0;
1987         }
1988         direct = !is_paging(vcpu);
1989         if (tdp_enabled)
1990                 direct = 1;
1991         for (i = 0; i < 4; ++i) {
1992                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1993
1994                 ASSERT(!VALID_PAGE(root));
1995                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1996                         pdptr = kvm_pdptr_read(vcpu, i);
1997                         if (!is_present_gpte(pdptr)) {
1998                                 vcpu->arch.mmu.pae_root[i] = 0;
1999                                 continue;
2000                         }
2001                         root_gfn = pdptr >> PAGE_SHIFT;
2002                 } else if (vcpu->arch.mmu.root_level == 0)
2003                         root_gfn = 0;
2004                 if (mmu_check_root(vcpu, root_gfn))
2005                         return 1;
2006                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2007                                       PT32_ROOT_LEVEL, direct,
2008                                       ACC_ALL, NULL);
2009                 root = __pa(sp->spt);
2010                 ++sp->root_count;
2011                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2012         }
2013         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2014         return 0;
2015 }
2016
2017 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2018 {
2019         int i;
2020         struct kvm_mmu_page *sp;
2021
2022         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2023                 return;
2024         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2025                 hpa_t root = vcpu->arch.mmu.root_hpa;
2026                 sp = page_header(root);
2027                 mmu_sync_children(vcpu, sp);
2028                 return;
2029         }
2030         for (i = 0; i < 4; ++i) {
2031                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2032
2033                 if (root && VALID_PAGE(root)) {
2034                         root &= PT64_BASE_ADDR_MASK;
2035                         sp = page_header(root);
2036                         mmu_sync_children(vcpu, sp);
2037                 }
2038         }
2039 }
2040
2041 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2042 {
2043         spin_lock(&vcpu->kvm->mmu_lock);
2044         mmu_sync_roots(vcpu);
2045         spin_unlock(&vcpu->kvm->mmu_lock);
2046 }
2047
2048 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2049 {
2050         return vaddr;
2051 }
2052
2053 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2054                                 u32 error_code)
2055 {
2056         gfn_t gfn;
2057         int r;
2058
2059         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2060         r = mmu_topup_memory_caches(vcpu);
2061         if (r)
2062                 return r;
2063
2064         ASSERT(vcpu);
2065         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2066
2067         gfn = gva >> PAGE_SHIFT;
2068
2069         return nonpaging_map(vcpu, gva & PAGE_MASK,
2070                              error_code & PFERR_WRITE_MASK, gfn);
2071 }
2072
2073 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2074                                 u32 error_code)
2075 {
2076         pfn_t pfn;
2077         int r;
2078         int largepage = 0;
2079         gfn_t gfn = gpa >> PAGE_SHIFT;
2080         unsigned long mmu_seq;
2081
2082         ASSERT(vcpu);
2083         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2084
2085         r = mmu_topup_memory_caches(vcpu);
2086         if (r)
2087                 return r;
2088
2089         if (is_largepage_backed(vcpu, gfn &
2090                         ~(KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL) - 1))) {
2091                 gfn &= ~(KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL) - 1);
2092                 largepage = 1;
2093         }
2094         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2095         smp_rmb();
2096         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2097         if (is_error_pfn(pfn)) {
2098                 kvm_release_pfn_clean(pfn);
2099                 return 1;
2100         }
2101         spin_lock(&vcpu->kvm->mmu_lock);
2102         if (mmu_notifier_retry(vcpu, mmu_seq))
2103                 goto out_unlock;
2104         kvm_mmu_free_some_pages(vcpu);
2105         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2106                          largepage, gfn, pfn);
2107         spin_unlock(&vcpu->kvm->mmu_lock);
2108
2109         return r;
2110
2111 out_unlock:
2112         spin_unlock(&vcpu->kvm->mmu_lock);
2113         kvm_release_pfn_clean(pfn);
2114         return 0;
2115 }
2116
2117 static void nonpaging_free(struct kvm_vcpu *vcpu)
2118 {
2119         mmu_free_roots(vcpu);
2120 }
2121
2122 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2123 {
2124         struct kvm_mmu *context = &vcpu->arch.mmu;
2125
2126         context->new_cr3 = nonpaging_new_cr3;
2127         context->page_fault = nonpaging_page_fault;
2128         context->gva_to_gpa = nonpaging_gva_to_gpa;
2129         context->free = nonpaging_free;
2130         context->prefetch_page = nonpaging_prefetch_page;
2131         context->sync_page = nonpaging_sync_page;
2132         context->invlpg = nonpaging_invlpg;
2133         context->root_level = 0;
2134         context->shadow_root_level = PT32E_ROOT_LEVEL;
2135         context->root_hpa = INVALID_PAGE;
2136         return 0;
2137 }
2138
2139 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2140 {
2141         ++vcpu->stat.tlb_flush;
2142         kvm_x86_ops->tlb_flush(vcpu);
2143 }
2144
2145 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2146 {
2147         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2148         mmu_free_roots(vcpu);
2149 }
2150
2151 static void inject_page_fault(struct kvm_vcpu *vcpu,
2152                               u64 addr,
2153                               u32 err_code)
2154 {
2155         kvm_inject_page_fault(vcpu, addr, err_code);
2156 }
2157
2158 static void paging_free(struct kvm_vcpu *vcpu)
2159 {
2160         nonpaging_free(vcpu);
2161 }
2162
2163 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2164 {
2165         int bit7;
2166
2167         bit7 = (gpte >> 7) & 1;
2168         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2169 }
2170
2171 #define PTTYPE 64
2172 #include "paging_tmpl.h"
2173 #undef PTTYPE
2174
2175 #define PTTYPE 32
2176 #include "paging_tmpl.h"
2177 #undef PTTYPE
2178
2179 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2180 {
2181         struct kvm_mmu *context = &vcpu->arch.mmu;
2182         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2183         u64 exb_bit_rsvd = 0;
2184
2185         if (!is_nx(vcpu))
2186                 exb_bit_rsvd = rsvd_bits(63, 63);
2187         switch (level) {
2188         case PT32_ROOT_LEVEL:
2189                 /* no rsvd bits for 2 level 4K page table entries */
2190                 context->rsvd_bits_mask[0][1] = 0;
2191                 context->rsvd_bits_mask[0][0] = 0;
2192                 if (is_cpuid_PSE36())
2193                         /* 36bits PSE 4MB page */
2194                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2195                 else
2196                         /* 32 bits PSE 4MB page */
2197                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2198                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2199                 break;
2200         case PT32E_ROOT_LEVEL:
2201                 context->rsvd_bits_mask[0][2] =
2202                         rsvd_bits(maxphyaddr, 63) |
2203                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2204                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2205                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2206                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2207                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2208                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2209                         rsvd_bits(maxphyaddr, 62) |
2210                         rsvd_bits(13, 20);              /* large page */
2211                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2212                 break;
2213         case PT64_ROOT_LEVEL:
2214                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2215                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2216                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2217                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2218                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2219                         rsvd_bits(maxphyaddr, 51);
2220                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2221                         rsvd_bits(maxphyaddr, 51);
2222                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2223                 context->rsvd_bits_mask[1][2] = context->rsvd_bits_mask[0][2];
2224                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2225                         rsvd_bits(maxphyaddr, 51) |
2226                         rsvd_bits(13, 20);              /* large page */
2227                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2228                 break;
2229         }
2230 }
2231
2232 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2233 {
2234         struct kvm_mmu *context = &vcpu->arch.mmu;
2235
2236         ASSERT(is_pae(vcpu));
2237         context->new_cr3 = paging_new_cr3;
2238         context->page_fault = paging64_page_fault;
2239         context->gva_to_gpa = paging64_gva_to_gpa;
2240         context->prefetch_page = paging64_prefetch_page;
2241         context->sync_page = paging64_sync_page;
2242         context->invlpg = paging64_invlpg;
2243         context->free = paging_free;
2244         context->root_level = level;
2245         context->shadow_root_level = level;
2246         context->root_hpa = INVALID_PAGE;
2247         return 0;
2248 }
2249
2250 static int paging64_init_context(struct kvm_vcpu *vcpu)
2251 {
2252         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2253         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2254 }
2255
2256 static int paging32_init_context(struct kvm_vcpu *vcpu)
2257 {
2258         struct kvm_mmu *context = &vcpu->arch.mmu;
2259
2260         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2261         context->new_cr3 = paging_new_cr3;
2262         context->page_fault = paging32_page_fault;
2263         context->gva_to_gpa = paging32_gva_to_gpa;
2264         context->free = paging_free;
2265         context->prefetch_page = paging32_prefetch_page;
2266         context->sync_page = paging32_sync_page;
2267         context->invlpg = paging32_invlpg;
2268         context->root_level = PT32_ROOT_LEVEL;
2269         context->shadow_root_level = PT32E_ROOT_LEVEL;
2270         context->root_hpa = INVALID_PAGE;
2271         return 0;
2272 }
2273
2274 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2275 {
2276         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2277         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2278 }
2279
2280 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2281 {
2282         struct kvm_mmu *context = &vcpu->arch.mmu;
2283
2284         context->new_cr3 = nonpaging_new_cr3;
2285         context->page_fault = tdp_page_fault;
2286         context->free = nonpaging_free;
2287         context->prefetch_page = nonpaging_prefetch_page;
2288         context->sync_page = nonpaging_sync_page;
2289         context->invlpg = nonpaging_invlpg;
2290         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2291         context->root_hpa = INVALID_PAGE;
2292
2293         if (!is_paging(vcpu)) {
2294                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2295                 context->root_level = 0;
2296         } else if (is_long_mode(vcpu)) {
2297                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2298                 context->gva_to_gpa = paging64_gva_to_gpa;
2299                 context->root_level = PT64_ROOT_LEVEL;
2300         } else if (is_pae(vcpu)) {
2301                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2302                 context->gva_to_gpa = paging64_gva_to_gpa;
2303                 context->root_level = PT32E_ROOT_LEVEL;
2304         } else {
2305                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2306                 context->gva_to_gpa = paging32_gva_to_gpa;
2307                 context->root_level = PT32_ROOT_LEVEL;
2308         }
2309
2310         return 0;
2311 }
2312
2313 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2314 {
2315         int r;
2316
2317         ASSERT(vcpu);
2318         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2319
2320         if (!is_paging(vcpu))
2321                 r = nonpaging_init_context(vcpu);
2322         else if (is_long_mode(vcpu))
2323                 r = paging64_init_context(vcpu);
2324         else if (is_pae(vcpu))
2325                 r = paging32E_init_context(vcpu);
2326         else
2327                 r = paging32_init_context(vcpu);
2328
2329         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2330
2331         return r;
2332 }
2333
2334 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2335 {
2336         vcpu->arch.update_pte.pfn = bad_pfn;
2337
2338         if (tdp_enabled)
2339                 return init_kvm_tdp_mmu(vcpu);
2340         else
2341                 return init_kvm_softmmu(vcpu);
2342 }
2343
2344 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2345 {
2346         ASSERT(vcpu);
2347         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2348                 vcpu->arch.mmu.free(vcpu);
2349                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2350         }
2351 }
2352
2353 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2354 {
2355         destroy_kvm_mmu(vcpu);
2356         return init_kvm_mmu(vcpu);
2357 }
2358 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2359
2360 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2361 {
2362         int r;
2363
2364         r = mmu_topup_memory_caches(vcpu);
2365         if (r)
2366                 goto out;
2367         spin_lock(&vcpu->kvm->mmu_lock);
2368         kvm_mmu_free_some_pages(vcpu);
2369         r = mmu_alloc_roots(vcpu);
2370         mmu_sync_roots(vcpu);
2371         spin_unlock(&vcpu->kvm->mmu_lock);
2372         if (r)
2373                 goto out;
2374         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2375         kvm_mmu_flush_tlb(vcpu);
2376 out:
2377         return r;
2378 }
2379 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2380
2381 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2382 {
2383         mmu_free_roots(vcpu);
2384 }
2385
2386 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2387                                   struct kvm_mmu_page *sp,
2388                                   u64 *spte)
2389 {
2390         u64 pte;
2391         struct kvm_mmu_page *child;
2392
2393         pte = *spte;
2394         if (is_shadow_present_pte(pte)) {
2395                 if (is_last_spte(pte, sp->role.level))
2396                         rmap_remove(vcpu->kvm, spte);
2397                 else {
2398                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2399                         mmu_page_remove_parent_pte(child, spte);
2400                 }
2401         }
2402         __set_spte(spte, shadow_trap_nonpresent_pte);
2403         if (is_large_pte(pte))
2404                 --vcpu->kvm->stat.lpages;
2405 }
2406
2407 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2408                                   struct kvm_mmu_page *sp,
2409                                   u64 *spte,
2410                                   const void *new)
2411 {
2412         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2413                 if (!vcpu->arch.update_pte.largepage ||
2414                     sp->role.glevels == PT32_ROOT_LEVEL) {
2415                         ++vcpu->kvm->stat.mmu_pde_zapped;
2416                         return;
2417                 }
2418         }
2419
2420         ++vcpu->kvm->stat.mmu_pte_updated;
2421         if (sp->role.glevels == PT32_ROOT_LEVEL)
2422                 paging32_update_pte(vcpu, sp, spte, new);
2423         else
2424                 paging64_update_pte(vcpu, sp, spte, new);
2425 }
2426
2427 static bool need_remote_flush(u64 old, u64 new)
2428 {
2429         if (!is_shadow_present_pte(old))
2430                 return false;
2431         if (!is_shadow_present_pte(new))
2432                 return true;
2433         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2434                 return true;
2435         old ^= PT64_NX_MASK;
2436         new ^= PT64_NX_MASK;
2437         return (old & ~new & PT64_PERM_MASK) != 0;
2438 }
2439
2440 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2441 {
2442         if (need_remote_flush(old, new))
2443                 kvm_flush_remote_tlbs(vcpu->kvm);
2444         else
2445                 kvm_mmu_flush_tlb(vcpu);
2446 }
2447
2448 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2449 {
2450         u64 *spte = vcpu->arch.last_pte_updated;
2451
2452         return !!(spte && (*spte & shadow_accessed_mask));
2453 }
2454
2455 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2456                                           const u8 *new, int bytes)
2457 {
2458         gfn_t gfn;
2459         int r;
2460         u64 gpte = 0;
2461         pfn_t pfn;
2462
2463         vcpu->arch.update_pte.largepage = 0;
2464
2465         if (bytes != 4 && bytes != 8)
2466                 return;
2467
2468         /*
2469          * Assume that the pte write on a page table of the same type
2470          * as the current vcpu paging mode.  This is nearly always true
2471          * (might be false while changing modes).  Note it is verified later
2472          * by update_pte().
2473          */
2474         if (is_pae(vcpu)) {
2475                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2476                 if ((bytes == 4) && (gpa % 4 == 0)) {
2477                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2478                         if (r)
2479                                 return;
2480                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2481                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2482                         memcpy((void *)&gpte, new, 8);
2483                 }
2484         } else {
2485                 if ((bytes == 4) && (gpa % 4 == 0))
2486                         memcpy((void *)&gpte, new, 4);
2487         }
2488         if (!is_present_gpte(gpte))
2489                 return;
2490         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2491
2492         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2493                 gfn &= ~(KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL) - 1);
2494                 vcpu->arch.update_pte.largepage = 1;
2495         }
2496         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2497         smp_rmb();
2498         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2499
2500         if (is_error_pfn(pfn)) {
2501                 kvm_release_pfn_clean(pfn);
2502                 return;
2503         }
2504         vcpu->arch.update_pte.gfn = gfn;
2505         vcpu->arch.update_pte.pfn = pfn;
2506 }
2507
2508 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2509 {
2510         u64 *spte = vcpu->arch.last_pte_updated;
2511
2512         if (spte
2513             && vcpu->arch.last_pte_gfn == gfn
2514             && shadow_accessed_mask
2515             && !(*spte & shadow_accessed_mask)
2516             && is_shadow_present_pte(*spte))
2517                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2518 }
2519
2520 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2521                        const u8 *new, int bytes,
2522                        bool guest_initiated)
2523 {
2524         gfn_t gfn = gpa >> PAGE_SHIFT;
2525         struct kvm_mmu_page *sp;
2526         struct hlist_node *node, *n;
2527         struct hlist_head *bucket;
2528         unsigned index;
2529         u64 entry, gentry;
2530         u64 *spte;
2531         unsigned offset = offset_in_page(gpa);
2532         unsigned pte_size;
2533         unsigned page_offset;
2534         unsigned misaligned;
2535         unsigned quadrant;
2536         int level;
2537         int flooded = 0;
2538         int npte;
2539         int r;
2540
2541         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2542         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2543         spin_lock(&vcpu->kvm->mmu_lock);
2544         kvm_mmu_access_page(vcpu, gfn);
2545         kvm_mmu_free_some_pages(vcpu);
2546         ++vcpu->kvm->stat.mmu_pte_write;
2547         kvm_mmu_audit(vcpu, "pre pte write");
2548         if (guest_initiated) {
2549                 if (gfn == vcpu->arch.last_pt_write_gfn
2550                     && !last_updated_pte_accessed(vcpu)) {
2551                         ++vcpu->arch.last_pt_write_count;
2552                         if (vcpu->arch.last_pt_write_count >= 3)
2553                                 flooded = 1;
2554                 } else {
2555                         vcpu->arch.last_pt_write_gfn = gfn;
2556                         vcpu->arch.last_pt_write_count = 1;
2557                         vcpu->arch.last_pte_updated = NULL;
2558                 }
2559         }
2560         index = kvm_page_table_hashfn(gfn);
2561         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2562         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2563                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2564                         continue;
2565                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2566                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2567                 misaligned |= bytes < 4;
2568                 if (misaligned || flooded) {
2569                         /*
2570                          * Misaligned accesses are too much trouble to fix
2571                          * up; also, they usually indicate a page is not used
2572                          * as a page table.
2573                          *
2574                          * If we're seeing too many writes to a page,
2575                          * it may no longer be a page table, or we may be
2576                          * forking, in which case it is better to unmap the
2577                          * page.
2578                          */
2579                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2580                                  gpa, bytes, sp->role.word);
2581                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2582                                 n = bucket->first;
2583                         ++vcpu->kvm->stat.mmu_flooded;
2584                         continue;
2585                 }
2586                 page_offset = offset;
2587                 level = sp->role.level;
2588                 npte = 1;
2589                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2590                         page_offset <<= 1;      /* 32->64 */
2591                         /*
2592                          * A 32-bit pde maps 4MB while the shadow pdes map
2593                          * only 2MB.  So we need to double the offset again
2594                          * and zap two pdes instead of one.
2595                          */
2596                         if (level == PT32_ROOT_LEVEL) {
2597                                 page_offset &= ~7; /* kill rounding error */
2598                                 page_offset <<= 1;
2599                                 npte = 2;
2600                         }
2601                         quadrant = page_offset >> PAGE_SHIFT;
2602                         page_offset &= ~PAGE_MASK;
2603                         if (quadrant != sp->role.quadrant)
2604                                 continue;
2605                 }
2606                 spte = &sp->spt[page_offset / sizeof(*spte)];
2607                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2608                         gentry = 0;
2609                         r = kvm_read_guest_atomic(vcpu->kvm,
2610                                                   gpa & ~(u64)(pte_size - 1),
2611                                                   &gentry, pte_size);
2612                         new = (const void *)&gentry;
2613                         if (r < 0)
2614                                 new = NULL;
2615                 }
2616                 while (npte--) {
2617                         entry = *spte;
2618                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2619                         if (new)
2620                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2621                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2622                         ++spte;
2623                 }
2624         }
2625         kvm_mmu_audit(vcpu, "post pte write");
2626         spin_unlock(&vcpu->kvm->mmu_lock);
2627         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2628                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2629                 vcpu->arch.update_pte.pfn = bad_pfn;
2630         }
2631 }
2632
2633 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2634 {
2635         gpa_t gpa;
2636         int r;
2637
2638         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2639
2640         spin_lock(&vcpu->kvm->mmu_lock);
2641         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2642         spin_unlock(&vcpu->kvm->mmu_lock);
2643         return r;
2644 }
2645 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2646
2647 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2648 {
2649         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2650                 struct kvm_mmu_page *sp;
2651
2652                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2653                                   struct kvm_mmu_page, link);
2654                 kvm_mmu_zap_page(vcpu->kvm, sp);
2655                 ++vcpu->kvm->stat.mmu_recycled;
2656         }
2657 }
2658
2659 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2660 {
2661         int r;
2662         enum emulation_result er;
2663
2664         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2665         if (r < 0)
2666                 goto out;
2667
2668         if (!r) {
2669                 r = 1;
2670                 goto out;
2671         }
2672
2673         r = mmu_topup_memory_caches(vcpu);
2674         if (r)
2675                 goto out;
2676
2677         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2678
2679         switch (er) {
2680         case EMULATE_DONE:
2681                 return 1;
2682         case EMULATE_DO_MMIO:
2683                 ++vcpu->stat.mmio_exits;
2684                 return 0;
2685         case EMULATE_FAIL:
2686                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2687                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2688                 return 0;
2689         default:
2690                 BUG();
2691         }
2692 out:
2693         return r;
2694 }
2695 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2696
2697 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2698 {
2699         vcpu->arch.mmu.invlpg(vcpu, gva);
2700         kvm_mmu_flush_tlb(vcpu);
2701         ++vcpu->stat.invlpg;
2702 }
2703 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2704
2705 void kvm_enable_tdp(void)
2706 {
2707         tdp_enabled = true;
2708 }
2709 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2710
2711 void kvm_disable_tdp(void)
2712 {
2713         tdp_enabled = false;
2714 }
2715 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2716
2717 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2718 {
2719         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2720 }
2721
2722 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2723 {
2724         struct page *page;
2725         int i;
2726
2727         ASSERT(vcpu);
2728
2729         if (vcpu->kvm->arch.n_requested_mmu_pages)
2730                 vcpu->kvm->arch.n_free_mmu_pages =
2731                                         vcpu->kvm->arch.n_requested_mmu_pages;
2732         else
2733                 vcpu->kvm->arch.n_free_mmu_pages =
2734                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2735         /*
2736          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2737          * Therefore we need to allocate shadow page tables in the first
2738          * 4GB of memory, which happens to fit the DMA32 zone.
2739          */
2740         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2741         if (!page)
2742                 goto error_1;
2743         vcpu->arch.mmu.pae_root = page_address(page);
2744         for (i = 0; i < 4; ++i)
2745                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2746
2747         return 0;
2748
2749 error_1:
2750         free_mmu_pages(vcpu);
2751         return -ENOMEM;
2752 }
2753
2754 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2755 {
2756         ASSERT(vcpu);
2757         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2758
2759         return alloc_mmu_pages(vcpu);
2760 }
2761
2762 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2763 {
2764         ASSERT(vcpu);
2765         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2766
2767         return init_kvm_mmu(vcpu);
2768 }
2769
2770 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2771 {
2772         ASSERT(vcpu);
2773
2774         destroy_kvm_mmu(vcpu);
2775         free_mmu_pages(vcpu);
2776         mmu_free_memory_caches(vcpu);
2777 }
2778
2779 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2780 {
2781         struct kvm_mmu_page *sp;
2782
2783         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2784                 int i;
2785                 u64 *pt;
2786
2787                 if (!test_bit(slot, sp->slot_bitmap))
2788                         continue;
2789
2790                 pt = sp->spt;
2791                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2792                         /* avoid RMW */
2793                         if (pt[i] & PT_WRITABLE_MASK)
2794                                 pt[i] &= ~PT_WRITABLE_MASK;
2795         }
2796         kvm_flush_remote_tlbs(kvm);
2797 }
2798
2799 void kvm_mmu_zap_all(struct kvm *kvm)
2800 {
2801         struct kvm_mmu_page *sp, *node;
2802
2803         spin_lock(&kvm->mmu_lock);
2804         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2805                 if (kvm_mmu_zap_page(kvm, sp))
2806                         node = container_of(kvm->arch.active_mmu_pages.next,
2807                                             struct kvm_mmu_page, link);
2808         spin_unlock(&kvm->mmu_lock);
2809
2810         kvm_flush_remote_tlbs(kvm);
2811 }
2812
2813 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2814 {
2815         struct kvm_mmu_page *page;
2816
2817         page = container_of(kvm->arch.active_mmu_pages.prev,
2818                             struct kvm_mmu_page, link);
2819         kvm_mmu_zap_page(kvm, page);
2820 }
2821
2822 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2823 {
2824         struct kvm *kvm;
2825         struct kvm *kvm_freed = NULL;
2826         int cache_count = 0;
2827
2828         spin_lock(&kvm_lock);
2829
2830         list_for_each_entry(kvm, &vm_list, vm_list) {
2831                 int npages;
2832
2833                 if (!down_read_trylock(&kvm->slots_lock))
2834                         continue;
2835                 spin_lock(&kvm->mmu_lock);
2836                 npages = kvm->arch.n_alloc_mmu_pages -
2837                          kvm->arch.n_free_mmu_pages;
2838                 cache_count += npages;
2839                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2840                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2841                         cache_count--;
2842                         kvm_freed = kvm;
2843                 }
2844                 nr_to_scan--;
2845
2846                 spin_unlock(&kvm->mmu_lock);
2847                 up_read(&kvm->slots_lock);
2848         }
2849         if (kvm_freed)
2850                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2851
2852         spin_unlock(&kvm_lock);
2853
2854         return cache_count;
2855 }
2856
2857 static struct shrinker mmu_shrinker = {
2858         .shrink = mmu_shrink,
2859         .seeks = DEFAULT_SEEKS * 10,
2860 };
2861
2862 static void mmu_destroy_caches(void)
2863 {
2864         if (pte_chain_cache)
2865                 kmem_cache_destroy(pte_chain_cache);
2866         if (rmap_desc_cache)
2867                 kmem_cache_destroy(rmap_desc_cache);
2868         if (mmu_page_header_cache)
2869                 kmem_cache_destroy(mmu_page_header_cache);
2870 }
2871
2872 void kvm_mmu_module_exit(void)
2873 {
2874         mmu_destroy_caches();
2875         unregister_shrinker(&mmu_shrinker);
2876 }
2877
2878 int kvm_mmu_module_init(void)
2879 {
2880         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2881                                             sizeof(struct kvm_pte_chain),
2882                                             0, 0, NULL);
2883         if (!pte_chain_cache)
2884                 goto nomem;
2885         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2886                                             sizeof(struct kvm_rmap_desc),
2887                                             0, 0, NULL);
2888         if (!rmap_desc_cache)
2889                 goto nomem;
2890
2891         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2892                                                   sizeof(struct kvm_mmu_page),
2893                                                   0, 0, NULL);
2894         if (!mmu_page_header_cache)
2895                 goto nomem;
2896
2897         register_shrinker(&mmu_shrinker);
2898
2899         return 0;
2900
2901 nomem:
2902         mmu_destroy_caches();
2903         return -ENOMEM;
2904 }
2905
2906 /*
2907  * Caculate mmu pages needed for kvm.
2908  */
2909 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2910 {
2911         int i;
2912         unsigned int nr_mmu_pages;
2913         unsigned int  nr_pages = 0;
2914
2915         for (i = 0; i < kvm->nmemslots; i++)
2916                 nr_pages += kvm->memslots[i].npages;
2917
2918         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2919         nr_mmu_pages = max(nr_mmu_pages,
2920                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2921
2922         return nr_mmu_pages;
2923 }
2924
2925 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2926                                 unsigned len)
2927 {
2928         if (len > buffer->len)
2929                 return NULL;
2930         return buffer->ptr;
2931 }
2932
2933 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2934                                 unsigned len)
2935 {
2936         void *ret;
2937
2938         ret = pv_mmu_peek_buffer(buffer, len);
2939         if (!ret)
2940                 return ret;
2941         buffer->ptr += len;
2942         buffer->len -= len;
2943         buffer->processed += len;
2944         return ret;
2945 }
2946
2947 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2948                              gpa_t addr, gpa_t value)
2949 {
2950         int bytes = 8;
2951         int r;
2952
2953         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2954                 bytes = 4;
2955
2956         r = mmu_topup_memory_caches(vcpu);
2957         if (r)
2958                 return r;
2959
2960         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2961                 return -EFAULT;
2962
2963         return 1;
2964 }
2965
2966 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2967 {
2968         kvm_set_cr3(vcpu, vcpu->arch.cr3);
2969         return 1;
2970 }
2971
2972 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2973 {
2974         spin_lock(&vcpu->kvm->mmu_lock);
2975         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2976         spin_unlock(&vcpu->kvm->mmu_lock);
2977         return 1;
2978 }
2979
2980 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2981                              struct kvm_pv_mmu_op_buffer *buffer)
2982 {
2983         struct kvm_mmu_op_header *header;
2984
2985         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2986         if (!header)
2987                 return 0;
2988         switch (header->op) {
2989         case KVM_MMU_OP_WRITE_PTE: {
2990                 struct kvm_mmu_op_write_pte *wpte;
2991
2992                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2993                 if (!wpte)
2994                         return 0;
2995                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2996                                         wpte->pte_val);
2997         }
2998         case KVM_MMU_OP_FLUSH_TLB: {
2999                 struct kvm_mmu_op_flush_tlb *ftlb;
3000
3001                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3002                 if (!ftlb)
3003                         return 0;
3004                 return kvm_pv_mmu_flush_tlb(vcpu);
3005         }
3006         case KVM_MMU_OP_RELEASE_PT: {
3007                 struct kvm_mmu_op_release_pt *rpt;
3008
3009                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3010                 if (!rpt)
3011                         return 0;
3012                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3013         }
3014         default: return 0;
3015         }
3016 }
3017
3018 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3019                   gpa_t addr, unsigned long *ret)
3020 {
3021         int r;
3022         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3023
3024         buffer->ptr = buffer->buf;
3025         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3026         buffer->processed = 0;
3027
3028         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3029         if (r)
3030                 goto out;
3031
3032         while (buffer->len) {
3033                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3034                 if (r < 0)
3035                         goto out;
3036                 if (r == 0)
3037                         break;
3038         }
3039
3040         r = 1;
3041 out:
3042         *ret = buffer->processed;
3043         return r;
3044 }
3045
3046 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3047 {
3048         struct kvm_shadow_walk_iterator iterator;
3049         int nr_sptes = 0;
3050
3051         spin_lock(&vcpu->kvm->mmu_lock);
3052         for_each_shadow_entry(vcpu, addr, iterator) {
3053                 sptes[iterator.level-1] = *iterator.sptep;
3054                 nr_sptes++;
3055                 if (!is_shadow_present_pte(*iterator.sptep))
3056                         break;
3057         }
3058         spin_unlock(&vcpu->kvm->mmu_lock);
3059
3060         return nr_sptes;
3061 }
3062 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3063
3064 #ifdef AUDIT
3065
3066 static const char *audit_msg;
3067
3068 static gva_t canonicalize(gva_t gva)
3069 {
3070 #ifdef CONFIG_X86_64
3071         gva = (long long)(gva << 16) >> 16;
3072 #endif
3073         return gva;
3074 }
3075
3076
3077 typedef void (*inspect_spte_fn) (struct kvm *kvm, struct kvm_mmu_page *sp,
3078                                  u64 *sptep);
3079
3080 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3081                             inspect_spte_fn fn)
3082 {
3083         int i;
3084
3085         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3086                 u64 ent = sp->spt[i];
3087
3088                 if (is_shadow_present_pte(ent)) {
3089                         if (!is_last_spte(ent, sp->role.level)) {
3090                                 struct kvm_mmu_page *child;
3091                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3092                                 __mmu_spte_walk(kvm, child, fn);
3093                         } else
3094                                 fn(kvm, sp, &sp->spt[i]);
3095                 }
3096         }
3097 }
3098
3099 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3100 {
3101         int i;
3102         struct kvm_mmu_page *sp;
3103
3104         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3105                 return;
3106         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3107                 hpa_t root = vcpu->arch.mmu.root_hpa;
3108                 sp = page_header(root);
3109                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3110                 return;
3111         }
3112         for (i = 0; i < 4; ++i) {
3113                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3114
3115                 if (root && VALID_PAGE(root)) {
3116                         root &= PT64_BASE_ADDR_MASK;
3117                         sp = page_header(root);
3118                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3119                 }
3120         }
3121         return;
3122 }
3123
3124 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3125                                 gva_t va, int level)
3126 {
3127         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3128         int i;
3129         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3130
3131         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3132                 u64 ent = pt[i];
3133
3134                 if (ent == shadow_trap_nonpresent_pte)
3135                         continue;
3136
3137                 va = canonicalize(va);
3138                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3139                         audit_mappings_page(vcpu, ent, va, level - 1);
3140                 else {
3141                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3142                         gfn_t gfn = gpa >> PAGE_SHIFT;
3143                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3144                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3145
3146                         if (is_error_pfn(pfn)) {
3147                                 kvm_release_pfn_clean(pfn);
3148                                 continue;
3149                         }
3150
3151                         if (is_shadow_present_pte(ent)
3152                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3153                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3154                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3155                                        audit_msg, vcpu->arch.mmu.root_level,
3156                                        va, gpa, hpa, ent,
3157                                        is_shadow_present_pte(ent));
3158                         else if (ent == shadow_notrap_nonpresent_pte
3159                                  && !is_error_hpa(hpa))
3160                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3161                                        " valid guest gva %lx\n", audit_msg, va);
3162                         kvm_release_pfn_clean(pfn);
3163
3164                 }
3165         }
3166 }
3167
3168 static void audit_mappings(struct kvm_vcpu *vcpu)
3169 {
3170         unsigned i;
3171
3172         if (vcpu->arch.mmu.root_level == 4)
3173                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3174         else
3175                 for (i = 0; i < 4; ++i)
3176                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3177                                 audit_mappings_page(vcpu,
3178                                                     vcpu->arch.mmu.pae_root[i],
3179                                                     i << 30,
3180                                                     2);
3181 }
3182
3183 static int count_rmaps(struct kvm_vcpu *vcpu)
3184 {
3185         int nmaps = 0;
3186         int i, j, k;
3187
3188         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3189                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3190                 struct kvm_rmap_desc *d;
3191
3192                 for (j = 0; j < m->npages; ++j) {
3193                         unsigned long *rmapp = &m->rmap[j];
3194
3195                         if (!*rmapp)
3196                                 continue;
3197                         if (!(*rmapp & 1)) {
3198                                 ++nmaps;
3199                                 continue;
3200                         }
3201                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3202                         while (d) {
3203                                 for (k = 0; k < RMAP_EXT; ++k)
3204                                         if (d->sptes[k])
3205                                                 ++nmaps;
3206                                         else
3207                                                 break;
3208                                 d = d->more;
3209                         }
3210                 }
3211         }
3212         return nmaps;
3213 }
3214
3215 void inspect_spte_has_rmap(struct kvm *kvm, struct kvm_mmu_page *sp, u64 *sptep)
3216 {
3217         unsigned long *rmapp;
3218         struct kvm_mmu_page *rev_sp;
3219         gfn_t gfn;
3220
3221         if (*sptep & PT_WRITABLE_MASK) {
3222                 rev_sp = page_header(__pa(sptep));
3223                 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3224
3225                 if (!gfn_to_memslot(kvm, gfn)) {
3226                         if (!printk_ratelimit())
3227                                 return;
3228                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3229                                          audit_msg, gfn);
3230                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3231                                         audit_msg, sptep - rev_sp->spt,
3232                                         rev_sp->gfn);
3233                         dump_stack();
3234                         return;
3235                 }
3236
3237                 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3238                                     is_large_pte(*sptep));
3239                 if (!*rmapp) {
3240                         if (!printk_ratelimit())
3241                                 return;
3242                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3243                                          audit_msg, *sptep);
3244                         dump_stack();
3245                 }
3246         }
3247
3248 }
3249
3250 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3251 {
3252         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3253 }
3254
3255 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3256 {
3257         struct kvm_mmu_page *sp;
3258         int i;
3259
3260         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3261                 u64 *pt = sp->spt;
3262
3263                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3264                         continue;
3265
3266                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3267                         u64 ent = pt[i];
3268
3269                         if (!(ent & PT_PRESENT_MASK))
3270                                 continue;
3271                         if (!(ent & PT_WRITABLE_MASK))
3272                                 continue;
3273                         inspect_spte_has_rmap(vcpu->kvm, sp, &pt[i]);
3274                 }
3275         }
3276         return;
3277 }
3278
3279 static void audit_rmap(struct kvm_vcpu *vcpu)
3280 {
3281         check_writable_mappings_rmap(vcpu);
3282         count_rmaps(vcpu);
3283 }
3284
3285 static void audit_write_protection(struct kvm_vcpu *vcpu)
3286 {
3287         struct kvm_mmu_page *sp;
3288         struct kvm_memory_slot *slot;
3289         unsigned long *rmapp;
3290         u64 *spte;
3291         gfn_t gfn;
3292
3293         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3294                 if (sp->role.direct)
3295                         continue;
3296                 if (sp->unsync)
3297                         continue;
3298
3299                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3300                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3301                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3302
3303                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3304                 while (spte) {
3305                         if (*spte & PT_WRITABLE_MASK)
3306                                 printk(KERN_ERR "%s: (%s) shadow page has "
3307                                 "writable mappings: gfn %lx role %x\n",
3308                                __func__, audit_msg, sp->gfn,
3309                                sp->role.word);
3310                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3311                 }
3312         }
3313 }
3314
3315 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3316 {
3317         int olddbg = dbg;
3318
3319         dbg = 0;
3320         audit_msg = msg;
3321         audit_rmap(vcpu);
3322         audit_write_protection(vcpu);
3323         if (strcmp("pre pte write", audit_msg) != 0)
3324                 audit_mappings(vcpu);
3325         audit_writable_sptes_have_rmaps(vcpu);
3326         dbg = olddbg;
3327 }
3328
3329 #endif