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 level)
 483 {
 484         struct kvm_memory_slot *slot;
 485         unsigned long idx;
 486
 487         slot = gfn_to_memslot(kvm, gfn);
 488         if (likely(level == PT_PAGE_TABLE_LEVEL))
 489                 return &slot->rmap[gfn - slot->base_gfn];
 490
 491         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
 492                 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
 493
 494         return &slot->lpage_info[level - 2][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)
 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, sp->role.level);
 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], sp->role.level);
 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 i, write_protected = 0;
 656
 657         gfn = unalias_gfn(kvm, gfn);
 658         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
 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         for (i = PT_DIRECTORY_LEVEL;
 681              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 682                 rmapp = gfn_to_rmap(kvm, gfn, i);
 683                 spte = rmap_next(kvm, rmapp, NULL);
 684                 while (spte) {
 685                         BUG_ON(!spte);
 686                         BUG_ON(!(*spte & PT_PRESENT_MASK));
 687                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 688                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 689                         if (is_writeble_pte(*spte)) {
 690                                 rmap_remove(kvm, spte);
 691                                 --kvm->stat.lpages;
 692                                 __set_spte(spte, shadow_trap_nonpresent_pte);
 693                                 spte = NULL;
 694                                 write_protected = 1;
 695                         }
 696                         spte = rmap_next(kvm, rmapp, spte);
 697                 }
 698         }
 699
 700         return write_protected;
 701 }
 702
 703 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 704 {
 705         u64 *spte;
 706         int need_tlb_flush = 0;
 707
 708         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 709                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 710                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 711                 rmap_remove(kvm, spte);
 712                 __set_spte(spte, shadow_trap_nonpresent_pte);
 713                 need_tlb_flush = 1;
 714         }
 715         return need_tlb_flush;
 716 }
 717
 718 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 719                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 720 {
 721         int i;
 722         int retval = 0;
 723
 724         /*
 725          * If mmap_sem isn't taken, we can look the memslots with only
 726          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 727          */
 728         for (i = 0; i < kvm->nmemslots; i++) {
 729                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 730                 unsigned long start = memslot->userspace_addr;
 731                 unsigned long end;
 732
 733                 /* mmu_lock protects userspace_addr */
 734                 if (!start)
 735                         continue;
 736
 737                 end = start + (memslot->npages << PAGE_SHIFT);
 738                 if (hva >= start && hva < end) {
 739                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 740                         int idx = gfn_offset /
 741                                   KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL);
 742                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 743                         retval |= handler(kvm,
 744                                         &memslot->lpage_info[0][idx].rmap_pde);
 745                 }
 746         }
 747
 748         return retval;
 749 }
 750
 751 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 752 {
 753         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 754 }
 755
 756 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 757 {
 758         u64 *spte;
 759         int young = 0;
 760
 761         /* always return old for EPT */
 762         if (!shadow_accessed_mask)
 763                 return 0;
 764
 765         spte = rmap_next(kvm, rmapp, NULL);
 766         while (spte) {
 767                 int _young;
 768                 u64 _spte = *spte;
 769                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 770                 _young = _spte & PT_ACCESSED_MASK;
 771                 if (_young) {
 772                         young = 1;
 773                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 774                 }
 775                 spte = rmap_next(kvm, rmapp, spte);
 776         }
 777         return young;
 778 }
 779
 780 #define RMAP_RECYCLE_THRESHOLD 1000
 781
 782 static void rmap_recycle(struct kvm_vcpu *vcpu, gfn_t gfn, int lpage)
 783 {
 784         unsigned long *rmapp;
 785
 786         gfn = unalias_gfn(vcpu->kvm, gfn);
 787         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 788
 789         kvm_unmap_rmapp(vcpu->kvm, rmapp);
 790         kvm_flush_remote_tlbs(vcpu->kvm);
 791 }
 792
 793 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 794 {
 795         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 796 }
 797
 798 #ifdef MMU_DEBUG
 799 static int is_empty_shadow_page(u64 *spt)
 800 {
 801         u64 *pos;
 802         u64 *end;
 803
 804         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 805                 if (is_shadow_present_pte(*pos)) {
 806                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 807                                pos, *pos);
 808                         return 0;
 809                 }
 810         return 1;
 811 }
 812 #endif
 813
 814 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 815 {
 816         ASSERT(is_empty_shadow_page(sp->spt));
 817         list_del(&sp->link);
 818         __free_page(virt_to_page(sp->spt));
 819         __free_page(virt_to_page(sp->gfns));
 820         kfree(sp);
 821         ++kvm->arch.n_free_mmu_pages;
 822 }
 823
 824 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 825 {
 826         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 827 }
 828
 829 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 830                                                u64 *parent_pte)
 831 {
 832         struct kvm_mmu_page *sp;
 833
 834         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 835         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 836         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 837         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 838         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 839         INIT_LIST_HEAD(&sp->oos_link);
 840         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 841         sp->multimapped = 0;
 842         sp->parent_pte = parent_pte;
 843         --vcpu->kvm->arch.n_free_mmu_pages;
 844         return sp;
 845 }
 846
 847 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 848                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 849 {
 850         struct kvm_pte_chain *pte_chain;
 851         struct hlist_node *node;
 852         int i;
 853
 854         if (!parent_pte)
 855                 return;
 856         if (!sp->multimapped) {
 857                 u64 *old = sp->parent_pte;
 858
 859                 if (!old) {
 860                         sp->parent_pte = parent_pte;
 861                         return;
 862                 }
 863                 sp->multimapped = 1;
 864                 pte_chain = mmu_alloc_pte_chain(vcpu);
 865                 INIT_HLIST_HEAD(&sp->parent_ptes);
 866                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 867                 pte_chain->parent_ptes[0] = old;
 868         }
 869         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 870                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 871                         continue;
 872                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 873                         if (!pte_chain->parent_ptes[i]) {
 874                                 pte_chain->parent_ptes[i] = parent_pte;
 875                                 return;
 876                         }
 877         }
 878         pte_chain = mmu_alloc_pte_chain(vcpu);
 879         BUG_ON(!pte_chain);
 880         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 881         pte_chain->parent_ptes[0] = parent_pte;
 882 }
 883
 884 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 885                                        u64 *parent_pte)
 886 {
 887         struct kvm_pte_chain *pte_chain;
 888         struct hlist_node *node;
 889         int i;
 890
 891         if (!sp->multimapped) {
 892                 BUG_ON(sp->parent_pte != parent_pte);
 893                 sp->parent_pte = NULL;
 894                 return;
 895         }
 896         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 897                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 898                         if (!pte_chain->parent_ptes[i])
 899                                 break;
 900                         if (pte_chain->parent_ptes[i] != parent_pte)
 901                                 continue;
 902                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 903                                 && pte_chain->parent_ptes[i + 1]) {
 904                                 pte_chain->parent_ptes[i]
 905                                         = pte_chain->parent_ptes[i + 1];
 906                                 ++i;
 907                         }
 908                         pte_chain->parent_ptes[i] = NULL;
 909                         if (i == 0) {
 910                                 hlist_del(&pte_chain->link);
 911                                 mmu_free_pte_chain(pte_chain);
 912                                 if (hlist_empty(&sp->parent_ptes)) {
 913                                         sp->multimapped = 0;
 914                                         sp->parent_pte = NULL;
 915                                 }
 916                         }
 917                         return;
 918                 }
 919         BUG();
 920 }
 921
 922
 923 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 924                             mmu_parent_walk_fn fn)
 925 {
 926         struct kvm_pte_chain *pte_chain;
 927         struct hlist_node *node;
 928         struct kvm_mmu_page *parent_sp;
 929         int i;
 930
 931         if (!sp->multimapped && sp->parent_pte) {
 932                 parent_sp = page_header(__pa(sp->parent_pte));
 933                 fn(vcpu, parent_sp);
 934                 mmu_parent_walk(vcpu, parent_sp, fn);
 935                 return;
 936         }
 937         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 938                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 939                         if (!pte_chain->parent_ptes[i])
 940                                 break;
 941                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 942                         fn(vcpu, parent_sp);
 943                         mmu_parent_walk(vcpu, parent_sp, fn);
 944                 }
 945 }
 946
 947 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 948 {
 949         unsigned int index;
 950         struct kvm_mmu_page *sp = page_header(__pa(spte));
 951
 952         index = spte - sp->spt;
 953         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
 954                 sp->unsync_children++;
 955         WARN_ON(!sp->unsync_children);
 956 }
 957
 958 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 959 {
 960         struct kvm_pte_chain *pte_chain;
 961         struct hlist_node *node;
 962         int i;
 963
 964         if (!sp->parent_pte)
 965                 return;
 966
 967         if (!sp->multimapped) {
 968                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 969                 return;
 970         }
 971
 972         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 973                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 974                         if (!pte_chain->parent_ptes[i])
 975                                 break;
 976                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 977                 }
 978 }
 979
 980 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 981 {
 982         kvm_mmu_update_parents_unsync(sp);
 983         return 1;
 984 }
 985
 986 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 987                                         struct kvm_mmu_page *sp)
 988 {
 989         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 990         kvm_mmu_update_parents_unsync(sp);
 991 }
 992
 993 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 994                                     struct kvm_mmu_page *sp)
 995 {
 996         int i;
 997
 998         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 999                 sp->spt[i] = shadow_trap_nonpresent_pte;
1000 }
1001
1002 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1003                                struct kvm_mmu_page *sp)
1004 {
1005         return 1;
1006 }
1007
1008 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1009 {
1010 }
1011
1012 #define KVM_PAGE_ARRAY_NR 16
1013
1014 struct kvm_mmu_pages {
1015         struct mmu_page_and_offset {
1016                 struct kvm_mmu_page *sp;
1017                 unsigned int idx;
1018         } page[KVM_PAGE_ARRAY_NR];
1019         unsigned int nr;
1020 };
1021
1022 #define for_each_unsync_children(bitmap, idx)           \
1023         for (idx = find_first_bit(bitmap, 512);         \
1024              idx < 512;                                 \
1025              idx = find_next_bit(bitmap, 512, idx+1))
1026
1027 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1028                          int idx)
1029 {
1030         int i;
1031
1032         if (sp->unsync)
1033                 for (i=0; i < pvec->nr; i++)
1034                         if (pvec->page[i].sp == sp)
1035                                 return 0;
1036
1037         pvec->page[pvec->nr].sp = sp;
1038         pvec->page[pvec->nr].idx = idx;
1039         pvec->nr++;
1040         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1041 }
1042
1043 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1044                            struct kvm_mmu_pages *pvec)
1045 {
1046         int i, ret, nr_unsync_leaf = 0;
1047
1048         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1049                 u64 ent = sp->spt[i];
1050
1051                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1052                         struct kvm_mmu_page *child;
1053                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1054
1055                         if (child->unsync_children) {
1056                                 if (mmu_pages_add(pvec, child, i))
1057                                         return -ENOSPC;
1058
1059                                 ret = __mmu_unsync_walk(child, pvec);
1060                                 if (!ret)
1061                                         __clear_bit(i, sp->unsync_child_bitmap);
1062                                 else if (ret > 0)
1063                                         nr_unsync_leaf += ret;
1064                                 else
1065                                         return ret;
1066                         }
1067
1068                         if (child->unsync) {
1069                                 nr_unsync_leaf++;
1070                                 if (mmu_pages_add(pvec, child, i))
1071                                         return -ENOSPC;
1072                         }
1073                 }
1074         }
1075
1076         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1077                 sp->unsync_children = 0;
1078
1079         return nr_unsync_leaf;
1080 }
1081
1082 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1083                            struct kvm_mmu_pages *pvec)
1084 {
1085         if (!sp->unsync_children)
1086                 return 0;
1087
1088         mmu_pages_add(pvec, sp, 0);
1089         return __mmu_unsync_walk(sp, pvec);
1090 }
1091
1092 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1093 {
1094         unsigned index;
1095         struct hlist_head *bucket;
1096         struct kvm_mmu_page *sp;
1097         struct hlist_node *node;
1098
1099         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1100         index = kvm_page_table_hashfn(gfn);
1101         bucket = &kvm->arch.mmu_page_hash[index];
1102         hlist_for_each_entry(sp, node, bucket, hash_link)
1103                 if (sp->gfn == gfn && !sp->role.direct
1104                     && !sp->role.invalid) {
1105                         pgprintk("%s: found role %x\n",
1106                                  __func__, sp->role.word);
1107                         return sp;
1108                 }
1109         return NULL;
1110 }
1111
1112 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1113 {
1114         WARN_ON(!sp->unsync);
1115         sp->unsync = 0;
1116         --kvm->stat.mmu_unsync;
1117 }
1118
1119 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1120
1121 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1122 {
1123         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1124                 kvm_mmu_zap_page(vcpu->kvm, sp);
1125                 return 1;
1126         }
1127
1128         trace_kvm_mmu_sync_page(sp);
1129         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1130                 kvm_flush_remote_tlbs(vcpu->kvm);
1131         kvm_unlink_unsync_page(vcpu->kvm, sp);
1132         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1133                 kvm_mmu_zap_page(vcpu->kvm, sp);
1134                 return 1;
1135         }
1136
1137         kvm_mmu_flush_tlb(vcpu);
1138         return 0;
1139 }
1140
1141 struct mmu_page_path {
1142         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1143         unsigned int idx[PT64_ROOT_LEVEL-1];
1144 };
1145
1146 #define for_each_sp(pvec, sp, parents, i)                       \
1147                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1148                         sp = pvec.page[i].sp;                   \
1149                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1150                         i = mmu_pages_next(&pvec, &parents, i))
1151
1152 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1153                           struct mmu_page_path *parents,
1154                           int i)
1155 {
1156         int n;
1157
1158         for (n = i+1; n < pvec->nr; n++) {
1159                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1160
1161                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1162                         parents->idx[0] = pvec->page[n].idx;
1163                         return n;
1164                 }
1165
1166                 parents->parent[sp->role.level-2] = sp;
1167                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1168         }
1169
1170         return n;
1171 }
1172
1173 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1174 {
1175         struct kvm_mmu_page *sp;
1176         unsigned int level = 0;
1177
1178         do {
1179                 unsigned int idx = parents->idx[level];
1180
1181                 sp = parents->parent[level];
1182                 if (!sp)
1183                         return;
1184
1185                 --sp->unsync_children;
1186                 WARN_ON((int)sp->unsync_children < 0);
1187                 __clear_bit(idx, sp->unsync_child_bitmap);
1188                 level++;
1189         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1190 }
1191
1192 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1193                                struct mmu_page_path *parents,
1194                                struct kvm_mmu_pages *pvec)
1195 {
1196         parents->parent[parent->role.level-1] = NULL;
1197         pvec->nr = 0;
1198 }
1199
1200 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1201                               struct kvm_mmu_page *parent)
1202 {
1203         int i;
1204         struct kvm_mmu_page *sp;
1205         struct mmu_page_path parents;
1206         struct kvm_mmu_pages pages;
1207
1208         kvm_mmu_pages_init(parent, &parents, &pages);
1209         while (mmu_unsync_walk(parent, &pages)) {
1210                 int protected = 0;
1211
1212                 for_each_sp(pages, sp, parents, i)
1213                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1214
1215                 if (protected)
1216                         kvm_flush_remote_tlbs(vcpu->kvm);
1217
1218                 for_each_sp(pages, sp, parents, i) {
1219                         kvm_sync_page(vcpu, sp);
1220                         mmu_pages_clear_parents(&parents);
1221                 }
1222                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1223                 kvm_mmu_pages_init(parent, &parents, &pages);
1224         }
1225 }
1226
1227 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1228                                              gfn_t gfn,
1229                                              gva_t gaddr,
1230                                              unsigned level,
1231                                              int direct,
1232                                              unsigned access,
1233                                              u64 *parent_pte)
1234 {
1235         union kvm_mmu_page_role role;
1236         unsigned index;
1237         unsigned quadrant;
1238         struct hlist_head *bucket;
1239         struct kvm_mmu_page *sp;
1240         struct hlist_node *node, *tmp;
1241
1242         role = vcpu->arch.mmu.base_role;
1243         role.level = level;
1244         role.direct = direct;
1245         role.access = access;
1246         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1247                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1248                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1249                 role.quadrant = quadrant;
1250         }
1251         index = kvm_page_table_hashfn(gfn);
1252         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1253         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1254                 if (sp->gfn == gfn) {
1255                         if (sp->unsync)
1256                                 if (kvm_sync_page(vcpu, sp))
1257                                         continue;
1258
1259                         if (sp->role.word != role.word)
1260                                 continue;
1261
1262                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1263                         if (sp->unsync_children) {
1264                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1265                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1266                         }
1267                         trace_kvm_mmu_get_page(sp, false);
1268                         return sp;
1269                 }
1270         ++vcpu->kvm->stat.mmu_cache_miss;
1271         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1272         if (!sp)
1273                 return sp;
1274         sp->gfn = gfn;
1275         sp->role = role;
1276         hlist_add_head(&sp->hash_link, bucket);
1277         if (!direct) {
1278                 if (rmap_write_protect(vcpu->kvm, gfn))
1279                         kvm_flush_remote_tlbs(vcpu->kvm);
1280                 account_shadowed(vcpu->kvm, gfn);
1281         }
1282         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1283                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1284         else
1285                 nonpaging_prefetch_page(vcpu, sp);
1286         trace_kvm_mmu_get_page(sp, true);
1287         return sp;
1288 }
1289
1290 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1291                              struct kvm_vcpu *vcpu, u64 addr)
1292 {
1293         iterator->addr = addr;
1294         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1295         iterator->level = vcpu->arch.mmu.shadow_root_level;
1296         if (iterator->level == PT32E_ROOT_LEVEL) {
1297                 iterator->shadow_addr
1298                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1299                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1300                 --iterator->level;
1301                 if (!iterator->shadow_addr)
1302                         iterator->level = 0;
1303         }
1304 }
1305
1306 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1307 {
1308         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1309                 return false;
1310
1311         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1312                 if (is_large_pte(*iterator->sptep))
1313                         return false;
1314
1315         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1316         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1317         return true;
1318 }
1319
1320 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1321 {
1322         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1323         --iterator->level;
1324 }
1325
1326 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1327                                          struct kvm_mmu_page *sp)
1328 {
1329         unsigned i;
1330         u64 *pt;
1331         u64 ent;
1332
1333         pt = sp->spt;
1334
1335         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1336                 ent = pt[i];
1337
1338                 if (is_shadow_present_pte(ent)) {
1339                         if (!is_last_spte(ent, sp->role.level)) {
1340                                 ent &= PT64_BASE_ADDR_MASK;
1341                                 mmu_page_remove_parent_pte(page_header(ent),
1342                                                            &pt[i]);
1343                         } else {
1344                                 if (is_large_pte(ent))
1345                                         --kvm->stat.lpages;
1346                                 rmap_remove(kvm, &pt[i]);
1347                         }
1348                 }
1349                 pt[i] = shadow_trap_nonpresent_pte;
1350         }
1351 }
1352
1353 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1354 {
1355         mmu_page_remove_parent_pte(sp, parent_pte);
1356 }
1357
1358 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1359 {
1360         int i;
1361         struct kvm_vcpu *vcpu;
1362
1363         kvm_for_each_vcpu(i, vcpu, kvm)
1364                 vcpu->arch.last_pte_updated = NULL;
1365 }
1366
1367 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1368 {
1369         u64 *parent_pte;
1370
1371         while (sp->multimapped || sp->parent_pte) {
1372                 if (!sp->multimapped)
1373                         parent_pte = sp->parent_pte;
1374                 else {
1375                         struct kvm_pte_chain *chain;
1376
1377                         chain = container_of(sp->parent_ptes.first,
1378                                              struct kvm_pte_chain, link);
1379                         parent_pte = chain->parent_ptes[0];
1380                 }
1381                 BUG_ON(!parent_pte);
1382                 kvm_mmu_put_page(sp, parent_pte);
1383                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1384         }
1385 }
1386
1387 static int mmu_zap_unsync_children(struct kvm *kvm,
1388                                    struct kvm_mmu_page *parent)
1389 {
1390         int i, zapped = 0;
1391         struct mmu_page_path parents;
1392         struct kvm_mmu_pages pages;
1393
1394         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1395                 return 0;
1396
1397         kvm_mmu_pages_init(parent, &parents, &pages);
1398         while (mmu_unsync_walk(parent, &pages)) {
1399                 struct kvm_mmu_page *sp;
1400
1401                 for_each_sp(pages, sp, parents, i) {
1402                         kvm_mmu_zap_page(kvm, sp);
1403                         mmu_pages_clear_parents(&parents);
1404                 }
1405                 zapped += pages.nr;
1406                 kvm_mmu_pages_init(parent, &parents, &pages);
1407         }
1408
1409         return zapped;
1410 }
1411
1412 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1413 {
1414         int ret;
1415
1416         trace_kvm_mmu_zap_page(sp);
1417         ++kvm->stat.mmu_shadow_zapped;
1418         ret = mmu_zap_unsync_children(kvm, sp);
1419         kvm_mmu_page_unlink_children(kvm, sp);
1420         kvm_mmu_unlink_parents(kvm, sp);
1421         kvm_flush_remote_tlbs(kvm);
1422         if (!sp->role.invalid && !sp->role.direct)
1423                 unaccount_shadowed(kvm, sp->gfn);
1424         if (sp->unsync)
1425                 kvm_unlink_unsync_page(kvm, sp);
1426         if (!sp->root_count) {
1427                 hlist_del(&sp->hash_link);
1428                 kvm_mmu_free_page(kvm, sp);
1429         } else {
1430                 sp->role.invalid = 1;
1431                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1432                 kvm_reload_remote_mmus(kvm);
1433         }
1434         kvm_mmu_reset_last_pte_updated(kvm);
1435         return ret;
1436 }
1437
1438 /*
1439  * Changing the number of mmu pages allocated to the vm
1440  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1441  */
1442 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1443 {
1444         int used_pages;
1445
1446         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1447         used_pages = max(0, used_pages);
1448
1449         /*
1450          * If we set the number of mmu pages to be smaller be than the
1451          * number of actived pages , we must to free some mmu pages before we
1452          * change the value
1453          */
1454
1455         if (used_pages > kvm_nr_mmu_pages) {
1456                 while (used_pages > kvm_nr_mmu_pages) {
1457                         struct kvm_mmu_page *page;
1458
1459                         page = container_of(kvm->arch.active_mmu_pages.prev,
1460                                             struct kvm_mmu_page, link);
1461                         kvm_mmu_zap_page(kvm, page);
1462                         used_pages--;
1463                 }
1464                 kvm->arch.n_free_mmu_pages = 0;
1465         }
1466         else
1467                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1468                                          - kvm->arch.n_alloc_mmu_pages;
1469
1470         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1471 }
1472
1473 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1474 {
1475         unsigned index;
1476         struct hlist_head *bucket;
1477         struct kvm_mmu_page *sp;
1478         struct hlist_node *node, *n;
1479         int r;
1480
1481         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1482         r = 0;
1483         index = kvm_page_table_hashfn(gfn);
1484         bucket = &kvm->arch.mmu_page_hash[index];
1485         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1486                 if (sp->gfn == gfn && !sp->role.direct) {
1487                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1488                                  sp->role.word);
1489                         r = 1;
1490                         if (kvm_mmu_zap_page(kvm, sp))
1491                                 n = bucket->first;
1492                 }
1493         return r;
1494 }
1495
1496 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1497 {
1498         unsigned index;
1499         struct hlist_head *bucket;
1500         struct kvm_mmu_page *sp;
1501         struct hlist_node *node, *nn;
1502
1503         index = kvm_page_table_hashfn(gfn);
1504         bucket = &kvm->arch.mmu_page_hash[index];
1505         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1506                 if (sp->gfn == gfn && !sp->role.direct
1507                     && !sp->role.invalid) {
1508                         pgprintk("%s: zap %lx %x\n",
1509                                  __func__, gfn, sp->role.word);
1510                         kvm_mmu_zap_page(kvm, sp);
1511                 }
1512         }
1513 }
1514
1515 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1516 {
1517         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1518         struct kvm_mmu_page *sp = page_header(__pa(pte));
1519
1520         __set_bit(slot, sp->slot_bitmap);
1521 }
1522
1523 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1524 {
1525         int i;
1526         u64 *pt = sp->spt;
1527
1528         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1529                 return;
1530
1531         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1532                 if (pt[i] == shadow_notrap_nonpresent_pte)
1533                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1534         }
1535 }
1536
1537 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1538 {
1539         struct page *page;
1540
1541         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1542
1543         if (gpa == UNMAPPED_GVA)
1544                 return NULL;
1545
1546         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1547
1548         return page;
1549 }
1550
1551 /*
1552  * The function is based on mtrr_type_lookup() in
1553  * arch/x86/kernel/cpu/mtrr/generic.c
1554  */
1555 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1556                          u64 start, u64 end)
1557 {
1558         int i;
1559         u64 base, mask;
1560         u8 prev_match, curr_match;
1561         int num_var_ranges = KVM_NR_VAR_MTRR;
1562
1563         if (!mtrr_state->enabled)
1564                 return 0xFF;
1565
1566         /* Make end inclusive end, instead of exclusive */
1567         end--;
1568
1569         /* Look in fixed ranges. Just return the type as per start */
1570         if (mtrr_state->have_fixed && (start < 0x100000)) {
1571                 int idx;
1572
1573                 if (start < 0x80000) {
1574                         idx = 0;
1575                         idx += (start >> 16);
1576                         return mtrr_state->fixed_ranges[idx];
1577                 } else if (start < 0xC0000) {
1578                         idx = 1 * 8;
1579                         idx += ((start - 0x80000) >> 14);
1580                         return mtrr_state->fixed_ranges[idx];
1581                 } else if (start < 0x1000000) {
1582                         idx = 3 * 8;
1583                         idx += ((start - 0xC0000) >> 12);
1584                         return mtrr_state->fixed_ranges[idx];
1585                 }
1586         }
1587
1588         /*
1589          * Look in variable ranges
1590          * Look of multiple ranges matching this address and pick type
1591          * as per MTRR precedence
1592          */
1593         if (!(mtrr_state->enabled & 2))
1594                 return mtrr_state->def_type;
1595
1596         prev_match = 0xFF;
1597         for (i = 0; i < num_var_ranges; ++i) {
1598                 unsigned short start_state, end_state;
1599
1600                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1601                         continue;
1602
1603                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1604                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1605                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1606                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1607
1608                 start_state = ((start & mask) == (base & mask));
1609                 end_state = ((end & mask) == (base & mask));
1610                 if (start_state != end_state)
1611                         return 0xFE;
1612
1613                 if ((start & mask) != (base & mask))
1614                         continue;
1615
1616                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1617                 if (prev_match == 0xFF) {
1618                         prev_match = curr_match;
1619                         continue;
1620                 }
1621
1622                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1623                     curr_match == MTRR_TYPE_UNCACHABLE)
1624                         return MTRR_TYPE_UNCACHABLE;
1625
1626                 if ((prev_match == MTRR_TYPE_WRBACK &&
1627                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1628                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1629                      curr_match == MTRR_TYPE_WRBACK)) {
1630                         prev_match = MTRR_TYPE_WRTHROUGH;
1631                         curr_match = MTRR_TYPE_WRTHROUGH;
1632                 }
1633
1634                 if (prev_match != curr_match)
1635                         return MTRR_TYPE_UNCACHABLE;
1636         }
1637
1638         if (prev_match != 0xFF)
1639                 return prev_match;
1640
1641         return mtrr_state->def_type;
1642 }
1643
1644 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1645 {
1646         u8 mtrr;
1647
1648         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1649                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1650         if (mtrr == 0xfe || mtrr == 0xff)
1651                 mtrr = MTRR_TYPE_WRBACK;
1652         return mtrr;
1653 }
1654 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1655
1656 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1657 {
1658         unsigned index;
1659         struct hlist_head *bucket;
1660         struct kvm_mmu_page *s;
1661         struct hlist_node *node, *n;
1662
1663         trace_kvm_mmu_unsync_page(sp);
1664         index = kvm_page_table_hashfn(sp->gfn);
1665         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1666         /* don't unsync if pagetable is shadowed with multiple roles */
1667         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1668                 if (s->gfn != sp->gfn || s->role.direct)
1669                         continue;
1670                 if (s->role.word != sp->role.word)
1671                         return 1;
1672         }
1673         ++vcpu->kvm->stat.mmu_unsync;
1674         sp->unsync = 1;
1675
1676         kvm_mmu_mark_parents_unsync(vcpu, sp);
1677
1678         mmu_convert_notrap(sp);
1679         return 0;
1680 }
1681
1682 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1683                                   bool can_unsync)
1684 {
1685         struct kvm_mmu_page *shadow;
1686
1687         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1688         if (shadow) {
1689                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1690                         return 1;
1691                 if (shadow->unsync)
1692                         return 0;
1693                 if (can_unsync && oos_shadow)
1694                         return kvm_unsync_page(vcpu, shadow);
1695                 return 1;
1696         }
1697         return 0;
1698 }
1699
1700 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1701                     unsigned pte_access, int user_fault,
1702                     int write_fault, int dirty, int largepage,
1703                     gfn_t gfn, pfn_t pfn, bool speculative,
1704                     bool can_unsync)
1705 {
1706         u64 spte;
1707         int ret = 0;
1708
1709         /*
1710          * We don't set the accessed bit, since we sometimes want to see
1711          * whether the guest actually used the pte (in order to detect
1712          * demand paging).
1713          */
1714         spte = shadow_base_present_pte | shadow_dirty_mask;
1715         if (!speculative)
1716                 spte |= shadow_accessed_mask;
1717         if (!dirty)
1718                 pte_access &= ~ACC_WRITE_MASK;
1719         if (pte_access & ACC_EXEC_MASK)
1720                 spte |= shadow_x_mask;
1721         else
1722                 spte |= shadow_nx_mask;
1723         if (pte_access & ACC_USER_MASK)
1724                 spte |= shadow_user_mask;
1725         if (largepage)
1726                 spte |= PT_PAGE_SIZE_MASK;
1727         if (tdp_enabled)
1728                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1729                         kvm_is_mmio_pfn(pfn));
1730
1731         spte |= (u64)pfn << PAGE_SHIFT;
1732
1733         if ((pte_access & ACC_WRITE_MASK)
1734             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1735
1736                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1737                         ret = 1;
1738                         spte = shadow_trap_nonpresent_pte;
1739                         goto set_pte;
1740                 }
1741
1742                 spte |= PT_WRITABLE_MASK;
1743
1744                 /*
1745                  * Optimization: for pte sync, if spte was writable the hash
1746                  * lookup is unnecessary (and expensive). Write protection
1747                  * is responsibility of mmu_get_page / kvm_sync_page.
1748                  * Same reasoning can be applied to dirty page accounting.
1749                  */
1750                 if (!can_unsync && is_writeble_pte(*sptep))
1751                         goto set_pte;
1752
1753                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1754                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1755                                  __func__, gfn);
1756                         ret = 1;
1757                         pte_access &= ~ACC_WRITE_MASK;
1758                         if (is_writeble_pte(spte))
1759                                 spte &= ~PT_WRITABLE_MASK;
1760                 }
1761         }
1762
1763         if (pte_access & ACC_WRITE_MASK)
1764                 mark_page_dirty(vcpu->kvm, gfn);
1765
1766 set_pte:
1767         __set_spte(sptep, spte);
1768         return ret;
1769 }
1770
1771 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1772                          unsigned pt_access, unsigned pte_access,
1773                          int user_fault, int write_fault, int dirty,
1774                          int *ptwrite, int largepage, gfn_t gfn,
1775                          pfn_t pfn, bool speculative)
1776 {
1777         int was_rmapped = 0;
1778         int was_writeble = is_writeble_pte(*sptep);
1779         int rmap_count;
1780
1781         pgprintk("%s: spte %llx access %x write_fault %d"
1782                  " user_fault %d gfn %lx\n",
1783                  __func__, *sptep, pt_access,
1784                  write_fault, user_fault, gfn);
1785
1786         if (is_rmap_spte(*sptep)) {
1787                 /*
1788                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1789                  * the parent of the now unreachable PTE.
1790                  */
1791                 if (largepage && !is_large_pte(*sptep)) {
1792                         struct kvm_mmu_page *child;
1793                         u64 pte = *sptep;
1794
1795                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1796                         mmu_page_remove_parent_pte(child, sptep);
1797                 } else if (pfn != spte_to_pfn(*sptep)) {
1798                         pgprintk("hfn old %lx new %lx\n",
1799                                  spte_to_pfn(*sptep), pfn);
1800                         rmap_remove(vcpu->kvm, sptep);
1801                 } else
1802                         was_rmapped = 1;
1803         }
1804         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1805                       dirty, largepage, gfn, pfn, speculative, true)) {
1806                 if (write_fault)
1807                         *ptwrite = 1;
1808                 kvm_x86_ops->tlb_flush(vcpu);
1809         }
1810
1811         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1812         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1813                  is_large_pte(*sptep)? "2MB" : "4kB",
1814                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1815                  *sptep, sptep);
1816         if (!was_rmapped && is_large_pte(*sptep))
1817                 ++vcpu->kvm->stat.lpages;
1818
1819         page_header_update_slot(vcpu->kvm, sptep, gfn);
1820         if (!was_rmapped) {
1821                 rmap_count = rmap_add(vcpu, sptep, gfn);
1822                 if (!is_rmap_spte(*sptep))
1823                         kvm_release_pfn_clean(pfn);
1824                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1825                         rmap_recycle(vcpu, gfn, largepage);
1826         } else {
1827                 if (was_writeble)
1828                         kvm_release_pfn_dirty(pfn);
1829                 else
1830                         kvm_release_pfn_clean(pfn);
1831         }
1832         if (speculative) {
1833                 vcpu->arch.last_pte_updated = sptep;
1834                 vcpu->arch.last_pte_gfn = gfn;
1835         }
1836 }
1837
1838 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1839 {
1840 }
1841
1842 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1843                         int largepage, gfn_t gfn, pfn_t pfn)
1844 {
1845         struct kvm_shadow_walk_iterator iterator;
1846         struct kvm_mmu_page *sp;
1847         int pt_write = 0;
1848         gfn_t pseudo_gfn;
1849
1850         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1851                 if (iterator.level == PT_PAGE_TABLE_LEVEL
1852                     || (largepage && iterator.level == PT_DIRECTORY_LEVEL)) {
1853                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1854                                      0, write, 1, &pt_write,
1855                                      largepage, gfn, pfn, false);
1856                         ++vcpu->stat.pf_fixed;
1857                         break;
1858                 }
1859
1860                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1861                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1862                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1863                                               iterator.level - 1,
1864                                               1, ACC_ALL, iterator.sptep);
1865                         if (!sp) {
1866                                 pgprintk("nonpaging_map: ENOMEM\n");
1867                                 kvm_release_pfn_clean(pfn);
1868                                 return -ENOMEM;
1869                         }
1870
1871                         __set_spte(iterator.sptep,
1872                                    __pa(sp->spt)
1873                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
1874                                    | shadow_user_mask | shadow_x_mask);
1875                 }
1876         }
1877         return pt_write;
1878 }
1879
1880 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1881 {
1882         int r;
1883         int largepage = 0;
1884         pfn_t pfn;
1885         unsigned long mmu_seq;
1886
1887         if (is_largepage_backed(vcpu, gfn &
1888                         ~(KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL) - 1))) {
1889                 gfn &= ~(KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL) - 1);
1890                 largepage = 1;
1891         }
1892
1893         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1894         smp_rmb();
1895         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1896
1897         /* mmio */
1898         if (is_error_pfn(pfn)) {
1899                 kvm_release_pfn_clean(pfn);
1900                 return 1;
1901         }
1902
1903         spin_lock(&vcpu->kvm->mmu_lock);
1904         if (mmu_notifier_retry(vcpu, mmu_seq))
1905                 goto out_unlock;
1906         kvm_mmu_free_some_pages(vcpu);
1907         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1908         spin_unlock(&vcpu->kvm->mmu_lock);
1909
1910
1911         return r;
1912
1913 out_unlock:
1914         spin_unlock(&vcpu->kvm->mmu_lock);
1915         kvm_release_pfn_clean(pfn);
1916         return 0;
1917 }
1918
1919
1920 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1921 {
1922         int i;
1923         struct kvm_mmu_page *sp;
1924
1925         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1926                 return;
1927         spin_lock(&vcpu->kvm->mmu_lock);
1928         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1929                 hpa_t root = vcpu->arch.mmu.root_hpa;
1930
1931                 sp = page_header(root);
1932                 --sp->root_count;
1933                 if (!sp->root_count && sp->role.invalid)
1934                         kvm_mmu_zap_page(vcpu->kvm, sp);
1935                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1936                 spin_unlock(&vcpu->kvm->mmu_lock);
1937                 return;
1938         }
1939         for (i = 0; i < 4; ++i) {
1940                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1941
1942                 if (root) {
1943                         root &= PT64_BASE_ADDR_MASK;
1944                         sp = page_header(root);
1945                         --sp->root_count;
1946                         if (!sp->root_count && sp->role.invalid)
1947                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1948                 }
1949                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1950         }
1951         spin_unlock(&vcpu->kvm->mmu_lock);
1952         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1953 }
1954
1955 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
1956 {
1957         int ret = 0;
1958
1959         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
1960                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
1961                 ret = 1;
1962         }
1963
1964         return ret;
1965 }
1966
1967 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
1968 {
1969         int i;
1970         gfn_t root_gfn;
1971         struct kvm_mmu_page *sp;
1972         int direct = 0;
1973         u64 pdptr;
1974
1975         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1976
1977         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1978                 hpa_t root = vcpu->arch.mmu.root_hpa;
1979
1980                 ASSERT(!VALID_PAGE(root));
1981                 if (tdp_enabled)
1982                         direct = 1;
1983                 if (mmu_check_root(vcpu, root_gfn))
1984                         return 1;
1985                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1986                                       PT64_ROOT_LEVEL, direct,
1987                                       ACC_ALL, NULL);
1988                 root = __pa(sp->spt);
1989                 ++sp->root_count;
1990                 vcpu->arch.mmu.root_hpa = root;
1991                 return 0;
1992         }
1993         direct = !is_paging(vcpu);
1994         if (tdp_enabled)
1995                 direct = 1;
1996         for (i = 0; i < 4; ++i) {
1997                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1998
1999                 ASSERT(!VALID_PAGE(root));
2000                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2001                         pdptr = kvm_pdptr_read(vcpu, i);
2002                         if (!is_present_gpte(pdptr)) {
2003                                 vcpu->arch.mmu.pae_root[i] = 0;
2004                                 continue;
2005                         }
2006                         root_gfn = pdptr >> PAGE_SHIFT;
2007                 } else if (vcpu->arch.mmu.root_level == 0)
2008                         root_gfn = 0;
2009                 if (mmu_check_root(vcpu, root_gfn))
2010                         return 1;
2011                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2012                                       PT32_ROOT_LEVEL, direct,
2013                                       ACC_ALL, NULL);
2014                 root = __pa(sp->spt);
2015                 ++sp->root_count;
2016                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2017         }
2018         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2019         return 0;
2020 }
2021
2022 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2023 {
2024         int i;
2025         struct kvm_mmu_page *sp;
2026
2027         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2028                 return;
2029         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2030                 hpa_t root = vcpu->arch.mmu.root_hpa;
2031                 sp = page_header(root);
2032                 mmu_sync_children(vcpu, sp);
2033                 return;
2034         }
2035         for (i = 0; i < 4; ++i) {
2036                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2037
2038                 if (root && VALID_PAGE(root)) {
2039                         root &= PT64_BASE_ADDR_MASK;
2040                         sp = page_header(root);
2041                         mmu_sync_children(vcpu, sp);
2042                 }
2043         }
2044 }
2045
2046 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2047 {
2048         spin_lock(&vcpu->kvm->mmu_lock);
2049         mmu_sync_roots(vcpu);
2050         spin_unlock(&vcpu->kvm->mmu_lock);
2051 }
2052
2053 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2054 {
2055         return vaddr;
2056 }
2057
2058 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2059                                 u32 error_code)
2060 {
2061         gfn_t gfn;
2062         int r;
2063
2064         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2065         r = mmu_topup_memory_caches(vcpu);
2066         if (r)
2067                 return r;
2068
2069         ASSERT(vcpu);
2070         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2071
2072         gfn = gva >> PAGE_SHIFT;
2073
2074         return nonpaging_map(vcpu, gva & PAGE_MASK,
2075                              error_code & PFERR_WRITE_MASK, gfn);
2076 }
2077
2078 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2079                                 u32 error_code)
2080 {
2081         pfn_t pfn;
2082         int r;
2083         int largepage = 0;
2084         gfn_t gfn = gpa >> PAGE_SHIFT;
2085         unsigned long mmu_seq;
2086
2087         ASSERT(vcpu);
2088         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2089
2090         r = mmu_topup_memory_caches(vcpu);
2091         if (r)
2092                 return r;
2093
2094         if (is_largepage_backed(vcpu, gfn &
2095                         ~(KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL) - 1))) {
2096                 gfn &= ~(KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL) - 1);
2097                 largepage = 1;
2098         }
2099         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2100         smp_rmb();
2101         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2102         if (is_error_pfn(pfn)) {
2103                 kvm_release_pfn_clean(pfn);
2104                 return 1;
2105         }
2106         spin_lock(&vcpu->kvm->mmu_lock);
2107         if (mmu_notifier_retry(vcpu, mmu_seq))
2108                 goto out_unlock;
2109         kvm_mmu_free_some_pages(vcpu);
2110         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2111                          largepage, gfn, pfn);
2112         spin_unlock(&vcpu->kvm->mmu_lock);
2113
2114         return r;
2115
2116 out_unlock:
2117         spin_unlock(&vcpu->kvm->mmu_lock);
2118         kvm_release_pfn_clean(pfn);
2119         return 0;
2120 }
2121
2122 static void nonpaging_free(struct kvm_vcpu *vcpu)
2123 {
2124         mmu_free_roots(vcpu);
2125 }
2126
2127 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2128 {
2129         struct kvm_mmu *context = &vcpu->arch.mmu;
2130
2131         context->new_cr3 = nonpaging_new_cr3;
2132         context->page_fault = nonpaging_page_fault;
2133         context->gva_to_gpa = nonpaging_gva_to_gpa;
2134         context->free = nonpaging_free;
2135         context->prefetch_page = nonpaging_prefetch_page;
2136         context->sync_page = nonpaging_sync_page;
2137         context->invlpg = nonpaging_invlpg;
2138         context->root_level = 0;
2139         context->shadow_root_level = PT32E_ROOT_LEVEL;
2140         context->root_hpa = INVALID_PAGE;
2141         return 0;
2142 }
2143
2144 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2145 {
2146         ++vcpu->stat.tlb_flush;
2147         kvm_x86_ops->tlb_flush(vcpu);
2148 }
2149
2150 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2151 {
2152         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2153         mmu_free_roots(vcpu);
2154 }
2155
2156 static void inject_page_fault(struct kvm_vcpu *vcpu,
2157                               u64 addr,
2158                               u32 err_code)
2159 {
2160         kvm_inject_page_fault(vcpu, addr, err_code);
2161 }
2162
2163 static void paging_free(struct kvm_vcpu *vcpu)
2164 {
2165         nonpaging_free(vcpu);
2166 }
2167
2168 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2169 {
2170         int bit7;
2171
2172         bit7 = (gpte >> 7) & 1;
2173         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2174 }
2175
2176 #define PTTYPE 64
2177 #include "paging_tmpl.h"
2178 #undef PTTYPE
2179
2180 #define PTTYPE 32
2181 #include "paging_tmpl.h"
2182 #undef PTTYPE
2183
2184 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2185 {
2186         struct kvm_mmu *context = &vcpu->arch.mmu;
2187         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2188         u64 exb_bit_rsvd = 0;
2189
2190         if (!is_nx(vcpu))
2191                 exb_bit_rsvd = rsvd_bits(63, 63);
2192         switch (level) {
2193         case PT32_ROOT_LEVEL:
2194                 /* no rsvd bits for 2 level 4K page table entries */
2195                 context->rsvd_bits_mask[0][1] = 0;
2196                 context->rsvd_bits_mask[0][0] = 0;
2197                 if (is_cpuid_PSE36())
2198                         /* 36bits PSE 4MB page */
2199                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2200                 else
2201                         /* 32 bits PSE 4MB page */
2202                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2203                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2204                 break;
2205         case PT32E_ROOT_LEVEL:
2206                 context->rsvd_bits_mask[0][2] =
2207                         rsvd_bits(maxphyaddr, 63) |
2208                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2209                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2210                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2211                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2212                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2213                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2214                         rsvd_bits(maxphyaddr, 62) |
2215                         rsvd_bits(13, 20);              /* large page */
2216                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2217                 break;
2218         case PT64_ROOT_LEVEL:
2219                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2220                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2221                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2222                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2223                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2224                         rsvd_bits(maxphyaddr, 51);
2225                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2226                         rsvd_bits(maxphyaddr, 51);
2227                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2228                 context->rsvd_bits_mask[1][2] = context->rsvd_bits_mask[0][2];
2229                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2230                         rsvd_bits(maxphyaddr, 51) |
2231                         rsvd_bits(13, 20);              /* large page */
2232                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2233                 break;
2234         }
2235 }
2236
2237 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2238 {
2239         struct kvm_mmu *context = &vcpu->arch.mmu;
2240
2241         ASSERT(is_pae(vcpu));
2242         context->new_cr3 = paging_new_cr3;
2243         context->page_fault = paging64_page_fault;
2244         context->gva_to_gpa = paging64_gva_to_gpa;
2245         context->prefetch_page = paging64_prefetch_page;
2246         context->sync_page = paging64_sync_page;
2247         context->invlpg = paging64_invlpg;
2248         context->free = paging_free;
2249         context->root_level = level;
2250         context->shadow_root_level = level;
2251         context->root_hpa = INVALID_PAGE;
2252         return 0;
2253 }
2254
2255 static int paging64_init_context(struct kvm_vcpu *vcpu)
2256 {
2257         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2258         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2259 }
2260
2261 static int paging32_init_context(struct kvm_vcpu *vcpu)
2262 {
2263         struct kvm_mmu *context = &vcpu->arch.mmu;
2264
2265         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2266         context->new_cr3 = paging_new_cr3;
2267         context->page_fault = paging32_page_fault;
2268         context->gva_to_gpa = paging32_gva_to_gpa;
2269         context->free = paging_free;
2270         context->prefetch_page = paging32_prefetch_page;
2271         context->sync_page = paging32_sync_page;
2272         context->invlpg = paging32_invlpg;
2273         context->root_level = PT32_ROOT_LEVEL;
2274         context->shadow_root_level = PT32E_ROOT_LEVEL;
2275         context->root_hpa = INVALID_PAGE;
2276         return 0;
2277 }
2278
2279 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2280 {
2281         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2282         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2283 }
2284
2285 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2286 {
2287         struct kvm_mmu *context = &vcpu->arch.mmu;
2288
2289         context->new_cr3 = nonpaging_new_cr3;
2290         context->page_fault = tdp_page_fault;
2291         context->free = nonpaging_free;
2292         context->prefetch_page = nonpaging_prefetch_page;
2293         context->sync_page = nonpaging_sync_page;
2294         context->invlpg = nonpaging_invlpg;
2295         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2296         context->root_hpa = INVALID_PAGE;
2297
2298         if (!is_paging(vcpu)) {
2299                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2300                 context->root_level = 0;
2301         } else if (is_long_mode(vcpu)) {
2302                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2303                 context->gva_to_gpa = paging64_gva_to_gpa;
2304                 context->root_level = PT64_ROOT_LEVEL;
2305         } else if (is_pae(vcpu)) {
2306                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2307                 context->gva_to_gpa = paging64_gva_to_gpa;
2308                 context->root_level = PT32E_ROOT_LEVEL;
2309         } else {
2310                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2311                 context->gva_to_gpa = paging32_gva_to_gpa;
2312                 context->root_level = PT32_ROOT_LEVEL;
2313         }
2314
2315         return 0;
2316 }
2317
2318 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2319 {
2320         int r;
2321
2322         ASSERT(vcpu);
2323         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2324
2325         if (!is_paging(vcpu))
2326                 r = nonpaging_init_context(vcpu);
2327         else if (is_long_mode(vcpu))
2328                 r = paging64_init_context(vcpu);
2329         else if (is_pae(vcpu))
2330                 r = paging32E_init_context(vcpu);
2331         else
2332                 r = paging32_init_context(vcpu);
2333
2334         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2335
2336         return r;
2337 }
2338
2339 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2340 {
2341         vcpu->arch.update_pte.pfn = bad_pfn;
2342
2343         if (tdp_enabled)
2344                 return init_kvm_tdp_mmu(vcpu);
2345         else
2346                 return init_kvm_softmmu(vcpu);
2347 }
2348
2349 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2350 {
2351         ASSERT(vcpu);
2352         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2353                 vcpu->arch.mmu.free(vcpu);
2354                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2355         }
2356 }
2357
2358 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2359 {
2360         destroy_kvm_mmu(vcpu);
2361         return init_kvm_mmu(vcpu);
2362 }
2363 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2364
2365 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2366 {
2367         int r;
2368
2369         r = mmu_topup_memory_caches(vcpu);
2370         if (r)
2371                 goto out;
2372         spin_lock(&vcpu->kvm->mmu_lock);
2373         kvm_mmu_free_some_pages(vcpu);
2374         r = mmu_alloc_roots(vcpu);
2375         mmu_sync_roots(vcpu);
2376         spin_unlock(&vcpu->kvm->mmu_lock);
2377         if (r)
2378                 goto out;
2379         /* set_cr3() should ensure TLB has been flushed */
2380         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2381 out:
2382         return r;
2383 }
2384 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2385
2386 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2387 {
2388         mmu_free_roots(vcpu);
2389 }
2390
2391 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2392                                   struct kvm_mmu_page *sp,
2393                                   u64 *spte)
2394 {
2395         u64 pte;
2396         struct kvm_mmu_page *child;
2397
2398         pte = *spte;
2399         if (is_shadow_present_pte(pte)) {
2400                 if (is_last_spte(pte, sp->role.level))
2401                         rmap_remove(vcpu->kvm, spte);
2402                 else {
2403                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2404                         mmu_page_remove_parent_pte(child, spte);
2405                 }
2406         }
2407         __set_spte(spte, shadow_trap_nonpresent_pte);
2408         if (is_large_pte(pte))
2409                 --vcpu->kvm->stat.lpages;
2410 }
2411
2412 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2413                                   struct kvm_mmu_page *sp,
2414                                   u64 *spte,
2415                                   const void *new)
2416 {
2417         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2418                 if (!vcpu->arch.update_pte.largepage ||
2419                     sp->role.glevels == PT32_ROOT_LEVEL) {
2420                         ++vcpu->kvm->stat.mmu_pde_zapped;
2421                         return;
2422                 }
2423         }
2424
2425         ++vcpu->kvm->stat.mmu_pte_updated;
2426         if (sp->role.glevels == PT32_ROOT_LEVEL)
2427                 paging32_update_pte(vcpu, sp, spte, new);
2428         else
2429                 paging64_update_pte(vcpu, sp, spte, new);
2430 }
2431
2432 static bool need_remote_flush(u64 old, u64 new)
2433 {
2434         if (!is_shadow_present_pte(old))
2435                 return false;
2436         if (!is_shadow_present_pte(new))
2437                 return true;
2438         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2439                 return true;
2440         old ^= PT64_NX_MASK;
2441         new ^= PT64_NX_MASK;
2442         return (old & ~new & PT64_PERM_MASK) != 0;
2443 }
2444
2445 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2446 {
2447         if (need_remote_flush(old, new))
2448                 kvm_flush_remote_tlbs(vcpu->kvm);
2449         else
2450                 kvm_mmu_flush_tlb(vcpu);
2451 }
2452
2453 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2454 {
2455         u64 *spte = vcpu->arch.last_pte_updated;
2456
2457         return !!(spte && (*spte & shadow_accessed_mask));
2458 }
2459
2460 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2461                                           const u8 *new, int bytes)
2462 {
2463         gfn_t gfn;
2464         int r;
2465         u64 gpte = 0;
2466         pfn_t pfn;
2467
2468         vcpu->arch.update_pte.largepage = 0;
2469
2470         if (bytes != 4 && bytes != 8)
2471                 return;
2472
2473         /*
2474          * Assume that the pte write on a page table of the same type
2475          * as the current vcpu paging mode.  This is nearly always true
2476          * (might be false while changing modes).  Note it is verified later
2477          * by update_pte().
2478          */
2479         if (is_pae(vcpu)) {
2480                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2481                 if ((bytes == 4) && (gpa % 4 == 0)) {
2482                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2483                         if (r)
2484                                 return;
2485                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2486                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2487                         memcpy((void *)&gpte, new, 8);
2488                 }
2489         } else {
2490                 if ((bytes == 4) && (gpa % 4 == 0))
2491                         memcpy((void *)&gpte, new, 4);
2492         }
2493         if (!is_present_gpte(gpte))
2494                 return;
2495         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2496
2497         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2498                 gfn &= ~(KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL) - 1);
2499                 vcpu->arch.update_pte.largepage = 1;
2500         }
2501         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2502         smp_rmb();
2503         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2504
2505         if (is_error_pfn(pfn)) {
2506                 kvm_release_pfn_clean(pfn);
2507                 return;
2508         }
2509         vcpu->arch.update_pte.gfn = gfn;
2510         vcpu->arch.update_pte.pfn = pfn;
2511 }
2512
2513 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2514 {
2515         u64 *spte = vcpu->arch.last_pte_updated;
2516
2517         if (spte
2518             && vcpu->arch.last_pte_gfn == gfn
2519             && shadow_accessed_mask
2520             && !(*spte & shadow_accessed_mask)
2521             && is_shadow_present_pte(*spte))
2522                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2523 }
2524
2525 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2526                        const u8 *new, int bytes,
2527                        bool guest_initiated)
2528 {
2529         gfn_t gfn = gpa >> PAGE_SHIFT;
2530         struct kvm_mmu_page *sp;
2531         struct hlist_node *node, *n;
2532         struct hlist_head *bucket;
2533         unsigned index;
2534         u64 entry, gentry;
2535         u64 *spte;
2536         unsigned offset = offset_in_page(gpa);
2537         unsigned pte_size;
2538         unsigned page_offset;
2539         unsigned misaligned;
2540         unsigned quadrant;
2541         int level;
2542         int flooded = 0;
2543         int npte;
2544         int r;
2545
2546         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2547         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2548         spin_lock(&vcpu->kvm->mmu_lock);
2549         kvm_mmu_access_page(vcpu, gfn);
2550         kvm_mmu_free_some_pages(vcpu);
2551         ++vcpu->kvm->stat.mmu_pte_write;
2552         kvm_mmu_audit(vcpu, "pre pte write");
2553         if (guest_initiated) {
2554                 if (gfn == vcpu->arch.last_pt_write_gfn
2555                     && !last_updated_pte_accessed(vcpu)) {
2556                         ++vcpu->arch.last_pt_write_count;
2557                         if (vcpu->arch.last_pt_write_count >= 3)
2558                                 flooded = 1;
2559                 } else {
2560                         vcpu->arch.last_pt_write_gfn = gfn;
2561                         vcpu->arch.last_pt_write_count = 1;
2562                         vcpu->arch.last_pte_updated = NULL;
2563                 }
2564         }
2565         index = kvm_page_table_hashfn(gfn);
2566         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2567         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2568                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2569                         continue;
2570                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2571                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2572                 misaligned |= bytes < 4;
2573                 if (misaligned || flooded) {
2574                         /*
2575                          * Misaligned accesses are too much trouble to fix
2576                          * up; also, they usually indicate a page is not used
2577                          * as a page table.
2578                          *
2579                          * If we're seeing too many writes to a page,
2580                          * it may no longer be a page table, or we may be
2581                          * forking, in which case it is better to unmap the
2582                          * page.
2583                          */
2584                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2585                                  gpa, bytes, sp->role.word);
2586                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2587                                 n = bucket->first;
2588                         ++vcpu->kvm->stat.mmu_flooded;
2589                         continue;
2590                 }
2591                 page_offset = offset;
2592                 level = sp->role.level;
2593                 npte = 1;
2594                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2595                         page_offset <<= 1;      /* 32->64 */
2596                         /*
2597                          * A 32-bit pde maps 4MB while the shadow pdes map
2598                          * only 2MB.  So we need to double the offset again
2599                          * and zap two pdes instead of one.
2600                          */
2601                         if (level == PT32_ROOT_LEVEL) {
2602                                 page_offset &= ~7; /* kill rounding error */
2603                                 page_offset <<= 1;
2604                                 npte = 2;
2605                         }
2606                         quadrant = page_offset >> PAGE_SHIFT;
2607                         page_offset &= ~PAGE_MASK;
2608                         if (quadrant != sp->role.quadrant)
2609                                 continue;
2610                 }
2611                 spte = &sp->spt[page_offset / sizeof(*spte)];
2612                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2613                         gentry = 0;
2614                         r = kvm_read_guest_atomic(vcpu->kvm,
2615                                                   gpa & ~(u64)(pte_size - 1),
2616                                                   &gentry, pte_size);
2617                         new = (const void *)&gentry;
2618                         if (r < 0)
2619                                 new = NULL;
2620                 }
2621                 while (npte--) {
2622                         entry = *spte;
2623                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2624                         if (new)
2625                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2626                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2627                         ++spte;
2628                 }
2629         }
2630         kvm_mmu_audit(vcpu, "post pte write");
2631         spin_unlock(&vcpu->kvm->mmu_lock);
2632         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2633                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2634                 vcpu->arch.update_pte.pfn = bad_pfn;
2635         }
2636 }
2637
2638 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2639 {
2640         gpa_t gpa;
2641         int r;
2642
2643         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2644
2645         spin_lock(&vcpu->kvm->mmu_lock);
2646         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2647         spin_unlock(&vcpu->kvm->mmu_lock);
2648         return r;
2649 }
2650 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2651
2652 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2653 {
2654         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2655                 struct kvm_mmu_page *sp;
2656
2657                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2658                                   struct kvm_mmu_page, link);
2659                 kvm_mmu_zap_page(vcpu->kvm, sp);
2660                 ++vcpu->kvm->stat.mmu_recycled;
2661         }
2662 }
2663
2664 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2665 {
2666         int r;
2667         enum emulation_result er;
2668
2669         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2670         if (r < 0)
2671                 goto out;
2672
2673         if (!r) {
2674                 r = 1;
2675                 goto out;
2676         }
2677
2678         r = mmu_topup_memory_caches(vcpu);
2679         if (r)
2680                 goto out;
2681
2682         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2683
2684         switch (er) {
2685         case EMULATE_DONE:
2686                 return 1;
2687         case EMULATE_DO_MMIO:
2688                 ++vcpu->stat.mmio_exits;
2689                 return 0;
2690         case EMULATE_FAIL:
2691                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2692                 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2693                 return 0;
2694         default:
2695                 BUG();
2696         }
2697 out:
2698         return r;
2699 }
2700 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2701
2702 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2703 {
2704         vcpu->arch.mmu.invlpg(vcpu, gva);
2705         kvm_mmu_flush_tlb(vcpu);
2706         ++vcpu->stat.invlpg;
2707 }
2708 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2709
2710 void kvm_enable_tdp(void)
2711 {
2712         tdp_enabled = true;
2713 }
2714 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2715
2716 void kvm_disable_tdp(void)
2717 {
2718         tdp_enabled = false;
2719 }
2720 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2721
2722 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2723 {
2724         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2725 }
2726
2727 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2728 {
2729         struct page *page;
2730         int i;
2731
2732         ASSERT(vcpu);
2733
2734         spin_lock(&vcpu->kvm->mmu_lock);
2735         if (vcpu->kvm->arch.n_requested_mmu_pages)
2736                 vcpu->kvm->arch.n_free_mmu_pages =
2737                                         vcpu->kvm->arch.n_requested_mmu_pages;
2738         else
2739                 vcpu->kvm->arch.n_free_mmu_pages =
2740                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2741         spin_unlock(&vcpu->kvm->mmu_lock);
2742         /*
2743          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2744          * Therefore we need to allocate shadow page tables in the first
2745          * 4GB of memory, which happens to fit the DMA32 zone.
2746          */
2747         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2748         if (!page)
2749                 goto error_1;
2750         vcpu->arch.mmu.pae_root = page_address(page);
2751         for (i = 0; i < 4; ++i)
2752                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2753
2754         return 0;
2755
2756 error_1:
2757         free_mmu_pages(vcpu);
2758         return -ENOMEM;
2759 }
2760
2761 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2762 {
2763         ASSERT(vcpu);
2764         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2765
2766         return alloc_mmu_pages(vcpu);
2767 }
2768
2769 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2770 {
2771         ASSERT(vcpu);
2772         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2773
2774         return init_kvm_mmu(vcpu);
2775 }
2776
2777 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2778 {
2779         ASSERT(vcpu);
2780
2781         destroy_kvm_mmu(vcpu);
2782         free_mmu_pages(vcpu);
2783         mmu_free_memory_caches(vcpu);
2784 }
2785
2786 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2787 {
2788         struct kvm_mmu_page *sp;
2789
2790         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2791                 int i;
2792                 u64 *pt;
2793
2794                 if (!test_bit(slot, sp->slot_bitmap))
2795                         continue;
2796
2797                 pt = sp->spt;
2798                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2799                         /* avoid RMW */
2800                         if (pt[i] & PT_WRITABLE_MASK)
2801                                 pt[i] &= ~PT_WRITABLE_MASK;
2802         }
2803         kvm_flush_remote_tlbs(kvm);
2804 }
2805
2806 void kvm_mmu_zap_all(struct kvm *kvm)
2807 {
2808         struct kvm_mmu_page *sp, *node;
2809
2810         spin_lock(&kvm->mmu_lock);
2811         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2812                 if (kvm_mmu_zap_page(kvm, sp))
2813                         node = container_of(kvm->arch.active_mmu_pages.next,
2814                                             struct kvm_mmu_page, link);
2815         spin_unlock(&kvm->mmu_lock);
2816
2817         kvm_flush_remote_tlbs(kvm);
2818 }
2819
2820 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2821 {
2822         struct kvm_mmu_page *page;
2823
2824         page = container_of(kvm->arch.active_mmu_pages.prev,
2825                             struct kvm_mmu_page, link);
2826         kvm_mmu_zap_page(kvm, page);
2827 }
2828
2829 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2830 {
2831         struct kvm *kvm;
2832         struct kvm *kvm_freed = NULL;
2833         int cache_count = 0;
2834
2835         spin_lock(&kvm_lock);
2836
2837         list_for_each_entry(kvm, &vm_list, vm_list) {
2838                 int npages;
2839
2840                 if (!down_read_trylock(&kvm->slots_lock))
2841                         continue;
2842                 spin_lock(&kvm->mmu_lock);
2843                 npages = kvm->arch.n_alloc_mmu_pages -
2844                          kvm->arch.n_free_mmu_pages;
2845                 cache_count += npages;
2846                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2847                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2848                         cache_count--;
2849                         kvm_freed = kvm;
2850                 }
2851                 nr_to_scan--;
2852
2853                 spin_unlock(&kvm->mmu_lock);
2854                 up_read(&kvm->slots_lock);
2855         }
2856         if (kvm_freed)
2857                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2858
2859         spin_unlock(&kvm_lock);
2860
2861         return cache_count;
2862 }
2863
2864 static struct shrinker mmu_shrinker = {
2865         .shrink = mmu_shrink,
2866         .seeks = DEFAULT_SEEKS * 10,
2867 };
2868
2869 static void mmu_destroy_caches(void)
2870 {
2871         if (pte_chain_cache)
2872                 kmem_cache_destroy(pte_chain_cache);
2873         if (rmap_desc_cache)
2874                 kmem_cache_destroy(rmap_desc_cache);
2875         if (mmu_page_header_cache)
2876                 kmem_cache_destroy(mmu_page_header_cache);
2877 }
2878
2879 void kvm_mmu_module_exit(void)
2880 {
2881         mmu_destroy_caches();
2882         unregister_shrinker(&mmu_shrinker);
2883 }
2884
2885 int kvm_mmu_module_init(void)
2886 {
2887         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2888                                             sizeof(struct kvm_pte_chain),
2889                                             0, 0, NULL);
2890         if (!pte_chain_cache)
2891                 goto nomem;
2892         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2893                                             sizeof(struct kvm_rmap_desc),
2894                                             0, 0, NULL);
2895         if (!rmap_desc_cache)
2896                 goto nomem;
2897
2898         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2899                                                   sizeof(struct kvm_mmu_page),
2900                                                   0, 0, NULL);
2901         if (!mmu_page_header_cache)
2902                 goto nomem;
2903
2904         register_shrinker(&mmu_shrinker);
2905
2906         return 0;
2907
2908 nomem:
2909         mmu_destroy_caches();
2910         return -ENOMEM;
2911 }
2912
2913 /*
2914  * Caculate mmu pages needed for kvm.
2915  */
2916 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2917 {
2918         int i;
2919         unsigned int nr_mmu_pages;
2920         unsigned int  nr_pages = 0;
2921
2922         for (i = 0; i < kvm->nmemslots; i++)
2923                 nr_pages += kvm->memslots[i].npages;
2924
2925         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2926         nr_mmu_pages = max(nr_mmu_pages,
2927                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2928
2929         return nr_mmu_pages;
2930 }
2931
2932 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2933                                 unsigned len)
2934 {
2935         if (len > buffer->len)
2936                 return NULL;
2937         return buffer->ptr;
2938 }
2939
2940 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2941                                 unsigned len)
2942 {
2943         void *ret;
2944
2945         ret = pv_mmu_peek_buffer(buffer, len);
2946         if (!ret)
2947                 return ret;
2948         buffer->ptr += len;
2949         buffer->len -= len;
2950         buffer->processed += len;
2951         return ret;
2952 }
2953
2954 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2955                              gpa_t addr, gpa_t value)
2956 {
2957         int bytes = 8;
2958         int r;
2959
2960         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2961                 bytes = 4;
2962
2963         r = mmu_topup_memory_caches(vcpu);
2964         if (r)
2965                 return r;
2966
2967         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2968                 return -EFAULT;
2969
2970         return 1;
2971 }
2972
2973 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2974 {
2975         kvm_set_cr3(vcpu, vcpu->arch.cr3);
2976         return 1;
2977 }
2978
2979 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2980 {
2981         spin_lock(&vcpu->kvm->mmu_lock);
2982         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2983         spin_unlock(&vcpu->kvm->mmu_lock);
2984         return 1;
2985 }
2986
2987 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2988                              struct kvm_pv_mmu_op_buffer *buffer)
2989 {
2990         struct kvm_mmu_op_header *header;
2991
2992         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2993         if (!header)
2994                 return 0;
2995         switch (header->op) {
2996         case KVM_MMU_OP_WRITE_PTE: {
2997                 struct kvm_mmu_op_write_pte *wpte;
2998
2999                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3000                 if (!wpte)
3001                         return 0;
3002                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3003                                         wpte->pte_val);
3004         }
3005         case KVM_MMU_OP_FLUSH_TLB: {
3006                 struct kvm_mmu_op_flush_tlb *ftlb;
3007
3008                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3009                 if (!ftlb)
3010                         return 0;
3011                 return kvm_pv_mmu_flush_tlb(vcpu);
3012         }
3013         case KVM_MMU_OP_RELEASE_PT: {
3014                 struct kvm_mmu_op_release_pt *rpt;
3015
3016                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3017                 if (!rpt)
3018                         return 0;
3019                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3020         }
3021         default: return 0;
3022         }
3023 }
3024
3025 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3026                   gpa_t addr, unsigned long *ret)
3027 {
3028         int r;
3029         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3030
3031         buffer->ptr = buffer->buf;
3032         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3033         buffer->processed = 0;
3034
3035         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3036         if (r)
3037                 goto out;
3038
3039         while (buffer->len) {
3040                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3041                 if (r < 0)
3042                         goto out;
3043                 if (r == 0)
3044                         break;
3045         }
3046
3047         r = 1;
3048 out:
3049         *ret = buffer->processed;
3050         return r;
3051 }
3052
3053 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3054 {
3055         struct kvm_shadow_walk_iterator iterator;
3056         int nr_sptes = 0;
3057
3058         spin_lock(&vcpu->kvm->mmu_lock);
3059         for_each_shadow_entry(vcpu, addr, iterator) {
3060                 sptes[iterator.level-1] = *iterator.sptep;
3061                 nr_sptes++;
3062                 if (!is_shadow_present_pte(*iterator.sptep))
3063                         break;
3064         }
3065         spin_unlock(&vcpu->kvm->mmu_lock);
3066
3067         return nr_sptes;
3068 }
3069 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3070
3071 #ifdef AUDIT
3072
3073 static const char *audit_msg;
3074
3075 static gva_t canonicalize(gva_t gva)
3076 {
3077 #ifdef CONFIG_X86_64
3078         gva = (long long)(gva << 16) >> 16;
3079 #endif
3080         return gva;
3081 }
3082
3083
3084 typedef void (*inspect_spte_fn) (struct kvm *kvm, struct kvm_mmu_page *sp,
3085                                  u64 *sptep);
3086
3087 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3088                             inspect_spte_fn fn)
3089 {
3090         int i;
3091
3092         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3093                 u64 ent = sp->spt[i];
3094
3095                 if (is_shadow_present_pte(ent)) {
3096                         if (!is_last_spte(ent, sp->role.level)) {
3097                                 struct kvm_mmu_page *child;
3098                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3099                                 __mmu_spte_walk(kvm, child, fn);
3100                         } else
3101                                 fn(kvm, sp, &sp->spt[i]);
3102                 }
3103         }
3104 }
3105
3106 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3107 {
3108         int i;
3109         struct kvm_mmu_page *sp;
3110
3111         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3112                 return;
3113         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3114                 hpa_t root = vcpu->arch.mmu.root_hpa;
3115                 sp = page_header(root);
3116                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3117                 return;
3118         }
3119         for (i = 0; i < 4; ++i) {
3120                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3121
3122                 if (root && VALID_PAGE(root)) {
3123                         root &= PT64_BASE_ADDR_MASK;
3124                         sp = page_header(root);
3125                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3126                 }
3127         }
3128         return;
3129 }
3130
3131 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3132                                 gva_t va, int level)
3133 {
3134         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3135         int i;
3136         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3137
3138         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3139                 u64 ent = pt[i];
3140
3141                 if (ent == shadow_trap_nonpresent_pte)
3142                         continue;
3143
3144                 va = canonicalize(va);
3145                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3146                         audit_mappings_page(vcpu, ent, va, level - 1);
3147                 else {
3148                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3149                         gfn_t gfn = gpa >> PAGE_SHIFT;
3150                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3151                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3152
3153                         if (is_error_pfn(pfn)) {
3154                                 kvm_release_pfn_clean(pfn);
3155                                 continue;
3156                         }
3157
3158                         if (is_shadow_present_pte(ent)
3159                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3160                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3161                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3162                                        audit_msg, vcpu->arch.mmu.root_level,
3163                                        va, gpa, hpa, ent,
3164                                        is_shadow_present_pte(ent));
3165                         else if (ent == shadow_notrap_nonpresent_pte
3166                                  && !is_error_hpa(hpa))
3167                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3168                                        " valid guest gva %lx\n", audit_msg, va);
3169                         kvm_release_pfn_clean(pfn);
3170
3171                 }
3172         }
3173 }
3174
3175 static void audit_mappings(struct kvm_vcpu *vcpu)
3176 {
3177         unsigned i;
3178
3179         if (vcpu->arch.mmu.root_level == 4)
3180                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3181         else
3182                 for (i = 0; i < 4; ++i)
3183                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3184                                 audit_mappings_page(vcpu,
3185                                                     vcpu->arch.mmu.pae_root[i],
3186                                                     i << 30,
3187                                                     2);
3188 }
3189
3190 static int count_rmaps(struct kvm_vcpu *vcpu)
3191 {
3192         int nmaps = 0;
3193         int i, j, k;
3194
3195         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3196                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3197                 struct kvm_rmap_desc *d;
3198
3199                 for (j = 0; j < m->npages; ++j) {
3200                         unsigned long *rmapp = &m->rmap[j];
3201
3202                         if (!*rmapp)
3203                                 continue;
3204                         if (!(*rmapp & 1)) {
3205                                 ++nmaps;
3206                                 continue;
3207                         }
3208                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3209                         while (d) {
3210                                 for (k = 0; k < RMAP_EXT; ++k)
3211                                         if (d->sptes[k])
3212                                                 ++nmaps;
3213                                         else
3214                                                 break;
3215                                 d = d->more;
3216                         }
3217                 }
3218         }
3219         return nmaps;
3220 }
3221
3222 void inspect_spte_has_rmap(struct kvm *kvm, struct kvm_mmu_page *sp, u64 *sptep)
3223 {
3224         unsigned long *rmapp;
3225         struct kvm_mmu_page *rev_sp;
3226         gfn_t gfn;
3227
3228         if (*sptep & PT_WRITABLE_MASK) {
3229                 rev_sp = page_header(__pa(sptep));
3230                 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3231
3232                 if (!gfn_to_memslot(kvm, gfn)) {
3233                         if (!printk_ratelimit())
3234                                 return;
3235                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3236                                          audit_msg, gfn);
3237                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3238                                         audit_msg, sptep - rev_sp->spt,
3239                                         rev_sp->gfn);
3240                         dump_stack();
3241                         return;
3242                 }
3243
3244                 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3245                                     is_large_pte(*sptep));
3246                 if (!*rmapp) {
3247                         if (!printk_ratelimit())
3248                                 return;
3249                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3250                                          audit_msg, *sptep);
3251                         dump_stack();
3252                 }
3253         }
3254
3255 }
3256
3257 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3258 {
3259         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3260 }
3261
3262 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3263 {
3264         struct kvm_mmu_page *sp;
3265         int i;
3266
3267         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3268                 u64 *pt = sp->spt;
3269
3270                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3271                         continue;
3272
3273                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3274                         u64 ent = pt[i];
3275
3276                         if (!(ent & PT_PRESENT_MASK))
3277                                 continue;
3278                         if (!(ent & PT_WRITABLE_MASK))
3279                                 continue;
3280                         inspect_spte_has_rmap(vcpu->kvm, sp, &pt[i]);
3281                 }
3282         }
3283         return;
3284 }
3285
3286 static void audit_rmap(struct kvm_vcpu *vcpu)
3287 {
3288         check_writable_mappings_rmap(vcpu);
3289         count_rmaps(vcpu);
3290 }
3291
3292 static void audit_write_protection(struct kvm_vcpu *vcpu)
3293 {
3294         struct kvm_mmu_page *sp;
3295         struct kvm_memory_slot *slot;
3296         unsigned long *rmapp;
3297         u64 *spte;
3298         gfn_t gfn;
3299
3300         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3301                 if (sp->role.direct)
3302                         continue;
3303                 if (sp->unsync)
3304                         continue;
3305
3306                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3307                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3308                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3309
3310                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3311                 while (spte) {
3312                         if (*spte & PT_WRITABLE_MASK)
3313                                 printk(KERN_ERR "%s: (%s) shadow page has "
3314                                 "writable mappings: gfn %lx role %x\n",
3315                                __func__, audit_msg, sp->gfn,
3316                                sp->role.word);
3317                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3318                 }
3319         }
3320 }
3321
3322 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3323 {
3324         int olddbg = dbg;
3325
3326         dbg = 0;
3327         audit_msg = msg;
3328         audit_rmap(vcpu);
3329         audit_write_protection(vcpu);
3330         if (strcmp("pre pte write", audit_msg) != 0)
3331                 audit_mappings(vcpu);
3332         audit_writable_sptes_have_rmaps(vcpu);
3333         dbg = olddbg;
3334 }
3335
3336 #endif