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  * Copyright 2010 Red Hat, Inc. and/or its affilates.
  11  *
  12  * Authors:
  13  *   Yaniv Kamay  <yaniv@qumranet.com>
  14  *   Avi Kivity   <avi@qumranet.com>
  15  *
  16  * This work is licensed under the terms of the GNU GPL, version 2.  See
  17  * the COPYING file in the top-level directory.
  18  *
  19  */
  20
  21 #include "mmu.h"
  22 #include "x86.h"
  23 #include "kvm_cache_regs.h"
  24
  25 #include <linux/kvm_host.h>
  26 #include <linux/types.h>
  27 #include <linux/string.h>
  28 #include <linux/mm.h>
  29 #include <linux/highmem.h>
  30 #include <linux/module.h>
  31 #include <linux/swap.h>
  32 #include <linux/hugetlb.h>
  33 #include <linux/compiler.h>
  34 #include <linux/srcu.h>
  35 #include <linux/slab.h>
  36 #include <linux/uaccess.h>
  37
  38 #include <asm/page.h>
  39 #include <asm/cmpxchg.h>
  40 #include <asm/io.h>
  41 #include <asm/vmx.h>
  42
  43 /*
  44  * When setting this variable to true it enables Two-Dimensional-Paging
  45  * where the hardware walks 2 page tables:
  46  * 1. the guest-virtual to guest-physical
  47  * 2. while doing 1. it walks guest-physical to host-physical
  48  * If the hardware supports that we don't need to do shadow paging.
  49  */
  50 bool tdp_enabled = false;
  51
  52 #undef MMU_DEBUG
  53
  54 #undef AUDIT
  55
  56 #ifdef AUDIT
  57 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  58 #else
  59 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  60 #endif
  61
  62 #ifdef MMU_DEBUG
  63
  64 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  65 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  66
  67 #else
  68
  69 #define pgprintk(x...) do { } while (0)
  70 #define rmap_printk(x...) do { } while (0)
  71
  72 #endif
  73
  74 #if defined(MMU_DEBUG) || defined(AUDIT)
  75 static int dbg = 0;
  76 module_param(dbg, bool, 0644);
  77 #endif
  78
  79 static int oos_shadow = 1;
  80 module_param(oos_shadow, bool, 0644);
  81
  82 #ifndef MMU_DEBUG
  83 #define ASSERT(x) do { } while (0)
  84 #else
  85 #define ASSERT(x)                                                       \
  86         if (!(x)) {                                                     \
  87                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  88                        __FILE__, __LINE__, #x);                         \
  89         }
  90 #endif
  91
  92 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  93 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  94
  95 #define PT64_LEVEL_BITS 9
  96
  97 #define PT64_LEVEL_SHIFT(level) \
  98                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  99
 100 #define PT64_LEVEL_MASK(level) \
 101                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 102
 103 #define PT64_INDEX(address, level)\
 104         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 105
 106
 107 #define PT32_LEVEL_BITS 10
 108
 109 #define PT32_LEVEL_SHIFT(level) \
 110                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 111
 112 #define PT32_LEVEL_MASK(level) \
 113                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 114 #define PT32_LVL_OFFSET_MASK(level) \
 115         (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 116                                                 * PT32_LEVEL_BITS))) - 1))
 117
 118 #define PT32_INDEX(address, level)\
 119         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 120
 121
 122 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 123 #define PT64_DIR_BASE_ADDR_MASK \
 124         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 125 #define PT64_LVL_ADDR_MASK(level) \
 126         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 127                                                 * PT64_LEVEL_BITS))) - 1))
 128 #define PT64_LVL_OFFSET_MASK(level) \
 129         (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
 130                                                 * PT64_LEVEL_BITS))) - 1))
 131
 132 #define PT32_BASE_ADDR_MASK PAGE_MASK
 133 #define PT32_DIR_BASE_ADDR_MASK \
 134         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 135 #define PT32_LVL_ADDR_MASK(level) \
 136         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
 137                                             * PT32_LEVEL_BITS))) - 1))
 138
 139 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 140                         | PT64_NX_MASK)
 141
 142 #define RMAP_EXT 4
 143
 144 #define ACC_EXEC_MASK    1
 145 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 146 #define ACC_USER_MASK    PT_USER_MASK
 147 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 148
 149 #include <trace/events/kvm.h>
 150
 151 #define CREATE_TRACE_POINTS
 152 #include "mmutrace.h"
 153
 154 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 155
 156 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 157
 158 struct kvm_rmap_desc {
 159         u64 *sptes[RMAP_EXT];
 160         struct kvm_rmap_desc *more;
 161 };
 162
 163 struct kvm_shadow_walk_iterator {
 164         u64 addr;
 165         hpa_t shadow_addr;
 166         int level;
 167         u64 *sptep;
 168         unsigned index;
 169 };
 170
 171 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
 172         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
 173              shadow_walk_okay(&(_walker));                      \
 174              shadow_walk_next(&(_walker)))
 175
 176 typedef void (*mmu_parent_walk_fn) (struct kvm_mmu_page *sp, u64 *spte);
 177
 178 static struct kmem_cache *pte_chain_cache;
 179 static struct kmem_cache *rmap_desc_cache;
 180 static struct kmem_cache *mmu_page_header_cache;
 181
 182 static u64 __read_mostly shadow_trap_nonpresent_pte;
 183 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 184 static u64 __read_mostly shadow_base_present_pte;
 185 static u64 __read_mostly shadow_nx_mask;
 186 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 187 static u64 __read_mostly shadow_user_mask;
 188 static u64 __read_mostly shadow_accessed_mask;
 189 static u64 __read_mostly shadow_dirty_mask;
 190
 191 static inline u64 rsvd_bits(int s, int e)
 192 {
 193         return ((1ULL << (e - s + 1)) - 1) << s;
 194 }
 195
 196 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 197 {
 198         shadow_trap_nonpresent_pte = trap_pte;
 199         shadow_notrap_nonpresent_pte = notrap_pte;
 200 }
 201 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 202
 203 void kvm_mmu_set_base_ptes(u64 base_pte)
 204 {
 205         shadow_base_present_pte = base_pte;
 206 }
 207 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 208
 209 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 210                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 211 {
 212         shadow_user_mask = user_mask;
 213         shadow_accessed_mask = accessed_mask;
 214         shadow_dirty_mask = dirty_mask;
 215         shadow_nx_mask = nx_mask;
 216         shadow_x_mask = x_mask;
 217 }
 218 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 219
 220 static bool is_write_protection(struct kvm_vcpu *vcpu)
 221 {
 222         return kvm_read_cr0_bits(vcpu, X86_CR0_WP);
 223 }
 224
 225 static int is_cpuid_PSE36(void)
 226 {
 227         return 1;
 228 }
 229
 230 static int is_nx(struct kvm_vcpu *vcpu)
 231 {
 232         return vcpu->arch.efer & EFER_NX;
 233 }
 234
 235 static int is_shadow_present_pte(u64 pte)
 236 {
 237         return pte != shadow_trap_nonpresent_pte
 238                 && pte != shadow_notrap_nonpresent_pte;
 239 }
 240
 241 static int is_large_pte(u64 pte)
 242 {
 243         return pte & PT_PAGE_SIZE_MASK;
 244 }
 245
 246 static int is_writable_pte(unsigned long pte)
 247 {
 248         return pte & PT_WRITABLE_MASK;
 249 }
 250
 251 static int is_dirty_gpte(unsigned long pte)
 252 {
 253         return pte & PT_DIRTY_MASK;
 254 }
 255
 256 static int is_rmap_spte(u64 pte)
 257 {
 258         return is_shadow_present_pte(pte);
 259 }
 260
 261 static int is_last_spte(u64 pte, int level)
 262 {
 263         if (level == PT_PAGE_TABLE_LEVEL)
 264                 return 1;
 265         if (is_large_pte(pte))
 266                 return 1;
 267         return 0;
 268 }
 269
 270 static pfn_t spte_to_pfn(u64 pte)
 271 {
 272         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 273 }
 274
 275 static gfn_t pse36_gfn_delta(u32 gpte)
 276 {
 277         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 278
 279         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 280 }
 281
 282 static void __set_spte(u64 *sptep, u64 spte)
 283 {
 284 #ifdef CONFIG_X86_64
 285         set_64bit((unsigned long *)sptep, spte);
 286 #else
 287         set_64bit((unsigned long long *)sptep, spte);
 288 #endif
 289 }
 290
 291 static u64 __xchg_spte(u64 *sptep, u64 new_spte)
 292 {
 293 #ifdef CONFIG_X86_64
 294         return xchg(sptep, new_spte);
 295 #else
 296         u64 old_spte;
 297
 298         do {
 299                 old_spte = *sptep;
 300         } while (cmpxchg64(sptep, old_spte, new_spte) != old_spte);
 301
 302         return old_spte;
 303 #endif
 304 }
 305
 306 static void update_spte(u64 *sptep, u64 new_spte)
 307 {
 308         u64 old_spte;
 309
 310         if (!shadow_accessed_mask || (new_spte & shadow_accessed_mask)) {
 311                 __set_spte(sptep, new_spte);
 312         } else {
 313                 old_spte = __xchg_spte(sptep, new_spte);
 314                 if (old_spte & shadow_accessed_mask)
 315                         mark_page_accessed(pfn_to_page(spte_to_pfn(old_spte)));
 316         }
 317 }
 318
 319 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 320                                   struct kmem_cache *base_cache, int min)
 321 {
 322         void *obj;
 323
 324         if (cache->nobjs >= min)
 325                 return 0;
 326         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 327                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 328                 if (!obj)
 329                         return -ENOMEM;
 330                 cache->objects[cache->nobjs++] = obj;
 331         }
 332         return 0;
 333 }
 334
 335 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
 336                                   struct kmem_cache *cache)
 337 {
 338         while (mc->nobjs)
 339                 kmem_cache_free(cache, mc->objects[--mc->nobjs]);
 340 }
 341
 342 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 343                                        int min)
 344 {
 345         struct page *page;
 346
 347         if (cache->nobjs >= min)
 348                 return 0;
 349         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 350                 page = alloc_page(GFP_KERNEL);
 351                 if (!page)
 352                         return -ENOMEM;
 353                 cache->objects[cache->nobjs++] = page_address(page);
 354         }
 355         return 0;
 356 }
 357
 358 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 359 {
 360         while (mc->nobjs)
 361                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 362 }
 363
 364 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 365 {
 366         int r;
 367
 368         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 369                                    pte_chain_cache, 4);
 370         if (r)
 371                 goto out;
 372         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 373                                    rmap_desc_cache, 4);
 374         if (r)
 375                 goto out;
 376         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 377         if (r)
 378                 goto out;
 379         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 380                                    mmu_page_header_cache, 4);
 381 out:
 382         return r;
 383 }
 384
 385 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 386 {
 387         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache, pte_chain_cache);
 388         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache, rmap_desc_cache);
 389         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 390         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
 391                                 mmu_page_header_cache);
 392 }
 393
 394 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 395                                     size_t size)
 396 {
 397         void *p;
 398
 399         BUG_ON(!mc->nobjs);
 400         p = mc->objects[--mc->nobjs];
 401         return p;
 402 }
 403
 404 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 405 {
 406         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 407                                       sizeof(struct kvm_pte_chain));
 408 }
 409
 410 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 411 {
 412         kmem_cache_free(pte_chain_cache, pc);
 413 }
 414
 415 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 416 {
 417         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 418                                       sizeof(struct kvm_rmap_desc));
 419 }
 420
 421 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 422 {
 423         kmem_cache_free(rmap_desc_cache, rd);
 424 }
 425
 426 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
 427 {
 428         if (!sp->role.direct)
 429                 return sp->gfns[index];
 430
 431         return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
 432 }
 433
 434 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
 435 {
 436         if (sp->role.direct)
 437                 BUG_ON(gfn != kvm_mmu_page_get_gfn(sp, index));
 438         else
 439                 sp->gfns[index] = gfn;
 440 }
 441
 442 /*
 443  * Return the pointer to the largepage write count for a given
 444  * gfn, handling slots that are not large page aligned.
 445  */
 446 static int *slot_largepage_idx(gfn_t gfn,
 447                                struct kvm_memory_slot *slot,
 448                                int level)
 449 {
 450         unsigned long idx;
 451
 452         idx = (gfn >> KVM_HPAGE_GFN_SHIFT(level)) -
 453               (slot->base_gfn >> KVM_HPAGE_GFN_SHIFT(level));
 454         return &slot->lpage_info[level - 2][idx].write_count;
 455 }
 456
 457 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 458 {
 459         struct kvm_memory_slot *slot;
 460         int *write_count;
 461         int i;
 462
 463         slot = gfn_to_memslot(kvm, gfn);
 464         for (i = PT_DIRECTORY_LEVEL;
 465              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 466                 write_count   = slot_largepage_idx(gfn, slot, i);
 467                 *write_count += 1;
 468         }
 469 }
 470
 471 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 472 {
 473         struct kvm_memory_slot *slot;
 474         int *write_count;
 475         int i;
 476
 477         slot = gfn_to_memslot(kvm, gfn);
 478         for (i = PT_DIRECTORY_LEVEL;
 479              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 480                 write_count   = slot_largepage_idx(gfn, slot, i);
 481                 *write_count -= 1;
 482                 WARN_ON(*write_count < 0);
 483         }
 484 }
 485
 486 static int has_wrprotected_page(struct kvm *kvm,
 487                                 gfn_t gfn,
 488                                 int level)
 489 {
 490         struct kvm_memory_slot *slot;
 491         int *largepage_idx;
 492
 493         slot = gfn_to_memslot(kvm, gfn);
 494         if (slot) {
 495                 largepage_idx = slot_largepage_idx(gfn, slot, level);
 496                 return *largepage_idx;
 497         }
 498
 499         return 1;
 500 }
 501
 502 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
 503 {
 504         unsigned long page_size;
 505         int i, ret = 0;
 506
 507         page_size = kvm_host_page_size(kvm, gfn);
 508
 509         for (i = PT_PAGE_TABLE_LEVEL;
 510              i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
 511                 if (page_size >= KVM_HPAGE_SIZE(i))
 512                         ret = i;
 513                 else
 514                         break;
 515         }
 516
 517         return ret;
 518 }
 519
 520 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 521 {
 522         struct kvm_memory_slot *slot;
 523         int host_level, level, max_level;
 524
 525         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 526         if (slot && slot->dirty_bitmap)
 527                 return PT_PAGE_TABLE_LEVEL;
 528
 529         host_level = host_mapping_level(vcpu->kvm, large_gfn);
 530
 531         if (host_level == PT_PAGE_TABLE_LEVEL)
 532                 return host_level;
 533
 534         max_level = kvm_x86_ops->get_lpage_level() < host_level ?
 535                 kvm_x86_ops->get_lpage_level() : host_level;
 536
 537         for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
 538                 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
 539                         break;
 540
 541         return level - 1;
 542 }
 543
 544 /*
 545  * Take gfn and return the reverse mapping to it.
 546  */
 547
 548 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
 549 {
 550         struct kvm_memory_slot *slot;
 551         unsigned long idx;
 552
 553         slot = gfn_to_memslot(kvm, gfn);
 554         if (likely(level == PT_PAGE_TABLE_LEVEL))
 555                 return &slot->rmap[gfn - slot->base_gfn];
 556
 557         idx = (gfn >> KVM_HPAGE_GFN_SHIFT(level)) -
 558                 (slot->base_gfn >> KVM_HPAGE_GFN_SHIFT(level));
 559
 560         return &slot->lpage_info[level - 2][idx].rmap_pde;
 561 }
 562
 563 /*
 564  * Reverse mapping data structures:
 565  *
 566  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 567  * that points to page_address(page).
 568  *
 569  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 570  * containing more mappings.
 571  *
 572  * Returns the number of rmap entries before the spte was added or zero if
 573  * the spte was not added.
 574  *
 575  */
 576 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 577 {
 578         struct kvm_mmu_page *sp;
 579         struct kvm_rmap_desc *desc;
 580         unsigned long *rmapp;
 581         int i, count = 0;
 582
 583         if (!is_rmap_spte(*spte))
 584                 return count;
 585         sp = page_header(__pa(spte));
 586         kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
 587         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 588         if (!*rmapp) {
 589                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 590                 *rmapp = (unsigned long)spte;
 591         } else if (!(*rmapp & 1)) {
 592                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 593                 desc = mmu_alloc_rmap_desc(vcpu);
 594                 desc->sptes[0] = (u64 *)*rmapp;
 595                 desc->sptes[1] = spte;
 596                 *rmapp = (unsigned long)desc | 1;
 597         } else {
 598                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 599                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 600                 while (desc->sptes[RMAP_EXT-1] && desc->more) {
 601                         desc = desc->more;
 602                         count += RMAP_EXT;
 603                 }
 604                 if (desc->sptes[RMAP_EXT-1]) {
 605                         desc->more = mmu_alloc_rmap_desc(vcpu);
 606                         desc = desc->more;
 607                 }
 608                 for (i = 0; desc->sptes[i]; ++i)
 609                         ;
 610                 desc->sptes[i] = spte;
 611         }
 612         return count;
 613 }
 614
 615 static void rmap_desc_remove_entry(unsigned long *rmapp,
 616                                    struct kvm_rmap_desc *desc,
 617                                    int i,
 618                                    struct kvm_rmap_desc *prev_desc)
 619 {
 620         int j;
 621
 622         for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
 623                 ;
 624         desc->sptes[i] = desc->sptes[j];
 625         desc->sptes[j] = NULL;
 626         if (j != 0)
 627                 return;
 628         if (!prev_desc && !desc->more)
 629                 *rmapp = (unsigned long)desc->sptes[0];
 630         else
 631                 if (prev_desc)
 632                         prev_desc->more = desc->more;
 633                 else
 634                         *rmapp = (unsigned long)desc->more | 1;
 635         mmu_free_rmap_desc(desc);
 636 }
 637
 638 static void rmap_remove(struct kvm *kvm, u64 *spte)
 639 {
 640         struct kvm_rmap_desc *desc;
 641         struct kvm_rmap_desc *prev_desc;
 642         struct kvm_mmu_page *sp;
 643         gfn_t gfn;
 644         unsigned long *rmapp;
 645         int i;
 646
 647         sp = page_header(__pa(spte));
 648         gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
 649         rmapp = gfn_to_rmap(kvm, gfn, sp->role.level);
 650         if (!*rmapp) {
 651                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 652                 BUG();
 653         } else if (!(*rmapp & 1)) {
 654                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 655                 if ((u64 *)*rmapp != spte) {
 656                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 657                                spte, *spte);
 658                         BUG();
 659                 }
 660                 *rmapp = 0;
 661         } else {
 662                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 663                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 664                 prev_desc = NULL;
 665                 while (desc) {
 666                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
 667                                 if (desc->sptes[i] == spte) {
 668                                         rmap_desc_remove_entry(rmapp,
 669                                                                desc, i,
 670                                                                prev_desc);
 671                                         return;
 672                                 }
 673                         prev_desc = desc;
 674                         desc = desc->more;
 675                 }
 676                 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
 677                 BUG();
 678         }
 679 }
 680
 681 static void drop_spte(struct kvm *kvm, u64 *sptep, u64 new_spte)
 682 {
 683         pfn_t pfn;
 684         u64 old_spte;
 685
 686         old_spte = __xchg_spte(sptep, new_spte);
 687         if (!is_rmap_spte(old_spte))
 688                 return;
 689         pfn = spte_to_pfn(old_spte);
 690         if (old_spte & shadow_accessed_mask)
 691                 kvm_set_pfn_accessed(pfn);
 692         if (is_writable_pte(old_spte))
 693                 kvm_set_pfn_dirty(pfn);
 694         rmap_remove(kvm, sptep);
 695 }
 696
 697 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 698 {
 699         struct kvm_rmap_desc *desc;
 700         u64 *prev_spte;
 701         int i;
 702
 703         if (!*rmapp)
 704                 return NULL;
 705         else if (!(*rmapp & 1)) {
 706                 if (!spte)
 707                         return (u64 *)*rmapp;
 708                 return NULL;
 709         }
 710         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 711         prev_spte = NULL;
 712         while (desc) {
 713                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
 714                         if (prev_spte == spte)
 715                                 return desc->sptes[i];
 716                         prev_spte = desc->sptes[i];
 717                 }
 718                 desc = desc->more;
 719         }
 720         return NULL;
 721 }
 722
 723 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 724 {
 725         unsigned long *rmapp;
 726         u64 *spte;
 727         int i, write_protected = 0;
 728
 729         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
 730
 731         spte = rmap_next(kvm, rmapp, NULL);
 732         while (spte) {
 733                 BUG_ON(!spte);
 734                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 735                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 736                 if (is_writable_pte(*spte)) {
 737                         update_spte(spte, *spte & ~PT_WRITABLE_MASK);
 738                         write_protected = 1;
 739                 }
 740                 spte = rmap_next(kvm, rmapp, spte);
 741         }
 742         if (write_protected) {
 743                 pfn_t pfn;
 744
 745                 spte = rmap_next(kvm, rmapp, NULL);
 746                 pfn = spte_to_pfn(*spte);
 747                 kvm_set_pfn_dirty(pfn);
 748         }
 749
 750         /* check for huge page mappings */
 751         for (i = PT_DIRECTORY_LEVEL;
 752              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
 753                 rmapp = gfn_to_rmap(kvm, gfn, i);
 754                 spte = rmap_next(kvm, rmapp, NULL);
 755                 while (spte) {
 756                         BUG_ON(!spte);
 757                         BUG_ON(!(*spte & PT_PRESENT_MASK));
 758                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 759                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 760                         if (is_writable_pte(*spte)) {
 761                                 drop_spte(kvm, spte,
 762                                           shadow_trap_nonpresent_pte);
 763                                 --kvm->stat.lpages;
 764                                 spte = NULL;
 765                                 write_protected = 1;
 766                         }
 767                         spte = rmap_next(kvm, rmapp, spte);
 768                 }
 769         }
 770
 771         return write_protected;
 772 }
 773
 774 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 775                            unsigned long data)
 776 {
 777         u64 *spte;
 778         int need_tlb_flush = 0;
 779
 780         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 781                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 782                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 783                 drop_spte(kvm, spte, shadow_trap_nonpresent_pte);
 784                 need_tlb_flush = 1;
 785         }
 786         return need_tlb_flush;
 787 }
 788
 789 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
 790                              unsigned long data)
 791 {
 792         int need_flush = 0;
 793         u64 *spte, new_spte, old_spte;
 794         pte_t *ptep = (pte_t *)data;
 795         pfn_t new_pfn;
 796
 797         WARN_ON(pte_huge(*ptep));
 798         new_pfn = pte_pfn(*ptep);
 799         spte = rmap_next(kvm, rmapp, NULL);
 800         while (spte) {
 801                 BUG_ON(!is_shadow_present_pte(*spte));
 802                 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
 803                 need_flush = 1;
 804                 if (pte_write(*ptep)) {
 805                         drop_spte(kvm, spte, shadow_trap_nonpresent_pte);
 806                         spte = rmap_next(kvm, rmapp, NULL);
 807                 } else {
 808                         new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
 809                         new_spte |= (u64)new_pfn << PAGE_SHIFT;
 810
 811                         new_spte &= ~PT_WRITABLE_MASK;
 812                         new_spte &= ~SPTE_HOST_WRITEABLE;
 813                         new_spte &= ~shadow_accessed_mask;
 814                         if (is_writable_pte(*spte))
 815                                 kvm_set_pfn_dirty(spte_to_pfn(*spte));
 816                         old_spte = __xchg_spte(spte, new_spte);
 817                         if (is_shadow_present_pte(old_spte)
 818                             && (old_spte & shadow_accessed_mask))
 819                                 mark_page_accessed(pfn_to_page(spte_to_pfn(old_spte)));
 820                         spte = rmap_next(kvm, rmapp, spte);
 821                 }
 822         }
 823         if (need_flush)
 824                 kvm_flush_remote_tlbs(kvm);
 825
 826         return 0;
 827 }
 828
 829 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 830                           unsigned long data,
 831                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 832                                          unsigned long data))
 833 {
 834         int i, j;
 835         int ret;
 836         int retval = 0;
 837         struct kvm_memslots *slots;
 838
 839         slots = kvm_memslots(kvm);
 840
 841         for (i = 0; i < slots->nmemslots; i++) {
 842                 struct kvm_memory_slot *memslot = &slots->memslots[i];
 843                 unsigned long start = memslot->userspace_addr;
 844                 unsigned long end;
 845
 846                 end = start + (memslot->npages << PAGE_SHIFT);
 847                 if (hva >= start && hva < end) {
 848                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 849
 850                         ret = handler(kvm, &memslot->rmap[gfn_offset], data);
 851
 852                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
 853                                 int idx = gfn_offset;
 854                                 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
 855                                 ret |= handler(kvm,
 856                                         &memslot->lpage_info[j][idx].rmap_pde,
 857                                         data);
 858                         }
 859                         trace_kvm_age_page(hva, memslot, ret);
 860                         retval |= ret;
 861                 }
 862         }
 863
 864         return retval;
 865 }
 866
 867 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 868 {
 869         return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
 870 }
 871
 872 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
 873 {
 874         kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
 875 }
 876
 877 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 878                          unsigned long data)
 879 {
 880         u64 *spte;
 881         int young = 0;
 882
 883         /*
 884          * Emulate the accessed bit for EPT, by checking if this page has
 885          * an EPT mapping, and clearing it if it does. On the next access,
 886          * a new EPT mapping will be established.
 887          * This has some overhead, but not as much as the cost of swapping
 888          * out actively used pages or breaking up actively used hugepages.
 889          */
 890         if (!shadow_accessed_mask)
 891                 return kvm_unmap_rmapp(kvm, rmapp, data);
 892
 893         spte = rmap_next(kvm, rmapp, NULL);
 894         while (spte) {
 895                 int _young;
 896                 u64 _spte = *spte;
 897                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 898                 _young = _spte & PT_ACCESSED_MASK;
 899                 if (_young) {
 900                         young = 1;
 901                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 902                 }
 903                 spte = rmap_next(kvm, rmapp, spte);
 904         }
 905         return young;
 906 }
 907
 908 #define RMAP_RECYCLE_THRESHOLD 1000
 909
 910 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 911 {
 912         unsigned long *rmapp;
 913         struct kvm_mmu_page *sp;
 914
 915         sp = page_header(__pa(spte));
 916
 917         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
 918
 919         kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
 920         kvm_flush_remote_tlbs(vcpu->kvm);
 921 }
 922
 923 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 924 {
 925         return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
 926 }
 927
 928 #ifdef MMU_DEBUG
 929 static int is_empty_shadow_page(u64 *spt)
 930 {
 931         u64 *pos;
 932         u64 *end;
 933
 934         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 935                 if (is_shadow_present_pte(*pos)) {
 936                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 937                                pos, *pos);
 938                         return 0;
 939                 }
 940         return 1;
 941 }
 942 #endif
 943
 944 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 945 {
 946         ASSERT(is_empty_shadow_page(sp->spt));
 947         hlist_del(&sp->hash_link);
 948         list_del(&sp->link);
 949         __free_page(virt_to_page(sp->spt));
 950         if (!sp->role.direct)
 951                 __free_page(virt_to_page(sp->gfns));
 952         kmem_cache_free(mmu_page_header_cache, sp);
 953         ++kvm->arch.n_free_mmu_pages;
 954 }
 955
 956 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 957 {
 958         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 959 }
 960
 961 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 962                                                u64 *parent_pte, int direct)
 963 {
 964         struct kvm_mmu_page *sp;
 965
 966         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 967         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 968         if (!direct)
 969                 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache,
 970                                                   PAGE_SIZE);
 971         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 972         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 973         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 974         sp->multimapped = 0;
 975         sp->parent_pte = parent_pte;
 976         --vcpu->kvm->arch.n_free_mmu_pages;
 977         return sp;
 978 }
 979
 980 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 981                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 982 {
 983         struct kvm_pte_chain *pte_chain;
 984         struct hlist_node *node;
 985         int i;
 986
 987         if (!parent_pte)
 988                 return;
 989         if (!sp->multimapped) {
 990                 u64 *old = sp->parent_pte;
 991
 992                 if (!old) {
 993                         sp->parent_pte = parent_pte;
 994                         return;
 995                 }
 996                 sp->multimapped = 1;
 997                 pte_chain = mmu_alloc_pte_chain(vcpu);
 998                 INIT_HLIST_HEAD(&sp->parent_ptes);
 999                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
1000                 pte_chain->parent_ptes[0] = old;
1001         }
1002         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
1003                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
1004                         continue;
1005                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
1006                         if (!pte_chain->parent_ptes[i]) {
1007                                 pte_chain->parent_ptes[i] = parent_pte;
1008                                 return;
1009                         }
1010         }
1011         pte_chain = mmu_alloc_pte_chain(vcpu);
1012         BUG_ON(!pte_chain);
1013         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
1014         pte_chain->parent_ptes[0] = parent_pte;
1015 }
1016
1017 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1018                                        u64 *parent_pte)
1019 {
1020         struct kvm_pte_chain *pte_chain;
1021         struct hlist_node *node;
1022         int i;
1023
1024         if (!sp->multimapped) {
1025                 BUG_ON(sp->parent_pte != parent_pte);
1026                 sp->parent_pte = NULL;
1027                 return;
1028         }
1029         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1030                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1031                         if (!pte_chain->parent_ptes[i])
1032                                 break;
1033                         if (pte_chain->parent_ptes[i] != parent_pte)
1034                                 continue;
1035                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
1036                                 && pte_chain->parent_ptes[i + 1]) {
1037                                 pte_chain->parent_ptes[i]
1038                                         = pte_chain->parent_ptes[i + 1];
1039                                 ++i;
1040                         }
1041                         pte_chain->parent_ptes[i] = NULL;
1042                         if (i == 0) {
1043                                 hlist_del(&pte_chain->link);
1044                                 mmu_free_pte_chain(pte_chain);
1045                                 if (hlist_empty(&sp->parent_ptes)) {
1046                                         sp->multimapped = 0;
1047                                         sp->parent_pte = NULL;
1048                                 }
1049                         }
1050                         return;
1051                 }
1052         BUG();
1053 }
1054
1055 static void mmu_parent_walk(struct kvm_mmu_page *sp, mmu_parent_walk_fn fn)
1056 {
1057         struct kvm_pte_chain *pte_chain;
1058         struct hlist_node *node;
1059         struct kvm_mmu_page *parent_sp;
1060         int i;
1061
1062         if (!sp->multimapped && sp->parent_pte) {
1063                 parent_sp = page_header(__pa(sp->parent_pte));
1064                 fn(parent_sp, sp->parent_pte);
1065                 return;
1066         }
1067
1068         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1069                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1070                         u64 *spte = pte_chain->parent_ptes[i];
1071
1072                         if (!spte)
1073                                 break;
1074                         parent_sp = page_header(__pa(spte));
1075                         fn(parent_sp, spte);
1076                 }
1077 }
1078
1079 static void mark_unsync(struct kvm_mmu_page *sp, u64 *spte);
1080 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1081 {
1082         mmu_parent_walk(sp, mark_unsync);
1083 }
1084
1085 static void mark_unsync(struct kvm_mmu_page *sp, u64 *spte)
1086 {
1087         unsigned int index;
1088
1089         index = spte - sp->spt;
1090         if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1091                 return;
1092         if (sp->unsync_children++)
1093                 return;
1094         kvm_mmu_mark_parents_unsync(sp);
1095 }
1096
1097 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1098                                     struct kvm_mmu_page *sp)
1099 {
1100         int i;
1101
1102         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1103                 sp->spt[i] = shadow_trap_nonpresent_pte;
1104 }
1105
1106 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1107                                struct kvm_mmu_page *sp, bool clear_unsync)
1108 {
1109         return 1;
1110 }
1111
1112 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1113 {
1114 }
1115
1116 #define KVM_PAGE_ARRAY_NR 16
1117
1118 struct kvm_mmu_pages {
1119         struct mmu_page_and_offset {
1120                 struct kvm_mmu_page *sp;
1121                 unsigned int idx;
1122         } page[KVM_PAGE_ARRAY_NR];
1123         unsigned int nr;
1124 };
1125
1126 #define for_each_unsync_children(bitmap, idx)           \
1127         for (idx = find_first_bit(bitmap, 512);         \
1128              idx < 512;                                 \
1129              idx = find_next_bit(bitmap, 512, idx+1))
1130
1131 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1132                          int idx)
1133 {
1134         int i;
1135
1136         if (sp->unsync)
1137                 for (i=0; i < pvec->nr; i++)
1138                         if (pvec->page[i].sp == sp)
1139                                 return 0;
1140
1141         pvec->page[pvec->nr].sp = sp;
1142         pvec->page[pvec->nr].idx = idx;
1143         pvec->nr++;
1144         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1145 }
1146
1147 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1148                            struct kvm_mmu_pages *pvec)
1149 {
1150         int i, ret, nr_unsync_leaf = 0;
1151
1152         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1153                 struct kvm_mmu_page *child;
1154                 u64 ent = sp->spt[i];
1155
1156                 if (!is_shadow_present_pte(ent) || is_large_pte(ent))
1157                         goto clear_child_bitmap;
1158
1159                 child = page_header(ent & PT64_BASE_ADDR_MASK);
1160
1161                 if (child->unsync_children) {
1162                         if (mmu_pages_add(pvec, child, i))
1163                                 return -ENOSPC;
1164
1165                         ret = __mmu_unsync_walk(child, pvec);
1166                         if (!ret)
1167                                 goto clear_child_bitmap;
1168                         else if (ret > 0)
1169                                 nr_unsync_leaf += ret;
1170                         else
1171                                 return ret;
1172                 } else if (child->unsync) {
1173                         nr_unsync_leaf++;
1174                         if (mmu_pages_add(pvec, child, i))
1175                                 return -ENOSPC;
1176                 } else
1177                          goto clear_child_bitmap;
1178
1179                 continue;
1180
1181 clear_child_bitmap:
1182                 __clear_bit(i, sp->unsync_child_bitmap);
1183                 sp->unsync_children--;
1184                 WARN_ON((int)sp->unsync_children < 0);
1185         }
1186
1187
1188         return nr_unsync_leaf;
1189 }
1190
1191 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1192                            struct kvm_mmu_pages *pvec)
1193 {
1194         if (!sp->unsync_children)
1195                 return 0;
1196
1197         mmu_pages_add(pvec, sp, 0);
1198         return __mmu_unsync_walk(sp, pvec);
1199 }
1200
1201 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1202 {
1203         WARN_ON(!sp->unsync);
1204         trace_kvm_mmu_sync_page(sp);
1205         sp->unsync = 0;
1206         --kvm->stat.mmu_unsync;
1207 }
1208
1209 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1210                                     struct list_head *invalid_list);
1211 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1212                                     struct list_head *invalid_list);
1213
1214 #define for_each_gfn_sp(kvm, sp, gfn, pos)                              \
1215   hlist_for_each_entry(sp, pos,                                         \
1216    &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link)   \
1217         if ((sp)->gfn != (gfn)) {} else
1218
1219 #define for_each_gfn_indirect_valid_sp(kvm, sp, gfn, pos)               \
1220   hlist_for_each_entry(sp, pos,                                         \
1221    &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link)   \
1222                 if ((sp)->gfn != (gfn) || (sp)->role.direct ||          \
1223                         (sp)->role.invalid) {} else
1224
1225 /* @sp->gfn should be write-protected at the call site */
1226 static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1227                            struct list_head *invalid_list, bool clear_unsync)
1228 {
1229         if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1230                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1231                 return 1;
1232         }
1233
1234         if (clear_unsync)
1235                 kvm_unlink_unsync_page(vcpu->kvm, sp);
1236
1237         if (vcpu->arch.mmu.sync_page(vcpu, sp, clear_unsync)) {
1238                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1239                 return 1;
1240         }
1241
1242         kvm_mmu_flush_tlb(vcpu);
1243         return 0;
1244 }
1245
1246 static int kvm_sync_page_transient(struct kvm_vcpu *vcpu,
1247                                    struct kvm_mmu_page *sp)
1248 {
1249         LIST_HEAD(invalid_list);
1250         int ret;
1251
1252         ret = __kvm_sync_page(vcpu, sp, &invalid_list, false);
1253         if (ret)
1254                 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1255
1256         return ret;
1257 }
1258
1259 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1260                          struct list_head *invalid_list)
1261 {
1262         return __kvm_sync_page(vcpu, sp, invalid_list, true);
1263 }
1264
1265 /* @gfn should be write-protected at the call site */
1266 static void kvm_sync_pages(struct kvm_vcpu *vcpu,  gfn_t gfn)
1267 {
1268         struct kvm_mmu_page *s;
1269         struct hlist_node *node;
1270         LIST_HEAD(invalid_list);
1271         bool flush = false;
1272
1273         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1274                 if (!s->unsync)
1275                         continue;
1276
1277                 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1278                 if ((s->role.cr4_pae != !!is_pae(vcpu)) ||
1279                         (vcpu->arch.mmu.sync_page(vcpu, s, true))) {
1280                         kvm_mmu_prepare_zap_page(vcpu->kvm, s, &invalid_list);
1281                         continue;
1282                 }
1283                 kvm_unlink_unsync_page(vcpu->kvm, s);
1284                 flush = true;
1285         }
1286
1287         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1288         if (flush)
1289                 kvm_mmu_flush_tlb(vcpu);
1290 }
1291
1292 struct mmu_page_path {
1293         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1294         unsigned int idx[PT64_ROOT_LEVEL-1];
1295 };
1296
1297 #define for_each_sp(pvec, sp, parents, i)                       \
1298                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1299                         sp = pvec.page[i].sp;                   \
1300                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1301                         i = mmu_pages_next(&pvec, &parents, i))
1302
1303 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1304                           struct mmu_page_path *parents,
1305                           int i)
1306 {
1307         int n;
1308
1309         for (n = i+1; n < pvec->nr; n++) {
1310                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1311
1312                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1313                         parents->idx[0] = pvec->page[n].idx;
1314                         return n;
1315                 }
1316
1317                 parents->parent[sp->role.level-2] = sp;
1318                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1319         }
1320
1321         return n;
1322 }
1323
1324 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1325 {
1326         struct kvm_mmu_page *sp;
1327         unsigned int level = 0;
1328
1329         do {
1330                 unsigned int idx = parents->idx[level];
1331
1332                 sp = parents->parent[level];
1333                 if (!sp)
1334                         return;
1335
1336                 --sp->unsync_children;
1337                 WARN_ON((int)sp->unsync_children < 0);
1338                 __clear_bit(idx, sp->unsync_child_bitmap);
1339                 level++;
1340         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1341 }
1342
1343 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1344                                struct mmu_page_path *parents,
1345                                struct kvm_mmu_pages *pvec)
1346 {
1347         parents->parent[parent->role.level-1] = NULL;
1348         pvec->nr = 0;
1349 }
1350
1351 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1352                               struct kvm_mmu_page *parent)
1353 {
1354         int i;
1355         struct kvm_mmu_page *sp;
1356         struct mmu_page_path parents;
1357         struct kvm_mmu_pages pages;
1358         LIST_HEAD(invalid_list);
1359
1360         kvm_mmu_pages_init(parent, &parents, &pages);
1361         while (mmu_unsync_walk(parent, &pages)) {
1362                 int protected = 0;
1363
1364                 for_each_sp(pages, sp, parents, i)
1365                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1366
1367                 if (protected)
1368                         kvm_flush_remote_tlbs(vcpu->kvm);
1369
1370                 for_each_sp(pages, sp, parents, i) {
1371                         kvm_sync_page(vcpu, sp, &invalid_list);
1372                         mmu_pages_clear_parents(&parents);
1373                 }
1374                 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1375                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1376                 kvm_mmu_pages_init(parent, &parents, &pages);
1377         }
1378 }
1379
1380 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1381                                              gfn_t gfn,
1382                                              gva_t gaddr,
1383                                              unsigned level,
1384                                              int direct,
1385                                              unsigned access,
1386                                              u64 *parent_pte)
1387 {
1388         union kvm_mmu_page_role role;
1389         unsigned quadrant;
1390         struct kvm_mmu_page *sp;
1391         struct hlist_node *node;
1392         bool need_sync = false;
1393
1394         role = vcpu->arch.mmu.base_role;
1395         role.level = level;
1396         role.direct = direct;
1397         if (role.direct)
1398                 role.cr4_pae = 0;
1399         role.access = access;
1400         if (!tdp_enabled && vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1401                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1402                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1403                 role.quadrant = quadrant;
1404         }
1405         for_each_gfn_sp(vcpu->kvm, sp, gfn, node) {
1406                 if (!need_sync && sp->unsync)
1407                         need_sync = true;
1408
1409                 if (sp->role.word != role.word)
1410                         continue;
1411
1412                 if (sp->unsync && kvm_sync_page_transient(vcpu, sp))
1413                         break;
1414
1415                 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1416                 if (sp->unsync_children) {
1417                         kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
1418                         kvm_mmu_mark_parents_unsync(sp);
1419                 } else if (sp->unsync)
1420                         kvm_mmu_mark_parents_unsync(sp);
1421
1422                 trace_kvm_mmu_get_page(sp, false);
1423                 return sp;
1424         }
1425         ++vcpu->kvm->stat.mmu_cache_miss;
1426         sp = kvm_mmu_alloc_page(vcpu, parent_pte, direct);
1427         if (!sp)
1428                 return sp;
1429         sp->gfn = gfn;
1430         sp->role = role;
1431         hlist_add_head(&sp->hash_link,
1432                 &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
1433         if (!direct) {
1434                 if (rmap_write_protect(vcpu->kvm, gfn))
1435                         kvm_flush_remote_tlbs(vcpu->kvm);
1436                 if (level > PT_PAGE_TABLE_LEVEL && need_sync)
1437                         kvm_sync_pages(vcpu, gfn);
1438
1439                 account_shadowed(vcpu->kvm, gfn);
1440         }
1441         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1442                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1443         else
1444                 nonpaging_prefetch_page(vcpu, sp);
1445         trace_kvm_mmu_get_page(sp, true);
1446         return sp;
1447 }
1448
1449 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1450                              struct kvm_vcpu *vcpu, u64 addr)
1451 {
1452         iterator->addr = addr;
1453         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1454         iterator->level = vcpu->arch.mmu.shadow_root_level;
1455         if (iterator->level == PT32E_ROOT_LEVEL) {
1456                 iterator->shadow_addr
1457                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1458                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1459                 --iterator->level;
1460                 if (!iterator->shadow_addr)
1461                         iterator->level = 0;
1462         }
1463 }
1464
1465 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1466 {
1467         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1468                 return false;
1469
1470         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1471                 if (is_large_pte(*iterator->sptep))
1472                         return false;
1473
1474         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1475         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1476         return true;
1477 }
1478
1479 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1480 {
1481         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1482         --iterator->level;
1483 }
1484
1485 static void link_shadow_page(u64 *sptep, struct kvm_mmu_page *sp)
1486 {
1487         u64 spte;
1488
1489         spte = __pa(sp->spt)
1490                 | PT_PRESENT_MASK | PT_ACCESSED_MASK
1491                 | PT_WRITABLE_MASK | PT_USER_MASK;
1492         __set_spte(sptep, spte);
1493 }
1494
1495 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1496 {
1497         if (is_large_pte(*sptep)) {
1498                 drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
1499                 kvm_flush_remote_tlbs(vcpu->kvm);
1500         }
1501 }
1502
1503 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1504                                          struct kvm_mmu_page *sp)
1505 {
1506         unsigned i;
1507         u64 *pt;
1508         u64 ent;
1509
1510         pt = sp->spt;
1511
1512         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1513                 ent = pt[i];
1514
1515                 if (is_shadow_present_pte(ent)) {
1516                         if (!is_last_spte(ent, sp->role.level)) {
1517                                 ent &= PT64_BASE_ADDR_MASK;
1518                                 mmu_page_remove_parent_pte(page_header(ent),
1519                                                            &pt[i]);
1520                         } else {
1521                                 if (is_large_pte(ent))
1522                                         --kvm->stat.lpages;
1523                                 drop_spte(kvm, &pt[i],
1524                                           shadow_trap_nonpresent_pte);
1525                         }
1526                 }
1527                 pt[i] = shadow_trap_nonpresent_pte;
1528         }
1529 }
1530
1531 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1532 {
1533         mmu_page_remove_parent_pte(sp, parent_pte);
1534 }
1535
1536 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1537 {
1538         int i;
1539         struct kvm_vcpu *vcpu;
1540
1541         kvm_for_each_vcpu(i, vcpu, kvm)
1542                 vcpu->arch.last_pte_updated = NULL;
1543 }
1544
1545 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1546 {
1547         u64 *parent_pte;
1548
1549         while (sp->multimapped || sp->parent_pte) {
1550                 if (!sp->multimapped)
1551                         parent_pte = sp->parent_pte;
1552                 else {
1553                         struct kvm_pte_chain *chain;
1554
1555                         chain = container_of(sp->parent_ptes.first,
1556                                              struct kvm_pte_chain, link);
1557                         parent_pte = chain->parent_ptes[0];
1558                 }
1559                 BUG_ON(!parent_pte);
1560                 kvm_mmu_put_page(sp, parent_pte);
1561                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1562         }
1563 }
1564
1565 static int mmu_zap_unsync_children(struct kvm *kvm,
1566                                    struct kvm_mmu_page *parent,
1567                                    struct list_head *invalid_list)
1568 {
1569         int i, zapped = 0;
1570         struct mmu_page_path parents;
1571         struct kvm_mmu_pages pages;
1572
1573         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1574                 return 0;
1575
1576         kvm_mmu_pages_init(parent, &parents, &pages);
1577         while (mmu_unsync_walk(parent, &pages)) {
1578                 struct kvm_mmu_page *sp;
1579
1580                 for_each_sp(pages, sp, parents, i) {
1581                         kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
1582                         mmu_pages_clear_parents(&parents);
1583                         zapped++;
1584                 }
1585                 kvm_mmu_pages_init(parent, &parents, &pages);
1586         }
1587
1588         return zapped;
1589 }
1590
1591 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1592                                     struct list_head *invalid_list)
1593 {
1594         int ret;
1595
1596         trace_kvm_mmu_prepare_zap_page(sp);
1597         ++kvm->stat.mmu_shadow_zapped;
1598         ret = mmu_zap_unsync_children(kvm, sp, invalid_list);
1599         kvm_mmu_page_unlink_children(kvm, sp);
1600         kvm_mmu_unlink_parents(kvm, sp);
1601         if (!sp->role.invalid && !sp->role.direct)
1602                 unaccount_shadowed(kvm, sp->gfn);
1603         if (sp->unsync)
1604                 kvm_unlink_unsync_page(kvm, sp);
1605         if (!sp->root_count) {
1606                 /* Count self */
1607                 ret++;
1608                 list_move(&sp->link, invalid_list);
1609         } else {
1610                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1611                 kvm_reload_remote_mmus(kvm);
1612         }
1613
1614         sp->role.invalid = 1;
1615         kvm_mmu_reset_last_pte_updated(kvm);
1616         return ret;
1617 }
1618
1619 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1620                                     struct list_head *invalid_list)
1621 {
1622         struct kvm_mmu_page *sp;
1623
1624         if (list_empty(invalid_list))
1625                 return;
1626
1627         kvm_flush_remote_tlbs(kvm);
1628
1629         do {
1630                 sp = list_first_entry(invalid_list, struct kvm_mmu_page, link);
1631                 WARN_ON(!sp->role.invalid || sp->root_count);
1632                 kvm_mmu_free_page(kvm, sp);
1633         } while (!list_empty(invalid_list));
1634
1635 }
1636
1637 /*
1638  * Changing the number of mmu pages allocated to the vm
1639  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1640  */
1641 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1642 {
1643         int used_pages;
1644         LIST_HEAD(invalid_list);
1645
1646         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1647         used_pages = max(0, used_pages);
1648
1649         /*
1650          * If we set the number of mmu pages to be smaller be than the
1651          * number of actived pages , we must to free some mmu pages before we
1652          * change the value
1653          */
1654
1655         if (used_pages > kvm_nr_mmu_pages) {
1656                 while (used_pages > kvm_nr_mmu_pages &&
1657                         !list_empty(&kvm->arch.active_mmu_pages)) {
1658                         struct kvm_mmu_page *page;
1659
1660                         page = container_of(kvm->arch.active_mmu_pages.prev,
1661                                             struct kvm_mmu_page, link);
1662                         used_pages -= kvm_mmu_prepare_zap_page(kvm, page,
1663                                                                &invalid_list);
1664                 }
1665                 kvm_mmu_commit_zap_page(kvm, &invalid_list);
1666                 kvm_nr_mmu_pages = used_pages;
1667                 kvm->arch.n_free_mmu_pages = 0;
1668         }
1669         else
1670                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1671                                          - kvm->arch.n_alloc_mmu_pages;
1672
1673         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1674 }
1675
1676 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1677 {
1678         struct kvm_mmu_page *sp;
1679         struct hlist_node *node;
1680         LIST_HEAD(invalid_list);
1681         int r;
1682
1683         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1684         r = 0;
1685
1686         for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1687                 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1688                          sp->role.word);
1689                 r = 1;
1690                 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1691         }
1692         kvm_mmu_commit_zap_page(kvm, &invalid_list);
1693         return r;
1694 }
1695
1696 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1697 {
1698         struct kvm_mmu_page *sp;
1699         struct hlist_node *node;
1700         LIST_HEAD(invalid_list);
1701
1702         for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1703                 pgprintk("%s: zap %lx %x\n",
1704                          __func__, gfn, sp->role.word);
1705                 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1706         }
1707         kvm_mmu_commit_zap_page(kvm, &invalid_list);
1708 }
1709
1710 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1711 {
1712         int slot = memslot_id(kvm, gfn);
1713         struct kvm_mmu_page *sp = page_header(__pa(pte));
1714
1715         __set_bit(slot, sp->slot_bitmap);
1716 }
1717
1718 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1719 {
1720         int i;
1721         u64 *pt = sp->spt;
1722
1723         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1724                 return;
1725
1726         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1727                 if (pt[i] == shadow_notrap_nonpresent_pte)
1728                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1729         }
1730 }
1731
1732 /*
1733  * The function is based on mtrr_type_lookup() in
1734  * arch/x86/kernel/cpu/mtrr/generic.c
1735  */
1736 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1737                          u64 start, u64 end)
1738 {
1739         int i;
1740         u64 base, mask;
1741         u8 prev_match, curr_match;
1742         int num_var_ranges = KVM_NR_VAR_MTRR;
1743
1744         if (!mtrr_state->enabled)
1745                 return 0xFF;
1746
1747         /* Make end inclusive end, instead of exclusive */
1748         end--;
1749
1750         /* Look in fixed ranges. Just return the type as per start */
1751         if (mtrr_state->have_fixed && (start < 0x100000)) {
1752                 int idx;
1753
1754                 if (start < 0x80000) {
1755                         idx = 0;
1756                         idx += (start >> 16);
1757                         return mtrr_state->fixed_ranges[idx];
1758                 } else if (start < 0xC0000) {
1759                         idx = 1 * 8;
1760                         idx += ((start - 0x80000) >> 14);
1761                         return mtrr_state->fixed_ranges[idx];
1762                 } else if (start < 0x1000000) {
1763                         idx = 3 * 8;
1764                         idx += ((start - 0xC0000) >> 12);
1765                         return mtrr_state->fixed_ranges[idx];
1766                 }
1767         }
1768
1769         /*
1770          * Look in variable ranges
1771          * Look of multiple ranges matching this address and pick type
1772          * as per MTRR precedence
1773          */
1774         if (!(mtrr_state->enabled & 2))
1775                 return mtrr_state->def_type;
1776
1777         prev_match = 0xFF;
1778         for (i = 0; i < num_var_ranges; ++i) {
1779                 unsigned short start_state, end_state;
1780
1781                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1782                         continue;
1783
1784                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1785                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1786                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1787                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1788
1789                 start_state = ((start & mask) == (base & mask));
1790                 end_state = ((end & mask) == (base & mask));
1791                 if (start_state != end_state)
1792                         return 0xFE;
1793
1794                 if ((start & mask) != (base & mask))
1795                         continue;
1796
1797                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1798                 if (prev_match == 0xFF) {
1799                         prev_match = curr_match;
1800                         continue;
1801                 }
1802
1803                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1804                     curr_match == MTRR_TYPE_UNCACHABLE)
1805                         return MTRR_TYPE_UNCACHABLE;
1806
1807                 if ((prev_match == MTRR_TYPE_WRBACK &&
1808                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1809                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1810                      curr_match == MTRR_TYPE_WRBACK)) {
1811                         prev_match = MTRR_TYPE_WRTHROUGH;
1812                         curr_match = MTRR_TYPE_WRTHROUGH;
1813                 }
1814
1815                 if (prev_match != curr_match)
1816                         return MTRR_TYPE_UNCACHABLE;
1817         }
1818
1819         if (prev_match != 0xFF)
1820                 return prev_match;
1821
1822         return mtrr_state->def_type;
1823 }
1824
1825 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1826 {
1827         u8 mtrr;
1828
1829         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1830                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1831         if (mtrr == 0xfe || mtrr == 0xff)
1832                 mtrr = MTRR_TYPE_WRBACK;
1833         return mtrr;
1834 }
1835 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1836
1837 static void __kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1838 {
1839         trace_kvm_mmu_unsync_page(sp);
1840         ++vcpu->kvm->stat.mmu_unsync;
1841         sp->unsync = 1;
1842
1843         kvm_mmu_mark_parents_unsync(sp);
1844         mmu_convert_notrap(sp);
1845 }
1846
1847 static void kvm_unsync_pages(struct kvm_vcpu *vcpu,  gfn_t gfn)
1848 {
1849         struct kvm_mmu_page *s;
1850         struct hlist_node *node;
1851
1852         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1853                 if (s->unsync)
1854                         continue;
1855                 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1856                 __kvm_unsync_page(vcpu, s);
1857         }
1858 }
1859
1860 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1861                                   bool can_unsync)
1862 {
1863         struct kvm_mmu_page *s;
1864         struct hlist_node *node;
1865         bool need_unsync = false;
1866
1867         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1868                 if (!can_unsync)
1869                         return 1;
1870
1871                 if (s->role.level != PT_PAGE_TABLE_LEVEL)
1872                         return 1;
1873
1874                 if (!need_unsync && !s->unsync) {
1875                         if (!oos_shadow)
1876                                 return 1;
1877                         need_unsync = true;
1878                 }
1879         }
1880         if (need_unsync)
1881                 kvm_unsync_pages(vcpu, gfn);
1882         return 0;
1883 }
1884
1885 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1886                     unsigned pte_access, int user_fault,
1887                     int write_fault, int dirty, int level,
1888                     gfn_t gfn, pfn_t pfn, bool speculative,
1889                     bool can_unsync, bool reset_host_protection)
1890 {
1891         u64 spte;
1892         int ret = 0;
1893
1894         /*
1895          * We don't set the accessed bit, since we sometimes want to see
1896          * whether the guest actually used the pte (in order to detect
1897          * demand paging).
1898          */
1899         spte = shadow_base_present_pte | shadow_dirty_mask;
1900         if (!speculative)
1901                 spte |= shadow_accessed_mask;
1902         if (!dirty)
1903                 pte_access &= ~ACC_WRITE_MASK;
1904         if (pte_access & ACC_EXEC_MASK)
1905                 spte |= shadow_x_mask;
1906         else
1907                 spte |= shadow_nx_mask;
1908         if (pte_access & ACC_USER_MASK)
1909                 spte |= shadow_user_mask;
1910         if (level > PT_PAGE_TABLE_LEVEL)
1911                 spte |= PT_PAGE_SIZE_MASK;
1912         if (tdp_enabled)
1913                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1914                         kvm_is_mmio_pfn(pfn));
1915
1916         if (reset_host_protection)
1917                 spte |= SPTE_HOST_WRITEABLE;
1918
1919         spte |= (u64)pfn << PAGE_SHIFT;
1920
1921         if ((pte_access & ACC_WRITE_MASK)
1922             || (!tdp_enabled && write_fault && !is_write_protection(vcpu)
1923                 && !user_fault)) {
1924
1925                 if (level > PT_PAGE_TABLE_LEVEL &&
1926                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1927                         ret = 1;
1928                         drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
1929                         goto done;
1930                 }
1931
1932                 spte |= PT_WRITABLE_MASK;
1933
1934                 if (!tdp_enabled && !(pte_access & ACC_WRITE_MASK))
1935                         spte &= ~PT_USER_MASK;
1936
1937                 /*
1938                  * Optimization: for pte sync, if spte was writable the hash
1939                  * lookup is unnecessary (and expensive). Write protection
1940                  * is responsibility of mmu_get_page / kvm_sync_page.
1941                  * Same reasoning can be applied to dirty page accounting.
1942                  */
1943                 if (!can_unsync && is_writable_pte(*sptep))
1944                         goto set_pte;
1945
1946                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1947                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1948                                  __func__, gfn);
1949                         ret = 1;
1950                         pte_access &= ~ACC_WRITE_MASK;
1951                         if (is_writable_pte(spte))
1952                                 spte &= ~PT_WRITABLE_MASK;
1953                 }
1954         }
1955
1956         if (pte_access & ACC_WRITE_MASK)
1957                 mark_page_dirty(vcpu->kvm, gfn);
1958
1959 set_pte:
1960         update_spte(sptep, spte);
1961 done:
1962         return ret;
1963 }
1964
1965 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1966                          unsigned pt_access, unsigned pte_access,
1967                          int user_fault, int write_fault, int dirty,
1968                          int *ptwrite, int level, gfn_t gfn,
1969                          pfn_t pfn, bool speculative,
1970                          bool reset_host_protection)
1971 {
1972         int was_rmapped = 0;
1973         int was_writable = is_writable_pte(*sptep);
1974         int rmap_count;
1975
1976         pgprintk("%s: spte %llx access %x write_fault %d"
1977                  " user_fault %d gfn %lx\n",
1978                  __func__, *sptep, pt_access,
1979                  write_fault, user_fault, gfn);
1980
1981         if (is_rmap_spte(*sptep)) {
1982                 /*
1983                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1984                  * the parent of the now unreachable PTE.
1985                  */
1986                 if (level > PT_PAGE_TABLE_LEVEL &&
1987                     !is_large_pte(*sptep)) {
1988                         struct kvm_mmu_page *child;
1989                         u64 pte = *sptep;
1990
1991                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1992                         mmu_page_remove_parent_pte(child, sptep);
1993                         __set_spte(sptep, shadow_trap_nonpresent_pte);
1994                         kvm_flush_remote_tlbs(vcpu->kvm);
1995                 } else if (pfn != spte_to_pfn(*sptep)) {
1996                         pgprintk("hfn old %lx new %lx\n",
1997                                  spte_to_pfn(*sptep), pfn);
1998                         drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
1999                         kvm_flush_remote_tlbs(vcpu->kvm);
2000                 } else
2001                         was_rmapped = 1;
2002         }
2003
2004         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
2005                       dirty, level, gfn, pfn, speculative, true,
2006                       reset_host_protection)) {
2007                 if (write_fault)
2008                         *ptwrite = 1;
2009                 kvm_mmu_flush_tlb(vcpu);
2010         }
2011
2012         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2013         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
2014                  is_large_pte(*sptep)? "2MB" : "4kB",
2015                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
2016                  *sptep, sptep);
2017         if (!was_rmapped && is_large_pte(*sptep))
2018                 ++vcpu->kvm->stat.lpages;
2019
2020         page_header_update_slot(vcpu->kvm, sptep, gfn);
2021         if (!was_rmapped) {
2022                 rmap_count = rmap_add(vcpu, sptep, gfn);
2023                 kvm_release_pfn_clean(pfn);
2024                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
2025                         rmap_recycle(vcpu, sptep, gfn);
2026         } else {
2027                 if (was_writable)
2028                         kvm_release_pfn_dirty(pfn);
2029                 else
2030                         kvm_release_pfn_clean(pfn);
2031         }
2032         if (speculative) {
2033                 vcpu->arch.last_pte_updated = sptep;
2034                 vcpu->arch.last_pte_gfn = gfn;
2035         }
2036 }
2037
2038 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
2039 {
2040 }
2041
2042 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
2043                         int level, gfn_t gfn, pfn_t pfn)
2044 {
2045         struct kvm_shadow_walk_iterator iterator;
2046         struct kvm_mmu_page *sp;
2047         int pt_write = 0;
2048         gfn_t pseudo_gfn;
2049
2050         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
2051                 if (iterator.level == level) {
2052                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
2053                                      0, write, 1, &pt_write,
2054                                      level, gfn, pfn, false, true);
2055                         ++vcpu->stat.pf_fixed;
2056                         break;
2057                 }
2058
2059                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
2060                         u64 base_addr = iterator.addr;
2061
2062                         base_addr &= PT64_LVL_ADDR_MASK(iterator.level);
2063                         pseudo_gfn = base_addr >> PAGE_SHIFT;
2064                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
2065                                               iterator.level - 1,
2066                                               1, ACC_ALL, iterator.sptep);
2067                         if (!sp) {
2068                                 pgprintk("nonpaging_map: ENOMEM\n");
2069                                 kvm_release_pfn_clean(pfn);
2070                                 return -ENOMEM;
2071                         }
2072
2073                         __set_spte(iterator.sptep,
2074                                    __pa(sp->spt)
2075                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
2076                                    | shadow_user_mask | shadow_x_mask);
2077                 }
2078         }
2079         return pt_write;
2080 }
2081
2082 static void kvm_send_hwpoison_signal(struct kvm *kvm, gfn_t gfn)
2083 {
2084         char buf[1];
2085         void __user *hva;
2086         int r;
2087
2088         /* Touch the page, so send SIGBUS */
2089         hva = (void __user *)gfn_to_hva(kvm, gfn);
2090         r = copy_from_user(buf, hva, 1);
2091 }
2092
2093 static int kvm_handle_bad_page(struct kvm *kvm, gfn_t gfn, pfn_t pfn)
2094 {
2095         kvm_release_pfn_clean(pfn);
2096         if (is_hwpoison_pfn(pfn)) {
2097                 kvm_send_hwpoison_signal(kvm, gfn);
2098                 return 0;
2099         } else if (is_fault_pfn(pfn))
2100                 return -EFAULT;
2101
2102         return 1;
2103 }
2104
2105 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
2106 {
2107         int r;
2108         int level;
2109         pfn_t pfn;
2110         unsigned long mmu_seq;
2111
2112         level = mapping_level(vcpu, gfn);
2113
2114         /*
2115          * This path builds a PAE pagetable - so we can map 2mb pages at
2116          * maximum. Therefore check if the level is larger than that.
2117          */
2118         if (level > PT_DIRECTORY_LEVEL)
2119                 level = PT_DIRECTORY_LEVEL;
2120
2121         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2122
2123         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2124         smp_rmb();
2125         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2126
2127         /* mmio */
2128         if (is_error_pfn(pfn))
2129                 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2130
2131         spin_lock(&vcpu->kvm->mmu_lock);
2132         if (mmu_notifier_retry(vcpu, mmu_seq))
2133                 goto out_unlock;
2134         kvm_mmu_free_some_pages(vcpu);
2135         r = __direct_map(vcpu, v, write, level, gfn, pfn);
2136         spin_unlock(&vcpu->kvm->mmu_lock);
2137
2138
2139         return r;
2140
2141 out_unlock:
2142         spin_unlock(&vcpu->kvm->mmu_lock);
2143         kvm_release_pfn_clean(pfn);
2144         return 0;
2145 }
2146
2147
2148 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2149 {
2150         int i;
2151         struct kvm_mmu_page *sp;
2152         LIST_HEAD(invalid_list);
2153
2154         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2155                 return;
2156         spin_lock(&vcpu->kvm->mmu_lock);
2157         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2158                 hpa_t root = vcpu->arch.mmu.root_hpa;
2159
2160                 sp = page_header(root);
2161                 --sp->root_count;
2162                 if (!sp->root_count && sp->role.invalid) {
2163                         kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
2164                         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2165                 }
2166                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2167                 spin_unlock(&vcpu->kvm->mmu_lock);
2168                 return;
2169         }
2170         for (i = 0; i < 4; ++i) {
2171                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2172
2173                 if (root) {
2174                         root &= PT64_BASE_ADDR_MASK;
2175                         sp = page_header(root);
2176                         --sp->root_count;
2177                         if (!sp->root_count && sp->role.invalid)
2178                                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2179                                                          &invalid_list);
2180                 }
2181                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2182         }
2183         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2184         spin_unlock(&vcpu->kvm->mmu_lock);
2185         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2186 }
2187
2188 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2189 {
2190         int ret = 0;
2191
2192         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2193                 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
2194                 ret = 1;
2195         }
2196
2197         return ret;
2198 }
2199
2200 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2201 {
2202         int i;
2203         gfn_t root_gfn;
2204         struct kvm_mmu_page *sp;
2205         int direct = 0;
2206         u64 pdptr;
2207
2208         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2209
2210         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2211                 hpa_t root = vcpu->arch.mmu.root_hpa;
2212
2213                 ASSERT(!VALID_PAGE(root));
2214                 if (mmu_check_root(vcpu, root_gfn))
2215                         return 1;
2216                 if (tdp_enabled) {
2217                         direct = 1;
2218                         root_gfn = 0;
2219                 }
2220                 spin_lock(&vcpu->kvm->mmu_lock);
2221                 kvm_mmu_free_some_pages(vcpu);
2222                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2223                                       PT64_ROOT_LEVEL, direct,
2224                                       ACC_ALL, NULL);
2225                 root = __pa(sp->spt);
2226                 ++sp->root_count;
2227                 spin_unlock(&vcpu->kvm->mmu_lock);
2228                 vcpu->arch.mmu.root_hpa = root;
2229                 return 0;
2230         }
2231         direct = !is_paging(vcpu);
2232         for (i = 0; i < 4; ++i) {
2233                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2234
2235                 ASSERT(!VALID_PAGE(root));
2236                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2237                         pdptr = kvm_pdptr_read(vcpu, i);
2238                         if (!is_present_gpte(pdptr)) {
2239                                 vcpu->arch.mmu.pae_root[i] = 0;
2240                                 continue;
2241                         }
2242                         root_gfn = pdptr >> PAGE_SHIFT;
2243                 } else if (vcpu->arch.mmu.root_level == 0)
2244                         root_gfn = 0;
2245                 if (mmu_check_root(vcpu, root_gfn))
2246                         return 1;
2247                 if (tdp_enabled) {
2248                         direct = 1;
2249                         root_gfn = i << 30;
2250                 }
2251                 spin_lock(&vcpu->kvm->mmu_lock);
2252                 kvm_mmu_free_some_pages(vcpu);
2253                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2254                                       PT32_ROOT_LEVEL, direct,
2255                                       ACC_ALL, NULL);
2256                 root = __pa(sp->spt);
2257                 ++sp->root_count;
2258                 spin_unlock(&vcpu->kvm->mmu_lock);
2259
2260                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2261         }
2262         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2263         return 0;
2264 }
2265
2266 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2267 {
2268         int i;
2269         struct kvm_mmu_page *sp;
2270
2271         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2272                 return;
2273         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2274                 hpa_t root = vcpu->arch.mmu.root_hpa;
2275                 sp = page_header(root);
2276                 mmu_sync_children(vcpu, sp);
2277                 return;
2278         }
2279         for (i = 0; i < 4; ++i) {
2280                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2281
2282                 if (root && VALID_PAGE(root)) {
2283                         root &= PT64_BASE_ADDR_MASK;
2284                         sp = page_header(root);
2285                         mmu_sync_children(vcpu, sp);
2286                 }
2287         }
2288 }
2289
2290 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2291 {
2292         spin_lock(&vcpu->kvm->mmu_lock);
2293         mmu_sync_roots(vcpu);
2294         spin_unlock(&vcpu->kvm->mmu_lock);
2295 }
2296
2297 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2298                                   u32 access, u32 *error)
2299 {
2300         if (error)
2301                 *error = 0;
2302         return vaddr;
2303 }
2304
2305 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2306                                 u32 error_code)
2307 {
2308         gfn_t gfn;
2309         int r;
2310
2311         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2312         r = mmu_topup_memory_caches(vcpu);
2313         if (r)
2314                 return r;
2315
2316         ASSERT(vcpu);
2317         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2318
2319         gfn = gva >> PAGE_SHIFT;
2320
2321         return nonpaging_map(vcpu, gva & PAGE_MASK,
2322                              error_code & PFERR_WRITE_MASK, gfn);
2323 }
2324
2325 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2326                                 u32 error_code)
2327 {
2328         pfn_t pfn;
2329         int r;
2330         int level;
2331         gfn_t gfn = gpa >> PAGE_SHIFT;
2332         unsigned long mmu_seq;
2333
2334         ASSERT(vcpu);
2335         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2336
2337         r = mmu_topup_memory_caches(vcpu);
2338         if (r)
2339                 return r;
2340
2341         level = mapping_level(vcpu, gfn);
2342
2343         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2344
2345         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2346         smp_rmb();
2347         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2348         if (is_error_pfn(pfn))
2349                 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2350         spin_lock(&vcpu->kvm->mmu_lock);
2351         if (mmu_notifier_retry(vcpu, mmu_seq))
2352                 goto out_unlock;
2353         kvm_mmu_free_some_pages(vcpu);
2354         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2355                          level, gfn, pfn);
2356         spin_unlock(&vcpu->kvm->mmu_lock);
2357
2358         return r;
2359
2360 out_unlock:
2361         spin_unlock(&vcpu->kvm->mmu_lock);
2362         kvm_release_pfn_clean(pfn);
2363         return 0;
2364 }
2365
2366 static void nonpaging_free(struct kvm_vcpu *vcpu)
2367 {
2368         mmu_free_roots(vcpu);
2369 }
2370
2371 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2372 {
2373         struct kvm_mmu *context = &vcpu->arch.mmu;
2374
2375         context->new_cr3 = nonpaging_new_cr3;
2376         context->page_fault = nonpaging_page_fault;
2377         context->gva_to_gpa = nonpaging_gva_to_gpa;
2378         context->free = nonpaging_free;
2379         context->prefetch_page = nonpaging_prefetch_page;
2380         context->sync_page = nonpaging_sync_page;
2381         context->invlpg = nonpaging_invlpg;
2382         context->root_level = 0;
2383         context->shadow_root_level = PT32E_ROOT_LEVEL;
2384         context->root_hpa = INVALID_PAGE;
2385         return 0;
2386 }
2387
2388 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2389 {
2390         ++vcpu->stat.tlb_flush;
2391         kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
2392 }
2393
2394 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2395 {
2396         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2397         mmu_free_roots(vcpu);
2398 }
2399
2400 static void inject_page_fault(struct kvm_vcpu *vcpu,
2401                               u64 addr,
2402                               u32 err_code)
2403 {
2404         kvm_inject_page_fault(vcpu, addr, err_code);
2405 }
2406
2407 static void paging_free(struct kvm_vcpu *vcpu)
2408 {
2409         nonpaging_free(vcpu);
2410 }
2411
2412 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2413 {
2414         int bit7;
2415
2416         bit7 = (gpte >> 7) & 1;
2417         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2418 }
2419
2420 #define PTTYPE 64
2421 #include "paging_tmpl.h"
2422 #undef PTTYPE
2423
2424 #define PTTYPE 32
2425 #include "paging_tmpl.h"
2426 #undef PTTYPE
2427
2428 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2429 {
2430         struct kvm_mmu *context = &vcpu->arch.mmu;
2431         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2432         u64 exb_bit_rsvd = 0;
2433
2434         if (!is_nx(vcpu))
2435                 exb_bit_rsvd = rsvd_bits(63, 63);
2436         switch (level) {
2437         case PT32_ROOT_LEVEL:
2438                 /* no rsvd bits for 2 level 4K page table entries */
2439                 context->rsvd_bits_mask[0][1] = 0;
2440                 context->rsvd_bits_mask[0][0] = 0;
2441                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2442
2443                 if (!is_pse(vcpu)) {
2444                         context->rsvd_bits_mask[1][1] = 0;
2445                         break;
2446                 }
2447
2448                 if (is_cpuid_PSE36())
2449                         /* 36bits PSE 4MB page */
2450                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2451                 else
2452                         /* 32 bits PSE 4MB page */
2453                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2454                 break;
2455         case PT32E_ROOT_LEVEL:
2456                 context->rsvd_bits_mask[0][2] =
2457                         rsvd_bits(maxphyaddr, 63) |
2458                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2459                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2460                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2461                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2462                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2463                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2464                         rsvd_bits(maxphyaddr, 62) |
2465                         rsvd_bits(13, 20);              /* large page */
2466                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2467                 break;
2468         case PT64_ROOT_LEVEL:
2469                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2470                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2471                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2472                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2473                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2474                         rsvd_bits(maxphyaddr, 51);
2475                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2476                         rsvd_bits(maxphyaddr, 51);
2477                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2478                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2479                         rsvd_bits(maxphyaddr, 51) |
2480                         rsvd_bits(13, 29);
2481                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2482                         rsvd_bits(maxphyaddr, 51) |
2483                         rsvd_bits(13, 20);              /* large page */
2484                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2485                 break;
2486         }
2487 }
2488
2489 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2490 {
2491         struct kvm_mmu *context = &vcpu->arch.mmu;
2492
2493         ASSERT(is_pae(vcpu));
2494         context->new_cr3 = paging_new_cr3;
2495         context->page_fault = paging64_page_fault;
2496         context->gva_to_gpa = paging64_gva_to_gpa;
2497         context->prefetch_page = paging64_prefetch_page;
2498         context->sync_page = paging64_sync_page;
2499         context->invlpg = paging64_invlpg;
2500         context->free = paging_free;
2501         context->root_level = level;
2502         context->shadow_root_level = level;
2503         context->root_hpa = INVALID_PAGE;
2504         return 0;
2505 }
2506
2507 static int paging64_init_context(struct kvm_vcpu *vcpu)
2508 {
2509         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2510         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2511 }
2512
2513 static int paging32_init_context(struct kvm_vcpu *vcpu)
2514 {
2515         struct kvm_mmu *context = &vcpu->arch.mmu;
2516
2517         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2518         context->new_cr3 = paging_new_cr3;
2519         context->page_fault = paging32_page_fault;
2520         context->gva_to_gpa = paging32_gva_to_gpa;
2521         context->free = paging_free;
2522         context->prefetch_page = paging32_prefetch_page;
2523         context->sync_page = paging32_sync_page;
2524         context->invlpg = paging32_invlpg;
2525         context->root_level = PT32_ROOT_LEVEL;
2526         context->shadow_root_level = PT32E_ROOT_LEVEL;
2527         context->root_hpa = INVALID_PAGE;
2528         return 0;
2529 }
2530
2531 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2532 {
2533         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2534         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2535 }
2536
2537 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2538 {
2539         struct kvm_mmu *context = &vcpu->arch.mmu;
2540
2541         context->new_cr3 = nonpaging_new_cr3;
2542         context->page_fault = tdp_page_fault;
2543         context->free = nonpaging_free;
2544         context->prefetch_page = nonpaging_prefetch_page;
2545         context->sync_page = nonpaging_sync_page;
2546         context->invlpg = nonpaging_invlpg;
2547         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2548         context->root_hpa = INVALID_PAGE;
2549
2550         if (!is_paging(vcpu)) {
2551                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2552                 context->root_level = 0;
2553         } else if (is_long_mode(vcpu)) {
2554                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2555                 context->gva_to_gpa = paging64_gva_to_gpa;
2556                 context->root_level = PT64_ROOT_LEVEL;
2557         } else if (is_pae(vcpu)) {
2558                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2559                 context->gva_to_gpa = paging64_gva_to_gpa;
2560                 context->root_level = PT32E_ROOT_LEVEL;
2561         } else {
2562                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2563                 context->gva_to_gpa = paging32_gva_to_gpa;
2564                 context->root_level = PT32_ROOT_LEVEL;
2565         }
2566
2567         return 0;
2568 }
2569
2570 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2571 {
2572         int r;
2573
2574         ASSERT(vcpu);
2575         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2576
2577         if (!is_paging(vcpu))
2578                 r = nonpaging_init_context(vcpu);
2579         else if (is_long_mode(vcpu))
2580                 r = paging64_init_context(vcpu);
2581         else if (is_pae(vcpu))
2582                 r = paging32E_init_context(vcpu);
2583         else
2584                 r = paging32_init_context(vcpu);
2585
2586         vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
2587         vcpu->arch.mmu.base_role.cr0_wp = is_write_protection(vcpu);
2588
2589         return r;
2590 }
2591
2592 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2593 {
2594         vcpu->arch.update_pte.pfn = bad_pfn;
2595
2596         if (tdp_enabled)
2597                 return init_kvm_tdp_mmu(vcpu);
2598         else
2599                 return init_kvm_softmmu(vcpu);
2600 }
2601
2602 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2603 {
2604         ASSERT(vcpu);
2605         if (VALID_PAGE(vcpu->arch.mmu.root_hpa))
2606                 /* mmu.free() should set root_hpa = INVALID_PAGE */
2607                 vcpu->arch.mmu.free(vcpu);
2608 }
2609
2610 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2611 {
2612         destroy_kvm_mmu(vcpu);
2613         return init_kvm_mmu(vcpu);
2614 }
2615 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2616
2617 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2618 {
2619         int r;
2620
2621         r = mmu_topup_memory_caches(vcpu);
2622         if (r)
2623                 goto out;
2624         r = mmu_alloc_roots(vcpu);
2625         spin_lock(&vcpu->kvm->mmu_lock);
2626         mmu_sync_roots(vcpu);
2627         spin_unlock(&vcpu->kvm->mmu_lock);
2628         if (r)
2629                 goto out;
2630         /* set_cr3() should ensure TLB has been flushed */
2631         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2632 out:
2633         return r;
2634 }
2635 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2636
2637 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2638 {
2639         mmu_free_roots(vcpu);
2640 }
2641
2642 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2643                                   struct kvm_mmu_page *sp,
2644                                   u64 *spte)
2645 {
2646         u64 pte;
2647         struct kvm_mmu_page *child;
2648
2649         pte = *spte;
2650         if (is_shadow_present_pte(pte)) {
2651                 if (is_last_spte(pte, sp->role.level))
2652                         drop_spte(vcpu->kvm, spte, shadow_trap_nonpresent_pte);
2653                 else {
2654                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2655                         mmu_page_remove_parent_pte(child, spte);
2656                 }
2657         }
2658         __set_spte(spte, shadow_trap_nonpresent_pte);
2659         if (is_large_pte(pte))
2660                 --vcpu->kvm->stat.lpages;
2661 }
2662
2663 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2664                                   struct kvm_mmu_page *sp,
2665                                   u64 *spte,
2666                                   const void *new)
2667 {
2668         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2669                 ++vcpu->kvm->stat.mmu_pde_zapped;
2670                 return;
2671         }
2672
2673         ++vcpu->kvm->stat.mmu_pte_updated;
2674         if (!sp->role.cr4_pae)
2675                 paging32_update_pte(vcpu, sp, spte, new);
2676         else
2677                 paging64_update_pte(vcpu, sp, spte, new);
2678 }
2679
2680 static bool need_remote_flush(u64 old, u64 new)
2681 {
2682         if (!is_shadow_present_pte(old))
2683                 return false;
2684         if (!is_shadow_present_pte(new))
2685                 return true;
2686         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2687                 return true;
2688         old ^= PT64_NX_MASK;
2689         new ^= PT64_NX_MASK;
2690         return (old & ~new & PT64_PERM_MASK) != 0;
2691 }
2692
2693 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, bool zap_page,
2694                                     bool remote_flush, bool local_flush)
2695 {
2696         if (zap_page)
2697                 return;
2698
2699         if (remote_flush)
2700                 kvm_flush_remote_tlbs(vcpu->kvm);
2701         else if (local_flush)
2702                 kvm_mmu_flush_tlb(vcpu);
2703 }
2704
2705 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2706 {
2707         u64 *spte = vcpu->arch.last_pte_updated;
2708
2709         return !!(spte && (*spte & shadow_accessed_mask));
2710 }
2711
2712 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2713                                           u64 gpte)
2714 {
2715         gfn_t gfn;
2716         pfn_t pfn;
2717
2718         if (!is_present_gpte(gpte))
2719                 return;
2720         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2721
2722         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2723         smp_rmb();
2724         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2725
2726         if (is_error_pfn(pfn)) {
2727                 kvm_release_pfn_clean(pfn);
2728                 return;
2729         }
2730         vcpu->arch.update_pte.gfn = gfn;
2731         vcpu->arch.update_pte.pfn = pfn;
2732 }
2733
2734 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2735 {
2736         u64 *spte = vcpu->arch.last_pte_updated;
2737
2738         if (spte
2739             && vcpu->arch.last_pte_gfn == gfn
2740             && shadow_accessed_mask
2741             && !(*spte & shadow_accessed_mask)
2742             && is_shadow_present_pte(*spte))
2743                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2744 }
2745
2746 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2747                        const u8 *new, int bytes,
2748                        bool guest_initiated)
2749 {
2750         gfn_t gfn = gpa >> PAGE_SHIFT;
2751         struct kvm_mmu_page *sp;
2752         struct hlist_node *node;
2753         LIST_HEAD(invalid_list);
2754         u64 entry, gentry;
2755         u64 *spte;
2756         unsigned offset = offset_in_page(gpa);
2757         unsigned pte_size;
2758         unsigned page_offset;
2759         unsigned misaligned;
2760         unsigned quadrant;
2761         int level;
2762         int flooded = 0;
2763         int npte;
2764         int r;
2765         int invlpg_counter;
2766         bool remote_flush, local_flush, zap_page;
2767
2768         zap_page = remote_flush = local_flush = false;
2769
2770         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2771
2772         invlpg_counter = atomic_read(&vcpu->kvm->arch.invlpg_counter);
2773
2774         /*
2775          * Assume that the pte write on a page table of the same type
2776          * as the current vcpu paging mode.  This is nearly always true
2777          * (might be false while changing modes).  Note it is verified later
2778          * by update_pte().
2779          */
2780         if ((is_pae(vcpu) && bytes == 4) || !new) {
2781                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2782                 if (is_pae(vcpu)) {
2783                         gpa &= ~(gpa_t)7;
2784                         bytes = 8;
2785                 }
2786                 r = kvm_read_guest(vcpu->kvm, gpa, &gentry, min(bytes, 8));
2787                 if (r)
2788                         gentry = 0;
2789                 new = (const u8 *)&gentry;
2790         }
2791
2792         switch (bytes) {
2793         case 4:
2794                 gentry = *(const u32 *)new;
2795                 break;
2796         case 8:
2797                 gentry = *(const u64 *)new;
2798                 break;
2799         default:
2800                 gentry = 0;
2801                 break;
2802         }
2803
2804         mmu_guess_page_from_pte_write(vcpu, gpa, gentry);
2805         spin_lock(&vcpu->kvm->mmu_lock);
2806         if (atomic_read(&vcpu->kvm->arch.invlpg_counter) != invlpg_counter)
2807                 gentry = 0;
2808         kvm_mmu_access_page(vcpu, gfn);
2809         kvm_mmu_free_some_pages(vcpu);
2810         ++vcpu->kvm->stat.mmu_pte_write;
2811         kvm_mmu_audit(vcpu, "pre pte write");
2812         if (guest_initiated) {
2813                 if (gfn == vcpu->arch.last_pt_write_gfn
2814                     && !last_updated_pte_accessed(vcpu)) {
2815                         ++vcpu->arch.last_pt_write_count;
2816                         if (vcpu->arch.last_pt_write_count >= 3)
2817                                 flooded = 1;
2818                 } else {
2819                         vcpu->arch.last_pt_write_gfn = gfn;
2820                         vcpu->arch.last_pt_write_count = 1;
2821                         vcpu->arch.last_pte_updated = NULL;
2822                 }
2823         }
2824
2825         for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn, node) {
2826                 pte_size = sp->role.cr4_pae ? 8 : 4;
2827                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2828                 misaligned |= bytes < 4;
2829                 if (misaligned || flooded) {
2830                         /*
2831                          * Misaligned accesses are too much trouble to fix
2832                          * up; also, they usually indicate a page is not used
2833                          * as a page table.
2834                          *
2835                          * If we're seeing too many writes to a page,
2836                          * it may no longer be a page table, or we may be
2837                          * forking, in which case it is better to unmap the
2838                          * page.
2839                          */
2840                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2841                                  gpa, bytes, sp->role.word);
2842                         zap_page |= !!kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2843                                                      &invalid_list);
2844                         ++vcpu->kvm->stat.mmu_flooded;
2845                         continue;
2846                 }
2847                 page_offset = offset;
2848                 level = sp->role.level;
2849                 npte = 1;
2850                 if (!sp->role.cr4_pae) {
2851                         page_offset <<= 1;      /* 32->64 */
2852                         /*
2853                          * A 32-bit pde maps 4MB while the shadow pdes map
2854                          * only 2MB.  So we need to double the offset again
2855                          * and zap two pdes instead of one.
2856                          */
2857                         if (level == PT32_ROOT_LEVEL) {
2858                                 page_offset &= ~7; /* kill rounding error */
2859                                 page_offset <<= 1;
2860                                 npte = 2;
2861                         }
2862                         quadrant = page_offset >> PAGE_SHIFT;
2863                         page_offset &= ~PAGE_MASK;
2864                         if (quadrant != sp->role.quadrant)
2865                                 continue;
2866                 }
2867                 local_flush = true;
2868                 spte = &sp->spt[page_offset / sizeof(*spte)];
2869                 while (npte--) {
2870                         entry = *spte;
2871                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2872                         if (gentry)
2873                                 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
2874                         if (!remote_flush && need_remote_flush(entry, *spte))
2875                                 remote_flush = true;
2876                         ++spte;
2877                 }
2878         }
2879         mmu_pte_write_flush_tlb(vcpu, zap_page, remote_flush, local_flush);
2880         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2881         kvm_mmu_audit(vcpu, "post pte write");
2882         spin_unlock(&vcpu->kvm->mmu_lock);
2883         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2884                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2885                 vcpu->arch.update_pte.pfn = bad_pfn;
2886         }
2887 }
2888
2889 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2890 {
2891         gpa_t gpa;
2892         int r;
2893
2894         if (tdp_enabled)
2895                 return 0;
2896
2897         gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2898
2899         spin_lock(&vcpu->kvm->mmu_lock);
2900         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2901         spin_unlock(&vcpu->kvm->mmu_lock);
2902         return r;
2903 }
2904 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2905
2906 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2907 {
2908         int free_pages;
2909         LIST_HEAD(invalid_list);
2910
2911         free_pages = vcpu->kvm->arch.n_free_mmu_pages;
2912         while (free_pages < KVM_REFILL_PAGES &&
2913                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2914                 struct kvm_mmu_page *sp;
2915
2916                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2917                                   struct kvm_mmu_page, link);
2918                 free_pages += kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2919                                                        &invalid_list);
2920                 ++vcpu->kvm->stat.mmu_recycled;
2921         }
2922         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2923 }
2924
2925 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2926 {
2927         int r;
2928         enum emulation_result er;
2929
2930         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2931         if (r < 0)
2932                 goto out;
2933
2934         if (!r) {
2935                 r = 1;
2936                 goto out;
2937         }
2938
2939         r = mmu_topup_memory_caches(vcpu);
2940         if (r)
2941                 goto out;
2942
2943         er = emulate_instruction(vcpu, cr2, error_code, 0);
2944
2945         switch (er) {
2946         case EMULATE_DONE:
2947                 return 1;
2948         case EMULATE_DO_MMIO:
2949                 ++vcpu->stat.mmio_exits;
2950                 /* fall through */
2951         case EMULATE_FAIL:
2952                 return 0;
2953         default:
2954                 BUG();
2955         }
2956 out:
2957         return r;
2958 }
2959 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2960
2961 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2962 {
2963         vcpu->arch.mmu.invlpg(vcpu, gva);
2964         kvm_mmu_flush_tlb(vcpu);
2965         ++vcpu->stat.invlpg;
2966 }
2967 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2968
2969 void kvm_enable_tdp(void)
2970 {
2971         tdp_enabled = true;
2972 }
2973 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2974
2975 void kvm_disable_tdp(void)
2976 {
2977         tdp_enabled = false;
2978 }
2979 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2980
2981 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2982 {
2983         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2984 }
2985
2986 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2987 {
2988         struct page *page;
2989         int i;
2990
2991         ASSERT(vcpu);
2992
2993         /*
2994          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2995          * Therefore we need to allocate shadow page tables in the first
2996          * 4GB of memory, which happens to fit the DMA32 zone.
2997          */
2998         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2999         if (!page)
3000                 return -ENOMEM;
3001
3002         vcpu->arch.mmu.pae_root = page_address(page);
3003         for (i = 0; i < 4; ++i)
3004                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
3005
3006         return 0;
3007 }
3008
3009 int kvm_mmu_create(struct kvm_vcpu *vcpu)
3010 {
3011         ASSERT(vcpu);
3012         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3013
3014         return alloc_mmu_pages(vcpu);
3015 }
3016
3017 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
3018 {
3019         ASSERT(vcpu);
3020         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3021
3022         return init_kvm_mmu(vcpu);
3023 }
3024
3025 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
3026 {
3027         ASSERT(vcpu);
3028
3029         destroy_kvm_mmu(vcpu);
3030         free_mmu_pages(vcpu);
3031         mmu_free_memory_caches(vcpu);
3032 }
3033
3034 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
3035 {
3036         struct kvm_mmu_page *sp;
3037
3038         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
3039                 int i;
3040                 u64 *pt;
3041
3042                 if (!test_bit(slot, sp->slot_bitmap))
3043                         continue;
3044
3045                 pt = sp->spt;
3046                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
3047                         /* avoid RMW */
3048                         if (is_writable_pte(pt[i]))
3049                                 pt[i] &= ~PT_WRITABLE_MASK;
3050         }
3051         kvm_flush_remote_tlbs(kvm);
3052 }
3053
3054 void kvm_mmu_zap_all(struct kvm *kvm)
3055 {
3056         struct kvm_mmu_page *sp, *node;
3057         LIST_HEAD(invalid_list);
3058
3059         spin_lock(&kvm->mmu_lock);
3060 restart:
3061         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
3062                 if (kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list))
3063                         goto restart;
3064
3065         kvm_mmu_commit_zap_page(kvm, &invalid_list);
3066         spin_unlock(&kvm->mmu_lock);
3067 }
3068
3069 static int kvm_mmu_remove_some_alloc_mmu_pages(struct kvm *kvm,
3070                                                struct list_head *invalid_list)
3071 {
3072         struct kvm_mmu_page *page;
3073
3074         page = container_of(kvm->arch.active_mmu_pages.prev,
3075                             struct kvm_mmu_page, link);
3076         return kvm_mmu_prepare_zap_page(kvm, page, invalid_list);
3077 }
3078
3079 static int mmu_shrink(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
3080 {
3081         struct kvm *kvm;
3082         struct kvm *kvm_freed = NULL;
3083         int cache_count = 0;
3084
3085         spin_lock(&kvm_lock);
3086
3087         list_for_each_entry(kvm, &vm_list, vm_list) {
3088                 int npages, idx, freed_pages;
3089                 LIST_HEAD(invalid_list);
3090
3091                 idx = srcu_read_lock(&kvm->srcu);
3092                 spin_lock(&kvm->mmu_lock);
3093                 npages = kvm->arch.n_alloc_mmu_pages -
3094                          kvm->arch.n_free_mmu_pages;
3095                 cache_count += npages;
3096                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
3097                         freed_pages = kvm_mmu_remove_some_alloc_mmu_pages(kvm,
3098                                                           &invalid_list);
3099                         cache_count -= freed_pages;
3100                         kvm_freed = kvm;
3101                 }
3102                 nr_to_scan--;
3103
3104                 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3105                 spin_unlock(&kvm->mmu_lock);
3106                 srcu_read_unlock(&kvm->srcu, idx);
3107         }
3108         if (kvm_freed)
3109                 list_move_tail(&kvm_freed->vm_list, &vm_list);
3110
3111         spin_unlock(&kvm_lock);
3112
3113         return cache_count;
3114 }
3115
3116 static struct shrinker mmu_shrinker = {
3117         .shrink = mmu_shrink,
3118         .seeks = DEFAULT_SEEKS * 10,
3119 };
3120
3121 static void mmu_destroy_caches(void)
3122 {
3123         if (pte_chain_cache)
3124                 kmem_cache_destroy(pte_chain_cache);
3125         if (rmap_desc_cache)
3126                 kmem_cache_destroy(rmap_desc_cache);
3127         if (mmu_page_header_cache)
3128                 kmem_cache_destroy(mmu_page_header_cache);
3129 }
3130
3131 void kvm_mmu_module_exit(void)
3132 {
3133         mmu_destroy_caches();
3134         unregister_shrinker(&mmu_shrinker);
3135 }
3136
3137 int kvm_mmu_module_init(void)
3138 {
3139         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
3140                                             sizeof(struct kvm_pte_chain),
3141                                             0, 0, NULL);
3142         if (!pte_chain_cache)
3143                 goto nomem;
3144         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
3145                                             sizeof(struct kvm_rmap_desc),
3146                                             0, 0, NULL);
3147         if (!rmap_desc_cache)
3148                 goto nomem;
3149
3150         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
3151                                                   sizeof(struct kvm_mmu_page),
3152                                                   0, 0, NULL);
3153         if (!mmu_page_header_cache)
3154                 goto nomem;
3155
3156         register_shrinker(&mmu_shrinker);
3157
3158         return 0;
3159
3160 nomem:
3161         mmu_destroy_caches();
3162         return -ENOMEM;
3163 }
3164
3165 /*
3166  * Caculate mmu pages needed for kvm.
3167  */
3168 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3169 {
3170         int i;
3171         unsigned int nr_mmu_pages;
3172         unsigned int  nr_pages = 0;
3173         struct kvm_memslots *slots;
3174
3175         slots = kvm_memslots(kvm);
3176
3177         for (i = 0; i < slots->nmemslots; i++)
3178                 nr_pages += slots->memslots[i].npages;
3179
3180         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3181         nr_mmu_pages = max(nr_mmu_pages,
3182                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3183
3184         return nr_mmu_pages;
3185 }
3186
3187 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3188                                 unsigned len)
3189 {
3190         if (len > buffer->len)
3191                 return NULL;
3192         return buffer->ptr;
3193 }
3194
3195 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3196                                 unsigned len)
3197 {
3198         void *ret;
3199
3200         ret = pv_mmu_peek_buffer(buffer, len);
3201         if (!ret)
3202                 return ret;
3203         buffer->ptr += len;
3204         buffer->len -= len;
3205         buffer->processed += len;
3206         return ret;
3207 }
3208
3209 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3210                              gpa_t addr, gpa_t value)
3211 {
3212         int bytes = 8;
3213         int r;
3214
3215         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3216                 bytes = 4;
3217
3218         r = mmu_topup_memory_caches(vcpu);
3219         if (r)
3220                 return r;
3221
3222         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3223                 return -EFAULT;
3224
3225         return 1;
3226 }
3227
3228 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3229 {
3230         (void)kvm_set_cr3(vcpu, vcpu->arch.cr3);
3231         return 1;
3232 }
3233
3234 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3235 {
3236         spin_lock(&vcpu->kvm->mmu_lock);
3237         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3238         spin_unlock(&vcpu->kvm->mmu_lock);
3239         return 1;
3240 }
3241
3242 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3243                              struct kvm_pv_mmu_op_buffer *buffer)
3244 {
3245         struct kvm_mmu_op_header *header;
3246
3247         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3248         if (!header)
3249                 return 0;
3250         switch (header->op) {
3251         case KVM_MMU_OP_WRITE_PTE: {
3252                 struct kvm_mmu_op_write_pte *wpte;
3253
3254                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3255                 if (!wpte)
3256                         return 0;
3257                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3258                                         wpte->pte_val);
3259         }
3260         case KVM_MMU_OP_FLUSH_TLB: {
3261                 struct kvm_mmu_op_flush_tlb *ftlb;
3262
3263                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3264                 if (!ftlb)
3265                         return 0;
3266                 return kvm_pv_mmu_flush_tlb(vcpu);
3267         }
3268         case KVM_MMU_OP_RELEASE_PT: {
3269                 struct kvm_mmu_op_release_pt *rpt;
3270
3271                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3272                 if (!rpt)
3273                         return 0;
3274                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3275         }
3276         default: return 0;
3277         }
3278 }
3279
3280 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3281                   gpa_t addr, unsigned long *ret)
3282 {
3283         int r;
3284         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3285
3286         buffer->ptr = buffer->buf;
3287         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3288         buffer->processed = 0;
3289
3290         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3291         if (r)
3292                 goto out;
3293
3294         while (buffer->len) {
3295                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3296                 if (r < 0)
3297                         goto out;
3298                 if (r == 0)
3299                         break;
3300         }
3301
3302         r = 1;
3303 out:
3304         *ret = buffer->processed;
3305         return r;
3306 }
3307
3308 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3309 {
3310         struct kvm_shadow_walk_iterator iterator;
3311         int nr_sptes = 0;
3312
3313         spin_lock(&vcpu->kvm->mmu_lock);
3314         for_each_shadow_entry(vcpu, addr, iterator) {
3315                 sptes[iterator.level-1] = *iterator.sptep;
3316                 nr_sptes++;
3317                 if (!is_shadow_present_pte(*iterator.sptep))
3318                         break;
3319         }
3320         spin_unlock(&vcpu->kvm->mmu_lock);
3321
3322         return nr_sptes;
3323 }
3324 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3325
3326 #ifdef AUDIT
3327
3328 static const char *audit_msg;
3329
3330 static gva_t canonicalize(gva_t gva)
3331 {
3332 #ifdef CONFIG_X86_64
3333         gva = (long long)(gva << 16) >> 16;
3334 #endif
3335         return gva;
3336 }
3337
3338
3339 typedef void (*inspect_spte_fn) (struct kvm *kvm, u64 *sptep);
3340
3341 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3342                             inspect_spte_fn fn)
3343 {
3344         int i;
3345
3346         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3347                 u64 ent = sp->spt[i];
3348
3349                 if (is_shadow_present_pte(ent)) {
3350                         if (!is_last_spte(ent, sp->role.level)) {
3351                                 struct kvm_mmu_page *child;
3352                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3353                                 __mmu_spte_walk(kvm, child, fn);
3354                         } else
3355                                 fn(kvm, &sp->spt[i]);
3356                 }
3357         }
3358 }
3359
3360 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3361 {
3362         int i;
3363         struct kvm_mmu_page *sp;
3364
3365         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3366                 return;
3367         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3368                 hpa_t root = vcpu->arch.mmu.root_hpa;
3369                 sp = page_header(root);
3370                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3371                 return;
3372         }
3373         for (i = 0; i < 4; ++i) {
3374                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3375
3376                 if (root && VALID_PAGE(root)) {
3377                         root &= PT64_BASE_ADDR_MASK;
3378                         sp = page_header(root);
3379                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3380                 }
3381         }
3382         return;
3383 }
3384
3385 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3386                                 gva_t va, int level)
3387 {
3388         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3389         int i;
3390         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3391
3392         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3393                 u64 ent = pt[i];
3394
3395                 if (ent == shadow_trap_nonpresent_pte)
3396                         continue;
3397
3398                 va = canonicalize(va);
3399                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3400                         audit_mappings_page(vcpu, ent, va, level - 1);
3401                 else {
3402                         gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, va, NULL);
3403                         gfn_t gfn = gpa >> PAGE_SHIFT;
3404                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3405                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3406
3407                         if (is_error_pfn(pfn)) {
3408                                 kvm_release_pfn_clean(pfn);
3409                                 continue;
3410                         }
3411
3412                         if (is_shadow_present_pte(ent)
3413                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3414                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3415                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3416                                        audit_msg, vcpu->arch.mmu.root_level,
3417                                        va, gpa, hpa, ent,
3418                                        is_shadow_present_pte(ent));
3419                         else if (ent == shadow_notrap_nonpresent_pte
3420                                  && !is_error_hpa(hpa))
3421                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3422                                        " valid guest gva %lx\n", audit_msg, va);
3423                         kvm_release_pfn_clean(pfn);
3424
3425                 }
3426         }
3427 }
3428
3429 static void audit_mappings(struct kvm_vcpu *vcpu)
3430 {
3431         unsigned i;
3432
3433         if (vcpu->arch.mmu.root_level == 4)
3434                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3435         else
3436                 for (i = 0; i < 4; ++i)
3437                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3438                                 audit_mappings_page(vcpu,
3439                                                     vcpu->arch.mmu.pae_root[i],
3440                                                     i << 30,
3441                                                     2);
3442 }
3443
3444 static int count_rmaps(struct kvm_vcpu *vcpu)
3445 {
3446         struct kvm *kvm = vcpu->kvm;
3447         struct kvm_memslots *slots;
3448         int nmaps = 0;
3449         int i, j, k, idx;
3450
3451         idx = srcu_read_lock(&kvm->srcu);
3452         slots = kvm_memslots(kvm);
3453         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3454                 struct kvm_memory_slot *m = &slots->memslots[i];
3455                 struct kvm_rmap_desc *d;
3456
3457                 for (j = 0; j < m->npages; ++j) {
3458                         unsigned long *rmapp = &m->rmap[j];
3459
3460                         if (!*rmapp)
3461                                 continue;
3462                         if (!(*rmapp & 1)) {
3463                                 ++nmaps;
3464                                 continue;
3465                         }
3466                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3467                         while (d) {
3468                                 for (k = 0; k < RMAP_EXT; ++k)
3469                                         if (d->sptes[k])
3470                                                 ++nmaps;
3471                                         else
3472                                                 break;
3473                                 d = d->more;
3474                         }
3475                 }
3476         }
3477         srcu_read_unlock(&kvm->srcu, idx);
3478         return nmaps;
3479 }
3480
3481 void inspect_spte_has_rmap(struct kvm *kvm, u64 *sptep)
3482 {
3483         unsigned long *rmapp;
3484         struct kvm_mmu_page *rev_sp;
3485         gfn_t gfn;
3486
3487         if (is_writable_pte(*sptep)) {
3488                 rev_sp = page_header(__pa(sptep));
3489                 gfn = kvm_mmu_page_get_gfn(rev_sp, sptep - rev_sp->spt);
3490
3491                 if (!gfn_to_memslot(kvm, gfn)) {
3492                         if (!printk_ratelimit())
3493                                 return;
3494                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3495                                          audit_msg, gfn);
3496                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3497                                audit_msg, (long int)(sptep - rev_sp->spt),
3498                                         rev_sp->gfn);
3499                         dump_stack();
3500                         return;
3501                 }
3502
3503                 rmapp = gfn_to_rmap(kvm, gfn, rev_sp->role.level);
3504                 if (!*rmapp) {
3505                         if (!printk_ratelimit())
3506                                 return;
3507                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3508                                          audit_msg, *sptep);
3509                         dump_stack();
3510                 }
3511         }
3512
3513 }
3514
3515 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3516 {
3517         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3518 }
3519
3520 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3521 {
3522         struct kvm_mmu_page *sp;
3523         int i;
3524
3525         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3526                 u64 *pt = sp->spt;
3527
3528                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3529                         continue;
3530
3531                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3532                         u64 ent = pt[i];
3533
3534                         if (!(ent & PT_PRESENT_MASK))
3535                                 continue;
3536                         if (!is_writable_pte(ent))
3537                                 continue;
3538                         inspect_spte_has_rmap(vcpu->kvm, &pt[i]);
3539                 }
3540         }
3541         return;
3542 }
3543
3544 static void audit_rmap(struct kvm_vcpu *vcpu)
3545 {
3546         check_writable_mappings_rmap(vcpu);
3547         count_rmaps(vcpu);
3548 }
3549
3550 static void audit_write_protection(struct kvm_vcpu *vcpu)
3551 {
3552         struct kvm_mmu_page *sp;
3553         struct kvm_memory_slot *slot;
3554         unsigned long *rmapp;
3555         u64 *spte;
3556         gfn_t gfn;
3557
3558         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3559                 if (sp->role.direct)
3560                         continue;
3561                 if (sp->unsync)
3562                         continue;
3563
3564                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
3565                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3566
3567                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3568                 while (spte) {
3569                         if (is_writable_pte(*spte))
3570                                 printk(KERN_ERR "%s: (%s) shadow page has "
3571                                 "writable mappings: gfn %lx role %x\n",
3572                                __func__, audit_msg, sp->gfn,
3573                                sp->role.word);
3574                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3575                 }
3576         }
3577 }
3578
3579 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3580 {
3581         int olddbg = dbg;
3582
3583         dbg = 0;
3584         audit_msg = msg;
3585         audit_rmap(vcpu);
3586         audit_write_protection(vcpu);
3587         if (strcmp("pre pte write", audit_msg) != 0)
3588                 audit_mappings(vcpu);
3589         audit_writable_sptes_have_rmaps(vcpu);
3590         dbg = olddbg;
3591 }
3592
3593 #endif