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

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "vmx.h"
  21 #include "kvm.h"
  22 #include "x86.h"
  23
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29
  30 #include <asm/page.h>
  31 #include <asm/cmpxchg.h>
  32
  33 #undef MMU_DEBUG
  34
  35 #undef AUDIT
  36
  37 #ifdef AUDIT
  38 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  39 #else
  40 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  41 #endif
  42
  43 #ifdef MMU_DEBUG
  44
  45 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  46 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  47
  48 #else
  49
  50 #define pgprintk(x...) do { } while (0)
  51 #define rmap_printk(x...) do { } while (0)
  52
  53 #endif
  54
  55 #if defined(MMU_DEBUG) || defined(AUDIT)
  56 static int dbg = 1;
  57 #endif
  58
  59 #ifndef MMU_DEBUG
  60 #define ASSERT(x) do { } while (0)
  61 #else
  62 #define ASSERT(x)                                                       \
  63         if (!(x)) {                                                     \
  64                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  65                        __FILE__, __LINE__, #x);                         \
  66         }
  67 #endif
  68
  69 #define PT64_PT_BITS 9
  70 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  71 #define PT32_PT_BITS 10
  72 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  73
  74 #define PT_WRITABLE_SHIFT 1
  75
  76 #define PT_PRESENT_MASK (1ULL << 0)
  77 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  78 #define PT_USER_MASK (1ULL << 2)
  79 #define PT_PWT_MASK (1ULL << 3)
  80 #define PT_PCD_MASK (1ULL << 4)
  81 #define PT_ACCESSED_MASK (1ULL << 5)
  82 #define PT_DIRTY_MASK (1ULL << 6)
  83 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  84 #define PT_PAT_MASK (1ULL << 7)
  85 #define PT_GLOBAL_MASK (1ULL << 8)
  86 #define PT64_NX_MASK (1ULL << 63)
  87
  88 #define PT_PAT_SHIFT 7
  89 #define PT_DIR_PAT_SHIFT 12
  90 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  91
  92 #define PT32_DIR_PSE36_SIZE 4
  93 #define PT32_DIR_PSE36_SHIFT 13
  94 #define PT32_DIR_PSE36_MASK \
  95         (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  96
  97
  98 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  99 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
 100
 101 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 102
 103 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 104
 105 #define PT64_LEVEL_BITS 9
 106
 107 #define PT64_LEVEL_SHIFT(level) \
 108                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
 109
 110 #define PT64_LEVEL_MASK(level) \
 111                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 112
 113 #define PT64_INDEX(address, level)\
 114         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 115
 116
 117 #define PT32_LEVEL_BITS 10
 118
 119 #define PT32_LEVEL_SHIFT(level) \
 120                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 121
 122 #define PT32_LEVEL_MASK(level) \
 123                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 124
 125 #define PT32_INDEX(address, level)\
 126         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 127
 128
 129 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 130 #define PT64_DIR_BASE_ADDR_MASK \
 131         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 132
 133 #define PT32_BASE_ADDR_MASK PAGE_MASK
 134 #define PT32_DIR_BASE_ADDR_MASK \
 135         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 136
 137
 138 #define PFERR_PRESENT_MASK (1U << 0)
 139 #define PFERR_WRITE_MASK (1U << 1)
 140 #define PFERR_USER_MASK (1U << 2)
 141 #define PFERR_FETCH_MASK (1U << 4)
 142
 143 #define PT64_ROOT_LEVEL 4
 144 #define PT32_ROOT_LEVEL 2
 145 #define PT32E_ROOT_LEVEL 3
 146
 147 #define PT_DIRECTORY_LEVEL 2
 148 #define PT_PAGE_TABLE_LEVEL 1
 149
 150 #define RMAP_EXT 4
 151
 152 struct kvm_rmap_desc {
 153         u64 *shadow_ptes[RMAP_EXT];
 154         struct kvm_rmap_desc *more;
 155 };
 156
 157 static struct kmem_cache *pte_chain_cache;
 158 static struct kmem_cache *rmap_desc_cache;
 159 static struct kmem_cache *mmu_page_header_cache;
 160
 161 static u64 __read_mostly shadow_trap_nonpresent_pte;
 162 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 163
 164 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 165 {
 166         shadow_trap_nonpresent_pte = trap_pte;
 167         shadow_notrap_nonpresent_pte = notrap_pte;
 168 }
 169 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 170
 171 static int is_write_protection(struct kvm_vcpu *vcpu)
 172 {
 173         return vcpu->cr0 & X86_CR0_WP;
 174 }
 175
 176 static int is_cpuid_PSE36(void)
 177 {
 178         return 1;
 179 }
 180
 181 static int is_nx(struct kvm_vcpu *vcpu)
 182 {
 183         return vcpu->shadow_efer & EFER_NX;
 184 }
 185
 186 static int is_present_pte(unsigned long pte)
 187 {
 188         return pte & PT_PRESENT_MASK;
 189 }
 190
 191 static int is_shadow_present_pte(u64 pte)
 192 {
 193         pte &= ~PT_SHADOW_IO_MARK;
 194         return pte != shadow_trap_nonpresent_pte
 195                 && pte != shadow_notrap_nonpresent_pte;
 196 }
 197
 198 static int is_writeble_pte(unsigned long pte)
 199 {
 200         return pte & PT_WRITABLE_MASK;
 201 }
 202
 203 static int is_dirty_pte(unsigned long pte)
 204 {
 205         return pte & PT_DIRTY_MASK;
 206 }
 207
 208 static int is_io_pte(unsigned long pte)
 209 {
 210         return pte & PT_SHADOW_IO_MARK;
 211 }
 212
 213 static int is_rmap_pte(u64 pte)
 214 {
 215         return pte != shadow_trap_nonpresent_pte
 216                 && pte != shadow_notrap_nonpresent_pte;
 217 }
 218
 219 static void set_shadow_pte(u64 *sptep, u64 spte)
 220 {
 221 #ifdef CONFIG_X86_64
 222         set_64bit((unsigned long *)sptep, spte);
 223 #else
 224         set_64bit((unsigned long long *)sptep, spte);
 225 #endif
 226 }
 227
 228 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 229                                   struct kmem_cache *base_cache, int min)
 230 {
 231         void *obj;
 232
 233         if (cache->nobjs >= min)
 234                 return 0;
 235         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 236                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 237                 if (!obj)
 238                         return -ENOMEM;
 239                 cache->objects[cache->nobjs++] = obj;
 240         }
 241         return 0;
 242 }
 243
 244 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 245 {
 246         while (mc->nobjs)
 247                 kfree(mc->objects[--mc->nobjs]);
 248 }
 249
 250 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 251                                        int min)
 252 {
 253         struct page *page;
 254
 255         if (cache->nobjs >= min)
 256                 return 0;
 257         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 258                 page = alloc_page(GFP_KERNEL);
 259                 if (!page)
 260                         return -ENOMEM;
 261                 set_page_private(page, 0);
 262                 cache->objects[cache->nobjs++] = page_address(page);
 263         }
 264         return 0;
 265 }
 266
 267 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 268 {
 269         while (mc->nobjs)
 270                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 271 }
 272
 273 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 274 {
 275         int r;
 276
 277         kvm_mmu_free_some_pages(vcpu);
 278         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 279                                    pte_chain_cache, 4);
 280         if (r)
 281                 goto out;
 282         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 283                                    rmap_desc_cache, 1);
 284         if (r)
 285                 goto out;
 286         r = mmu_topup_memory_cache_page(&vcpu->mmu_page_cache, 8);
 287         if (r)
 288                 goto out;
 289         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 290                                    mmu_page_header_cache, 4);
 291 out:
 292         return r;
 293 }
 294
 295 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 296 {
 297         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 298         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 299         mmu_free_memory_cache_page(&vcpu->mmu_page_cache);
 300         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 301 }
 302
 303 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 304                                     size_t size)
 305 {
 306         void *p;
 307
 308         BUG_ON(!mc->nobjs);
 309         p = mc->objects[--mc->nobjs];
 310         memset(p, 0, size);
 311         return p;
 312 }
 313
 314 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 315 {
 316         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 317                                       sizeof(struct kvm_pte_chain));
 318 }
 319
 320 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 321 {
 322         kfree(pc);
 323 }
 324
 325 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 326 {
 327         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 328                                       sizeof(struct kvm_rmap_desc));
 329 }
 330
 331 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 332 {
 333         kfree(rd);
 334 }
 335
 336 /*
 337  * Take gfn and return the reverse mapping to it.
 338  * Note: gfn must be unaliased before this function get called
 339  */
 340
 341 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
 342 {
 343         struct kvm_memory_slot *slot;
 344
 345         slot = gfn_to_memslot(kvm, gfn);
 346         return &slot->rmap[gfn - slot->base_gfn];
 347 }
 348
 349 /*
 350  * Reverse mapping data structures:
 351  *
 352  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 353  * that points to page_address(page).
 354  *
 355  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 356  * containing more mappings.
 357  */
 358 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 359 {
 360         struct kvm_mmu_page *page;
 361         struct kvm_rmap_desc *desc;
 362         unsigned long *rmapp;
 363         int i;
 364
 365         if (!is_rmap_pte(*spte))
 366                 return;
 367         gfn = unalias_gfn(vcpu->kvm, gfn);
 368         page = page_header(__pa(spte));
 369         page->gfns[spte - page->spt] = gfn;
 370         rmapp = gfn_to_rmap(vcpu->kvm, gfn);
 371         if (!*rmapp) {
 372                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 373                 *rmapp = (unsigned long)spte;
 374         } else if (!(*rmapp & 1)) {
 375                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 376                 desc = mmu_alloc_rmap_desc(vcpu);
 377                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 378                 desc->shadow_ptes[1] = spte;
 379                 *rmapp = (unsigned long)desc | 1;
 380         } else {
 381                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 382                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 383                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 384                         desc = desc->more;
 385                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 386                         desc->more = mmu_alloc_rmap_desc(vcpu);
 387                         desc = desc->more;
 388                 }
 389                 for (i = 0; desc->shadow_ptes[i]; ++i)
 390                         ;
 391                 desc->shadow_ptes[i] = spte;
 392         }
 393 }
 394
 395 static void rmap_desc_remove_entry(unsigned long *rmapp,
 396                                    struct kvm_rmap_desc *desc,
 397                                    int i,
 398                                    struct kvm_rmap_desc *prev_desc)
 399 {
 400         int j;
 401
 402         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 403                 ;
 404         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 405         desc->shadow_ptes[j] = NULL;
 406         if (j != 0)
 407                 return;
 408         if (!prev_desc && !desc->more)
 409                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 410         else
 411                 if (prev_desc)
 412                         prev_desc->more = desc->more;
 413                 else
 414                         *rmapp = (unsigned long)desc->more | 1;
 415         mmu_free_rmap_desc(desc);
 416 }
 417
 418 static void rmap_remove(struct kvm *kvm, u64 *spte)
 419 {
 420         struct kvm_rmap_desc *desc;
 421         struct kvm_rmap_desc *prev_desc;
 422         struct kvm_mmu_page *page;
 423         struct page *release_page;
 424         unsigned long *rmapp;
 425         int i;
 426
 427         if (!is_rmap_pte(*spte))
 428                 return;
 429         page = page_header(__pa(spte));
 430         release_page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 431         if (is_writeble_pte(*spte))
 432                 kvm_release_page_dirty(release_page);
 433         else
 434                 kvm_release_page_clean(release_page);
 435         rmapp = gfn_to_rmap(kvm, page->gfns[spte - page->spt]);
 436         if (!*rmapp) {
 437                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 438                 BUG();
 439         } else if (!(*rmapp & 1)) {
 440                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 441                 if ((u64 *)*rmapp != spte) {
 442                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 443                                spte, *spte);
 444                         BUG();
 445                 }
 446                 *rmapp = 0;
 447         } else {
 448                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 449                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 450                 prev_desc = NULL;
 451                 while (desc) {
 452                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 453                                 if (desc->shadow_ptes[i] == spte) {
 454                                         rmap_desc_remove_entry(rmapp,
 455                                                                desc, i,
 456                                                                prev_desc);
 457                                         return;
 458                                 }
 459                         prev_desc = desc;
 460                         desc = desc->more;
 461                 }
 462                 BUG();
 463         }
 464 }
 465
 466 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 467 {
 468         struct kvm_rmap_desc *desc;
 469         struct kvm_rmap_desc *prev_desc;
 470         u64 *prev_spte;
 471         int i;
 472
 473         if (!*rmapp)
 474                 return NULL;
 475         else if (!(*rmapp & 1)) {
 476                 if (!spte)
 477                         return (u64 *)*rmapp;
 478                 return NULL;
 479         }
 480         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 481         prev_desc = NULL;
 482         prev_spte = NULL;
 483         while (desc) {
 484                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 485                         if (prev_spte == spte)
 486                                 return desc->shadow_ptes[i];
 487                         prev_spte = desc->shadow_ptes[i];
 488                 }
 489                 desc = desc->more;
 490         }
 491         return NULL;
 492 }
 493
 494 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 495 {
 496         unsigned long *rmapp;
 497         u64 *spte;
 498
 499         gfn = unalias_gfn(kvm, gfn);
 500         rmapp = gfn_to_rmap(kvm, gfn);
 501
 502         spte = rmap_next(kvm, rmapp, NULL);
 503         while (spte) {
 504                 BUG_ON(!spte);
 505                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 506                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 507                 if (is_writeble_pte(*spte))
 508                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 509                 kvm_flush_remote_tlbs(kvm);
 510                 spte = rmap_next(kvm, rmapp, spte);
 511         }
 512 }
 513
 514 #ifdef MMU_DEBUG
 515 static int is_empty_shadow_page(u64 *spt)
 516 {
 517         u64 *pos;
 518         u64 *end;
 519
 520         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 521                 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
 522                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 523                                pos, *pos);
 524                         return 0;
 525                 }
 526         return 1;
 527 }
 528 #endif
 529
 530 static void kvm_mmu_free_page(struct kvm *kvm,
 531                               struct kvm_mmu_page *page_head)
 532 {
 533         ASSERT(is_empty_shadow_page(page_head->spt));
 534         list_del(&page_head->link);
 535         __free_page(virt_to_page(page_head->spt));
 536         __free_page(virt_to_page(page_head->gfns));
 537         kfree(page_head);
 538         ++kvm->n_free_mmu_pages;
 539 }
 540
 541 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 542 {
 543         return gfn;
 544 }
 545
 546 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 547                                                u64 *parent_pte)
 548 {
 549         struct kvm_mmu_page *page;
 550
 551         if (!vcpu->kvm->n_free_mmu_pages)
 552                 return NULL;
 553
 554         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 555                                       sizeof *page);
 556         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 557         page->gfns = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 558         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 559         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 560         ASSERT(is_empty_shadow_page(page->spt));
 561         page->slot_bitmap = 0;
 562         page->multimapped = 0;
 563         page->parent_pte = parent_pte;
 564         --vcpu->kvm->n_free_mmu_pages;
 565         return page;
 566 }
 567
 568 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 569                                     struct kvm_mmu_page *page, u64 *parent_pte)
 570 {
 571         struct kvm_pte_chain *pte_chain;
 572         struct hlist_node *node;
 573         int i;
 574
 575         if (!parent_pte)
 576                 return;
 577         if (!page->multimapped) {
 578                 u64 *old = page->parent_pte;
 579
 580                 if (!old) {
 581                         page->parent_pte = parent_pte;
 582                         return;
 583                 }
 584                 page->multimapped = 1;
 585                 pte_chain = mmu_alloc_pte_chain(vcpu);
 586                 INIT_HLIST_HEAD(&page->parent_ptes);
 587                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 588                 pte_chain->parent_ptes[0] = old;
 589         }
 590         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 591                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 592                         continue;
 593                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 594                         if (!pte_chain->parent_ptes[i]) {
 595                                 pte_chain->parent_ptes[i] = parent_pte;
 596                                 return;
 597                         }
 598         }
 599         pte_chain = mmu_alloc_pte_chain(vcpu);
 600         BUG_ON(!pte_chain);
 601         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 602         pte_chain->parent_ptes[0] = parent_pte;
 603 }
 604
 605 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 606                                        u64 *parent_pte)
 607 {
 608         struct kvm_pte_chain *pte_chain;
 609         struct hlist_node *node;
 610         int i;
 611
 612         if (!page->multimapped) {
 613                 BUG_ON(page->parent_pte != parent_pte);
 614                 page->parent_pte = NULL;
 615                 return;
 616         }
 617         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 618                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 619                         if (!pte_chain->parent_ptes[i])
 620                                 break;
 621                         if (pte_chain->parent_ptes[i] != parent_pte)
 622                                 continue;
 623                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 624                                 && pte_chain->parent_ptes[i + 1]) {
 625                                 pte_chain->parent_ptes[i]
 626                                         = pte_chain->parent_ptes[i + 1];
 627                                 ++i;
 628                         }
 629                         pte_chain->parent_ptes[i] = NULL;
 630                         if (i == 0) {
 631                                 hlist_del(&pte_chain->link);
 632                                 mmu_free_pte_chain(pte_chain);
 633                                 if (hlist_empty(&page->parent_ptes)) {
 634                                         page->multimapped = 0;
 635                                         page->parent_pte = NULL;
 636                                 }
 637                         }
 638                         return;
 639                 }
 640         BUG();
 641 }
 642
 643 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm,
 644                                                 gfn_t gfn)
 645 {
 646         unsigned index;
 647         struct hlist_head *bucket;
 648         struct kvm_mmu_page *page;
 649         struct hlist_node *node;
 650
 651         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 652         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 653         bucket = &kvm->mmu_page_hash[index];
 654         hlist_for_each_entry(page, node, bucket, hash_link)
 655                 if (page->gfn == gfn && !page->role.metaphysical) {
 656                         pgprintk("%s: found role %x\n",
 657                                  __FUNCTION__, page->role.word);
 658                         return page;
 659                 }
 660         return NULL;
 661 }
 662
 663 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 664                                              gfn_t gfn,
 665                                              gva_t gaddr,
 666                                              unsigned level,
 667                                              int metaphysical,
 668                                              unsigned hugepage_access,
 669                                              u64 *parent_pte)
 670 {
 671         union kvm_mmu_page_role role;
 672         unsigned index;
 673         unsigned quadrant;
 674         struct hlist_head *bucket;
 675         struct kvm_mmu_page *page;
 676         struct hlist_node *node;
 677
 678         role.word = 0;
 679         role.glevels = vcpu->mmu.root_level;
 680         role.level = level;
 681         role.metaphysical = metaphysical;
 682         role.hugepage_access = hugepage_access;
 683         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 684                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 685                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 686                 role.quadrant = quadrant;
 687         }
 688         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 689                  gfn, role.word);
 690         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 691         bucket = &vcpu->kvm->mmu_page_hash[index];
 692         hlist_for_each_entry(page, node, bucket, hash_link)
 693                 if (page->gfn == gfn && page->role.word == role.word) {
 694                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 695                         pgprintk("%s: found\n", __FUNCTION__);
 696                         return page;
 697                 }
 698         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 699         if (!page)
 700                 return page;
 701         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 702         page->gfn = gfn;
 703         page->role = role;
 704         hlist_add_head(&page->hash_link, bucket);
 705         vcpu->mmu.prefetch_page(vcpu, page);
 706         if (!metaphysical)
 707                 rmap_write_protect(vcpu->kvm, gfn);
 708         return page;
 709 }
 710
 711 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 712                                          struct kvm_mmu_page *page)
 713 {
 714         unsigned i;
 715         u64 *pt;
 716         u64 ent;
 717
 718         pt = page->spt;
 719
 720         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 721                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 722                         if (is_shadow_present_pte(pt[i]))
 723                                 rmap_remove(kvm, &pt[i]);
 724                         pt[i] = shadow_trap_nonpresent_pte;
 725                 }
 726                 kvm_flush_remote_tlbs(kvm);
 727                 return;
 728         }
 729
 730         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 731                 ent = pt[i];
 732
 733                 pt[i] = shadow_trap_nonpresent_pte;
 734                 if (!is_shadow_present_pte(ent))
 735                         continue;
 736                 ent &= PT64_BASE_ADDR_MASK;
 737                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 738         }
 739         kvm_flush_remote_tlbs(kvm);
 740 }
 741
 742 static void kvm_mmu_put_page(struct kvm_mmu_page *page,
 743                              u64 *parent_pte)
 744 {
 745         mmu_page_remove_parent_pte(page, parent_pte);
 746 }
 747
 748 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 749 {
 750         int i;
 751
 752         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 753                 if (kvm->vcpus[i])
 754                         kvm->vcpus[i]->last_pte_updated = NULL;
 755 }
 756
 757 static void kvm_mmu_zap_page(struct kvm *kvm,
 758                              struct kvm_mmu_page *page)
 759 {
 760         u64 *parent_pte;
 761
 762         ++kvm->stat.mmu_shadow_zapped;
 763         while (page->multimapped || page->parent_pte) {
 764                 if (!page->multimapped)
 765                         parent_pte = page->parent_pte;
 766                 else {
 767                         struct kvm_pte_chain *chain;
 768
 769                         chain = container_of(page->parent_ptes.first,
 770                                              struct kvm_pte_chain, link);
 771                         parent_pte = chain->parent_ptes[0];
 772                 }
 773                 BUG_ON(!parent_pte);
 774                 kvm_mmu_put_page(page, parent_pte);
 775                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 776         }
 777         kvm_mmu_page_unlink_children(kvm, page);
 778         if (!page->root_count) {
 779                 hlist_del(&page->hash_link);
 780                 kvm_mmu_free_page(kvm, page);
 781         } else
 782                 list_move(&page->link, &kvm->active_mmu_pages);
 783         kvm_mmu_reset_last_pte_updated(kvm);
 784 }
 785
 786 /*
 787  * Changing the number of mmu pages allocated to the vm
 788  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 789  */
 790 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 791 {
 792         /*
 793          * If we set the number of mmu pages to be smaller be than the
 794          * number of actived pages , we must to free some mmu pages before we
 795          * change the value
 796          */
 797
 798         if ((kvm->n_alloc_mmu_pages - kvm->n_free_mmu_pages) >
 799             kvm_nr_mmu_pages) {
 800                 int n_used_mmu_pages = kvm->n_alloc_mmu_pages
 801                                        - kvm->n_free_mmu_pages;
 802
 803                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 804                         struct kvm_mmu_page *page;
 805
 806                         page = container_of(kvm->active_mmu_pages.prev,
 807                                             struct kvm_mmu_page, link);
 808                         kvm_mmu_zap_page(kvm, page);
 809                         n_used_mmu_pages--;
 810                 }
 811                 kvm->n_free_mmu_pages = 0;
 812         }
 813         else
 814                 kvm->n_free_mmu_pages += kvm_nr_mmu_pages
 815                                          - kvm->n_alloc_mmu_pages;
 816
 817         kvm->n_alloc_mmu_pages = kvm_nr_mmu_pages;
 818 }
 819
 820 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 821 {
 822         unsigned index;
 823         struct hlist_head *bucket;
 824         struct kvm_mmu_page *page;
 825         struct hlist_node *node, *n;
 826         int r;
 827
 828         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 829         r = 0;
 830         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 831         bucket = &kvm->mmu_page_hash[index];
 832         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 833                 if (page->gfn == gfn && !page->role.metaphysical) {
 834                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 835                                  page->role.word);
 836                         kvm_mmu_zap_page(kvm, page);
 837                         r = 1;
 838                 }
 839         return r;
 840 }
 841
 842 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
 843 {
 844         struct kvm_mmu_page *page;
 845
 846         while ((page = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
 847                 pgprintk("%s: zap %lx %x\n",
 848                          __FUNCTION__, gfn, page->role.word);
 849                 kvm_mmu_zap_page(kvm, page);
 850         }
 851 }
 852
 853 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 854 {
 855         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 856         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 857
 858         __set_bit(slot, &page_head->slot_bitmap);
 859 }
 860
 861 hpa_t gpa_to_hpa(struct kvm *kvm, gpa_t gpa)
 862 {
 863         struct page *page;
 864         hpa_t hpa;
 865
 866         ASSERT((gpa & HPA_ERR_MASK) == 0);
 867         page = gfn_to_page(kvm, gpa >> PAGE_SHIFT);
 868         hpa = ((hpa_t)page_to_pfn(page) << PAGE_SHIFT) | (gpa & (PAGE_SIZE-1));
 869         if (is_error_page(page))
 870                 return hpa | HPA_ERR_MASK;
 871         return hpa;
 872 }
 873
 874 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 875 {
 876         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 877
 878         if (gpa == UNMAPPED_GVA)
 879                 return UNMAPPED_GVA;
 880         return gpa_to_hpa(vcpu->kvm, gpa);
 881 }
 882
 883 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 884 {
 885         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 886
 887         if (gpa == UNMAPPED_GVA)
 888                 return NULL;
 889         return pfn_to_page(gpa_to_hpa(vcpu->kvm, gpa) >> PAGE_SHIFT);
 890 }
 891
 892 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 893 {
 894 }
 895
 896 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 897 {
 898         int level = PT32E_ROOT_LEVEL;
 899         hpa_t table_addr = vcpu->mmu.root_hpa;
 900         struct page *page;
 901
 902         page = pfn_to_page(p >> PAGE_SHIFT);
 903         for (; ; level--) {
 904                 u32 index = PT64_INDEX(v, level);
 905                 u64 *table;
 906                 u64 pte;
 907
 908                 ASSERT(VALID_PAGE(table_addr));
 909                 table = __va(table_addr);
 910
 911                 if (level == 1) {
 912                         int was_rmapped;
 913
 914                         pte = table[index];
 915                         was_rmapped = is_rmap_pte(pte);
 916                         if (is_shadow_present_pte(pte) && is_writeble_pte(pte)) {
 917                                 kvm_release_page_clean(page);
 918                                 return 0;
 919                         }
 920                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 921                         page_header_update_slot(vcpu->kvm, table, v);
 922                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 923                                                                 PT_USER_MASK;
 924                         if (!was_rmapped)
 925                                 rmap_add(vcpu, &table[index], v >> PAGE_SHIFT);
 926                         else
 927                                 kvm_release_page_clean(page);
 928
 929                         return 0;
 930                 }
 931
 932                 if (table[index] == shadow_trap_nonpresent_pte) {
 933                         struct kvm_mmu_page *new_table;
 934                         gfn_t pseudo_gfn;
 935
 936                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 937                                 >> PAGE_SHIFT;
 938                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 939                                                      v, level - 1,
 940                                                      1, 3, &table[index]);
 941                         if (!new_table) {
 942                                 pgprintk("nonpaging_map: ENOMEM\n");
 943                                 kvm_release_page_clean(page);
 944                                 return -ENOMEM;
 945                         }
 946
 947                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 948                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 949                 }
 950                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 951         }
 952 }
 953
 954 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 955                                     struct kvm_mmu_page *sp)
 956 {
 957         int i;
 958
 959         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 960                 sp->spt[i] = shadow_trap_nonpresent_pte;
 961 }
 962
 963 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 964 {
 965         int i;
 966         struct kvm_mmu_page *page;
 967
 968         if (!VALID_PAGE(vcpu->mmu.root_hpa))
 969                 return;
 970 #ifdef CONFIG_X86_64
 971         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 972                 hpa_t root = vcpu->mmu.root_hpa;
 973
 974                 page = page_header(root);
 975                 --page->root_count;
 976                 vcpu->mmu.root_hpa = INVALID_PAGE;
 977                 return;
 978         }
 979 #endif
 980         for (i = 0; i < 4; ++i) {
 981                 hpa_t root = vcpu->mmu.pae_root[i];
 982
 983                 if (root) {
 984                         root &= PT64_BASE_ADDR_MASK;
 985                         page = page_header(root);
 986                         --page->root_count;
 987                 }
 988                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 989         }
 990         vcpu->mmu.root_hpa = INVALID_PAGE;
 991 }
 992
 993 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 994 {
 995         int i;
 996         gfn_t root_gfn;
 997         struct kvm_mmu_page *page;
 998
 999         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
1000
1001 #ifdef CONFIG_X86_64
1002         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1003                 hpa_t root = vcpu->mmu.root_hpa;
1004
1005                 ASSERT(!VALID_PAGE(root));
1006                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
1007                                         PT64_ROOT_LEVEL, 0, 0, NULL);
1008                 root = __pa(page->spt);
1009                 ++page->root_count;
1010                 vcpu->mmu.root_hpa = root;
1011                 return;
1012         }
1013 #endif
1014         for (i = 0; i < 4; ++i) {
1015                 hpa_t root = vcpu->mmu.pae_root[i];
1016
1017                 ASSERT(!VALID_PAGE(root));
1018                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
1019                         if (!is_present_pte(vcpu->pdptrs[i])) {
1020                                 vcpu->mmu.pae_root[i] = 0;
1021                                 continue;
1022                         }
1023                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
1024                 } else if (vcpu->mmu.root_level == 0)
1025                         root_gfn = 0;
1026                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1027                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
1028                                         0, NULL);
1029                 root = __pa(page->spt);
1030                 ++page->root_count;
1031                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
1032         }
1033         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
1034 }
1035
1036 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1037 {
1038         return vaddr;
1039 }
1040
1041 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1042                                u32 error_code)
1043 {
1044         gpa_t addr = gva;
1045         hpa_t paddr;
1046         int r;
1047
1048         r = mmu_topup_memory_caches(vcpu);
1049         if (r)
1050                 return r;
1051
1052         ASSERT(vcpu);
1053         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
1054
1055
1056         paddr = gpa_to_hpa(vcpu->kvm, addr & PT64_BASE_ADDR_MASK);
1057
1058         if (is_error_hpa(paddr)) {
1059                 kvm_release_page_clean(pfn_to_page((paddr & PT64_BASE_ADDR_MASK)
1060                                        >> PAGE_SHIFT));
1061                 return 1;
1062         }
1063
1064         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
1065 }
1066
1067 static void nonpaging_free(struct kvm_vcpu *vcpu)
1068 {
1069         mmu_free_roots(vcpu);
1070 }
1071
1072 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1073 {
1074         struct kvm_mmu *context = &vcpu->mmu;
1075
1076         context->new_cr3 = nonpaging_new_cr3;
1077         context->page_fault = nonpaging_page_fault;
1078         context->gva_to_gpa = nonpaging_gva_to_gpa;
1079         context->free = nonpaging_free;
1080         context->prefetch_page = nonpaging_prefetch_page;
1081         context->root_level = 0;
1082         context->shadow_root_level = PT32E_ROOT_LEVEL;
1083         context->root_hpa = INVALID_PAGE;
1084         return 0;
1085 }
1086
1087 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1088 {
1089         ++vcpu->stat.tlb_flush;
1090         kvm_x86_ops->tlb_flush(vcpu);
1091 }
1092
1093 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1094 {
1095         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
1096         mmu_free_roots(vcpu);
1097 }
1098
1099 static void inject_page_fault(struct kvm_vcpu *vcpu,
1100                               u64 addr,
1101                               u32 err_code)
1102 {
1103         kvm_x86_ops->inject_page_fault(vcpu, addr, err_code);
1104 }
1105
1106 static void paging_free(struct kvm_vcpu *vcpu)
1107 {
1108         nonpaging_free(vcpu);
1109 }
1110
1111 #define PTTYPE 64
1112 #include "paging_tmpl.h"
1113 #undef PTTYPE
1114
1115 #define PTTYPE 32
1116 #include "paging_tmpl.h"
1117 #undef PTTYPE
1118
1119 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1120 {
1121         struct kvm_mmu *context = &vcpu->mmu;
1122
1123         ASSERT(is_pae(vcpu));
1124         context->new_cr3 = paging_new_cr3;
1125         context->page_fault = paging64_page_fault;
1126         context->gva_to_gpa = paging64_gva_to_gpa;
1127         context->prefetch_page = paging64_prefetch_page;
1128         context->free = paging_free;
1129         context->root_level = level;
1130         context->shadow_root_level = level;
1131         context->root_hpa = INVALID_PAGE;
1132         return 0;
1133 }
1134
1135 static int paging64_init_context(struct kvm_vcpu *vcpu)
1136 {
1137         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1138 }
1139
1140 static int paging32_init_context(struct kvm_vcpu *vcpu)
1141 {
1142         struct kvm_mmu *context = &vcpu->mmu;
1143
1144         context->new_cr3 = paging_new_cr3;
1145         context->page_fault = paging32_page_fault;
1146         context->gva_to_gpa = paging32_gva_to_gpa;
1147         context->free = paging_free;
1148         context->prefetch_page = paging32_prefetch_page;
1149         context->root_level = PT32_ROOT_LEVEL;
1150         context->shadow_root_level = PT32E_ROOT_LEVEL;
1151         context->root_hpa = INVALID_PAGE;
1152         return 0;
1153 }
1154
1155 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1156 {
1157         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1158 }
1159
1160 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1161 {
1162         ASSERT(vcpu);
1163         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1164
1165         if (!is_paging(vcpu))
1166                 return nonpaging_init_context(vcpu);
1167         else if (is_long_mode(vcpu))
1168                 return paging64_init_context(vcpu);
1169         else if (is_pae(vcpu))
1170                 return paging32E_init_context(vcpu);
1171         else
1172                 return paging32_init_context(vcpu);
1173 }
1174
1175 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1176 {
1177         ASSERT(vcpu);
1178         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1179                 vcpu->mmu.free(vcpu);
1180                 vcpu->mmu.root_hpa = INVALID_PAGE;
1181         }
1182 }
1183
1184 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1185 {
1186         destroy_kvm_mmu(vcpu);
1187         return init_kvm_mmu(vcpu);
1188 }
1189 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1190
1191 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1192 {
1193         int r;
1194
1195         mutex_lock(&vcpu->kvm->lock);
1196         r = mmu_topup_memory_caches(vcpu);
1197         if (r)
1198                 goto out;
1199         mmu_alloc_roots(vcpu);
1200         kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1201         kvm_mmu_flush_tlb(vcpu);
1202 out:
1203         mutex_unlock(&vcpu->kvm->lock);
1204         return r;
1205 }
1206 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1207
1208 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1209 {
1210         mmu_free_roots(vcpu);
1211 }
1212
1213 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1214                                   struct kvm_mmu_page *page,
1215                                   u64 *spte)
1216 {
1217         u64 pte;
1218         struct kvm_mmu_page *child;
1219
1220         pte = *spte;
1221         if (is_shadow_present_pte(pte)) {
1222                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1223                         rmap_remove(vcpu->kvm, spte);
1224                 else {
1225                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1226                         mmu_page_remove_parent_pte(child, spte);
1227                 }
1228         }
1229         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1230         kvm_flush_remote_tlbs(vcpu->kvm);
1231 }
1232
1233 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1234                                   struct kvm_mmu_page *page,
1235                                   u64 *spte,
1236                                   const void *new, int bytes,
1237                                   int offset_in_pte)
1238 {
1239         if (page->role.level != PT_PAGE_TABLE_LEVEL) {
1240                 ++vcpu->kvm->stat.mmu_pde_zapped;
1241                 return;
1242         }
1243
1244         ++vcpu->kvm->stat.mmu_pte_updated;
1245         if (page->role.glevels == PT32_ROOT_LEVEL)
1246                 paging32_update_pte(vcpu, page, spte, new, bytes,
1247                                     offset_in_pte);
1248         else
1249                 paging64_update_pte(vcpu, page, spte, new, bytes,
1250                                     offset_in_pte);
1251 }
1252
1253 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1254 {
1255         u64 *spte = vcpu->last_pte_updated;
1256
1257         return !!(spte && (*spte & PT_ACCESSED_MASK));
1258 }
1259
1260 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1261                        const u8 *new, int bytes)
1262 {
1263         gfn_t gfn = gpa >> PAGE_SHIFT;
1264         struct kvm_mmu_page *page;
1265         struct hlist_node *node, *n;
1266         struct hlist_head *bucket;
1267         unsigned index;
1268         u64 *spte;
1269         unsigned offset = offset_in_page(gpa);
1270         unsigned pte_size;
1271         unsigned page_offset;
1272         unsigned misaligned;
1273         unsigned quadrant;
1274         int level;
1275         int flooded = 0;
1276         int npte;
1277
1278         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1279         ++vcpu->kvm->stat.mmu_pte_write;
1280         kvm_mmu_audit(vcpu, "pre pte write");
1281         if (gfn == vcpu->last_pt_write_gfn
1282             && !last_updated_pte_accessed(vcpu)) {
1283                 ++vcpu->last_pt_write_count;
1284                 if (vcpu->last_pt_write_count >= 3)
1285                         flooded = 1;
1286         } else {
1287                 vcpu->last_pt_write_gfn = gfn;
1288                 vcpu->last_pt_write_count = 1;
1289                 vcpu->last_pte_updated = NULL;
1290         }
1291         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1292         bucket = &vcpu->kvm->mmu_page_hash[index];
1293         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1294                 if (page->gfn != gfn || page->role.metaphysical)
1295                         continue;
1296                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1297                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1298                 misaligned |= bytes < 4;
1299                 if (misaligned || flooded) {
1300                         /*
1301                          * Misaligned accesses are too much trouble to fix
1302                          * up; also, they usually indicate a page is not used
1303                          * as a page table.
1304                          *
1305                          * If we're seeing too many writes to a page,
1306                          * it may no longer be a page table, or we may be
1307                          * forking, in which case it is better to unmap the
1308                          * page.
1309                          */
1310                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1311                                  gpa, bytes, page->role.word);
1312                         kvm_mmu_zap_page(vcpu->kvm, page);
1313                         ++vcpu->kvm->stat.mmu_flooded;
1314                         continue;
1315                 }
1316                 page_offset = offset;
1317                 level = page->role.level;
1318                 npte = 1;
1319                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1320                         page_offset <<= 1;      /* 32->64 */
1321                         /*
1322                          * A 32-bit pde maps 4MB while the shadow pdes map
1323                          * only 2MB.  So we need to double the offset again
1324                          * and zap two pdes instead of one.
1325                          */
1326                         if (level == PT32_ROOT_LEVEL) {
1327                                 page_offset &= ~7; /* kill rounding error */
1328                                 page_offset <<= 1;
1329                                 npte = 2;
1330                         }
1331                         quadrant = page_offset >> PAGE_SHIFT;
1332                         page_offset &= ~PAGE_MASK;
1333                         if (quadrant != page->role.quadrant)
1334                                 continue;
1335                 }
1336                 spte = &page->spt[page_offset / sizeof(*spte)];
1337                 while (npte--) {
1338                         mmu_pte_write_zap_pte(vcpu, page, spte);
1339                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes,
1340                                               page_offset & (pte_size - 1));
1341                         ++spte;
1342                 }
1343         }
1344         kvm_mmu_audit(vcpu, "post pte write");
1345 }
1346
1347 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1348 {
1349         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1350
1351         return kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1352 }
1353
1354 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1355 {
1356         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1357                 struct kvm_mmu_page *page;
1358
1359                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1360                                     struct kvm_mmu_page, link);
1361                 kvm_mmu_zap_page(vcpu->kvm, page);
1362                 ++vcpu->kvm->stat.mmu_recycled;
1363         }
1364 }
1365
1366 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1367 {
1368         int r;
1369         enum emulation_result er;
1370
1371         mutex_lock(&vcpu->kvm->lock);
1372         r = vcpu->mmu.page_fault(vcpu, cr2, error_code);
1373         if (r < 0)
1374                 goto out;
1375
1376         if (!r) {
1377                 r = 1;
1378                 goto out;
1379         }
1380
1381         r = mmu_topup_memory_caches(vcpu);
1382         if (r)
1383                 goto out;
1384
1385         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1386         mutex_unlock(&vcpu->kvm->lock);
1387
1388         switch (er) {
1389         case EMULATE_DONE:
1390                 return 1;
1391         case EMULATE_DO_MMIO:
1392                 ++vcpu->stat.mmio_exits;
1393                 return 0;
1394         case EMULATE_FAIL:
1395                 kvm_report_emulation_failure(vcpu, "pagetable");
1396                 return 1;
1397         default:
1398                 BUG();
1399         }
1400 out:
1401         mutex_unlock(&vcpu->kvm->lock);
1402         return r;
1403 }
1404 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1405
1406 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1407 {
1408         struct kvm_mmu_page *page;
1409
1410         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1411                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1412                                     struct kvm_mmu_page, link);
1413                 kvm_mmu_zap_page(vcpu->kvm, page);
1414         }
1415         free_page((unsigned long)vcpu->mmu.pae_root);
1416 }
1417
1418 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1419 {
1420         struct page *page;
1421         int i;
1422
1423         ASSERT(vcpu);
1424
1425         if (vcpu->kvm->n_requested_mmu_pages)
1426                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_requested_mmu_pages;
1427         else
1428                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_alloc_mmu_pages;
1429         /*
1430          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1431          * Therefore we need to allocate shadow page tables in the first
1432          * 4GB of memory, which happens to fit the DMA32 zone.
1433          */
1434         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1435         if (!page)
1436                 goto error_1;
1437         vcpu->mmu.pae_root = page_address(page);
1438         for (i = 0; i < 4; ++i)
1439                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1440
1441         return 0;
1442
1443 error_1:
1444         free_mmu_pages(vcpu);
1445         return -ENOMEM;
1446 }
1447
1448 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1449 {
1450         ASSERT(vcpu);
1451         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1452
1453         return alloc_mmu_pages(vcpu);
1454 }
1455
1456 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1457 {
1458         ASSERT(vcpu);
1459         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1460
1461         return init_kvm_mmu(vcpu);
1462 }
1463
1464 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1465 {
1466         ASSERT(vcpu);
1467
1468         destroy_kvm_mmu(vcpu);
1469         free_mmu_pages(vcpu);
1470         mmu_free_memory_caches(vcpu);
1471 }
1472
1473 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1474 {
1475         struct kvm_mmu_page *page;
1476
1477         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1478                 int i;
1479                 u64 *pt;
1480
1481                 if (!test_bit(slot, &page->slot_bitmap))
1482                         continue;
1483
1484                 pt = page->spt;
1485                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1486                         /* avoid RMW */
1487                         if (pt[i] & PT_WRITABLE_MASK)
1488                                 pt[i] &= ~PT_WRITABLE_MASK;
1489         }
1490 }
1491
1492 void kvm_mmu_zap_all(struct kvm *kvm)
1493 {
1494         struct kvm_mmu_page *page, *node;
1495
1496         list_for_each_entry_safe(page, node, &kvm->active_mmu_pages, link)
1497                 kvm_mmu_zap_page(kvm, page);
1498
1499         kvm_flush_remote_tlbs(kvm);
1500 }
1501
1502 void kvm_mmu_module_exit(void)
1503 {
1504         if (pte_chain_cache)
1505                 kmem_cache_destroy(pte_chain_cache);
1506         if (rmap_desc_cache)
1507                 kmem_cache_destroy(rmap_desc_cache);
1508         if (mmu_page_header_cache)
1509                 kmem_cache_destroy(mmu_page_header_cache);
1510 }
1511
1512 int kvm_mmu_module_init(void)
1513 {
1514         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1515                                             sizeof(struct kvm_pte_chain),
1516                                             0, 0, NULL);
1517         if (!pte_chain_cache)
1518                 goto nomem;
1519         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1520                                             sizeof(struct kvm_rmap_desc),
1521                                             0, 0, NULL);
1522         if (!rmap_desc_cache)
1523                 goto nomem;
1524
1525         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1526                                                   sizeof(struct kvm_mmu_page),
1527                                                   0, 0, NULL);
1528         if (!mmu_page_header_cache)
1529                 goto nomem;
1530
1531         return 0;
1532
1533 nomem:
1534         kvm_mmu_module_exit();
1535         return -ENOMEM;
1536 }
1537
1538 /*
1539  * Caculate mmu pages needed for kvm.
1540  */
1541 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
1542 {
1543         int i;
1544         unsigned int nr_mmu_pages;
1545         unsigned int  nr_pages = 0;
1546
1547         for (i = 0; i < kvm->nmemslots; i++)
1548                 nr_pages += kvm->memslots[i].npages;
1549
1550         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
1551         nr_mmu_pages = max(nr_mmu_pages,
1552                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
1553
1554         return nr_mmu_pages;
1555 }
1556
1557 #ifdef AUDIT
1558
1559 static const char *audit_msg;
1560
1561 static gva_t canonicalize(gva_t gva)
1562 {
1563 #ifdef CONFIG_X86_64
1564         gva = (long long)(gva << 16) >> 16;
1565 #endif
1566         return gva;
1567 }
1568
1569 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1570                                 gva_t va, int level)
1571 {
1572         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1573         int i;
1574         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1575
1576         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1577                 u64 ent = pt[i];
1578
1579                 if (ent == shadow_trap_nonpresent_pte)
1580                         continue;
1581
1582                 va = canonicalize(va);
1583                 if (level > 1) {
1584                         if (ent == shadow_notrap_nonpresent_pte)
1585                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1586                                        " in nonleaf level: levels %d gva %lx"
1587                                        " level %d pte %llx\n", audit_msg,
1588                                        vcpu->mmu.root_level, va, level, ent);
1589
1590                         audit_mappings_page(vcpu, ent, va, level - 1);
1591                 } else {
1592                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1593                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1594                         struct page *page;
1595
1596                         if (is_shadow_present_pte(ent)
1597                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1598                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1599                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1600                                        audit_msg, vcpu->mmu.root_level,
1601                                        va, gpa, hpa, ent,
1602                                        is_shadow_present_pte(ent));
1603                         else if (ent == shadow_notrap_nonpresent_pte
1604                                  && !is_error_hpa(hpa))
1605                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1606                                        " valid guest gva %lx\n", audit_msg, va);
1607                         page = pfn_to_page((gpa & PT64_BASE_ADDR_MASK)
1608                                            >> PAGE_SHIFT);
1609                         kvm_release_page_clean(page);
1610
1611                 }
1612         }
1613 }
1614
1615 static void audit_mappings(struct kvm_vcpu *vcpu)
1616 {
1617         unsigned i;
1618
1619         if (vcpu->mmu.root_level == 4)
1620                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1621         else
1622                 for (i = 0; i < 4; ++i)
1623                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1624                                 audit_mappings_page(vcpu,
1625                                                     vcpu->mmu.pae_root[i],
1626                                                     i << 30,
1627                                                     2);
1628 }
1629
1630 static int count_rmaps(struct kvm_vcpu *vcpu)
1631 {
1632         int nmaps = 0;
1633         int i, j, k;
1634
1635         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1636                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1637                 struct kvm_rmap_desc *d;
1638
1639                 for (j = 0; j < m->npages; ++j) {
1640                         unsigned long *rmapp = &m->rmap[j];
1641
1642                         if (!*rmapp)
1643                                 continue;
1644                         if (!(*rmapp & 1)) {
1645                                 ++nmaps;
1646                                 continue;
1647                         }
1648                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1649                         while (d) {
1650                                 for (k = 0; k < RMAP_EXT; ++k)
1651                                         if (d->shadow_ptes[k])
1652                                                 ++nmaps;
1653                                         else
1654                                                 break;
1655                                 d = d->more;
1656                         }
1657                 }
1658         }
1659         return nmaps;
1660 }
1661
1662 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1663 {
1664         int nmaps = 0;
1665         struct kvm_mmu_page *page;
1666         int i;
1667
1668         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1669                 u64 *pt = page->spt;
1670
1671                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1672                         continue;
1673
1674                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1675                         u64 ent = pt[i];
1676
1677                         if (!(ent & PT_PRESENT_MASK))
1678                                 continue;
1679                         if (!(ent & PT_WRITABLE_MASK))
1680                                 continue;
1681                         ++nmaps;
1682                 }
1683         }
1684         return nmaps;
1685 }
1686
1687 static void audit_rmap(struct kvm_vcpu *vcpu)
1688 {
1689         int n_rmap = count_rmaps(vcpu);
1690         int n_actual = count_writable_mappings(vcpu);
1691
1692         if (n_rmap != n_actual)
1693                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1694                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1695 }
1696
1697 static void audit_write_protection(struct kvm_vcpu *vcpu)
1698 {
1699         struct kvm_mmu_page *page;
1700         struct kvm_memory_slot *slot;
1701         unsigned long *rmapp;
1702         gfn_t gfn;
1703
1704         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1705                 if (page->role.metaphysical)
1706                         continue;
1707
1708                 slot = gfn_to_memslot(vcpu->kvm, page->gfn);
1709                 gfn = unalias_gfn(vcpu->kvm, page->gfn);
1710                 rmapp = &slot->rmap[gfn - slot->base_gfn];
1711                 if (*rmapp)
1712                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1713                                " mappings: gfn %lx role %x\n",
1714                                __FUNCTION__, audit_msg, page->gfn,
1715                                page->role.word);
1716         }
1717 }
1718
1719 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1720 {
1721         int olddbg = dbg;
1722
1723         dbg = 0;
1724         audit_msg = msg;
1725         audit_rmap(vcpu);
1726         audit_write_protection(vcpu);
1727         audit_mappings(vcpu);
1728         dbg = olddbg;
1729 }
1730
1731 #endif