drivers/kvm/paging_tmpl.h

   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 /*
  21  * We need the mmu code to access both 32-bit and 64-bit guest ptes,
  22  * so the code in this file is compiled twice, once per pte size.
  23  */
  24
  25 #if PTTYPE == 64
  26         #define pt_element_t u64
  27         #define guest_walker guest_walker64
  28         #define FNAME(name) paging##64_##name
  29         #define PT_BASE_ADDR_MASK PT64_BASE_ADDR_MASK
  30         #define PT_DIR_BASE_ADDR_MASK PT64_DIR_BASE_ADDR_MASK
  31         #define PT_INDEX(addr, level) PT64_INDEX(addr, level)
  32         #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
  33         #define PT_LEVEL_MASK(level) PT64_LEVEL_MASK(level)
  34         #define PT_PTE_COPY_MASK PT64_PTE_COPY_MASK
  35         #ifdef CONFIG_X86_64
  36         #define PT_MAX_FULL_LEVELS 4
  37         #else
  38         #define PT_MAX_FULL_LEVELS 2
  39         #endif
  40 #elif PTTYPE == 32
  41         #define pt_element_t u32
  42         #define guest_walker guest_walker32
  43         #define FNAME(name) paging##32_##name
  44         #define PT_BASE_ADDR_MASK PT32_BASE_ADDR_MASK
  45         #define PT_DIR_BASE_ADDR_MASK PT32_DIR_BASE_ADDR_MASK
  46         #define PT_INDEX(addr, level) PT32_INDEX(addr, level)
  47         #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
  48         #define PT_LEVEL_MASK(level) PT32_LEVEL_MASK(level)
  49         #define PT_PTE_COPY_MASK PT32_PTE_COPY_MASK
  50         #define PT_MAX_FULL_LEVELS 2
  51 #else
  52         #error Invalid PTTYPE value
  53 #endif
  54
  55 /*
  56  * The guest_walker structure emulates the behavior of the hardware page
  57  * table walker.
  58  */
  59 struct guest_walker {
  60         int level;
  61         gfn_t table_gfn[PT_MAX_FULL_LEVELS];
  62         pt_element_t *table;
  63         pt_element_t *ptep;
  64         pt_element_t inherited_ar;
  65         gfn_t gfn;
  66         u32 error_code;
  67 };
  68
  69 /*
  70  * Fetch a guest pte for a guest virtual address
  71  */
  72 static int FNAME(walk_addr)(struct guest_walker *walker,
  73                             struct kvm_vcpu *vcpu, gva_t addr,
  74                             int write_fault, int user_fault, int fetch_fault)
  75 {
  76         hpa_t hpa;
  77         struct kvm_memory_slot *slot;
  78         pt_element_t *ptep;
  79         pt_element_t root;
  80         gfn_t table_gfn;
  81
  82         pgprintk("%s: addr %lx\n", __FUNCTION__, addr);
  83         walker->level = vcpu->mmu.root_level;
  84         walker->table = NULL;
  85         root = vcpu->cr3;
  86 #if PTTYPE == 64
  87         if (!is_long_mode(vcpu)) {
  88                 walker->ptep = &vcpu->pdptrs[(addr >> 30) & 3];
  89                 root = *walker->ptep;
  90                 if (!(root & PT_PRESENT_MASK))
  91                         goto not_present;
  92                 --walker->level;
  93         }
  94 #endif
  95         table_gfn = (root & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
  96         walker->table_gfn[walker->level - 1] = table_gfn;
  97         pgprintk("%s: table_gfn[%d] %lx\n", __FUNCTION__,
  98                  walker->level - 1, table_gfn);
  99         slot = gfn_to_memslot(vcpu->kvm, table_gfn);
 100         hpa = safe_gpa_to_hpa(vcpu, root & PT64_BASE_ADDR_MASK);
 101         walker->table = kmap_atomic(pfn_to_page(hpa >> PAGE_SHIFT), KM_USER0);
 102
 103         ASSERT((!is_long_mode(vcpu) && is_pae(vcpu)) ||
 104                (vcpu->cr3 & ~(PAGE_MASK | CR3_FLAGS_MASK)) == 0);
 105
 106         walker->inherited_ar = PT_USER_MASK | PT_WRITABLE_MASK;
 107
 108         for (;;) {
 109                 int index = PT_INDEX(addr, walker->level);
 110                 hpa_t paddr;
 111
 112                 ptep = &walker->table[index];
 113                 ASSERT(((unsigned long)walker->table & PAGE_MASK) ==
 114                        ((unsigned long)ptep & PAGE_MASK));
 115
 116                 if (!is_present_pte(*ptep))
 117                         goto not_present;
 118
 119                 if (write_fault && !is_writeble_pte(*ptep))
 120                         if (user_fault || is_write_protection(vcpu))
 121                                 goto access_error;
 122
 123                 if (user_fault && !(*ptep & PT_USER_MASK))
 124                         goto access_error;
 125
 126 #if PTTYPE == 64
 127                 if (fetch_fault && is_nx(vcpu) && (*ptep & PT64_NX_MASK))
 128                         goto access_error;
 129 #endif
 130
 131                 if (!(*ptep & PT_ACCESSED_MASK)) {
 132                         mark_page_dirty(vcpu->kvm, table_gfn);
 133                         *ptep |= PT_ACCESSED_MASK;
 134                 }
 135
 136                 if (walker->level == PT_PAGE_TABLE_LEVEL) {
 137                         walker->gfn = (*ptep & PT_BASE_ADDR_MASK)
 138                                 >> PAGE_SHIFT;
 139                         break;
 140                 }
 141
 142                 if (walker->level == PT_DIRECTORY_LEVEL
 143                     && (*ptep & PT_PAGE_SIZE_MASK)
 144                     && (PTTYPE == 64 || is_pse(vcpu))) {
 145                         walker->gfn = (*ptep & PT_DIR_BASE_ADDR_MASK)
 146                                 >> PAGE_SHIFT;
 147                         walker->gfn += PT_INDEX(addr, PT_PAGE_TABLE_LEVEL);
 148                         break;
 149                 }
 150
 151                 walker->inherited_ar &= walker->table[index];
 152                 table_gfn = (*ptep & PT_BASE_ADDR_MASK) >> PAGE_SHIFT;
 153                 paddr = safe_gpa_to_hpa(vcpu, *ptep & PT_BASE_ADDR_MASK);
 154                 kunmap_atomic(walker->table, KM_USER0);
 155                 walker->table = kmap_atomic(pfn_to_page(paddr >> PAGE_SHIFT),
 156                                             KM_USER0);
 157                 --walker->level;
 158                 walker->table_gfn[walker->level - 1 ] = table_gfn;
 159                 pgprintk("%s: table_gfn[%d] %lx\n", __FUNCTION__,
 160                          walker->level - 1, table_gfn);
 161         }
 162         walker->ptep = ptep;
 163         pgprintk("%s: pte %llx\n", __FUNCTION__, (u64)*ptep);
 164         return 1;
 165
 166 not_present:
 167         walker->error_code = 0;
 168         goto err;
 169
 170 access_error:
 171         walker->error_code = PFERR_PRESENT_MASK;
 172
 173 err:
 174         if (write_fault)
 175                 walker->error_code |= PFERR_WRITE_MASK;
 176         if (user_fault)
 177                 walker->error_code |= PFERR_USER_MASK;
 178         if (fetch_fault)
 179                 walker->error_code |= PFERR_FETCH_MASK;
 180         return 0;
 181 }
 182
 183 static void FNAME(release_walker)(struct guest_walker *walker)
 184 {
 185         if (walker->table)
 186                 kunmap_atomic(walker->table, KM_USER0);
 187 }
 188
 189 static void FNAME(mark_pagetable_dirty)(struct kvm *kvm,
 190                                         struct guest_walker *walker)
 191 {
 192         mark_page_dirty(kvm, walker->table_gfn[walker->level - 1]);
 193 }
 194
 195 static void FNAME(set_pte_common)(struct kvm_vcpu *vcpu,
 196                                   u64 *shadow_pte,
 197                                   gpa_t gaddr,
 198                                   pt_element_t *gpte,
 199                                   u64 access_bits,
 200                                   gfn_t gfn)
 201 {
 202         hpa_t paddr;
 203         int dirty = *gpte & PT_DIRTY_MASK;
 204
 205         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 206         if (!dirty)
 207                 access_bits &= ~PT_WRITABLE_MASK;
 208
 209         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 210
 211         *shadow_pte |= access_bits;
 212
 213         if (is_error_hpa(paddr)) {
 214                 *shadow_pte |= gaddr;
 215                 *shadow_pte |= PT_SHADOW_IO_MARK;
 216                 *shadow_pte &= ~PT_PRESENT_MASK;
 217                 return;
 218         }
 219
 220         *shadow_pte |= paddr;
 221
 222         if (access_bits & PT_WRITABLE_MASK) {
 223                 struct kvm_mmu_page *shadow;
 224
 225                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 226                 if (shadow) {
 227                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 228                                  __FUNCTION__, gfn);
 229                         access_bits &= ~PT_WRITABLE_MASK;
 230                         if (is_writeble_pte(*shadow_pte)) {
 231                                     *shadow_pte &= ~PT_WRITABLE_MASK;
 232                                     kvm_arch_ops->tlb_flush(vcpu);
 233                         }
 234                 }
 235         }
 236
 237         if (access_bits & PT_WRITABLE_MASK)
 238                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 239
 240         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 241         rmap_add(vcpu, shadow_pte);
 242 }
 243
 244 static void FNAME(set_pte)(struct kvm_vcpu *vcpu, pt_element_t *gpte,
 245                            u64 *shadow_pte, u64 access_bits, gfn_t gfn)
 246 {
 247         ASSERT(*shadow_pte == 0);
 248         access_bits &= *gpte;
 249         *shadow_pte = (*gpte & PT_PTE_COPY_MASK);
 250         FNAME(set_pte_common)(vcpu, shadow_pte, *gpte & PT_BASE_ADDR_MASK,
 251                               gpte, access_bits, gfn);
 252 }
 253
 254 static void FNAME(update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *page,
 255                               u64 *spte, const void *pte, int bytes)
 256 {
 257         pt_element_t gpte;
 258
 259         if (bytes < sizeof(pt_element_t))
 260                 return;
 261         gpte = *(const pt_element_t *)pte;
 262         if (~gpte & (PT_PRESENT_MASK | PT_ACCESSED_MASK))
 263                 return;
 264         pgprintk("%s: gpte %llx spte %p\n", __FUNCTION__, (u64)gpte, spte);
 265         FNAME(set_pte)(vcpu, &gpte, spte, PT_USER_MASK | PT_WRITABLE_MASK,
 266                        (gpte & PT_BASE_ADDR_MASK) >> PAGE_SHIFT);
 267 }
 268
 269 static void FNAME(set_pde)(struct kvm_vcpu *vcpu, pt_element_t *gpde,
 270                            u64 *shadow_pte, u64 access_bits, gfn_t gfn)
 271 {
 272         gpa_t gaddr;
 273
 274         ASSERT(*shadow_pte == 0);
 275         access_bits &= *gpde;
 276         gaddr = (gpa_t)gfn << PAGE_SHIFT;
 277         if (PTTYPE == 32 && is_cpuid_PSE36())
 278                 gaddr |= (*gpde & PT32_DIR_PSE36_MASK) <<
 279                         (32 - PT32_DIR_PSE36_SHIFT);
 280         *shadow_pte = *gpde & PT_PTE_COPY_MASK;
 281         FNAME(set_pte_common)(vcpu, shadow_pte, gaddr,
 282                               gpde, access_bits, gfn);
 283 }
 284
 285 /*
 286  * Fetch a shadow pte for a specific level in the paging hierarchy.
 287  */
 288 static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
 289                               struct guest_walker *walker)
 290 {
 291         hpa_t shadow_addr;
 292         int level;
 293         u64 *shadow_ent;
 294         u64 *prev_shadow_ent = NULL;
 295         pt_element_t *guest_ent = walker->ptep;
 296
 297         if (!is_present_pte(*guest_ent))
 298                 return NULL;
 299
 300         shadow_addr = vcpu->mmu.root_hpa;
 301         level = vcpu->mmu.shadow_root_level;
 302         if (level == PT32E_ROOT_LEVEL) {
 303                 shadow_addr = vcpu->mmu.pae_root[(addr >> 30) & 3];
 304                 shadow_addr &= PT64_BASE_ADDR_MASK;
 305                 --level;
 306         }
 307
 308         for (; ; level--) {
 309                 u32 index = SHADOW_PT_INDEX(addr, level);
 310                 struct kvm_mmu_page *shadow_page;
 311                 u64 shadow_pte;
 312                 int metaphysical;
 313                 gfn_t table_gfn;
 314                 unsigned hugepage_access = 0;
 315
 316                 shadow_ent = ((u64 *)__va(shadow_addr)) + index;
 317                 if (is_present_pte(*shadow_ent) || is_io_pte(*shadow_ent)) {
 318                         if (level == PT_PAGE_TABLE_LEVEL)
 319                                 return shadow_ent;
 320                         shadow_addr = *shadow_ent & PT64_BASE_ADDR_MASK;
 321                         prev_shadow_ent = shadow_ent;
 322                         continue;
 323                 }
 324
 325                 if (level == PT_PAGE_TABLE_LEVEL)
 326                         break;
 327
 328                 if (level - 1 == PT_PAGE_TABLE_LEVEL
 329                     && walker->level == PT_DIRECTORY_LEVEL) {
 330                         metaphysical = 1;
 331                         hugepage_access = *guest_ent;
 332                         hugepage_access &= PT_USER_MASK | PT_WRITABLE_MASK;
 333                         hugepage_access >>= PT_WRITABLE_SHIFT;
 334                         table_gfn = (*guest_ent & PT_BASE_ADDR_MASK)
 335                                 >> PAGE_SHIFT;
 336                 } else {
 337                         metaphysical = 0;
 338                         table_gfn = walker->table_gfn[level - 2];
 339                 }
 340                 shadow_page = kvm_mmu_get_page(vcpu, table_gfn, addr, level-1,
 341                                                metaphysical, hugepage_access,
 342                                                shadow_ent);
 343                 shadow_addr = __pa(shadow_page->spt);
 344                 shadow_pte = shadow_addr | PT_PRESENT_MASK | PT_ACCESSED_MASK
 345                         | PT_WRITABLE_MASK | PT_USER_MASK;
 346                 *shadow_ent = shadow_pte;
 347                 prev_shadow_ent = shadow_ent;
 348         }
 349
 350         if (walker->level == PT_DIRECTORY_LEVEL) {
 351                 if (prev_shadow_ent)
 352                         *prev_shadow_ent |= PT_SHADOW_PS_MARK;
 353                 FNAME(set_pde)(vcpu, guest_ent, shadow_ent,
 354                                walker->inherited_ar, walker->gfn);
 355         } else {
 356                 ASSERT(walker->level == PT_PAGE_TABLE_LEVEL);
 357                 FNAME(set_pte)(vcpu, guest_ent, shadow_ent,
 358                                walker->inherited_ar,
 359                                walker->gfn);
 360         }
 361         return shadow_ent;
 362 }
 363
 364 /*
 365  * The guest faulted for write.  We need to
 366  *
 367  * - check write permissions
 368  * - update the guest pte dirty bit
 369  * - update our own dirty page tracking structures
 370  */
 371 static int FNAME(fix_write_pf)(struct kvm_vcpu *vcpu,
 372                                u64 *shadow_ent,
 373                                struct guest_walker *walker,
 374                                gva_t addr,
 375                                int user,
 376                                int *write_pt)
 377 {
 378         pt_element_t *guest_ent;
 379         int writable_shadow;
 380         gfn_t gfn;
 381         struct kvm_mmu_page *page;
 382
 383         if (is_writeble_pte(*shadow_ent))
 384                 return !user || (*shadow_ent & PT_USER_MASK);
 385
 386         writable_shadow = *shadow_ent & PT_SHADOW_WRITABLE_MASK;
 387         if (user) {
 388                 /*
 389                  * User mode access.  Fail if it's a kernel page or a read-only
 390                  * page.
 391                  */
 392                 if (!(*shadow_ent & PT_SHADOW_USER_MASK) || !writable_shadow)
 393                         return 0;
 394                 ASSERT(*shadow_ent & PT_USER_MASK);
 395         } else
 396                 /*
 397                  * Kernel mode access.  Fail if it's a read-only page and
 398                  * supervisor write protection is enabled.
 399                  */
 400                 if (!writable_shadow) {
 401                         if (is_write_protection(vcpu))
 402                                 return 0;
 403                         *shadow_ent &= ~PT_USER_MASK;
 404                 }
 405
 406         guest_ent = walker->ptep;
 407
 408         if (!is_present_pte(*guest_ent)) {
 409                 *shadow_ent = 0;
 410                 return 0;
 411         }
 412
 413         gfn = walker->gfn;
 414
 415         if (user) {
 416                 /*
 417                  * Usermode page faults won't be for page table updates.
 418                  */
 419                 while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
 420                         pgprintk("%s: zap %lx %x\n",
 421                                  __FUNCTION__, gfn, page->role.word);
 422                         kvm_mmu_zap_page(vcpu, page);
 423                 }
 424         } else if (kvm_mmu_lookup_page(vcpu, gfn)) {
 425                 pgprintk("%s: found shadow page for %lx, marking ro\n",
 426                          __FUNCTION__, gfn);
 427                 mark_page_dirty(vcpu->kvm, gfn);
 428                 FNAME(mark_pagetable_dirty)(vcpu->kvm, walker);
 429                 *guest_ent |= PT_DIRTY_MASK;
 430                 *write_pt = 1;
 431                 return 0;
 432         }
 433         mark_page_dirty(vcpu->kvm, gfn);
 434         *shadow_ent |= PT_WRITABLE_MASK;
 435         FNAME(mark_pagetable_dirty)(vcpu->kvm, walker);
 436         *guest_ent |= PT_DIRTY_MASK;
 437         rmap_add(vcpu, shadow_ent);
 438
 439         return 1;
 440 }
 441
 442 /*
 443  * Page fault handler.  There are several causes for a page fault:
 444  *   - there is no shadow pte for the guest pte
 445  *   - write access through a shadow pte marked read only so that we can set
 446  *     the dirty bit
 447  *   - write access to a shadow pte marked read only so we can update the page
 448  *     dirty bitmap, when userspace requests it
 449  *   - mmio access; in this case we will never install a present shadow pte
 450  *   - normal guest page fault due to the guest pte marked not present, not
 451  *     writable, or not executable
 452  *
 453  *  Returns: 1 if we need to emulate the instruction, 0 otherwise, or
 454  *           a negative value on error.
 455  */
 456 static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,
 457                                u32 error_code)
 458 {
 459         int write_fault = error_code & PFERR_WRITE_MASK;
 460         int user_fault = error_code & PFERR_USER_MASK;
 461         int fetch_fault = error_code & PFERR_FETCH_MASK;
 462         struct guest_walker walker;
 463         u64 *shadow_pte;
 464         int fixed;
 465         int write_pt = 0;
 466         int r;
 467
 468         pgprintk("%s: addr %lx err %x\n", __FUNCTION__, addr, error_code);
 469         kvm_mmu_audit(vcpu, "pre page fault");
 470
 471         r = mmu_topup_memory_caches(vcpu);
 472         if (r)
 473                 return r;
 474
 475         /*
 476          * Look up the shadow pte for the faulting address.
 477          */
 478         r = FNAME(walk_addr)(&walker, vcpu, addr, write_fault, user_fault,
 479                              fetch_fault);
 480
 481         /*
 482          * The page is not mapped by the guest.  Let the guest handle it.
 483          */
 484         if (!r) {
 485                 pgprintk("%s: guest page fault\n", __FUNCTION__);
 486                 inject_page_fault(vcpu, addr, walker.error_code);
 487                 FNAME(release_walker)(&walker);
 488                 vcpu->last_pt_write_count = 0; /* reset fork detector */
 489                 return 0;
 490         }
 491
 492         shadow_pte = FNAME(fetch)(vcpu, addr, &walker);
 493         pgprintk("%s: shadow pte %p %llx\n", __FUNCTION__,
 494                  shadow_pte, *shadow_pte);
 495
 496         /*
 497          * Update the shadow pte.
 498          */
 499         if (write_fault)
 500                 fixed = FNAME(fix_write_pf)(vcpu, shadow_pte, &walker, addr,
 501                                             user_fault, &write_pt);
 502         else
 503                 fixed = fix_read_pf(shadow_pte);
 504
 505         pgprintk("%s: updated shadow pte %p %llx\n", __FUNCTION__,
 506                  shadow_pte, *shadow_pte);
 507
 508         FNAME(release_walker)(&walker);
 509
 510         if (!write_pt)
 511                 vcpu->last_pt_write_count = 0; /* reset fork detector */
 512
 513         /*
 514          * mmio: emulate if accessible, otherwise its a guest fault.
 515          */
 516         if (is_io_pte(*shadow_pte))
 517                 return 1;
 518
 519         ++vcpu->stat.pf_fixed;
 520         kvm_mmu_audit(vcpu, "post page fault (fixed)");
 521
 522         return write_pt;
 523 }
 524
 525 static gpa_t FNAME(gva_to_gpa)(struct kvm_vcpu *vcpu, gva_t vaddr)
 526 {
 527         struct guest_walker walker;
 528         gpa_t gpa = UNMAPPED_GVA;
 529         int r;
 530
 531         r = FNAME(walk_addr)(&walker, vcpu, vaddr, 0, 0, 0);
 532
 533         if (r) {
 534                 gpa = (gpa_t)walker.gfn << PAGE_SHIFT;
 535                 gpa |= vaddr & ~PAGE_MASK;
 536         }
 537
 538         FNAME(release_walker)(&walker);
 539         return gpa;
 540 }
 541
 542 #undef pt_element_t
 543 #undef guest_walker
 544 #undef FNAME
 545 #undef PT_BASE_ADDR_MASK
 546 #undef PT_INDEX
 547 #undef SHADOW_PT_INDEX
 548 #undef PT_LEVEL_MASK
 549 #undef PT_PTE_COPY_MASK
 550 #undef PT_NON_PTE_COPY_MASK
 551 #undef PT_DIR_BASE_ADDR_MASK
 552 #undef PT_MAX_FULL_LEVELS