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         #ifdef CONFIG_X86_64
  35         #define PT_MAX_FULL_LEVELS 4
  36         #else
  37         #define PT_MAX_FULL_LEVELS 2
  38         #endif
  39 #elif PTTYPE == 32
  40         #define pt_element_t u32
  41         #define guest_walker guest_walker32
  42         #define FNAME(name) paging##32_##name
  43         #define PT_BASE_ADDR_MASK PT32_BASE_ADDR_MASK
  44         #define PT_DIR_BASE_ADDR_MASK PT32_DIR_BASE_ADDR_MASK
  45         #define PT_INDEX(addr, level) PT32_INDEX(addr, level)
  46         #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
  47         #define PT_LEVEL_MASK(level) PT32_LEVEL_MASK(level)
  48         #define PT_MAX_FULL_LEVELS 2
  49 #else
  50         #error Invalid PTTYPE value
  51 #endif
  52
  53 /*
  54  * The guest_walker structure emulates the behavior of the hardware page
  55  * table walker.
  56  */
  57 struct guest_walker {
  58         int level;
  59         gfn_t table_gfn[PT_MAX_FULL_LEVELS];
  60         pt_element_t *table;
  61         pt_element_t pte;
  62         pt_element_t *ptep;
  63         struct page *page;
  64         int index;
  65         pt_element_t inherited_ar;
  66         gfn_t gfn;
  67         u32 error_code;
  68 };
  69
  70 /*
  71  * Fetch a guest pte for a guest virtual address
  72  */
  73 static int FNAME(walk_addr)(struct guest_walker *walker,
  74                             struct kvm_vcpu *vcpu, gva_t addr,
  75                             int write_fault, int user_fault, int fetch_fault)
  76 {
  77         hpa_t hpa;
  78         struct kvm_memory_slot *slot;
  79         pt_element_t *ptep;
  80         pt_element_t root;
  81         gfn_t table_gfn;
  82
  83         pgprintk("%s: addr %lx\n", __FUNCTION__, addr);
  84         walker->level = vcpu->mmu.root_level;
  85         walker->table = NULL;
  86         walker->page = NULL;
  87         walker->ptep = NULL;
  88         root = vcpu->cr3;
  89 #if PTTYPE == 64
  90         if (!is_long_mode(vcpu)) {
  91                 walker->ptep = &vcpu->pdptrs[(addr >> 30) & 3];
  92                 root = *walker->ptep;
  93                 walker->pte = root;
  94                 if (!(root & PT_PRESENT_MASK))
  95                         goto not_present;
  96                 --walker->level;
  97         }
  98 #endif
  99         table_gfn = (root & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 100         walker->table_gfn[walker->level - 1] = table_gfn;
 101         pgprintk("%s: table_gfn[%d] %lx\n", __FUNCTION__,
 102                  walker->level - 1, table_gfn);
 103         slot = gfn_to_memslot(vcpu->kvm, table_gfn);
 104         hpa = safe_gpa_to_hpa(vcpu, root & PT64_BASE_ADDR_MASK);
 105         walker->page = pfn_to_page(hpa >> PAGE_SHIFT);
 106         walker->table = kmap_atomic(walker->page, KM_USER0);
 107
 108         ASSERT((!is_long_mode(vcpu) && is_pae(vcpu)) ||
 109                (vcpu->cr3 & CR3_NONPAE_RESERVED_BITS) == 0);
 110
 111         walker->inherited_ar = PT_USER_MASK | PT_WRITABLE_MASK;
 112
 113         for (;;) {
 114                 int index = PT_INDEX(addr, walker->level);
 115                 hpa_t paddr;
 116
 117                 ptep = &walker->table[index];
 118                 walker->index = index;
 119                 ASSERT(((unsigned long)walker->table & PAGE_MASK) ==
 120                        ((unsigned long)ptep & PAGE_MASK));
 121
 122                 if (!is_present_pte(*ptep))
 123                         goto not_present;
 124
 125                 if (write_fault && !is_writeble_pte(*ptep))
 126                         if (user_fault || is_write_protection(vcpu))
 127                                 goto access_error;
 128
 129                 if (user_fault && !(*ptep & PT_USER_MASK))
 130                         goto access_error;
 131
 132 #if PTTYPE == 64
 133                 if (fetch_fault && is_nx(vcpu) && (*ptep & PT64_NX_MASK))
 134                         goto access_error;
 135 #endif
 136
 137                 if (!(*ptep & PT_ACCESSED_MASK)) {
 138                         mark_page_dirty(vcpu->kvm, table_gfn);
 139                         *ptep |= PT_ACCESSED_MASK;
 140                 }
 141
 142                 if (walker->level == PT_PAGE_TABLE_LEVEL) {
 143                         walker->gfn = (*ptep & PT_BASE_ADDR_MASK)
 144                                 >> PAGE_SHIFT;
 145                         break;
 146                 }
 147
 148                 if (walker->level == PT_DIRECTORY_LEVEL
 149                     && (*ptep & PT_PAGE_SIZE_MASK)
 150                     && (PTTYPE == 64 || is_pse(vcpu))) {
 151                         walker->gfn = (*ptep & PT_DIR_BASE_ADDR_MASK)
 152                                 >> PAGE_SHIFT;
 153                         walker->gfn += PT_INDEX(addr, PT_PAGE_TABLE_LEVEL);
 154                         break;
 155                 }
 156
 157                 walker->inherited_ar &= walker->table[index];
 158                 table_gfn = (*ptep & PT_BASE_ADDR_MASK) >> PAGE_SHIFT;
 159                 kunmap_atomic(walker->table, KM_USER0);
 160                 paddr = safe_gpa_to_hpa(vcpu, table_gfn << PAGE_SHIFT);
 161                 walker->page = pfn_to_page(paddr >> PAGE_SHIFT);
 162                 walker->table = kmap_atomic(walker->page, KM_USER0);
 163                 --walker->level;
 164                 walker->table_gfn[walker->level - 1 ] = table_gfn;
 165                 pgprintk("%s: table_gfn[%d] %lx\n", __FUNCTION__,
 166                          walker->level - 1, table_gfn);
 167         }
 168         walker->pte = *ptep;
 169         if (walker->page)
 170                 walker->ptep = NULL;
 171         if (walker->table)
 172                 kunmap_atomic(walker->table, KM_USER0);
 173         pgprintk("%s: pte %llx\n", __FUNCTION__, (u64)*ptep);
 174         return 1;
 175
 176 not_present:
 177         walker->error_code = 0;
 178         goto err;
 179
 180 access_error:
 181         walker->error_code = PFERR_PRESENT_MASK;
 182
 183 err:
 184         if (write_fault)
 185                 walker->error_code |= PFERR_WRITE_MASK;
 186         if (user_fault)
 187                 walker->error_code |= PFERR_USER_MASK;
 188         if (fetch_fault)
 189                 walker->error_code |= PFERR_FETCH_MASK;
 190         if (walker->table)
 191                 kunmap_atomic(walker->table, KM_USER0);
 192         return 0;
 193 }
 194
 195 static void FNAME(mark_pagetable_dirty)(struct kvm *kvm,
 196                                         struct guest_walker *walker)
 197 {
 198         mark_page_dirty(kvm, walker->table_gfn[walker->level - 1]);
 199 }
 200
 201 static void FNAME(set_pte_common)(struct kvm_vcpu *vcpu,
 202                                   u64 *shadow_pte,
 203                                   gpa_t gaddr,
 204                                   pt_element_t gpte,
 205                                   u64 access_bits,
 206                                   int user_fault,
 207                                   int write_fault,
 208                                   int *ptwrite,
 209                                   struct guest_walker *walker,
 210                                   gfn_t gfn)
 211 {
 212         hpa_t paddr;
 213         int dirty = gpte & PT_DIRTY_MASK;
 214         u64 spte = *shadow_pte;
 215         int was_rmapped = is_rmap_pte(spte);
 216
 217         pgprintk("%s: spte %llx gpte %llx access %llx write_fault %d"
 218                  " user_fault %d gfn %lx\n",
 219                  __FUNCTION__, spte, (u64)gpte, access_bits,
 220                  write_fault, user_fault, gfn);
 221
 222         if (write_fault && !dirty) {
 223                 pt_element_t *guest_ent, *tmp = NULL;
 224
 225                 if (walker->ptep)
 226                         guest_ent = walker->ptep;
 227                 else {
 228                         tmp = kmap_atomic(walker->page, KM_USER0);
 229                         guest_ent = &tmp[walker->index];
 230                 }
 231
 232                 *guest_ent |= PT_DIRTY_MASK;
 233                 if (!walker->ptep)
 234                         kunmap_atomic(tmp, KM_USER0);
 235                 dirty = 1;
 236                 FNAME(mark_pagetable_dirty)(vcpu->kvm, walker);
 237         }
 238
 239         spte |= PT_PRESENT_MASK | PT_ACCESSED_MASK | PT_DIRTY_MASK;
 240         spte |= gpte & PT64_NX_MASK;
 241         if (!dirty)
 242                 access_bits &= ~PT_WRITABLE_MASK;
 243
 244         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 245
 246         spte |= PT_PRESENT_MASK;
 247         if (access_bits & PT_USER_MASK)
 248                 spte |= PT_USER_MASK;
 249
 250         if (is_error_hpa(paddr)) {
 251                 spte |= gaddr;
 252                 spte |= PT_SHADOW_IO_MARK;
 253                 spte &= ~PT_PRESENT_MASK;
 254                 set_shadow_pte(shadow_pte, spte);
 255                 return;
 256         }
 257
 258         spte |= paddr;
 259
 260         if ((access_bits & PT_WRITABLE_MASK)
 261             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
 262                 struct kvm_mmu_page *shadow;
 263
 264                 spte |= PT_WRITABLE_MASK;
 265                 if (user_fault) {
 266                         mmu_unshadow(vcpu, gfn);
 267                         goto unshadowed;
 268                 }
 269
 270                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 271                 if (shadow) {
 272                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 273                                  __FUNCTION__, gfn);
 274                         access_bits &= ~PT_WRITABLE_MASK;
 275                         if (is_writeble_pte(spte)) {
 276                                 spte &= ~PT_WRITABLE_MASK;
 277                                 kvm_x86_ops->tlb_flush(vcpu);
 278                         }
 279                         if (write_fault)
 280                                 *ptwrite = 1;
 281                 }
 282         }
 283
 284 unshadowed:
 285
 286         if (access_bits & PT_WRITABLE_MASK)
 287                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 288
 289         set_shadow_pte(shadow_pte, spte);
 290         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 291         if (!was_rmapped)
 292                 rmap_add(vcpu, shadow_pte);
 293 }
 294
 295 static void FNAME(set_pte)(struct kvm_vcpu *vcpu, pt_element_t gpte,
 296                            u64 *shadow_pte, u64 access_bits,
 297                            int user_fault, int write_fault, int *ptwrite,
 298                            struct guest_walker *walker, gfn_t gfn)
 299 {
 300         access_bits &= gpte;
 301         FNAME(set_pte_common)(vcpu, shadow_pte, gpte & PT_BASE_ADDR_MASK,
 302                               gpte, access_bits, user_fault, write_fault,
 303                               ptwrite, walker, gfn);
 304 }
 305
 306 static void FNAME(update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *page,
 307                               u64 *spte, const void *pte, int bytes)
 308 {
 309         pt_element_t gpte;
 310
 311         if (bytes < sizeof(pt_element_t))
 312                 return;
 313         gpte = *(const pt_element_t *)pte;
 314         if (~gpte & (PT_PRESENT_MASK | PT_ACCESSED_MASK))
 315                 return;
 316         pgprintk("%s: gpte %llx spte %p\n", __FUNCTION__, (u64)gpte, spte);
 317         FNAME(set_pte)(vcpu, gpte, spte, PT_USER_MASK | PT_WRITABLE_MASK, 0,
 318                        0, NULL, NULL,
 319                        (gpte & PT_BASE_ADDR_MASK) >> PAGE_SHIFT);
 320 }
 321
 322 static void FNAME(set_pde)(struct kvm_vcpu *vcpu, pt_element_t gpde,
 323                            u64 *shadow_pte, u64 access_bits,
 324                            int user_fault, int write_fault, int *ptwrite,
 325                            struct guest_walker *walker, gfn_t gfn)
 326 {
 327         gpa_t gaddr;
 328
 329         access_bits &= gpde;
 330         gaddr = (gpa_t)gfn << PAGE_SHIFT;
 331         if (PTTYPE == 32 && is_cpuid_PSE36())
 332                 gaddr |= (gpde & PT32_DIR_PSE36_MASK) <<
 333                         (32 - PT32_DIR_PSE36_SHIFT);
 334         FNAME(set_pte_common)(vcpu, shadow_pte, gaddr,
 335                               gpde, access_bits, user_fault, write_fault,
 336                               ptwrite, walker, gfn);
 337 }
 338
 339 /*
 340  * Fetch a shadow pte for a specific level in the paging hierarchy.
 341  */
 342 static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
 343                          struct guest_walker *walker,
 344                          int user_fault, int write_fault, int *ptwrite)
 345 {
 346         hpa_t shadow_addr;
 347         int level;
 348         u64 *shadow_ent;
 349         u64 *prev_shadow_ent = NULL;
 350
 351         if (!is_present_pte(walker->pte))
 352                 return NULL;
 353
 354         shadow_addr = vcpu->mmu.root_hpa;
 355         level = vcpu->mmu.shadow_root_level;
 356         if (level == PT32E_ROOT_LEVEL) {
 357                 shadow_addr = vcpu->mmu.pae_root[(addr >> 30) & 3];
 358                 shadow_addr &= PT64_BASE_ADDR_MASK;
 359                 --level;
 360         }
 361
 362         for (; ; level--) {
 363                 u32 index = SHADOW_PT_INDEX(addr, level);
 364                 struct kvm_mmu_page *shadow_page;
 365                 u64 shadow_pte;
 366                 int metaphysical;
 367                 gfn_t table_gfn;
 368                 unsigned hugepage_access = 0;
 369
 370                 shadow_ent = ((u64 *)__va(shadow_addr)) + index;
 371                 if (is_present_pte(*shadow_ent) || is_io_pte(*shadow_ent)) {
 372                         if (level == PT_PAGE_TABLE_LEVEL)
 373                                 break;
 374                         shadow_addr = *shadow_ent & PT64_BASE_ADDR_MASK;
 375                         prev_shadow_ent = shadow_ent;
 376                         continue;
 377                 }
 378
 379                 if (level == PT_PAGE_TABLE_LEVEL)
 380                         break;
 381
 382                 if (level - 1 == PT_PAGE_TABLE_LEVEL
 383                     && walker->level == PT_DIRECTORY_LEVEL) {
 384                         metaphysical = 1;
 385                         hugepage_access = walker->pte;
 386                         hugepage_access &= PT_USER_MASK | PT_WRITABLE_MASK;
 387                         if (walker->pte & PT64_NX_MASK)
 388                                 hugepage_access |= (1 << 2);
 389                         hugepage_access >>= PT_WRITABLE_SHIFT;
 390                         table_gfn = (walker->pte & PT_BASE_ADDR_MASK)
 391                                 >> PAGE_SHIFT;
 392                 } else {
 393                         metaphysical = 0;
 394                         table_gfn = walker->table_gfn[level - 2];
 395                 }
 396                 shadow_page = kvm_mmu_get_page(vcpu, table_gfn, addr, level-1,
 397                                                metaphysical, hugepage_access,
 398                                                shadow_ent);
 399                 shadow_addr = __pa(shadow_page->spt);
 400                 shadow_pte = shadow_addr | PT_PRESENT_MASK | PT_ACCESSED_MASK
 401                         | PT_WRITABLE_MASK | PT_USER_MASK;
 402                 *shadow_ent = shadow_pte;
 403                 prev_shadow_ent = shadow_ent;
 404         }
 405
 406         if (walker->level == PT_DIRECTORY_LEVEL) {
 407                 FNAME(set_pde)(vcpu, walker->pte, shadow_ent,
 408                                walker->inherited_ar, user_fault, write_fault,
 409                                ptwrite, walker, walker->gfn);
 410         } else {
 411                 ASSERT(walker->level == PT_PAGE_TABLE_LEVEL);
 412                 FNAME(set_pte)(vcpu, walker->pte, shadow_ent,
 413                                walker->inherited_ar, user_fault, write_fault,
 414                                ptwrite, walker, walker->gfn);
 415         }
 416         return shadow_ent;
 417 }
 418
 419 /*
 420  * Page fault handler.  There are several causes for a page fault:
 421  *   - there is no shadow pte for the guest pte
 422  *   - write access through a shadow pte marked read only so that we can set
 423  *     the dirty bit
 424  *   - write access to a shadow pte marked read only so we can update the page
 425  *     dirty bitmap, when userspace requests it
 426  *   - mmio access; in this case we will never install a present shadow pte
 427  *   - normal guest page fault due to the guest pte marked not present, not
 428  *     writable, or not executable
 429  *
 430  *  Returns: 1 if we need to emulate the instruction, 0 otherwise, or
 431  *           a negative value on error.
 432  */
 433 static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,
 434                                u32 error_code)
 435 {
 436         int write_fault = error_code & PFERR_WRITE_MASK;
 437         int user_fault = error_code & PFERR_USER_MASK;
 438         int fetch_fault = error_code & PFERR_FETCH_MASK;
 439         struct guest_walker walker;
 440         u64 *shadow_pte;
 441         int write_pt = 0;
 442         int r;
 443
 444         pgprintk("%s: addr %lx err %x\n", __FUNCTION__, addr, error_code);
 445         kvm_mmu_audit(vcpu, "pre page fault");
 446
 447         r = mmu_topup_memory_caches(vcpu);
 448         if (r)
 449                 return r;
 450
 451         /*
 452          * Look up the shadow pte for the faulting address.
 453          */
 454         r = FNAME(walk_addr)(&walker, vcpu, addr, write_fault, user_fault,
 455                              fetch_fault);
 456
 457         /*
 458          * The page is not mapped by the guest.  Let the guest handle it.
 459          */
 460         if (!r) {
 461                 pgprintk("%s: guest page fault\n", __FUNCTION__);
 462                 inject_page_fault(vcpu, addr, walker.error_code);
 463                 vcpu->last_pt_write_count = 0; /* reset fork detector */
 464                 return 0;
 465         }
 466
 467         shadow_pte = FNAME(fetch)(vcpu, addr, &walker, user_fault, write_fault,
 468                                   &write_pt);
 469         pgprintk("%s: shadow pte %p %llx ptwrite %d\n", __FUNCTION__,
 470                  shadow_pte, *shadow_pte, write_pt);
 471
 472         if (!write_pt)
 473                 vcpu->last_pt_write_count = 0; /* reset fork detector */
 474
 475         /*
 476          * mmio: emulate if accessible, otherwise its a guest fault.
 477          */
 478         if (is_io_pte(*shadow_pte))
 479                 return 1;
 480
 481         ++vcpu->stat.pf_fixed;
 482         kvm_mmu_audit(vcpu, "post page fault (fixed)");
 483
 484         return write_pt;
 485 }
 486
 487 static gpa_t FNAME(gva_to_gpa)(struct kvm_vcpu *vcpu, gva_t vaddr)
 488 {
 489         struct guest_walker walker;
 490         gpa_t gpa = UNMAPPED_GVA;
 491         int r;
 492
 493         r = FNAME(walk_addr)(&walker, vcpu, vaddr, 0, 0, 0);
 494
 495         if (r) {
 496                 gpa = (gpa_t)walker.gfn << PAGE_SHIFT;
 497                 gpa |= vaddr & ~PAGE_MASK;
 498         }
 499
 500         return gpa;
 501 }
 502
 503 #undef pt_element_t
 504 #undef guest_walker
 505 #undef FNAME
 506 #undef PT_BASE_ADDR_MASK
 507 #undef PT_INDEX
 508 #undef SHADOW_PT_INDEX
 509 #undef PT_LEVEL_MASK
 510 #undef PT_DIR_BASE_ADDR_MASK
 511 #undef PT_MAX_FULL_LEVELS