arch/x86/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_LEVEL_BITS PT64_LEVEL_BITS
  35         #ifdef CONFIG_X86_64
  36         #define PT_MAX_FULL_LEVELS 4
  37         #define CMPXCHG cmpxchg
  38         #else
  39         #define CMPXCHG cmpxchg64
  40         #define PT_MAX_FULL_LEVELS 2
  41         #endif
  42 #elif PTTYPE == 32
  43         #define pt_element_t u32
  44         #define guest_walker guest_walker32
  45         #define FNAME(name) paging##32_##name
  46         #define PT_BASE_ADDR_MASK PT32_BASE_ADDR_MASK
  47         #define PT_DIR_BASE_ADDR_MASK PT32_DIR_BASE_ADDR_MASK
  48         #define PT_INDEX(addr, level) PT32_INDEX(addr, level)
  49         #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
  50         #define PT_LEVEL_MASK(level) PT32_LEVEL_MASK(level)
  51         #define PT_LEVEL_BITS PT32_LEVEL_BITS
  52         #define PT_MAX_FULL_LEVELS 2
  53         #define CMPXCHG cmpxchg
  54 #else
  55         #error Invalid PTTYPE value
  56 #endif
  57
  58 #define gpte_to_gfn FNAME(gpte_to_gfn)
  59 #define gpte_to_gfn_pde FNAME(gpte_to_gfn_pde)
  60
  61 /*
  62  * The guest_walker structure emulates the behavior of the hardware page
  63  * table walker.
  64  */
  65 struct guest_walker {
  66         int level;
  67         gfn_t table_gfn[PT_MAX_FULL_LEVELS];
  68         pt_element_t ptes[PT_MAX_FULL_LEVELS];
  69         gpa_t pte_gpa[PT_MAX_FULL_LEVELS];
  70         unsigned pt_access;
  71         unsigned pte_access;
  72         gfn_t gfn;
  73         u32 error_code;
  74 };
  75
  76 static gfn_t gpte_to_gfn(pt_element_t gpte)
  77 {
  78         return (gpte & PT_BASE_ADDR_MASK) >> PAGE_SHIFT;
  79 }
  80
  81 static gfn_t gpte_to_gfn_pde(pt_element_t gpte)
  82 {
  83         return (gpte & PT_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
  84 }
  85
  86 static bool FNAME(cmpxchg_gpte)(struct kvm *kvm,
  87                          gfn_t table_gfn, unsigned index,
  88                          pt_element_t orig_pte, pt_element_t new_pte)
  89 {
  90         pt_element_t ret;
  91         pt_element_t *table;
  92         struct page *page;
  93
  94         down_read(&current->mm->mmap_sem);
  95         page = gfn_to_page(kvm, table_gfn);
  96         up_read(&current->mm->mmap_sem);
  97
  98         table = kmap_atomic(page, KM_USER0);
  99
 100         ret = CMPXCHG(&table[index], orig_pte, new_pte);
 101
 102         kunmap_atomic(table, KM_USER0);
 103
 104         kvm_release_page_dirty(page);
 105
 106         return (ret != orig_pte);
 107 }
 108
 109 static unsigned FNAME(gpte_access)(struct kvm_vcpu *vcpu, pt_element_t gpte)
 110 {
 111         unsigned access;
 112
 113         access = (gpte & (PT_WRITABLE_MASK | PT_USER_MASK)) | ACC_EXEC_MASK;
 114 #if PTTYPE == 64
 115         if (is_nx(vcpu))
 116                 access &= ~(gpte >> PT64_NX_SHIFT);
 117 #endif
 118         return access;
 119 }
 120
 121 /*
 122  * Fetch a guest pte for a guest virtual address
 123  */
 124 static int FNAME(walk_addr)(struct guest_walker *walker,
 125                             struct kvm_vcpu *vcpu, gva_t addr,
 126                             int write_fault, int user_fault, int fetch_fault)
 127 {
 128         pt_element_t pte;
 129         gfn_t table_gfn;
 130         unsigned index, pt_access, pte_access;
 131         gpa_t pte_gpa;
 132
 133         pgprintk("%s: addr %lx\n", __func__, addr);
 134 walk:
 135         walker->level = vcpu->arch.mmu.root_level;
 136         pte = vcpu->arch.cr3;
 137 #if PTTYPE == 64
 138         if (!is_long_mode(vcpu)) {
 139                 pte = vcpu->arch.pdptrs[(addr >> 30) & 3];
 140                 if (!is_present_pte(pte))
 141                         goto not_present;
 142                 --walker->level;
 143         }
 144 #endif
 145         ASSERT((!is_long_mode(vcpu) && is_pae(vcpu)) ||
 146                (vcpu->arch.cr3 & CR3_NONPAE_RESERVED_BITS) == 0);
 147
 148         pt_access = ACC_ALL;
 149
 150         for (;;) {
 151                 index = PT_INDEX(addr, walker->level);
 152
 153                 table_gfn = gpte_to_gfn(pte);
 154                 pte_gpa = gfn_to_gpa(table_gfn);
 155                 pte_gpa += index * sizeof(pt_element_t);
 156                 walker->table_gfn[walker->level - 1] = table_gfn;
 157                 walker->pte_gpa[walker->level - 1] = pte_gpa;
 158                 pgprintk("%s: table_gfn[%d] %lx\n", __func__,
 159                          walker->level - 1, table_gfn);
 160
 161                 kvm_read_guest(vcpu->kvm, pte_gpa, &pte, sizeof(pte));
 162
 163                 if (!is_present_pte(pte))
 164                         goto not_present;
 165
 166                 if (write_fault && !is_writeble_pte(pte))
 167                         if (user_fault || is_write_protection(vcpu))
 168                                 goto access_error;
 169
 170                 if (user_fault && !(pte & PT_USER_MASK))
 171                         goto access_error;
 172
 173 #if PTTYPE == 64
 174                 if (fetch_fault && is_nx(vcpu) && (pte & PT64_NX_MASK))
 175                         goto access_error;
 176 #endif
 177
 178                 if (!(pte & PT_ACCESSED_MASK)) {
 179                         mark_page_dirty(vcpu->kvm, table_gfn);
 180                         if (FNAME(cmpxchg_gpte)(vcpu->kvm, table_gfn,
 181                             index, pte, pte|PT_ACCESSED_MASK))
 182                                 goto walk;
 183                         pte |= PT_ACCESSED_MASK;
 184                 }
 185
 186                 pte_access = pt_access & FNAME(gpte_access)(vcpu, pte);
 187
 188                 walker->ptes[walker->level - 1] = pte;
 189
 190                 if (walker->level == PT_PAGE_TABLE_LEVEL) {
 191                         walker->gfn = gpte_to_gfn(pte);
 192                         break;
 193                 }
 194
 195                 if (walker->level == PT_DIRECTORY_LEVEL
 196                     && (pte & PT_PAGE_SIZE_MASK)
 197                     && (PTTYPE == 64 || is_pse(vcpu))) {
 198                         walker->gfn = gpte_to_gfn_pde(pte);
 199                         walker->gfn += PT_INDEX(addr, PT_PAGE_TABLE_LEVEL);
 200                         if (PTTYPE == 32 && is_cpuid_PSE36())
 201                                 walker->gfn += pse36_gfn_delta(pte);
 202                         break;
 203                 }
 204
 205                 pt_access = pte_access;
 206                 --walker->level;
 207         }
 208
 209         if (write_fault && !is_dirty_pte(pte)) {
 210                 bool ret;
 211
 212                 mark_page_dirty(vcpu->kvm, table_gfn);
 213                 ret = FNAME(cmpxchg_gpte)(vcpu->kvm, table_gfn, index, pte,
 214                             pte|PT_DIRTY_MASK);
 215                 if (ret)
 216                         goto walk;
 217                 pte |= PT_DIRTY_MASK;
 218                 kvm_mmu_pte_write(vcpu, pte_gpa, (u8 *)&pte, sizeof(pte));
 219                 walker->ptes[walker->level - 1] = pte;
 220         }
 221
 222         walker->pt_access = pt_access;
 223         walker->pte_access = pte_access;
 224         pgprintk("%s: pte %llx pte_access %x pt_access %x\n",
 225                  __func__, (u64)pte, pt_access, pte_access);
 226         return 1;
 227
 228 not_present:
 229         walker->error_code = 0;
 230         goto err;
 231
 232 access_error:
 233         walker->error_code = PFERR_PRESENT_MASK;
 234
 235 err:
 236         if (write_fault)
 237                 walker->error_code |= PFERR_WRITE_MASK;
 238         if (user_fault)
 239                 walker->error_code |= PFERR_USER_MASK;
 240         if (fetch_fault)
 241                 walker->error_code |= PFERR_FETCH_MASK;
 242         return 0;
 243 }
 244
 245 static void FNAME(update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *page,
 246                               u64 *spte, const void *pte)
 247 {
 248         pt_element_t gpte;
 249         unsigned pte_access;
 250         pfn_t pfn;
 251         int largepage = vcpu->arch.update_pte.largepage;
 252
 253         gpte = *(const pt_element_t *)pte;
 254         if (~gpte & (PT_PRESENT_MASK | PT_ACCESSED_MASK)) {
 255                 if (!is_present_pte(gpte))
 256                         set_shadow_pte(spte, shadow_notrap_nonpresent_pte);
 257                 return;
 258         }
 259         pgprintk("%s: gpte %llx spte %p\n", __func__, (u64)gpte, spte);
 260         pte_access = page->role.access & FNAME(gpte_access)(vcpu, gpte);
 261         if (gpte_to_gfn(gpte) != vcpu->arch.update_pte.gfn)
 262                 return;
 263         pfn = vcpu->arch.update_pte.pfn;
 264         if (is_error_pfn(pfn))
 265                 return;
 266         if (mmu_notifier_retry(vcpu, vcpu->arch.update_pte.mmu_seq))
 267                 return;
 268         kvm_get_pfn(pfn);
 269         mmu_set_spte(vcpu, spte, page->role.access, pte_access, 0, 0,
 270                      gpte & PT_DIRTY_MASK, NULL, largepage, gpte_to_gfn(gpte),
 271                      pfn, true);
 272 }
 273
 274 /*
 275  * Fetch a shadow pte for a specific level in the paging hierarchy.
 276  */
 277 static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
 278                          struct guest_walker *walker,
 279                          int user_fault, int write_fault, int largepage,
 280                          int *ptwrite, pfn_t pfn)
 281 {
 282         hpa_t shadow_addr;
 283         int level;
 284         u64 *shadow_ent;
 285         unsigned access = walker->pt_access;
 286
 287         if (!is_present_pte(walker->ptes[walker->level - 1]))
 288                 return NULL;
 289
 290         shadow_addr = vcpu->arch.mmu.root_hpa;
 291         level = vcpu->arch.mmu.shadow_root_level;
 292         if (level == PT32E_ROOT_LEVEL) {
 293                 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
 294                 shadow_addr &= PT64_BASE_ADDR_MASK;
 295                 --level;
 296         }
 297
 298         for (; ; level--) {
 299                 u32 index = SHADOW_PT_INDEX(addr, level);
 300                 struct kvm_mmu_page *shadow_page;
 301                 u64 shadow_pte;
 302                 int metaphysical;
 303                 gfn_t table_gfn;
 304
 305                 shadow_ent = ((u64 *)__va(shadow_addr)) + index;
 306                 if (level == PT_PAGE_TABLE_LEVEL)
 307                         break;
 308
 309                 if (largepage && level == PT_DIRECTORY_LEVEL)
 310                         break;
 311
 312                 if (is_shadow_present_pte(*shadow_ent)
 313                     && !is_large_pte(*shadow_ent)) {
 314                         shadow_addr = *shadow_ent & PT64_BASE_ADDR_MASK;
 315                         continue;
 316                 }
 317
 318                 if (is_large_pte(*shadow_ent))
 319                         rmap_remove(vcpu->kvm, shadow_ent);
 320
 321                 if (level - 1 == PT_PAGE_TABLE_LEVEL
 322                     && walker->level == PT_DIRECTORY_LEVEL) {
 323                         metaphysical = 1;
 324                         if (!is_dirty_pte(walker->ptes[level - 1]))
 325                                 access &= ~ACC_WRITE_MASK;
 326                         table_gfn = gpte_to_gfn(walker->ptes[level - 1]);
 327                 } else {
 328                         metaphysical = 0;
 329                         table_gfn = walker->table_gfn[level - 2];
 330                 }
 331                 shadow_page = kvm_mmu_get_page(vcpu, table_gfn, addr, level-1,
 332                                                metaphysical, access,
 333                                                shadow_ent);
 334                 if (!metaphysical) {
 335                         int r;
 336                         pt_element_t curr_pte;
 337                         r = kvm_read_guest_atomic(vcpu->kvm,
 338                                                   walker->pte_gpa[level - 2],
 339                                                   &curr_pte, sizeof(curr_pte));
 340                         if (r || curr_pte != walker->ptes[level - 2]) {
 341                                 kvm_release_pfn_clean(pfn);
 342                                 return NULL;
 343                         }
 344                 }
 345                 shadow_addr = __pa(shadow_page->spt);
 346                 shadow_pte = shadow_addr | PT_PRESENT_MASK | PT_ACCESSED_MASK
 347                         | PT_WRITABLE_MASK | PT_USER_MASK;
 348                 set_shadow_pte(shadow_ent, shadow_pte);
 349         }
 350
 351         mmu_set_spte(vcpu, shadow_ent, access, walker->pte_access & access,
 352                      user_fault, write_fault,
 353                      walker->ptes[walker->level-1] & PT_DIRTY_MASK,
 354                      ptwrite, largepage, walker->gfn, pfn, false);
 355
 356         return shadow_ent;
 357 }
 358
 359 /*
 360  * Page fault handler.  There are several causes for a page fault:
 361  *   - there is no shadow pte for the guest pte
 362  *   - write access through a shadow pte marked read only so that we can set
 363  *     the dirty bit
 364  *   - write access to a shadow pte marked read only so we can update the page
 365  *     dirty bitmap, when userspace requests it
 366  *   - mmio access; in this case we will never install a present shadow pte
 367  *   - normal guest page fault due to the guest pte marked not present, not
 368  *     writable, or not executable
 369  *
 370  *  Returns: 1 if we need to emulate the instruction, 0 otherwise, or
 371  *           a negative value on error.
 372  */
 373 static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,
 374                                u32 error_code)
 375 {
 376         int write_fault = error_code & PFERR_WRITE_MASK;
 377         int user_fault = error_code & PFERR_USER_MASK;
 378         int fetch_fault = error_code & PFERR_FETCH_MASK;
 379         struct guest_walker walker;
 380         u64 *shadow_pte;
 381         int write_pt = 0;
 382         int r;
 383         pfn_t pfn;
 384         int largepage = 0;
 385         unsigned long mmu_seq;
 386
 387         pgprintk("%s: addr %lx err %x\n", __func__, addr, error_code);
 388         kvm_mmu_audit(vcpu, "pre page fault");
 389
 390         r = mmu_topup_memory_caches(vcpu);
 391         if (r)
 392                 return r;
 393
 394         /*
 395          * Look up the shadow pte for the faulting address.
 396          */
 397         r = FNAME(walk_addr)(&walker, vcpu, addr, write_fault, user_fault,
 398                              fetch_fault);
 399
 400         /*
 401          * The page is not mapped by the guest.  Let the guest handle it.
 402          */
 403         if (!r) {
 404                 pgprintk("%s: guest page fault\n", __func__);
 405                 inject_page_fault(vcpu, addr, walker.error_code);
 406                 vcpu->arch.last_pt_write_count = 0; /* reset fork detector */
 407                 return 0;
 408         }
 409
 410         down_read(&current->mm->mmap_sem);
 411         if (walker.level == PT_DIRECTORY_LEVEL) {
 412                 gfn_t large_gfn;
 413                 large_gfn = walker.gfn & ~(KVM_PAGES_PER_HPAGE-1);
 414                 if (is_largepage_backed(vcpu, large_gfn)) {
 415                         walker.gfn = large_gfn;
 416                         largepage = 1;
 417                 }
 418         }
 419         mmu_seq = vcpu->kvm->mmu_notifier_seq;
 420         /* implicit mb(), we'll read before PT lock is unlocked */
 421         pfn = gfn_to_pfn(vcpu->kvm, walker.gfn);
 422         up_read(&current->mm->mmap_sem);
 423
 424         /* mmio */
 425         if (is_error_pfn(pfn)) {
 426                 pgprintk("gfn %lx is mmio\n", walker.gfn);
 427                 kvm_release_pfn_clean(pfn);
 428                 return 1;
 429         }
 430
 431         spin_lock(&vcpu->kvm->mmu_lock);
 432         if (mmu_notifier_retry(vcpu, mmu_seq))
 433                 goto out_unlock;
 434         kvm_mmu_free_some_pages(vcpu);
 435         shadow_pte = FNAME(fetch)(vcpu, addr, &walker, user_fault, write_fault,
 436                                   largepage, &write_pt, pfn);
 437
 438         pgprintk("%s: shadow pte %p %llx ptwrite %d\n", __func__,
 439                  shadow_pte, *shadow_pte, write_pt);
 440
 441         if (!write_pt)
 442                 vcpu->arch.last_pt_write_count = 0; /* reset fork detector */
 443
 444         ++vcpu->stat.pf_fixed;
 445         kvm_mmu_audit(vcpu, "post page fault (fixed)");
 446         spin_unlock(&vcpu->kvm->mmu_lock);
 447
 448         return write_pt;
 449
 450 out_unlock:
 451         spin_unlock(&vcpu->kvm->mmu_lock);
 452         kvm_release_pfn_clean(pfn);
 453         return 0;
 454 }
 455
 456 static gpa_t FNAME(gva_to_gpa)(struct kvm_vcpu *vcpu, gva_t vaddr)
 457 {
 458         struct guest_walker walker;
 459         gpa_t gpa = UNMAPPED_GVA;
 460         int r;
 461
 462         r = FNAME(walk_addr)(&walker, vcpu, vaddr, 0, 0, 0);
 463
 464         if (r) {
 465                 gpa = gfn_to_gpa(walker.gfn);
 466                 gpa |= vaddr & ~PAGE_MASK;
 467         }
 468
 469         return gpa;
 470 }
 471
 472 static void FNAME(prefetch_page)(struct kvm_vcpu *vcpu,
 473                                  struct kvm_mmu_page *sp)
 474 {
 475         int i, j, offset, r;
 476         pt_element_t pt[256 / sizeof(pt_element_t)];
 477         gpa_t pte_gpa;
 478
 479         if (sp->role.metaphysical
 480             || (PTTYPE == 32 && sp->role.level > PT_PAGE_TABLE_LEVEL)) {
 481                 nonpaging_prefetch_page(vcpu, sp);
 482                 return;
 483         }
 484
 485         pte_gpa = gfn_to_gpa(sp->gfn);
 486         if (PTTYPE == 32) {
 487                 offset = sp->role.quadrant << PT64_LEVEL_BITS;
 488                 pte_gpa += offset * sizeof(pt_element_t);
 489         }
 490
 491         for (i = 0; i < PT64_ENT_PER_PAGE; i += ARRAY_SIZE(pt)) {
 492                 r = kvm_read_guest_atomic(vcpu->kvm, pte_gpa, pt, sizeof pt);
 493                 pte_gpa += ARRAY_SIZE(pt) * sizeof(pt_element_t);
 494                 for (j = 0; j < ARRAY_SIZE(pt); ++j)
 495                         if (r || is_present_pte(pt[j]))
 496                                 sp->spt[i+j] = shadow_trap_nonpresent_pte;
 497                         else
 498                                 sp->spt[i+j] = shadow_notrap_nonpresent_pte;
 499         }
 500 }
 501
 502 #undef pt_element_t
 503 #undef guest_walker
 504 #undef FNAME
 505 #undef PT_BASE_ADDR_MASK
 506 #undef PT_INDEX
 507 #undef SHADOW_PT_INDEX
 508 #undef PT_LEVEL_MASK
 509 #undef PT_DIR_BASE_ADDR_MASK
 510 #undef PT_LEVEL_BITS
 511 #undef PT_MAX_FULL_LEVELS
 512 #undef gpte_to_gfn
 513 #undef gpte_to_gfn_pde
 514 #undef CMPXCHG