arch/x86/mm/pgtable.c

   1 #include <linux/mm.h>
   2 #include <linux/gfp.h>
   3 #include <asm/pgalloc.h>
   4 #include <asm/pgtable.h>
   5 #include <asm/tlb.h>
   6 #include <asm/fixmap.h>
   7 #include <asm/mtrr.h>
   8
   9 #define PGALLOC_GFP GFP_KERNEL | __GFP_NOTRACK | __GFP_REPEAT | __GFP_ZERO
  10
  11 #ifdef CONFIG_HIGHPTE
  12 #define PGALLOC_USER_GFP __GFP_HIGHMEM
  13 #else
  14 #define PGALLOC_USER_GFP 0
  15 #endif
  16
  17 gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
  18
  19 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
  20 {
  21         return (pte_t *)__get_free_page(PGALLOC_GFP);
  22 }
  23
  24 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
  25 {
  26         struct page *pte;
  27
  28         pte = alloc_pages(__userpte_alloc_gfp, 0);
  29         if (!pte)
  30                 return NULL;
  31         if (!pgtable_page_ctor(pte)) {
  32                 __free_page(pte);
  33                 return NULL;
  34         }
  35         return pte;
  36 }
  37
  38 static int __init setup_userpte(char *arg)
  39 {
  40         if (!arg)
  41                 return -EINVAL;
  42
  43         /*
  44          * "userpte=nohigh" disables allocation of user pagetables in
  45          * high memory.
  46          */
  47         if (strcmp(arg, "nohigh") == 0)
  48                 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
  49         else
  50                 return -EINVAL;
  51         return 0;
  52 }
  53 early_param("userpte", setup_userpte);
  54
  55 void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
  56 {
  57         pgtable_page_dtor(pte);
  58         paravirt_release_pte(page_to_pfn(pte));
  59         tlb_remove_page(tlb, pte);
  60 }
  61
  62 #if CONFIG_PGTABLE_LEVELS > 2
  63 void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
  64 {
  65         struct page *page = virt_to_page(pmd);
  66         paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
  67         /*
  68          * NOTE! For PAE, any changes to the top page-directory-pointer-table
  69          * entries need a full cr3 reload to flush.
  70          */
  71 #ifdef CONFIG_X86_PAE
  72         tlb->need_flush_all = 1;
  73 #endif
  74         pgtable_pmd_page_dtor(page);
  75         tlb_remove_page(tlb, page);
  76 }
  77
  78 #if CONFIG_PGTABLE_LEVELS > 3
  79 void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
  80 {
  81         paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
  82         tlb_remove_page(tlb, virt_to_page(pud));
  83 }
  84 #endif  /* CONFIG_PGTABLE_LEVELS > 3 */
  85 #endif  /* CONFIG_PGTABLE_LEVELS > 2 */
  86
  87 static inline void pgd_list_add(pgd_t *pgd)
  88 {
  89         struct page *page = virt_to_page(pgd);
  90
  91         list_add(&page->lru, &pgd_list);
  92 }
  93
  94 static inline void pgd_list_del(pgd_t *pgd)
  95 {
  96         struct page *page = virt_to_page(pgd);
  97
  98         list_del(&page->lru);
  99 }
 100
 101 #define UNSHARED_PTRS_PER_PGD                           \
 102         (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
 103
 104
 105 static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
 106 {
 107         BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
 108         virt_to_page(pgd)->index = (pgoff_t)mm;
 109 }
 110
 111 struct mm_struct *pgd_page_get_mm(struct page *page)
 112 {
 113         return (struct mm_struct *)page->index;
 114 }
 115
 116 static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
 117 {
 118         /* If the pgd points to a shared pagetable level (either the
 119            ptes in non-PAE, or shared PMD in PAE), then just copy the
 120            references from swapper_pg_dir. */
 121         if (CONFIG_PGTABLE_LEVELS == 2 ||
 122             (CONFIG_PGTABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
 123             CONFIG_PGTABLE_LEVELS == 4) {
 124                 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
 125                                 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
 126                                 KERNEL_PGD_PTRS);
 127         }
 128
 129         /* list required to sync kernel mapping updates */
 130         if (!SHARED_KERNEL_PMD) {
 131                 pgd_set_mm(pgd, mm);
 132                 pgd_list_add(pgd);
 133         }
 134 }
 135
 136 static void pgd_dtor(pgd_t *pgd)
 137 {
 138         if (SHARED_KERNEL_PMD)
 139                 return;
 140
 141         spin_lock(&pgd_lock);
 142         pgd_list_del(pgd);
 143         spin_unlock(&pgd_lock);
 144 }
 145
 146 /*
 147  * List of all pgd's needed for non-PAE so it can invalidate entries
 148  * in both cached and uncached pgd's; not needed for PAE since the
 149  * kernel pmd is shared. If PAE were not to share the pmd a similar
 150  * tactic would be needed. This is essentially codepath-based locking
 151  * against pageattr.c; it is the unique case in which a valid change
 152  * of kernel pagetables can't be lazily synchronized by vmalloc faults.
 153  * vmalloc faults work because attached pagetables are never freed.
 154  * -- nyc
 155  */
 156
 157 #ifdef CONFIG_X86_PAE
 158 /*
 159  * In PAE mode, we need to do a cr3 reload (=tlb flush) when
 160  * updating the top-level pagetable entries to guarantee the
 161  * processor notices the update.  Since this is expensive, and
 162  * all 4 top-level entries are used almost immediately in a
 163  * new process's life, we just pre-populate them here.
 164  *
 165  * Also, if we're in a paravirt environment where the kernel pmd is
 166  * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
 167  * and initialize the kernel pmds here.
 168  */
 169 #define PREALLOCATED_PMDS       UNSHARED_PTRS_PER_PGD
 170
 171 void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
 172 {
 173         paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
 174
 175         /* Note: almost everything apart from _PAGE_PRESENT is
 176            reserved at the pmd (PDPT) level. */
 177         set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
 178
 179         /*
 180          * According to Intel App note "TLBs, Paging-Structure Caches,
 181          * and Their Invalidation", April 2007, document 317080-001,
 182          * section 8.1: in PAE mode we explicitly have to flush the
 183          * TLB via cr3 if the top-level pgd is changed...
 184          */
 185         flush_tlb_mm(mm);
 186 }
 187 #else  /* !CONFIG_X86_PAE */
 188
 189 /* No need to prepopulate any pagetable entries in non-PAE modes. */
 190 #define PREALLOCATED_PMDS       0
 191
 192 #endif  /* CONFIG_X86_PAE */
 193
 194 static void free_pmds(struct mm_struct *mm, pmd_t *pmds[])
 195 {
 196         int i;
 197
 198         for(i = 0; i < PREALLOCATED_PMDS; i++)
 199                 if (pmds[i]) {
 200                         pgtable_pmd_page_dtor(virt_to_page(pmds[i]));
 201                         free_page((unsigned long)pmds[i]);
 202                         mm_dec_nr_pmds(mm);
 203                 }
 204 }
 205
 206 static int preallocate_pmds(struct mm_struct *mm, pmd_t *pmds[])
 207 {
 208         int i;
 209         bool failed = false;
 210
 211         for(i = 0; i < PREALLOCATED_PMDS; i++) {
 212                 pmd_t *pmd = (pmd_t *)__get_free_page(PGALLOC_GFP);
 213                 if (!pmd)
 214                         failed = true;
 215                 if (pmd && !pgtable_pmd_page_ctor(virt_to_page(pmd))) {
 216                         free_page((unsigned long)pmd);
 217                         pmd = NULL;
 218                         failed = true;
 219                 }
 220                 if (pmd)
 221                         mm_inc_nr_pmds(mm);
 222                 pmds[i] = pmd;
 223         }
 224
 225         if (failed) {
 226                 free_pmds(mm, pmds);
 227                 return -ENOMEM;
 228         }
 229
 230         return 0;
 231 }
 232
 233 /*
 234  * Mop up any pmd pages which may still be attached to the pgd.
 235  * Normally they will be freed by munmap/exit_mmap, but any pmd we
 236  * preallocate which never got a corresponding vma will need to be
 237  * freed manually.
 238  */
 239 static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
 240 {
 241         int i;
 242
 243         for(i = 0; i < PREALLOCATED_PMDS; i++) {
 244                 pgd_t pgd = pgdp[i];
 245
 246                 if (pgd_val(pgd) != 0) {
 247                         pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
 248
 249                         pgdp[i] = native_make_pgd(0);
 250
 251                         paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
 252                         pmd_free(mm, pmd);
 253                         mm_dec_nr_pmds(mm);
 254                 }
 255         }
 256 }
 257
 258 static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
 259 {
 260         pud_t *pud;
 261         int i;
 262
 263         if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
 264                 return;
 265
 266         pud = pud_offset(pgd, 0);
 267
 268         for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
 269                 pmd_t *pmd = pmds[i];
 270
 271                 if (i >= KERNEL_PGD_BOUNDARY)
 272                         memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
 273                                sizeof(pmd_t) * PTRS_PER_PMD);
 274
 275                 pud_populate(mm, pud, pmd);
 276         }
 277 }
 278
 279 /*
 280  * Xen paravirt assumes pgd table should be in one page. 64 bit kernel also
 281  * assumes that pgd should be in one page.
 282  *
 283  * But kernel with PAE paging that is not running as a Xen domain
 284  * only needs to allocate 32 bytes for pgd instead of one page.
 285  */
 286 #ifdef CONFIG_X86_PAE
 287
 288 #include <linux/slab.h>
 289
 290 #define PGD_SIZE        (PTRS_PER_PGD * sizeof(pgd_t))
 291 #define PGD_ALIGN       32
 292
 293 static struct kmem_cache *pgd_cache;
 294
 295 static int __init pgd_cache_init(void)
 296 {
 297         /*
 298          * When PAE kernel is running as a Xen domain, it does not use
 299          * shared kernel pmd. And this requires a whole page for pgd.
 300          */
 301         if (!SHARED_KERNEL_PMD)
 302                 return 0;
 303
 304         /*
 305          * when PAE kernel is not running as a Xen domain, it uses
 306          * shared kernel pmd. Shared kernel pmd does not require a whole
 307          * page for pgd. We are able to just allocate a 32-byte for pgd.
 308          * During boot time, we create a 32-byte slab for pgd table allocation.
 309          */
 310         pgd_cache = kmem_cache_create("pgd_cache", PGD_SIZE, PGD_ALIGN,
 311                                       SLAB_PANIC, NULL);
 312         if (!pgd_cache)
 313                 return -ENOMEM;
 314
 315         return 0;
 316 }
 317 core_initcall(pgd_cache_init);
 318
 319 static inline pgd_t *_pgd_alloc(void)
 320 {
 321         /*
 322          * If no SHARED_KERNEL_PMD, PAE kernel is running as a Xen domain.
 323          * We allocate one page for pgd.
 324          */
 325         if (!SHARED_KERNEL_PMD)
 326                 return (pgd_t *)__get_free_page(PGALLOC_GFP);
 327
 328         /*
 329          * Now PAE kernel is not running as a Xen domain. We can allocate
 330          * a 32-byte slab for pgd to save memory space.
 331          */
 332         return kmem_cache_alloc(pgd_cache, PGALLOC_GFP);
 333 }
 334
 335 static inline void _pgd_free(pgd_t *pgd)
 336 {
 337         if (!SHARED_KERNEL_PMD)
 338                 free_page((unsigned long)pgd);
 339         else
 340                 kmem_cache_free(pgd_cache, pgd);
 341 }
 342 #else
 343 static inline pgd_t *_pgd_alloc(void)
 344 {
 345         return (pgd_t *)__get_free_page(PGALLOC_GFP);
 346 }
 347
 348 static inline void _pgd_free(pgd_t *pgd)
 349 {
 350         free_page((unsigned long)pgd);
 351 }
 352 #endif /* CONFIG_X86_PAE */
 353
 354 pgd_t *pgd_alloc(struct mm_struct *mm)
 355 {
 356         pgd_t *pgd;
 357         pmd_t *pmds[PREALLOCATED_PMDS];
 358
 359         pgd = _pgd_alloc();
 360
 361         if (pgd == NULL)
 362                 goto out;
 363
 364         mm->pgd = pgd;
 365
 366         if (preallocate_pmds(mm, pmds) != 0)
 367                 goto out_free_pgd;
 368
 369         if (paravirt_pgd_alloc(mm) != 0)
 370                 goto out_free_pmds;
 371
 372         /*
 373          * Make sure that pre-populating the pmds is atomic with
 374          * respect to anything walking the pgd_list, so that they
 375          * never see a partially populated pgd.
 376          */
 377         spin_lock(&pgd_lock);
 378
 379         pgd_ctor(mm, pgd);
 380         pgd_prepopulate_pmd(mm, pgd, pmds);
 381
 382         spin_unlock(&pgd_lock);
 383
 384         return pgd;
 385
 386 out_free_pmds:
 387         free_pmds(mm, pmds);
 388 out_free_pgd:
 389         _pgd_free(pgd);
 390 out:
 391         return NULL;
 392 }
 393
 394 void pgd_free(struct mm_struct *mm, pgd_t *pgd)
 395 {
 396         pgd_mop_up_pmds(mm, pgd);
 397         pgd_dtor(pgd);
 398         paravirt_pgd_free(mm, pgd);
 399         _pgd_free(pgd);
 400 }
 401
 402 /*
 403  * Used to set accessed or dirty bits in the page table entries
 404  * on other architectures. On x86, the accessed and dirty bits
 405  * are tracked by hardware. However, do_wp_page calls this function
 406  * to also make the pte writeable at the same time the dirty bit is
 407  * set. In that case we do actually need to write the PTE.
 408  */
 409 int ptep_set_access_flags(struct vm_area_struct *vma,
 410                           unsigned long address, pte_t *ptep,
 411                           pte_t entry, int dirty)
 412 {
 413         int changed = !pte_same(*ptep, entry);
 414
 415         if (changed && dirty) {
 416                 *ptep = entry;
 417                 pte_update(vma->vm_mm, address, ptep);
 418         }
 419
 420         return changed;
 421 }
 422
 423 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 424 int pmdp_set_access_flags(struct vm_area_struct *vma,
 425                           unsigned long address, pmd_t *pmdp,
 426                           pmd_t entry, int dirty)
 427 {
 428         int changed = !pmd_same(*pmdp, entry);
 429
 430         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 431
 432         if (changed && dirty) {
 433                 *pmdp = entry;
 434                 /*
 435                  * We had a write-protection fault here and changed the pmd
 436                  * to to more permissive. No need to flush the TLB for that,
 437                  * #PF is architecturally guaranteed to do that and in the
 438                  * worst-case we'll generate a spurious fault.
 439                  */
 440         }
 441
 442         return changed;
 443 }
 444 #endif
 445
 446 int ptep_test_and_clear_young(struct vm_area_struct *vma,
 447                               unsigned long addr, pte_t *ptep)
 448 {
 449         int ret = 0;
 450
 451         if (pte_young(*ptep))
 452                 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
 453                                          (unsigned long *) &ptep->pte);
 454
 455         if (ret)
 456                 pte_update(vma->vm_mm, addr, ptep);
 457
 458         return ret;
 459 }
 460
 461 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 462 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
 463                               unsigned long addr, pmd_t *pmdp)
 464 {
 465         int ret = 0;
 466
 467         if (pmd_young(*pmdp))
 468                 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
 469                                          (unsigned long *)pmdp);
 470
 471         return ret;
 472 }
 473 #endif
 474
 475 int ptep_clear_flush_young(struct vm_area_struct *vma,
 476                            unsigned long address, pte_t *ptep)
 477 {
 478         /*
 479          * On x86 CPUs, clearing the accessed bit without a TLB flush
 480          * doesn't cause data corruption. [ It could cause incorrect
 481          * page aging and the (mistaken) reclaim of hot pages, but the
 482          * chance of that should be relatively low. ]
 483          *
 484          * So as a performance optimization don't flush the TLB when
 485          * clearing the accessed bit, it will eventually be flushed by
 486          * a context switch or a VM operation anyway. [ In the rare
 487          * event of it not getting flushed for a long time the delay
 488          * shouldn't really matter because there's no real memory
 489          * pressure for swapout to react to. ]
 490          */
 491         return ptep_test_and_clear_young(vma, address, ptep);
 492 }
 493
 494 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 495 int pmdp_clear_flush_young(struct vm_area_struct *vma,
 496                            unsigned long address, pmd_t *pmdp)
 497 {
 498         int young;
 499
 500         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 501
 502         young = pmdp_test_and_clear_young(vma, address, pmdp);
 503         if (young)
 504                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 505
 506         return young;
 507 }
 508 #endif
 509
 510 /**
 511  * reserve_top_address - reserves a hole in the top of kernel address space
 512  * @reserve - size of hole to reserve
 513  *
 514  * Can be used to relocate the fixmap area and poke a hole in the top
 515  * of kernel address space to make room for a hypervisor.
 516  */
 517 void __init reserve_top_address(unsigned long reserve)
 518 {
 519 #ifdef CONFIG_X86_32
 520         BUG_ON(fixmaps_set > 0);
 521         __FIXADDR_TOP = round_down(-reserve, 1 << PMD_SHIFT) - PAGE_SIZE;
 522         printk(KERN_INFO "Reserving virtual address space above 0x%08lx (rounded to 0x%08lx)\n",
 523                -reserve, __FIXADDR_TOP + PAGE_SIZE);
 524 #endif
 525 }
 526
 527 int fixmaps_set;
 528
 529 void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
 530 {
 531         unsigned long address = __fix_to_virt(idx);
 532
 533         if (idx >= __end_of_fixed_addresses) {
 534                 BUG();
 535                 return;
 536         }
 537         set_pte_vaddr(address, pte);
 538         fixmaps_set++;
 539 }
 540
 541 void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
 542                        pgprot_t flags)
 543 {
 544         __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
 545 }
 546
 547 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
 548 /**
 549  * pud_set_huge - setup kernel PUD mapping
 550  *
 551  * MTRRs can override PAT memory types with 4KiB granularity. Therefore, this
 552  * function sets up a huge page only if any of the following conditions are met:
 553  *
 554  * - MTRRs are disabled, or
 555  *
 556  * - MTRRs are enabled and the range is completely covered by a single MTRR, or
 557  *
 558  * - MTRRs are enabled and the corresponding MTRR memory type is WB, which
 559  *   has no effect on the requested PAT memory type.
 560  *
 561  * Callers should try to decrease page size (1GB -> 2MB -> 4K) if the bigger
 562  * page mapping attempt fails.
 563  *
 564  * Returns 1 on success and 0 on failure.
 565  */
 566 int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
 567 {
 568         u8 mtrr, uniform;
 569
 570         mtrr = mtrr_type_lookup(addr, addr + PUD_SIZE, &uniform);
 571         if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
 572             (mtrr != MTRR_TYPE_WRBACK))
 573                 return 0;
 574
 575         prot = pgprot_4k_2_large(prot);
 576
 577         set_pte((pte_t *)pud, pfn_pte(
 578                 (u64)addr >> PAGE_SHIFT,
 579                 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 580
 581         return 1;
 582 }
 583
 584 /**
 585  * pmd_set_huge - setup kernel PMD mapping
 586  *
 587  * See text over pud_set_huge() above.
 588  *
 589  * Returns 1 on success and 0 on failure.
 590  */
 591 int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
 592 {
 593         u8 mtrr, uniform;
 594
 595         mtrr = mtrr_type_lookup(addr, addr + PMD_SIZE, &uniform);
 596         if ((mtrr != MTRR_TYPE_INVALID) && (!uniform) &&
 597             (mtrr != MTRR_TYPE_WRBACK)) {
 598                 pr_warn_once("%s: Cannot satisfy [mem %#010llx-%#010llx] with a huge-page mapping due to MTRR override.\n",
 599                              __func__, addr, addr + PMD_SIZE);
 600                 return 0;
 601         }
 602
 603         prot = pgprot_4k_2_large(prot);
 604
 605         set_pte((pte_t *)pmd, pfn_pte(
 606                 (u64)addr >> PAGE_SHIFT,
 607                 __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 608
 609         return 1;
 610 }
 611
 612 /**
 613  * pud_clear_huge - clear kernel PUD mapping when it is set
 614  *
 615  * Returns 1 on success and 0 on failure (no PUD map is found).
 616  */
 617 int pud_clear_huge(pud_t *pud)
 618 {
 619         if (pud_large(*pud)) {
 620                 pud_clear(pud);
 621                 return 1;
 622         }
 623
 624         return 0;
 625 }
 626
 627 /**
 628  * pmd_clear_huge - clear kernel PMD mapping when it is set
 629  *
 630  * Returns 1 on success and 0 on failure (no PMD map is found).
 631  */
 632 int pmd_clear_huge(pmd_t *pmd)
 633 {
 634         if (pmd_large(*pmd)) {
 635                 pmd_clear(pmd);
 636                 return 1;
 637         }
 638
 639         return 0;
 640 }
 641 #endif  /* CONFIG_HAVE_ARCH_HUGE_VMAP */