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
   8 #define PGALLOC_GFP GFP_KERNEL | __GFP_NOTRACK | __GFP_REPEAT | __GFP_ZERO
   9
  10 #ifdef CONFIG_HIGHPTE
  11 #define PGALLOC_USER_GFP __GFP_HIGHMEM
  12 #else
  13 #define PGALLOC_USER_GFP 0
  14 #endif
  15
  16 gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
  17
  18 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
  19 {
  20         return (pte_t *)__get_free_page(PGALLOC_GFP);
  21 }
  22
  23 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
  24 {
  25         struct page *pte;
  26
  27         pte = alloc_pages(__userpte_alloc_gfp, 0);
  28         if (pte)
  29                 pgtable_page_ctor(pte);
  30         return pte;
  31 }
  32
  33 static int __init setup_userpte(char *arg)
  34 {
  35         if (!arg)
  36                 return -EINVAL;
  37
  38         /*
  39          * "userpte=nohigh" disables allocation of user pagetables in
  40          * high memory.
  41          */
  42         if (strcmp(arg, "nohigh") == 0)
  43                 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
  44         else
  45                 return -EINVAL;
  46         return 0;
  47 }
  48 early_param("userpte", setup_userpte);
  49
  50 void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
  51 {
  52         pgtable_page_dtor(pte);
  53         paravirt_release_pte(page_to_pfn(pte));
  54         tlb_remove_page(tlb, pte);
  55 }
  56
  57 #if PAGETABLE_LEVELS > 2
  58 void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
  59 {
  60         paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
  61         tlb_remove_page(tlb, virt_to_page(pmd));
  62 }
  63
  64 #if PAGETABLE_LEVELS > 3
  65 void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
  66 {
  67         paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
  68         tlb_remove_page(tlb, virt_to_page(pud));
  69 }
  70 #endif  /* PAGETABLE_LEVELS > 3 */
  71 #endif  /* PAGETABLE_LEVELS > 2 */
  72
  73 static inline void pgd_list_add(pgd_t *pgd)
  74 {
  75         struct page *page = virt_to_page(pgd);
  76
  77         list_add(&page->lru, &pgd_list);
  78 }
  79
  80 static inline void pgd_list_del(pgd_t *pgd)
  81 {
  82         struct page *page = virt_to_page(pgd);
  83
  84         list_del(&page->lru);
  85 }
  86
  87 #define UNSHARED_PTRS_PER_PGD                           \
  88         (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
  89
  90
  91 static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
  92 {
  93         BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
  94         virt_to_page(pgd)->index = (pgoff_t)mm;
  95 }
  96
  97 struct mm_struct *pgd_page_get_mm(struct page *page)
  98 {
  99         return (struct mm_struct *)page->index;
 100 }
 101
 102 static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
 103 {
 104         /* If the pgd points to a shared pagetable level (either the
 105            ptes in non-PAE, or shared PMD in PAE), then just copy the
 106            references from swapper_pg_dir. */
 107         if (PAGETABLE_LEVELS == 2 ||
 108             (PAGETABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
 109             PAGETABLE_LEVELS == 4) {
 110                 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
 111                                 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
 112                                 KERNEL_PGD_PTRS);
 113         }
 114
 115         /* list required to sync kernel mapping updates */
 116         if (!SHARED_KERNEL_PMD) {
 117                 pgd_set_mm(pgd, mm);
 118                 pgd_list_add(pgd);
 119         }
 120 }
 121
 122 static void pgd_dtor(pgd_t *pgd)
 123 {
 124         if (SHARED_KERNEL_PMD)
 125                 return;
 126
 127         spin_lock(&pgd_lock);
 128         pgd_list_del(pgd);
 129         spin_unlock(&pgd_lock);
 130 }
 131
 132 /*
 133  * List of all pgd's needed for non-PAE so it can invalidate entries
 134  * in both cached and uncached pgd's; not needed for PAE since the
 135  * kernel pmd is shared. If PAE were not to share the pmd a similar
 136  * tactic would be needed. This is essentially codepath-based locking
 137  * against pageattr.c; it is the unique case in which a valid change
 138  * of kernel pagetables can't be lazily synchronized by vmalloc faults.
 139  * vmalloc faults work because attached pagetables are never freed.
 140  * -- nyc
 141  */
 142
 143 #ifdef CONFIG_X86_PAE
 144 /*
 145  * In PAE mode, we need to do a cr3 reload (=tlb flush) when
 146  * updating the top-level pagetable entries to guarantee the
 147  * processor notices the update.  Since this is expensive, and
 148  * all 4 top-level entries are used almost immediately in a
 149  * new process's life, we just pre-populate them here.
 150  *
 151  * Also, if we're in a paravirt environment where the kernel pmd is
 152  * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
 153  * and initialize the kernel pmds here.
 154  */
 155 #define PREALLOCATED_PMDS       UNSHARED_PTRS_PER_PGD
 156
 157 void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
 158 {
 159         paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
 160
 161         /* Note: almost everything apart from _PAGE_PRESENT is
 162            reserved at the pmd (PDPT) level. */
 163         set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
 164
 165         /*
 166          * According to Intel App note "TLBs, Paging-Structure Caches,
 167          * and Their Invalidation", April 2007, document 317080-001,
 168          * section 8.1: in PAE mode we explicitly have to flush the
 169          * TLB via cr3 if the top-level pgd is changed...
 170          */
 171         flush_tlb_mm(mm);
 172 }
 173 #else  /* !CONFIG_X86_PAE */
 174
 175 /* No need to prepopulate any pagetable entries in non-PAE modes. */
 176 #define PREALLOCATED_PMDS       0
 177
 178 #endif  /* CONFIG_X86_PAE */
 179
 180 static void free_pmds(pmd_t *pmds[])
 181 {
 182         int i;
 183
 184         for(i = 0; i < PREALLOCATED_PMDS; i++)
 185                 if (pmds[i])
 186                         free_page((unsigned long)pmds[i]);
 187 }
 188
 189 static int preallocate_pmds(pmd_t *pmds[])
 190 {
 191         int i;
 192         bool failed = false;
 193
 194         for(i = 0; i < PREALLOCATED_PMDS; i++) {
 195                 pmd_t *pmd = (pmd_t *)__get_free_page(PGALLOC_GFP);
 196                 if (pmd == NULL)
 197                         failed = true;
 198                 pmds[i] = pmd;
 199         }
 200
 201         if (failed) {
 202                 free_pmds(pmds);
 203                 return -ENOMEM;
 204         }
 205
 206         return 0;
 207 }
 208
 209 /*
 210  * Mop up any pmd pages which may still be attached to the pgd.
 211  * Normally they will be freed by munmap/exit_mmap, but any pmd we
 212  * preallocate which never got a corresponding vma will need to be
 213  * freed manually.
 214  */
 215 static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
 216 {
 217         int i;
 218
 219         for(i = 0; i < PREALLOCATED_PMDS; i++) {
 220                 pgd_t pgd = pgdp[i];
 221
 222                 if (pgd_val(pgd) != 0) {
 223                         pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
 224
 225                         pgdp[i] = native_make_pgd(0);
 226
 227                         paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
 228                         pmd_free(mm, pmd);
 229                 }
 230         }
 231 }
 232
 233 static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
 234 {
 235         pud_t *pud;
 236         unsigned long addr;
 237         int i;
 238
 239         if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
 240                 return;
 241
 242         pud = pud_offset(pgd, 0);
 243
 244         for (addr = i = 0; i < PREALLOCATED_PMDS;
 245              i++, pud++, addr += PUD_SIZE) {
 246                 pmd_t *pmd = pmds[i];
 247
 248                 if (i >= KERNEL_PGD_BOUNDARY)
 249                         memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
 250                                sizeof(pmd_t) * PTRS_PER_PMD);
 251
 252                 pud_populate(mm, pud, pmd);
 253         }
 254 }
 255
 256 pgd_t *pgd_alloc(struct mm_struct *mm)
 257 {
 258         pgd_t *pgd;
 259         pmd_t *pmds[PREALLOCATED_PMDS];
 260
 261         pgd = (pgd_t *)__get_free_page(PGALLOC_GFP);
 262
 263         if (pgd == NULL)
 264                 goto out;
 265
 266         mm->pgd = pgd;
 267
 268         if (preallocate_pmds(pmds) != 0)
 269                 goto out_free_pgd;
 270
 271         if (paravirt_pgd_alloc(mm) != 0)
 272                 goto out_free_pmds;
 273
 274         /*
 275          * Make sure that pre-populating the pmds is atomic with
 276          * respect to anything walking the pgd_list, so that they
 277          * never see a partially populated pgd.
 278          */
 279         spin_lock(&pgd_lock);
 280
 281         pgd_ctor(mm, pgd);
 282         pgd_prepopulate_pmd(mm, pgd, pmds);
 283
 284         spin_unlock(&pgd_lock);
 285
 286         return pgd;
 287
 288 out_free_pmds:
 289         free_pmds(pmds);
 290 out_free_pgd:
 291         free_page((unsigned long)pgd);
 292 out:
 293         return NULL;
 294 }
 295
 296 void pgd_free(struct mm_struct *mm, pgd_t *pgd)
 297 {
 298         pgd_mop_up_pmds(mm, pgd);
 299         pgd_dtor(pgd);
 300         paravirt_pgd_free(mm, pgd);
 301         free_page((unsigned long)pgd);
 302 }
 303
 304 /*
 305  * Used to set accessed or dirty bits in the page table entries
 306  * on other architectures. On x86, the accessed and dirty bits
 307  * are tracked by hardware. However, do_wp_page calls this function
 308  * to also make the pte writeable at the same time the dirty bit is
 309  * set. In that case we do actually need to write the PTE.
 310  */
 311 int ptep_set_access_flags(struct vm_area_struct *vma,
 312                           unsigned long address, pte_t *ptep,
 313                           pte_t entry, int dirty)
 314 {
 315         int changed = !pte_same(*ptep, entry);
 316
 317         if (changed && dirty) {
 318                 *ptep = entry;
 319                 pte_update_defer(vma->vm_mm, address, ptep);
 320         }
 321
 322         return changed;
 323 }
 324
 325 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 326 int pmdp_set_access_flags(struct vm_area_struct *vma,
 327                           unsigned long address, pmd_t *pmdp,
 328                           pmd_t entry, int dirty)
 329 {
 330         int changed = !pmd_same(*pmdp, entry);
 331
 332         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 333
 334         if (changed && dirty) {
 335                 *pmdp = entry;
 336                 pmd_update_defer(vma->vm_mm, address, pmdp);
 337                 /*
 338                  * We had a write-protection fault here and changed the pmd
 339                  * to to more permissive. No need to flush the TLB for that,
 340                  * #PF is architecturally guaranteed to do that and in the
 341                  * worst-case we'll generate a spurious fault.
 342                  */
 343         }
 344
 345         return changed;
 346 }
 347 #endif
 348
 349 int ptep_test_and_clear_young(struct vm_area_struct *vma,
 350                               unsigned long addr, pte_t *ptep)
 351 {
 352         int ret = 0;
 353
 354         if (pte_young(*ptep))
 355                 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
 356                                          (unsigned long *) &ptep->pte);
 357
 358         if (ret)
 359                 pte_update(vma->vm_mm, addr, ptep);
 360
 361         return ret;
 362 }
 363
 364 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 365 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
 366                               unsigned long addr, pmd_t *pmdp)
 367 {
 368         int ret = 0;
 369
 370         if (pmd_young(*pmdp))
 371                 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
 372                                          (unsigned long *)pmdp);
 373
 374         if (ret)
 375                 pmd_update(vma->vm_mm, addr, pmdp);
 376
 377         return ret;
 378 }
 379 #endif
 380
 381 int ptep_clear_flush_young(struct vm_area_struct *vma,
 382                            unsigned long address, pte_t *ptep)
 383 {
 384         int young;
 385
 386         young = ptep_test_and_clear_young(vma, address, ptep);
 387         if (young)
 388                 flush_tlb_page(vma, address);
 389
 390         return young;
 391 }
 392
 393 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 394 int pmdp_clear_flush_young(struct vm_area_struct *vma,
 395                            unsigned long address, pmd_t *pmdp)
 396 {
 397         int young;
 398
 399         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 400
 401         young = pmdp_test_and_clear_young(vma, address, pmdp);
 402         if (young)
 403                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 404
 405         return young;
 406 }
 407
 408 void pmdp_splitting_flush(struct vm_area_struct *vma,
 409                           unsigned long address, pmd_t *pmdp)
 410 {
 411         int set;
 412         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 413         set = !test_and_set_bit(_PAGE_BIT_SPLITTING,
 414                                 (unsigned long *)pmdp);
 415         if (set) {
 416                 pmd_update(vma->vm_mm, address, pmdp);
 417                 /* need tlb flush only to serialize against gup-fast */
 418                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 419         }
 420 }
 421 #endif
 422
 423 /**
 424  * reserve_top_address - reserves a hole in the top of kernel address space
 425  * @reserve - size of hole to reserve
 426  *
 427  * Can be used to relocate the fixmap area and poke a hole in the top
 428  * of kernel address space to make room for a hypervisor.
 429  */
 430 void __init reserve_top_address(unsigned long reserve)
 431 {
 432 #ifdef CONFIG_X86_32
 433         BUG_ON(fixmaps_set > 0);
 434         printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
 435                (int)-reserve);
 436         __FIXADDR_TOP = -reserve - PAGE_SIZE;
 437 #endif
 438 }
 439
 440 int fixmaps_set;
 441
 442 void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
 443 {
 444         unsigned long address = __fix_to_virt(idx);
 445
 446         if (idx >= __end_of_fixed_addresses) {
 447                 BUG();
 448                 return;
 449         }
 450         set_pte_vaddr(address, pte);
 451         fixmaps_set++;
 452 }
 453
 454 void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
 455                        pgprot_t flags)
 456 {
 457         __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
 458 }