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         /*
  62          * NOTE! For PAE, any changes to the top page-directory-pointer-table
  63          * entries need a full cr3 reload to flush.
  64          */
  65 #ifdef CONFIG_X86_PAE
  66         tlb->need_flush_all = 1;
  67 #endif
  68         tlb_remove_page(tlb, virt_to_page(pmd));
  69 }
  70
  71 #if PAGETABLE_LEVELS > 3
  72 void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
  73 {
  74         paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
  75         tlb_remove_page(tlb, virt_to_page(pud));
  76 }
  77 #endif  /* PAGETABLE_LEVELS > 3 */
  78 #endif  /* PAGETABLE_LEVELS > 2 */
  79
  80 static inline void pgd_list_add(pgd_t *pgd)
  81 {
  82         struct page *page = virt_to_page(pgd);
  83
  84         list_add(&page->lru, &pgd_list);
  85 }
  86
  87 static inline void pgd_list_del(pgd_t *pgd)
  88 {
  89         struct page *page = virt_to_page(pgd);
  90
  91         list_del(&page->lru);
  92 }
  93
  94 #define UNSHARED_PTRS_PER_PGD                           \
  95         (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
  96
  97
  98 static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
  99 {
 100         BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
 101         virt_to_page(pgd)->index = (pgoff_t)mm;
 102 }
 103
 104 struct mm_struct *pgd_page_get_mm(struct page *page)
 105 {
 106         return (struct mm_struct *)page->index;
 107 }
 108
 109 static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
 110 {
 111         /* If the pgd points to a shared pagetable level (either the
 112            ptes in non-PAE, or shared PMD in PAE), then just copy the
 113            references from swapper_pg_dir. */
 114         if (PAGETABLE_LEVELS == 2 ||
 115             (PAGETABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
 116             PAGETABLE_LEVELS == 4) {
 117                 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
 118                                 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
 119                                 KERNEL_PGD_PTRS);
 120         }
 121
 122         /* list required to sync kernel mapping updates */
 123         if (!SHARED_KERNEL_PMD) {
 124                 pgd_set_mm(pgd, mm);
 125                 pgd_list_add(pgd);
 126         }
 127 }
 128
 129 static void pgd_dtor(pgd_t *pgd)
 130 {
 131         if (SHARED_KERNEL_PMD)
 132                 return;
 133
 134         spin_lock(&pgd_lock);
 135         pgd_list_del(pgd);
 136         spin_unlock(&pgd_lock);
 137 }
 138
 139 /*
 140  * List of all pgd's needed for non-PAE so it can invalidate entries
 141  * in both cached and uncached pgd's; not needed for PAE since the
 142  * kernel pmd is shared. If PAE were not to share the pmd a similar
 143  * tactic would be needed. This is essentially codepath-based locking
 144  * against pageattr.c; it is the unique case in which a valid change
 145  * of kernel pagetables can't be lazily synchronized by vmalloc faults.
 146  * vmalloc faults work because attached pagetables are never freed.
 147  * -- nyc
 148  */
 149
 150 #ifdef CONFIG_X86_PAE
 151 /*
 152  * In PAE mode, we need to do a cr3 reload (=tlb flush) when
 153  * updating the top-level pagetable entries to guarantee the
 154  * processor notices the update.  Since this is expensive, and
 155  * all 4 top-level entries are used almost immediately in a
 156  * new process's life, we just pre-populate them here.
 157  *
 158  * Also, if we're in a paravirt environment where the kernel pmd is
 159  * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
 160  * and initialize the kernel pmds here.
 161  */
 162 #define PREALLOCATED_PMDS       UNSHARED_PTRS_PER_PGD
 163
 164 void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
 165 {
 166         paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
 167
 168         /* Note: almost everything apart from _PAGE_PRESENT is
 169            reserved at the pmd (PDPT) level. */
 170         set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
 171
 172         /*
 173          * According to Intel App note "TLBs, Paging-Structure Caches,
 174          * and Their Invalidation", April 2007, document 317080-001,
 175          * section 8.1: in PAE mode we explicitly have to flush the
 176          * TLB via cr3 if the top-level pgd is changed...
 177          */
 178         flush_tlb_mm(mm);
 179 }
 180 #else  /* !CONFIG_X86_PAE */
 181
 182 /* No need to prepopulate any pagetable entries in non-PAE modes. */
 183 #define PREALLOCATED_PMDS       0
 184
 185 #endif  /* CONFIG_X86_PAE */
 186
 187 static void free_pmds(pmd_t *pmds[])
 188 {
 189         int i;
 190
 191         for(i = 0; i < PREALLOCATED_PMDS; i++)
 192                 if (pmds[i]) {
 193                         pgtable_pmd_page_dtor(virt_to_page(pmds[i]));
 194                         free_page((unsigned long)pmds[i]);
 195                 }
 196 }
 197
 198 static int preallocate_pmds(pmd_t *pmds[])
 199 {
 200         int i;
 201         bool failed = false;
 202
 203         for(i = 0; i < PREALLOCATED_PMDS; i++) {
 204                 pmd_t *pmd = (pmd_t *)__get_free_page(PGALLOC_GFP);
 205                 if (!pmd)
 206                         failed = true;
 207                 if (pmd && !pgtable_pmd_page_ctor(virt_to_page(pmd))) {
 208                         free_page((unsigned long)pmds[i]);
 209                         pmd = NULL;
 210                         failed = true;
 211                 }
 212                 pmds[i] = pmd;
 213         }
 214
 215         if (failed) {
 216                 free_pmds(pmds);
 217                 return -ENOMEM;
 218         }
 219
 220         return 0;
 221 }
 222
 223 /*
 224  * Mop up any pmd pages which may still be attached to the pgd.
 225  * Normally they will be freed by munmap/exit_mmap, but any pmd we
 226  * preallocate which never got a corresponding vma will need to be
 227  * freed manually.
 228  */
 229 static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
 230 {
 231         int i;
 232
 233         for(i = 0; i < PREALLOCATED_PMDS; i++) {
 234                 pgd_t pgd = pgdp[i];
 235
 236                 if (pgd_val(pgd) != 0) {
 237                         pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
 238
 239                         pgdp[i] = native_make_pgd(0);
 240
 241                         paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
 242                         pmd_free(mm, pmd);
 243                 }
 244         }
 245 }
 246
 247 static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
 248 {
 249         pud_t *pud;
 250         int i;
 251
 252         if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
 253                 return;
 254
 255         pud = pud_offset(pgd, 0);
 256
 257         for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
 258                 pmd_t *pmd = pmds[i];
 259
 260                 if (i >= KERNEL_PGD_BOUNDARY)
 261                         memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
 262                                sizeof(pmd_t) * PTRS_PER_PMD);
 263
 264                 pud_populate(mm, pud, pmd);
 265         }
 266 }
 267
 268 pgd_t *pgd_alloc(struct mm_struct *mm)
 269 {
 270         pgd_t *pgd;
 271         pmd_t *pmds[PREALLOCATED_PMDS];
 272
 273         pgd = (pgd_t *)__get_free_page(PGALLOC_GFP);
 274
 275         if (pgd == NULL)
 276                 goto out;
 277
 278         mm->pgd = pgd;
 279
 280         if (preallocate_pmds(pmds) != 0)
 281                 goto out_free_pgd;
 282
 283         if (paravirt_pgd_alloc(mm) != 0)
 284                 goto out_free_pmds;
 285
 286         /*
 287          * Make sure that pre-populating the pmds is atomic with
 288          * respect to anything walking the pgd_list, so that they
 289          * never see a partially populated pgd.
 290          */
 291         spin_lock(&pgd_lock);
 292
 293         pgd_ctor(mm, pgd);
 294         pgd_prepopulate_pmd(mm, pgd, pmds);
 295
 296         spin_unlock(&pgd_lock);
 297
 298         return pgd;
 299
 300 out_free_pmds:
 301         free_pmds(pmds);
 302 out_free_pgd:
 303         free_page((unsigned long)pgd);
 304 out:
 305         return NULL;
 306 }
 307
 308 void pgd_free(struct mm_struct *mm, pgd_t *pgd)
 309 {
 310         pgd_mop_up_pmds(mm, pgd);
 311         pgd_dtor(pgd);
 312         paravirt_pgd_free(mm, pgd);
 313         free_page((unsigned long)pgd);
 314 }
 315
 316 /*
 317  * Used to set accessed or dirty bits in the page table entries
 318  * on other architectures. On x86, the accessed and dirty bits
 319  * are tracked by hardware. However, do_wp_page calls this function
 320  * to also make the pte writeable at the same time the dirty bit is
 321  * set. In that case we do actually need to write the PTE.
 322  */
 323 int ptep_set_access_flags(struct vm_area_struct *vma,
 324                           unsigned long address, pte_t *ptep,
 325                           pte_t entry, int dirty)
 326 {
 327         int changed = !pte_same(*ptep, entry);
 328
 329         if (changed && dirty) {
 330                 *ptep = entry;
 331                 pte_update_defer(vma->vm_mm, address, ptep);
 332         }
 333
 334         return changed;
 335 }
 336
 337 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 338 int pmdp_set_access_flags(struct vm_area_struct *vma,
 339                           unsigned long address, pmd_t *pmdp,
 340                           pmd_t entry, int dirty)
 341 {
 342         int changed = !pmd_same(*pmdp, entry);
 343
 344         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 345
 346         if (changed && dirty) {
 347                 *pmdp = entry;
 348                 pmd_update_defer(vma->vm_mm, address, pmdp);
 349                 /*
 350                  * We had a write-protection fault here and changed the pmd
 351                  * to to more permissive. No need to flush the TLB for that,
 352                  * #PF is architecturally guaranteed to do that and in the
 353                  * worst-case we'll generate a spurious fault.
 354                  */
 355         }
 356
 357         return changed;
 358 }
 359 #endif
 360
 361 int ptep_test_and_clear_young(struct vm_area_struct *vma,
 362                               unsigned long addr, pte_t *ptep)
 363 {
 364         int ret = 0;
 365
 366         if (pte_young(*ptep))
 367                 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
 368                                          (unsigned long *) &ptep->pte);
 369
 370         if (ret)
 371                 pte_update(vma->vm_mm, addr, ptep);
 372
 373         return ret;
 374 }
 375
 376 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 377 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
 378                               unsigned long addr, pmd_t *pmdp)
 379 {
 380         int ret = 0;
 381
 382         if (pmd_young(*pmdp))
 383                 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
 384                                          (unsigned long *)pmdp);
 385
 386         if (ret)
 387                 pmd_update(vma->vm_mm, addr, pmdp);
 388
 389         return ret;
 390 }
 391 #endif
 392
 393 int ptep_clear_flush_young(struct vm_area_struct *vma,
 394                            unsigned long address, pte_t *ptep)
 395 {
 396         int young;
 397
 398         young = ptep_test_and_clear_young(vma, address, ptep);
 399         if (young)
 400                 flush_tlb_page(vma, address);
 401
 402         return young;
 403 }
 404
 405 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 406 int pmdp_clear_flush_young(struct vm_area_struct *vma,
 407                            unsigned long address, pmd_t *pmdp)
 408 {
 409         int young;
 410
 411         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 412
 413         young = pmdp_test_and_clear_young(vma, address, pmdp);
 414         if (young)
 415                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 416
 417         return young;
 418 }
 419
 420 void pmdp_splitting_flush(struct vm_area_struct *vma,
 421                           unsigned long address, pmd_t *pmdp)
 422 {
 423         int set;
 424         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 425         set = !test_and_set_bit(_PAGE_BIT_SPLITTING,
 426                                 (unsigned long *)pmdp);
 427         if (set) {
 428                 pmd_update(vma->vm_mm, address, pmdp);
 429                 /* need tlb flush only to serialize against gup-fast */
 430                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 431         }
 432 }
 433 #endif
 434
 435 /**
 436  * reserve_top_address - reserves a hole in the top of kernel address space
 437  * @reserve - size of hole to reserve
 438  *
 439  * Can be used to relocate the fixmap area and poke a hole in the top
 440  * of kernel address space to make room for a hypervisor.
 441  */
 442 void __init reserve_top_address(unsigned long reserve)
 443 {
 444 #ifdef CONFIG_X86_32
 445         BUG_ON(fixmaps_set > 0);
 446         printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
 447                (int)-reserve);
 448         __FIXADDR_TOP = -reserve - PAGE_SIZE;
 449 #endif
 450 }
 451
 452 int fixmaps_set;
 453
 454 void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
 455 {
 456         unsigned long address = __fix_to_virt(idx);
 457
 458         if (idx >= __end_of_fixed_addresses) {
 459                 BUG();
 460                 return;
 461         }
 462         set_pte_vaddr(address, pte);
 463         fixmaps_set++;
 464 }
 465
 466 void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
 467                        pgprot_t flags)
 468 {
 469         __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
 470 }