arch/x86/mm/hugetlbpage.c

   1 /*
   2  * IA-32 Huge TLB Page Support for Kernel.
   3  *
   4  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
   5  */
   6
   7 #include <linux/init.h>
   8 #include <linux/fs.h>
   9 #include <linux/mm.h>
  10 #include <linux/hugetlb.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/err.h>
  14 #include <linux/sysctl.h>
  15 #include <asm/mman.h>
  16 #include <asm/tlb.h>
  17 #include <asm/tlbflush.h>
  18 #include <asm/pgalloc.h>
  19
  20 static unsigned long page_table_shareable(struct vm_area_struct *svma,
  21                                 struct vm_area_struct *vma,
  22                                 unsigned long addr, pgoff_t idx)
  23 {
  24         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
  25                                 svma->vm_start;
  26         unsigned long sbase = saddr & PUD_MASK;
  27         unsigned long s_end = sbase + PUD_SIZE;
  28
  29         /*
  30          * match the virtual addresses, permission and the alignment of the
  31          * page table page.
  32          */
  33         if (pmd_index(addr) != pmd_index(saddr) ||
  34             vma->vm_flags != svma->vm_flags ||
  35             sbase < svma->vm_start || svma->vm_end < s_end)
  36                 return 0;
  37
  38         return saddr;
  39 }
  40
  41 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
  42 {
  43         unsigned long base = addr & PUD_MASK;
  44         unsigned long end = base + PUD_SIZE;
  45
  46         /*
  47          * check on proper vm_flags and page table alignment
  48          */
  49         if (vma->vm_flags & VM_MAYSHARE &&
  50             vma->vm_start <= base && end <= vma->vm_end)
  51                 return 1;
  52         return 0;
  53 }
  54
  55 /*
  56  * search for a shareable pmd page for hugetlb.
  57  */
  58 static void huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
  59 {
  60         struct vm_area_struct *vma = find_vma(mm, addr);
  61         struct address_space *mapping = vma->vm_file->f_mapping;
  62         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
  63                         vma->vm_pgoff;
  64         struct prio_tree_iter iter;
  65         struct vm_area_struct *svma;
  66         unsigned long saddr;
  67         pte_t *spte = NULL;
  68
  69         if (!vma_shareable(vma, addr))
  70                 return;
  71
  72         spin_lock(&mapping->i_mmap_lock);
  73         vma_prio_tree_foreach(svma, &iter, &mapping->i_mmap, idx, idx) {
  74                 if (svma == vma)
  75                         continue;
  76
  77                 saddr = page_table_shareable(svma, vma, addr, idx);
  78                 if (saddr) {
  79                         spte = huge_pte_offset(svma->vm_mm, saddr);
  80                         if (spte) {
  81                                 get_page(virt_to_page(spte));
  82                                 break;
  83                         }
  84                 }
  85         }
  86
  87         if (!spte)
  88                 goto out;
  89
  90         spin_lock(&mm->page_table_lock);
  91         if (pud_none(*pud))
  92                 pud_populate(mm, pud, (pmd_t *)((unsigned long)spte & PAGE_MASK));
  93         else
  94                 put_page(virt_to_page(spte));
  95         spin_unlock(&mm->page_table_lock);
  96 out:
  97         spin_unlock(&mapping->i_mmap_lock);
  98 }
  99
 100 /*
 101  * unmap huge page backed by shared pte.
 102  *
 103  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
 104  * indicated by page_count > 1, unmap is achieved by clearing pud and
 105  * decrementing the ref count. If count == 1, the pte page is not shared.
 106  *
 107  * called with vma->vm_mm->page_table_lock held.
 108  *
 109  * returns: 1 successfully unmapped a shared pte page
 110  *          0 the underlying pte page is not shared, or it is the last user
 111  */
 112 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
 113 {
 114         pgd_t *pgd = pgd_offset(mm, *addr);
 115         pud_t *pud = pud_offset(pgd, *addr);
 116
 117         BUG_ON(page_count(virt_to_page(ptep)) == 0);
 118         if (page_count(virt_to_page(ptep)) == 1)
 119                 return 0;
 120
 121         pud_clear(pud);
 122         put_page(virt_to_page(ptep));
 123         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
 124         return 1;
 125 }
 126
 127 pte_t *huge_pte_alloc(struct mm_struct *mm,
 128                         unsigned long addr, unsigned long sz)
 129 {
 130         pgd_t *pgd;
 131         pud_t *pud;
 132         pte_t *pte = NULL;
 133
 134         pgd = pgd_offset(mm, addr);
 135         pud = pud_alloc(mm, pgd, addr);
 136         if (pud) {
 137                 if (pud_none(*pud))
 138                         huge_pmd_share(mm, addr, pud);
 139                 pte = (pte_t *) pmd_alloc(mm, pud, addr);
 140         }
 141         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
 142
 143         return pte;
 144 }
 145
 146 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 147 {
 148         pgd_t *pgd;
 149         pud_t *pud;
 150         pmd_t *pmd = NULL;
 151
 152         pgd = pgd_offset(mm, addr);
 153         if (pgd_present(*pgd)) {
 154                 pud = pud_offset(pgd, addr);
 155                 if (pud_present(*pud))
 156                         pmd = pmd_offset(pud, addr);
 157         }
 158         return (pte_t *) pmd;
 159 }
 160
 161 #if 0   /* This is just for testing */
 162 struct page *
 163 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 164 {
 165         unsigned long start = address;
 166         int length = 1;
 167         int nr;
 168         struct page *page;
 169         struct vm_area_struct *vma;
 170
 171         vma = find_vma(mm, addr);
 172         if (!vma || !is_vm_hugetlb_page(vma))
 173                 return ERR_PTR(-EINVAL);
 174
 175         pte = huge_pte_offset(mm, address);
 176
 177         /* hugetlb should be locked, and hence, prefaulted */
 178         WARN_ON(!pte || pte_none(*pte));
 179
 180         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 181
 182         WARN_ON(!PageHead(page));
 183
 184         return page;
 185 }
 186
 187 int pmd_huge(pmd_t pmd)
 188 {
 189         return 0;
 190 }
 191
 192 int pud_huge(pud_t pud)
 193 {
 194         return 0;
 195 }
 196
 197 struct page *
 198 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 199                 pmd_t *pmd, int write)
 200 {
 201         return NULL;
 202 }
 203
 204 #else
 205
 206 struct page *
 207 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 208 {
 209         return ERR_PTR(-EINVAL);
 210 }
 211
 212 int pmd_huge(pmd_t pmd)
 213 {
 214         return !!(pmd_val(pmd) & _PAGE_PSE);
 215 }
 216
 217 int pud_huge(pud_t pud)
 218 {
 219         return 0;
 220 }
 221
 222 struct page *
 223 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 224                 pmd_t *pmd, int write)
 225 {
 226         struct page *page;
 227
 228         page = pte_page(*(pte_t *)pmd);
 229         if (page)
 230                 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
 231         return page;
 232 }
 233
 234 struct page *
 235 follow_huge_pud(struct mm_struct *mm, unsigned long address,
 236                 pud_t *pud, int write)
 237 {
 238         struct page *page;
 239
 240         page = pte_page(*(pte_t *)pud);
 241         if (page)
 242                 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
 243         return page;
 244 }
 245
 246 #endif
 247
 248 /* x86_64 also uses this file */
 249
 250 #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 251 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
 252                 unsigned long addr, unsigned long len,
 253                 unsigned long pgoff, unsigned long flags)
 254 {
 255         struct mm_struct *mm = current->mm;
 256         struct vm_area_struct *vma;
 257         unsigned long start_addr;
 258
 259         if (len > mm->cached_hole_size) {
 260                 start_addr = mm->free_area_cache;
 261         } else {
 262                 start_addr = TASK_UNMAPPED_BASE;
 263                 mm->cached_hole_size = 0;
 264         }
 265
 266 full_search:
 267         addr = ALIGN(start_addr, HPAGE_SIZE);
 268
 269         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
 270                 /* At this point:  (!vma || addr < vma->vm_end). */
 271                 if (TASK_SIZE - len < addr) {
 272                         /*
 273                          * Start a new search - just in case we missed
 274                          * some holes.
 275                          */
 276                         if (start_addr != TASK_UNMAPPED_BASE) {
 277                                 start_addr = TASK_UNMAPPED_BASE;
 278                                 mm->cached_hole_size = 0;
 279                                 goto full_search;
 280                         }
 281                         return -ENOMEM;
 282                 }
 283                 if (!vma || addr + len <= vma->vm_start) {
 284                         mm->free_area_cache = addr + len;
 285                         return addr;
 286                 }
 287                 if (addr + mm->cached_hole_size < vma->vm_start)
 288                         mm->cached_hole_size = vma->vm_start - addr;
 289                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 290         }
 291 }
 292
 293 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
 294                 unsigned long addr0, unsigned long len,
 295                 unsigned long pgoff, unsigned long flags)
 296 {
 297         struct mm_struct *mm = current->mm;
 298         struct vm_area_struct *vma, *prev_vma;
 299         unsigned long base = mm->mmap_base, addr = addr0;
 300         unsigned long largest_hole = mm->cached_hole_size;
 301         int first_time = 1;
 302
 303         /* don't allow allocations above current base */
 304         if (mm->free_area_cache > base)
 305                 mm->free_area_cache = base;
 306
 307         if (len <= largest_hole) {
 308                 largest_hole = 0;
 309                 mm->free_area_cache  = base;
 310         }
 311 try_again:
 312         /* make sure it can fit in the remaining address space */
 313         if (mm->free_area_cache < len)
 314                 goto fail;
 315
 316         /* either no address requested or cant fit in requested address hole */
 317         addr = (mm->free_area_cache - len) & HPAGE_MASK;
 318         do {
 319                 /*
 320                  * Lookup failure means no vma is above this address,
 321                  * i.e. return with success:
 322                  */
 323                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
 324                         return addr;
 325
 326                 /*
 327                  * new region fits between prev_vma->vm_end and
 328                  * vma->vm_start, use it:
 329                  */
 330                 if (addr + len <= vma->vm_start &&
 331                             (!prev_vma || (addr >= prev_vma->vm_end))) {
 332                         /* remember the address as a hint for next time */
 333                         mm->cached_hole_size = largest_hole;
 334                         return (mm->free_area_cache = addr);
 335                 } else {
 336                         /* pull free_area_cache down to the first hole */
 337                         if (mm->free_area_cache == vma->vm_end) {
 338                                 mm->free_area_cache = vma->vm_start;
 339                                 mm->cached_hole_size = largest_hole;
 340                         }
 341                 }
 342
 343                 /* remember the largest hole we saw so far */
 344                 if (addr + largest_hole < vma->vm_start)
 345                         largest_hole = vma->vm_start - addr;
 346
 347                 /* try just below the current vma->vm_start */
 348                 addr = (vma->vm_start - len) & HPAGE_MASK;
 349         } while (len <= vma->vm_start);
 350
 351 fail:
 352         /*
 353          * if hint left us with no space for the requested
 354          * mapping then try again:
 355          */
 356         if (first_time) {
 357                 mm->free_area_cache = base;
 358                 largest_hole = 0;
 359                 first_time = 0;
 360                 goto try_again;
 361         }
 362         /*
 363          * A failed mmap() very likely causes application failure,
 364          * so fall back to the bottom-up function here. This scenario
 365          * can happen with large stack limits and large mmap()
 366          * allocations.
 367          */
 368         mm->free_area_cache = TASK_UNMAPPED_BASE;
 369         mm->cached_hole_size = ~0UL;
 370         addr = hugetlb_get_unmapped_area_bottomup(file, addr0,
 371                         len, pgoff, flags);
 372
 373         /*
 374          * Restore the topdown base:
 375          */
 376         mm->free_area_cache = base;
 377         mm->cached_hole_size = ~0UL;
 378
 379         return addr;
 380 }
 381
 382 unsigned long
 383 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 384                 unsigned long len, unsigned long pgoff, unsigned long flags)
 385 {
 386         struct mm_struct *mm = current->mm;
 387         struct vm_area_struct *vma;
 388
 389         if (len & ~HPAGE_MASK)
 390                 return -EINVAL;
 391         if (len > TASK_SIZE)
 392                 return -ENOMEM;
 393
 394         if (flags & MAP_FIXED) {
 395                 if (prepare_hugepage_range(file, addr, len))
 396                         return -EINVAL;
 397                 return addr;
 398         }
 399
 400         if (addr) {
 401                 addr = ALIGN(addr, HPAGE_SIZE);
 402                 vma = find_vma(mm, addr);
 403                 if (TASK_SIZE - len >= addr &&
 404                     (!vma || addr + len <= vma->vm_start))
 405                         return addr;
 406         }
 407         if (mm->get_unmapped_area == arch_get_unmapped_area)
 408                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 409                                 pgoff, flags);
 410         else
 411                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 412                                 pgoff, flags);
 413 }
 414
 415 #endif /*HAVE_ARCH_HUGETLB_UNMAPPED_AREA*/
 416