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/err.h>
  13 #include <linux/sysctl.h>
  14 #include <asm/mman.h>
  15 #include <asm/tlb.h>
  16 #include <asm/tlbflush.h>
  17 #include <asm/pgalloc.h>
  18
  19 static unsigned long page_table_shareable(struct vm_area_struct *svma,
  20                                 struct vm_area_struct *vma,
  21                                 unsigned long addr, pgoff_t idx)
  22 {
  23         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
  24                                 svma->vm_start;
  25         unsigned long sbase = saddr & PUD_MASK;
  26         unsigned long s_end = sbase + PUD_SIZE;
  27
  28         /* Allow segments to share if only one is marked locked */
  29         unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
  30         unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
  31
  32         /*
  33          * match the virtual addresses, permission and the alignment of the
  34          * page table page.
  35          */
  36         if (pmd_index(addr) != pmd_index(saddr) ||
  37             vm_flags != svm_flags ||
  38             sbase < svma->vm_start || svma->vm_end < s_end)
  39                 return 0;
  40
  41         return saddr;
  42 }
  43
  44 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
  45 {
  46         unsigned long base = addr & PUD_MASK;
  47         unsigned long end = base + PUD_SIZE;
  48
  49         /*
  50          * check on proper vm_flags and page table alignment
  51          */
  52         if (vma->vm_flags & VM_MAYSHARE &&
  53             vma->vm_start <= base && end <= vma->vm_end)
  54                 return 1;
  55         return 0;
  56 }
  57
  58 /*
  59  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
  60  * and returns the corresponding pte. While this is not necessary for the
  61  * !shared pmd case because we can allocate the pmd later as well, it makes the
  62  * code much cleaner. pmd allocation is essential for the shared case because
  63  * pud has to be populated inside the same i_mmap_mutex section - otherwise
  64  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
  65  * bad pmd for sharing.
  66  */
  67 static pte_t *
  68 huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
  69 {
  70         struct vm_area_struct *vma = find_vma(mm, addr);
  71         struct address_space *mapping = vma->vm_file->f_mapping;
  72         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
  73                         vma->vm_pgoff;
  74         struct prio_tree_iter iter;
  75         struct vm_area_struct *svma;
  76         unsigned long saddr;
  77         pte_t *spte = NULL;
  78         pte_t *pte;
  79
  80         if (!vma_shareable(vma, addr))
  81                 return (pte_t *)pmd_alloc(mm, pud, addr);
  82
  83         mutex_lock(&mapping->i_mmap_mutex);
  84         vma_prio_tree_foreach(svma, &iter, &mapping->i_mmap, idx, idx) {
  85                 if (svma == vma)
  86                         continue;
  87
  88                 saddr = page_table_shareable(svma, vma, addr, idx);
  89                 if (saddr) {
  90                         spte = huge_pte_offset(svma->vm_mm, saddr);
  91                         if (spte) {
  92                                 get_page(virt_to_page(spte));
  93                                 break;
  94                         }
  95                 }
  96         }
  97
  98         if (!spte)
  99                 goto out;
 100
 101         spin_lock(&mm->page_table_lock);
 102         if (pud_none(*pud))
 103                 pud_populate(mm, pud, (pmd_t *)((unsigned long)spte & PAGE_MASK));
 104         else
 105                 put_page(virt_to_page(spte));
 106         spin_unlock(&mm->page_table_lock);
 107 out:
 108         pte = (pte_t *)pmd_alloc(mm, pud, addr);
 109         mutex_unlock(&mapping->i_mmap_mutex);
 110         return pte;
 111 }
 112
 113 /*
 114  * unmap huge page backed by shared pte.
 115  *
 116  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
 117  * indicated by page_count > 1, unmap is achieved by clearing pud and
 118  * decrementing the ref count. If count == 1, the pte page is not shared.
 119  *
 120  * called with vma->vm_mm->page_table_lock held.
 121  *
 122  * returns: 1 successfully unmapped a shared pte page
 123  *          0 the underlying pte page is not shared, or it is the last user
 124  */
 125 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
 126 {
 127         pgd_t *pgd = pgd_offset(mm, *addr);
 128         pud_t *pud = pud_offset(pgd, *addr);
 129
 130         BUG_ON(page_count(virt_to_page(ptep)) == 0);
 131         if (page_count(virt_to_page(ptep)) == 1)
 132                 return 0;
 133
 134         pud_clear(pud);
 135         put_page(virt_to_page(ptep));
 136         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
 137         return 1;
 138 }
 139
 140 pte_t *huge_pte_alloc(struct mm_struct *mm,
 141                         unsigned long addr, unsigned long sz)
 142 {
 143         pgd_t *pgd;
 144         pud_t *pud;
 145         pte_t *pte = NULL;
 146
 147         pgd = pgd_offset(mm, addr);
 148         pud = pud_alloc(mm, pgd, addr);
 149         if (pud) {
 150                 if (sz == PUD_SIZE) {
 151                         pte = (pte_t *)pud;
 152                 } else {
 153                         BUG_ON(sz != PMD_SIZE);
 154                         if (pud_none(*pud))
 155                                 pte = huge_pmd_share(mm, addr, pud);
 156                         else
 157                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
 158                 }
 159         }
 160         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
 161
 162         return pte;
 163 }
 164
 165 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 166 {
 167         pgd_t *pgd;
 168         pud_t *pud;
 169         pmd_t *pmd = NULL;
 170
 171         pgd = pgd_offset(mm, addr);
 172         if (pgd_present(*pgd)) {
 173                 pud = pud_offset(pgd, addr);
 174                 if (pud_present(*pud)) {
 175                         if (pud_large(*pud))
 176                                 return (pte_t *)pud;
 177                         pmd = pmd_offset(pud, addr);
 178                 }
 179         }
 180         return (pte_t *) pmd;
 181 }
 182
 183 #if 0   /* This is just for testing */
 184 struct page *
 185 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 186 {
 187         unsigned long start = address;
 188         int length = 1;
 189         int nr;
 190         struct page *page;
 191         struct vm_area_struct *vma;
 192
 193         vma = find_vma(mm, addr);
 194         if (!vma || !is_vm_hugetlb_page(vma))
 195                 return ERR_PTR(-EINVAL);
 196
 197         pte = huge_pte_offset(mm, address);
 198
 199         /* hugetlb should be locked, and hence, prefaulted */
 200         WARN_ON(!pte || pte_none(*pte));
 201
 202         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 203
 204         WARN_ON(!PageHead(page));
 205
 206         return page;
 207 }
 208
 209 int pmd_huge(pmd_t pmd)
 210 {
 211         return 0;
 212 }
 213
 214 int pud_huge(pud_t pud)
 215 {
 216         return 0;
 217 }
 218
 219 struct page *
 220 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 221                 pmd_t *pmd, int write)
 222 {
 223         return NULL;
 224 }
 225
 226 #else
 227
 228 struct page *
 229 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 230 {
 231         return ERR_PTR(-EINVAL);
 232 }
 233
 234 int pmd_huge(pmd_t pmd)
 235 {
 236         return !!(pmd_val(pmd) & _PAGE_PSE);
 237 }
 238
 239 int pud_huge(pud_t pud)
 240 {
 241         return !!(pud_val(pud) & _PAGE_PSE);
 242 }
 243
 244 struct page *
 245 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 246                 pmd_t *pmd, int write)
 247 {
 248         struct page *page;
 249
 250         page = pte_page(*(pte_t *)pmd);
 251         if (page)
 252                 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
 253         return page;
 254 }
 255
 256 struct page *
 257 follow_huge_pud(struct mm_struct *mm, unsigned long address,
 258                 pud_t *pud, int write)
 259 {
 260         struct page *page;
 261
 262         page = pte_page(*(pte_t *)pud);
 263         if (page)
 264                 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
 265         return page;
 266 }
 267
 268 #endif
 269
 270 /* x86_64 also uses this file */
 271
 272 #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 273 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
 274                 unsigned long addr, unsigned long len,
 275                 unsigned long pgoff, unsigned long flags)
 276 {
 277         struct hstate *h = hstate_file(file);
 278         struct mm_struct *mm = current->mm;
 279         struct vm_area_struct *vma;
 280         unsigned long start_addr;
 281
 282         if (len > mm->cached_hole_size) {
 283                 start_addr = mm->free_area_cache;
 284         } else {
 285                 start_addr = TASK_UNMAPPED_BASE;
 286                 mm->cached_hole_size = 0;
 287         }
 288
 289 full_search:
 290         addr = ALIGN(start_addr, huge_page_size(h));
 291
 292         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
 293                 /* At this point:  (!vma || addr < vma->vm_end). */
 294                 if (TASK_SIZE - len < addr) {
 295                         /*
 296                          * Start a new search - just in case we missed
 297                          * some holes.
 298                          */
 299                         if (start_addr != TASK_UNMAPPED_BASE) {
 300                                 start_addr = TASK_UNMAPPED_BASE;
 301                                 mm->cached_hole_size = 0;
 302                                 goto full_search;
 303                         }
 304                         return -ENOMEM;
 305                 }
 306                 if (!vma || addr + len <= vma->vm_start) {
 307                         mm->free_area_cache = addr + len;
 308                         return addr;
 309                 }
 310                 if (addr + mm->cached_hole_size < vma->vm_start)
 311                         mm->cached_hole_size = vma->vm_start - addr;
 312                 addr = ALIGN(vma->vm_end, huge_page_size(h));
 313         }
 314 }
 315
 316 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
 317                 unsigned long addr0, unsigned long len,
 318                 unsigned long pgoff, unsigned long flags)
 319 {
 320         struct hstate *h = hstate_file(file);
 321         struct mm_struct *mm = current->mm;
 322         struct vm_area_struct *vma;
 323         unsigned long base = mm->mmap_base;
 324         unsigned long addr = addr0;
 325         unsigned long largest_hole = mm->cached_hole_size;
 326         unsigned long start_addr;
 327
 328         /* don't allow allocations above current base */
 329         if (mm->free_area_cache > base)
 330                 mm->free_area_cache = base;
 331
 332         if (len <= largest_hole) {
 333                 largest_hole = 0;
 334                 mm->free_area_cache  = base;
 335         }
 336 try_again:
 337         start_addr = mm->free_area_cache;
 338
 339         /* make sure it can fit in the remaining address space */
 340         if (mm->free_area_cache < len)
 341                 goto fail;
 342
 343         /* either no address requested or can't fit in requested address hole */
 344         addr = (mm->free_area_cache - len) & huge_page_mask(h);
 345         do {
 346                 /*
 347                  * Lookup failure means no vma is above this address,
 348                  * i.e. return with success:
 349                  */
 350                 vma = find_vma(mm, addr);
 351                 if (!vma)
 352                         return addr;
 353
 354                 if (addr + len <= vma->vm_start) {
 355                         /* remember the address as a hint for next time */
 356                         mm->cached_hole_size = largest_hole;
 357                         return (mm->free_area_cache = addr);
 358                 } else if (mm->free_area_cache == vma->vm_end) {
 359                         /* pull free_area_cache down to the first hole */
 360                         mm->free_area_cache = vma->vm_start;
 361                         mm->cached_hole_size = largest_hole;
 362                 }
 363
 364                 /* remember the largest hole we saw so far */
 365                 if (addr + largest_hole < vma->vm_start)
 366                         largest_hole = vma->vm_start - addr;
 367
 368                 /* try just below the current vma->vm_start */
 369                 addr = (vma->vm_start - len) & huge_page_mask(h);
 370         } while (len <= vma->vm_start);
 371
 372 fail:
 373         /*
 374          * if hint left us with no space for the requested
 375          * mapping then try again:
 376          */
 377         if (start_addr != base) {
 378                 mm->free_area_cache = base;
 379                 largest_hole = 0;
 380                 goto try_again;
 381         }
 382         /*
 383          * A failed mmap() very likely causes application failure,
 384          * so fall back to the bottom-up function here. This scenario
 385          * can happen with large stack limits and large mmap()
 386          * allocations.
 387          */
 388         mm->free_area_cache = TASK_UNMAPPED_BASE;
 389         mm->cached_hole_size = ~0UL;
 390         addr = hugetlb_get_unmapped_area_bottomup(file, addr0,
 391                         len, pgoff, flags);
 392
 393         /*
 394          * Restore the topdown base:
 395          */
 396         mm->free_area_cache = base;
 397         mm->cached_hole_size = ~0UL;
 398
 399         return addr;
 400 }
 401
 402 unsigned long
 403 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 404                 unsigned long len, unsigned long pgoff, unsigned long flags)
 405 {
 406         struct hstate *h = hstate_file(file);
 407         struct mm_struct *mm = current->mm;
 408         struct vm_area_struct *vma;
 409
 410         if (len & ~huge_page_mask(h))
 411                 return -EINVAL;
 412         if (len > TASK_SIZE)
 413                 return -ENOMEM;
 414
 415         if (flags & MAP_FIXED) {
 416                 if (prepare_hugepage_range(file, addr, len))
 417                         return -EINVAL;
 418                 return addr;
 419         }
 420
 421         if (addr) {
 422                 addr = ALIGN(addr, huge_page_size(h));
 423                 vma = find_vma(mm, addr);
 424                 if (TASK_SIZE - len >= addr &&
 425                     (!vma || addr + len <= vma->vm_start))
 426                         return addr;
 427         }
 428         if (mm->get_unmapped_area == arch_get_unmapped_area)
 429                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 430                                 pgoff, flags);
 431         else
 432                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 433                                 pgoff, flags);
 434 }
 435
 436 #endif /*HAVE_ARCH_HUGETLB_UNMAPPED_AREA*/
 437
 438 #ifdef CONFIG_X86_64
 439 static __init int setup_hugepagesz(char *opt)
 440 {
 441         unsigned long ps = memparse(opt, &opt);
 442         if (ps == PMD_SIZE) {
 443                 hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT);
 444         } else if (ps == PUD_SIZE && cpu_has_gbpages) {
 445                 hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
 446         } else {
 447                 printk(KERN_ERR "hugepagesz: Unsupported page size %lu M\n",
 448                         ps >> 20);
 449                 return 0;
 450         }
 451         return 1;
 452 }
 453 __setup("hugepagesz=", setup_hugepagesz);
 454 #endif