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