arch/i386/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/config.h>
   8 #include <linux/init.h>
   9 #include <linux/fs.h>
  10 #include <linux/mm.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/smp_lock.h>
  14 #include <linux/slab.h>
  15 #include <linux/err.h>
  16 #include <linux/sysctl.h>
  17 #include <asm/mman.h>
  18 #include <asm/tlb.h>
  19 #include <asm/tlbflush.h>
  20
  21 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
  22 {
  23         pgd_t *pgd;
  24         pud_t *pud;
  25         pmd_t *pmd = NULL;
  26
  27         pgd = pgd_offset(mm, addr);
  28         pud = pud_alloc(mm, pgd, addr);
  29         pmd = pmd_alloc(mm, pud, addr);
  30         return (pte_t *) pmd;
  31 }
  32
  33 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
  34 {
  35         pgd_t *pgd;
  36         pud_t *pud;
  37         pmd_t *pmd = NULL;
  38
  39         pgd = pgd_offset(mm, addr);
  40         pud = pud_offset(pgd, addr);
  41         pmd = pmd_offset(pud, addr);
  42         return (pte_t *) pmd;
  43 }
  44
  45 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma, struct page *page, pte_t * page_table, int write_access)
  46 {
  47         pte_t entry;
  48
  49         vx_rsspages_add(mm, HPAGE_SIZE / PAGE_SIZE);
  50         if (write_access) {
  51                 entry =
  52                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
  53         } else
  54                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
  55         entry = pte_mkyoung(entry);
  56         mk_pte_huge(entry);
  57         set_pte(page_table, entry);
  58 }
  59
  60 /*
  61  * This function checks for proper alignment of input addr and len parameters.
  62  */
  63 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
  64 {
  65         if (len & ~HPAGE_MASK)
  66                 return -EINVAL;
  67         if (addr & ~HPAGE_MASK)
  68                 return -EINVAL;
  69         return 0;
  70 }
  71
  72 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
  73                         struct vm_area_struct *vma)
  74 {
  75         pte_t *src_pte, *dst_pte, entry;
  76         struct page *ptepage;
  77         unsigned long addr = vma->vm_start;
  78         unsigned long end = vma->vm_end;
  79
  80         while (addr < end) {
  81                 dst_pte = huge_pte_alloc(dst, addr);
  82                 if (!dst_pte)
  83                         goto nomem;
  84                 src_pte = huge_pte_offset(src, addr);
  85                 entry = *src_pte;
  86                 ptepage = pte_page(entry);
  87                 get_page(ptepage);
  88                 set_pte(dst_pte, entry);
  89                 vx_rsspages_add(dst, HPAGE_SIZE / PAGE_SIZE);
  90                 addr += HPAGE_SIZE;
  91         }
  92         return 0;
  93
  94 nomem:
  95         return -ENOMEM;
  96 }
  97
  98 int
  99 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 100                     struct page **pages, struct vm_area_struct **vmas,
 101                     unsigned long *position, int *length, int i)
 102 {
 103         unsigned long vpfn, vaddr = *position;
 104         int remainder = *length;
 105
 106         WARN_ON(!is_vm_hugetlb_page(vma));
 107
 108         vpfn = vaddr/PAGE_SIZE;
 109         while (vaddr < vma->vm_end && remainder) {
 110
 111                 if (pages) {
 112                         pte_t *pte;
 113                         struct page *page;
 114
 115                         pte = huge_pte_offset(mm, vaddr);
 116
 117                         /* hugetlb should be locked, and hence, prefaulted */
 118                         WARN_ON(!pte || pte_none(*pte));
 119
 120                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 121
 122                         WARN_ON(!PageCompound(page));
 123
 124                         get_page(page);
 125                         pages[i] = page;
 126                 }
 127
 128                 if (vmas)
 129                         vmas[i] = vma;
 130
 131                 vaddr += PAGE_SIZE;
 132                 ++vpfn;
 133                 --remainder;
 134                 ++i;
 135         }
 136
 137         *length = remainder;
 138         *position = vaddr;
 139
 140         return i;
 141 }
 142
 143 #if 0   /* This is just for testing */
 144 struct page *
 145 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 146 {
 147         unsigned long start = address;
 148         int length = 1;
 149         int nr;
 150         struct page *page;
 151         struct vm_area_struct *vma;
 152
 153         vma = find_vma(mm, addr);
 154         if (!vma || !is_vm_hugetlb_page(vma))
 155                 return ERR_PTR(-EINVAL);
 156
 157         pte = huge_pte_offset(mm, address);
 158
 159         /* hugetlb should be locked, and hence, prefaulted */
 160         WARN_ON(!pte || pte_none(*pte));
 161
 162         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 163
 164         WARN_ON(!PageCompound(page));
 165
 166         return page;
 167 }
 168
 169 int pmd_huge(pmd_t pmd)
 170 {
 171         return 0;
 172 }
 173
 174 struct page *
 175 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 176                 pmd_t *pmd, int write)
 177 {
 178         return NULL;
 179 }
 180
 181 #else
 182
 183 struct page *
 184 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 185 {
 186         return ERR_PTR(-EINVAL);
 187 }
 188
 189 int pmd_huge(pmd_t pmd)
 190 {
 191         return !!(pmd_val(pmd) & _PAGE_PSE);
 192 }
 193
 194 struct page *
 195 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 196                 pmd_t *pmd, int write)
 197 {
 198         struct page *page;
 199
 200         page = pte_page(*(pte_t *)pmd);
 201         if (page)
 202                 page += ((address & ~HPAGE_MASK) >> PAGE_SHIFT);
 203         return page;
 204 }
 205 #endif
 206
 207 void unmap_hugepage_range(struct vm_area_struct *vma,
 208                 unsigned long start, unsigned long end)
 209 {
 210         struct mm_struct *mm = vma->vm_mm;
 211         unsigned long address;
 212         pte_t pte;
 213         struct page *page;
 214
 215         BUG_ON(start & (HPAGE_SIZE - 1));
 216         BUG_ON(end & (HPAGE_SIZE - 1));
 217
 218         for (address = start; address < end; address += HPAGE_SIZE) {
 219                 pte = ptep_get_and_clear(huge_pte_offset(mm, address));
 220                 if (pte_none(pte))
 221                         continue;
 222                 page = pte_page(pte);
 223                 put_page(page);
 224         }
 225         vx_rsspages_sub(mm, (end - start) >> PAGE_SHIFT);
 226         flush_tlb_range(vma, start, end);
 227 }
 228
 229 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 230 {
 231         struct mm_struct *mm = current->mm;
 232         unsigned long addr;
 233         int ret = 0;
 234
 235         BUG_ON(vma->vm_start & ~HPAGE_MASK);
 236         BUG_ON(vma->vm_end & ~HPAGE_MASK);
 237
 238         spin_lock(&mm->page_table_lock);
 239         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 240                 unsigned long idx;
 241                 pte_t *pte = huge_pte_alloc(mm, addr);
 242                 struct page *page;
 243
 244                 if (!pte) {
 245                         ret = -ENOMEM;
 246                         goto out;
 247                 }
 248
 249                 if (!pte_none(*pte)) {
 250                         pmd_t *pmd = (pmd_t *) pte;
 251
 252                         page = pmd_page(*pmd);
 253                         pmd_clear(pmd);
 254                         mm->nr_ptes--;
 255                         dec_page_state(nr_page_table_pages);
 256                         page_cache_release(page);
 257                 }
 258
 259                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 260                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 261                 page = find_get_page(mapping, idx);
 262                 if (!page) {
 263                         /* charge the fs quota first */
 264                         if (hugetlb_get_quota(mapping)) {
 265                                 ret = -ENOMEM;
 266                                 goto out;
 267                         }
 268                         page = alloc_huge_page();
 269                         if (!page) {
 270                                 hugetlb_put_quota(mapping);
 271                                 ret = -ENOMEM;
 272                                 goto out;
 273                         }
 274                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 275                         if (! ret) {
 276                                 unlock_page(page);
 277                         } else {
 278                                 hugetlb_put_quota(mapping);
 279                                 free_huge_page(page);
 280                                 goto out;
 281                         }
 282                 }
 283                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
 284         }
 285 out:
 286         spin_unlock(&mm->page_table_lock);
 287         return ret;
 288 }
 289
 290 /* x86_64 also uses this file */
 291
 292 #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 293 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
 294                 unsigned long addr, unsigned long len,
 295                 unsigned long pgoff, unsigned long flags)
 296 {
 297         struct mm_struct *mm = current->mm;
 298         struct vm_area_struct *vma;
 299         unsigned long start_addr;
 300
 301         start_addr = mm->free_area_cache;
 302
 303 full_search:
 304         addr = ALIGN(start_addr, HPAGE_SIZE);
 305
 306         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
 307                 /* At this point:  (!vma || addr < vma->vm_end). */
 308                 if (TASK_SIZE - len < addr) {
 309                         /*
 310                          * Start a new search - just in case we missed
 311                          * some holes.
 312                          */
 313                         if (start_addr != TASK_UNMAPPED_BASE) {
 314                                 start_addr = TASK_UNMAPPED_BASE;
 315                                 goto full_search;
 316                         }
 317                         return -ENOMEM;
 318                 }
 319                 if (!vma || addr + len <= vma->vm_start) {
 320                         mm->free_area_cache = addr + len;
 321                         return addr;
 322                 }
 323                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 324         }
 325 }
 326
 327 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
 328                 unsigned long addr0, unsigned long len,
 329                 unsigned long pgoff, unsigned long flags)
 330 {
 331         struct mm_struct *mm = current->mm;
 332         struct vm_area_struct *vma, *prev_vma;
 333         unsigned long base = mm->mmap_base, addr = addr0;
 334         int first_time = 1;
 335
 336         /* don't allow allocations above current base */
 337         if (mm->free_area_cache > base)
 338                 mm->free_area_cache = base;
 339
 340 try_again:
 341         /* make sure it can fit in the remaining address space */
 342         if (mm->free_area_cache < len)
 343                 goto fail;
 344
 345         /* either no address requested or cant fit in requested address hole */
 346         addr = (mm->free_area_cache - len) & HPAGE_MASK;
 347         do {
 348                 /*
 349                  * Lookup failure means no vma is above this address,
 350                  * i.e. return with success:
 351                  */
 352                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
 353                         return addr;
 354
 355                 /*
 356                  * new region fits between prev_vma->vm_end and
 357                  * vma->vm_start, use it:
 358                  */
 359                 if (addr + len <= vma->vm_start &&
 360                                 (!prev_vma || (addr >= prev_vma->vm_end)))
 361                         /* remember the address as a hint for next time */
 362                         return (mm->free_area_cache = addr);
 363                 else
 364                         /* pull free_area_cache down to the first hole */
 365                         if (mm->free_area_cache == vma->vm_end)
 366                                 mm->free_area_cache = vma->vm_start;
 367
 368                 /* try just below the current vma->vm_start */
 369                 addr = (vma->vm_start - len) & HPAGE_MASK;
 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 (first_time) {
 378                 mm->free_area_cache = base;
 379                 first_time = 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         addr = hugetlb_get_unmapped_area_bottomup(file, addr0,
 390                         len, pgoff, flags);
 391
 392         /*
 393          * Restore the topdown base:
 394          */
 395         mm->free_area_cache = base;
 396
 397         return addr;
 398 }
 399
 400 unsigned long
 401 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 402                 unsigned long len, unsigned long pgoff, unsigned long flags)
 403 {
 404         struct mm_struct *mm = current->mm;
 405         struct vm_area_struct *vma;
 406
 407         if (len & ~HPAGE_MASK)
 408                 return -EINVAL;
 409         if (len > TASK_SIZE)
 410                 return -ENOMEM;
 411
 412         if (addr) {
 413                 addr = ALIGN(addr, HPAGE_SIZE);
 414                 vma = find_vma(mm, addr);
 415                 if (TASK_SIZE - len >= addr &&
 416                     (!vma || addr + len <= vma->vm_start))
 417                         return addr;
 418         }
 419         if (mm->get_unmapped_area == arch_get_unmapped_area)
 420                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 421                                 pgoff, flags);
 422         else
 423                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 424                                 pgoff, flags);
 425 }
 426
 427 #endif /*HAVE_ARCH_HUGETLB_UNMAPPED_AREA*/
 428