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         pmd_t *pmd = NULL;
  25
  26         pgd = pgd_offset(mm, addr);
  27         pmd = pmd_alloc(mm, pgd, addr);
  28         return (pte_t *) pmd;
  29 }
  30
  31 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
  32 {
  33         pgd_t *pgd;
  34         pmd_t *pmd = NULL;
  35
  36         pgd = pgd_offset(mm, addr);
  37         pmd = pmd_offset(pgd, addr);
  38         return (pte_t *) pmd;
  39 }
  40
  41 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma, struct page *page, pte_t * page_table, int write_access)
  42 {
  43         pte_t entry;
  44
  45         // mm->rss += (HPAGE_SIZE / PAGE_SIZE);
  46         vx_rsspages_add(mm, HPAGE_SIZE / PAGE_SIZE);
  47         if (write_access) {
  48                 entry =
  49                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
  50         } else
  51                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
  52         entry = pte_mkyoung(entry);
  53         mk_pte_huge(entry);
  54         set_pte(page_table, entry);
  55 }
  56
  57 /*
  58  * This function checks for proper alignment of input addr and len parameters.
  59  */
  60 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
  61 {
  62         if (len & ~HPAGE_MASK)
  63                 return -EINVAL;
  64         if (addr & ~HPAGE_MASK)
  65                 return -EINVAL;
  66         return 0;
  67 }
  68
  69 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
  70                         struct vm_area_struct *vma)
  71 {
  72         pte_t *src_pte, *dst_pte, entry;
  73         struct page *ptepage;
  74         unsigned long addr = vma->vm_start;
  75         unsigned long end = vma->vm_end;
  76
  77         while (addr < end) {
  78                 dst_pte = huge_pte_alloc(dst, addr);
  79                 if (!dst_pte)
  80                         goto nomem;
  81                 src_pte = huge_pte_offset(src, addr);
  82                 entry = *src_pte;
  83                 ptepage = pte_page(entry);
  84                 get_page(ptepage);
  85                 set_pte(dst_pte, entry);
  86                 // dst->rss += (HPAGE_SIZE / PAGE_SIZE);
  87                 vx_rsspages_add(dst, HPAGE_SIZE / PAGE_SIZE);
  88                 addr += HPAGE_SIZE;
  89         }
  90         return 0;
  91
  92 nomem:
  93         return -ENOMEM;
  94 }
  95
  96 int
  97 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
  98                     struct page **pages, struct vm_area_struct **vmas,
  99                     unsigned long *position, int *length, int i)
 100 {
 101         unsigned long vpfn, vaddr = *position;
 102         int remainder = *length;
 103
 104         WARN_ON(!is_vm_hugetlb_page(vma));
 105
 106         vpfn = vaddr/PAGE_SIZE;
 107         while (vaddr < vma->vm_end && remainder) {
 108
 109                 if (pages) {
 110                         pte_t *pte;
 111                         struct page *page;
 112
 113                         pte = huge_pte_offset(mm, vaddr);
 114
 115                         /* hugetlb should be locked, and hence, prefaulted */
 116                         WARN_ON(!pte || pte_none(*pte));
 117
 118                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 119
 120                         WARN_ON(!PageCompound(page));
 121
 122                         get_page(page);
 123                         pages[i] = page;
 124                 }
 125
 126                 if (vmas)
 127                         vmas[i] = vma;
 128
 129                 vaddr += PAGE_SIZE;
 130                 ++vpfn;
 131                 --remainder;
 132                 ++i;
 133         }
 134
 135         *length = remainder;
 136         *position = vaddr;
 137
 138         return i;
 139 }
 140
 141 #if 0   /* This is just for testing */
 142 struct page *
 143 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 144 {
 145         unsigned long start = address;
 146         int length = 1;
 147         int nr;
 148         struct page *page;
 149         struct vm_area_struct *vma;
 150
 151         vma = find_vma(mm, addr);
 152         if (!vma || !is_vm_hugetlb_page(vma))
 153                 return ERR_PTR(-EINVAL);
 154
 155         pte = huge_pte_offset(mm, address);
 156
 157         /* hugetlb should be locked, and hence, prefaulted */
 158         WARN_ON(!pte || pte_none(*pte));
 159
 160         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 161
 162         WARN_ON(!PageCompound(page));
 163
 164         return page;
 165 }
 166
 167 int pmd_huge(pmd_t pmd)
 168 {
 169         return 0;
 170 }
 171
 172 struct page *
 173 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 174                 pmd_t *pmd, int write)
 175 {
 176         return NULL;
 177 }
 178
 179 #else
 180
 181 struct page *
 182 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 183 {
 184         return ERR_PTR(-EINVAL);
 185 }
 186
 187 int pmd_huge(pmd_t pmd)
 188 {
 189         return !!(pmd_val(pmd) & _PAGE_PSE);
 190 }
 191
 192 struct page *
 193 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 194                 pmd_t *pmd, int write)
 195 {
 196         struct page *page;
 197
 198         page = pte_page(*(pte_t *)pmd);
 199         if (page)
 200                 page += ((address & ~HPAGE_MASK) >> PAGE_SHIFT);
 201         return page;
 202 }
 203 #endif
 204
 205 void unmap_hugepage_range(struct vm_area_struct *vma,
 206                 unsigned long start, unsigned long end)
 207 {
 208         struct mm_struct *mm = vma->vm_mm;
 209         unsigned long address;
 210         pte_t pte;
 211         struct page *page;
 212
 213         BUG_ON(start & (HPAGE_SIZE - 1));
 214         BUG_ON(end & (HPAGE_SIZE - 1));
 215
 216         for (address = start; address < end; address += HPAGE_SIZE) {
 217                 pte = ptep_get_and_clear(huge_pte_offset(mm, address));
 218                 if (pte_none(pte))
 219                         continue;
 220                 page = pte_page(pte);
 221                 put_page(page);
 222         }
 223         // mm->rss -= (end - start) >> PAGE_SHIFT;
 224         vx_rsspages_sub(mm, (end - start) >> PAGE_SHIFT);
 225         flush_tlb_range(vma, start, end);
 226 }
 227
 228 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 229 {
 230         struct mm_struct *mm = current->mm;
 231         unsigned long addr;
 232         int ret = 0;
 233
 234         BUG_ON(vma->vm_start & ~HPAGE_MASK);
 235         BUG_ON(vma->vm_end & ~HPAGE_MASK);
 236
 237         spin_lock(&mm->page_table_lock);
 238         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 239                 unsigned long idx;
 240                 pte_t *pte = huge_pte_alloc(mm, addr);
 241                 struct page *page;
 242
 243                 if (!pte) {
 244                         ret = -ENOMEM;
 245                         goto out;
 246                 }
 247
 248                 if (!pte_none(*pte)) {
 249                         pmd_t *pmd = (pmd_t *) pte;
 250
 251                         page = pmd_page(*pmd);
 252                         pmd_clear(pmd);
 253                         mm->nr_ptes--;
 254                         dec_page_state(nr_page_table_pages);
 255                         page_cache_release(page);
 256                 }
 257
 258                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 259                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 260                 page = find_get_page(mapping, idx);
 261                 if (!page) {
 262                         /* charge the fs quota first */
 263                         if (hugetlb_get_quota(mapping)) {
 264                                 ret = -ENOMEM;
 265                                 goto out;
 266                         }
 267                         page = alloc_huge_page();
 268                         if (!page) {
 269                                 hugetlb_put_quota(mapping);
 270                                 ret = -ENOMEM;
 271                                 goto out;
 272                         }
 273                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 274                         if (! ret) {
 275                                 unlock_page(page);
 276                         } else {
 277                                 hugetlb_put_quota(mapping);
 278                                 free_huge_page(page);
 279                                 goto out;
 280                         }
 281                 }
 282                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
 283         }
 284 out:
 285         spin_unlock(&mm->page_table_lock);
 286         return ret;
 287 }
 288
 289 /* x86_64 also uses this file */
 290
 291 #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 292 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
 293                 unsigned long addr, unsigned long len,
 294                 unsigned long pgoff, unsigned long flags)
 295 {
 296         struct mm_struct *mm = current->mm;
 297         struct vm_area_struct *vma;
 298         unsigned long start_addr;
 299
 300         start_addr = mm->free_area_cache;
 301
 302 full_search:
 303         addr = ALIGN(start_addr, HPAGE_SIZE);
 304
 305         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
 306                 /* At this point:  (!vma || addr < vma->vm_end). */
 307                 if (TASK_SIZE - len < addr) {
 308                         /*
 309                          * Start a new search - just in case we missed
 310                          * some holes.
 311                          */
 312                         if (start_addr != TASK_UNMAPPED_BASE) {
 313                                 start_addr = TASK_UNMAPPED_BASE;
 314                                 goto full_search;
 315                         }
 316                         return -ENOMEM;
 317                 }
 318                 if (!vma || addr + len <= vma->vm_start) {
 319                         mm->free_area_cache = addr + len;
 320                         return addr;
 321                 }
 322                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 323         }
 324 }
 325
 326 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
 327                 unsigned long addr0, unsigned long len,
 328                 unsigned long pgoff, unsigned long flags)
 329 {
 330         struct mm_struct *mm = current->mm;
 331         struct vm_area_struct *vma, *prev_vma;
 332         unsigned long base = mm->mmap_base, addr = addr0;
 333         int first_time = 1;
 334
 335         /* don't allow allocations above current base */
 336         if (mm->free_area_cache > base)
 337                 mm->free_area_cache = base;
 338
 339 try_again:
 340         /* make sure it can fit in the remaining address space */
 341         if (mm->free_area_cache < len)
 342                 goto fail;
 343
 344         /* either no address requested or cant fit in requested address hole */
 345         addr = (mm->free_area_cache - len) & HPAGE_MASK;
 346         do {
 347                 /*
 348                  * Lookup failure means no vma is above this address,
 349                  * i.e. return with success:
 350                  */
 351                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
 352                         return addr;
 353
 354                 /*
 355                  * new region fits between prev_vma->vm_end and
 356                  * vma->vm_start, use it:
 357                  */
 358                 if (addr + len <= vma->vm_start &&
 359                                 (!prev_vma || (addr >= prev_vma->vm_end)))
 360                         /* remember the address as a hint for next time */
 361                         return (mm->free_area_cache = addr);
 362                 else
 363                         /* pull free_area_cache down to the first hole */
 364                         if (mm->free_area_cache == vma->vm_end)
 365                                 mm->free_area_cache = vma->vm_start;
 366
 367                 /* try just below the current vma->vm_start */
 368                 addr = (vma->vm_start - len) & HPAGE_MASK;
 369         } while (len <= vma->vm_start);
 370
 371 fail:
 372         /*
 373          * if hint left us with no space for the requested
 374          * mapping then try again:
 375          */
 376         if (first_time) {
 377                 mm->free_area_cache = base;
 378                 first_time = 0;
 379                 goto try_again;
 380         }
 381         /*
 382          * A failed mmap() very likely causes application failure,
 383          * so fall back to the bottom-up function here. This scenario
 384          * can happen with large stack limits and large mmap()
 385          * allocations.
 386          */
 387         mm->free_area_cache = TASK_UNMAPPED_BASE;
 388         addr = hugetlb_get_unmapped_area_bottomup(file, addr0,
 389                         len, pgoff, flags);
 390
 391         /*
 392          * Restore the topdown base:
 393          */
 394         mm->free_area_cache = base;
 395
 396         return addr;
 397 }
 398
 399 unsigned long
 400 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 401                 unsigned long len, unsigned long pgoff, unsigned long flags)
 402 {
 403         struct mm_struct *mm = current->mm;
 404         struct vm_area_struct *vma;
 405
 406         if (len & ~HPAGE_MASK)
 407                 return -EINVAL;
 408         if (len > TASK_SIZE)
 409                 return -ENOMEM;
 410
 411         if (addr) {
 412                 addr = ALIGN(addr, HPAGE_SIZE);
 413                 vma = find_vma(mm, addr);
 414                 if (TASK_SIZE - len >= addr &&
 415                     (!vma || addr + len <= vma->vm_start))
 416                         return addr;
 417         }
 418         if (mm->get_unmapped_area == arch_get_unmapped_area)
 419                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 420                                 pgoff, flags);
 421         else
 422                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 423                                 pgoff, flags);
 424 }
 425
 426 #endif /*HAVE_ARCH_HUGETLB_UNMAPPED_AREA*/
 427