arch/ia64/mm/hugetlbpage.c

   1 /*
   2  * IA-64 Huge TLB Page Support for Kernel.
   3  *
   4  * Copyright (C) 2002-2004 Rohit Seth <rohit.seth@intel.com>
   5  * Copyright (C) 2003-2004 Ken Chen <kenneth.w.chen@intel.com>
   6  *
   7  * Sep, 2003: add numa support
   8  * Feb, 2004: dynamic hugetlb page size via boot parameter
   9  */
  10
  11 #include <linux/config.h>
  12 #include <linux/init.h>
  13 #include <linux/fs.h>
  14 #include <linux/mm.h>
  15 #include <linux/hugetlb.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/smp_lock.h>
  18 #include <linux/slab.h>
  19 #include <linux/sysctl.h>
  20 #include <asm/mman.h>
  21 #include <asm/pgalloc.h>
  22 #include <asm/tlb.h>
  23 #include <asm/tlbflush.h>
  24
  25 unsigned int hpage_shift=HPAGE_SHIFT_DEFAULT;
  26
  27 static pte_t *
  28 huge_pte_alloc (struct mm_struct *mm, unsigned long addr)
  29 {
  30         unsigned long taddr = htlbpage_to_page(addr);
  31         pgd_t *pgd;
  32         pmd_t *pmd;
  33         pte_t *pte = NULL;
  34
  35         pgd = pgd_offset(mm, taddr);
  36         pmd = pmd_alloc(mm, pgd, taddr);
  37         if (pmd)
  38                 pte = pte_alloc_map(mm, pmd, taddr);
  39         return pte;
  40 }
  41
  42 static pte_t *
  43 huge_pte_offset (struct mm_struct *mm, unsigned long addr)
  44 {
  45         unsigned long taddr = htlbpage_to_page(addr);
  46         pgd_t *pgd;
  47         pmd_t *pmd;
  48         pte_t *pte = NULL;
  49
  50         pgd = pgd_offset(mm, taddr);
  51         if (pgd_present(*pgd)) {
  52                 pmd = pmd_offset(pgd, taddr);
  53                 if (pmd_present(*pmd))
  54                         pte = pte_offset_map(pmd, taddr);
  55         }
  56
  57         return pte;
  58 }
  59
  60 #define mk_pte_huge(entry) { pte_val(entry) |= _PAGE_P; }
  61
  62 static void
  63 set_huge_pte (struct mm_struct *mm, struct vm_area_struct *vma,
  64               struct page *page, pte_t * page_table, int write_access)
  65 {
  66         pte_t entry;
  67
  68         mm->rss += (HPAGE_SIZE / PAGE_SIZE);
  69         if (write_access) {
  70                 entry =
  71                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
  72         } else
  73                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
  74         entry = pte_mkyoung(entry);
  75         mk_pte_huge(entry);
  76         set_pte(page_table, entry);
  77         return;
  78 }
  79 /*
  80  * This function checks for proper alignment of input addr and len parameters.
  81  */
  82 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
  83 {
  84         if (len & ~HPAGE_MASK)
  85                 return -EINVAL;
  86         if (addr & ~HPAGE_MASK)
  87                 return -EINVAL;
  88         if (REGION_NUMBER(addr) != REGION_HPAGE)
  89                 return -EINVAL;
  90
  91         return 0;
  92 }
  93
  94 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
  95                         struct vm_area_struct *vma)
  96 {
  97         pte_t *src_pte, *dst_pte, entry;
  98         struct page *ptepage;
  99         unsigned long addr = vma->vm_start;
 100         unsigned long end = vma->vm_end;
 101
 102         while (addr < end) {
 103                 dst_pte = huge_pte_alloc(dst, addr);
 104                 if (!dst_pte)
 105                         goto nomem;
 106                 src_pte = huge_pte_offset(src, addr);
 107                 entry = *src_pte;
 108                 ptepage = pte_page(entry);
 109                 get_page(ptepage);
 110                 set_pte(dst_pte, entry);
 111                 dst->rss += (HPAGE_SIZE / PAGE_SIZE);
 112                 addr += HPAGE_SIZE;
 113         }
 114         return 0;
 115 nomem:
 116         return -ENOMEM;
 117 }
 118
 119 int
 120 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 121                     struct page **pages, struct vm_area_struct **vmas,
 122                     unsigned long *st, int *length, int i)
 123 {
 124         pte_t *ptep, pte;
 125         unsigned long start = *st;
 126         unsigned long pstart;
 127         int len = *length;
 128         struct page *page;
 129
 130         do {
 131                 pstart = start & HPAGE_MASK;
 132                 ptep = huge_pte_offset(mm, start);
 133                 pte = *ptep;
 134
 135 back1:
 136                 page = pte_page(pte);
 137                 if (pages) {
 138                         page += ((start & ~HPAGE_MASK) >> PAGE_SHIFT);
 139                         get_page(page);
 140                         pages[i] = page;
 141                 }
 142                 if (vmas)
 143                         vmas[i] = vma;
 144                 i++;
 145                 len--;
 146                 start += PAGE_SIZE;
 147                 if (((start & HPAGE_MASK) == pstart) && len &&
 148                                 (start < vma->vm_end))
 149                         goto back1;
 150         } while (len && start < vma->vm_end);
 151         *length = len;
 152         *st = start;
 153         return i;
 154 }
 155
 156 struct page *follow_huge_addr(struct mm_struct *mm, unsigned long addr, int write)
 157 {
 158         struct page *page;
 159         pte_t *ptep;
 160
 161         if (! mm->used_hugetlb)
 162                 return ERR_PTR(-EINVAL);
 163         if (REGION_NUMBER(addr) != REGION_HPAGE)
 164                 return ERR_PTR(-EINVAL);
 165
 166         ptep = huge_pte_offset(mm, addr);
 167         if (!ptep || pte_none(*ptep))
 168                 return NULL;
 169         page = pte_page(*ptep);
 170         page += ((addr & ~HPAGE_MASK) >> PAGE_SHIFT);
 171         return page;
 172 }
 173 int pmd_huge(pmd_t pmd)
 174 {
 175         return 0;
 176 }
 177 struct page *
 178 follow_huge_pmd(struct mm_struct *mm, unsigned long address, pmd_t *pmd, int write)
 179 {
 180         return NULL;
 181 }
 182
 183 /*
 184  * Same as generic free_pgtables(), except constant PGDIR_* and pgd_offset
 185  * are hugetlb region specific.
 186  */
 187 void hugetlb_free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *prev,
 188         unsigned long start, unsigned long end)
 189 {
 190         unsigned long first = start & HUGETLB_PGDIR_MASK;
 191         unsigned long last = end + HUGETLB_PGDIR_SIZE - 1;
 192         unsigned long start_index, end_index;
 193         struct mm_struct *mm = tlb->mm;
 194
 195         if (!prev) {
 196                 prev = mm->mmap;
 197                 if (!prev)
 198                         goto no_mmaps;
 199                 if (prev->vm_end > start) {
 200                         if (last > prev->vm_start)
 201                                 last = prev->vm_start;
 202                         goto no_mmaps;
 203                 }
 204         }
 205         for (;;) {
 206                 struct vm_area_struct *next = prev->vm_next;
 207
 208                 if (next) {
 209                         if (next->vm_start < start) {
 210                                 prev = next;
 211                                 continue;
 212                         }
 213                         if (last > next->vm_start)
 214                                 last = next->vm_start;
 215                 }
 216                 if (prev->vm_end > first)
 217                         first = prev->vm_end + HUGETLB_PGDIR_SIZE - 1;
 218                 break;
 219         }
 220 no_mmaps:
 221         if (last < first)       /* for arches with discontiguous pgd indices */
 222                 return;
 223         /*
 224          * If the PGD bits are not consecutive in the virtual address, the
 225          * old method of shifting the VA >> by PGDIR_SHIFT doesn't work.
 226          */
 227
 228         start_index = pgd_index(htlbpage_to_page(first));
 229         end_index = pgd_index(htlbpage_to_page(last));
 230
 231         if (end_index > start_index) {
 232                 clear_page_tables(tlb, start_index, end_index - start_index);
 233         }
 234 }
 235
 236 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
 237 {
 238         struct mm_struct *mm = vma->vm_mm;
 239         unsigned long address;
 240         pte_t *pte;
 241         struct page *page;
 242
 243         BUG_ON(start & (HPAGE_SIZE - 1));
 244         BUG_ON(end & (HPAGE_SIZE - 1));
 245
 246         for (address = start; address < end; address += HPAGE_SIZE) {
 247                 pte = huge_pte_offset(mm, address);
 248                 if (pte_none(*pte))
 249                         continue;
 250                 page = pte_page(*pte);
 251                 put_page(page);
 252                 pte_clear(pte);
 253         }
 254         mm->rss -= (end - start) >> PAGE_SHIFT;
 255         flush_tlb_range(vma, start, end);
 256 }
 257
 258 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 259 {
 260         struct mm_struct *mm = current->mm;
 261         unsigned long addr;
 262         int ret = 0;
 263
 264         BUG_ON(vma->vm_start & ~HPAGE_MASK);
 265         BUG_ON(vma->vm_end & ~HPAGE_MASK);
 266
 267         spin_lock(&mm->page_table_lock);
 268         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 269                 unsigned long idx;
 270                 pte_t *pte = huge_pte_alloc(mm, addr);
 271                 struct page *page;
 272
 273                 if (!pte) {
 274                         ret = -ENOMEM;
 275                         goto out;
 276                 }
 277                 if (!pte_none(*pte))
 278                         continue;
 279
 280                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 281                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 282                 page = find_get_page(mapping, idx);
 283                 if (!page) {
 284                         /* charge the fs quota first */
 285                         if (hugetlb_get_quota(mapping)) {
 286                                 ret = -ENOMEM;
 287                                 goto out;
 288                         }
 289                         page = alloc_huge_page();
 290                         if (!page) {
 291                                 hugetlb_put_quota(mapping);
 292                                 ret = -ENOMEM;
 293                                 goto out;
 294                         }
 295                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 296                         unlock_page(page);
 297                         if (ret) {
 298                                 hugetlb_put_quota(mapping);
 299                                 free_huge_page(page);
 300                                 goto out;
 301                         }
 302                 }
 303                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
 304         }
 305 out:
 306         spin_unlock(&mm->page_table_lock);
 307         return ret;
 308 }
 309
 310 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
 311                 unsigned long pgoff, unsigned long flags)
 312 {
 313         struct vm_area_struct *vmm;
 314
 315         if (len > RGN_MAP_LIMIT)
 316                 return -ENOMEM;
 317         if (len & ~HPAGE_MASK)
 318                 return -EINVAL;
 319         /* This code assumes that REGION_HPAGE != 0. */
 320         if ((REGION_NUMBER(addr) != REGION_HPAGE) || (addr & (HPAGE_SIZE - 1)))
 321                 addr = HPAGE_REGION_BASE;
 322         else
 323                 addr = ALIGN(addr, HPAGE_SIZE);
 324         for (vmm = find_vma(current->mm, addr); ; vmm = vmm->vm_next) {
 325                 /* At this point:  (!vmm || addr < vmm->vm_end). */
 326                 if (REGION_OFFSET(addr) + len > RGN_MAP_LIMIT)
 327                         return -ENOMEM;
 328                 if (!vmm || (addr + len) <= vmm->vm_start)
 329                         return addr;
 330                 addr = ALIGN(vmm->vm_end, HPAGE_SIZE);
 331         }
 332 }
 333
 334 static int __init hugetlb_setup_sz(char *str)
 335 {
 336         u64 tr_pages;
 337         unsigned long long size;
 338
 339         if (ia64_pal_vm_page_size(&tr_pages, NULL) != 0)
 340                 /*
 341                  * shouldn't happen, but just in case.
 342                  */
 343                 tr_pages = 0x15557000UL;
 344
 345         size = memparse(str, &str);
 346         if (*str || (size & (size-1)) || !(tr_pages & size) ||
 347                 size <= PAGE_SIZE ||
 348                 size >= (1UL << PAGE_SHIFT << MAX_ORDER)) {
 349                 printk(KERN_WARNING "Invalid huge page size specified\n");
 350                 return 1;
 351         }
 352
 353         hpage_shift = __ffs(size);
 354         /*
 355          * boot cpu already executed ia64_mmu_init, and has HPAGE_SHIFT_DEFAULT
 356          * override here with new page shift.
 357          */
 358         ia64_set_rr(HPAGE_REGION_BASE, hpage_shift << 2);
 359         return 1;
 360 }
 361 __setup("hugepagesz=", hugetlb_setup_sz);