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         vx_rsspages_add(mm, HPAGE_SIZE / PAGE_SIZE);
  70         if (write_access) {
  71                 entry =
  72                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
  73         } else
  74                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
  75         entry = pte_mkyoung(entry);
  76         mk_pte_huge(entry);
  77         set_pte(page_table, entry);
  78         return;
  79 }
  80 /*
  81  * This function checks for proper alignment of input addr and len parameters.
  82  */
  83 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
  84 {
  85         if (len & ~HPAGE_MASK)
  86                 return -EINVAL;
  87         if (addr & ~HPAGE_MASK)
  88                 return -EINVAL;
  89         if (REGION_NUMBER(addr) != REGION_HPAGE)
  90                 return -EINVAL;
  91
  92         return 0;
  93 }
  94
  95 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
  96                         struct vm_area_struct *vma)
  97 {
  98         pte_t *src_pte, *dst_pte, entry;
  99         struct page *ptepage;
 100         unsigned long addr = vma->vm_start;
 101         unsigned long end = vma->vm_end;
 102
 103         while (addr < end) {
 104                 dst_pte = huge_pte_alloc(dst, addr);
 105                 if (!dst_pte)
 106                         goto nomem;
 107                 src_pte = huge_pte_offset(src, addr);
 108                 entry = *src_pte;
 109                 ptepage = pte_page(entry);
 110                 get_page(ptepage);
 111                 set_pte(dst_pte, entry);
 112                 // dst->rss += (HPAGE_SIZE / PAGE_SIZE);
 113                 vx_rsspages_add(dst, HPAGE_SIZE / PAGE_SIZE);
 114                 addr += HPAGE_SIZE;
 115         }
 116         return 0;
 117 nomem:
 118         return -ENOMEM;
 119 }
 120
 121 int
 122 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 123                     struct page **pages, struct vm_area_struct **vmas,
 124                     unsigned long *st, int *length, int i)
 125 {
 126         pte_t *ptep, pte;
 127         unsigned long start = *st;
 128         unsigned long pstart;
 129         int len = *length;
 130         struct page *page;
 131
 132         do {
 133                 pstart = start & HPAGE_MASK;
 134                 ptep = huge_pte_offset(mm, start);
 135                 pte = *ptep;
 136
 137 back1:
 138                 page = pte_page(pte);
 139                 if (pages) {
 140                         page += ((start & ~HPAGE_MASK) >> PAGE_SHIFT);
 141                         get_page(page);
 142                         pages[i] = page;
 143                 }
 144                 if (vmas)
 145                         vmas[i] = vma;
 146                 i++;
 147                 len--;
 148                 start += PAGE_SIZE;
 149                 if (((start & HPAGE_MASK) == pstart) && len &&
 150                                 (start < vma->vm_end))
 151                         goto back1;
 152         } while (len && start < vma->vm_end);
 153         *length = len;
 154         *st = start;
 155         return i;
 156 }
 157
 158 struct page *follow_huge_addr(struct mm_struct *mm, unsigned long addr, int write)
 159 {
 160         struct page *page;
 161         pte_t *ptep;
 162
 163         if (! mm->used_hugetlb)
 164                 return ERR_PTR(-EINVAL);
 165         if (REGION_NUMBER(addr) != REGION_HPAGE)
 166                 return ERR_PTR(-EINVAL);
 167
 168         ptep = huge_pte_offset(mm, addr);
 169         if (!ptep || pte_none(*ptep))
 170                 return NULL;
 171         page = pte_page(*ptep);
 172         page += ((addr & ~HPAGE_MASK) >> PAGE_SHIFT);
 173         return page;
 174 }
 175 int pmd_huge(pmd_t pmd)
 176 {
 177         return 0;
 178 }
 179 struct page *
 180 follow_huge_pmd(struct mm_struct *mm, unsigned long address, pmd_t *pmd, int write)
 181 {
 182         return NULL;
 183 }
 184
 185 /*
 186  * Same as generic free_pgtables(), except constant PGDIR_* and pgd_offset
 187  * are hugetlb region specific.
 188  */
 189 void hugetlb_free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *prev,
 190         unsigned long start, unsigned long end)
 191 {
 192         unsigned long first = start & HUGETLB_PGDIR_MASK;
 193         unsigned long last = end + HUGETLB_PGDIR_SIZE - 1;
 194         unsigned long start_index, end_index;
 195         struct mm_struct *mm = tlb->mm;
 196
 197         if (!prev) {
 198                 prev = mm->mmap;
 199                 if (!prev)
 200                         goto no_mmaps;
 201                 if (prev->vm_end > start) {
 202                         if (last > prev->vm_start)
 203                                 last = prev->vm_start;
 204                         goto no_mmaps;
 205                 }
 206         }
 207         for (;;) {
 208                 struct vm_area_struct *next = prev->vm_next;
 209
 210                 if (next) {
 211                         if (next->vm_start < start) {
 212                                 prev = next;
 213                                 continue;
 214                         }
 215                         if (last > next->vm_start)
 216                                 last = next->vm_start;
 217                 }
 218                 if (prev->vm_end > first)
 219                         first = prev->vm_end + HUGETLB_PGDIR_SIZE - 1;
 220                 break;
 221         }
 222 no_mmaps:
 223         if (last < first)       /* for arches with discontiguous pgd indices */
 224                 return;
 225         /*
 226          * If the PGD bits are not consecutive in the virtual address, the
 227          * old method of shifting the VA >> by PGDIR_SHIFT doesn't work.
 228          */
 229
 230         start_index = pgd_index(htlbpage_to_page(first));
 231         end_index = pgd_index(htlbpage_to_page(last));
 232
 233         if (end_index > start_index) {
 234                 clear_page_tables(tlb, start_index, end_index - start_index);
 235         }
 236 }
 237
 238 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
 239 {
 240         struct mm_struct *mm = vma->vm_mm;
 241         unsigned long address;
 242         pte_t *pte;
 243         struct page *page;
 244
 245         BUG_ON(start & (HPAGE_SIZE - 1));
 246         BUG_ON(end & (HPAGE_SIZE - 1));
 247
 248         for (address = start; address < end; address += HPAGE_SIZE) {
 249                 pte = huge_pte_offset(mm, address);
 250                 if (pte_none(*pte))
 251                         continue;
 252                 page = pte_page(*pte);
 253                 put_page(page);
 254                 pte_clear(pte);
 255         }
 256         // mm->rss -= (end - start) >> PAGE_SHIFT;
 257         vx_rsspages_sub(mm, (end - start) >> PAGE_SHIFT);
 258         flush_tlb_range(vma, start, end);
 259 }
 260
 261 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 262 {
 263         struct mm_struct *mm = current->mm;
 264         unsigned long addr;
 265         int ret = 0;
 266
 267         BUG_ON(vma->vm_start & ~HPAGE_MASK);
 268         BUG_ON(vma->vm_end & ~HPAGE_MASK);
 269
 270         spin_lock(&mm->page_table_lock);
 271         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 272                 unsigned long idx;
 273                 pte_t *pte = huge_pte_alloc(mm, addr);
 274                 struct page *page;
 275
 276                 if (!pte) {
 277                         ret = -ENOMEM;
 278                         goto out;
 279                 }
 280                 if (!pte_none(*pte))
 281                         continue;
 282
 283                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 284                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 285                 page = find_get_page(mapping, idx);
 286                 if (!page) {
 287                         /* charge the fs quota first */
 288                         if (hugetlb_get_quota(mapping)) {
 289                                 ret = -ENOMEM;
 290                                 goto out;
 291                         }
 292                         page = alloc_huge_page();
 293                         if (!page) {
 294                                 hugetlb_put_quota(mapping);
 295                                 ret = -ENOMEM;
 296                                 goto out;
 297                         }
 298                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 299                         unlock_page(page);
 300                         if (ret) {
 301                                 hugetlb_put_quota(mapping);
 302                                 free_huge_page(page);
 303                                 goto out;
 304                         }
 305                 }
 306                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
 307         }
 308 out:
 309         spin_unlock(&mm->page_table_lock);
 310         return ret;
 311 }
 312
 313 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
 314                 unsigned long pgoff, unsigned long flags)
 315 {
 316         struct vm_area_struct *vmm;
 317
 318         if (len > RGN_MAP_LIMIT)
 319                 return -ENOMEM;
 320         if (len & ~HPAGE_MASK)
 321                 return -EINVAL;
 322         /* This code assumes that REGION_HPAGE != 0. */
 323         if ((REGION_NUMBER(addr) != REGION_HPAGE) || (addr & (HPAGE_SIZE - 1)))
 324                 addr = HPAGE_REGION_BASE;
 325         else
 326                 addr = ALIGN(addr, HPAGE_SIZE);
 327         for (vmm = find_vma(current->mm, addr); ; vmm = vmm->vm_next) {
 328                 /* At this point:  (!vmm || addr < vmm->vm_end). */
 329                 if (REGION_OFFSET(addr) + len > RGN_MAP_LIMIT)
 330                         return -ENOMEM;
 331                 if (!vmm || (addr + len) <= vmm->vm_start)
 332                         return addr;
 333                 addr = ALIGN(vmm->vm_end, HPAGE_SIZE);
 334         }
 335 }
 336
 337 static int __init hugetlb_setup_sz(char *str)
 338 {
 339         u64 tr_pages;
 340         unsigned long long size;
 341
 342         if (ia64_pal_vm_page_size(&tr_pages, NULL) != 0)
 343                 /*
 344                  * shouldn't happen, but just in case.
 345                  */
 346                 tr_pages = 0x15557000UL;
 347
 348         size = memparse(str, &str);
 349         if (*str || (size & (size-1)) || !(tr_pages & size) ||
 350                 size <= PAGE_SIZE ||
 351                 size >= (1UL << PAGE_SHIFT << MAX_ORDER)) {
 352                 printk(KERN_WARNING "Invalid huge page size specified\n");
 353                 return 1;
 354         }
 355
 356         hpage_shift = __ffs(size);
 357         /*
 358          * boot cpu already executed ia64_mmu_init, and has HPAGE_SHIFT_DEFAULT
 359          * override here with new page shift.
 360          */
 361         ia64_set_rr(HPAGE_REGION_BASE, hpage_shift << 2);
 362         return 1;
 363 }
 364 __setup("hugepagesz=", hugetlb_setup_sz);