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