arch/sparc/mm/hugetlbpage.c

   1 /*
   2  * SPARC64 Huge TLB page support.
   3  *
   4  * Copyright (C) 2002, 2003, 2006 David S. Miller (davem@davemloft.net)
   5  */
   6
   7 #include <linux/init.h>
   8 #include <linux/module.h>
   9 #include <linux/fs.h>
  10 #include <linux/mm.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/sysctl.h>
  14
  15 #include <asm/mman.h>
  16 #include <asm/pgalloc.h>
  17 #include <asm/tlb.h>
  18 #include <asm/tlbflush.h>
  19 #include <asm/cacheflush.h>
  20 #include <asm/mmu_context.h>
  21
  22 /* Slightly simplified from the non-hugepage variant because by
  23  * definition we don't have to worry about any page coloring stuff
  24  */
  25 #define VA_EXCLUDE_START (0x0000080000000000UL - (1UL << 32UL))
  26 #define VA_EXCLUDE_END   (0xfffff80000000000UL + (1UL << 32UL))
  27
  28 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *filp,
  29                                                         unsigned long addr,
  30                                                         unsigned long len,
  31                                                         unsigned long pgoff,
  32                                                         unsigned long flags)
  33 {
  34         struct mm_struct *mm = current->mm;
  35         struct vm_area_struct * vma;
  36         unsigned long task_size = TASK_SIZE;
  37         unsigned long start_addr;
  38
  39         if (test_thread_flag(TIF_32BIT))
  40                 task_size = STACK_TOP32;
  41         if (unlikely(len >= VA_EXCLUDE_START))
  42                 return -ENOMEM;
  43
  44         if (len > mm->cached_hole_size) {
  45                 start_addr = addr = mm->free_area_cache;
  46         } else {
  47                 start_addr = addr = TASK_UNMAPPED_BASE;
  48                 mm->cached_hole_size = 0;
  49         }
  50
  51         task_size -= len;
  52
  53 full_search:
  54         addr = ALIGN(addr, HPAGE_SIZE);
  55
  56         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
  57                 /* At this point:  (!vma || addr < vma->vm_end). */
  58                 if (addr < VA_EXCLUDE_START &&
  59                     (addr + len) >= VA_EXCLUDE_START) {
  60                         addr = VA_EXCLUDE_END;
  61                         vma = find_vma(mm, VA_EXCLUDE_END);
  62                 }
  63                 if (unlikely(task_size < addr)) {
  64                         if (start_addr != TASK_UNMAPPED_BASE) {
  65                                 start_addr = addr = TASK_UNMAPPED_BASE;
  66                                 mm->cached_hole_size = 0;
  67                                 goto full_search;
  68                         }
  69                         return -ENOMEM;
  70                 }
  71                 if (likely(!vma || addr + len <= vma->vm_start)) {
  72                         /*
  73                          * Remember the place where we stopped the search:
  74                          */
  75                         mm->free_area_cache = addr + len;
  76                         return addr;
  77                 }
  78                 if (addr + mm->cached_hole_size < vma->vm_start)
  79                         mm->cached_hole_size = vma->vm_start - addr;
  80
  81                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
  82         }
  83 }
  84
  85 static unsigned long
  86 hugetlb_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
  87                                   const unsigned long len,
  88                                   const unsigned long pgoff,
  89                                   const unsigned long flags)
  90 {
  91         struct vm_area_struct *vma;
  92         struct mm_struct *mm = current->mm;
  93         unsigned long addr = addr0;
  94
  95         /* This should only ever run for 32-bit processes.  */
  96         BUG_ON(!test_thread_flag(TIF_32BIT));
  97
  98         /* check if free_area_cache is useful for us */
  99         if (len <= mm->cached_hole_size) {
 100                 mm->cached_hole_size = 0;
 101                 mm->free_area_cache = mm->mmap_base;
 102         }
 103
 104         /* either no address requested or can't fit in requested address hole */
 105         addr = mm->free_area_cache & HPAGE_MASK;
 106
 107         /* make sure it can fit in the remaining address space */
 108         if (likely(addr > len)) {
 109                 vma = find_vma(mm, addr-len);
 110                 if (!vma || addr <= vma->vm_start) {
 111                         /* remember the address as a hint for next time */
 112                         return (mm->free_area_cache = addr-len);
 113                 }
 114         }
 115
 116         if (unlikely(mm->mmap_base < len))
 117                 goto bottomup;
 118
 119         addr = (mm->mmap_base-len) & HPAGE_MASK;
 120
 121         do {
 122                 /*
 123                  * Lookup failure means no vma is above this address,
 124                  * else if new region fits below vma->vm_start,
 125                  * return with success:
 126                  */
 127                 vma = find_vma(mm, addr);
 128                 if (likely(!vma || addr+len <= vma->vm_start)) {
 129                         /* remember the address as a hint for next time */
 130                         return (mm->free_area_cache = addr);
 131                 }
 132
 133                 /* remember the largest hole we saw so far */
 134                 if (addr + mm->cached_hole_size < vma->vm_start)
 135                         mm->cached_hole_size = vma->vm_start - addr;
 136
 137                 /* try just below the current vma->vm_start */
 138                 addr = (vma->vm_start-len) & HPAGE_MASK;
 139         } while (likely(len < vma->vm_start));
 140
 141 bottomup:
 142         /*
 143          * A failed mmap() very likely causes application failure,
 144          * so fall back to the bottom-up function here. This scenario
 145          * can happen with large stack limits and large mmap()
 146          * allocations.
 147          */
 148         mm->cached_hole_size = ~0UL;
 149         mm->free_area_cache = TASK_UNMAPPED_BASE;
 150         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
 151         /*
 152          * Restore the topdown base:
 153          */
 154         mm->free_area_cache = mm->mmap_base;
 155         mm->cached_hole_size = ~0UL;
 156
 157         return addr;
 158 }
 159
 160 unsigned long
 161 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 162                 unsigned long len, unsigned long pgoff, unsigned long flags)
 163 {
 164         struct mm_struct *mm = current->mm;
 165         struct vm_area_struct *vma;
 166         unsigned long task_size = TASK_SIZE;
 167
 168         if (test_thread_flag(TIF_32BIT))
 169                 task_size = STACK_TOP32;
 170
 171         if (len & ~HPAGE_MASK)
 172                 return -EINVAL;
 173         if (len > task_size)
 174                 return -ENOMEM;
 175
 176         if (flags & MAP_FIXED) {
 177                 if (prepare_hugepage_range(file, addr, len))
 178                         return -EINVAL;
 179                 return addr;
 180         }
 181
 182         if (addr) {
 183                 addr = ALIGN(addr, HPAGE_SIZE);
 184                 vma = find_vma(mm, addr);
 185                 if (task_size - len >= addr &&
 186                     (!vma || addr + len <= vma->vm_start))
 187                         return addr;
 188         }
 189         if (mm->get_unmapped_area == arch_get_unmapped_area)
 190                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 191                                 pgoff, flags);
 192         else
 193                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 194                                 pgoff, flags);
 195 }
 196
 197 pte_t *huge_pte_alloc(struct mm_struct *mm,
 198                         unsigned long addr, unsigned long sz)
 199 {
 200         pgd_t *pgd;
 201         pud_t *pud;
 202         pmd_t *pmd;
 203         pte_t *pte = NULL;
 204
 205         /* We must align the address, because our caller will run
 206          * set_huge_pte_at() on whatever we return, which writes out
 207          * all of the sub-ptes for the hugepage range.  So we have
 208          * to give it the first such sub-pte.
 209          */
 210         addr &= HPAGE_MASK;
 211
 212         pgd = pgd_offset(mm, addr);
 213         pud = pud_alloc(mm, pgd, addr);
 214         if (pud) {
 215                 pmd = pmd_alloc(mm, pud, addr);
 216                 if (pmd)
 217                         pte = pte_alloc_map(mm, NULL, pmd, addr);
 218         }
 219         return pte;
 220 }
 221
 222 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 223 {
 224         pgd_t *pgd;
 225         pud_t *pud;
 226         pmd_t *pmd;
 227         pte_t *pte = NULL;
 228
 229         addr &= HPAGE_MASK;
 230
 231         pgd = pgd_offset(mm, addr);
 232         if (!pgd_none(*pgd)) {
 233                 pud = pud_offset(pgd, addr);
 234                 if (!pud_none(*pud)) {
 235                         pmd = pmd_offset(pud, addr);
 236                         if (!pmd_none(*pmd))
 237                                 pte = pte_offset_map(pmd, addr);
 238                 }
 239         }
 240         return pte;
 241 }
 242
 243 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
 244 {
 245         return 0;
 246 }
 247
 248 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
 249                      pte_t *ptep, pte_t entry)
 250 {
 251         int i;
 252
 253         if (!pte_present(*ptep) && pte_present(entry))
 254                 mm->context.huge_pte_count++;
 255
 256         addr &= HPAGE_MASK;
 257         for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 258                 set_pte_at(mm, addr, ptep, entry);
 259                 ptep++;
 260                 addr += PAGE_SIZE;
 261                 pte_val(entry) += PAGE_SIZE;
 262         }
 263 }
 264
 265 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
 266                               pte_t *ptep)
 267 {
 268         pte_t entry;
 269         int i;
 270
 271         entry = *ptep;
 272         if (pte_present(entry))
 273                 mm->context.huge_pte_count--;
 274
 275         addr &= HPAGE_MASK;
 276
 277         for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 278                 pte_clear(mm, addr, ptep);
 279                 addr += PAGE_SIZE;
 280                 ptep++;
 281         }
 282
 283         return entry;
 284 }
 285
 286 struct page *follow_huge_addr(struct mm_struct *mm,
 287                               unsigned long address, int write)
 288 {
 289         return ERR_PTR(-EINVAL);
 290 }
 291
 292 int pmd_huge(pmd_t pmd)
 293 {
 294         return 0;
 295 }
 296
 297 int pud_huge(pud_t pud)
 298 {
 299         return 0;
 300 }
 301
 302 struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 303                              pmd_t *pmd, int write)
 304 {
 305         return NULL;
 306 }
 307
 308 static void context_reload(void *__data)
 309 {
 310         struct mm_struct *mm = __data;
 311
 312         if (mm == current->mm)
 313                 load_secondary_context(mm);
 314 }
 315
 316 void hugetlb_prefault_arch_hook(struct mm_struct *mm)
 317 {
 318         struct tsb_config *tp = &mm->context.tsb_block[MM_TSB_HUGE];
 319
 320         if (likely(tp->tsb != NULL))
 321                 return;
 322
 323         tsb_grow(mm, MM_TSB_HUGE, 0);
 324         tsb_context_switch(mm);
 325         smp_tsb_sync(mm);
 326
 327         /* On UltraSPARC-III+ and later, configure the second half of
 328          * the Data-TLB for huge pages.
 329          */
 330         if (tlb_type == cheetah_plus) {
 331                 unsigned long ctx;
 332
 333                 spin_lock(&ctx_alloc_lock);
 334                 ctx = mm->context.sparc64_ctx_val;
 335                 ctx &= ~CTX_PGSZ_MASK;
 336                 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
 337                 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
 338
 339                 if (ctx != mm->context.sparc64_ctx_val) {
 340                         /* When changing the page size fields, we
 341                          * must perform a context flush so that no
 342                          * stale entries match.  This flush must
 343                          * occur with the original context register
 344                          * settings.
 345                          */
 346                         do_flush_tlb_mm(mm);
 347
 348                         /* Reload the context register of all processors
 349                          * also executing in this address space.
 350                          */
 351                         mm->context.sparc64_ctx_val = ctx;
 352                         on_each_cpu(context_reload, mm, 0);
 353                 }
 354                 spin_unlock(&ctx_alloc_lock);
 355         }
 356 }