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