arch/sparc64/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/config.h>
   8 #include <linux/init.h>
   9 #include <linux/module.h>
  10 #include <linux/fs.h>
  11 #include <linux/mm.h>
  12 #include <linux/hugetlb.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/smp_lock.h>
  15 #include <linux/slab.h>
  16 #include <linux/sysctl.h>
  17
  18 #include <asm/mman.h>
  19 #include <asm/pgalloc.h>
  20 #include <asm/tlb.h>
  21 #include <asm/tlbflush.h>
  22 #include <asm/cacheflush.h>
  23 #include <asm/mmu_context.h>
  24
  25 /* Slightly simplified from the non-hugepage variant because by
  26  * definition we don't have to worry about any page coloring stuff
  27  */
  28 #define VA_EXCLUDE_START (0x0000080000000000UL - (1UL << 32UL))
  29 #define VA_EXCLUDE_END   (0xfffff80000000000UL + (1UL << 32UL))
  30
  31 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *filp,
  32                                                         unsigned long addr,
  33                                                         unsigned long len,
  34                                                         unsigned long pgoff,
  35                                                         unsigned long flags)
  36 {
  37         struct mm_struct *mm = current->mm;
  38         struct vm_area_struct * vma;
  39         unsigned long task_size = TASK_SIZE;
  40         unsigned long start_addr;
  41
  42         if (test_thread_flag(TIF_32BIT))
  43                 task_size = STACK_TOP32;
  44         if (unlikely(len >= VA_EXCLUDE_START))
  45                 return -ENOMEM;
  46
  47         if (len > mm->cached_hole_size) {
  48                 start_addr = addr = mm->free_area_cache;
  49         } else {
  50                 start_addr = addr = TASK_UNMAPPED_BASE;
  51                 mm->cached_hole_size = 0;
  52         }
  53
  54         task_size -= len;
  55
  56 full_search:
  57         addr = ALIGN(addr, HPAGE_SIZE);
  58
  59         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
  60                 /* At this point:  (!vma || addr < vma->vm_end). */
  61                 if (addr < VA_EXCLUDE_START &&
  62                     (addr + len) >= VA_EXCLUDE_START) {
  63                         addr = VA_EXCLUDE_END;
  64                         vma = find_vma(mm, VA_EXCLUDE_END);
  65                 }
  66                 if (unlikely(task_size < addr)) {
  67                         if (start_addr != TASK_UNMAPPED_BASE) {
  68                                 start_addr = addr = TASK_UNMAPPED_BASE;
  69                                 mm->cached_hole_size = 0;
  70                                 goto full_search;
  71                         }
  72                         return -ENOMEM;
  73                 }
  74                 if (likely(!vma || addr + len <= vma->vm_start)) {
  75                         /*
  76                          * Remember the place where we stopped the search:
  77                          */
  78                         mm->free_area_cache = addr + len;
  79                         return addr;
  80                 }
  81                 if (addr + mm->cached_hole_size < vma->vm_start)
  82                         mm->cached_hole_size = vma->vm_start - addr;
  83
  84                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
  85         }
  86 }
  87
  88 static unsigned long
  89 hugetlb_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
  90                                   const unsigned long len,
  91                                   const unsigned long pgoff,
  92                                   const unsigned long flags)
  93 {
  94         struct vm_area_struct *vma;
  95         struct mm_struct *mm = current->mm;
  96         unsigned long addr = addr0;
  97
  98         /* This should only ever run for 32-bit processes.  */
  99         BUG_ON(!test_thread_flag(TIF_32BIT));
 100
 101         /* check if free_area_cache is useful for us */
 102         if (len <= mm->cached_hole_size) {
 103                 mm->cached_hole_size = 0;
 104                 mm->free_area_cache = mm->mmap_base;
 105         }
 106
 107         /* either no address requested or can't fit in requested address hole */
 108         addr = mm->free_area_cache & HPAGE_MASK;
 109
 110         /* make sure it can fit in the remaining address space */
 111         if (likely(addr > len)) {
 112                 vma = find_vma(mm, addr-len);
 113                 if (!vma || addr <= vma->vm_start) {
 114                         /* remember the address as a hint for next time */
 115                         return (mm->free_area_cache = addr-len);
 116                 }
 117         }
 118
 119         if (unlikely(mm->mmap_base < len))
 120                 goto bottomup;
 121
 122         addr = (mm->mmap_base-len) & HPAGE_MASK;
 123
 124         do {
 125                 /*
 126                  * Lookup failure means no vma is above this address,
 127                  * else if new region fits below vma->vm_start,
 128                  * return with success:
 129                  */
 130                 vma = find_vma(mm, addr);
 131                 if (likely(!vma || addr+len <= vma->vm_start)) {
 132                         /* remember the address as a hint for next time */
 133                         return (mm->free_area_cache = addr);
 134                 }
 135
 136                 /* remember the largest hole we saw so far */
 137                 if (addr + mm->cached_hole_size < vma->vm_start)
 138                         mm->cached_hole_size = vma->vm_start - addr;
 139
 140                 /* try just below the current vma->vm_start */
 141                 addr = (vma->vm_start-len) & HPAGE_MASK;
 142         } while (likely(len < vma->vm_start));
 143
 144 bottomup:
 145         /*
 146          * A failed mmap() very likely causes application failure,
 147          * so fall back to the bottom-up function here. This scenario
 148          * can happen with large stack limits and large mmap()
 149          * allocations.
 150          */
 151         mm->cached_hole_size = ~0UL;
 152         mm->free_area_cache = TASK_UNMAPPED_BASE;
 153         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
 154         /*
 155          * Restore the topdown base:
 156          */
 157         mm->free_area_cache = mm->mmap_base;
 158         mm->cached_hole_size = ~0UL;
 159
 160         return addr;
 161 }
 162
 163 unsigned long
 164 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 165                 unsigned long len, unsigned long pgoff, unsigned long flags)
 166 {
 167         struct mm_struct *mm = current->mm;
 168         struct vm_area_struct *vma;
 169         unsigned long task_size = TASK_SIZE;
 170
 171         if (test_thread_flag(TIF_32BIT))
 172                 task_size = STACK_TOP32;
 173
 174         if (len & ~HPAGE_MASK)
 175                 return -EINVAL;
 176         if (len > task_size)
 177                 return -ENOMEM;
 178
 179         if (addr) {
 180                 addr = ALIGN(addr, HPAGE_SIZE);
 181                 vma = find_vma(mm, addr);
 182                 if (task_size - len >= addr &&
 183                     (!vma || addr + len <= vma->vm_start))
 184                         return addr;
 185         }
 186         if (mm->get_unmapped_area == arch_get_unmapped_area)
 187                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 188                                 pgoff, flags);
 189         else
 190                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 191                                 pgoff, flags);
 192 }
 193
 194 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
 195 {
 196         pgd_t *pgd;
 197         pud_t *pud;
 198         pmd_t *pmd;
 199         pte_t *pte = NULL;
 200
 201         /* We must align the address, because our caller will run
 202          * set_huge_pte_at() on whatever we return, which writes out
 203          * all of the sub-ptes for the hugepage range.  So we have
 204          * to give it the first such sub-pte.
 205          */
 206         addr &= HPAGE_MASK;
 207
 208         pgd = pgd_offset(mm, addr);
 209         pud = pud_alloc(mm, pgd, addr);
 210         if (pud) {
 211                 pmd = pmd_alloc(mm, pud, addr);
 212                 if (pmd)
 213                         pte = pte_alloc_map(mm, pmd, addr);
 214         }
 215         return pte;
 216 }
 217
 218 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 219 {
 220         pgd_t *pgd;
 221         pud_t *pud;
 222         pmd_t *pmd;
 223         pte_t *pte = NULL;
 224
 225         addr &= HPAGE_MASK;
 226
 227         pgd = pgd_offset(mm, addr);
 228         if (!pgd_none(*pgd)) {
 229                 pud = pud_offset(pgd, addr);
 230                 if (!pud_none(*pud)) {
 231                         pmd = pmd_offset(pud, addr);
 232                         if (!pmd_none(*pmd))
 233                                 pte = pte_offset_map(pmd, addr);
 234                 }
 235         }
 236         return pte;
 237 }
 238
 239 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
 240                      pte_t *ptep, pte_t entry)
 241 {
 242         int i;
 243
 244         if (!pte_present(*ptep) && pte_present(entry))
 245                 mm->context.huge_pte_count++;
 246
 247         for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 248                 set_pte_at(mm, addr, ptep, entry);
 249                 ptep++;
 250                 addr += PAGE_SIZE;
 251                 pte_val(entry) += PAGE_SIZE;
 252         }
 253 }
 254
 255 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
 256                               pte_t *ptep)
 257 {
 258         pte_t entry;
 259         int i;
 260
 261         entry = *ptep;
 262         if (pte_present(entry))
 263                 mm->context.huge_pte_count--;
 264
 265         for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 266                 pte_clear(mm, addr, ptep);
 267                 addr += PAGE_SIZE;
 268                 ptep++;
 269         }
 270
 271         return entry;
 272 }
 273
 274 struct page *follow_huge_addr(struct mm_struct *mm,
 275                               unsigned long address, int write)
 276 {
 277         return ERR_PTR(-EINVAL);
 278 }
 279
 280 int pmd_huge(pmd_t pmd)
 281 {
 282         return 0;
 283 }
 284
 285 struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 286                              pmd_t *pmd, int write)
 287 {
 288         return NULL;
 289 }
 290
 291 static void context_reload(void *__data)
 292 {
 293         struct mm_struct *mm = __data;
 294
 295         if (mm == current->mm)
 296                 load_secondary_context(mm);
 297 }
 298
 299 void hugetlb_prefault_arch_hook(struct mm_struct *mm)
 300 {
 301         struct tsb_config *tp = &mm->context.tsb_block[MM_TSB_HUGE];
 302
 303         if (likely(tp->tsb != NULL))
 304                 return;
 305
 306         tsb_grow(mm, MM_TSB_HUGE, 0);
 307         tsb_context_switch(mm);
 308         smp_tsb_sync(mm);
 309
 310         /* On UltraSPARC-III+ and later, configure the second half of
 311          * the Data-TLB for huge pages.
 312          */
 313         if (tlb_type == cheetah_plus) {
 314                 unsigned long ctx;
 315
 316                 spin_lock(&ctx_alloc_lock);
 317                 ctx = mm->context.sparc64_ctx_val;
 318                 ctx &= ~CTX_PGSZ_MASK;
 319                 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
 320                 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
 321
 322                 if (ctx != mm->context.sparc64_ctx_val) {
 323                         /* When changing the page size fields, we
 324                          * must perform a context flush so that no
 325                          * stale entries match.  This flush must
 326                          * occur with the original context register
 327                          * settings.
 328                          */
 329                         do_flush_tlb_mm(mm);
 330
 331                         /* Reload the context register of all processors
 332                          * also executing in this address space.
 333                          */
 334                         mm->context.sparc64_ctx_val = ctx;
 335                         on_each_cpu(context_reload, mm, 0, 0);
 336                 }
 337                 spin_unlock(&ctx_alloc_lock);
 338         }
 339 }