arch/sparc/mm/gup.c

   1 /*
   2  * Lockless get_user_pages_fast for sparc, cribbed from powerpc
   3  *
   4  * Copyright (C) 2008 Nick Piggin
   5  * Copyright (C) 2008 Novell Inc.
   6  */
   7
   8 #include <linux/sched.h>
   9 #include <linux/mm.h>
  10 #include <linux/vmstat.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/rwsem.h>
  13 #include <asm/pgtable.h>
  14
  15 /*
  16  * The performance critical leaf functions are made noinline otherwise gcc
  17  * inlines everything into a single function which results in too much
  18  * register pressure.
  19  */
  20 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
  21                 unsigned long end, int write, struct page **pages, int *nr)
  22 {
  23         unsigned long mask, result;
  24         pte_t *ptep;
  25
  26         if (tlb_type == hypervisor) {
  27                 result = _PAGE_PRESENT_4V|_PAGE_P_4V;
  28                 if (write)
  29                         result |= _PAGE_WRITE_4V;
  30         } else {
  31                 result = _PAGE_PRESENT_4U|_PAGE_P_4U;
  32                 if (write)
  33                         result |= _PAGE_WRITE_4U;
  34         }
  35         mask = result | _PAGE_SPECIAL;
  36
  37         ptep = pte_offset_kernel(&pmd, addr);
  38         do {
  39                 struct page *page, *head;
  40                 pte_t pte = *ptep;
  41
  42                 if ((pte_val(pte) & mask) != result)
  43                         return 0;
  44                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
  45
  46                 /* The hugepage case is simplified on sparc64 because
  47                  * we encode the sub-page pfn offsets into the
  48                  * hugepage PTEs.  We could optimize this in the future
  49                  * use page_cache_add_speculative() for the hugepage case.
  50                  */
  51                 page = pte_page(pte);
  52                 head = compound_head(page);
  53                 if (!page_cache_get_speculative(head))
  54                         return 0;
  55                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
  56                         put_page(head);
  57                         return 0;
  58                 }
  59
  60                 pages[*nr] = page;
  61                 (*nr)++;
  62         } while (ptep++, addr += PAGE_SIZE, addr != end);
  63
  64         return 1;
  65 }
  66
  67 static int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
  68                         unsigned long end, int write, struct page **pages,
  69                         int *nr)
  70 {
  71         struct page *head, *page;
  72         int refs;
  73
  74         if (!(pmd_val(pmd) & _PAGE_VALID))
  75                 return 0;
  76
  77         if (write && !pmd_write(pmd))
  78                 return 0;
  79
  80         refs = 0;
  81         page = pmd_page(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
  82         head = compound_head(page);
  83         do {
  84                 VM_BUG_ON(compound_head(page) != head);
  85                 pages[*nr] = page;
  86                 (*nr)++;
  87                 page++;
  88                 refs++;
  89         } while (addr += PAGE_SIZE, addr != end);
  90
  91         if (!page_cache_add_speculative(head, refs)) {
  92                 *nr -= refs;
  93                 return 0;
  94         }
  95
  96         if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
  97                 *nr -= refs;
  98                 while (refs--)
  99                         put_page(head);
 100                 return 0;
 101         }
 102
 103         return 1;
 104 }
 105
 106 static int gup_huge_pud(pud_t *pudp, pud_t pud, unsigned long addr,
 107                         unsigned long end, int write, struct page **pages,
 108                         int *nr)
 109 {
 110         struct page *head, *page;
 111         int refs;
 112
 113         if (!(pud_val(pud) & _PAGE_VALID))
 114                 return 0;
 115
 116         if (write && !pud_write(pud))
 117                 return 0;
 118
 119         refs = 0;
 120         page = pud_page(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 121         head = compound_head(page);
 122         do {
 123                 VM_BUG_ON(compound_head(page) != head);
 124                 pages[*nr] = page;
 125                 (*nr)++;
 126                 page++;
 127                 refs++;
 128         } while (addr += PAGE_SIZE, addr != end);
 129
 130         if (!page_cache_add_speculative(head, refs)) {
 131                 *nr -= refs;
 132                 return 0;
 133         }
 134
 135         if (unlikely(pud_val(pud) != pud_val(*pudp))) {
 136                 *nr -= refs;
 137                 while (refs--)
 138                         put_page(head);
 139                 return 0;
 140         }
 141
 142         return 1;
 143 }
 144
 145 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
 146                 int write, struct page **pages, int *nr)
 147 {
 148         unsigned long next;
 149         pmd_t *pmdp;
 150
 151         pmdp = pmd_offset(&pud, addr);
 152         do {
 153                 pmd_t pmd = *pmdp;
 154
 155                 next = pmd_addr_end(addr, end);
 156                 if (pmd_none(pmd))
 157                         return 0;
 158                 if (unlikely(pmd_large(pmd))) {
 159                         if (!gup_huge_pmd(pmdp, pmd, addr, next,
 160                                           write, pages, nr))
 161                                 return 0;
 162                 } else if (!gup_pte_range(pmd, addr, next, write,
 163                                           pages, nr))
 164                         return 0;
 165         } while (pmdp++, addr = next, addr != end);
 166
 167         return 1;
 168 }
 169
 170 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
 171                 int write, struct page **pages, int *nr)
 172 {
 173         unsigned long next;
 174         pud_t *pudp;
 175
 176         pudp = pud_offset(&pgd, addr);
 177         do {
 178                 pud_t pud = *pudp;
 179
 180                 next = pud_addr_end(addr, end);
 181                 if (pud_none(pud))
 182                         return 0;
 183                 if (unlikely(pud_large(pud))) {
 184                         if (!gup_huge_pud(pudp, pud, addr, next,
 185                                           write, pages, nr))
 186                                 return 0;
 187                 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
 188                         return 0;
 189         } while (pudp++, addr = next, addr != end);
 190
 191         return 1;
 192 }
 193
 194 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
 195                           struct page **pages)
 196 {
 197         struct mm_struct *mm = current->mm;
 198         unsigned long addr, len, end;
 199         unsigned long next, flags;
 200         pgd_t *pgdp;
 201         int nr = 0;
 202
 203         start &= PAGE_MASK;
 204         addr = start;
 205         len = (unsigned long) nr_pages << PAGE_SHIFT;
 206         end = start + len;
 207
 208         local_irq_save(flags);
 209         pgdp = pgd_offset(mm, addr);
 210         do {
 211                 pgd_t pgd = *pgdp;
 212
 213                 next = pgd_addr_end(addr, end);
 214                 if (pgd_none(pgd))
 215                         break;
 216                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 217                         break;
 218         } while (pgdp++, addr = next, addr != end);
 219         local_irq_restore(flags);
 220
 221         return nr;
 222 }
 223
 224 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 225                         struct page **pages)
 226 {
 227         struct mm_struct *mm = current->mm;
 228         unsigned long addr, len, end;
 229         unsigned long next;
 230         pgd_t *pgdp;
 231         int nr = 0;
 232
 233         start &= PAGE_MASK;
 234         addr = start;
 235         len = (unsigned long) nr_pages << PAGE_SHIFT;
 236         end = start + len;
 237
 238         /*
 239          * XXX: batch / limit 'nr', to avoid large irq off latency
 240          * needs some instrumenting to determine the common sizes used by
 241          * important workloads (eg. DB2), and whether limiting the batch size
 242          * will decrease performance.
 243          *
 244          * It seems like we're in the clear for the moment. Direct-IO is
 245          * the main guy that batches up lots of get_user_pages, and even
 246          * they are limited to 64-at-a-time which is not so many.
 247          */
 248         /*
 249          * This doesn't prevent pagetable teardown, but does prevent
 250          * the pagetables from being freed on sparc.
 251          *
 252          * So long as we atomically load page table pointers versus teardown,
 253          * we can follow the address down to the the page and take a ref on it.
 254          */
 255         local_irq_disable();
 256
 257         pgdp = pgd_offset(mm, addr);
 258         do {
 259                 pgd_t pgd = *pgdp;
 260
 261                 next = pgd_addr_end(addr, end);
 262                 if (pgd_none(pgd))
 263                         goto slow;
 264                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 265                         goto slow;
 266         } while (pgdp++, addr = next, addr != end);
 267
 268         local_irq_enable();
 269
 270         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 271         return nr;
 272
 273         {
 274                 int ret;
 275
 276 slow:
 277                 local_irq_enable();
 278
 279                 /* Try to get the remaining pages with get_user_pages */
 280                 start += nr << PAGE_SHIFT;
 281                 pages += nr;
 282
 283                 ret = get_user_pages_unlocked(start,
 284                         (end - start) >> PAGE_SHIFT, pages,
 285                         write ? FOLL_WRITE : 0);
 286
 287                 /* Have to be a bit careful with return values */
 288                 if (nr > 0) {
 289                         if (ret < 0)
 290                                 ret = nr;
 291                         else
 292                                 ret += nr;
 293                 }
 294
 295                 return ret;
 296         }
 297 }