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                 if (head != page)
  60                         get_huge_page_tail(page);
  61
  62                 pages[*nr] = page;
  63                 (*nr)++;
  64         } while (ptep++, addr += PAGE_SIZE, addr != end);
  65
  66         return 1;
  67 }
  68
  69 static int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
  70                         unsigned long end, int write, struct page **pages,
  71                         int *nr)
  72 {
  73         struct page *head, *page, *tail;
  74         u32 mask;
  75         int refs;
  76
  77         mask = PMD_HUGE_PRESENT;
  78         if (write)
  79                 mask |= PMD_HUGE_WRITE;
  80         if ((pmd_val(pmd) & mask) != mask)
  81                 return 0;
  82
  83         refs = 0;
  84         head = pmd_page(pmd);
  85         page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
  86         tail = page;
  87         do {
  88                 VM_BUG_ON(compound_head(page) != head);
  89                 pages[*nr] = page;
  90                 (*nr)++;
  91                 page++;
  92                 refs++;
  93         } while (addr += PAGE_SIZE, addr != end);
  94
  95         if (!page_cache_add_speculative(head, refs)) {
  96                 *nr -= refs;
  97                 return 0;
  98         }
  99
 100         if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
 101                 *nr -= refs;
 102                 while (refs--)
 103                         put_page(head);
 104                 return 0;
 105         }
 106
 107         /* Any tail page need their mapcount reference taken before we
 108          * return.
 109          */
 110         while (refs--) {
 111                 if (PageTail(tail))
 112                         get_huge_page_tail(tail);
 113                 tail++;
 114         }
 115
 116         return 1;
 117 }
 118
 119 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
 120                 int write, struct page **pages, int *nr)
 121 {
 122         unsigned long next;
 123         pmd_t *pmdp;
 124
 125         pmdp = pmd_offset(&pud, addr);
 126         do {
 127                 pmd_t pmd = *pmdp;
 128
 129                 next = pmd_addr_end(addr, end);
 130                 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
 131                         return 0;
 132                 if (unlikely(pmd_large(pmd))) {
 133                         if (!gup_huge_pmd(pmdp, pmd, addr, next,
 134                                           write, pages, nr))
 135                                 return 0;
 136                 } else if (!gup_pte_range(pmd, addr, next, write,
 137                                           pages, nr))
 138                         return 0;
 139         } while (pmdp++, addr = next, addr != end);
 140
 141         return 1;
 142 }
 143
 144 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
 145                 int write, struct page **pages, int *nr)
 146 {
 147         unsigned long next;
 148         pud_t *pudp;
 149
 150         pudp = pud_offset(&pgd, addr);
 151         do {
 152                 pud_t pud = *pudp;
 153
 154                 next = pud_addr_end(addr, end);
 155                 if (pud_none(pud))
 156                         return 0;
 157                 if (!gup_pmd_range(pud, addr, next, write, pages, nr))
 158                         return 0;
 159         } while (pudp++, addr = next, addr != end);
 160
 161         return 1;
 162 }
 163
 164 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 165                         struct page **pages)
 166 {
 167         struct mm_struct *mm = current->mm;
 168         unsigned long addr, len, end;
 169         unsigned long next;
 170         pgd_t *pgdp;
 171         int nr = 0;
 172
 173         start &= PAGE_MASK;
 174         addr = start;
 175         len = (unsigned long) nr_pages << PAGE_SHIFT;
 176         end = start + len;
 177
 178         /*
 179          * XXX: batch / limit 'nr', to avoid large irq off latency
 180          * needs some instrumenting to determine the common sizes used by
 181          * important workloads (eg. DB2), and whether limiting the batch size
 182          * will decrease performance.
 183          *
 184          * It seems like we're in the clear for the moment. Direct-IO is
 185          * the main guy that batches up lots of get_user_pages, and even
 186          * they are limited to 64-at-a-time which is not so many.
 187          */
 188         /*
 189          * This doesn't prevent pagetable teardown, but does prevent
 190          * the pagetables from being freed on sparc.
 191          *
 192          * So long as we atomically load page table pointers versus teardown,
 193          * we can follow the address down to the the page and take a ref on it.
 194          */
 195         local_irq_disable();
 196
 197         pgdp = pgd_offset(mm, addr);
 198         do {
 199                 pgd_t pgd = *pgdp;
 200
 201                 next = pgd_addr_end(addr, end);
 202                 if (pgd_none(pgd))
 203                         goto slow;
 204                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 205                         goto slow;
 206         } while (pgdp++, addr = next, addr != end);
 207
 208         local_irq_enable();
 209
 210         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 211         return nr;
 212
 213         {
 214                 int ret;
 215
 216 slow:
 217                 local_irq_enable();
 218
 219                 /* Try to get the remaining pages with get_user_pages */
 220                 start += nr << PAGE_SHIFT;
 221                 pages += nr;
 222
 223                 down_read(&mm->mmap_sem);
 224                 ret = get_user_pages(current, mm, start,
 225                         (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
 226                 up_read(&mm->mmap_sem);
 227
 228                 /* Have to be a bit careful with return values */
 229                 if (nr > 0) {
 230                         if (ret < 0)
 231                                 ret = nr;
 232                         else
 233                                 ret += nr;
 234                 }
 235
 236                 return ret;
 237         }
 238 }