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 static inline void get_huge_page_tail(struct page *page)
  16 {
  17         /*
  18          * __split_huge_page_refcount() cannot run
  19          * from under us.
  20          */
  21         VM_BUG_ON(page_mapcount(page) < 0);
  22         VM_BUG_ON(atomic_read(&page->_count) != 0);
  23         atomic_inc(&page->_mapcount);
  24 }
  25
  26 /*
  27  * The performance critical leaf functions are made noinline otherwise gcc
  28  * inlines everything into a single function which results in too much
  29  * register pressure.
  30  */
  31 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
  32                 unsigned long end, int write, struct page **pages, int *nr)
  33 {
  34         unsigned long mask, result;
  35         pte_t *ptep;
  36
  37         if (tlb_type == hypervisor) {
  38                 result = _PAGE_PRESENT_4V|_PAGE_P_4V;
  39                 if (write)
  40                         result |= _PAGE_WRITE_4V;
  41         } else {
  42                 result = _PAGE_PRESENT_4U|_PAGE_P_4U;
  43                 if (write)
  44                         result |= _PAGE_WRITE_4U;
  45         }
  46         mask = result | _PAGE_SPECIAL;
  47
  48         ptep = pte_offset_kernel(&pmd, addr);
  49         do {
  50                 struct page *page, *head;
  51                 pte_t pte = *ptep;
  52
  53                 if ((pte_val(pte) & mask) != result)
  54                         return 0;
  55                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
  56
  57                 /* The hugepage case is simplified on sparc64 because
  58                  * we encode the sub-page pfn offsets into the
  59                  * hugepage PTEs.  We could optimize this in the future
  60                  * use page_cache_add_speculative() for the hugepage case.
  61                  */
  62                 page = pte_page(pte);
  63                 head = compound_head(page);
  64                 if (!page_cache_get_speculative(head))
  65                         return 0;
  66                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
  67                         put_page(head);
  68                         return 0;
  69                 }
  70                 if (head != page)
  71                         get_huge_page_tail(page);
  72
  73                 pages[*nr] = page;
  74                 (*nr)++;
  75         } while (ptep++, addr += PAGE_SIZE, addr != end);
  76
  77         return 1;
  78 }
  79
  80 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
  81                 int write, struct page **pages, int *nr)
  82 {
  83         unsigned long next;
  84         pmd_t *pmdp;
  85
  86         pmdp = pmd_offset(&pud, addr);
  87         do {
  88                 pmd_t pmd = *pmdp;
  89
  90                 next = pmd_addr_end(addr, end);
  91                 if (pmd_none(pmd))
  92                         return 0;
  93                 if (!gup_pte_range(pmd, addr, next, write, pages, nr))
  94                         return 0;
  95         } while (pmdp++, addr = next, addr != end);
  96
  97         return 1;
  98 }
  99
 100 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
 101                 int write, struct page **pages, int *nr)
 102 {
 103         unsigned long next;
 104         pud_t *pudp;
 105
 106         pudp = pud_offset(&pgd, addr);
 107         do {
 108                 pud_t pud = *pudp;
 109
 110                 next = pud_addr_end(addr, end);
 111                 if (pud_none(pud))
 112                         return 0;
 113                 if (!gup_pmd_range(pud, addr, next, write, pages, nr))
 114                         return 0;
 115         } while (pudp++, addr = next, addr != end);
 116
 117         return 1;
 118 }
 119
 120 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 121                         struct page **pages)
 122 {
 123         struct mm_struct *mm = current->mm;
 124         unsigned long addr, len, end;
 125         unsigned long next;
 126         pgd_t *pgdp;
 127         int nr = 0;
 128
 129         start &= PAGE_MASK;
 130         addr = start;
 131         len = (unsigned long) nr_pages << PAGE_SHIFT;
 132         end = start + len;
 133
 134         /*
 135          * XXX: batch / limit 'nr', to avoid large irq off latency
 136          * needs some instrumenting to determine the common sizes used by
 137          * important workloads (eg. DB2), and whether limiting the batch size
 138          * will decrease performance.
 139          *
 140          * It seems like we're in the clear for the moment. Direct-IO is
 141          * the main guy that batches up lots of get_user_pages, and even
 142          * they are limited to 64-at-a-time which is not so many.
 143          */
 144         /*
 145          * This doesn't prevent pagetable teardown, but does prevent
 146          * the pagetables from being freed on sparc.
 147          *
 148          * So long as we atomically load page table pointers versus teardown,
 149          * we can follow the address down to the the page and take a ref on it.
 150          */
 151         local_irq_disable();
 152
 153         pgdp = pgd_offset(mm, addr);
 154         do {
 155                 pgd_t pgd = *pgdp;
 156
 157                 next = pgd_addr_end(addr, end);
 158                 if (pgd_none(pgd))
 159                         goto slow;
 160                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 161                         goto slow;
 162         } while (pgdp++, addr = next, addr != end);
 163
 164         local_irq_enable();
 165
 166         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 167         return nr;
 168
 169         {
 170                 int ret;
 171
 172 slow:
 173                 local_irq_enable();
 174
 175                 /* Try to get the remaining pages with get_user_pages */
 176                 start += nr << PAGE_SHIFT;
 177                 pages += nr;
 178
 179                 down_read(&mm->mmap_sem);
 180                 ret = get_user_pages(current, mm, start,
 181                         (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
 182                 up_read(&mm->mmap_sem);
 183
 184                 /* Have to be a bit careful with return values */
 185                 if (nr > 0) {
 186                         if (ret < 0)
 187                                 ret = nr;
 188                         else
 189                                 ret += nr;
 190                 }
 191
 192                 return ret;
 193         }
 194 }