arch/powerpc/mm/gup.c

   1 /*
   2  * Lockless get_user_pages_fast for powerpc
   3  *
   4  * Copyright (C) 2008 Nick Piggin
   5  * Copyright (C) 2008 Novell Inc.
   6  */
   7 #undef DEBUG
   8
   9 #include <linux/sched.h>
  10 #include <linux/mm.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/vmstat.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/rwsem.h>
  15 #include <asm/pgtable.h>
  16
  17 #ifdef __HAVE_ARCH_PTE_SPECIAL
  18
  19 /*
  20  * The performance critical leaf functions are made noinline otherwise gcc
  21  * inlines everything into a single function which results in too much
  22  * register pressure.
  23  */
  24 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
  25                 unsigned long end, int write, struct page **pages, int *nr)
  26 {
  27         unsigned long mask, result;
  28         pte_t *ptep;
  29
  30         result = _PAGE_PRESENT|_PAGE_USER;
  31         if (write)
  32                 result |= _PAGE_RW;
  33         mask = result | _PAGE_SPECIAL;
  34
  35         ptep = pte_offset_kernel(&pmd, addr);
  36         do {
  37                 pte_t pte = *ptep;
  38                 struct page *page;
  39
  40                 if ((pte_val(pte) & mask) != result)
  41                         return 0;
  42                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
  43                 page = pte_page(pte);
  44                 if (!page_cache_get_speculative(page))
  45                         return 0;
  46                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
  47                         put_page(page);
  48                         return 0;
  49                 }
  50                 pages[*nr] = page;
  51                 (*nr)++;
  52
  53         } while (ptep++, addr += PAGE_SIZE, addr != end);
  54
  55         return 1;
  56 }
  57
  58 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
  59                 int write, struct page **pages, int *nr)
  60 {
  61         unsigned long next;
  62         pmd_t *pmdp;
  63
  64         pmdp = pmd_offset(&pud, addr);
  65         do {
  66                 pmd_t pmd = *pmdp;
  67
  68                 next = pmd_addr_end(addr, end);
  69                 if (pmd_none(pmd))
  70                         return 0;
  71                 if (pmd_huge(pmd)) {
  72                         if (!gup_hugepte((pte_t *)pmdp, PMD_SIZE, addr, next,
  73                                          write, pages, nr))
  74                                 return 0;
  75                 } else if (is_hugepd(pmdp)) {
  76                         if (!gup_hugepd((hugepd_t *)pmdp, PMD_SHIFT,
  77                                         addr, next, write, pages, nr))
  78                                 return 0;
  79                 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
  80                         return 0;
  81         } while (pmdp++, addr = next, addr != end);
  82
  83         return 1;
  84 }
  85
  86 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
  87                 int write, struct page **pages, int *nr)
  88 {
  89         unsigned long next;
  90         pud_t *pudp;
  91
  92         pudp = pud_offset(&pgd, addr);
  93         do {
  94                 pud_t pud = *pudp;
  95
  96                 next = pud_addr_end(addr, end);
  97                 if (pud_none(pud))
  98                         return 0;
  99                 if (pud_huge(pud)) {
 100                         if (!gup_hugepte((pte_t *)pudp, PUD_SIZE, addr, next,
 101                                          write, pages, nr))
 102                                 return 0;
 103                 } else if (is_hugepd(pudp)) {
 104                         if (!gup_hugepd((hugepd_t *)pudp, PUD_SHIFT,
 105                                         addr, next, write, pages, nr))
 106                                 return 0;
 107                 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
 108                         return 0;
 109         } while (pudp++, addr = next, addr != end);
 110
 111         return 1;
 112 }
 113
 114 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 115                         struct page **pages)
 116 {
 117         struct mm_struct *mm = current->mm;
 118         unsigned long addr, len, end;
 119         unsigned long next;
 120         pgd_t *pgdp;
 121         int nr = 0;
 122
 123         pr_devel("%s(%lx,%x,%s)\n", __func__, start, nr_pages, write ? "write" : "read");
 124
 125         start &= PAGE_MASK;
 126         addr = start;
 127         len = (unsigned long) nr_pages << PAGE_SHIFT;
 128         end = start + len;
 129
 130         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
 131                                         start, len)))
 132                 goto slow_irqon;
 133
 134         pr_devel("  aligned: %lx .. %lx\n", start, end);
 135
 136         /*
 137          * XXX: batch / limit 'nr', to avoid large irq off latency
 138          * needs some instrumenting to determine the common sizes used by
 139          * important workloads (eg. DB2), and whether limiting the batch size
 140          * will decrease performance.
 141          *
 142          * It seems like we're in the clear for the moment. Direct-IO is
 143          * the main guy that batches up lots of get_user_pages, and even
 144          * they are limited to 64-at-a-time which is not so many.
 145          */
 146         /*
 147          * This doesn't prevent pagetable teardown, but does prevent
 148          * the pagetables from being freed on powerpc.
 149          *
 150          * So long as we atomically load page table pointers versus teardown,
 151          * we can follow the address down to the the page and take a ref on it.
 152          */
 153         local_irq_disable();
 154
 155         pgdp = pgd_offset(mm, addr);
 156         do {
 157                 pgd_t pgd = *pgdp;
 158
 159                 pr_devel("  %016lx: normal pgd %p\n", addr,
 160                          (void *)pgd_val(pgd));
 161                 next = pgd_addr_end(addr, end);
 162                 if (pgd_none(pgd))
 163                         goto slow;
 164                 if (pgd_huge(pgd)) {
 165                         if (!gup_hugepte((pte_t *)pgdp, PGDIR_SIZE, addr, next,
 166                                          write, pages, &nr))
 167                                 goto slow;
 168                 } else if (is_hugepd(pgdp)) {
 169                         if (!gup_hugepd((hugepd_t *)pgdp, PGDIR_SHIFT,
 170                                         addr, next, write, pages, &nr))
 171                                 goto slow;
 172                 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 173                         goto slow;
 174         } while (pgdp++, addr = next, addr != end);
 175
 176         local_irq_enable();
 177
 178         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 179         return nr;
 180
 181         {
 182                 int ret;
 183
 184 slow:
 185                 local_irq_enable();
 186 slow_irqon:
 187                 pr_devel("  slow path ! nr = %d\n", nr);
 188
 189                 /* Try to get the remaining pages with get_user_pages */
 190                 start += nr << PAGE_SHIFT;
 191                 pages += nr;
 192
 193                 down_read(&mm->mmap_sem);
 194                 ret = get_user_pages(current, mm, start,
 195                         (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
 196                 up_read(&mm->mmap_sem);
 197
 198                 /* Have to be a bit careful with return values */
 199                 if (nr > 0) {
 200                         if (ret < 0)
 201                                 ret = nr;
 202                         else
 203                                 ret += nr;
 204                 }
 205
 206                 return ret;
 207         }
 208 }
 209
 210 #endif /* __HAVE_ARCH_PTE_SPECIAL */