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                 /*
  70                  * If we find a splitting transparent hugepage we
  71                  * return zero. That will result in taking the slow
  72                  * path which will call wait_split_huge_page()
  73                  * if the pmd is still in splitting state
  74                  */
  75                 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
  76                         return 0;
  77                 if (pmd_huge(pmd) || pmd_large(pmd)) {
  78                         if (!gup_hugepte((pte_t *)pmdp, PMD_SIZE, addr, next,
  79                                          write, pages, nr))
  80                                 return 0;
  81                 } else if (is_hugepd(pmdp)) {
  82                         if (!gup_hugepd((hugepd_t *)pmdp, PMD_SHIFT,
  83                                         addr, next, write, pages, nr))
  84                                 return 0;
  85                 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
  86                         return 0;
  87         } while (pmdp++, addr = next, addr != end);
  88
  89         return 1;
  90 }
  91
  92 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
  93                 int write, struct page **pages, int *nr)
  94 {
  95         unsigned long next;
  96         pud_t *pudp;
  97
  98         pudp = pud_offset(&pgd, addr);
  99         do {
 100                 pud_t pud = *pudp;
 101
 102                 next = pud_addr_end(addr, end);
 103                 if (pud_none(pud))
 104                         return 0;
 105                 if (pud_huge(pud)) {
 106                         if (!gup_hugepte((pte_t *)pudp, PUD_SIZE, addr, next,
 107                                          write, pages, nr))
 108                                 return 0;
 109                 } else if (is_hugepd(pudp)) {
 110                         if (!gup_hugepd((hugepd_t *)pudp, PUD_SHIFT,
 111                                         addr, next, write, pages, nr))
 112                                 return 0;
 113                 } else 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         pr_devel("%s(%lx,%x,%s)\n", __func__, start, nr_pages, write ? "write" : "read");
 130
 131         start &= PAGE_MASK;
 132         addr = start;
 133         len = (unsigned long) nr_pages << PAGE_SHIFT;
 134         end = start + len;
 135
 136         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
 137                                         start, len)))
 138                 goto slow_irqon;
 139
 140         pr_devel("  aligned: %lx .. %lx\n", start, end);
 141
 142         /*
 143          * XXX: batch / limit 'nr', to avoid large irq off latency
 144          * needs some instrumenting to determine the common sizes used by
 145          * important workloads (eg. DB2), and whether limiting the batch size
 146          * will decrease performance.
 147          *
 148          * It seems like we're in the clear for the moment. Direct-IO is
 149          * the main guy that batches up lots of get_user_pages, and even
 150          * they are limited to 64-at-a-time which is not so many.
 151          */
 152         /*
 153          * This doesn't prevent pagetable teardown, but does prevent
 154          * the pagetables from being freed on powerpc.
 155          *
 156          * So long as we atomically load page table pointers versus teardown,
 157          * we can follow the address down to the the page and take a ref on it.
 158          */
 159         local_irq_disable();
 160
 161         pgdp = pgd_offset(mm, addr);
 162         do {
 163                 pgd_t pgd = *pgdp;
 164
 165                 pr_devel("  %016lx: normal pgd %p\n", addr,
 166                          (void *)pgd_val(pgd));
 167                 next = pgd_addr_end(addr, end);
 168                 if (pgd_none(pgd))
 169                         goto slow;
 170                 if (pgd_huge(pgd)) {
 171                         if (!gup_hugepte((pte_t *)pgdp, PGDIR_SIZE, addr, next,
 172                                          write, pages, &nr))
 173                                 goto slow;
 174                 } else if (is_hugepd(pgdp)) {
 175                         if (!gup_hugepd((hugepd_t *)pgdp, PGDIR_SHIFT,
 176                                         addr, next, write, pages, &nr))
 177                                 goto slow;
 178                 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 179                         goto slow;
 180         } while (pgdp++, addr = next, addr != end);
 181
 182         local_irq_enable();
 183
 184         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 185         return nr;
 186
 187         {
 188                 int ret;
 189
 190 slow:
 191                 local_irq_enable();
 192 slow_irqon:
 193                 pr_devel("  slow path ! nr = %d\n", nr);
 194
 195                 /* Try to get the remaining pages with get_user_pages */
 196                 start += nr << PAGE_SHIFT;
 197                 pages += nr;
 198
 199                 down_read(&mm->mmap_sem);
 200                 ret = get_user_pages(current, mm, start,
 201                         (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
 202                 up_read(&mm->mmap_sem);
 203
 204                 /* Have to be a bit careful with return values */
 205                 if (nr > 0) {
 206                         if (ret < 0)
 207                                 ret = nr;
 208                         else
 209                                 ret += nr;
 210                 }
 211
 212                 return ret;
 213         }
 214 }
 215
 216 #endif /* __HAVE_ARCH_PTE_SPECIAL */