arch/s390/mm/gup.c

   1 /*
   2  *  Lockless get_user_pages_fast for s390
   3  *
   4  *  Copyright IBM Corp. 2010
   5  *  Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
   6  */
   7 #include <linux/sched.h>
   8 #include <linux/mm.h>
   9 #include <linux/hugetlb.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 inline int gup_pte_range(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
  21                 unsigned long end, int write, struct page **pages, int *nr)
  22 {
  23         unsigned long mask;
  24         pte_t *ptep, pte;
  25         struct page *page;
  26
  27         mask = (write ? _PAGE_RO : 0) | _PAGE_INVALID | _PAGE_SPECIAL;
  28
  29         ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr);
  30         do {
  31                 pte = *ptep;
  32                 barrier();
  33                 if ((pte_val(pte) & mask) != 0)
  34                         return 0;
  35                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
  36                 page = pte_page(pte);
  37                 if (!page_cache_get_speculative(page))
  38                         return 0;
  39                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
  40                         put_page(page);
  41                         return 0;
  42                 }
  43                 pages[*nr] = page;
  44                 (*nr)++;
  45
  46         } while (ptep++, addr += PAGE_SIZE, addr != end);
  47
  48         return 1;
  49 }
  50
  51 static inline int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
  52                 unsigned long end, int write, struct page **pages, int *nr)
  53 {
  54         unsigned long mask, result;
  55         struct page *head, *page, *tail;
  56         int refs;
  57
  58         result = write ? 0 : _SEGMENT_ENTRY_RO;
  59         mask = result | _SEGMENT_ENTRY_INV;
  60         if ((pmd_val(pmd) & mask) != result)
  61                 return 0;
  62         VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT));
  63
  64         refs = 0;
  65         head = pmd_page(pmd);
  66         page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
  67         tail = page;
  68         do {
  69                 VM_BUG_ON(compound_head(page) != head);
  70                 pages[*nr] = page;
  71                 (*nr)++;
  72                 page++;
  73                 refs++;
  74         } while (addr += PAGE_SIZE, addr != end);
  75
  76         if (!page_cache_add_speculative(head, refs)) {
  77                 *nr -= refs;
  78                 return 0;
  79         }
  80
  81         if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
  82                 *nr -= refs;
  83                 while (refs--)
  84                         put_page(head);
  85                 return 0;
  86         }
  87
  88         /*
  89          * Any tail page need their mapcount reference taken before we
  90          * return.
  91          */
  92         while (refs--) {
  93                 if (PageTail(tail))
  94                         get_huge_page_tail(tail);
  95                 tail++;
  96         }
  97
  98         return 1;
  99 }
 100
 101
 102 static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
 103                 unsigned long end, int write, struct page **pages, int *nr)
 104 {
 105         unsigned long next;
 106         pmd_t *pmdp, pmd;
 107
 108         pmdp = (pmd_t *) pudp;
 109 #ifdef CONFIG_64BIT
 110         if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
 111                 pmdp = (pmd_t *) pud_deref(pud);
 112         pmdp += pmd_index(addr);
 113 #endif
 114         do {
 115                 pmd = *pmdp;
 116                 barrier();
 117                 next = pmd_addr_end(addr, end);
 118                 if (pmd_none(pmd))
 119                         return 0;
 120                 if (unlikely(pmd_huge(pmd))) {
 121                         if (!gup_huge_pmd(pmdp, pmd, addr, next,
 122                                           write, pages, nr))
 123                                 return 0;
 124                 } else if (!gup_pte_range(pmdp, pmd, addr, next,
 125                                           write, pages, nr))
 126                         return 0;
 127         } while (pmdp++, addr = next, addr != end);
 128
 129         return 1;
 130 }
 131
 132 static inline int gup_pud_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
 133                 unsigned long end, int write, struct page **pages, int *nr)
 134 {
 135         unsigned long next;
 136         pud_t *pudp, pud;
 137
 138         pudp = (pud_t *) pgdp;
 139 #ifdef CONFIG_64BIT
 140         if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2)
 141                 pudp = (pud_t *) pgd_deref(pgd);
 142         pudp += pud_index(addr);
 143 #endif
 144         do {
 145                 pud = *pudp;
 146                 barrier();
 147                 next = pud_addr_end(addr, end);
 148                 if (pud_none(pud))
 149                         return 0;
 150                 if (!gup_pmd_range(pudp, pud, addr, next, write, pages, nr))
 151                         return 0;
 152         } while (pudp++, addr = next, addr != end);
 153
 154         return 1;
 155 }
 156
 157 /**
 158  * get_user_pages_fast() - pin user pages in memory
 159  * @start:      starting user address
 160  * @nr_pages:   number of pages from start to pin
 161  * @write:      whether pages will be written to
 162  * @pages:      array that receives pointers to the pages pinned.
 163  *              Should be at least nr_pages long.
 164  *
 165  * Attempt to pin user pages in memory without taking mm->mmap_sem.
 166  * If not successful, it will fall back to taking the lock and
 167  * calling get_user_pages().
 168  *
 169  * Returns number of pages pinned. This may be fewer than the number
 170  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 171  * were pinned, returns -errno.
 172  */
 173 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 174                         struct page **pages)
 175 {
 176         struct mm_struct *mm = current->mm;
 177         unsigned long addr, len, end;
 178         unsigned long next;
 179         pgd_t *pgdp, pgd;
 180         int nr = 0;
 181
 182         start &= PAGE_MASK;
 183         addr = start;
 184         len = (unsigned long) nr_pages << PAGE_SHIFT;
 185         end = start + len;
 186         if (end < start)
 187                 goto slow_irqon;
 188
 189         /*
 190          * local_irq_disable() doesn't prevent pagetable teardown, but does
 191          * prevent the pagetables from being freed on s390.
 192          *
 193          * So long as we atomically load page table pointers versus teardown,
 194          * we can follow the address down to the the page and take a ref on it.
 195          */
 196         local_irq_disable();
 197         pgdp = pgd_offset(mm, addr);
 198         do {
 199                 pgd = *pgdp;
 200                 barrier();
 201                 next = pgd_addr_end(addr, end);
 202                 if (pgd_none(pgd))
 203                         goto slow;
 204                 if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr))
 205                         goto slow;
 206         } while (pgdp++, addr = next, addr != end);
 207         local_irq_enable();
 208
 209         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 210         return nr;
 211
 212         {
 213                 int ret;
 214 slow:
 215                 local_irq_enable();
 216 slow_irqon:
 217                 /* Try to get the remaining pages with get_user_pages */
 218                 start += nr << PAGE_SHIFT;
 219                 pages += nr;
 220
 221                 down_read(&mm->mmap_sem);
 222                 ret = get_user_pages(current, mm, start,
 223                         (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
 224                 up_read(&mm->mmap_sem);
 225
 226                 /* Have to be a bit careful with return values */
 227                 if (nr > 0) {
 228                         if (ret < 0)
 229                                 ret = nr;
 230                         else
 231                                 ret += nr;
 232                 }
 233
 234                 return ret;
 235         }
 236 }