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_PROTECT : 0) | _PAGE_INVALID | _PAGE_SPECIAL;
  28
  29         ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr);
  30         do {
  31                 pte = *ptep;
  32                 barrier();
  33                 /* Similar to the PMD case, NUMA hinting must take slow path */
  34                 if (pte_protnone(pte))
  35                         return 0;
  36                 if ((pte_val(pte) & mask) != 0)
  37                         return 0;
  38                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
  39                 page = pte_page(pte);
  40                 if (!page_cache_get_speculative(page))
  41                         return 0;
  42                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
  43                         put_page(page);
  44                         return 0;
  45                 }
  46                 pages[*nr] = page;
  47                 (*nr)++;
  48
  49         } while (ptep++, addr += PAGE_SIZE, addr != end);
  50
  51         return 1;
  52 }
  53
  54 static inline int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
  55                 unsigned long end, int write, struct page **pages, int *nr)
  56 {
  57         unsigned long mask, result;
  58         struct page *head, *page;
  59         int refs;
  60
  61         result = write ? 0 : _SEGMENT_ENTRY_PROTECT;
  62         mask = result | _SEGMENT_ENTRY_INVALID;
  63         if ((pmd_val(pmd) & mask) != result)
  64                 return 0;
  65         VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT));
  66
  67         refs = 0;
  68         head = pmd_page(pmd);
  69         page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
  70         do {
  71                 VM_BUG_ON(compound_head(page) != head);
  72                 pages[*nr] = page;
  73                 (*nr)++;
  74                 page++;
  75                 refs++;
  76         } while (addr += PAGE_SIZE, addr != end);
  77
  78         if (!page_cache_add_speculative(head, refs)) {
  79                 *nr -= refs;
  80                 return 0;
  81         }
  82
  83         if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
  84                 *nr -= refs;
  85                 while (refs--)
  86                         put_page(head);
  87                 return 0;
  88         }
  89
  90         return 1;
  91 }
  92
  93
  94 static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
  95                 unsigned long end, int write, struct page **pages, int *nr)
  96 {
  97         unsigned long next;
  98         pmd_t *pmdp, pmd;
  99
 100         pmdp = (pmd_t *) pudp;
 101         if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
 102                 pmdp = (pmd_t *) pud_deref(pud);
 103         pmdp += pmd_index(addr);
 104         do {
 105                 pmd = *pmdp;
 106                 barrier();
 107                 next = pmd_addr_end(addr, end);
 108                 if (pmd_none(pmd))
 109                         return 0;
 110                 if (unlikely(pmd_large(pmd))) {
 111                         /*
 112                          * NUMA hinting faults need to be handled in the GUP
 113                          * slowpath for accounting purposes and so that they
 114                          * can be serialised against THP migration.
 115                          */
 116                         if (pmd_protnone(pmd))
 117                                 return 0;
 118                         if (!gup_huge_pmd(pmdp, pmd, addr, next,
 119                                           write, pages, nr))
 120                                 return 0;
 121                 } else if (!gup_pte_range(pmdp, pmd, addr, next,
 122                                           write, pages, nr))
 123                         return 0;
 124         } while (pmdp++, addr = next, addr != end);
 125
 126         return 1;
 127 }
 128
 129 static inline int gup_pud_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
 130                 unsigned long end, int write, struct page **pages, int *nr)
 131 {
 132         unsigned long next;
 133         pud_t *pudp, pud;
 134
 135         pudp = (pud_t *) pgdp;
 136         if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2)
 137                 pudp = (pud_t *) pgd_deref(pgd);
 138         pudp += pud_index(addr);
 139         do {
 140                 pud = *pudp;
 141                 barrier();
 142                 next = pud_addr_end(addr, end);
 143                 if (pud_none(pud))
 144                         return 0;
 145                 if (!gup_pmd_range(pudp, pud, addr, next, write, pages, nr))
 146                         return 0;
 147         } while (pudp++, addr = next, addr != end);
 148
 149         return 1;
 150 }
 151
 152 /*
 153  * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
 154  * back to the regular GUP.
 155  */
 156 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
 157                           struct page **pages)
 158 {
 159         struct mm_struct *mm = current->mm;
 160         unsigned long addr, len, end;
 161         unsigned long next, flags;
 162         pgd_t *pgdp, pgd;
 163         int nr = 0;
 164
 165         start &= PAGE_MASK;
 166         addr = start;
 167         len = (unsigned long) nr_pages << PAGE_SHIFT;
 168         end = start + len;
 169         if ((end <= start) || (end > TASK_SIZE))
 170                 return 0;
 171         /*
 172          * local_irq_save() doesn't prevent pagetable teardown, but does
 173          * prevent the pagetables from being freed on s390.
 174          *
 175          * So long as we atomically load page table pointers versus teardown,
 176          * we can follow the address down to the the page and take a ref on it.
 177          */
 178         local_irq_save(flags);
 179         pgdp = pgd_offset(mm, addr);
 180         do {
 181                 pgd = *pgdp;
 182                 barrier();
 183                 next = pgd_addr_end(addr, end);
 184                 if (pgd_none(pgd))
 185                         break;
 186                 if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr))
 187                         break;
 188         } while (pgdp++, addr = next, addr != end);
 189         local_irq_restore(flags);
 190
 191         return nr;
 192 }
 193
 194 /**
 195  * get_user_pages_fast() - pin user pages in memory
 196  * @start:      starting user address
 197  * @nr_pages:   number of pages from start to pin
 198  * @write:      whether pages will be written to
 199  * @pages:      array that receives pointers to the pages pinned.
 200  *              Should be at least nr_pages long.
 201  *
 202  * Attempt to pin user pages in memory without taking mm->mmap_sem.
 203  * If not successful, it will fall back to taking the lock and
 204  * calling get_user_pages().
 205  *
 206  * Returns number of pages pinned. This may be fewer than the number
 207  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 208  * were pinned, returns -errno.
 209  */
 210 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 211                         struct page **pages)
 212 {
 213         int nr, ret;
 214
 215         might_sleep();
 216         start &= PAGE_MASK;
 217         nr = __get_user_pages_fast(start, nr_pages, write, pages);
 218         if (nr == nr_pages)
 219                 return nr;
 220
 221         /* Try to get the remaining pages with get_user_pages */
 222         start += nr << PAGE_SHIFT;
 223         pages += nr;
 224         ret = get_user_pages_unlocked(start, nr_pages - nr, write, 0, pages);
 225         /* Have to be a bit careful with return values */
 226         if (nr > 0)
 227                 ret = (ret < 0) ? nr : ret + nr;
 228         return ret;
 229 }