arch/sh/mm/gup.c

   1 /*
   2  * Lockless get_user_pages_fast for SuperH
   3  *
   4  * Copyright (C) 2009 - 2010  Paul Mundt
   5  *
   6  * Cloned from the x86 and PowerPC versions, by:
   7  *
   8  *      Copyright (C) 2008 Nick Piggin
   9  *      Copyright (C) 2008 Novell Inc.
  10  */
  11 #include <linux/sched.h>
  12 #include <linux/mm.h>
  13 #include <linux/vmstat.h>
  14 #include <linux/highmem.h>
  15 #include <asm/pgtable.h>
  16
  17 static inline pte_t gup_get_pte(pte_t *ptep)
  18 {
  19 #ifndef CONFIG_X2TLB
  20         return READ_ONCE(*ptep);
  21 #else
  22         /*
  23          * With get_user_pages_fast, we walk down the pagetables without
  24          * taking any locks.  For this we would like to load the pointers
  25          * atomically, but that is not possible with 64-bit PTEs.  What
  26          * we do have is the guarantee that a pte will only either go
  27          * from not present to present, or present to not present or both
  28          * -- it will not switch to a completely different present page
  29          * without a TLB flush in between; something that we are blocking
  30          * by holding interrupts off.
  31          *
  32          * Setting ptes from not present to present goes:
  33          * ptep->pte_high = h;
  34          * smp_wmb();
  35          * ptep->pte_low = l;
  36          *
  37          * And present to not present goes:
  38          * ptep->pte_low = 0;
  39          * smp_wmb();
  40          * ptep->pte_high = 0;
  41          *
  42          * We must ensure here that the load of pte_low sees l iff pte_high
  43          * sees h. We load pte_high *after* loading pte_low, which ensures we
  44          * don't see an older value of pte_high.  *Then* we recheck pte_low,
  45          * which ensures that we haven't picked up a changed pte high. We might
  46          * have got rubbish values from pte_low and pte_high, but we are
  47          * guaranteed that pte_low will not have the present bit set *unless*
  48          * it is 'l'. And get_user_pages_fast only operates on present ptes, so
  49          * we're safe.
  50          *
  51          * gup_get_pte should not be used or copied outside gup.c without being
  52          * very careful -- it does not atomically load the pte or anything that
  53          * is likely to be useful for you.
  54          */
  55         pte_t pte;
  56
  57 retry:
  58         pte.pte_low = ptep->pte_low;
  59         smp_rmb();
  60         pte.pte_high = ptep->pte_high;
  61         smp_rmb();
  62         if (unlikely(pte.pte_low != ptep->pte_low))
  63                 goto retry;
  64
  65         return pte;
  66 #endif
  67 }
  68
  69 /*
  70  * The performance critical leaf functions are made noinline otherwise gcc
  71  * inlines everything into a single function which results in too much
  72  * register pressure.
  73  */
  74 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
  75                 unsigned long end, int write, struct page **pages, int *nr)
  76 {
  77         u64 mask, result;
  78         pte_t *ptep;
  79
  80 #ifdef CONFIG_X2TLB
  81         result = _PAGE_PRESENT | _PAGE_EXT(_PAGE_EXT_KERN_READ | _PAGE_EXT_USER_READ);
  82         if (write)
  83                 result |= _PAGE_EXT(_PAGE_EXT_KERN_WRITE | _PAGE_EXT_USER_WRITE);
  84 #elif defined(CONFIG_SUPERH64)
  85         result = _PAGE_PRESENT | _PAGE_USER | _PAGE_READ;
  86         if (write)
  87                 result |= _PAGE_WRITE;
  88 #else
  89         result = _PAGE_PRESENT | _PAGE_USER;
  90         if (write)
  91                 result |= _PAGE_RW;
  92 #endif
  93
  94         mask = result | _PAGE_SPECIAL;
  95
  96         ptep = pte_offset_map(&pmd, addr);
  97         do {
  98                 pte_t pte = gup_get_pte(ptep);
  99                 struct page *page;
 100
 101                 if ((pte_val(pte) & mask) != result) {
 102                         pte_unmap(ptep);
 103                         return 0;
 104                 }
 105                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 106                 page = pte_page(pte);
 107                 get_page(page);
 108                 __flush_anon_page(page, addr);
 109                 flush_dcache_page(page);
 110                 pages[*nr] = page;
 111                 (*nr)++;
 112
 113         } while (ptep++, addr += PAGE_SIZE, addr != end);
 114         pte_unmap(ptep - 1);
 115
 116         return 1;
 117 }
 118
 119 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
 120                 int write, struct page **pages, int *nr)
 121 {
 122         unsigned long next;
 123         pmd_t *pmdp;
 124
 125         pmdp = pmd_offset(&pud, addr);
 126         do {
 127                 pmd_t pmd = *pmdp;
 128
 129                 next = pmd_addr_end(addr, end);
 130                 if (pmd_none(pmd))
 131                         return 0;
 132                 if (!gup_pte_range(pmd, addr, next, write, pages, nr))
 133                         return 0;
 134         } while (pmdp++, addr = next, addr != end);
 135
 136         return 1;
 137 }
 138
 139 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
 140                         int write, struct page **pages, int *nr)
 141 {
 142         unsigned long next;
 143         pud_t *pudp;
 144
 145         pudp = pud_offset(&pgd, addr);
 146         do {
 147                 pud_t pud = *pudp;
 148
 149                 next = pud_addr_end(addr, end);
 150                 if (pud_none(pud))
 151                         return 0;
 152                 if (!gup_pmd_range(pud, addr, next, write, pages, nr))
 153                         return 0;
 154         } while (pudp++, addr = next, addr != end);
 155
 156         return 1;
 157 }
 158
 159 /*
 160  * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
 161  * back to the regular GUP.
 162  */
 163 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
 164                           struct page **pages)
 165 {
 166         struct mm_struct *mm = current->mm;
 167         unsigned long addr, len, end;
 168         unsigned long next;
 169         unsigned long flags;
 170         pgd_t *pgdp;
 171         int nr = 0;
 172
 173         start &= PAGE_MASK;
 174         addr = start;
 175         len = (unsigned long) nr_pages << PAGE_SHIFT;
 176         end = start + len;
 177         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
 178                                         (void __user *)start, len)))
 179                 return 0;
 180
 181         /*
 182          * This doesn't prevent pagetable teardown, but does prevent
 183          * the pagetables and pages from being freed.
 184          */
 185         local_irq_save(flags);
 186         pgdp = pgd_offset(mm, addr);
 187         do {
 188                 pgd_t pgd = *pgdp;
 189
 190                 next = pgd_addr_end(addr, end);
 191                 if (pgd_none(pgd))
 192                         break;
 193                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 194                         break;
 195         } while (pgdp++, addr = next, addr != end);
 196         local_irq_restore(flags);
 197
 198         return nr;
 199 }
 200
 201 /**
 202  * get_user_pages_fast() - pin user pages in memory
 203  * @start:      starting user address
 204  * @nr_pages:   number of pages from start to pin
 205  * @write:      whether pages will be written to
 206  * @pages:      array that receives pointers to the pages pinned.
 207  *              Should be at least nr_pages long.
 208  *
 209  * Attempt to pin user pages in memory without taking mm->mmap_sem.
 210  * If not successful, it will fall back to taking the lock and
 211  * calling get_user_pages().
 212  *
 213  * Returns number of pages pinned. This may be fewer than the number
 214  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 215  * were pinned, returns -errno.
 216  */
 217 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 218                         struct page **pages)
 219 {
 220         struct mm_struct *mm = current->mm;
 221         unsigned long addr, len, end;
 222         unsigned long next;
 223         pgd_t *pgdp;
 224         int nr = 0;
 225
 226         start &= PAGE_MASK;
 227         addr = start;
 228         len = (unsigned long) nr_pages << PAGE_SHIFT;
 229
 230         end = start + len;
 231         if (end < start)
 232                 goto slow_irqon;
 233
 234         local_irq_disable();
 235         pgdp = pgd_offset(mm, addr);
 236         do {
 237                 pgd_t pgd = *pgdp;
 238
 239                 next = pgd_addr_end(addr, end);
 240                 if (pgd_none(pgd))
 241                         goto slow;
 242                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 243                         goto slow;
 244         } while (pgdp++, addr = next, addr != end);
 245         local_irq_enable();
 246
 247         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 248         return nr;
 249
 250         {
 251                 int ret;
 252
 253 slow:
 254                 local_irq_enable();
 255 slow_irqon:
 256                 /* Try to get the remaining pages with get_user_pages */
 257                 start += nr << PAGE_SHIFT;
 258                 pages += nr;
 259
 260                 ret = get_user_pages_unlocked(start,
 261                         (end - start) >> PAGE_SHIFT, pages,
 262                         write ? FOLL_WRITE : 0);
 263
 264                 /* Have to be a bit careful with return values */
 265                 if (nr > 0) {
 266                         if (ret < 0)
 267                                 ret = nr;
 268                         else
 269                                 ret += nr;
 270                 }
 271
 272                 return ret;
 273         }
 274 }