kernel/mm/pgtable-generic.c

   1 /*
   2  *  mm/pgtable-generic.c
   3  *
   4  *  Generic pgtable methods declared in asm-generic/pgtable.h
   5  *
   6  *  Copyright (C) 2010  Linus Torvalds
   7  */
   8
   9 #include <linux/pagemap.h>
  10 #include <asm/tlb.h>
  11 #include <asm-generic/pgtable.h>
  12
  13 /*
  14  * If a p?d_bad entry is found while walking page tables, report
  15  * the error, before resetting entry to p?d_none.  Usually (but
  16  * very seldom) called out from the p?d_none_or_clear_bad macros.
  17  */
  18
  19 void pgd_clear_bad(pgd_t *pgd)
  20 {
  21         pgd_ERROR(*pgd);
  22         pgd_clear(pgd);
  23 }
  24
  25 void pud_clear_bad(pud_t *pud)
  26 {
  27         pud_ERROR(*pud);
  28         pud_clear(pud);
  29 }
  30
  31 void pmd_clear_bad(pmd_t *pmd)
  32 {
  33         pmd_ERROR(*pmd);
  34         pmd_clear(pmd);
  35 }
  36
  37 #ifndef __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
  38 /*
  39  * Only sets the access flags (dirty, accessed), as well as write
  40  * permission. Furthermore, we know it always gets set to a "more
  41  * permissive" setting, which allows most architectures to optimize
  42  * this. We return whether the PTE actually changed, which in turn
  43  * instructs the caller to do things like update__mmu_cache.  This
  44  * used to be done in the caller, but sparc needs minor faults to
  45  * force that call on sun4c so we changed this macro slightly
  46  */
  47 int ptep_set_access_flags(struct vm_area_struct *vma,
  48                           unsigned long address, pte_t *ptep,
  49                           pte_t entry, int dirty)
  50 {
  51         int changed = !pte_same(*ptep, entry);
  52         if (changed) {
  53                 set_pte_at(vma->vm_mm, address, ptep, entry);
  54                 flush_tlb_fix_spurious_fault(vma, address);
  55         }
  56         return changed;
  57 }
  58 #endif
  59
  60 #ifndef __HAVE_ARCH_PTEP_CLEAR_YOUNG_FLUSH
  61 int ptep_clear_flush_young(struct vm_area_struct *vma,
  62                            unsigned long address, pte_t *ptep)
  63 {
  64         int young;
  65         young = ptep_test_and_clear_young(vma, address, ptep);
  66         if (young)
  67                 flush_tlb_page(vma, address);
  68         return young;
  69 }
  70 #endif
  71
  72 #ifndef __HAVE_ARCH_PTEP_CLEAR_FLUSH
  73 pte_t ptep_clear_flush(struct vm_area_struct *vma, unsigned long address,
  74                        pte_t *ptep)
  75 {
  76         struct mm_struct *mm = (vma)->vm_mm;
  77         pte_t pte;
  78         pte = ptep_get_and_clear(mm, address, ptep);
  79         if (pte_accessible(mm, pte))
  80                 flush_tlb_page(vma, address);
  81         return pte;
  82 }
  83 #endif
  84
  85 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
  86
  87 #ifndef __HAVE_ARCH_FLUSH_PMD_TLB_RANGE
  88
  89 /*
  90  * ARCHes with special requirements for evicting THP backing TLB entries can
  91  * implement this. Otherwise also, it can help optimize normal TLB flush in
  92  * THP regime. stock flush_tlb_range() typically has optimization to nuke the
  93  * entire TLB TLB if flush span is greater than a threshhold, which will
  94  * likely be true for a single huge page. Thus a single thp flush will
  95  * invalidate the entire TLB which is not desitable.
  96  * e.g. see arch/arc: flush_pmd_tlb_range
  97  */
  98 #define flush_pmd_tlb_range(vma, addr, end)     flush_tlb_range(vma, addr, end)
  99 #endif
 100
 101 #ifndef __HAVE_ARCH_PMDP_SET_ACCESS_FLAGS
 102 int pmdp_set_access_flags(struct vm_area_struct *vma,
 103                           unsigned long address, pmd_t *pmdp,
 104                           pmd_t entry, int dirty)
 105 {
 106         int changed = !pmd_same(*pmdp, entry);
 107         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 108         if (changed) {
 109                 set_pmd_at(vma->vm_mm, address, pmdp, entry);
 110                 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 111         }
 112         return changed;
 113 }
 114 #endif
 115
 116 #ifndef __HAVE_ARCH_PMDP_CLEAR_YOUNG_FLUSH
 117 int pmdp_clear_flush_young(struct vm_area_struct *vma,
 118                            unsigned long address, pmd_t *pmdp)
 119 {
 120         int young;
 121         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 122         young = pmdp_test_and_clear_young(vma, address, pmdp);
 123         if (young)
 124                 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 125         return young;
 126 }
 127 #endif
 128
 129 #ifndef __HAVE_ARCH_PMDP_HUGE_CLEAR_FLUSH
 130 pmd_t pmdp_huge_clear_flush(struct vm_area_struct *vma, unsigned long address,
 131                             pmd_t *pmdp)
 132 {
 133         pmd_t pmd;
 134         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 135         VM_BUG_ON(!pmd_trans_huge(*pmdp));
 136         pmd = pmdp_huge_get_and_clear(vma->vm_mm, address, pmdp);
 137         flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 138         return pmd;
 139 }
 140 #endif
 141
 142 #ifndef __HAVE_ARCH_PMDP_SPLITTING_FLUSH
 143 void pmdp_splitting_flush(struct vm_area_struct *vma, unsigned long address,
 144                           pmd_t *pmdp)
 145 {
 146         pmd_t pmd = pmd_mksplitting(*pmdp);
 147         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 148         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
 149         /* tlb flush only to serialize against gup-fast */
 150         flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 151 }
 152 #endif
 153
 154 #ifndef __HAVE_ARCH_PGTABLE_DEPOSIT
 155 void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
 156                                 pgtable_t pgtable)
 157 {
 158         assert_spin_locked(pmd_lockptr(mm, pmdp));
 159
 160         /* FIFO */
 161         if (!pmd_huge_pte(mm, pmdp))
 162                 INIT_LIST_HEAD(&pgtable->lru);
 163         else
 164                 list_add(&pgtable->lru, &pmd_huge_pte(mm, pmdp)->lru);
 165         pmd_huge_pte(mm, pmdp) = pgtable;
 166 }
 167 #endif
 168
 169 #ifndef __HAVE_ARCH_PGTABLE_WITHDRAW
 170 /* no "address" argument so destroys page coloring of some arch */
 171 pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
 172 {
 173         pgtable_t pgtable;
 174
 175         assert_spin_locked(pmd_lockptr(mm, pmdp));
 176
 177         /* FIFO */
 178         pgtable = pmd_huge_pte(mm, pmdp);
 179         if (list_empty(&pgtable->lru))
 180                 pmd_huge_pte(mm, pmdp) = NULL;
 181         else {
 182                 pmd_huge_pte(mm, pmdp) = list_entry(pgtable->lru.next,
 183                                               struct page, lru);
 184                 list_del(&pgtable->lru);
 185         }
 186         return pgtable;
 187 }
 188 #endif
 189
 190 #ifndef __HAVE_ARCH_PMDP_INVALIDATE
 191 void pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
 192                      pmd_t *pmdp)
 193 {
 194         pmd_t entry = *pmdp;
 195         set_pmd_at(vma->vm_mm, address, pmdp, pmd_mknotpresent(entry));
 196         flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 197 }
 198 #endif
 199
 200 #ifndef pmdp_collapse_flush
 201 pmd_t pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
 202                           pmd_t *pmdp)
 203 {
 204         /*
 205          * pmd and hugepage pte format are same. So we could
 206          * use the same function.
 207          */
 208         pmd_t pmd;
 209
 210         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 211         VM_BUG_ON(pmd_trans_huge(*pmdp));
 212         pmd = pmdp_huge_get_and_clear(vma->vm_mm, address, pmdp);
 213
 214         /* collapse entails shooting down ptes not pmd */
 215         flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 216         return pmd;
 217 }
 218 #endif
 219 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */