mm/pgtable-generic.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  mm/pgtable-generic.c
   4  *
   5  *  Generic pgtable methods declared in linux/pgtable.h
   6  *
   7  *  Copyright (C) 2010  Linus Torvalds
   8  */
   9
  10 #include <linux/pagemap.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/pgtable.h>
  13 #include <linux/swap.h>
  14 #include <linux/swapops.h>
  15 #include <linux/mm_inline.h>
  16 #include <asm/tlb.h>
  17
  18 /*
  19  * If a p?d_bad entry is found while walking page tables, report
  20  * the error, before resetting entry to p?d_none.  Usually (but
  21  * very seldom) called out from the p?d_none_or_clear_bad macros.
  22  */
  23
  24 void pgd_clear_bad(pgd_t *pgd)
  25 {
  26         pgd_ERROR(*pgd);
  27         pgd_clear(pgd);
  28 }
  29
  30 #ifndef __PAGETABLE_P4D_FOLDED
  31 void p4d_clear_bad(p4d_t *p4d)
  32 {
  33         p4d_ERROR(*p4d);
  34         p4d_clear(p4d);
  35 }
  36 #endif
  37
  38 #ifndef __PAGETABLE_PUD_FOLDED
  39 void pud_clear_bad(pud_t *pud)
  40 {
  41         pud_ERROR(*pud);
  42         pud_clear(pud);
  43 }
  44 #endif
  45
  46 /*
  47  * Note that the pmd variant below can't be stub'ed out just as for p4d/pud
  48  * above. pmd folding is special and typically pmd_* macros refer to upper
  49  * level even when folded
  50  */
  51 void pmd_clear_bad(pmd_t *pmd)
  52 {
  53         pmd_ERROR(*pmd);
  54         pmd_clear(pmd);
  55 }
  56
  57 #ifndef __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
  58 /*
  59  * Only sets the access flags (dirty, accessed), as well as write
  60  * permission. Furthermore, we know it always gets set to a "more
  61  * permissive" setting, which allows most architectures to optimize
  62  * this. We return whether the PTE actually changed, which in turn
  63  * instructs the caller to do things like update__mmu_cache.  This
  64  * used to be done in the caller, but sparc needs minor faults to
  65  * force that call on sun4c so we changed this macro slightly
  66  */
  67 int ptep_set_access_flags(struct vm_area_struct *vma,
  68                           unsigned long address, pte_t *ptep,
  69                           pte_t entry, int dirty)
  70 {
  71         int changed = !pte_same(ptep_get(ptep), entry);
  72         if (changed) {
  73                 set_pte_at(vma->vm_mm, address, ptep, entry);
  74                 flush_tlb_fix_spurious_fault(vma, address, ptep);
  75         }
  76         return changed;
  77 }
  78 #endif
  79
  80 #ifndef __HAVE_ARCH_PTEP_CLEAR_YOUNG_FLUSH
  81 int ptep_clear_flush_young(struct vm_area_struct *vma,
  82                            unsigned long address, pte_t *ptep)
  83 {
  84         int young;
  85         young = ptep_test_and_clear_young(vma, address, ptep);
  86         if (young)
  87                 flush_tlb_page(vma, address);
  88         return young;
  89 }
  90 #endif
  91
  92 #ifndef __HAVE_ARCH_PTEP_CLEAR_FLUSH
  93 pte_t ptep_clear_flush(struct vm_area_struct *vma, unsigned long address,
  94                        pte_t *ptep)
  95 {
  96         struct mm_struct *mm = (vma)->vm_mm;
  97         pte_t pte;
  98         pte = ptep_get_and_clear(mm, address, ptep);
  99         if (pte_accessible(mm, pte))
 100                 flush_tlb_page(vma, address);
 101         return pte;
 102 }
 103 #endif
 104
 105 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 106
 107 #ifndef __HAVE_ARCH_PMDP_SET_ACCESS_FLAGS
 108 int pmdp_set_access_flags(struct vm_area_struct *vma,
 109                           unsigned long address, pmd_t *pmdp,
 110                           pmd_t entry, int dirty)
 111 {
 112         int changed = !pmd_same(*pmdp, entry);
 113         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 114         if (changed) {
 115                 set_pmd_at(vma->vm_mm, address, pmdp, entry);
 116                 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 117         }
 118         return changed;
 119 }
 120 #endif
 121
 122 #ifndef __HAVE_ARCH_PMDP_CLEAR_YOUNG_FLUSH
 123 int pmdp_clear_flush_young(struct vm_area_struct *vma,
 124                            unsigned long address, pmd_t *pmdp)
 125 {
 126         int young;
 127         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 128         young = pmdp_test_and_clear_young(vma, address, pmdp);
 129         if (young)
 130                 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 131         return young;
 132 }
 133 #endif
 134
 135 #ifndef __HAVE_ARCH_PMDP_HUGE_CLEAR_FLUSH
 136 pmd_t pmdp_huge_clear_flush(struct vm_area_struct *vma, unsigned long address,
 137                             pmd_t *pmdp)
 138 {
 139         pmd_t pmd;
 140         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 141         VM_BUG_ON(pmd_present(*pmdp) && !pmd_trans_huge(*pmdp) &&
 142                            !pmd_devmap(*pmdp));
 143         pmd = pmdp_huge_get_and_clear(vma->vm_mm, address, pmdp);
 144         flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 145         return pmd;
 146 }
 147
 148 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 149 pud_t pudp_huge_clear_flush(struct vm_area_struct *vma, unsigned long address,
 150                             pud_t *pudp)
 151 {
 152         pud_t pud;
 153
 154         VM_BUG_ON(address & ~HPAGE_PUD_MASK);
 155         VM_BUG_ON(!pud_trans_huge(*pudp) && !pud_devmap(*pudp));
 156         pud = pudp_huge_get_and_clear(vma->vm_mm, address, pudp);
 157         flush_pud_tlb_range(vma, address, address + HPAGE_PUD_SIZE);
 158         return pud;
 159 }
 160 #endif
 161 #endif
 162
 163 #ifndef __HAVE_ARCH_PGTABLE_DEPOSIT
 164 void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
 165                                 pgtable_t pgtable)
 166 {
 167         assert_spin_locked(pmd_lockptr(mm, pmdp));
 168
 169         /* FIFO */
 170         if (!pmd_huge_pte(mm, pmdp))
 171                 INIT_LIST_HEAD(&pgtable->lru);
 172         else
 173                 list_add(&pgtable->lru, &pmd_huge_pte(mm, pmdp)->lru);
 174         pmd_huge_pte(mm, pmdp) = pgtable;
 175 }
 176 #endif
 177
 178 #ifndef __HAVE_ARCH_PGTABLE_WITHDRAW
 179 /* no "address" argument so destroys page coloring of some arch */
 180 pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
 181 {
 182         pgtable_t pgtable;
 183
 184         assert_spin_locked(pmd_lockptr(mm, pmdp));
 185
 186         /* FIFO */
 187         pgtable = pmd_huge_pte(mm, pmdp);
 188         pmd_huge_pte(mm, pmdp) = list_first_entry_or_null(&pgtable->lru,
 189                                                           struct page, lru);
 190         if (pmd_huge_pte(mm, pmdp))
 191                 list_del(&pgtable->lru);
 192         return pgtable;
 193 }
 194 #endif
 195
 196 #ifndef __HAVE_ARCH_PMDP_INVALIDATE
 197 pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
 198                      pmd_t *pmdp)
 199 {
 200         pmd_t old = pmdp_establish(vma, address, pmdp, pmd_mkinvalid(*pmdp));
 201         flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 202         return old;
 203 }
 204 #endif
 205
 206 #ifndef __HAVE_ARCH_PMDP_INVALIDATE_AD
 207 pmd_t pmdp_invalidate_ad(struct vm_area_struct *vma, unsigned long address,
 208                          pmd_t *pmdp)
 209 {
 210         return pmdp_invalidate(vma, address, pmdp);
 211 }
 212 #endif
 213
 214 #ifndef pmdp_collapse_flush
 215 pmd_t pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
 216                           pmd_t *pmdp)
 217 {
 218         /*
 219          * pmd and hugepage pte format are same. So we could
 220          * use the same function.
 221          */
 222         pmd_t pmd;
 223
 224         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 225         VM_BUG_ON(pmd_trans_huge(*pmdp));
 226         pmd = pmdp_huge_get_and_clear(vma->vm_mm, address, pmdp);
 227
 228         /* collapse entails shooting down ptes not pmd */
 229         flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 230         return pmd;
 231 }
 232 #endif
 233 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 234
 235 pte_t *__pte_offset_map(pmd_t *pmd, unsigned long addr, pmd_t *pmdvalp)
 236 {
 237         pmd_t pmdval;
 238
 239         /* rcu_read_lock() to be added later */
 240         pmdval = pmdp_get_lockless(pmd);
 241         if (pmdvalp)
 242                 *pmdvalp = pmdval;
 243         if (unlikely(pmd_none(pmdval) || is_pmd_migration_entry(pmdval)))
 244                 goto nomap;
 245         if (unlikely(pmd_trans_huge(pmdval) || pmd_devmap(pmdval)))
 246                 goto nomap;
 247         if (unlikely(pmd_bad(pmdval))) {
 248                 pmd_clear_bad(pmd);
 249                 goto nomap;
 250         }
 251         return __pte_map(&pmdval, addr);
 252 nomap:
 253         /* rcu_read_unlock() to be added later */
 254         return NULL;
 255 }
 256
 257 pte_t *pte_offset_map_nolock(struct mm_struct *mm, pmd_t *pmd,
 258                              unsigned long addr, spinlock_t **ptlp)
 259 {
 260         pmd_t pmdval;
 261         pte_t *pte;
 262
 263         pte = __pte_offset_map(pmd, addr, &pmdval);
 264         if (likely(pte))
 265                 *ptlp = pte_lockptr(mm, &pmdval);
 266         return pte;
 267 }
 268
 269 pte_t *__pte_offset_map_lock(struct mm_struct *mm, pmd_t *pmd,
 270                              unsigned long addr, spinlock_t **ptlp)
 271 {
 272         spinlock_t *ptl;
 273         pmd_t pmdval;
 274         pte_t *pte;
 275 again:
 276         pte = __pte_offset_map(pmd, addr, &pmdval);
 277         if (unlikely(!pte))
 278                 return pte;
 279         ptl = pte_lockptr(mm, &pmdval);
 280         spin_lock(ptl);
 281         if (likely(pmd_same(pmdval, pmdp_get_lockless(pmd)))) {
 282                 *ptlp = ptl;
 283                 return pte;
 284         }
 285         pte_unmap_unlock(pte, ptl);
 286         goto again;
 287 }