mm/pagewalk.c

   1 #include <linux/mm.h>
   2 #include <linux/highmem.h>
   3 #include <linux/sched.h>
   4 #include <linux/hugetlb.h>
   5
   6 static int walk_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
   7                           struct mm_walk *walk)
   8 {
   9         pte_t *pte;
  10         int err = 0;
  11
  12         pte = pte_offset_map(pmd, addr);
  13         for (;;) {
  14                 err = walk->pte_entry(pte, addr, addr + PAGE_SIZE, walk);
  15                 if (err)
  16                        break;
  17                 addr += PAGE_SIZE;
  18                 if (addr == end)
  19                         break;
  20                 pte++;
  21         }
  22
  23         pte_unmap(pte);
  24         return err;
  25 }
  26
  27 static int walk_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
  28                           struct mm_walk *walk)
  29 {
  30         pmd_t *pmd;
  31         unsigned long next;
  32         int err = 0;
  33
  34         pmd = pmd_offset(pud, addr);
  35         do {
  36 again:
  37                 next = pmd_addr_end(addr, end);
  38                 if (pmd_none(*pmd)) {
  39                         if (walk->pte_hole)
  40                                 err = walk->pte_hole(addr, next, walk);
  41                         if (err)
  42                                 break;
  43                         continue;
  44                 }
  45                 /*
  46                  * This implies that each ->pmd_entry() handler
  47                  * needs to know about pmd_trans_huge() pmds
  48                  */
  49                 if (walk->pmd_entry)
  50                         err = walk->pmd_entry(pmd, addr, next, walk);
  51                 if (err)
  52                         break;
  53
  54                 /*
  55                  * Check this here so we only break down trans_huge
  56                  * pages when we _need_ to
  57                  */
  58                 if (!walk->pte_entry)
  59                         continue;
  60
  61                 split_huge_page_pmd(walk->mm, pmd);
  62                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
  63                         goto again;
  64                 err = walk_pte_range(pmd, addr, next, walk);
  65                 if (err)
  66                         break;
  67         } while (pmd++, addr = next, addr != end);
  68
  69         return err;
  70 }
  71
  72 static int walk_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end,
  73                           struct mm_walk *walk)
  74 {
  75         pud_t *pud;
  76         unsigned long next;
  77         int err = 0;
  78
  79         pud = pud_offset(pgd, addr);
  80         do {
  81                 next = pud_addr_end(addr, end);
  82                 if (pud_none_or_clear_bad(pud)) {
  83                         if (walk->pte_hole)
  84                                 err = walk->pte_hole(addr, next, walk);
  85                         if (err)
  86                                 break;
  87                         continue;
  88                 }
  89                 if (walk->pud_entry)
  90                         err = walk->pud_entry(pud, addr, next, walk);
  91                 if (!err && (walk->pmd_entry || walk->pte_entry))
  92                         err = walk_pmd_range(pud, addr, next, walk);
  93                 if (err)
  94                         break;
  95         } while (pud++, addr = next, addr != end);
  96
  97         return err;
  98 }
  99
 100 #ifdef CONFIG_HUGETLB_PAGE
 101 static unsigned long hugetlb_entry_end(struct hstate *h, unsigned long addr,
 102                                        unsigned long end)
 103 {
 104         unsigned long boundary = (addr & huge_page_mask(h)) + huge_page_size(h);
 105         return boundary < end ? boundary : end;
 106 }
 107
 108 static int walk_hugetlb_range(struct vm_area_struct *vma,
 109                               unsigned long addr, unsigned long end,
 110                               struct mm_walk *walk)
 111 {
 112         struct hstate *h = hstate_vma(vma);
 113         unsigned long next;
 114         unsigned long hmask = huge_page_mask(h);
 115         pte_t *pte;
 116         int err = 0;
 117
 118         do {
 119                 next = hugetlb_entry_end(h, addr, end);
 120                 pte = huge_pte_offset(walk->mm, addr & hmask);
 121                 if (pte && walk->hugetlb_entry)
 122                         err = walk->hugetlb_entry(pte, hmask, addr, next, walk);
 123                 if (err)
 124                         return err;
 125         } while (addr = next, addr != end);
 126
 127         return 0;
 128 }
 129
 130 static struct vm_area_struct* hugetlb_vma(unsigned long addr, struct mm_walk *walk)
 131 {
 132         struct vm_area_struct *vma;
 133
 134         /* We don't need vma lookup at all. */
 135         if (!walk->hugetlb_entry)
 136                 return NULL;
 137
 138         VM_BUG_ON(!rwsem_is_locked(&walk->mm->mmap_sem));
 139         vma = find_vma(walk->mm, addr);
 140         if (vma && vma->vm_start <= addr && is_vm_hugetlb_page(vma))
 141                 return vma;
 142
 143         return NULL;
 144 }
 145
 146 #else /* CONFIG_HUGETLB_PAGE */
 147 static struct vm_area_struct* hugetlb_vma(unsigned long addr, struct mm_walk *walk)
 148 {
 149         return NULL;
 150 }
 151
 152 static int walk_hugetlb_range(struct vm_area_struct *vma,
 153                               unsigned long addr, unsigned long end,
 154                               struct mm_walk *walk)
 155 {
 156         return 0;
 157 }
 158
 159 #endif /* CONFIG_HUGETLB_PAGE */
 160
 161
 162
 163 /**
 164  * walk_page_range - walk a memory map's page tables with a callback
 165  * @addr: starting address
 166  * @end: ending address
 167  * @walk: set of callbacks to invoke for each level of the tree
 168  *
 169  * Recursively walk the page table for the memory area in a VMA,
 170  * calling supplied callbacks. Callbacks are called in-order (first
 171  * PGD, first PUD, first PMD, first PTE, second PTE... second PMD,
 172  * etc.). If lower-level callbacks are omitted, walking depth is reduced.
 173  *
 174  * Each callback receives an entry pointer and the start and end of the
 175  * associated range, and a copy of the original mm_walk for access to
 176  * the ->private or ->mm fields.
 177  *
 178  * Usually no locks are taken, but splitting transparent huge page may
 179  * take page table lock. And the bottom level iterator will map PTE
 180  * directories from highmem if necessary.
 181  *
 182  * If any callback returns a non-zero value, the walk is aborted and
 183  * the return value is propagated back to the caller. Otherwise 0 is returned.
 184  *
 185  * walk->mm->mmap_sem must be held for at least read if walk->hugetlb_entry
 186  * is !NULL.
 187  */
 188 int walk_page_range(unsigned long addr, unsigned long end,
 189                     struct mm_walk *walk)
 190 {
 191         pgd_t *pgd;
 192         unsigned long next;
 193         int err = 0;
 194
 195         if (addr >= end)
 196                 return err;
 197
 198         if (!walk->mm)
 199                 return -EINVAL;
 200
 201         pgd = pgd_offset(walk->mm, addr);
 202         do {
 203                 struct vm_area_struct *vma;
 204
 205                 next = pgd_addr_end(addr, end);
 206
 207                 /*
 208                  * handle hugetlb vma individually because pagetable walk for
 209                  * the hugetlb page is dependent on the architecture and
 210                  * we can't handled it in the same manner as non-huge pages.
 211                  */
 212                 vma = hugetlb_vma(addr, walk);
 213                 if (vma) {
 214                         if (vma->vm_end < next)
 215                                 next = vma->vm_end;
 216                         /*
 217                          * Hugepage is very tightly coupled with vma, so
 218                          * walk through hugetlb entries within a given vma.
 219                          */
 220                         err = walk_hugetlb_range(vma, addr, next, walk);
 221                         if (err)
 222                                 break;
 223                         pgd = pgd_offset(walk->mm, next);
 224                         continue;
 225                 }
 226
 227                 if (pgd_none_or_clear_bad(pgd)) {
 228                         if (walk->pte_hole)
 229                                 err = walk->pte_hole(addr, next, walk);
 230                         if (err)
 231                                 break;
 232                         pgd++;
 233                         continue;
 234                 }
 235                 if (walk->pgd_entry)
 236                         err = walk->pgd_entry(pgd, addr, next, walk);
 237                 if (!err &&
 238                     (walk->pud_entry || walk->pmd_entry || walk->pte_entry))
 239                         err = walk_pud_range(pgd, addr, next, walk);
 240                 if (err)
 241                         break;
 242                 pgd++;
 243         } while (addr = next, addr != end);
 244
 245         return err;
 246 }