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 #else /* CONFIG_HUGETLB_PAGE */
 131 static int walk_hugetlb_range(struct vm_area_struct *vma,
 132                               unsigned long addr, unsigned long end,
 133                               struct mm_walk *walk)
 134 {
 135         return 0;
 136 }
 137
 138 #endif /* CONFIG_HUGETLB_PAGE */
 139
 140
 141
 142 /**
 143  * walk_page_range - walk a memory map's page tables with a callback
 144  * @mm: memory map to walk
 145  * @addr: starting address
 146  * @end: ending address
 147  * @walk: set of callbacks to invoke for each level of the tree
 148  *
 149  * Recursively walk the page table for the memory area in a VMA,
 150  * calling supplied callbacks. Callbacks are called in-order (first
 151  * PGD, first PUD, first PMD, first PTE, second PTE... second PMD,
 152  * etc.). If lower-level callbacks are omitted, walking depth is reduced.
 153  *
 154  * Each callback receives an entry pointer and the start and end of the
 155  * associated range, and a copy of the original mm_walk for access to
 156  * the ->private or ->mm fields.
 157  *
 158  * Usually no locks are taken, but splitting transparent huge page may
 159  * take page table lock. And the bottom level iterator will map PTE
 160  * directories from highmem if necessary.
 161  *
 162  * If any callback returns a non-zero value, the walk is aborted and
 163  * the return value is propagated back to the caller. Otherwise 0 is returned.
 164  *
 165  * walk->mm->mmap_sem must be held for at least read if walk->hugetlb_entry
 166  * is !NULL.
 167  */
 168 int walk_page_range(unsigned long addr, unsigned long end,
 169                     struct mm_walk *walk)
 170 {
 171         pgd_t *pgd;
 172         unsigned long next;
 173         int err = 0;
 174
 175         if (addr >= end)
 176                 return err;
 177
 178         if (!walk->mm)
 179                 return -EINVAL;
 180
 181         VM_BUG_ON(!rwsem_is_locked(&walk->mm->mmap_sem));
 182
 183         pgd = pgd_offset(walk->mm, addr);
 184         do {
 185                 struct vm_area_struct *vma = NULL;
 186
 187                 next = pgd_addr_end(addr, end);
 188
 189                 /*
 190                  * This function was not intended to be vma based.
 191                  * But there are vma special cases to be handled:
 192                  * - hugetlb vma's
 193                  * - VM_PFNMAP vma's
 194                  */
 195                 vma = find_vma(walk->mm, addr);
 196                 if (vma) {
 197                         /*
 198                          * There are no page structures backing a VM_PFNMAP
 199                          * range, so do not allow split_huge_page_pmd().
 200                          */
 201                         if ((vma->vm_start <= addr) &&
 202                             (vma->vm_flags & VM_PFNMAP)) {
 203                                 if (walk->pte_hole)
 204                                         err = walk->pte_hole(addr, next, walk);
 205                                 if (err)
 206                                         break;
 207                                 pgd = pgd_offset(walk->mm, next);
 208                                 continue;
 209                         }
 210                         /*
 211                          * Handle hugetlb vma individually because pagetable
 212                          * walk for the hugetlb page is dependent on the
 213                          * architecture and we can't handled it in the same
 214                          * manner as non-huge pages.
 215                          */
 216                         if (walk->hugetlb_entry && (vma->vm_start <= addr) &&
 217                             is_vm_hugetlb_page(vma)) {
 218                                 if (vma->vm_end < next)
 219                                         next = vma->vm_end;
 220                                 /*
 221                                  * Hugepage is very tightly coupled with vma,
 222                                  * so walk through hugetlb entries within a
 223                                  * given vma.
 224                                  */
 225                                 err = walk_hugetlb_range(vma, addr, next, walk);
 226                                 if (err)
 227                                         break;
 228                                 pgd = pgd_offset(walk->mm, next);
 229                                 continue;
 230                         }
 231                 }
 232
 233                 if (pgd_none_or_clear_bad(pgd)) {
 234                         if (walk->pte_hole)
 235                                 err = walk->pte_hole(addr, next, walk);
 236                         if (err)
 237                                 break;
 238                         pgd++;
 239                         continue;
 240                 }
 241                 if (walk->pgd_entry)
 242                         err = walk->pgd_entry(pgd, addr, next, walk);
 243                 if (!err &&
 244                     (walk->pud_entry || walk->pmd_entry || walk->pte_entry))
 245                         err = walk_pud_range(pgd, addr, next, walk);
 246                 if (err)
 247                         break;
 248                 pgd++;
 249         } while (addr = next, addr != end);
 250
 251         return err;
 252 }