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_mm(walk->mm, addr, pmd);
  62                 if (pmd_trans_unstable(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->pmd_entry || walk->pte_entry)
  90                         err = walk_pmd_range(pud, addr, next, walk);
  91                 if (err)
  92                         break;
  93         } while (pud++, addr = next, addr != end);
  94
  95         return err;
  96 }
  97
  98 static int walk_pgd_range(unsigned long addr, unsigned long end,
  99                           struct mm_walk *walk)
 100 {
 101         pgd_t *pgd;
 102         unsigned long next;
 103         int err = 0;
 104
 105         pgd = pgd_offset(walk->mm, addr);
 106         do {
 107                 next = pgd_addr_end(addr, end);
 108                 if (pgd_none_or_clear_bad(pgd)) {
 109                         if (walk->pte_hole)
 110                                 err = walk->pte_hole(addr, next, walk);
 111                         if (err)
 112                                 break;
 113                         continue;
 114                 }
 115                 if (walk->pmd_entry || walk->pte_entry)
 116                         err = walk_pud_range(pgd, addr, next, walk);
 117                 if (err)
 118                         break;
 119         } while (pgd++, addr = next, addr != end);
 120
 121         return err;
 122 }
 123
 124 #ifdef CONFIG_HUGETLB_PAGE
 125 static unsigned long hugetlb_entry_end(struct hstate *h, unsigned long addr,
 126                                        unsigned long end)
 127 {
 128         unsigned long boundary = (addr & huge_page_mask(h)) + huge_page_size(h);
 129         return boundary < end ? boundary : end;
 130 }
 131
 132 static int walk_hugetlb_range(unsigned long addr, unsigned long end,
 133                               struct mm_walk *walk)
 134 {
 135         struct vm_area_struct *vma = walk->vma;
 136         struct hstate *h = hstate_vma(vma);
 137         unsigned long next;
 138         unsigned long hmask = huge_page_mask(h);
 139         pte_t *pte;
 140         int err = 0;
 141
 142         do {
 143                 next = hugetlb_entry_end(h, addr, end);
 144                 pte = huge_pte_offset(walk->mm, addr & hmask);
 145                 if (pte && walk->hugetlb_entry)
 146                         err = walk->hugetlb_entry(pte, hmask, addr, next, walk);
 147                 if (err)
 148                         break;
 149         } while (addr = next, addr != end);
 150
 151         return err;
 152 }
 153
 154 #else /* CONFIG_HUGETLB_PAGE */
 155 static int walk_hugetlb_range(unsigned long addr, unsigned long end,
 156                               struct mm_walk *walk)
 157 {
 158         return 0;
 159 }
 160
 161 #endif /* CONFIG_HUGETLB_PAGE */
 162
 163 /*
 164  * Decide whether we really walk over the current vma on [@start, @end)
 165  * or skip it via the returned value. Return 0 if we do walk over the
 166  * current vma, and return 1 if we skip the vma. Negative values means
 167  * error, where we abort the current walk.
 168  *
 169  * Default check (only VM_PFNMAP check for now) is used when the caller
 170  * doesn't define test_walk() callback.
 171  */
 172 static int walk_page_test(unsigned long start, unsigned long end,
 173                         struct mm_walk *walk)
 174 {
 175         struct vm_area_struct *vma = walk->vma;
 176
 177         if (walk->test_walk)
 178                 return walk->test_walk(start, end, walk);
 179
 180         /*
 181          * Do not walk over vma(VM_PFNMAP), because we have no valid struct
 182          * page backing a VM_PFNMAP range. See also commit a9ff785e4437.
 183          */
 184         if (vma->vm_flags & VM_PFNMAP)
 185                 return 1;
 186         return 0;
 187 }
 188
 189 static int __walk_page_range(unsigned long start, unsigned long end,
 190                         struct mm_walk *walk)
 191 {
 192         int err = 0;
 193         struct vm_area_struct *vma = walk->vma;
 194
 195         if (vma && is_vm_hugetlb_page(vma)) {
 196                 if (walk->hugetlb_entry)
 197                         err = walk_hugetlb_range(start, end, walk);
 198         } else
 199                 err = walk_pgd_range(start, end, walk);
 200
 201         return err;
 202 }
 203
 204 /**
 205  * walk_page_range - walk page table with caller specific callbacks
 206  *
 207  * Recursively walk the page table tree of the process represented by @walk->mm
 208  * within the virtual address range [@start, @end). During walking, we can do
 209  * some caller-specific works for each entry, by setting up pmd_entry(),
 210  * pte_entry(), and/or hugetlb_entry(). If you don't set up for some of these
 211  * callbacks, the associated entries/pages are just ignored.
 212  * The return values of these callbacks are commonly defined like below:
 213  *  - 0  : succeeded to handle the current entry, and if you don't reach the
 214  *         end address yet, continue to walk.
 215  *  - >0 : succeeded to handle the current entry, and return to the caller
 216  *         with caller specific value.
 217  *  - <0 : failed to handle the current entry, and return to the caller
 218  *         with error code.
 219  *
 220  * Before starting to walk page table, some callers want to check whether
 221  * they really want to walk over the current vma, typically by checking
 222  * its vm_flags. walk_page_test() and @walk->test_walk() are used for this
 223  * purpose.
 224  *
 225  * struct mm_walk keeps current values of some common data like vma and pmd,
 226  * which are useful for the access from callbacks. If you want to pass some
 227  * caller-specific data to callbacks, @walk->private should be helpful.
 228  *
 229  * Locking:
 230  *   Callers of walk_page_range() and walk_page_vma() should hold
 231  *   @walk->mm->mmap_sem, because these function traverse vma list and/or
 232  *   access to vma's data.
 233  */
 234 int walk_page_range(unsigned long start, unsigned long end,
 235                     struct mm_walk *walk)
 236 {
 237         int err = 0;
 238         unsigned long next;
 239         struct vm_area_struct *vma;
 240
 241         if (start >= end)
 242                 return -EINVAL;
 243
 244         if (!walk->mm)
 245                 return -EINVAL;
 246
 247         VM_BUG_ON_MM(!rwsem_is_locked(&walk->mm->mmap_sem), walk->mm);
 248
 249         vma = find_vma(walk->mm, start);
 250         do {
 251                 if (!vma) { /* after the last vma */
 252                         walk->vma = NULL;
 253                         next = end;
 254                 } else if (start < vma->vm_start) { /* outside vma */
 255                         walk->vma = NULL;
 256                         next = min(end, vma->vm_start);
 257                 } else { /* inside vma */
 258                         walk->vma = vma;
 259                         next = min(end, vma->vm_end);
 260                         vma = vma->vm_next;
 261
 262                         err = walk_page_test(start, next, walk);
 263                         if (err > 0)
 264                                 continue;
 265                         if (err < 0)
 266                                 break;
 267                 }
 268                 if (walk->vma || walk->pte_hole)
 269                         err = __walk_page_range(start, next, walk);
 270                 if (err)
 271                         break;
 272         } while (start = next, start < end);
 273         return err;
 274 }
 275
 276 int walk_page_vma(struct vm_area_struct *vma, struct mm_walk *walk)
 277 {
 278         int err;
 279
 280         if (!walk->mm)
 281                 return -EINVAL;
 282
 283         VM_BUG_ON(!rwsem_is_locked(&walk->mm->mmap_sem));
 284         VM_BUG_ON(!vma);
 285         walk->vma = vma;
 286         err = walk_page_test(vma->vm_start, vma->vm_end, walk);
 287         if (err > 0)
 288                 return 0;
 289         if (err < 0)
 290                 return err;
 291         return __walk_page_range(vma->vm_start, vma->vm_end, walk);
 292 }