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) || !walk->vma) {
  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_pmd(walk->vma, pmd, addr);
  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
 146                 if (pte)
 147                         err = walk->hugetlb_entry(pte, hmask, addr, next, walk);
 148                 else if (walk->pte_hole)
 149                         err = walk->pte_hole(addr, next, walk);
 150
 151                 if (err)
 152                         break;
 153         } while (addr = next, addr != end);
 154
 155         return err;
 156 }
 157
 158 #else /* CONFIG_HUGETLB_PAGE */
 159 static int walk_hugetlb_range(unsigned long addr, unsigned long end,
 160                               struct mm_walk *walk)
 161 {
 162         return 0;
 163 }
 164
 165 #endif /* CONFIG_HUGETLB_PAGE */
 166
 167 /*
 168  * Decide whether we really walk over the current vma on [@start, @end)
 169  * or skip it via the returned value. Return 0 if we do walk over the
 170  * current vma, and return 1 if we skip the vma. Negative values means
 171  * error, where we abort the current walk.
 172  */
 173 static int walk_page_test(unsigned long start, unsigned long end,
 174                         struct mm_walk *walk)
 175 {
 176         struct vm_area_struct *vma = walk->vma;
 177
 178         if (walk->test_walk)
 179                 return walk->test_walk(start, end, walk);
 180
 181         /*
 182          * vma(VM_PFNMAP) doesn't have any valid struct pages behind VM_PFNMAP
 183          * range, so we don't walk over it as we do for normal vmas. However,
 184          * Some callers are interested in handling hole range and they don't
 185          * want to just ignore any single address range. Such users certainly
 186          * define their ->pte_hole() callbacks, so let's delegate them to handle
 187          * vma(VM_PFNMAP).
 188          */
 189         if (vma->vm_flags & VM_PFNMAP) {
 190                 int err = 1;
 191                 if (walk->pte_hole)
 192                         err = walk->pte_hole(start, end, walk);
 193                 return err ? err : 1;
 194         }
 195         return 0;
 196 }
 197
 198 static int __walk_page_range(unsigned long start, unsigned long end,
 199                         struct mm_walk *walk)
 200 {
 201         int err = 0;
 202         struct vm_area_struct *vma = walk->vma;
 203
 204         if (vma && is_vm_hugetlb_page(vma)) {
 205                 if (walk->hugetlb_entry)
 206                         err = walk_hugetlb_range(start, end, walk);
 207         } else
 208                 err = walk_pgd_range(start, end, walk);
 209
 210         return err;
 211 }
 212
 213 /**
 214  * walk_page_range - walk page table with caller specific callbacks
 215  *
 216  * Recursively walk the page table tree of the process represented by @walk->mm
 217  * within the virtual address range [@start, @end). During walking, we can do
 218  * some caller-specific works for each entry, by setting up pmd_entry(),
 219  * pte_entry(), and/or hugetlb_entry(). If you don't set up for some of these
 220  * callbacks, the associated entries/pages are just ignored.
 221  * The return values of these callbacks are commonly defined like below:
 222  *  - 0  : succeeded to handle the current entry, and if you don't reach the
 223  *         end address yet, continue to walk.
 224  *  - >0 : succeeded to handle the current entry, and return to the caller
 225  *         with caller specific value.
 226  *  - <0 : failed to handle the current entry, and return to the caller
 227  *         with error code.
 228  *
 229  * Before starting to walk page table, some callers want to check whether
 230  * they really want to walk over the current vma, typically by checking
 231  * its vm_flags. walk_page_test() and @walk->test_walk() are used for this
 232  * purpose.
 233  *
 234  * struct mm_walk keeps current values of some common data like vma and pmd,
 235  * which are useful for the access from callbacks. If you want to pass some
 236  * caller-specific data to callbacks, @walk->private should be helpful.
 237  *
 238  * Locking:
 239  *   Callers of walk_page_range() and walk_page_vma() should hold
 240  *   @walk->mm->mmap_sem, because these function traverse vma list and/or
 241  *   access to vma's data.
 242  */
 243 int walk_page_range(unsigned long start, unsigned long end,
 244                     struct mm_walk *walk)
 245 {
 246         int err = 0;
 247         unsigned long next;
 248         struct vm_area_struct *vma;
 249
 250         if (start >= end)
 251                 return -EINVAL;
 252
 253         if (!walk->mm)
 254                 return -EINVAL;
 255
 256         VM_BUG_ON_MM(!rwsem_is_locked(&walk->mm->mmap_sem), walk->mm);
 257
 258         vma = find_vma(walk->mm, start);
 259         do {
 260                 if (!vma) { /* after the last vma */
 261                         walk->vma = NULL;
 262                         next = end;
 263                 } else if (start < vma->vm_start) { /* outside vma */
 264                         walk->vma = NULL;
 265                         next = min(end, vma->vm_start);
 266                 } else { /* inside vma */
 267                         walk->vma = vma;
 268                         next = min(end, vma->vm_end);
 269                         vma = vma->vm_next;
 270
 271                         err = walk_page_test(start, next, walk);
 272                         if (err > 0) {
 273                                 /*
 274                                  * positive return values are purely for
 275                                  * controlling the pagewalk, so should never
 276                                  * be passed to the callers.
 277                                  */
 278                                 err = 0;
 279                                 continue;
 280                         }
 281                         if (err < 0)
 282                                 break;
 283                 }
 284                 if (walk->vma || walk->pte_hole)
 285                         err = __walk_page_range(start, next, walk);
 286                 if (err)
 287                         break;
 288         } while (start = next, start < end);
 289         return err;
 290 }
 291
 292 int walk_page_vma(struct vm_area_struct *vma, struct mm_walk *walk)
 293 {
 294         int err;
 295
 296         if (!walk->mm)
 297                 return -EINVAL;
 298
 299         VM_BUG_ON(!rwsem_is_locked(&walk->mm->mmap_sem));
 300         VM_BUG_ON(!vma);
 301         walk->vma = vma;
 302         err = walk_page_test(vma->vm_start, vma->vm_end, walk);
 303         if (err > 0)
 304                 return 0;
 305         if (err < 0)
 306                 return err;
 307         return __walk_page_range(vma->vm_start, vma->vm_end, walk);
 308 }