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(p4d_t *p4d, 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(p4d, addr);
  80         do {
  81  again:
  82                 next = pud_addr_end(addr, end);
  83                 if (pud_none(*pud) || !walk->vma) {
  84                         if (walk->pte_hole)
  85                                 err = walk->pte_hole(addr, next, walk);
  86                         if (err)
  87                                 break;
  88                         continue;
  89                 }
  90
  91                 if (walk->pud_entry) {
  92                         spinlock_t *ptl = pud_trans_huge_lock(pud, walk->vma);
  93
  94                         if (ptl) {
  95                                 err = walk->pud_entry(pud, addr, next, walk);
  96                                 spin_unlock(ptl);
  97                                 if (err)
  98                                         break;
  99                                 continue;
 100                         }
 101                 }
 102
 103                 split_huge_pud(walk->vma, pud, addr);
 104                 if (pud_none(*pud))
 105                         goto again;
 106
 107                 if (walk->pmd_entry || walk->pte_entry)
 108                         err = walk_pmd_range(pud, addr, next, walk);
 109                 if (err)
 110                         break;
 111         } while (pud++, addr = next, addr != end);
 112
 113         return err;
 114 }
 115
 116 static int walk_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 117                           struct mm_walk *walk)
 118 {
 119         p4d_t *p4d;
 120         unsigned long next;
 121         int err = 0;
 122
 123         p4d = p4d_offset(pgd, addr);
 124         do {
 125                 next = p4d_addr_end(addr, end);
 126                 if (p4d_none_or_clear_bad(p4d)) {
 127                         if (walk->pte_hole)
 128                                 err = walk->pte_hole(addr, next, walk);
 129                         if (err)
 130                                 break;
 131                         continue;
 132                 }
 133                 if (walk->pmd_entry || walk->pte_entry)
 134                         err = walk_pud_range(p4d, addr, next, walk);
 135                 if (err)
 136                         break;
 137         } while (p4d++, addr = next, addr != end);
 138
 139         return err;
 140 }
 141
 142 static int walk_pgd_range(unsigned long addr, unsigned long end,
 143                           struct mm_walk *walk)
 144 {
 145         pgd_t *pgd;
 146         unsigned long next;
 147         int err = 0;
 148
 149         pgd = pgd_offset(walk->mm, addr);
 150         do {
 151                 next = pgd_addr_end(addr, end);
 152                 if (pgd_none_or_clear_bad(pgd)) {
 153                         if (walk->pte_hole)
 154                                 err = walk->pte_hole(addr, next, walk);
 155                         if (err)
 156                                 break;
 157                         continue;
 158                 }
 159                 if (walk->pmd_entry || walk->pte_entry)
 160                         err = walk_p4d_range(pgd, addr, next, walk);
 161                 if (err)
 162                         break;
 163         } while (pgd++, addr = next, addr != end);
 164
 165         return err;
 166 }
 167
 168 #ifdef CONFIG_HUGETLB_PAGE
 169 static unsigned long hugetlb_entry_end(struct hstate *h, unsigned long addr,
 170                                        unsigned long end)
 171 {
 172         unsigned long boundary = (addr & huge_page_mask(h)) + huge_page_size(h);
 173         return boundary < end ? boundary : end;
 174 }
 175
 176 static int walk_hugetlb_range(unsigned long addr, unsigned long end,
 177                               struct mm_walk *walk)
 178 {
 179         struct vm_area_struct *vma = walk->vma;
 180         struct hstate *h = hstate_vma(vma);
 181         unsigned long next;
 182         unsigned long hmask = huge_page_mask(h);
 183         pte_t *pte;
 184         int err = 0;
 185
 186         do {
 187                 next = hugetlb_entry_end(h, addr, end);
 188                 pte = huge_pte_offset(walk->mm, addr & hmask);
 189                 if (pte && walk->hugetlb_entry)
 190                         err = walk->hugetlb_entry(pte, hmask, addr, next, walk);
 191                 if (err)
 192                         break;
 193         } while (addr = next, addr != end);
 194
 195         return err;
 196 }
 197
 198 #else /* CONFIG_HUGETLB_PAGE */
 199 static int walk_hugetlb_range(unsigned long addr, unsigned long end,
 200                               struct mm_walk *walk)
 201 {
 202         return 0;
 203 }
 204
 205 #endif /* CONFIG_HUGETLB_PAGE */
 206
 207 /*
 208  * Decide whether we really walk over the current vma on [@start, @end)
 209  * or skip it via the returned value. Return 0 if we do walk over the
 210  * current vma, and return 1 if we skip the vma. Negative values means
 211  * error, where we abort the current walk.
 212  */
 213 static int walk_page_test(unsigned long start, unsigned long end,
 214                         struct mm_walk *walk)
 215 {
 216         struct vm_area_struct *vma = walk->vma;
 217
 218         if (walk->test_walk)
 219                 return walk->test_walk(start, end, walk);
 220
 221         /*
 222          * vma(VM_PFNMAP) doesn't have any valid struct pages behind VM_PFNMAP
 223          * range, so we don't walk over it as we do for normal vmas. However,
 224          * Some callers are interested in handling hole range and they don't
 225          * want to just ignore any single address range. Such users certainly
 226          * define their ->pte_hole() callbacks, so let's delegate them to handle
 227          * vma(VM_PFNMAP).
 228          */
 229         if (vma->vm_flags & VM_PFNMAP) {
 230                 int err = 1;
 231                 if (walk->pte_hole)
 232                         err = walk->pte_hole(start, end, walk);
 233                 return err ? err : 1;
 234         }
 235         return 0;
 236 }
 237
 238 static int __walk_page_range(unsigned long start, unsigned long end,
 239                         struct mm_walk *walk)
 240 {
 241         int err = 0;
 242         struct vm_area_struct *vma = walk->vma;
 243
 244         if (vma && is_vm_hugetlb_page(vma)) {
 245                 if (walk->hugetlb_entry)
 246                         err = walk_hugetlb_range(start, end, walk);
 247         } else
 248                 err = walk_pgd_range(start, end, walk);
 249
 250         return err;
 251 }
 252
 253 /**
 254  * walk_page_range - walk page table with caller specific callbacks
 255  *
 256  * Recursively walk the page table tree of the process represented by @walk->mm
 257  * within the virtual address range [@start, @end). During walking, we can do
 258  * some caller-specific works for each entry, by setting up pmd_entry(),
 259  * pte_entry(), and/or hugetlb_entry(). If you don't set up for some of these
 260  * callbacks, the associated entries/pages are just ignored.
 261  * The return values of these callbacks are commonly defined like below:
 262  *  - 0  : succeeded to handle the current entry, and if you don't reach the
 263  *         end address yet, continue to walk.
 264  *  - >0 : succeeded to handle the current entry, and return to the caller
 265  *         with caller specific value.
 266  *  - <0 : failed to handle the current entry, and return to the caller
 267  *         with error code.
 268  *
 269  * Before starting to walk page table, some callers want to check whether
 270  * they really want to walk over the current vma, typically by checking
 271  * its vm_flags. walk_page_test() and @walk->test_walk() are used for this
 272  * purpose.
 273  *
 274  * struct mm_walk keeps current values of some common data like vma and pmd,
 275  * which are useful for the access from callbacks. If you want to pass some
 276  * caller-specific data to callbacks, @walk->private should be helpful.
 277  *
 278  * Locking:
 279  *   Callers of walk_page_range() and walk_page_vma() should hold
 280  *   @walk->mm->mmap_sem, because these function traverse vma list and/or
 281  *   access to vma's data.
 282  */
 283 int walk_page_range(unsigned long start, unsigned long end,
 284                     struct mm_walk *walk)
 285 {
 286         int err = 0;
 287         unsigned long next;
 288         struct vm_area_struct *vma;
 289
 290         if (start >= end)
 291                 return -EINVAL;
 292
 293         if (!walk->mm)
 294                 return -EINVAL;
 295
 296         VM_BUG_ON_MM(!rwsem_is_locked(&walk->mm->mmap_sem), walk->mm);
 297
 298         vma = find_vma(walk->mm, start);
 299         do {
 300                 if (!vma) { /* after the last vma */
 301                         walk->vma = NULL;
 302                         next = end;
 303                 } else if (start < vma->vm_start) { /* outside vma */
 304                         walk->vma = NULL;
 305                         next = min(end, vma->vm_start);
 306                 } else { /* inside vma */
 307                         walk->vma = vma;
 308                         next = min(end, vma->vm_end);
 309                         vma = vma->vm_next;
 310
 311                         err = walk_page_test(start, next, walk);
 312                         if (err > 0) {
 313                                 /*
 314                                  * positive return values are purely for
 315                                  * controlling the pagewalk, so should never
 316                                  * be passed to the callers.
 317                                  */
 318                                 err = 0;
 319                                 continue;
 320                         }
 321                         if (err < 0)
 322                                 break;
 323                 }
 324                 if (walk->vma || walk->pte_hole)
 325                         err = __walk_page_range(start, next, walk);
 326                 if (err)
 327                         break;
 328         } while (start = next, start < end);
 329         return err;
 330 }
 331
 332 int walk_page_vma(struct vm_area_struct *vma, struct mm_walk *walk)
 333 {
 334         int err;
 335
 336         if (!walk->mm)
 337                 return -EINVAL;
 338
 339         VM_BUG_ON(!rwsem_is_locked(&walk->mm->mmap_sem));
 340         VM_BUG_ON(!vma);
 341         walk->vma = vma;
 342         err = walk_page_test(vma->vm_start, vma->vm_end, walk);
 343         if (err > 0)
 344                 return 0;
 345         if (err < 0)
 346                 return err;
 347         return __walk_page_range(vma->vm_start, vma->vm_end, walk);
 348 }