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         unsigned long sz = huge_page_size(h);
 184         pte_t *pte;
 185         int err = 0;
 186
 187         do {
 188                 next = hugetlb_entry_end(h, addr, end);
 189                 pte = huge_pte_offset(walk->mm, addr & hmask, sz);
 190                 if (pte && walk->hugetlb_entry)
 191                         err = walk->hugetlb_entry(pte, hmask, addr, next, walk);
 192                 if (err)
 193                         break;
 194         } while (addr = next, addr != end);
 195
 196         return err;
 197 }
 198
 199 #else /* CONFIG_HUGETLB_PAGE */
 200 static int walk_hugetlb_range(unsigned long addr, unsigned long end,
 201                               struct mm_walk *walk)
 202 {
 203         return 0;
 204 }
 205
 206 #endif /* CONFIG_HUGETLB_PAGE */
 207
 208 /*
 209  * Decide whether we really walk over the current vma on [@start, @end)
 210  * or skip it via the returned value. Return 0 if we do walk over the
 211  * current vma, and return 1 if we skip the vma. Negative values means
 212  * error, where we abort the current walk.
 213  */
 214 static int walk_page_test(unsigned long start, unsigned long end,
 215                         struct mm_walk *walk)
 216 {
 217         struct vm_area_struct *vma = walk->vma;
 218
 219         if (walk->test_walk)
 220                 return walk->test_walk(start, end, walk);
 221
 222         /*
 223          * vma(VM_PFNMAP) doesn't have any valid struct pages behind VM_PFNMAP
 224          * range, so we don't walk over it as we do for normal vmas. However,
 225          * Some callers are interested in handling hole range and they don't
 226          * want to just ignore any single address range. Such users certainly
 227          * define their ->pte_hole() callbacks, so let's delegate them to handle
 228          * vma(VM_PFNMAP).
 229          */
 230         if (vma->vm_flags & VM_PFNMAP) {
 231                 int err = 1;
 232                 if (walk->pte_hole)
 233                         err = walk->pte_hole(start, end, walk);
 234                 return err ? err : 1;
 235         }
 236         return 0;
 237 }
 238
 239 static int __walk_page_range(unsigned long start, unsigned long end,
 240                         struct mm_walk *walk)
 241 {
 242         int err = 0;
 243         struct vm_area_struct *vma = walk->vma;
 244
 245         if (vma && is_vm_hugetlb_page(vma)) {
 246                 if (walk->hugetlb_entry)
 247                         err = walk_hugetlb_range(start, end, walk);
 248         } else
 249                 err = walk_pgd_range(start, end, walk);
 250
 251         return err;
 252 }
 253
 254 /**
 255  * walk_page_range - walk page table with caller specific callbacks
 256  *
 257  * Recursively walk the page table tree of the process represented by @walk->mm
 258  * within the virtual address range [@start, @end). During walking, we can do
 259  * some caller-specific works for each entry, by setting up pmd_entry(),
 260  * pte_entry(), and/or hugetlb_entry(). If you don't set up for some of these
 261  * callbacks, the associated entries/pages are just ignored.
 262  * The return values of these callbacks are commonly defined like below:
 263  *  - 0  : succeeded to handle the current entry, and if you don't reach the
 264  *         end address yet, continue to walk.
 265  *  - >0 : succeeded to handle the current entry, and return to the caller
 266  *         with caller specific value.
 267  *  - <0 : failed to handle the current entry, and return to the caller
 268  *         with error code.
 269  *
 270  * Before starting to walk page table, some callers want to check whether
 271  * they really want to walk over the current vma, typically by checking
 272  * its vm_flags. walk_page_test() and @walk->test_walk() are used for this
 273  * purpose.
 274  *
 275  * struct mm_walk keeps current values of some common data like vma and pmd,
 276  * which are useful for the access from callbacks. If you want to pass some
 277  * caller-specific data to callbacks, @walk->private should be helpful.
 278  *
 279  * Locking:
 280  *   Callers of walk_page_range() and walk_page_vma() should hold
 281  *   @walk->mm->mmap_sem, because these function traverse vma list and/or
 282  *   access to vma's data.
 283  */
 284 int walk_page_range(unsigned long start, unsigned long end,
 285                     struct mm_walk *walk)
 286 {
 287         int err = 0;
 288         unsigned long next;
 289         struct vm_area_struct *vma;
 290
 291         if (start >= end)
 292                 return -EINVAL;
 293
 294         if (!walk->mm)
 295                 return -EINVAL;
 296
 297         VM_BUG_ON_MM(!rwsem_is_locked(&walk->mm->mmap_sem), walk->mm);
 298
 299         vma = find_vma(walk->mm, start);
 300         do {
 301                 if (!vma) { /* after the last vma */
 302                         walk->vma = NULL;
 303                         next = end;
 304                 } else if (start < vma->vm_start) { /* outside vma */
 305                         walk->vma = NULL;
 306                         next = min(end, vma->vm_start);
 307                 } else { /* inside vma */
 308                         walk->vma = vma;
 309                         next = min(end, vma->vm_end);
 310                         vma = vma->vm_next;
 311
 312                         err = walk_page_test(start, next, walk);
 313                         if (err > 0) {
 314                                 /*
 315                                  * positive return values are purely for
 316                                  * controlling the pagewalk, so should never
 317                                  * be passed to the callers.
 318                                  */
 319                                 err = 0;
 320                                 continue;
 321                         }
 322                         if (err < 0)
 323                                 break;
 324                 }
 325                 if (walk->vma || walk->pte_hole)
 326                         err = __walk_page_range(start, next, walk);
 327                 if (err)
 328                         break;
 329         } while (start = next, start < end);
 330         return err;
 331 }
 332
 333 int walk_page_vma(struct vm_area_struct *vma, struct mm_walk *walk)
 334 {
 335         int err;
 336
 337         if (!walk->mm)
 338                 return -EINVAL;
 339
 340         VM_BUG_ON(!rwsem_is_locked(&walk->mm->mmap_sem));
 341         VM_BUG_ON(!vma);
 342         walk->vma = vma;
 343         err = walk_page_test(vma->vm_start, vma->vm_end, walk);
 344         if (err > 0)
 345                 return 0;
 346         if (err < 0)
 347                 return err;
 348         return __walk_page_range(vma->vm_start, vma->vm_end, walk);
 349 }