mm/mremap.c

   1 /*
   2  *      linux/mm/remap.c
   3  *
   4  *      (C) Copyright 1996 Linus Torvalds
   5  */
   6
   7 #include <linux/slab.h>
   8 #include <linux/smp_lock.h>
   9 #include <linux/shm.h>
  10 #include <linux/mman.h>
  11 #include <linux/swap.h>
  12
  13 #include <asm/uaccess.h>
  14 #include <asm/pgalloc.h>
  15
  16 extern int vm_enough_memory(long pages);
  17
  18 static inline pte_t *get_one_pte(struct mm_struct *mm, unsigned long addr)
  19 {
  20         pgd_t * pgd;
  21         pmd_t * pmd;
  22         pte_t * pte = NULL;
  23
  24         pgd = pgd_offset(mm, addr);
  25         if (pgd_none(*pgd))
  26                 goto end;
  27         if (pgd_bad(*pgd)) {
  28                 pgd_ERROR(*pgd);
  29                 pgd_clear(pgd);
  30                 goto end;
  31         }
  32
  33         pmd = pmd_offset(pgd, addr);
  34         if (pmd_none(*pmd))
  35                 goto end;
  36         if (pmd_bad(*pmd)) {
  37                 pmd_ERROR(*pmd);
  38                 pmd_clear(pmd);
  39                 goto end;
  40         }
  41
  42         pte = pte_offset(pmd, addr);
  43         if (pte_none(*pte))
  44                 pte = NULL;
  45 end:
  46         return pte;
  47 }
  48
  49 static inline pte_t *alloc_one_pte(struct mm_struct *mm, unsigned long addr)
  50 {
  51         pmd_t * pmd;
  52         pte_t * pte = NULL;
  53
  54         pmd = pmd_alloc(mm, pgd_offset(mm, addr), addr);
  55         if (pmd)
  56                 pte = pte_alloc(mm, pmd, addr);
  57         return pte;
  58 }
  59
  60 static inline int copy_one_pte(struct mm_struct *mm, pte_t * src, pte_t * dst)
  61 {
  62         int error = 0;
  63         pte_t pte;
  64
  65         if (!pte_none(*src)) {
  66                 pte = ptep_get_and_clear(src);
  67                 if (!dst) {
  68                         /* No dest?  We must put it back. */
  69                         dst = src;
  70                         error++;
  71                 }
  72                 set_pte(dst, pte);
  73         }
  74         return error;
  75 }
  76
  77 static int move_one_page(struct mm_struct *mm, unsigned long old_addr, unsigned long new_addr)
  78 {
  79         int error = 0;
  80         pte_t * src;
  81
  82         spin_lock(&mm->page_table_lock);
  83         src = get_one_pte(mm, old_addr);
  84         if (src)
  85                 error = copy_one_pte(mm, src, alloc_one_pte(mm, new_addr));
  86         spin_unlock(&mm->page_table_lock);
  87         return error;
  88 }
  89
  90 static int move_page_tables(struct mm_struct * mm,
  91         unsigned long new_addr, unsigned long old_addr, unsigned long len)
  92 {
  93         unsigned long offset = len;
  94
  95         flush_cache_range(mm, old_addr, old_addr + len);
  96
  97         /*
  98          * This is not the clever way to do this, but we're taking the
  99          * easy way out on the assumption that most remappings will be
 100          * only a few pages.. This also makes error recovery easier.
 101          */
 102         while (offset) {
 103                 offset -= PAGE_SIZE;
 104                 if (move_one_page(mm, old_addr + offset, new_addr + offset))
 105                         goto oops_we_failed;
 106         }
 107         flush_tlb_range(mm, old_addr, old_addr + len);
 108         return 0;
 109
 110         /*
 111          * Ok, the move failed because we didn't have enough pages for
 112          * the new page table tree. This is unlikely, but we have to
 113          * take the possibility into account. In that case we just move
 114          * all the pages back (this will work, because we still have
 115          * the old page tables)
 116          */
 117 oops_we_failed:
 118         flush_cache_range(mm, new_addr, new_addr + len);
 119         while ((offset += PAGE_SIZE) < len)
 120                 move_one_page(mm, new_addr + offset, old_addr + offset);
 121         zap_page_range(mm, new_addr, len);
 122         return -1;
 123 }
 124
 125 static inline unsigned long move_vma(struct vm_area_struct * vma,
 126         unsigned long addr, unsigned long old_len, unsigned long new_len,
 127         unsigned long new_addr)
 128 {
 129         struct mm_struct * mm = vma->vm_mm;
 130         struct vm_area_struct * new_vma, * next, * prev;
 131         int allocated_vma;
 132
 133         new_vma = NULL;
 134         next = find_vma_prev(mm, new_addr, &prev);
 135         if (next) {
 136                 if (prev && prev->vm_end == new_addr &&
 137                     can_vma_merge(prev, vma->vm_flags) && !vma->vm_file && !(vma->vm_flags & VM_SHARED)) {
 138                         spin_lock(&mm->page_table_lock);
 139                         prev->vm_end = new_addr + new_len;
 140                         spin_unlock(&mm->page_table_lock);
 141                         new_vma = prev;
 142                         if (next != prev->vm_next)
 143                                 BUG();
 144                         if (prev->vm_end == next->vm_start && can_vma_merge(next, prev->vm_flags)) {
 145                                 spin_lock(&mm->page_table_lock);
 146                                 prev->vm_end = next->vm_end;
 147                                 __vma_unlink(mm, next, prev);
 148                                 spin_unlock(&mm->page_table_lock);
 149
 150                                 mm->map_count--;
 151                                 kmem_cache_free(vm_area_cachep, next);
 152                         }
 153                 } else if (next->vm_start == new_addr + new_len &&
 154                            can_vma_merge(next, vma->vm_flags) && !vma->vm_file && !(vma->vm_flags & VM_SHARED)) {
 155                         spin_lock(&mm->page_table_lock);
 156                         next->vm_start = new_addr;
 157                         spin_unlock(&mm->page_table_lock);
 158                         new_vma = next;
 159                 }
 160         } else {
 161                 prev = find_vma(mm, new_addr-1);
 162                 if (prev && prev->vm_end == new_addr &&
 163                     can_vma_merge(prev, vma->vm_flags) && !vma->vm_file && !(vma->vm_flags & VM_SHARED)) {
 164                         spin_lock(&mm->page_table_lock);
 165                         prev->vm_end = new_addr + new_len;
 166                         spin_unlock(&mm->page_table_lock);
 167                         new_vma = prev;
 168                 }
 169         }
 170
 171         allocated_vma = 0;
 172         if (!new_vma) {
 173                 new_vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
 174                 if (!new_vma)
 175                         goto out;
 176                 allocated_vma = 1;
 177         }
 178
 179         if (!move_page_tables(current->mm, new_addr, addr, old_len)) {
 180                 if (allocated_vma) {
 181                         *new_vma = *vma;
 182                         new_vma->vm_start = new_addr;
 183                         new_vma->vm_end = new_addr+new_len;
 184                         new_vma->vm_pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
 185                         new_vma->vm_raend = 0;
 186                         if (new_vma->vm_file)
 187                                 get_file(new_vma->vm_file);
 188                         if (new_vma->vm_ops && new_vma->vm_ops->open)
 189                                 new_vma->vm_ops->open(new_vma);
 190                         insert_vm_struct(current->mm, new_vma);
 191                 }
 192                 do_munmap(current->mm, addr, old_len);
 193                 current->mm->total_vm += new_len >> PAGE_SHIFT;
 194                 if (new_vma->vm_flags & VM_LOCKED) {
 195                         current->mm->locked_vm += new_len >> PAGE_SHIFT;
 196                         make_pages_present(new_vma->vm_start,
 197                                            new_vma->vm_end);
 198                 }
 199                 return new_addr;
 200         }
 201         if (allocated_vma)
 202                 kmem_cache_free(vm_area_cachep, new_vma);
 203  out:
 204         return -ENOMEM;
 205 }
 206
 207 /*
 208  * Expand (or shrink) an existing mapping, potentially moving it at the
 209  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 210  *
 211  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 212  * This option implies MREMAP_MAYMOVE.
 213  */
 214 unsigned long do_mremap(unsigned long addr,
 215         unsigned long old_len, unsigned long new_len,
 216         unsigned long flags, unsigned long new_addr)
 217 {
 218         struct vm_area_struct *vma;
 219         unsigned long ret = -EINVAL;
 220
 221         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 222                 goto out;
 223
 224         if (addr & ~PAGE_MASK)
 225                 goto out;
 226
 227         old_len = PAGE_ALIGN(old_len);
 228         new_len = PAGE_ALIGN(new_len);
 229
 230         /* new_addr is only valid if MREMAP_FIXED is specified */
 231         if (flags & MREMAP_FIXED) {
 232                 if (new_addr & ~PAGE_MASK)
 233                         goto out;
 234                 if (!(flags & MREMAP_MAYMOVE))
 235                         goto out;
 236
 237                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 238                         goto out;
 239
 240                 /* Check if the location we're moving into overlaps the
 241                  * old location at all, and fail if it does.
 242                  */
 243                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 244                         goto out;
 245
 246                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 247                         goto out;
 248
 249                 do_munmap(current->mm, new_addr, new_len);
 250         }
 251
 252         /*
 253          * Always allow a shrinking remap: that just unmaps
 254          * the unnecessary pages..
 255          */
 256         ret = addr;
 257         if (old_len >= new_len) {
 258                 do_munmap(current->mm, addr+new_len, old_len - new_len);
 259                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 260                         goto out;
 261         }
 262
 263         /*
 264          * Ok, we need to grow..  or relocate.
 265          */
 266         ret = -EFAULT;
 267         vma = find_vma(current->mm, addr);
 268         if (!vma || vma->vm_start > addr)
 269                 goto out;
 270         /* We can't remap across vm area boundaries */
 271         if (old_len > vma->vm_end - addr)
 272                 goto out;
 273         if (vma->vm_flags & VM_DONTEXPAND) {
 274                 if (new_len > old_len)
 275                         goto out;
 276         }
 277         if (vma->vm_flags & VM_LOCKED) {
 278                 unsigned long locked = current->mm->locked_vm << PAGE_SHIFT;
 279                 locked += new_len - old_len;
 280                 ret = -EAGAIN;
 281                 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
 282                         goto out;
 283         }
 284         ret = -ENOMEM;
 285         if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 286             > current->rlim[RLIMIT_AS].rlim_cur)
 287                 goto out;
 288         /* Private writable mapping? Check memory availability.. */
 289         if ((vma->vm_flags & (VM_SHARED | VM_WRITE)) == VM_WRITE &&
 290             !(flags & MAP_NORESERVE)                             &&
 291             !vm_enough_memory((new_len - old_len) >> PAGE_SHIFT))
 292                 goto out;
 293
 294         /* old_len exactly to the end of the area..
 295          * And we're not relocating the area.
 296          */
 297         if (old_len == vma->vm_end - addr &&
 298             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 299             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 300                 unsigned long max_addr = TASK_SIZE;
 301                 if (vma->vm_next)
 302                         max_addr = vma->vm_next->vm_start;
 303                 /* can we just expand the current mapping? */
 304                 if (max_addr - addr >= new_len) {
 305                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 306                         spin_lock(&vma->vm_mm->page_table_lock);
 307                         vma->vm_end = addr + new_len;
 308                         spin_unlock(&vma->vm_mm->page_table_lock);
 309                         current->mm->total_vm += pages;
 310                         if (vma->vm_flags & VM_LOCKED) {
 311                                 current->mm->locked_vm += pages;
 312                                 make_pages_present(addr + old_len,
 313                                                    addr + new_len);
 314                         }
 315                         ret = addr;
 316                         goto out;
 317                 }
 318         }
 319
 320         /*
 321          * We weren't able to just expand or shrink the area,
 322          * we need to create a new one and move it..
 323          */
 324         ret = -ENOMEM;
 325         if (flags & MREMAP_MAYMOVE) {
 326                 if (!(flags & MREMAP_FIXED)) {
 327                         unsigned long map_flags = 0;
 328                         if (vma->vm_flags & VM_SHARED)
 329                                 map_flags |= MAP_SHARED;
 330
 331                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len, vma->vm_pgoff, map_flags);
 332                         ret = new_addr;
 333                         if (new_addr & ~PAGE_MASK)
 334                                 goto out;
 335                 }
 336                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 337         }
 338 out:
 339         return ret;
 340 }
 341
 342 asmlinkage unsigned long sys_mremap(unsigned long addr,
 343         unsigned long old_len, unsigned long new_len,
 344         unsigned long flags, unsigned long new_addr)
 345 {
 346         unsigned long ret;
 347
 348         down_write(&current->mm->mmap_sem);
 349         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 350         up_write(&current->mm->mmap_sem);
 351         return ret;
 352 }