mm/mremap.c

   1 /*
   2  *      mm/mremap.c
   3  *
   4  *      (C) Copyright 1996 Linus Torvalds
   5  *
   6  *      Address space accounting code   <alan@redhat.com>
   7  *      (C) Copyright 2002 Red Hat Inc, All Rights Reserved
   8  */
   9
  10 #include <linux/mm.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/slab.h>
  13 #include <linux/shm.h>
  14 #include <linux/mman.h>
  15 #include <linux/swap.h>
  16 #include <linux/fs.h>
  17 #include <linux/highmem.h>
  18 #include <linux/security.h>
  19
  20 #include <asm/uaccess.h>
  21 #include <asm/pgalloc.h>
  22 #include <asm/cacheflush.h>
  23 #include <asm/tlbflush.h>
  24
  25 static pte_t *get_one_pte_map_nested(struct mm_struct *mm, unsigned long addr)
  26 {
  27         pgd_t *pgd;
  28         pmd_t *pmd;
  29         pte_t *pte = NULL;
  30
  31         pgd = pgd_offset(mm, addr);
  32         if (pgd_none(*pgd))
  33                 goto end;
  34         if (pgd_bad(*pgd)) {
  35                 pgd_ERROR(*pgd);
  36                 pgd_clear(pgd);
  37                 goto end;
  38         }
  39
  40         pmd = pmd_offset(pgd, addr);
  41         if (pmd_none(*pmd))
  42                 goto end;
  43         if (pmd_bad(*pmd)) {
  44                 pmd_ERROR(*pmd);
  45                 pmd_clear(pmd);
  46                 goto end;
  47         }
  48
  49         pte = pte_offset_map_nested(pmd, addr);
  50         if (pte_none(*pte)) {
  51                 pte_unmap_nested(pte);
  52                 pte = NULL;
  53         }
  54 end:
  55         return pte;
  56 }
  57
  58 static pte_t *get_one_pte_map(struct mm_struct *mm, unsigned long addr)
  59 {
  60         pgd_t *pgd;
  61         pmd_t *pmd;
  62
  63         pgd = pgd_offset(mm, addr);
  64         if (pgd_none(*pgd))
  65                 return NULL;
  66         pmd = pmd_offset(pgd, addr);
  67         if (!pmd_present(*pmd))
  68                 return NULL;
  69         return pte_offset_map(pmd, addr);
  70 }
  71
  72 static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
  73 {
  74         pmd_t *pmd;
  75         pte_t *pte = NULL;
  76
  77         pmd = pmd_alloc(mm, pgd_offset(mm, addr), addr);
  78         if (pmd)
  79                 pte = pte_alloc_map(mm, pmd, addr);
  80         return pte;
  81 }
  82
  83 static int
  84 move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
  85                 unsigned long new_addr)
  86 {
  87         struct address_space *mapping = NULL;
  88         struct mm_struct *mm = vma->vm_mm;
  89         int error = 0;
  90         pte_t *src, *dst;
  91
  92         if (vma->vm_file) {
  93                 /*
  94                  * Subtle point from Rajesh Venkatasubramanian: before
  95                  * moving file-based ptes, we must lock vmtruncate out,
  96                  * since it might clean the dst vma before the src vma,
  97                  * and we propagate stale pages into the dst afterward.
  98                  */
  99                 mapping = vma->vm_file->f_mapping;
 100                 spin_lock(&mapping->i_mmap_lock);
 101         }
 102         spin_lock(&mm->page_table_lock);
 103
 104         src = get_one_pte_map_nested(mm, old_addr);
 105         if (src) {
 106                 /*
 107                  * Look to see whether alloc_one_pte_map needs to perform a
 108                  * memory allocation.  If it does then we need to drop the
 109                  * atomic kmap
 110                  */
 111                 dst = get_one_pte_map(mm, new_addr);
 112                 if (unlikely(!dst)) {
 113                         pte_unmap_nested(src);
 114                         if (mapping)
 115                                 spin_unlock(&mapping->i_mmap_lock);
 116                         dst = alloc_one_pte_map(mm, new_addr);
 117                         if (mapping && !spin_trylock(&mapping->i_mmap_lock)) {
 118                                 spin_unlock(&mm->page_table_lock);
 119                                 spin_lock(&mapping->i_mmap_lock);
 120                                 spin_lock(&mm->page_table_lock);
 121                         }
 122                         src = get_one_pte_map_nested(mm, old_addr);
 123                 }
 124                 /*
 125                  * Since alloc_one_pte_map can drop and re-acquire
 126                  * page_table_lock, we should re-check the src entry...
 127                  */
 128                 if (src) {
 129                         if (dst) {
 130                                 pte_t pte;
 131                                 pte = ptep_clear_flush(vma, old_addr, src);
 132                                 set_pte(dst, pte);
 133                         } else
 134                                 error = -ENOMEM;
 135                         pte_unmap_nested(src);
 136                 }
 137                 if (dst)
 138                         pte_unmap(dst);
 139         }
 140         spin_unlock(&mm->page_table_lock);
 141         if (mapping)
 142                 spin_unlock(&mapping->i_mmap_lock);
 143         return error;
 144 }
 145
 146 static unsigned long move_page_tables(struct vm_area_struct *vma,
 147                 unsigned long new_addr, unsigned long old_addr,
 148                 unsigned long len)
 149 {
 150         unsigned long offset;
 151
 152         flush_cache_range(vma, old_addr, old_addr + len);
 153
 154         /*
 155          * This is not the clever way to do this, but we're taking the
 156          * easy way out on the assumption that most remappings will be
 157          * only a few pages.. This also makes error recovery easier.
 158          */
 159         for (offset = 0; offset < len; offset += PAGE_SIZE) {
 160                 if (move_one_page(vma, old_addr+offset, new_addr+offset) < 0)
 161                         break;
 162                 cond_resched();
 163         }
 164         return offset;
 165 }
 166
 167 static unsigned long move_vma(struct vm_area_struct *vma,
 168                 unsigned long old_addr, unsigned long old_len,
 169                 unsigned long new_len, unsigned long new_addr)
 170 {
 171         struct mm_struct *mm = vma->vm_mm;
 172         struct vm_area_struct *new_vma;
 173         unsigned long vm_flags = vma->vm_flags;
 174         unsigned long new_pgoff;
 175         unsigned long moved_len;
 176         unsigned long excess = 0;
 177         int split = 0;
 178
 179         /*
 180          * We'd prefer to avoid failure later on in do_munmap:
 181          * which may split one vma into three before unmapping.
 182          */
 183         if (mm->map_count >= sysctl_max_map_count - 3)
 184                 return -ENOMEM;
 185
 186         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 187         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 188         if (!new_vma)
 189                 return -ENOMEM;
 190
 191         moved_len = move_page_tables(vma, new_addr, old_addr, old_len);
 192         if (moved_len < old_len) {
 193                 /*
 194                  * On error, move entries back from new area to old,
 195                  * which will succeed since page tables still there,
 196                  * and then proceed to unmap new area instead of old.
 197                  */
 198                 move_page_tables(new_vma, old_addr, new_addr, moved_len);
 199                 vma = new_vma;
 200                 old_len = new_len;
 201                 old_addr = new_addr;
 202                 new_addr = -ENOMEM;
 203         }
 204
 205         /* Conceal VM_ACCOUNT so old reservation is not undone */
 206         if (vm_flags & VM_ACCOUNT) {
 207                 vma->vm_flags &= ~VM_ACCOUNT;
 208                 excess = vma->vm_end - vma->vm_start - old_len;
 209                 if (old_addr > vma->vm_start &&
 210                     old_addr + old_len < vma->vm_end)
 211                         split = 1;
 212         }
 213
 214         if (do_munmap(mm, old_addr, old_len) < 0) {
 215                 /* OOM: unable to split vma, just get accounts right */
 216                 vm_unacct_memory(excess >> PAGE_SHIFT);
 217                 excess = 0;
 218         }
 219
 220         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 221         if (excess) {
 222                 vma->vm_flags |= VM_ACCOUNT;
 223                 if (split)
 224                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 225         }
 226
 227         // mm->total_vm += new_len >> PAGE_SHIFT;
 228         vx_vmpages_add(mm, new_len >> PAGE_SHIFT);
 229         if (vm_flags & VM_LOCKED) {
 230                 // mm->locked_vm += new_len >> PAGE_SHIFT;
 231                 vx_vmlocked_add(mm, new_len >> PAGE_SHIFT);
 232                 if (new_len > old_len)
 233                         make_pages_present(new_addr + old_len,
 234                                            new_addr + new_len);
 235         }
 236
 237         return new_addr;
 238 }
 239
 240 /*
 241  * Expand (or shrink) an existing mapping, potentially moving it at the
 242  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 243  *
 244  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 245  * This option implies MREMAP_MAYMOVE.
 246  */
 247 unsigned long do_mremap(unsigned long addr,
 248         unsigned long old_len, unsigned long new_len,
 249         unsigned long flags, unsigned long new_addr)
 250 {
 251         struct vm_area_struct *vma;
 252         unsigned long ret = -EINVAL;
 253         unsigned long charged = 0;
 254
 255         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 256                 goto out;
 257
 258         if (addr & ~PAGE_MASK)
 259                 goto out;
 260
 261         old_len = PAGE_ALIGN(old_len);
 262         new_len = PAGE_ALIGN(new_len);
 263
 264         /*
 265          * We allow a zero old-len as a special case
 266          * for DOS-emu "duplicate shm area" thing. But
 267          * a zero new-len is nonsensical.
 268          */
 269         if (!new_len)
 270                 goto out;
 271
 272         /* new_addr is only valid if MREMAP_FIXED is specified */
 273         if (flags & MREMAP_FIXED) {
 274                 if (new_addr & ~PAGE_MASK)
 275                         goto out;
 276                 if (!(flags & MREMAP_MAYMOVE))
 277                         goto out;
 278
 279                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 280                         goto out;
 281
 282                 /* Check if the location we're moving into overlaps the
 283                  * old location at all, and fail if it does.
 284                  */
 285                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 286                         goto out;
 287
 288                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 289                         goto out;
 290
 291                 ret = do_munmap(current->mm, new_addr, new_len);
 292                 if (ret)
 293                         goto out;
 294         }
 295
 296         /*
 297          * Always allow a shrinking remap: that just unmaps
 298          * the unnecessary pages..
 299          * do_munmap does all the needed commit accounting
 300          */
 301         if (old_len >= new_len) {
 302                 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
 303                 if (ret && old_len != new_len)
 304                         goto out;
 305                 ret = addr;
 306                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 307                         goto out;
 308                 old_len = new_len;
 309         }
 310
 311         /*
 312          * Ok, we need to grow..  or relocate.
 313          */
 314         ret = -EFAULT;
 315         vma = find_vma(current->mm, addr);
 316         if (!vma || vma->vm_start > addr)
 317                 goto out;
 318         if (is_vm_hugetlb_page(vma)) {
 319                 ret = -EINVAL;
 320                 goto out;
 321         }
 322         /* We can't remap across vm area boundaries */
 323         if (old_len > vma->vm_end - addr)
 324                 goto out;
 325         if (vma->vm_flags & VM_DONTEXPAND) {
 326                 if (new_len > old_len)
 327                         goto out;
 328         }
 329         if (vma->vm_flags & VM_LOCKED) {
 330                 unsigned long locked = current->mm->locked_vm << PAGE_SHIFT;
 331                 locked += new_len - old_len;
 332                 ret = -EAGAIN;
 333                 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
 334                         goto out;
 335                 ret = -ENOMEM;
 336                 if (!vx_vmlocked_avail(current->mm,
 337                         (new_len - old_len) >> PAGE_SHIFT))
 338                         goto out;
 339         }
 340         ret = -ENOMEM;
 341         if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 342             > current->rlim[RLIMIT_AS].rlim_cur)
 343                 goto out;
 344         /* check context space, maybe only Private writable mapping? */
 345         if (!vx_vmpages_avail(current->mm, (new_len - old_len) >> PAGE_SHIFT))
 346                 goto out;
 347
 348         if (vma->vm_flags & VM_ACCOUNT) {
 349                 charged = (new_len - old_len) >> PAGE_SHIFT;
 350                 if (security_vm_enough_memory(charged))
 351                         goto out_nc;
 352         }
 353
 354         /* old_len exactly to the end of the area..
 355          * And we're not relocating the area.
 356          */
 357         if (old_len == vma->vm_end - addr &&
 358             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 359             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 360                 unsigned long max_addr = TASK_SIZE;
 361                 if (vma->vm_next)
 362                         max_addr = vma->vm_next->vm_start;
 363                 /* can we just expand the current mapping? */
 364                 if (max_addr - addr >= new_len) {
 365                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 366
 367                         vma_adjust(vma, vma->vm_start,
 368                                 addr + new_len, vma->vm_pgoff, NULL);
 369
 370                         // current->mm->total_vm += pages;
 371                         vx_vmpages_add(current->mm, pages);
 372                         if (vma->vm_flags & VM_LOCKED) {
 373                                 // current->mm->locked_vm += pages;
 374                                 vx_vmlocked_add(vma->vm_mm, pages);
 375                                 make_pages_present(addr + old_len,
 376                                                    addr + new_len);
 377                         }
 378                         ret = addr;
 379                         goto out;
 380                 }
 381         }
 382
 383         /*
 384          * We weren't able to just expand or shrink the area,
 385          * we need to create a new one and move it..
 386          */
 387         ret = -ENOMEM;
 388         if (flags & MREMAP_MAYMOVE) {
 389                 if (!(flags & MREMAP_FIXED)) {
 390                         unsigned long map_flags = 0;
 391                         if (vma->vm_flags & VM_MAYSHARE)
 392                                 map_flags |= MAP_SHARED;
 393
 394                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 395                                                 vma->vm_pgoff, map_flags);
 396                         ret = new_addr;
 397                         if (new_addr & ~PAGE_MASK)
 398                                 goto out;
 399                 }
 400                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 401         }
 402 out:
 403         if (ret & ~PAGE_MASK)
 404                 vm_unacct_memory(charged);
 405 out_nc:
 406         return ret;
 407 }
 408
 409 asmlinkage unsigned long sys_mremap(unsigned long addr,
 410         unsigned long old_len, unsigned long new_len,
 411         unsigned long flags, unsigned long new_addr)
 412 {
 413         unsigned long ret;
 414
 415         down_write(&current->mm->mmap_sem);
 416         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 417         up_write(&current->mm->mmap_sem);
 418         return ret;
 419 }