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 #include <linux/vs_memory.h>
  20 #include <linux/syscalls.h>
  21
  22 #include <asm/uaccess.h>
  23 #include <asm/cacheflush.h>
  24 #include <asm/tlbflush.h>
  25
  26 static pte_t *get_one_pte_map_nested(struct mm_struct *mm, unsigned long addr)
  27 {
  28         pgd_t *pgd;
  29         pmd_t *pmd;
  30         pte_t *pte = NULL;
  31
  32         pgd = pgd_offset(mm, addr);
  33         if (pgd_none(*pgd))
  34                 goto end;
  35         if (pgd_bad(*pgd)) {
  36                 pgd_ERROR(*pgd);
  37                 pgd_clear(pgd);
  38                 goto end;
  39         }
  40
  41         pmd = pmd_offset(pgd, addr);
  42         if (pmd_none(*pmd))
  43                 goto end;
  44         if (pmd_bad(*pmd)) {
  45                 pmd_ERROR(*pmd);
  46                 pmd_clear(pmd);
  47                 goto end;
  48         }
  49
  50         pte = pte_offset_map_nested(pmd, addr);
  51         if (pte_none(*pte)) {
  52                 pte_unmap_nested(pte);
  53                 pte = NULL;
  54         }
  55 end:
  56         return pte;
  57 }
  58
  59 static pte_t *get_one_pte_map(struct mm_struct *mm, unsigned long addr)
  60 {
  61         pgd_t *pgd;
  62         pmd_t *pmd;
  63
  64         pgd = pgd_offset(mm, addr);
  65         if (pgd_none(*pgd))
  66                 return NULL;
  67         pmd = pmd_offset(pgd, addr);
  68         if (!pmd_present(*pmd))
  69                 return NULL;
  70         return pte_offset_map(pmd, addr);
  71 }
  72
  73 static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
  74 {
  75         pmd_t *pmd;
  76         pte_t *pte = NULL;
  77
  78         pmd = pmd_alloc(mm, pgd_offset(mm, addr), addr);
  79         if (pmd)
  80                 pte = pte_alloc_map(mm, pmd, addr);
  81         return pte;
  82 }
  83
  84 static int
  85 move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
  86                 unsigned long new_addr)
  87 {
  88         struct address_space *mapping = NULL;
  89         struct mm_struct *mm = vma->vm_mm;
  90         int error = 0;
  91         pte_t *src, *dst;
  92
  93         if (vma->vm_file) {
  94                 /*
  95                  * Subtle point from Rajesh Venkatasubramanian: before
  96                  * moving file-based ptes, we must lock vmtruncate out,
  97                  * since it might clean the dst vma before the src vma,
  98                  * and we propagate stale pages into the dst afterward.
  99                  */
 100                 mapping = vma->vm_file->f_mapping;
 101                 spin_lock(&mapping->i_mmap_lock);
 102         }
 103         spin_lock(&mm->page_table_lock);
 104
 105         src = get_one_pte_map_nested(mm, old_addr);
 106         if (src) {
 107                 /*
 108                  * Look to see whether alloc_one_pte_map needs to perform a
 109                  * memory allocation.  If it does then we need to drop the
 110                  * atomic kmap
 111                  */
 112                 dst = get_one_pte_map(mm, new_addr);
 113                 if (unlikely(!dst)) {
 114                         pte_unmap_nested(src);
 115                         if (mapping)
 116                                 spin_unlock(&mapping->i_mmap_lock);
 117                         dst = alloc_one_pte_map(mm, new_addr);
 118                         if (mapping && !spin_trylock(&mapping->i_mmap_lock)) {
 119                                 spin_unlock(&mm->page_table_lock);
 120                                 spin_lock(&mapping->i_mmap_lock);
 121                                 spin_lock(&mm->page_table_lock);
 122                         }
 123                         src = get_one_pte_map_nested(mm, old_addr);
 124                 }
 125                 /*
 126                  * Since alloc_one_pte_map can drop and re-acquire
 127                  * page_table_lock, we should re-check the src entry...
 128                  */
 129                 if (src) {
 130                         if (dst) {
 131                                 pte_t pte;
 132                                 pte = ptep_clear_flush(vma, old_addr, src);
 133                                 set_pte(dst, pte);
 134                         } else
 135                                 error = -ENOMEM;
 136                         pte_unmap_nested(src);
 137                 }
 138                 if (dst)
 139                         pte_unmap(dst);
 140         }
 141         spin_unlock(&mm->page_table_lock);
 142         if (mapping)
 143                 spin_unlock(&mapping->i_mmap_lock);
 144         return error;
 145 }
 146
 147 static unsigned long move_page_tables(struct vm_area_struct *vma,
 148                 unsigned long new_addr, unsigned long old_addr,
 149                 unsigned long len)
 150 {
 151         unsigned long offset;
 152
 153         flush_cache_range(vma, old_addr, old_addr + len);
 154
 155         /*
 156          * This is not the clever way to do this, but we're taking the
 157          * easy way out on the assumption that most remappings will be
 158          * only a few pages.. This also makes error recovery easier.
 159          */
 160         for (offset = 0; offset < len; offset += PAGE_SIZE) {
 161                 if (move_one_page(vma, old_addr+offset, new_addr+offset) < 0)
 162                         break;
 163                 cond_resched();
 164         }
 165         return offset;
 166 }
 167
 168 static unsigned long move_vma(struct vm_area_struct *vma,
 169                 unsigned long old_addr, unsigned long old_len,
 170                 unsigned long new_len, unsigned long new_addr)
 171 {
 172         struct mm_struct *mm = vma->vm_mm;
 173         struct vm_area_struct *new_vma;
 174         unsigned long vm_flags = vma->vm_flags;
 175         unsigned long new_pgoff;
 176         unsigned long moved_len;
 177         unsigned long excess = 0;
 178         int split = 0;
 179
 180         /*
 181          * We'd prefer to avoid failure later on in do_munmap:
 182          * which may split one vma into three before unmapping.
 183          */
 184         if (mm->map_count >= sysctl_max_map_count - 3)
 185                 return -ENOMEM;
 186
 187         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 188         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 189         if (!new_vma)
 190                 return -ENOMEM;
 191
 192         moved_len = move_page_tables(vma, new_addr, old_addr, old_len);
 193         if (moved_len < old_len) {
 194                 /*
 195                  * On error, move entries back from new area to old,
 196                  * which will succeed since page tables still there,
 197                  * and then proceed to unmap new area instead of old.
 198                  */
 199                 move_page_tables(new_vma, old_addr, new_addr, moved_len);
 200                 vma = new_vma;
 201                 old_len = new_len;
 202                 old_addr = new_addr;
 203                 new_addr = -ENOMEM;
 204         }
 205
 206         /* Conceal VM_ACCOUNT so old reservation is not undone */
 207         if (vm_flags & VM_ACCOUNT) {
 208                 vma->vm_flags &= ~VM_ACCOUNT;
 209                 excess = vma->vm_end - vma->vm_start - old_len;
 210                 if (old_addr > vma->vm_start &&
 211                     old_addr + old_len < vma->vm_end)
 212                         split = 1;
 213         }
 214
 215         if (do_munmap(mm, old_addr, old_len) < 0) {
 216                 /* OOM: unable to split vma, just get accounts right */
 217                 vm_unacct_memory(excess >> PAGE_SHIFT);
 218                 excess = 0;
 219         }
 220
 221         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 222         if (excess) {
 223                 vma->vm_flags |= VM_ACCOUNT;
 224                 if (split)
 225                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 226         }
 227
 228         // mm->total_vm += new_len >> PAGE_SHIFT;
 229         vx_vmpages_add(mm, new_len >> PAGE_SHIFT);
 230         __vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 231         if (vm_flags & VM_LOCKED) {
 232                 // mm->locked_vm += new_len >> PAGE_SHIFT;
 233                 vx_vmlocked_add(mm, new_len >> PAGE_SHIFT);
 234                 if (new_len > old_len)
 235                         make_pages_present(new_addr + old_len,
 236                                            new_addr + new_len);
 237         }
 238
 239         return new_addr;
 240 }
 241
 242 /*
 243  * Expand (or shrink) an existing mapping, potentially moving it at the
 244  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 245  *
 246  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 247  * This option implies MREMAP_MAYMOVE.
 248  */
 249 unsigned long do_mremap(unsigned long addr,
 250         unsigned long old_len, unsigned long new_len,
 251         unsigned long flags, unsigned long new_addr)
 252 {
 253         struct vm_area_struct *vma;
 254         unsigned long ret = -EINVAL;
 255         unsigned long charged = 0;
 256
 257         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 258                 goto out;
 259
 260         if (addr & ~PAGE_MASK)
 261                 goto out;
 262
 263         old_len = PAGE_ALIGN(old_len);
 264         new_len = PAGE_ALIGN(new_len);
 265
 266         /*
 267          * We allow a zero old-len as a special case
 268          * for DOS-emu "duplicate shm area" thing. But
 269          * a zero new-len is nonsensical.
 270          */
 271         if (!new_len)
 272                 goto out;
 273
 274         /* new_addr is only valid if MREMAP_FIXED is specified */
 275         if (flags & MREMAP_FIXED) {
 276                 if (new_addr & ~PAGE_MASK)
 277                         goto out;
 278                 if (!(flags & MREMAP_MAYMOVE))
 279                         goto out;
 280
 281                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 282                         goto out;
 283
 284                 /* Check if the location we're moving into overlaps the
 285                  * old location at all, and fail if it does.
 286                  */
 287                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 288                         goto out;
 289
 290                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 291                         goto out;
 292
 293                 ret = do_munmap(current->mm, new_addr, new_len);
 294                 if (ret)
 295                         goto out;
 296         }
 297
 298         /*
 299          * Always allow a shrinking remap: that just unmaps
 300          * the unnecessary pages..
 301          * do_munmap does all the needed commit accounting
 302          */
 303         if (old_len >= new_len) {
 304                 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
 305                 if (ret && old_len != new_len)
 306                         goto out;
 307                 ret = addr;
 308                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 309                         goto out;
 310                 old_len = new_len;
 311         }
 312
 313         /*
 314          * Ok, we need to grow..  or relocate.
 315          */
 316         ret = -EFAULT;
 317         vma = find_vma(current->mm, addr);
 318         if (!vma || vma->vm_start > addr)
 319                 goto out;
 320         if (is_vm_hugetlb_page(vma)) {
 321                 ret = -EINVAL;
 322                 goto out;
 323         }
 324         /* We can't remap across vm area boundaries */
 325         if (old_len > vma->vm_end - addr)
 326                 goto out;
 327         if (vma->vm_flags & VM_DONTEXPAND) {
 328                 if (new_len > old_len)
 329                         goto out;
 330         }
 331         if (vma->vm_flags & VM_LOCKED) {
 332                 unsigned long locked, lock_limit;
 333                 locked = current->mm->locked_vm << PAGE_SHIFT;
 334                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 335                 locked += new_len - old_len;
 336                 ret = -EAGAIN;
 337                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 338                         goto out;
 339                 ret = -ENOMEM;
 340                 if (!vx_vmlocked_avail(current->mm,
 341                         (new_len - old_len) >> PAGE_SHIFT))
 342                         goto out;
 343         }
 344         ret = -ENOMEM;
 345         if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 346             > current->signal->rlim[RLIMIT_AS].rlim_cur)
 347                 goto out;
 348         /* check context space, maybe only Private writable mapping? */
 349         if (!vx_vmpages_avail(current->mm, (new_len - old_len) >> PAGE_SHIFT))
 350                 goto out;
 351
 352         if (vma->vm_flags & VM_ACCOUNT) {
 353                 charged = (new_len - old_len) >> PAGE_SHIFT;
 354                 if (security_vm_enough_memory(charged))
 355                         goto out_nc;
 356         }
 357
 358         /* old_len exactly to the end of the area..
 359          * And we're not relocating the area.
 360          */
 361         if (old_len == vma->vm_end - addr &&
 362             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 363             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 364                 unsigned long max_addr = TASK_SIZE;
 365                 if (vma->vm_next)
 366                         max_addr = vma->vm_next->vm_start;
 367                 /* can we just expand the current mapping? */
 368                 if (max_addr - addr >= new_len) {
 369                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 370
 371                         vma_adjust(vma, vma->vm_start,
 372                                 addr + new_len, vma->vm_pgoff, NULL);
 373
 374                         // current->mm->total_vm += pages;
 375                         vx_vmpages_add(current->mm, pages);
 376                         __vm_stat_account(vma->vm_mm, vma->vm_flags,
 377                                                         vma->vm_file, pages);
 378                         if (vma->vm_flags & VM_LOCKED) {
 379                                 // current->mm->locked_vm += pages;
 380                                 vx_vmlocked_add(vma->vm_mm, pages);
 381                                 make_pages_present(addr + old_len,
 382                                                    addr + new_len);
 383                         }
 384                         ret = addr;
 385                         goto out;
 386                 }
 387         }
 388
 389         /*
 390          * We weren't able to just expand or shrink the area,
 391          * we need to create a new one and move it..
 392          */
 393         ret = -ENOMEM;
 394         if (flags & MREMAP_MAYMOVE) {
 395                 if (!(flags & MREMAP_FIXED)) {
 396                         unsigned long map_flags = 0;
 397                         if (vma->vm_flags & VM_MAYSHARE)
 398                                 map_flags |= MAP_SHARED;
 399
 400                         new_addr = get_unmapped_area_prot(vma->vm_file, 0, new_len,
 401                                 vma->vm_pgoff, map_flags, vma->vm_flags & VM_EXEC);
 402                         ret = new_addr;
 403                         if (new_addr & ~PAGE_MASK)
 404                                 goto out;
 405                 }
 406                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 407         }
 408 out:
 409         if (ret & ~PAGE_MASK)
 410                 vm_unacct_memory(charged);
 411 out_nc:
 412         return ret;
 413 }
 414
 415 asmlinkage unsigned long sys_mremap(unsigned long addr,
 416         unsigned long old_len, unsigned long new_len,
 417         unsigned long flags, unsigned long new_addr)
 418 {
 419         unsigned long ret;
 420
 421         down_write(&current->mm->mmap_sem);
 422         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 423         up_write(&current->mm->mmap_sem);
 424         return ret;
 425 }