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/syscalls.h>
  20 #include <linux/vs_memory.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         pud_t *pud;
  30         pmd_t *pmd;
  31         pte_t *pte = NULL;
  32
  33         pgd = pgd_offset(mm, addr);
  34         if (pgd_none_or_clear_bad(pgd))
  35                 goto end;
  36
  37         pud = pud_offset(pgd, addr);
  38         if (pud_none_or_clear_bad(pud))
  39                 goto end;
  40
  41         pmd = pmd_offset(pud, addr);
  42         if (pmd_none_or_clear_bad(pmd))
  43                 goto end;
  44
  45         pte = pte_offset_map_nested(pmd, addr);
  46         if (pte_none(*pte)) {
  47                 pte_unmap_nested(pte);
  48                 pte = NULL;
  49         }
  50 end:
  51         return pte;
  52 }
  53
  54 static pte_t *get_one_pte_map(struct mm_struct *mm, unsigned long addr)
  55 {
  56         pgd_t *pgd;
  57         pud_t *pud;
  58         pmd_t *pmd;
  59
  60         pgd = pgd_offset(mm, addr);
  61         if (pgd_none_or_clear_bad(pgd))
  62                 return NULL;
  63
  64         pud = pud_offset(pgd, addr);
  65         if (pud_none_or_clear_bad(pud))
  66                 return NULL;
  67
  68         pmd = pmd_offset(pud, addr);
  69         if (pmd_none_or_clear_bad(pmd))
  70                 return NULL;
  71
  72         return pte_offset_map(pmd, addr);
  73 }
  74
  75 static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
  76 {
  77         pgd_t *pgd;
  78         pud_t *pud;
  79         pmd_t *pmd;
  80         pte_t *pte = NULL;
  81
  82         pgd = pgd_offset(mm, addr);
  83
  84         pud = pud_alloc(mm, pgd, addr);
  85         if (!pud)
  86                 return NULL;
  87         pmd = pmd_alloc(mm, pud, addr);
  88         if (pmd)
  89                 pte = pte_alloc_map(mm, pmd, addr);
  90         return pte;
  91 }
  92
  93 static int
  94 move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
  95                 struct vm_area_struct *new_vma, unsigned long new_addr)
  96 {
  97         struct address_space *mapping = NULL;
  98         struct mm_struct *mm = vma->vm_mm;
  99         int error = 0;
 100         pte_t *src, *dst;
 101
 102         if (vma->vm_file) {
 103                 /*
 104                  * Subtle point from Rajesh Venkatasubramanian: before
 105                  * moving file-based ptes, we must lock vmtruncate out,
 106                  * since it might clean the dst vma before the src vma,
 107                  * and we propagate stale pages into the dst afterward.
 108                  */
 109                 mapping = vma->vm_file->f_mapping;
 110                 spin_lock(&mapping->i_mmap_lock);
 111                 if (new_vma->vm_truncate_count &&
 112                     new_vma->vm_truncate_count != vma->vm_truncate_count)
 113                         new_vma->vm_truncate_count = 0;
 114         }
 115         spin_lock(&mm->page_table_lock);
 116
 117         src = get_one_pte_map_nested(mm, old_addr);
 118         if (src) {
 119                 /*
 120                  * Look to see whether alloc_one_pte_map needs to perform a
 121                  * memory allocation.  If it does then we need to drop the
 122                  * atomic kmap
 123                  */
 124                 dst = get_one_pte_map(mm, new_addr);
 125                 if (unlikely(!dst)) {
 126                         pte_unmap_nested(src);
 127                         if (mapping)
 128                                 spin_unlock(&mapping->i_mmap_lock);
 129                         dst = alloc_one_pte_map(mm, new_addr);
 130                         if (mapping && !spin_trylock(&mapping->i_mmap_lock)) {
 131                                 spin_unlock(&mm->page_table_lock);
 132                                 spin_lock(&mapping->i_mmap_lock);
 133                                 spin_lock(&mm->page_table_lock);
 134                         }
 135                         src = get_one_pte_map_nested(mm, old_addr);
 136                 }
 137                 /*
 138                  * Since alloc_one_pte_map can drop and re-acquire
 139                  * page_table_lock, we should re-check the src entry...
 140                  */
 141                 if (src) {
 142                         if (dst) {
 143                                 pte_t pte;
 144                                 pte = ptep_clear_flush(vma, old_addr, src);
 145                                 set_pte_at(mm, new_addr, dst, pte);
 146                         } else
 147                                 error = -ENOMEM;
 148                         pte_unmap_nested(src);
 149                 }
 150                 if (dst)
 151                         pte_unmap(dst);
 152         }
 153         spin_unlock(&mm->page_table_lock);
 154         if (mapping)
 155                 spin_unlock(&mapping->i_mmap_lock);
 156         return error;
 157 }
 158
 159 static unsigned long move_page_tables(struct vm_area_struct *vma,
 160                 unsigned long old_addr, struct vm_area_struct *new_vma,
 161                 unsigned long new_addr, unsigned long len)
 162 {
 163         unsigned long offset;
 164
 165         flush_cache_range(vma, old_addr, old_addr + len);
 166
 167         /*
 168          * This is not the clever way to do this, but we're taking the
 169          * easy way out on the assumption that most remappings will be
 170          * only a few pages.. This also makes error recovery easier.
 171          */
 172         for (offset = 0; offset < len; offset += PAGE_SIZE) {
 173                 if (move_one_page(vma, old_addr + offset,
 174                                 new_vma, new_addr + offset) < 0)
 175                         break;
 176                 cond_resched();
 177         }
 178         return offset;
 179 }
 180
 181 static unsigned long move_vma(struct vm_area_struct *vma,
 182                 unsigned long old_addr, unsigned long old_len,
 183                 unsigned long new_len, unsigned long new_addr)
 184 {
 185         struct mm_struct *mm = vma->vm_mm;
 186         struct vm_area_struct *new_vma;
 187         unsigned long vm_flags = vma->vm_flags;
 188         unsigned long new_pgoff;
 189         unsigned long moved_len;
 190         unsigned long excess = 0;
 191         int split = 0;
 192
 193         /*
 194          * We'd prefer to avoid failure later on in do_munmap:
 195          * which may split one vma into three before unmapping.
 196          */
 197         if (mm->map_count >= sysctl_max_map_count - 3)
 198                 return -ENOMEM;
 199
 200         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 201         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 202         if (!new_vma)
 203                 return -ENOMEM;
 204
 205         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
 206         if (moved_len < old_len) {
 207                 /*
 208                  * On error, move entries back from new area to old,
 209                  * which will succeed since page tables still there,
 210                  * and then proceed to unmap new area instead of old.
 211                  */
 212                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
 213                 vma = new_vma;
 214                 old_len = new_len;
 215                 old_addr = new_addr;
 216                 new_addr = -ENOMEM;
 217         }
 218
 219         /* Conceal VM_ACCOUNT so old reservation is not undone */
 220         if (vm_flags & VM_ACCOUNT) {
 221                 vma->vm_flags &= ~VM_ACCOUNT;
 222                 excess = vma->vm_end - vma->vm_start - old_len;
 223                 if (old_addr > vma->vm_start &&
 224                     old_addr + old_len < vma->vm_end)
 225                         split = 1;
 226         }
 227
 228         /*
 229          * if we failed to move page tables we still do total_vm increment
 230          * since do_munmap() will decrement it by old_len == new_len
 231          */
 232         vx_vmpages_add(mm, new_len >> PAGE_SHIFT);
 233
 234         if (do_munmap(mm, old_addr, old_len) < 0) {
 235                 /* OOM: unable to split vma, just get accounts right */
 236                 vm_unacct_memory(excess >> PAGE_SHIFT);
 237                 excess = 0;
 238         }
 239
 240         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 241         if (excess) {
 242                 vma->vm_flags |= VM_ACCOUNT;
 243                 if (split)
 244                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 245         }
 246
 247         vx_vmpages_add(mm, new_len >> PAGE_SHIFT);
 248         __vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 249         if (vm_flags & VM_LOCKED) {
 250                 vx_vmlocked_add(mm, new_len >> PAGE_SHIFT);
 251                 if (new_len > old_len)
 252                         make_pages_present(new_addr + old_len,
 253                                            new_addr + new_len);
 254         }
 255
 256         return new_addr;
 257 }
 258
 259 /*
 260  * Expand (or shrink) an existing mapping, potentially moving it at the
 261  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 262  *
 263  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 264  * This option implies MREMAP_MAYMOVE.
 265  */
 266 unsigned long do_mremap(unsigned long addr,
 267         unsigned long old_len, unsigned long new_len,
 268         unsigned long flags, unsigned long new_addr)
 269 {
 270         struct vm_area_struct *vma;
 271         unsigned long ret = -EINVAL;
 272         unsigned long charged = 0;
 273
 274         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 275                 goto out;
 276
 277         if (addr & ~PAGE_MASK)
 278                 goto out;
 279
 280         old_len = PAGE_ALIGN(old_len);
 281         new_len = PAGE_ALIGN(new_len);
 282
 283         /*
 284          * We allow a zero old-len as a special case
 285          * for DOS-emu "duplicate shm area" thing. But
 286          * a zero new-len is nonsensical.
 287          */
 288         if (!new_len)
 289                 goto out;
 290
 291         /* new_addr is only valid if MREMAP_FIXED is specified */
 292         if (flags & MREMAP_FIXED) {
 293                 if (new_addr & ~PAGE_MASK)
 294                         goto out;
 295                 if (!(flags & MREMAP_MAYMOVE))
 296                         goto out;
 297
 298                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 299                         goto out;
 300
 301                 /* Check if the location we're moving into overlaps the
 302                  * old location at all, and fail if it does.
 303                  */
 304                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 305                         goto out;
 306
 307                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 308                         goto out;
 309
 310                 ret = do_munmap(current->mm, new_addr, new_len);
 311                 if (ret)
 312                         goto out;
 313         }
 314
 315         /*
 316          * Always allow a shrinking remap: that just unmaps
 317          * the unnecessary pages..
 318          * do_munmap does all the needed commit accounting
 319          */
 320         if (old_len >= new_len) {
 321                 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
 322                 if (ret && old_len != new_len)
 323                         goto out;
 324                 ret = addr;
 325                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 326                         goto out;
 327                 old_len = new_len;
 328         }
 329
 330         /*
 331          * Ok, we need to grow..  or relocate.
 332          */
 333         ret = -EFAULT;
 334         vma = find_vma(current->mm, addr);
 335         if (!vma || vma->vm_start > addr)
 336                 goto out;
 337         if (is_vm_hugetlb_page(vma)) {
 338                 ret = -EINVAL;
 339                 goto out;
 340         }
 341         /* We can't remap across vm area boundaries */
 342         if (old_len > vma->vm_end - addr)
 343                 goto out;
 344         if (vma->vm_flags & VM_DONTEXPAND) {
 345                 if (new_len > old_len)
 346                         goto out;
 347         }
 348         if (vma->vm_flags & VM_LOCKED) {
 349                 unsigned long locked, lock_limit;
 350                 locked = current->mm->locked_vm << PAGE_SHIFT;
 351                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 352                 locked += new_len - old_len;
 353                 ret = -EAGAIN;
 354                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 355                         goto out;
 356                 if (!vx_vmlocked_avail(current->mm,
 357                         (new_len - old_len) >> PAGE_SHIFT))
 358                         goto out;
 359         }
 360         if (!may_expand_vm(current->mm, (new_len - old_len) >> PAGE_SHIFT)) {
 361                 ret = -ENOMEM;
 362                 goto out;
 363         }
 364
 365         /* check context space, maybe only Private writable mapping? */
 366         if (!vx_vmpages_avail(current->mm, (new_len - old_len) >> PAGE_SHIFT))
 367                 goto out;
 368
 369         if (vma->vm_flags & VM_ACCOUNT) {
 370                 charged = (new_len - old_len) >> PAGE_SHIFT;
 371                 if (security_vm_enough_memory(charged))
 372                         goto out_nc;
 373         }
 374
 375         /* old_len exactly to the end of the area..
 376          * And we're not relocating the area.
 377          */
 378         if (old_len == vma->vm_end - addr &&
 379             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 380             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 381                 unsigned long max_addr = TASK_SIZE;
 382                 if (vma->vm_next)
 383                         max_addr = vma->vm_next->vm_start;
 384                 /* can we just expand the current mapping? */
 385                 if (max_addr - addr >= new_len) {
 386                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 387
 388                         vma_adjust(vma, vma->vm_start,
 389                                 addr + new_len, vma->vm_pgoff, NULL);
 390
 391                         vx_vmpages_add(current->mm, pages);
 392                         __vm_stat_account(vma->vm_mm, vma->vm_flags,
 393                                                         vma->vm_file, pages);
 394                         if (vma->vm_flags & VM_LOCKED) {
 395                                 vx_vmlocked_add(vma->vm_mm, pages);
 396                                 make_pages_present(addr + old_len,
 397                                                    addr + new_len);
 398                         }
 399                         ret = addr;
 400                         goto out;
 401                 }
 402         }
 403
 404         /*
 405          * We weren't able to just expand or shrink the area,
 406          * we need to create a new one and move it..
 407          */
 408         ret = -ENOMEM;
 409         if (flags & MREMAP_MAYMOVE) {
 410                 if (!(flags & MREMAP_FIXED)) {
 411                         unsigned long map_flags = 0;
 412                         if (vma->vm_flags & VM_MAYSHARE)
 413                                 map_flags |= MAP_SHARED;
 414
 415                         new_addr = get_unmapped_area_prot(vma->vm_file, 0, new_len,
 416                                 vma->vm_pgoff, map_flags, vma->vm_flags & VM_EXEC);
 417                         ret = new_addr;
 418                         if (new_addr & ~PAGE_MASK)
 419                                 goto out;
 420                 }
 421                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 422         }
 423 out:
 424         if (ret & ~PAGE_MASK)
 425                 vm_unacct_memory(charged);
 426 out_nc:
 427         return ret;
 428 }
 429
 430 asmlinkage unsigned long sys_mremap(unsigned long addr,
 431         unsigned long old_len, unsigned long new_len,
 432         unsigned long flags, unsigned long new_addr)
 433 {
 434         unsigned long ret;
 435
 436         down_write(&current->mm->mmap_sem);
 437         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 438         up_write(&current->mm->mmap_sem);
 439         return ret;
 440 }