SAFE public projects git trees. - safe/jmp/linux-2.6/blob - mm/mremap.c

   1 /*
   2  *      mm/mremap.c
   3  *
   4  *      (C) Copyright 1996 Linus Torvalds
   5  *
   6  *      Address space accounting code   <alan@lxorguk.ukuu.org.uk>
   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/ksm.h>
  15 #include <linux/mman.h>
  16 #include <linux/swap.h>
  17 #include <linux/capability.h>
  18 #include <linux/fs.h>
  19 #include <linux/highmem.h>
  20 #include <linux/security.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/mmu_notifier.h>
  23
  24 #include <asm/uaccess.h>
  25 #include <asm/cacheflush.h>
  26 #include <asm/tlbflush.h>
  27
  28 #include "internal.h"
  29
  30 #ifndef arch_mmap_check
  31 #define arch_mmap_check(addr, len, flags)       (0)
  32 #endif
  33
  34 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  35 {
  36         pgd_t *pgd;
  37         pud_t *pud;
  38         pmd_t *pmd;
  39
  40         pgd = pgd_offset(mm, addr);
  41         if (pgd_none_or_clear_bad(pgd))
  42                 return NULL;
  43
  44         pud = pud_offset(pgd, addr);
  45         if (pud_none_or_clear_bad(pud))
  46                 return NULL;
  47
  48         pmd = pmd_offset(pud, addr);
  49         if (pmd_none_or_clear_bad(pmd))
  50                 return NULL;
  51
  52         return pmd;
  53 }
  54
  55 static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
  56 {
  57         pgd_t *pgd;
  58         pud_t *pud;
  59         pmd_t *pmd;
  60
  61         pgd = pgd_offset(mm, addr);
  62         pud = pud_alloc(mm, pgd, addr);
  63         if (!pud)
  64                 return NULL;
  65
  66         pmd = pmd_alloc(mm, pud, addr);
  67         if (!pmd)
  68                 return NULL;
  69
  70         if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
  71                 return NULL;
  72
  73         return pmd;
  74 }
  75
  76 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
  77                 unsigned long old_addr, unsigned long old_end,
  78                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
  79                 unsigned long new_addr)
  80 {
  81         struct address_space *mapping = NULL;
  82         struct mm_struct *mm = vma->vm_mm;
  83         pte_t *old_pte, *new_pte, pte;
  84         spinlock_t *old_ptl, *new_ptl;
  85         unsigned long old_start;
  86
  87         old_start = old_addr;
  88         mmu_notifier_invalidate_range_start(vma->vm_mm,
  89                                             old_start, old_end);
  90         if (vma->vm_file) {
  91                 /*
  92                  * Subtle point from Rajesh Venkatasubramanian: before
  93                  * moving file-based ptes, we must lock truncate_pagecache
  94                  * out, since it might clean the dst vma before the src vma,
  95                  * and we propagate stale pages into the dst afterward.
  96                  */
  97                 mapping = vma->vm_file->f_mapping;
  98                 spin_lock(&mapping->i_mmap_lock);
  99                 if (new_vma->vm_truncate_count &&
 100                     new_vma->vm_truncate_count != vma->vm_truncate_count)
 101                         new_vma->vm_truncate_count = 0;
 102         }
 103
 104         /*
 105          * We don't have to worry about the ordering of src and dst
 106          * pte locks because exclusive mmap_sem prevents deadlock.
 107          */
 108         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 109         new_pte = pte_offset_map_nested(new_pmd, new_addr);
 110         new_ptl = pte_lockptr(mm, new_pmd);
 111         if (new_ptl != old_ptl)
 112                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 113         arch_enter_lazy_mmu_mode();
 114
 115         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 116                                    new_pte++, new_addr += PAGE_SIZE) {
 117                 if (pte_none(*old_pte))
 118                         continue;
 119                 pte = ptep_clear_flush(vma, old_addr, old_pte);
 120                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 121                 set_pte_at(mm, new_addr, new_pte, pte);
 122         }
 123
 124         arch_leave_lazy_mmu_mode();
 125         if (new_ptl != old_ptl)
 126                 spin_unlock(new_ptl);
 127         pte_unmap_nested(new_pte - 1);
 128         pte_unmap_unlock(old_pte - 1, old_ptl);
 129         if (mapping)
 130                 spin_unlock(&mapping->i_mmap_lock);
 131         mmu_notifier_invalidate_range_end(vma->vm_mm, old_start, old_end);
 132 }
 133
 134 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 135
 136 unsigned long move_page_tables(struct vm_area_struct *vma,
 137                 unsigned long old_addr, struct vm_area_struct *new_vma,
 138                 unsigned long new_addr, unsigned long len)
 139 {
 140         unsigned long extent, next, old_end;
 141         pmd_t *old_pmd, *new_pmd;
 142
 143         old_end = old_addr + len;
 144         flush_cache_range(vma, old_addr, old_end);
 145
 146         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 147                 cond_resched();
 148                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 149                 if (next - 1 > old_end)
 150                         next = old_end;
 151                 extent = next - old_addr;
 152                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 153                 if (!old_pmd)
 154                         continue;
 155                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
 156                 if (!new_pmd)
 157                         break;
 158                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 159                 if (extent > next - new_addr)
 160                         extent = next - new_addr;
 161                 if (extent > LATENCY_LIMIT)
 162                         extent = LATENCY_LIMIT;
 163                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 164                                 new_vma, new_pmd, new_addr);
 165         }
 166
 167         return len + old_addr - old_end;        /* how much done */
 168 }
 169
 170 static unsigned long move_vma(struct vm_area_struct *vma,
 171                 unsigned long old_addr, unsigned long old_len,
 172                 unsigned long new_len, unsigned long new_addr)
 173 {
 174         struct mm_struct *mm = vma->vm_mm;
 175         struct vm_area_struct *new_vma;
 176         unsigned long vm_flags = vma->vm_flags;
 177         unsigned long new_pgoff;
 178         unsigned long moved_len;
 179         unsigned long excess = 0;
 180         unsigned long hiwater_vm;
 181         int split = 0;
 182         int err;
 183
 184         /*
 185          * We'd prefer to avoid failure later on in do_munmap:
 186          * which may split one vma into three before unmapping.
 187          */
 188         if (mm->map_count >= sysctl_max_map_count - 3)
 189                 return -ENOMEM;
 190
 191         /*
 192          * Advise KSM to break any KSM pages in the area to be moved:
 193          * it would be confusing if they were to turn up at the new
 194          * location, where they happen to coincide with different KSM
 195          * pages recently unmapped.  But leave vma->vm_flags as it was,
 196          * so KSM can come around to merge on vma and new_vma afterwards.
 197          */
 198         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 199                                                 MADV_UNMERGEABLE, &vm_flags);
 200         if (err)
 201                 return err;
 202
 203         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 204         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 205         if (!new_vma)
 206                 return -ENOMEM;
 207
 208         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
 209         if (moved_len < old_len) {
 210                 /*
 211                  * On error, move entries back from new area to old,
 212                  * which will succeed since page tables still there,
 213                  * and then proceed to unmap new area instead of old.
 214                  */
 215                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
 216                 vma = new_vma;
 217                 old_len = new_len;
 218                 old_addr = new_addr;
 219                 new_addr = -ENOMEM;
 220         }
 221
 222         /* Conceal VM_ACCOUNT so old reservation is not undone */
 223         if (vm_flags & VM_ACCOUNT) {
 224                 vma->vm_flags &= ~VM_ACCOUNT;
 225                 excess = vma->vm_end - vma->vm_start - old_len;
 226                 if (old_addr > vma->vm_start &&
 227                     old_addr + old_len < vma->vm_end)
 228                         split = 1;
 229         }
 230
 231         /*
 232          * If we failed to move page tables we still do total_vm increment
 233          * since do_munmap() will decrement it by old_len == new_len.
 234          *
 235          * Since total_vm is about to be raised artificially high for a
 236          * moment, we need to restore high watermark afterwards: if stats
 237          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 238          * If this were a serious issue, we'd add a flag to do_munmap().
 239          */
 240         hiwater_vm = mm->hiwater_vm;
 241         mm->total_vm += new_len >> PAGE_SHIFT;
 242         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 243
 244         if (do_munmap(mm, old_addr, old_len) < 0) {
 245                 /* OOM: unable to split vma, just get accounts right */
 246                 vm_unacct_memory(excess >> PAGE_SHIFT);
 247                 excess = 0;
 248         }
 249         mm->hiwater_vm = hiwater_vm;
 250
 251         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 252         if (excess) {
 253                 vma->vm_flags |= VM_ACCOUNT;
 254                 if (split)
 255                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 256         }
 257
 258         if (vm_flags & VM_LOCKED) {
 259                 mm->locked_vm += new_len >> PAGE_SHIFT;
 260                 if (new_len > old_len)
 261                         mlock_vma_pages_range(new_vma, new_addr + old_len,
 262                                                        new_addr + new_len);
 263         }
 264
 265         return new_addr;
 266 }
 267
 268 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 269         unsigned long old_len, unsigned long new_len, unsigned long *p)
 270 {
 271         struct mm_struct *mm = current->mm;
 272         struct vm_area_struct *vma = find_vma(mm, addr);
 273
 274         if (!vma || vma->vm_start > addr)
 275                 goto Efault;
 276
 277         if (is_vm_hugetlb_page(vma))
 278                 goto Einval;
 279
 280         /* We can't remap across vm area boundaries */
 281         if (old_len > vma->vm_end - addr)
 282                 goto Efault;
 283
 284         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)) {
 285                 if (new_len > old_len)
 286                         goto Efault;
 287         }
 288
 289         if (vma->vm_flags & VM_LOCKED) {
 290                 unsigned long locked, lock_limit;
 291                 locked = mm->locked_vm << PAGE_SHIFT;
 292                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 293                 locked += new_len - old_len;
 294                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 295                         goto Eagain;
 296         }
 297
 298         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
 299                 goto Enomem;
 300
 301         if (vma->vm_flags & VM_ACCOUNT) {
 302                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 303                 if (security_vm_enough_memory(charged))
 304                         goto Efault;
 305                 *p = charged;
 306         }
 307
 308         return vma;
 309
 310 Efault: /* very odd choice for most of the cases, but... */
 311         return ERR_PTR(-EFAULT);
 312 Einval:
 313         return ERR_PTR(-EINVAL);
 314 Enomem:
 315         return ERR_PTR(-ENOMEM);
 316 Eagain:
 317         return ERR_PTR(-EAGAIN);
 318 }
 319
 320 static unsigned long mremap_to(unsigned long addr,
 321         unsigned long old_len, unsigned long new_addr,
 322         unsigned long new_len)
 323 {
 324         struct mm_struct *mm = current->mm;
 325         struct vm_area_struct *vma;
 326         unsigned long ret = -EINVAL;
 327         unsigned long charged = 0;
 328         unsigned long map_flags;
 329
 330         if (new_addr & ~PAGE_MASK)
 331                 goto out;
 332
 333         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 334                 goto out;
 335
 336         /* Check if the location we're moving into overlaps the
 337          * old location at all, and fail if it does.
 338          */
 339         if ((new_addr <= addr) && (new_addr+new_len) > addr)
 340                 goto out;
 341
 342         if ((addr <= new_addr) && (addr+old_len) > new_addr)
 343                 goto out;
 344
 345         ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
 346         if (ret)
 347                 goto out;
 348
 349         ret = do_munmap(mm, new_addr, new_len);
 350         if (ret)
 351                 goto out;
 352
 353         if (old_len >= new_len) {
 354                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 355                 if (ret && old_len != new_len)
 356                         goto out;
 357                 old_len = new_len;
 358         }
 359
 360         vma = vma_to_resize(addr, old_len, new_len, &charged);
 361         if (IS_ERR(vma)) {
 362                 ret = PTR_ERR(vma);
 363                 goto out;
 364         }
 365
 366         map_flags = MAP_FIXED;
 367         if (vma->vm_flags & VM_MAYSHARE)
 368                 map_flags |= MAP_SHARED;
 369         ret = arch_mmap_check(new_addr, new_len, map_flags);
 370         if (ret)
 371                 goto out1;
 372         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 373                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 374                                 map_flags);
 375         if (ret & ~PAGE_MASK)
 376                 goto out1;
 377
 378         ret = move_vma(vma, addr, old_len, new_len, new_addr);
 379         if (!(ret & ~PAGE_MASK))
 380                 goto out;
 381 out1:
 382         vm_unacct_memory(charged);
 383
 384 out:
 385         return ret;
 386 }
 387
 388 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 389 {
 390         unsigned long end = vma->vm_end + delta;
 391         unsigned long max_addr = TASK_SIZE;
 392         if (vma->vm_next)
 393                 max_addr = vma->vm_next->vm_start;
 394         if (max_addr < end || end < vma->vm_end)
 395                 return 0;
 396         if (arch_mmap_check(vma->vm_start, end - vma->vm_start, MAP_FIXED))
 397                 return 0;
 398         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 399                               0, MAP_FIXED) & ~PAGE_MASK)
 400                 return 0;
 401         return 1;
 402 }
 403
 404 /*
 405  * Expand (or shrink) an existing mapping, potentially moving it at the
 406  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 407  *
 408  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 409  * This option implies MREMAP_MAYMOVE.
 410  */
 411 unsigned long do_mremap(unsigned long addr,
 412         unsigned long old_len, unsigned long new_len,
 413         unsigned long flags, unsigned long new_addr)
 414 {
 415         struct mm_struct *mm = current->mm;
 416         struct vm_area_struct *vma;
 417         unsigned long ret = -EINVAL;
 418         unsigned long charged = 0;
 419
 420         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 421                 goto out;
 422
 423         if (addr & ~PAGE_MASK)
 424                 goto out;
 425
 426         old_len = PAGE_ALIGN(old_len);
 427         new_len = PAGE_ALIGN(new_len);
 428
 429         /*
 430          * We allow a zero old-len as a special case
 431          * for DOS-emu "duplicate shm area" thing. But
 432          * a zero new-len is nonsensical.
 433          */
 434         if (!new_len)
 435                 goto out;
 436
 437         if (flags & MREMAP_FIXED) {
 438                 if (flags & MREMAP_MAYMOVE)
 439                         ret = mremap_to(addr, old_len, new_addr, new_len);
 440                 goto out;
 441         }
 442
 443         /*
 444          * Always allow a shrinking remap: that just unmaps
 445          * the unnecessary pages..
 446          * do_munmap does all the needed commit accounting
 447          */
 448         if (old_len >= new_len) {
 449                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 450                 if (ret && old_len != new_len)
 451                         goto out;
 452                 ret = addr;
 453                 goto out;
 454         }
 455
 456         /*
 457          * Ok, we need to grow..
 458          */
 459         vma = vma_to_resize(addr, old_len, new_len, &charged);
 460         if (IS_ERR(vma)) {
 461                 ret = PTR_ERR(vma);
 462                 goto out;
 463         }
 464
 465         /* old_len exactly to the end of the area..
 466          */
 467         if (old_len == vma->vm_end - addr) {
 468                 /* can we just expand the current mapping? */
 469                 if (vma_expandable(vma, new_len - old_len)) {
 470                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 471
 472                         vma_adjust(vma, vma->vm_start,
 473                                 addr + new_len, vma->vm_pgoff, NULL);
 474
 475                         mm->total_vm += pages;
 476                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
 477                         if (vma->vm_flags & VM_LOCKED) {
 478                                 mm->locked_vm += pages;
 479                                 mlock_vma_pages_range(vma, addr + old_len,
 480                                                    addr + new_len);
 481                         }
 482                         ret = addr;
 483                         goto out;
 484                 }
 485         }
 486
 487         /*
 488          * We weren't able to just expand or shrink the area,
 489          * we need to create a new one and move it..
 490          */
 491         ret = -ENOMEM;
 492         if (flags & MREMAP_MAYMOVE) {
 493                 unsigned long map_flags = 0;
 494                 if (vma->vm_flags & VM_MAYSHARE)
 495                         map_flags |= MAP_SHARED;
 496
 497                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 498                                         vma->vm_pgoff +
 499                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 500                                         map_flags);
 501                 if (new_addr & ~PAGE_MASK) {
 502                         ret = new_addr;
 503                         goto out;
 504                 }
 505
 506                 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
 507                 if (ret)
 508                         goto out;
 509                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 510         }
 511 out:
 512         if (ret & ~PAGE_MASK)
 513                 vm_unacct_memory(charged);
 514         return ret;
 515 }
 516
 517 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 518                 unsigned long, new_len, unsigned long, flags,
 519                 unsigned long, new_addr)
 520 {
 521         unsigned long ret;
 522
 523         down_write(&current->mm->mmap_sem);
 524         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 525         up_write(&current->mm->mmap_sem);
 526         return ret;
 527 }