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/shm.h>
  13 #include <linux/ksm.h>
  14 #include <linux/mman.h>
  15 #include <linux/swap.h>
  16 #include <linux/capability.h>
  17 #include <linux/fs.h>
  18 #include <linux/swapops.h>
  19 #include <linux/highmem.h>
  20 #include <linux/security.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/mmu_notifier.h>
  23 #include <linux/uaccess.h>
  24 #include <linux/mm-arch-hooks.h>
  25
  26 #include <asm/cacheflush.h>
  27 #include <asm/tlbflush.h>
  28
  29 #include "internal.h"
  30
  31 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  32 {
  33         pgd_t *pgd;
  34         pud_t *pud;
  35         pmd_t *pmd;
  36
  37         pgd = pgd_offset(mm, addr);
  38         if (pgd_none_or_clear_bad(pgd))
  39                 return NULL;
  40
  41         pud = pud_offset(pgd, addr);
  42         if (pud_none_or_clear_bad(pud))
  43                 return NULL;
  44
  45         pmd = pmd_offset(pud, addr);
  46         if (pmd_none(*pmd))
  47                 return NULL;
  48
  49         return pmd;
  50 }
  51
  52 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
  53                             unsigned long addr)
  54 {
  55         pgd_t *pgd;
  56         pud_t *pud;
  57         pmd_t *pmd;
  58
  59         pgd = pgd_offset(mm, addr);
  60         pud = pud_alloc(mm, pgd, addr);
  61         if (!pud)
  62                 return NULL;
  63
  64         pmd = pmd_alloc(mm, pud, addr);
  65         if (!pmd)
  66                 return NULL;
  67
  68         VM_BUG_ON(pmd_trans_huge(*pmd));
  69
  70         return pmd;
  71 }
  72
  73 static void take_rmap_locks(struct vm_area_struct *vma)
  74 {
  75         if (vma->vm_file)
  76                 i_mmap_lock_write(vma->vm_file->f_mapping);
  77         if (vma->anon_vma)
  78                 anon_vma_lock_write(vma->anon_vma);
  79 }
  80
  81 static void drop_rmap_locks(struct vm_area_struct *vma)
  82 {
  83         if (vma->anon_vma)
  84                 anon_vma_unlock_write(vma->anon_vma);
  85         if (vma->vm_file)
  86                 i_mmap_unlock_write(vma->vm_file->f_mapping);
  87 }
  88
  89 static pte_t move_soft_dirty_pte(pte_t pte)
  90 {
  91         /*
  92          * Set soft dirty bit so we can notice
  93          * in userspace the ptes were moved.
  94          */
  95 #ifdef CONFIG_MEM_SOFT_DIRTY
  96         if (pte_present(pte))
  97                 pte = pte_mksoft_dirty(pte);
  98         else if (is_swap_pte(pte))
  99                 pte = pte_swp_mksoft_dirty(pte);
 100 #endif
 101         return pte;
 102 }
 103
 104 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
 105                 unsigned long old_addr, unsigned long old_end,
 106                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
 107                 unsigned long new_addr, bool need_rmap_locks, bool *need_flush)
 108 {
 109         struct mm_struct *mm = vma->vm_mm;
 110         pte_t *old_pte, *new_pte, pte;
 111         spinlock_t *old_ptl, *new_ptl;
 112         bool force_flush = false;
 113         unsigned long len = old_end - old_addr;
 114
 115         /*
 116          * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
 117          * locks to ensure that rmap will always observe either the old or the
 118          * new ptes. This is the easiest way to avoid races with
 119          * truncate_pagecache(), page migration, etc...
 120          *
 121          * When need_rmap_locks is false, we use other ways to avoid
 122          * such races:
 123          *
 124          * - During exec() shift_arg_pages(), we use a specially tagged vma
 125          *   which rmap call sites look for using is_vma_temporary_stack().
 126          *
 127          * - During mremap(), new_vma is often known to be placed after vma
 128          *   in rmap traversal order. This ensures rmap will always observe
 129          *   either the old pte, or the new pte, or both (the page table locks
 130          *   serialize access to individual ptes, but only rmap traversal
 131          *   order guarantees that we won't miss both the old and new ptes).
 132          */
 133         if (need_rmap_locks)
 134                 take_rmap_locks(vma);
 135
 136         /*
 137          * We don't have to worry about the ordering of src and dst
 138          * pte locks because exclusive mmap_sem prevents deadlock.
 139          */
 140         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 141         new_pte = pte_offset_map(new_pmd, new_addr);
 142         new_ptl = pte_lockptr(mm, new_pmd);
 143         if (new_ptl != old_ptl)
 144                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 145         arch_enter_lazy_mmu_mode();
 146
 147         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 148                                    new_pte++, new_addr += PAGE_SIZE) {
 149                 if (pte_none(*old_pte))
 150                         continue;
 151
 152                 /*
 153                  * We are remapping a dirty PTE, make sure to
 154                  * flush TLB before we drop the PTL for the
 155                  * old PTE or we may race with page_mkclean().
 156                  */
 157                 if (pte_present(*old_pte) && pte_dirty(*old_pte))
 158                         force_flush = true;
 159                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
 160                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 161                 pte = move_soft_dirty_pte(pte);
 162                 set_pte_at(mm, new_addr, new_pte, pte);
 163         }
 164
 165         arch_leave_lazy_mmu_mode();
 166         if (new_ptl != old_ptl)
 167                 spin_unlock(new_ptl);
 168         pte_unmap(new_pte - 1);
 169         if (force_flush)
 170                 flush_tlb_range(vma, old_end - len, old_end);
 171         else
 172                 *need_flush = true;
 173         pte_unmap_unlock(old_pte - 1, old_ptl);
 174         if (need_rmap_locks)
 175                 drop_rmap_locks(vma);
 176 }
 177
 178 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 179
 180 unsigned long move_page_tables(struct vm_area_struct *vma,
 181                 unsigned long old_addr, struct vm_area_struct *new_vma,
 182                 unsigned long new_addr, unsigned long len,
 183                 bool need_rmap_locks)
 184 {
 185         unsigned long extent, next, old_end;
 186         pmd_t *old_pmd, *new_pmd;
 187         bool need_flush = false;
 188         unsigned long mmun_start;       /* For mmu_notifiers */
 189         unsigned long mmun_end;         /* For mmu_notifiers */
 190
 191         old_end = old_addr + len;
 192         flush_cache_range(vma, old_addr, old_end);
 193
 194         mmun_start = old_addr;
 195         mmun_end   = old_end;
 196         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 197
 198         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 199                 cond_resched();
 200                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 201                 /* even if next overflowed, extent below will be ok */
 202                 extent = next - old_addr;
 203                 if (extent > old_end - old_addr)
 204                         extent = old_end - old_addr;
 205                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 206                 if (!old_pmd)
 207                         continue;
 208                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 209                 if (!new_pmd)
 210                         break;
 211                 if (pmd_trans_huge(*old_pmd)) {
 212                         if (extent == HPAGE_PMD_SIZE) {
 213                                 bool moved;
 214                                 /* See comment in move_ptes() */
 215                                 if (need_rmap_locks)
 216                                         take_rmap_locks(vma);
 217                                 moved = move_huge_pmd(vma, old_addr, new_addr,
 218                                                     old_end, old_pmd, new_pmd,
 219                                                     &need_flush);
 220                                 if (need_rmap_locks)
 221                                         drop_rmap_locks(vma);
 222                                 if (moved)
 223                                         continue;
 224                         }
 225                         split_huge_pmd(vma, old_pmd, old_addr);
 226                         if (pmd_trans_unstable(old_pmd))
 227                                 continue;
 228                 }
 229                 if (pte_alloc(new_vma->vm_mm, new_pmd, new_addr))
 230                         break;
 231                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 232                 if (extent > next - new_addr)
 233                         extent = next - new_addr;
 234                 if (extent > LATENCY_LIMIT)
 235                         extent = LATENCY_LIMIT;
 236                 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
 237                           new_pmd, new_addr, need_rmap_locks, &need_flush);
 238         }
 239         if (need_flush)
 240                 flush_tlb_range(vma, old_end-len, old_addr);
 241
 242         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 243
 244         return len + old_addr - old_end;        /* how much done */
 245 }
 246
 247 static unsigned long move_vma(struct vm_area_struct *vma,
 248                 unsigned long old_addr, unsigned long old_len,
 249                 unsigned long new_len, unsigned long new_addr, bool *locked)
 250 {
 251         struct mm_struct *mm = vma->vm_mm;
 252         struct vm_area_struct *new_vma;
 253         unsigned long vm_flags = vma->vm_flags;
 254         unsigned long new_pgoff;
 255         unsigned long moved_len;
 256         unsigned long excess = 0;
 257         unsigned long hiwater_vm;
 258         int split = 0;
 259         int err;
 260         bool need_rmap_locks;
 261
 262         /*
 263          * We'd prefer to avoid failure later on in do_munmap:
 264          * which may split one vma into three before unmapping.
 265          */
 266         if (mm->map_count >= sysctl_max_map_count - 3)
 267                 return -ENOMEM;
 268
 269         /*
 270          * Advise KSM to break any KSM pages in the area to be moved:
 271          * it would be confusing if they were to turn up at the new
 272          * location, where they happen to coincide with different KSM
 273          * pages recently unmapped.  But leave vma->vm_flags as it was,
 274          * so KSM can come around to merge on vma and new_vma afterwards.
 275          */
 276         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 277                                                 MADV_UNMERGEABLE, &vm_flags);
 278         if (err)
 279                 return err;
 280
 281         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 282         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 283                            &need_rmap_locks);
 284         if (!new_vma)
 285                 return -ENOMEM;
 286
 287         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 288                                      need_rmap_locks);
 289         if (moved_len < old_len) {
 290                 err = -ENOMEM;
 291         } else if (vma->vm_ops && vma->vm_ops->mremap) {
 292                 err = vma->vm_ops->mremap(new_vma);
 293         }
 294
 295         if (unlikely(err)) {
 296                 /*
 297                  * On error, move entries back from new area to old,
 298                  * which will succeed since page tables still there,
 299                  * and then proceed to unmap new area instead of old.
 300                  */
 301                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 302                                  true);
 303                 vma = new_vma;
 304                 old_len = new_len;
 305                 old_addr = new_addr;
 306                 new_addr = err;
 307         } else {
 308                 arch_remap(mm, old_addr, old_addr + old_len,
 309                            new_addr, new_addr + new_len);
 310         }
 311
 312         /* Conceal VM_ACCOUNT so old reservation is not undone */
 313         if (vm_flags & VM_ACCOUNT) {
 314                 vma->vm_flags &= ~VM_ACCOUNT;
 315                 excess = vma->vm_end - vma->vm_start - old_len;
 316                 if (old_addr > vma->vm_start &&
 317                     old_addr + old_len < vma->vm_end)
 318                         split = 1;
 319         }
 320
 321         /*
 322          * If we failed to move page tables we still do total_vm increment
 323          * since do_munmap() will decrement it by old_len == new_len.
 324          *
 325          * Since total_vm is about to be raised artificially high for a
 326          * moment, we need to restore high watermark afterwards: if stats
 327          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 328          * If this were a serious issue, we'd add a flag to do_munmap().
 329          */
 330         hiwater_vm = mm->hiwater_vm;
 331         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 332
 333         /* Tell pfnmap has moved from this vma */
 334         if (unlikely(vma->vm_flags & VM_PFNMAP))
 335                 untrack_pfn_moved(vma);
 336
 337         if (do_munmap(mm, old_addr, old_len) < 0) {
 338                 /* OOM: unable to split vma, just get accounts right */
 339                 vm_unacct_memory(excess >> PAGE_SHIFT);
 340                 excess = 0;
 341         }
 342         mm->hiwater_vm = hiwater_vm;
 343
 344         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 345         if (excess) {
 346                 vma->vm_flags |= VM_ACCOUNT;
 347                 if (split)
 348                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 349         }
 350
 351         if (vm_flags & VM_LOCKED) {
 352                 mm->locked_vm += new_len >> PAGE_SHIFT;
 353                 *locked = true;
 354         }
 355
 356         return new_addr;
 357 }
 358
 359 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 360         unsigned long old_len, unsigned long new_len, unsigned long *p)
 361 {
 362         struct mm_struct *mm = current->mm;
 363         struct vm_area_struct *vma = find_vma(mm, addr);
 364         unsigned long pgoff;
 365
 366         if (!vma || vma->vm_start > addr)
 367                 return ERR_PTR(-EFAULT);
 368
 369         if (is_vm_hugetlb_page(vma))
 370                 return ERR_PTR(-EINVAL);
 371
 372         /* We can't remap across vm area boundaries */
 373         if (old_len > vma->vm_end - addr)
 374                 return ERR_PTR(-EFAULT);
 375
 376         if (new_len == old_len)
 377                 return vma;
 378
 379         /* Need to be careful about a growing mapping */
 380         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 381         pgoff += vma->vm_pgoff;
 382         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 383                 return ERR_PTR(-EINVAL);
 384
 385         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 386                 return ERR_PTR(-EFAULT);
 387
 388         if (vma->vm_flags & VM_LOCKED) {
 389                 unsigned long locked, lock_limit;
 390                 locked = mm->locked_vm << PAGE_SHIFT;
 391                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 392                 locked += new_len - old_len;
 393                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 394                         return ERR_PTR(-EAGAIN);
 395         }
 396
 397         if (!may_expand_vm(mm, vma->vm_flags,
 398                                 (new_len - old_len) >> PAGE_SHIFT))
 399                 return ERR_PTR(-ENOMEM);
 400
 401         if (vma->vm_flags & VM_ACCOUNT) {
 402                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 403                 if (security_vm_enough_memory_mm(mm, charged))
 404                         return ERR_PTR(-ENOMEM);
 405                 *p = charged;
 406         }
 407
 408         return vma;
 409 }
 410
 411 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 412                 unsigned long new_addr, unsigned long new_len, bool *locked)
 413 {
 414         struct mm_struct *mm = current->mm;
 415         struct vm_area_struct *vma;
 416         unsigned long ret = -EINVAL;
 417         unsigned long charged = 0;
 418         unsigned long map_flags;
 419
 420         if (offset_in_page(new_addr))
 421                 goto out;
 422
 423         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 424                 goto out;
 425
 426         /* Ensure the old/new locations do not overlap */
 427         if (addr + old_len > new_addr && new_addr + new_len > addr)
 428                 goto out;
 429
 430         ret = do_munmap(mm, new_addr, new_len);
 431         if (ret)
 432                 goto out;
 433
 434         if (old_len >= new_len) {
 435                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 436                 if (ret && old_len != new_len)
 437                         goto out;
 438                 old_len = new_len;
 439         }
 440
 441         vma = vma_to_resize(addr, old_len, new_len, &charged);
 442         if (IS_ERR(vma)) {
 443                 ret = PTR_ERR(vma);
 444                 goto out;
 445         }
 446
 447         map_flags = MAP_FIXED;
 448         if (vma->vm_flags & VM_MAYSHARE)
 449                 map_flags |= MAP_SHARED;
 450
 451         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 452                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 453                                 map_flags);
 454         if (offset_in_page(ret))
 455                 goto out1;
 456
 457         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
 458         if (!(offset_in_page(ret)))
 459                 goto out;
 460 out1:
 461         vm_unacct_memory(charged);
 462
 463 out:
 464         return ret;
 465 }
 466
 467 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 468 {
 469         unsigned long end = vma->vm_end + delta;
 470         if (end < vma->vm_end) /* overflow */
 471                 return 0;
 472         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 473                 return 0;
 474         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 475                               0, MAP_FIXED) & ~PAGE_MASK)
 476                 return 0;
 477         return 1;
 478 }
 479
 480 /*
 481  * Expand (or shrink) an existing mapping, potentially moving it at the
 482  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 483  *
 484  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 485  * This option implies MREMAP_MAYMOVE.
 486  */
 487 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 488                 unsigned long, new_len, unsigned long, flags,
 489                 unsigned long, new_addr)
 490 {
 491         struct mm_struct *mm = current->mm;
 492         struct vm_area_struct *vma;
 493         unsigned long ret = -EINVAL;
 494         unsigned long charged = 0;
 495         bool locked = false;
 496
 497         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 498                 return ret;
 499
 500         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 501                 return ret;
 502
 503         if (offset_in_page(addr))
 504                 return ret;
 505
 506         old_len = PAGE_ALIGN(old_len);
 507         new_len = PAGE_ALIGN(new_len);
 508
 509         /*
 510          * We allow a zero old-len as a special case
 511          * for DOS-emu "duplicate shm area" thing. But
 512          * a zero new-len is nonsensical.
 513          */
 514         if (!new_len)
 515                 return ret;
 516
 517         if (down_write_killable(&current->mm->mmap_sem))
 518                 return -EINTR;
 519
 520         if (flags & MREMAP_FIXED) {
 521                 ret = mremap_to(addr, old_len, new_addr, new_len,
 522                                 &locked);
 523                 goto out;
 524         }
 525
 526         /*
 527          * Always allow a shrinking remap: that just unmaps
 528          * the unnecessary pages..
 529          * do_munmap does all the needed commit accounting
 530          */
 531         if (old_len >= new_len) {
 532                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 533                 if (ret && old_len != new_len)
 534                         goto out;
 535                 ret = addr;
 536                 goto out;
 537         }
 538
 539         /*
 540          * Ok, we need to grow..
 541          */
 542         vma = vma_to_resize(addr, old_len, new_len, &charged);
 543         if (IS_ERR(vma)) {
 544                 ret = PTR_ERR(vma);
 545                 goto out;
 546         }
 547
 548         /* old_len exactly to the end of the area..
 549          */
 550         if (old_len == vma->vm_end - addr) {
 551                 /* can we just expand the current mapping? */
 552                 if (vma_expandable(vma, new_len - old_len)) {
 553                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 554
 555                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 556                                        vma->vm_pgoff, NULL)) {
 557                                 ret = -ENOMEM;
 558                                 goto out;
 559                         }
 560
 561                         vm_stat_account(mm, vma->vm_flags, pages);
 562                         if (vma->vm_flags & VM_LOCKED) {
 563                                 mm->locked_vm += pages;
 564                                 locked = true;
 565                                 new_addr = addr;
 566                         }
 567                         ret = addr;
 568                         goto out;
 569                 }
 570         }
 571
 572         /*
 573          * We weren't able to just expand or shrink the area,
 574          * we need to create a new one and move it..
 575          */
 576         ret = -ENOMEM;
 577         if (flags & MREMAP_MAYMOVE) {
 578                 unsigned long map_flags = 0;
 579                 if (vma->vm_flags & VM_MAYSHARE)
 580                         map_flags |= MAP_SHARED;
 581
 582                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 583                                         vma->vm_pgoff +
 584                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 585                                         map_flags);
 586                 if (offset_in_page(new_addr)) {
 587                         ret = new_addr;
 588                         goto out;
 589                 }
 590
 591                 ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
 592         }
 593 out:
 594         if (offset_in_page(ret)) {
 595                 vm_unacct_memory(charged);
 596                 locked = 0;
 597         }
 598         up_write(&current->mm->mmap_sem);
 599         if (locked && new_len > old_len)
 600                 mm_populate(new_addr + old_len, new_len - old_len);
 601         return ret;
 602 }