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                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
 153                 /*
 154                  * If we are remapping a dirty PTE, make sure
 155                  * to flush TLB before we drop the PTL for the
 156                  * old PTE or we may race with page_mkclean().
 157                  *
 158                  * This check has to be done after we removed the
 159                  * old PTE from page tables or another thread may
 160                  * dirty it after the check and before the removal.
 161                  */
 162                 if (pte_present(pte) && pte_dirty(pte))
 163                         force_flush = true;
 164                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 165                 pte = move_soft_dirty_pte(pte);
 166                 set_pte_at(mm, new_addr, new_pte, pte);
 167         }
 168
 169         arch_leave_lazy_mmu_mode();
 170         if (new_ptl != old_ptl)
 171                 spin_unlock(new_ptl);
 172         pte_unmap(new_pte - 1);
 173         if (force_flush)
 174                 flush_tlb_range(vma, old_end - len, old_end);
 175         else
 176                 *need_flush = true;
 177         pte_unmap_unlock(old_pte - 1, old_ptl);
 178         if (need_rmap_locks)
 179                 drop_rmap_locks(vma);
 180 }
 181
 182 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 183
 184 unsigned long move_page_tables(struct vm_area_struct *vma,
 185                 unsigned long old_addr, struct vm_area_struct *new_vma,
 186                 unsigned long new_addr, unsigned long len,
 187                 bool need_rmap_locks)
 188 {
 189         unsigned long extent, next, old_end;
 190         pmd_t *old_pmd, *new_pmd;
 191         bool need_flush = false;
 192         unsigned long mmun_start;       /* For mmu_notifiers */
 193         unsigned long mmun_end;         /* For mmu_notifiers */
 194
 195         old_end = old_addr + len;
 196         flush_cache_range(vma, old_addr, old_end);
 197
 198         mmun_start = old_addr;
 199         mmun_end   = old_end;
 200         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 201
 202         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 203                 cond_resched();
 204                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 205                 /* even if next overflowed, extent below will be ok */
 206                 extent = next - old_addr;
 207                 if (extent > old_end - old_addr)
 208                         extent = old_end - old_addr;
 209                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 210                 if (!old_pmd)
 211                         continue;
 212                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 213                 if (!new_pmd)
 214                         break;
 215                 if (pmd_trans_huge(*old_pmd)) {
 216                         if (extent == HPAGE_PMD_SIZE) {
 217                                 bool moved;
 218                                 /* See comment in move_ptes() */
 219                                 if (need_rmap_locks)
 220                                         take_rmap_locks(vma);
 221                                 moved = move_huge_pmd(vma, old_addr, new_addr,
 222                                                     old_end, old_pmd, new_pmd,
 223                                                     &need_flush);
 224                                 if (need_rmap_locks)
 225                                         drop_rmap_locks(vma);
 226                                 if (moved)
 227                                         continue;
 228                         }
 229                         split_huge_pmd(vma, old_pmd, old_addr);
 230                         if (pmd_trans_unstable(old_pmd))
 231                                 continue;
 232                 }
 233                 if (pte_alloc(new_vma->vm_mm, new_pmd, new_addr))
 234                         break;
 235                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 236                 if (extent > next - new_addr)
 237                         extent = next - new_addr;
 238                 if (extent > LATENCY_LIMIT)
 239                         extent = LATENCY_LIMIT;
 240                 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
 241                           new_pmd, new_addr, need_rmap_locks, &need_flush);
 242         }
 243         if (need_flush)
 244                 flush_tlb_range(vma, old_end-len, old_addr);
 245
 246         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 247
 248         return len + old_addr - old_end;        /* how much done */
 249 }
 250
 251 static unsigned long move_vma(struct vm_area_struct *vma,
 252                 unsigned long old_addr, unsigned long old_len,
 253                 unsigned long new_len, unsigned long new_addr, bool *locked)
 254 {
 255         struct mm_struct *mm = vma->vm_mm;
 256         struct vm_area_struct *new_vma;
 257         unsigned long vm_flags = vma->vm_flags;
 258         unsigned long new_pgoff;
 259         unsigned long moved_len;
 260         unsigned long excess = 0;
 261         unsigned long hiwater_vm;
 262         int split = 0;
 263         int err;
 264         bool need_rmap_locks;
 265
 266         /*
 267          * We'd prefer to avoid failure later on in do_munmap:
 268          * which may split one vma into three before unmapping.
 269          */
 270         if (mm->map_count >= sysctl_max_map_count - 3)
 271                 return -ENOMEM;
 272
 273         /*
 274          * Advise KSM to break any KSM pages in the area to be moved:
 275          * it would be confusing if they were to turn up at the new
 276          * location, where they happen to coincide with different KSM
 277          * pages recently unmapped.  But leave vma->vm_flags as it was,
 278          * so KSM can come around to merge on vma and new_vma afterwards.
 279          */
 280         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 281                                                 MADV_UNMERGEABLE, &vm_flags);
 282         if (err)
 283                 return err;
 284
 285         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 286         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 287                            &need_rmap_locks);
 288         if (!new_vma)
 289                 return -ENOMEM;
 290
 291         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 292                                      need_rmap_locks);
 293         if (moved_len < old_len) {
 294                 err = -ENOMEM;
 295         } else if (vma->vm_ops && vma->vm_ops->mremap) {
 296                 err = vma->vm_ops->mremap(new_vma);
 297         }
 298
 299         if (unlikely(err)) {
 300                 /*
 301                  * On error, move entries back from new area to old,
 302                  * which will succeed since page tables still there,
 303                  * and then proceed to unmap new area instead of old.
 304                  */
 305                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 306                                  true);
 307                 vma = new_vma;
 308                 old_len = new_len;
 309                 old_addr = new_addr;
 310                 new_addr = err;
 311         } else {
 312                 arch_remap(mm, old_addr, old_addr + old_len,
 313                            new_addr, new_addr + new_len);
 314         }
 315
 316         /* Conceal VM_ACCOUNT so old reservation is not undone */
 317         if (vm_flags & VM_ACCOUNT) {
 318                 vma->vm_flags &= ~VM_ACCOUNT;
 319                 excess = vma->vm_end - vma->vm_start - old_len;
 320                 if (old_addr > vma->vm_start &&
 321                     old_addr + old_len < vma->vm_end)
 322                         split = 1;
 323         }
 324
 325         /*
 326          * If we failed to move page tables we still do total_vm increment
 327          * since do_munmap() will decrement it by old_len == new_len.
 328          *
 329          * Since total_vm is about to be raised artificially high for a
 330          * moment, we need to restore high watermark afterwards: if stats
 331          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 332          * If this were a serious issue, we'd add a flag to do_munmap().
 333          */
 334         hiwater_vm = mm->hiwater_vm;
 335         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 336
 337         /* Tell pfnmap has moved from this vma */
 338         if (unlikely(vma->vm_flags & VM_PFNMAP))
 339                 untrack_pfn_moved(vma);
 340
 341         if (do_munmap(mm, old_addr, old_len) < 0) {
 342                 /* OOM: unable to split vma, just get accounts right */
 343                 vm_unacct_memory(excess >> PAGE_SHIFT);
 344                 excess = 0;
 345         }
 346         mm->hiwater_vm = hiwater_vm;
 347
 348         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 349         if (excess) {
 350                 vma->vm_flags |= VM_ACCOUNT;
 351                 if (split)
 352                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 353         }
 354
 355         if (vm_flags & VM_LOCKED) {
 356                 mm->locked_vm += new_len >> PAGE_SHIFT;
 357                 *locked = true;
 358         }
 359
 360         return new_addr;
 361 }
 362
 363 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 364         unsigned long old_len, unsigned long new_len, unsigned long *p)
 365 {
 366         struct mm_struct *mm = current->mm;
 367         struct vm_area_struct *vma = find_vma(mm, addr);
 368         unsigned long pgoff;
 369
 370         if (!vma || vma->vm_start > addr)
 371                 return ERR_PTR(-EFAULT);
 372
 373         if (is_vm_hugetlb_page(vma))
 374                 return ERR_PTR(-EINVAL);
 375
 376         /* We can't remap across vm area boundaries */
 377         if (old_len > vma->vm_end - addr)
 378                 return ERR_PTR(-EFAULT);
 379
 380         if (new_len == old_len)
 381                 return vma;
 382
 383         /* Need to be careful about a growing mapping */
 384         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 385         pgoff += vma->vm_pgoff;
 386         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 387                 return ERR_PTR(-EINVAL);
 388
 389         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 390                 return ERR_PTR(-EFAULT);
 391
 392         if (vma->vm_flags & VM_LOCKED) {
 393                 unsigned long locked, lock_limit;
 394                 locked = mm->locked_vm << PAGE_SHIFT;
 395                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 396                 locked += new_len - old_len;
 397                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 398                         return ERR_PTR(-EAGAIN);
 399         }
 400
 401         if (!may_expand_vm(mm, vma->vm_flags,
 402                                 (new_len - old_len) >> PAGE_SHIFT))
 403                 return ERR_PTR(-ENOMEM);
 404
 405         if (vma->vm_flags & VM_ACCOUNT) {
 406                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 407                 if (security_vm_enough_memory_mm(mm, charged))
 408                         return ERR_PTR(-ENOMEM);
 409                 *p = charged;
 410         }
 411
 412         return vma;
 413 }
 414
 415 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 416                 unsigned long new_addr, unsigned long new_len, bool *locked)
 417 {
 418         struct mm_struct *mm = current->mm;
 419         struct vm_area_struct *vma;
 420         unsigned long ret = -EINVAL;
 421         unsigned long charged = 0;
 422         unsigned long map_flags;
 423
 424         if (offset_in_page(new_addr))
 425                 goto out;
 426
 427         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 428                 goto out;
 429
 430         /* Ensure the old/new locations do not overlap */
 431         if (addr + old_len > new_addr && new_addr + new_len > addr)
 432                 goto out;
 433
 434         ret = do_munmap(mm, new_addr, new_len);
 435         if (ret)
 436                 goto out;
 437
 438         if (old_len >= new_len) {
 439                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 440                 if (ret && old_len != new_len)
 441                         goto out;
 442                 old_len = new_len;
 443         }
 444
 445         vma = vma_to_resize(addr, old_len, new_len, &charged);
 446         if (IS_ERR(vma)) {
 447                 ret = PTR_ERR(vma);
 448                 goto out;
 449         }
 450
 451         map_flags = MAP_FIXED;
 452         if (vma->vm_flags & VM_MAYSHARE)
 453                 map_flags |= MAP_SHARED;
 454
 455         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 456                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 457                                 map_flags);
 458         if (offset_in_page(ret))
 459                 goto out1;
 460
 461         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
 462         if (!(offset_in_page(ret)))
 463                 goto out;
 464 out1:
 465         vm_unacct_memory(charged);
 466
 467 out:
 468         return ret;
 469 }
 470
 471 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 472 {
 473         unsigned long end = vma->vm_end + delta;
 474         if (end < vma->vm_end) /* overflow */
 475                 return 0;
 476         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 477                 return 0;
 478         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 479                               0, MAP_FIXED) & ~PAGE_MASK)
 480                 return 0;
 481         return 1;
 482 }
 483
 484 /*
 485  * Expand (or shrink) an existing mapping, potentially moving it at the
 486  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 487  *
 488  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 489  * This option implies MREMAP_MAYMOVE.
 490  */
 491 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 492                 unsigned long, new_len, unsigned long, flags,
 493                 unsigned long, new_addr)
 494 {
 495         struct mm_struct *mm = current->mm;
 496         struct vm_area_struct *vma;
 497         unsigned long ret = -EINVAL;
 498         unsigned long charged = 0;
 499         bool locked = false;
 500
 501         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 502                 return ret;
 503
 504         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 505                 return ret;
 506
 507         if (offset_in_page(addr))
 508                 return ret;
 509
 510         old_len = PAGE_ALIGN(old_len);
 511         new_len = PAGE_ALIGN(new_len);
 512
 513         /*
 514          * We allow a zero old-len as a special case
 515          * for DOS-emu "duplicate shm area" thing. But
 516          * a zero new-len is nonsensical.
 517          */
 518         if (!new_len)
 519                 return ret;
 520
 521         if (down_write_killable(&current->mm->mmap_sem))
 522                 return -EINTR;
 523
 524         if (flags & MREMAP_FIXED) {
 525                 ret = mremap_to(addr, old_len, new_addr, new_len,
 526                                 &locked);
 527                 goto out;
 528         }
 529
 530         /*
 531          * Always allow a shrinking remap: that just unmaps
 532          * the unnecessary pages..
 533          * do_munmap does all the needed commit accounting
 534          */
 535         if (old_len >= new_len) {
 536                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 537                 if (ret && old_len != new_len)
 538                         goto out;
 539                 ret = addr;
 540                 goto out;
 541         }
 542
 543         /*
 544          * Ok, we need to grow..
 545          */
 546         vma = vma_to_resize(addr, old_len, new_len, &charged);
 547         if (IS_ERR(vma)) {
 548                 ret = PTR_ERR(vma);
 549                 goto out;
 550         }
 551
 552         /* old_len exactly to the end of the area..
 553          */
 554         if (old_len == vma->vm_end - addr) {
 555                 /* can we just expand the current mapping? */
 556                 if (vma_expandable(vma, new_len - old_len)) {
 557                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 558
 559                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 560                                        vma->vm_pgoff, NULL)) {
 561                                 ret = -ENOMEM;
 562                                 goto out;
 563                         }
 564
 565                         vm_stat_account(mm, vma->vm_flags, pages);
 566                         if (vma->vm_flags & VM_LOCKED) {
 567                                 mm->locked_vm += pages;
 568                                 locked = true;
 569                                 new_addr = addr;
 570                         }
 571                         ret = addr;
 572                         goto out;
 573                 }
 574         }
 575
 576         /*
 577          * We weren't able to just expand or shrink the area,
 578          * we need to create a new one and move it..
 579          */
 580         ret = -ENOMEM;
 581         if (flags & MREMAP_MAYMOVE) {
 582                 unsigned long map_flags = 0;
 583                 if (vma->vm_flags & VM_MAYSHARE)
 584                         map_flags |= MAP_SHARED;
 585
 586                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 587                                         vma->vm_pgoff +
 588                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 589                                         map_flags);
 590                 if (offset_in_page(new_addr)) {
 591                         ret = new_addr;
 592                         goto out;
 593                 }
 594
 595                 ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
 596         }
 597 out:
 598         if (offset_in_page(ret)) {
 599                 vm_unacct_memory(charged);
 600                 locked = 0;
 601         }
 602         up_write(&current->mm->mmap_sem);
 603         if (locked && new_len > old_len)
 604                 mm_populate(new_addr + old_len, new_len - old_len);
 605         return ret;
 606 }