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