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 #include <linux/userfaultfd_k.h>
  26
  27 #include <asm/cacheflush.h>
  28 #include <asm/tlbflush.h>
  29
  30 #include "internal.h"
  31
  32 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  33 {
  34         pgd_t *pgd;
  35         pud_t *pud;
  36         pmd_t *pmd;
  37
  38         pgd = pgd_offset(mm, addr);
  39         if (pgd_none_or_clear_bad(pgd))
  40                 return NULL;
  41
  42         pud = pud_offset(pgd, addr);
  43         if (pud_none_or_clear_bad(pud))
  44                 return NULL;
  45
  46         pmd = pmd_offset(pud, addr);
  47         if (pmd_none(*pmd))
  48                 return NULL;
  49
  50         return pmd;
  51 }
  52
  53 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
  54                             unsigned long addr)
  55 {
  56         pgd_t *pgd;
  57         pud_t *pud;
  58         pmd_t *pmd;
  59
  60         pgd = pgd_offset(mm, addr);
  61         pud = pud_alloc(mm, pgd, addr);
  62         if (!pud)
  63                 return NULL;
  64
  65         pmd = pmd_alloc(mm, pud, addr);
  66         if (!pmd)
  67                 return NULL;
  68
  69         VM_BUG_ON(pmd_trans_huge(*pmd));
  70
  71         return pmd;
  72 }
  73
  74 static void take_rmap_locks(struct vm_area_struct *vma)
  75 {
  76         if (vma->vm_file)
  77                 i_mmap_lock_write(vma->vm_file->f_mapping);
  78         if (vma->anon_vma)
  79                 anon_vma_lock_write(vma->anon_vma);
  80 }
  81
  82 static void drop_rmap_locks(struct vm_area_struct *vma)
  83 {
  84         if (vma->anon_vma)
  85                 anon_vma_unlock_write(vma->anon_vma);
  86         if (vma->vm_file)
  87                 i_mmap_unlock_write(vma->vm_file->f_mapping);
  88 }
  89
  90 static pte_t move_soft_dirty_pte(pte_t pte)
  91 {
  92         /*
  93          * Set soft dirty bit so we can notice
  94          * in userspace the ptes were moved.
  95          */
  96 #ifdef CONFIG_MEM_SOFT_DIRTY
  97         if (pte_present(pte))
  98                 pte = pte_mksoft_dirty(pte);
  99         else if (is_swap_pte(pte))
 100                 pte = pte_swp_mksoft_dirty(pte);
 101 #endif
 102         return pte;
 103 }
 104
 105 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
 106                 unsigned long old_addr, unsigned long old_end,
 107                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
 108                 unsigned long new_addr, bool need_rmap_locks, bool *need_flush)
 109 {
 110         struct mm_struct *mm = vma->vm_mm;
 111         pte_t *old_pte, *new_pte, pte;
 112         spinlock_t *old_ptl, *new_ptl;
 113         bool force_flush = false;
 114         unsigned long len = old_end - old_addr;
 115
 116         /*
 117          * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
 118          * locks to ensure that rmap will always observe either the old or the
 119          * new ptes. This is the easiest way to avoid races with
 120          * truncate_pagecache(), page migration, etc...
 121          *
 122          * When need_rmap_locks is false, we use other ways to avoid
 123          * such races:
 124          *
 125          * - During exec() shift_arg_pages(), we use a specially tagged vma
 126          *   which rmap call sites look for using is_vma_temporary_stack().
 127          *
 128          * - During mremap(), new_vma is often known to be placed after vma
 129          *   in rmap traversal order. This ensures rmap will always observe
 130          *   either the old pte, or the new pte, or both (the page table locks
 131          *   serialize access to individual ptes, but only rmap traversal
 132          *   order guarantees that we won't miss both the old and new ptes).
 133          */
 134         if (need_rmap_locks)
 135                 take_rmap_locks(vma);
 136
 137         /*
 138          * We don't have to worry about the ordering of src and dst
 139          * pte locks because exclusive mmap_sem prevents deadlock.
 140          */
 141         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 142         new_pte = pte_offset_map(new_pmd, new_addr);
 143         new_ptl = pte_lockptr(mm, new_pmd);
 144         if (new_ptl != old_ptl)
 145                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 146         arch_enter_lazy_mmu_mode();
 147
 148         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 149                                    new_pte++, new_addr += PAGE_SIZE) {
 150                 if (pte_none(*old_pte))
 151                         continue;
 152
 153                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
 154                 /*
 155                  * If we are remapping a dirty PTE, make sure
 156                  * to flush TLB before we drop the PTL for the
 157                  * old PTE or we may race with page_mkclean().
 158                  *
 159                  * This check has to be done after we removed the
 160                  * old PTE from page tables or another thread may
 161                  * dirty it after the check and before the removal.
 162                  */
 163                 if (pte_present(pte) && pte_dirty(pte))
 164                         force_flush = true;
 165                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 166                 pte = move_soft_dirty_pte(pte);
 167                 set_pte_at(mm, new_addr, new_pte, pte);
 168         }
 169
 170         arch_leave_lazy_mmu_mode();
 171         if (new_ptl != old_ptl)
 172                 spin_unlock(new_ptl);
 173         pte_unmap(new_pte - 1);
 174         if (force_flush)
 175                 flush_tlb_range(vma, old_end - len, old_end);
 176         else
 177                 *need_flush = true;
 178         pte_unmap_unlock(old_pte - 1, old_ptl);
 179         if (need_rmap_locks)
 180                 drop_rmap_locks(vma);
 181 }
 182
 183 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 184
 185 unsigned long move_page_tables(struct vm_area_struct *vma,
 186                 unsigned long old_addr, struct vm_area_struct *new_vma,
 187                 unsigned long new_addr, unsigned long len,
 188                 bool need_rmap_locks)
 189 {
 190         unsigned long extent, next, old_end;
 191         pmd_t *old_pmd, *new_pmd;
 192         bool need_flush = false;
 193         unsigned long mmun_start;       /* For mmu_notifiers */
 194         unsigned long mmun_end;         /* For mmu_notifiers */
 195
 196         old_end = old_addr + len;
 197         flush_cache_range(vma, old_addr, old_end);
 198
 199         mmun_start = old_addr;
 200         mmun_end   = old_end;
 201         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 202
 203         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 204                 cond_resched();
 205                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 206                 /* even if next overflowed, extent below will be ok */
 207                 extent = next - old_addr;
 208                 if (extent > old_end - old_addr)
 209                         extent = old_end - old_addr;
 210                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 211                 if (!old_pmd)
 212                         continue;
 213                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 214                 if (!new_pmd)
 215                         break;
 216                 if (pmd_trans_huge(*old_pmd)) {
 217                         if (extent == HPAGE_PMD_SIZE) {
 218                                 bool moved;
 219                                 /* See comment in move_ptes() */
 220                                 if (need_rmap_locks)
 221                                         take_rmap_locks(vma);
 222                                 moved = move_huge_pmd(vma, old_addr, new_addr,
 223                                                     old_end, old_pmd, new_pmd,
 224                                                     &need_flush);
 225                                 if (need_rmap_locks)
 226                                         drop_rmap_locks(vma);
 227                                 if (moved)
 228                                         continue;
 229                         }
 230                         split_huge_pmd(vma, old_pmd, old_addr);
 231                         if (pmd_trans_unstable(old_pmd))
 232                                 continue;
 233                 }
 234                 if (pte_alloc(new_vma->vm_mm, new_pmd, new_addr))
 235                         break;
 236                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 237                 if (extent > next - new_addr)
 238                         extent = next - new_addr;
 239                 if (extent > LATENCY_LIMIT)
 240                         extent = LATENCY_LIMIT;
 241                 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
 242                           new_pmd, new_addr, need_rmap_locks, &need_flush);
 243         }
 244         if (need_flush)
 245                 flush_tlb_range(vma, old_end-len, old_addr);
 246
 247         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 248
 249         return len + old_addr - old_end;        /* how much done */
 250 }
 251
 252 static unsigned long move_vma(struct vm_area_struct *vma,
 253                 unsigned long old_addr, unsigned long old_len,
 254                 unsigned long new_len, unsigned long new_addr,
 255                 bool *locked, struct vm_userfaultfd_ctx *uf)
 256 {
 257         struct mm_struct *mm = vma->vm_mm;
 258         struct vm_area_struct *new_vma;
 259         unsigned long vm_flags = vma->vm_flags;
 260         unsigned long new_pgoff;
 261         unsigned long moved_len;
 262         unsigned long excess = 0;
 263         unsigned long hiwater_vm;
 264         int split = 0;
 265         int err;
 266         bool need_rmap_locks;
 267
 268         /*
 269          * We'd prefer to avoid failure later on in do_munmap:
 270          * which may split one vma into three before unmapping.
 271          */
 272         if (mm->map_count >= sysctl_max_map_count - 3)
 273                 return -ENOMEM;
 274
 275         /*
 276          * Advise KSM to break any KSM pages in the area to be moved:
 277          * it would be confusing if they were to turn up at the new
 278          * location, where they happen to coincide with different KSM
 279          * pages recently unmapped.  But leave vma->vm_flags as it was,
 280          * so KSM can come around to merge on vma and new_vma afterwards.
 281          */
 282         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 283                                                 MADV_UNMERGEABLE, &vm_flags);
 284         if (err)
 285                 return err;
 286
 287         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 288         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 289                            &need_rmap_locks);
 290         if (!new_vma)
 291                 return -ENOMEM;
 292
 293         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 294                                      need_rmap_locks);
 295         if (moved_len < old_len) {
 296                 err = -ENOMEM;
 297         } else if (vma->vm_ops && vma->vm_ops->mremap) {
 298                 err = vma->vm_ops->mremap(new_vma);
 299         }
 300
 301         if (unlikely(err)) {
 302                 /*
 303                  * On error, move entries back from new area to old,
 304                  * which will succeed since page tables still there,
 305                  * and then proceed to unmap new area instead of old.
 306                  */
 307                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 308                                  true);
 309                 vma = new_vma;
 310                 old_len = new_len;
 311                 old_addr = new_addr;
 312                 new_addr = err;
 313         } else {
 314                 mremap_userfaultfd_prep(new_vma, uf);
 315                 arch_remap(mm, old_addr, old_addr + old_len,
 316                            new_addr, new_addr + new_len);
 317         }
 318
 319         /* Conceal VM_ACCOUNT so old reservation is not undone */
 320         if (vm_flags & VM_ACCOUNT) {
 321                 vma->vm_flags &= ~VM_ACCOUNT;
 322                 excess = vma->vm_end - vma->vm_start - old_len;
 323                 if (old_addr > vma->vm_start &&
 324                     old_addr + old_len < vma->vm_end)
 325                         split = 1;
 326         }
 327
 328         /*
 329          * If we failed to move page tables we still do total_vm increment
 330          * since do_munmap() will decrement it by old_len == new_len.
 331          *
 332          * Since total_vm is about to be raised artificially high for a
 333          * moment, we need to restore high watermark afterwards: if stats
 334          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 335          * If this were a serious issue, we'd add a flag to do_munmap().
 336          */
 337         hiwater_vm = mm->hiwater_vm;
 338         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 339
 340         /* Tell pfnmap has moved from this vma */
 341         if (unlikely(vma->vm_flags & VM_PFNMAP))
 342                 untrack_pfn_moved(vma);
 343
 344         if (do_munmap(mm, old_addr, old_len) < 0) {
 345                 /* OOM: unable to split vma, just get accounts right */
 346                 vm_unacct_memory(excess >> PAGE_SHIFT);
 347                 excess = 0;
 348         }
 349         mm->hiwater_vm = hiwater_vm;
 350
 351         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 352         if (excess) {
 353                 vma->vm_flags |= VM_ACCOUNT;
 354                 if (split)
 355                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 356         }
 357
 358         if (vm_flags & VM_LOCKED) {
 359                 mm->locked_vm += new_len >> PAGE_SHIFT;
 360                 *locked = true;
 361         }
 362
 363         return new_addr;
 364 }
 365
 366 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 367         unsigned long old_len, unsigned long new_len, unsigned long *p)
 368 {
 369         struct mm_struct *mm = current->mm;
 370         struct vm_area_struct *vma = find_vma(mm, addr);
 371         unsigned long pgoff;
 372
 373         if (!vma || vma->vm_start > addr)
 374                 return ERR_PTR(-EFAULT);
 375
 376         if (is_vm_hugetlb_page(vma))
 377                 return ERR_PTR(-EINVAL);
 378
 379         /* We can't remap across vm area boundaries */
 380         if (old_len > vma->vm_end - addr)
 381                 return ERR_PTR(-EFAULT);
 382
 383         if (new_len == old_len)
 384                 return vma;
 385
 386         /* Need to be careful about a growing mapping */
 387         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 388         pgoff += vma->vm_pgoff;
 389         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 390                 return ERR_PTR(-EINVAL);
 391
 392         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 393                 return ERR_PTR(-EFAULT);
 394
 395         if (vma->vm_flags & VM_LOCKED) {
 396                 unsigned long locked, lock_limit;
 397                 locked = mm->locked_vm << PAGE_SHIFT;
 398                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 399                 locked += new_len - old_len;
 400                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 401                         return ERR_PTR(-EAGAIN);
 402         }
 403
 404         if (!may_expand_vm(mm, vma->vm_flags,
 405                                 (new_len - old_len) >> PAGE_SHIFT))
 406                 return ERR_PTR(-ENOMEM);
 407
 408         if (vma->vm_flags & VM_ACCOUNT) {
 409                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 410                 if (security_vm_enough_memory_mm(mm, charged))
 411                         return ERR_PTR(-ENOMEM);
 412                 *p = charged;
 413         }
 414
 415         return vma;
 416 }
 417
 418 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 419                 unsigned long new_addr, unsigned long new_len, bool *locked,
 420                 struct vm_userfaultfd_ctx *uf)
 421 {
 422         struct mm_struct *mm = current->mm;
 423         struct vm_area_struct *vma;
 424         unsigned long ret = -EINVAL;
 425         unsigned long charged = 0;
 426         unsigned long map_flags;
 427
 428         if (offset_in_page(new_addr))
 429                 goto out;
 430
 431         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 432                 goto out;
 433
 434         /* Ensure the old/new locations do not overlap */
 435         if (addr + old_len > new_addr && new_addr + new_len > addr)
 436                 goto out;
 437
 438         ret = do_munmap(mm, new_addr, new_len);
 439         if (ret)
 440                 goto out;
 441
 442         if (old_len >= new_len) {
 443                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 444                 if (ret && old_len != new_len)
 445                         goto out;
 446                 old_len = new_len;
 447         }
 448
 449         vma = vma_to_resize(addr, old_len, new_len, &charged);
 450         if (IS_ERR(vma)) {
 451                 ret = PTR_ERR(vma);
 452                 goto out;
 453         }
 454
 455         map_flags = MAP_FIXED;
 456         if (vma->vm_flags & VM_MAYSHARE)
 457                 map_flags |= MAP_SHARED;
 458
 459         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 460                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 461                                 map_flags);
 462         if (offset_in_page(ret))
 463                 goto out1;
 464
 465         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, uf);
 466         if (!(offset_in_page(ret)))
 467                 goto out;
 468 out1:
 469         vm_unacct_memory(charged);
 470
 471 out:
 472         return ret;
 473 }
 474
 475 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 476 {
 477         unsigned long end = vma->vm_end + delta;
 478         if (end < vma->vm_end) /* overflow */
 479                 return 0;
 480         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 481                 return 0;
 482         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 483                               0, MAP_FIXED) & ~PAGE_MASK)
 484                 return 0;
 485         return 1;
 486 }
 487
 488 /*
 489  * Expand (or shrink) an existing mapping, potentially moving it at the
 490  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 491  *
 492  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 493  * This option implies MREMAP_MAYMOVE.
 494  */
 495 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 496                 unsigned long, new_len, unsigned long, flags,
 497                 unsigned long, new_addr)
 498 {
 499         struct mm_struct *mm = current->mm;
 500         struct vm_area_struct *vma;
 501         unsigned long ret = -EINVAL;
 502         unsigned long charged = 0;
 503         bool locked = false;
 504         struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
 505
 506         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 507                 return ret;
 508
 509         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 510                 return ret;
 511
 512         if (offset_in_page(addr))
 513                 return ret;
 514
 515         old_len = PAGE_ALIGN(old_len);
 516         new_len = PAGE_ALIGN(new_len);
 517
 518         /*
 519          * We allow a zero old-len as a special case
 520          * for DOS-emu "duplicate shm area" thing. But
 521          * a zero new-len is nonsensical.
 522          */
 523         if (!new_len)
 524                 return ret;
 525
 526         if (down_write_killable(&current->mm->mmap_sem))
 527                 return -EINTR;
 528
 529         if (flags & MREMAP_FIXED) {
 530                 ret = mremap_to(addr, old_len, new_addr, new_len,
 531                                 &locked, &uf);
 532                 goto out;
 533         }
 534
 535         /*
 536          * Always allow a shrinking remap: that just unmaps
 537          * the unnecessary pages..
 538          * do_munmap does all the needed commit accounting
 539          */
 540         if (old_len >= new_len) {
 541                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 542                 if (ret && old_len != new_len)
 543                         goto out;
 544                 ret = addr;
 545                 goto out;
 546         }
 547
 548         /*
 549          * Ok, we need to grow..
 550          */
 551         vma = vma_to_resize(addr, old_len, new_len, &charged);
 552         if (IS_ERR(vma)) {
 553                 ret = PTR_ERR(vma);
 554                 goto out;
 555         }
 556
 557         /* old_len exactly to the end of the area..
 558          */
 559         if (old_len == vma->vm_end - addr) {
 560                 /* can we just expand the current mapping? */
 561                 if (vma_expandable(vma, new_len - old_len)) {
 562                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 563
 564                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 565                                        vma->vm_pgoff, NULL)) {
 566                                 ret = -ENOMEM;
 567                                 goto out;
 568                         }
 569
 570                         vm_stat_account(mm, vma->vm_flags, pages);
 571                         if (vma->vm_flags & VM_LOCKED) {
 572                                 mm->locked_vm += pages;
 573                                 locked = true;
 574                                 new_addr = addr;
 575                         }
 576                         ret = addr;
 577                         goto out;
 578                 }
 579         }
 580
 581         /*
 582          * We weren't able to just expand or shrink the area,
 583          * we need to create a new one and move it..
 584          */
 585         ret = -ENOMEM;
 586         if (flags & MREMAP_MAYMOVE) {
 587                 unsigned long map_flags = 0;
 588                 if (vma->vm_flags & VM_MAYSHARE)
 589                         map_flags |= MAP_SHARED;
 590
 591                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 592                                         vma->vm_pgoff +
 593                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 594                                         map_flags);
 595                 if (offset_in_page(new_addr)) {
 596                         ret = new_addr;
 597                         goto out;
 598                 }
 599
 600                 ret = move_vma(vma, addr, old_len, new_len, new_addr,
 601                                &locked, &uf);
 602         }
 603 out:
 604         if (offset_in_page(ret)) {
 605                 vm_unacct_memory(charged);
 606                 locked = 0;
 607         }
 608         up_write(&current->mm->mmap_sem);
 609         if (locked && new_len > old_len)
 610                 mm_populate(new_addr + old_len, new_len - old_len);
 611         mremap_userfaultfd_complete(&uf, addr, new_addr, old_len);
 612         return ret;
 613 }