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