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