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                 struct list_head *uf_unmap)
 257 {
 258         struct mm_struct *mm = vma->vm_mm;
 259         struct vm_area_struct *new_vma;
 260         unsigned long vm_flags = vma->vm_flags;
 261         unsigned long new_pgoff;
 262         unsigned long moved_len;
 263         unsigned long excess = 0;
 264         unsigned long hiwater_vm;
 265         int split = 0;
 266         int err;
 267         bool need_rmap_locks;
 268
 269         /*
 270          * We'd prefer to avoid failure later on in do_munmap:
 271          * which may split one vma into three before unmapping.
 272          */
 273         if (mm->map_count >= sysctl_max_map_count - 3)
 274                 return -ENOMEM;
 275
 276         /*
 277          * Advise KSM to break any KSM pages in the area to be moved:
 278          * it would be confusing if they were to turn up at the new
 279          * location, where they happen to coincide with different KSM
 280          * pages recently unmapped.  But leave vma->vm_flags as it was,
 281          * so KSM can come around to merge on vma and new_vma afterwards.
 282          */
 283         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 284                                                 MADV_UNMERGEABLE, &vm_flags);
 285         if (err)
 286                 return err;
 287
 288         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 289         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 290                            &need_rmap_locks);
 291         if (!new_vma)
 292                 return -ENOMEM;
 293
 294         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 295                                      need_rmap_locks);
 296         if (moved_len < old_len) {
 297                 err = -ENOMEM;
 298         } else if (vma->vm_ops && vma->vm_ops->mremap) {
 299                 err = vma->vm_ops->mremap(new_vma);
 300         }
 301
 302         if (unlikely(err)) {
 303                 /*
 304                  * On error, move entries back from new area to old,
 305                  * which will succeed since page tables still there,
 306                  * and then proceed to unmap new area instead of old.
 307                  */
 308                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 309                                  true);
 310                 vma = new_vma;
 311                 old_len = new_len;
 312                 old_addr = new_addr;
 313                 new_addr = err;
 314         } else {
 315                 mremap_userfaultfd_prep(new_vma, uf);
 316                 arch_remap(mm, old_addr, old_addr + old_len,
 317                            new_addr, new_addr + new_len);
 318         }
 319
 320         /* Conceal VM_ACCOUNT so old reservation is not undone */
 321         if (vm_flags & VM_ACCOUNT) {
 322                 vma->vm_flags &= ~VM_ACCOUNT;
 323                 excess = vma->vm_end - vma->vm_start - old_len;
 324                 if (old_addr > vma->vm_start &&
 325                     old_addr + old_len < vma->vm_end)
 326                         split = 1;
 327         }
 328
 329         /*
 330          * If we failed to move page tables we still do total_vm increment
 331          * since do_munmap() will decrement it by old_len == new_len.
 332          *
 333          * Since total_vm is about to be raised artificially high for a
 334          * moment, we need to restore high watermark afterwards: if stats
 335          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 336          * If this were a serious issue, we'd add a flag to do_munmap().
 337          */
 338         hiwater_vm = mm->hiwater_vm;
 339         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 340
 341         /* Tell pfnmap has moved from this vma */
 342         if (unlikely(vma->vm_flags & VM_PFNMAP))
 343                 untrack_pfn_moved(vma);
 344
 345         if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
 346                 /* OOM: unable to split vma, just get accounts right */
 347                 vm_unacct_memory(excess >> PAGE_SHIFT);
 348                 excess = 0;
 349         }
 350         mm->hiwater_vm = hiwater_vm;
 351
 352         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 353         if (excess) {
 354                 vma->vm_flags |= VM_ACCOUNT;
 355                 if (split)
 356                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 357         }
 358
 359         if (vm_flags & VM_LOCKED) {
 360                 mm->locked_vm += new_len >> PAGE_SHIFT;
 361                 *locked = true;
 362         }
 363
 364         return new_addr;
 365 }
 366
 367 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 368         unsigned long old_len, unsigned long new_len, unsigned long *p)
 369 {
 370         struct mm_struct *mm = current->mm;
 371         struct vm_area_struct *vma = find_vma(mm, addr);
 372         unsigned long pgoff;
 373
 374         if (!vma || vma->vm_start > addr)
 375                 return ERR_PTR(-EFAULT);
 376
 377         if (is_vm_hugetlb_page(vma))
 378                 return ERR_PTR(-EINVAL);
 379
 380         /* We can't remap across vm area boundaries */
 381         if (old_len > vma->vm_end - addr)
 382                 return ERR_PTR(-EFAULT);
 383
 384         if (new_len == old_len)
 385                 return vma;
 386
 387         /* Need to be careful about a growing mapping */
 388         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 389         pgoff += vma->vm_pgoff;
 390         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 391                 return ERR_PTR(-EINVAL);
 392
 393         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 394                 return ERR_PTR(-EFAULT);
 395
 396         if (vma->vm_flags & VM_LOCKED) {
 397                 unsigned long locked, lock_limit;
 398                 locked = mm->locked_vm << PAGE_SHIFT;
 399                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 400                 locked += new_len - old_len;
 401                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 402                         return ERR_PTR(-EAGAIN);
 403         }
 404
 405         if (!may_expand_vm(mm, vma->vm_flags,
 406                                 (new_len - old_len) >> PAGE_SHIFT))
 407                 return ERR_PTR(-ENOMEM);
 408
 409         if (vma->vm_flags & VM_ACCOUNT) {
 410                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 411                 if (security_vm_enough_memory_mm(mm, charged))
 412                         return ERR_PTR(-ENOMEM);
 413                 *p = charged;
 414         }
 415
 416         return vma;
 417 }
 418
 419 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 420                 unsigned long new_addr, unsigned long new_len, bool *locked,
 421                 struct vm_userfaultfd_ctx *uf,
 422                 struct list_head *uf_unmap)
 423 {
 424         struct mm_struct *mm = current->mm;
 425         struct vm_area_struct *vma;
 426         unsigned long ret = -EINVAL;
 427         unsigned long charged = 0;
 428         unsigned long map_flags;
 429
 430         if (offset_in_page(new_addr))
 431                 goto out;
 432
 433         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 434                 goto out;
 435
 436         /* Ensure the old/new locations do not overlap */
 437         if (addr + old_len > new_addr && new_addr + new_len > addr)
 438                 goto out;
 439
 440         ret = do_munmap(mm, new_addr, new_len, NULL);
 441         if (ret)
 442                 goto out;
 443
 444         if (old_len >= new_len) {
 445                 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
 446                 if (ret && old_len != new_len)
 447                         goto out;
 448                 old_len = new_len;
 449         }
 450
 451         vma = vma_to_resize(addr, old_len, new_len, &charged);
 452         if (IS_ERR(vma)) {
 453                 ret = PTR_ERR(vma);
 454                 goto out;
 455         }
 456
 457         map_flags = MAP_FIXED;
 458         if (vma->vm_flags & VM_MAYSHARE)
 459                 map_flags |= MAP_SHARED;
 460
 461         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 462                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 463                                 map_flags);
 464         if (offset_in_page(ret))
 465                 goto out1;
 466
 467         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, uf,
 468                        uf_unmap);
 469         if (!(offset_in_page(ret)))
 470                 goto out;
 471 out1:
 472         vm_unacct_memory(charged);
 473
 474 out:
 475         return ret;
 476 }
 477
 478 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 479 {
 480         unsigned long end = vma->vm_end + delta;
 481         if (end < vma->vm_end) /* overflow */
 482                 return 0;
 483         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 484                 return 0;
 485         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 486                               0, MAP_FIXED) & ~PAGE_MASK)
 487                 return 0;
 488         return 1;
 489 }
 490
 491 /*
 492  * Expand (or shrink) an existing mapping, potentially moving it at the
 493  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 494  *
 495  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 496  * This option implies MREMAP_MAYMOVE.
 497  */
 498 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 499                 unsigned long, new_len, unsigned long, flags,
 500                 unsigned long, new_addr)
 501 {
 502         struct mm_struct *mm = current->mm;
 503         struct vm_area_struct *vma;
 504         unsigned long ret = -EINVAL;
 505         unsigned long charged = 0;
 506         bool locked = false;
 507         struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
 508         LIST_HEAD(uf_unmap);
 509
 510         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 511                 return ret;
 512
 513         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 514                 return ret;
 515
 516         if (offset_in_page(addr))
 517                 return ret;
 518
 519         old_len = PAGE_ALIGN(old_len);
 520         new_len = PAGE_ALIGN(new_len);
 521
 522         /*
 523          * We allow a zero old-len as a special case
 524          * for DOS-emu "duplicate shm area" thing. But
 525          * a zero new-len is nonsensical.
 526          */
 527         if (!new_len)
 528                 return ret;
 529
 530         if (down_write_killable(&current->mm->mmap_sem))
 531                 return -EINTR;
 532
 533         if (flags & MREMAP_FIXED) {
 534                 ret = mremap_to(addr, old_len, new_addr, new_len,
 535                                 &locked, &uf, &uf_unmap);
 536                 goto out;
 537         }
 538
 539         /*
 540          * Always allow a shrinking remap: that just unmaps
 541          * the unnecessary pages..
 542          * do_munmap does all the needed commit accounting
 543          */
 544         if (old_len >= new_len) {
 545                 ret = do_munmap(mm, addr+new_len, old_len - new_len, &uf_unmap);
 546                 if (ret && old_len != new_len)
 547                         goto out;
 548                 ret = addr;
 549                 goto out;
 550         }
 551
 552         /*
 553          * Ok, we need to grow..
 554          */
 555         vma = vma_to_resize(addr, old_len, new_len, &charged);
 556         if (IS_ERR(vma)) {
 557                 ret = PTR_ERR(vma);
 558                 goto out;
 559         }
 560
 561         /* old_len exactly to the end of the area..
 562          */
 563         if (old_len == vma->vm_end - addr) {
 564                 /* can we just expand the current mapping? */
 565                 if (vma_expandable(vma, new_len - old_len)) {
 566                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 567
 568                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 569                                        vma->vm_pgoff, NULL)) {
 570                                 ret = -ENOMEM;
 571                                 goto out;
 572                         }
 573
 574                         vm_stat_account(mm, vma->vm_flags, pages);
 575                         if (vma->vm_flags & VM_LOCKED) {
 576                                 mm->locked_vm += pages;
 577                                 locked = true;
 578                                 new_addr = addr;
 579                         }
 580                         ret = addr;
 581                         goto out;
 582                 }
 583         }
 584
 585         /*
 586          * We weren't able to just expand or shrink the area,
 587          * we need to create a new one and move it..
 588          */
 589         ret = -ENOMEM;
 590         if (flags & MREMAP_MAYMOVE) {
 591                 unsigned long map_flags = 0;
 592                 if (vma->vm_flags & VM_MAYSHARE)
 593                         map_flags |= MAP_SHARED;
 594
 595                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 596                                         vma->vm_pgoff +
 597                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 598                                         map_flags);
 599                 if (offset_in_page(new_addr)) {
 600                         ret = new_addr;
 601                         goto out;
 602                 }
 603
 604                 ret = move_vma(vma, addr, old_len, new_len, new_addr,
 605                                &locked, &uf, &uf_unmap);
 606         }
 607 out:
 608         if (offset_in_page(ret)) {
 609                 vm_unacct_memory(charged);
 610                 locked = 0;
 611         }
 612         up_write(&current->mm->mmap_sem);
 613         if (locked && new_len > old_len)
 614                 mm_populate(new_addr + old_len, new_len - old_len);
 615         mremap_userfaultfd_complete(&uf, addr, new_addr, old_len);
 616         userfaultfd_unmap_complete(mm, &uf_unmap);
 617         return ret;
 618 }