mm/mremap.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *      mm/mremap.c
   4  *
   5  *      (C) Copyright 1996 Linus Torvalds
   6  *
   7  *      Address space accounting code   <alan@lxorguk.ukuu.org.uk>
   8  *      (C) Copyright 2002 Red Hat Inc, All Rights Reserved
   9  */
  10
  11 #include <linux/mm.h>
  12 #include <linux/hugetlb.h>
  13 #include <linux/shm.h>
  14 #include <linux/ksm.h>
  15 #include <linux/mman.h>
  16 #include <linux/swap.h>
  17 #include <linux/capability.h>
  18 #include <linux/fs.h>
  19 #include <linux/swapops.h>
  20 #include <linux/highmem.h>
  21 #include <linux/security.h>
  22 #include <linux/syscalls.h>
  23 #include <linux/mmu_notifier.h>
  24 #include <linux/uaccess.h>
  25 #include <linux/mm-arch-hooks.h>
  26 #include <linux/userfaultfd_k.h>
  27
  28 #include <asm/cacheflush.h>
  29 #include <asm/tlbflush.h>
  30
  31 #include "internal.h"
  32
  33 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  34 {
  35         pgd_t *pgd;
  36         p4d_t *p4d;
  37         pud_t *pud;
  38         pmd_t *pmd;
  39
  40         pgd = pgd_offset(mm, addr);
  41         if (pgd_none_or_clear_bad(pgd))
  42                 return NULL;
  43
  44         p4d = p4d_offset(pgd, addr);
  45         if (p4d_none_or_clear_bad(p4d))
  46                 return NULL;
  47
  48         pud = pud_offset(p4d, addr);
  49         if (pud_none_or_clear_bad(pud))
  50                 return NULL;
  51
  52         pmd = pmd_offset(pud, addr);
  53         if (pmd_none(*pmd))
  54                 return NULL;
  55
  56         return pmd;
  57 }
  58
  59 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
  60                             unsigned long addr)
  61 {
  62         pgd_t *pgd;
  63         p4d_t *p4d;
  64         pud_t *pud;
  65         pmd_t *pmd;
  66
  67         pgd = pgd_offset(mm, addr);
  68         p4d = p4d_alloc(mm, pgd, addr);
  69         if (!p4d)
  70                 return NULL;
  71         pud = pud_alloc(mm, p4d, addr);
  72         if (!pud)
  73                 return NULL;
  74
  75         pmd = pmd_alloc(mm, pud, addr);
  76         if (!pmd)
  77                 return NULL;
  78
  79         VM_BUG_ON(pmd_trans_huge(*pmd));
  80
  81         return pmd;
  82 }
  83
  84 static void take_rmap_locks(struct vm_area_struct *vma)
  85 {
  86         if (vma->vm_file)
  87                 i_mmap_lock_write(vma->vm_file->f_mapping);
  88         if (vma->anon_vma)
  89                 anon_vma_lock_write(vma->anon_vma);
  90 }
  91
  92 static void drop_rmap_locks(struct vm_area_struct *vma)
  93 {
  94         if (vma->anon_vma)
  95                 anon_vma_unlock_write(vma->anon_vma);
  96         if (vma->vm_file)
  97                 i_mmap_unlock_write(vma->vm_file->f_mapping);
  98 }
  99
 100 static pte_t move_soft_dirty_pte(pte_t pte)
 101 {
 102         /*
 103          * Set soft dirty bit so we can notice
 104          * in userspace the ptes were moved.
 105          */
 106 #ifdef CONFIG_MEM_SOFT_DIRTY
 107         if (pte_present(pte))
 108                 pte = pte_mksoft_dirty(pte);
 109         else if (is_swap_pte(pte))
 110                 pte = pte_swp_mksoft_dirty(pte);
 111 #endif
 112         return pte;
 113 }
 114
 115 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
 116                 unsigned long old_addr, unsigned long old_end,
 117                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
 118                 unsigned long new_addr, bool need_rmap_locks)
 119 {
 120         struct mm_struct *mm = vma->vm_mm;
 121         pte_t *old_pte, *new_pte, pte;
 122         spinlock_t *old_ptl, *new_ptl;
 123         bool force_flush = false;
 124         unsigned long len = old_end - old_addr;
 125
 126         /*
 127          * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
 128          * locks to ensure that rmap will always observe either the old or the
 129          * new ptes. This is the easiest way to avoid races with
 130          * truncate_pagecache(), page migration, etc...
 131          *
 132          * When need_rmap_locks is false, we use other ways to avoid
 133          * such races:
 134          *
 135          * - During exec() shift_arg_pages(), we use a specially tagged vma
 136          *   which rmap call sites look for using vma_is_temporary_stack().
 137          *
 138          * - During mremap(), new_vma is often known to be placed after vma
 139          *   in rmap traversal order. This ensures rmap will always observe
 140          *   either the old pte, or the new pte, or both (the page table locks
 141          *   serialize access to individual ptes, but only rmap traversal
 142          *   order guarantees that we won't miss both the old and new ptes).
 143          */
 144         if (need_rmap_locks)
 145                 take_rmap_locks(vma);
 146
 147         /*
 148          * We don't have to worry about the ordering of src and dst
 149          * pte locks because exclusive mmap_lock prevents deadlock.
 150          */
 151         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 152         new_pte = pte_offset_map(new_pmd, new_addr);
 153         new_ptl = pte_lockptr(mm, new_pmd);
 154         if (new_ptl != old_ptl)
 155                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 156         flush_tlb_batched_pending(vma->vm_mm);
 157         arch_enter_lazy_mmu_mode();
 158
 159         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 160                                    new_pte++, new_addr += PAGE_SIZE) {
 161                 if (pte_none(*old_pte))
 162                         continue;
 163
 164 #ifdef CONFIG_FINEGRAINED_THP
 165                 if (pte_cont(*old_pte)) {
 166                         /*
 167                          * Contiguous ptes will be moved,
 168                          * and we cannot ensure their alignment.
 169                          * So, simply split them.
 170                          */
 171                         split_huge_pte_address(vma, old_addr, false, NULL);
 172                 }
 173 #endif /* CONFIG_FINEGRAINED_THP */
 174
 175                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
 176                 /*
 177                  * If we are remapping a valid PTE, make sure
 178                  * to flush TLB before we drop the PTL for the
 179                  * PTE.
 180                  *
 181                  * NOTE! Both old and new PTL matter: the old one
 182                  * for racing with page_mkclean(), the new one to
 183                  * make sure the physical page stays valid until
 184                  * the TLB entry for the old mapping has been
 185                  * flushed.
 186                  */
 187                 if (pte_present(pte))
 188                         force_flush = true;
 189                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 190                 pte = move_soft_dirty_pte(pte);
 191                 set_pte_at(mm, new_addr, new_pte, pte);
 192         }
 193
 194         arch_leave_lazy_mmu_mode();
 195         if (force_flush)
 196                 flush_tlb_range(vma, old_end - len, old_end);
 197         if (new_ptl != old_ptl)
 198                 spin_unlock(new_ptl);
 199         pte_unmap(new_pte - 1);
 200         pte_unmap_unlock(old_pte - 1, old_ptl);
 201         if (need_rmap_locks)
 202                 drop_rmap_locks(vma);
 203 }
 204
 205 #ifdef CONFIG_HAVE_MOVE_PMD
 206 static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
 207                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
 208 {
 209         spinlock_t *old_ptl, *new_ptl;
 210         struct mm_struct *mm = vma->vm_mm;
 211         pmd_t pmd;
 212
 213         /*
 214          * The destination pmd shouldn't be established, free_pgtables()
 215          * should have released it.
 216          *
 217          * However, there's a case during execve() where we use mremap
 218          * to move the initial stack, and in that case the target area
 219          * may overlap the source area (always moving down).
 220          *
 221          * If everything is PMD-aligned, that works fine, as moving
 222          * each pmd down will clear the source pmd. But if we first
 223          * have a few 4kB-only pages that get moved down, and then
 224          * hit the "now the rest is PMD-aligned, let's do everything
 225          * one pmd at a time", we will still have the old (now empty
 226          * of any 4kB pages, but still there) PMD in the page table
 227          * tree.
 228          *
 229          * Warn on it once - because we really should try to figure
 230          * out how to do this better - but then say "I won't move
 231          * this pmd".
 232          *
 233          * One alternative might be to just unmap the target pmd at
 234          * this point, and verify that it really is empty. We'll see.
 235          */
 236         if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
 237                 return false;
 238
 239         /*
 240          * We don't have to worry about the ordering of src and dst
 241          * ptlocks because exclusive mmap_lock prevents deadlock.
 242          */
 243         old_ptl = pmd_lock(vma->vm_mm, old_pmd);
 244         new_ptl = pmd_lockptr(mm, new_pmd);
 245         if (new_ptl != old_ptl)
 246                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 247
 248         /* Clear the pmd */
 249         pmd = *old_pmd;
 250         pmd_clear(old_pmd);
 251
 252         VM_BUG_ON(!pmd_none(*new_pmd));
 253
 254         /* Set the new pmd */
 255         set_pmd_at(mm, new_addr, new_pmd, pmd);
 256         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
 257         if (new_ptl != old_ptl)
 258                 spin_unlock(new_ptl);
 259         spin_unlock(old_ptl);
 260
 261         return true;
 262 }
 263 #endif
 264
 265 unsigned long move_page_tables(struct vm_area_struct *vma,
 266                 unsigned long old_addr, struct vm_area_struct *new_vma,
 267                 unsigned long new_addr, unsigned long len,
 268                 bool need_rmap_locks)
 269 {
 270         unsigned long extent, next, old_end;
 271         struct mmu_notifier_range range;
 272         pmd_t *old_pmd, *new_pmd;
 273
 274         old_end = old_addr + len;
 275         flush_cache_range(vma, old_addr, old_end);
 276
 277         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
 278                                 old_addr, old_end);
 279         mmu_notifier_invalidate_range_start(&range);
 280
 281         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 282                 cond_resched();
 283                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 284                 /* even if next overflowed, extent below will be ok */
 285                 extent = next - old_addr;
 286                 if (extent > old_end - old_addr)
 287                         extent = old_end - old_addr;
 288                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 289                 if (extent > next - new_addr)
 290                         extent = next - new_addr;
 291                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 292                 if (!old_pmd)
 293                         continue;
 294                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 295                 if (!new_pmd)
 296                         break;
 297                 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) || pmd_devmap(*old_pmd)) {
 298                         if (extent == HPAGE_PMD_SIZE) {
 299                                 bool moved;
 300                                 /* See comment in move_ptes() */
 301                                 if (need_rmap_locks)
 302                                         take_rmap_locks(vma);
 303                                 moved = move_huge_pmd(vma, old_addr, new_addr,
 304                                                       old_pmd, new_pmd);
 305                                 if (need_rmap_locks)
 306                                         drop_rmap_locks(vma);
 307                                 if (moved)
 308                                         continue;
 309                         }
 310                         split_huge_pmd(vma, old_pmd, old_addr);
 311                         if (pmd_trans_unstable(old_pmd))
 312                                 continue;
 313                 } else if (extent == PMD_SIZE) {
 314 #ifdef CONFIG_HAVE_MOVE_PMD
 315                         /*
 316                          * If the extent is PMD-sized, try to speed the move by
 317                          * moving at the PMD level if possible.
 318                          */
 319                         bool moved;
 320
 321                         if (need_rmap_locks)
 322                                 take_rmap_locks(vma);
 323                         moved = move_normal_pmd(vma, old_addr, new_addr,
 324                                                 old_pmd, new_pmd);
 325                         if (need_rmap_locks)
 326                                 drop_rmap_locks(vma);
 327                         if (moved)
 328                                 continue;
 329 #endif
 330                 }
 331
 332                 if (pte_alloc(new_vma->vm_mm, new_pmd))
 333                         break;
 334                 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
 335                           new_pmd, new_addr, need_rmap_locks);
 336         }
 337
 338         mmu_notifier_invalidate_range_end(&range);
 339
 340         return len + old_addr - old_end;        /* how much done */
 341 }
 342
 343 static unsigned long move_vma(struct vm_area_struct *vma,
 344                 unsigned long old_addr, unsigned long old_len,
 345                 unsigned long new_len, unsigned long new_addr,
 346                 bool *locked, unsigned long flags,
 347                 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
 348 {
 349         struct mm_struct *mm = vma->vm_mm;
 350         struct vm_area_struct *new_vma;
 351         unsigned long vm_flags = vma->vm_flags;
 352         unsigned long new_pgoff;
 353         unsigned long moved_len;
 354         unsigned long excess = 0;
 355         unsigned long hiwater_vm;
 356         int split = 0;
 357         int err;
 358         bool need_rmap_locks;
 359
 360         /*
 361          * We'd prefer to avoid failure later on in do_munmap:
 362          * which may split one vma into three before unmapping.
 363          */
 364         if (mm->map_count >= sysctl_max_map_count - 3)
 365                 return -ENOMEM;
 366
 367         /*
 368          * Advise KSM to break any KSM pages in the area to be moved:
 369          * it would be confusing if they were to turn up at the new
 370          * location, where they happen to coincide with different KSM
 371          * pages recently unmapped.  But leave vma->vm_flags as it was,
 372          * so KSM can come around to merge on vma and new_vma afterwards.
 373          */
 374         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 375                                                 MADV_UNMERGEABLE, &vm_flags);
 376         if (err)
 377                 return err;
 378
 379         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 380         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 381                            &need_rmap_locks);
 382         if (!new_vma)
 383                 return -ENOMEM;
 384
 385         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 386                                      need_rmap_locks);
 387         if (moved_len < old_len) {
 388                 err = -ENOMEM;
 389         } else if (vma->vm_ops && vma->vm_ops->mremap) {
 390                 err = vma->vm_ops->mremap(new_vma);
 391         }
 392
 393         if (unlikely(err)) {
 394                 /*
 395                  * On error, move entries back from new area to old,
 396                  * which will succeed since page tables still there,
 397                  * and then proceed to unmap new area instead of old.
 398                  */
 399                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 400                                  true);
 401                 vma = new_vma;
 402                 old_len = new_len;
 403                 old_addr = new_addr;
 404                 new_addr = err;
 405         } else {
 406                 mremap_userfaultfd_prep(new_vma, uf);
 407                 arch_remap(mm, old_addr, old_addr + old_len,
 408                            new_addr, new_addr + new_len);
 409         }
 410
 411         /* Conceal VM_ACCOUNT so old reservation is not undone */
 412         if (vm_flags & VM_ACCOUNT) {
 413                 vma->vm_flags &= ~VM_ACCOUNT;
 414                 excess = vma->vm_end - vma->vm_start - old_len;
 415                 if (old_addr > vma->vm_start &&
 416                     old_addr + old_len < vma->vm_end)
 417                         split = 1;
 418         }
 419
 420         /*
 421          * If we failed to move page tables we still do total_vm increment
 422          * since do_munmap() will decrement it by old_len == new_len.
 423          *
 424          * Since total_vm is about to be raised artificially high for a
 425          * moment, we need to restore high watermark afterwards: if stats
 426          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 427          * If this were a serious issue, we'd add a flag to do_munmap().
 428          */
 429         hiwater_vm = mm->hiwater_vm;
 430         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 431
 432         /* Tell pfnmap has moved from this vma */
 433         if (unlikely(vma->vm_flags & VM_PFNMAP))
 434                 untrack_pfn_moved(vma);
 435
 436         if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
 437                 if (vm_flags & VM_ACCOUNT) {
 438                         /* Always put back VM_ACCOUNT since we won't unmap */
 439                         vma->vm_flags |= VM_ACCOUNT;
 440
 441                         vm_acct_memory(new_len >> PAGE_SHIFT);
 442                 }
 443
 444                 /*
 445                  * VMAs can actually be merged back together in copy_vma
 446                  * calling merge_vma. This can happen with anonymous vmas
 447                  * which have not yet been faulted, so if we were to consider
 448                  * this VMA split we'll end up adding VM_ACCOUNT on the
 449                  * next VMA, which is completely unrelated if this VMA
 450                  * was re-merged.
 451                  */
 452                 if (split && new_vma == vma)
 453                         split = 0;
 454
 455                 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
 456                 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
 457
 458                 /* Because we won't unmap we don't need to touch locked_vm */
 459                 goto out;
 460         }
 461
 462         if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
 463                 /* OOM: unable to split vma, just get accounts right */
 464                 vm_unacct_memory(excess >> PAGE_SHIFT);
 465                 excess = 0;
 466         }
 467
 468         if (vm_flags & VM_LOCKED) {
 469                 mm->locked_vm += new_len >> PAGE_SHIFT;
 470                 *locked = true;
 471         }
 472 out:
 473         mm->hiwater_vm = hiwater_vm;
 474
 475         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 476         if (excess) {
 477                 vma->vm_flags |= VM_ACCOUNT;
 478                 if (split)
 479                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 480         }
 481
 482         return new_addr;
 483 }
 484
 485 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 486         unsigned long old_len, unsigned long new_len, unsigned long flags,
 487         unsigned long *p)
 488 {
 489         struct mm_struct *mm = current->mm;
 490         struct vm_area_struct *vma = find_vma(mm, addr);
 491         unsigned long pgoff;
 492
 493         if (!vma || vma->vm_start > addr)
 494                 return ERR_PTR(-EFAULT);
 495
 496         /*
 497          * !old_len is a special case where an attempt is made to 'duplicate'
 498          * a mapping.  This makes no sense for private mappings as it will
 499          * instead create a fresh/new mapping unrelated to the original.  This
 500          * is contrary to the basic idea of mremap which creates new mappings
 501          * based on the original.  There are no known use cases for this
 502          * behavior.  As a result, fail such attempts.
 503          */
 504         if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
 505                 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap.  This is not supported.\n", current->comm, current->pid);
 506                 return ERR_PTR(-EINVAL);
 507         }
 508
 509         if (flags & MREMAP_DONTUNMAP && (!vma_is_anonymous(vma) ||
 510                         vma->vm_flags & VM_SHARED))
 511                 return ERR_PTR(-EINVAL);
 512
 513         if (is_vm_hugetlb_page(vma))
 514                 return ERR_PTR(-EINVAL);
 515
 516         /* We can't remap across vm area boundaries */
 517         if (old_len > vma->vm_end - addr)
 518                 return ERR_PTR(-EFAULT);
 519
 520         if (new_len == old_len)
 521                 return vma;
 522
 523         /* Need to be careful about a growing mapping */
 524         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 525         pgoff += vma->vm_pgoff;
 526         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 527                 return ERR_PTR(-EINVAL);
 528
 529         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 530                 return ERR_PTR(-EFAULT);
 531
 532         if (vma->vm_flags & VM_LOCKED) {
 533                 unsigned long locked, lock_limit;
 534                 locked = mm->locked_vm << PAGE_SHIFT;
 535                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 536                 locked += new_len - old_len;
 537                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 538                         return ERR_PTR(-EAGAIN);
 539         }
 540
 541         if (!may_expand_vm(mm, vma->vm_flags,
 542                                 (new_len - old_len) >> PAGE_SHIFT))
 543                 return ERR_PTR(-ENOMEM);
 544
 545         if (vma->vm_flags & VM_ACCOUNT) {
 546                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 547                 if (security_vm_enough_memory_mm(mm, charged))
 548                         return ERR_PTR(-ENOMEM);
 549                 *p = charged;
 550         }
 551
 552         return vma;
 553 }
 554
 555 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 556                 unsigned long new_addr, unsigned long new_len, bool *locked,
 557                 unsigned long flags, struct vm_userfaultfd_ctx *uf,
 558                 struct list_head *uf_unmap_early,
 559                 struct list_head *uf_unmap)
 560 {
 561         struct mm_struct *mm = current->mm;
 562         struct vm_area_struct *vma;
 563         unsigned long ret = -EINVAL;
 564         unsigned long charged = 0;
 565         unsigned long map_flags = 0;
 566
 567         if (offset_in_page(new_addr))
 568                 goto out;
 569
 570         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 571                 goto out;
 572
 573         /* Ensure the old/new locations do not overlap */
 574         if (addr + old_len > new_addr && new_addr + new_len > addr)
 575                 goto out;
 576
 577         /*
 578          * move_vma() need us to stay 4 maps below the threshold, otherwise
 579          * it will bail out at the very beginning.
 580          * That is a problem if we have already unmaped the regions here
 581          * (new_addr, and old_addr), because userspace will not know the
 582          * state of the vma's after it gets -ENOMEM.
 583          * So, to avoid such scenario we can pre-compute if the whole
 584          * operation has high chances to success map-wise.
 585          * Worst-scenario case is when both vma's (new_addr and old_addr) get
 586          * split in 3 before unmaping it.
 587          * That means 2 more maps (1 for each) to the ones we already hold.
 588          * Check whether current map count plus 2 still leads us to 4 maps below
 589          * the threshold, otherwise return -ENOMEM here to be more safe.
 590          */
 591         if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
 592                 return -ENOMEM;
 593
 594         if (flags & MREMAP_FIXED) {
 595                 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
 596                 if (ret)
 597                         goto out;
 598         }
 599
 600         if (old_len >= new_len) {
 601                 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
 602                 if (ret && old_len != new_len)
 603                         goto out;
 604                 old_len = new_len;
 605         }
 606
 607         vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
 608         if (IS_ERR(vma)) {
 609                 ret = PTR_ERR(vma);
 610                 goto out;
 611         }
 612
 613         /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
 614         if (flags & MREMAP_DONTUNMAP &&
 615                 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
 616                 ret = -ENOMEM;
 617                 goto out;
 618         }
 619
 620         if (flags & MREMAP_FIXED)
 621                 map_flags |= MAP_FIXED;
 622
 623         if (vma->vm_flags & VM_MAYSHARE)
 624                 map_flags |= MAP_SHARED;
 625
 626         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 627                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 628                                 map_flags);
 629         if (IS_ERR_VALUE(ret))
 630                 goto out1;
 631
 632         /* We got a new mapping */
 633         if (!(flags & MREMAP_FIXED))
 634                 new_addr = ret;
 635
 636         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
 637                        uf_unmap);
 638
 639         if (!(offset_in_page(ret)))
 640                 goto out;
 641
 642 out1:
 643         vm_unacct_memory(charged);
 644
 645 out:
 646         return ret;
 647 }
 648
 649 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 650 {
 651         unsigned long end = vma->vm_end + delta;
 652         if (end < vma->vm_end) /* overflow */
 653                 return 0;
 654         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 655                 return 0;
 656         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 657                               0, MAP_FIXED) & ~PAGE_MASK)
 658                 return 0;
 659         return 1;
 660 }
 661
 662 /*
 663  * Expand (or shrink) an existing mapping, potentially moving it at the
 664  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 665  *
 666  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 667  * This option implies MREMAP_MAYMOVE.
 668  */
 669 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 670                 unsigned long, new_len, unsigned long, flags,
 671                 unsigned long, new_addr)
 672 {
 673         struct mm_struct *mm = current->mm;
 674         struct vm_area_struct *vma;
 675         unsigned long ret = -EINVAL;
 676         unsigned long charged = 0;
 677         bool locked = false;
 678         bool downgraded = false;
 679         struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
 680         LIST_HEAD(uf_unmap_early);
 681         LIST_HEAD(uf_unmap);
 682
 683         /*
 684          * There is a deliberate asymmetry here: we strip the pointer tag
 685          * from the old address but leave the new address alone. This is
 686          * for consistency with mmap(), where we prevent the creation of
 687          * aliasing mappings in userspace by leaving the tag bits of the
 688          * mapping address intact. A non-zero tag will cause the subsequent
 689          * range checks to reject the address as invalid.
 690          *
 691          * See Documentation/arm64/tagged-address-abi.rst for more information.
 692          */
 693         addr = untagged_addr(addr);
 694
 695         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
 696                 return ret;
 697
 698         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 699                 return ret;
 700
 701         /*
 702          * MREMAP_DONTUNMAP is always a move and it does not allow resizing
 703          * in the process.
 704          */
 705         if (flags & MREMAP_DONTUNMAP &&
 706                         (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
 707                 return ret;
 708
 709
 710         if (offset_in_page(addr))
 711                 return ret;
 712
 713         old_len = PAGE_ALIGN(old_len);
 714         new_len = PAGE_ALIGN(new_len);
 715
 716         /*
 717          * We allow a zero old-len as a special case
 718          * for DOS-emu "duplicate shm area" thing. But
 719          * a zero new-len is nonsensical.
 720          */
 721         if (!new_len)
 722                 return ret;
 723
 724         if (mmap_write_lock_killable(current->mm))
 725                 return -EINTR;
 726
 727         if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
 728                 ret = mremap_to(addr, old_len, new_addr, new_len,
 729                                 &locked, flags, &uf, &uf_unmap_early,
 730                                 &uf_unmap);
 731                 goto out;
 732         }
 733
 734         /*
 735          * Always allow a shrinking remap: that just unmaps
 736          * the unnecessary pages..
 737          * __do_munmap does all the needed commit accounting, and
 738          * downgrades mmap_lock to read if so directed.
 739          */
 740         if (old_len >= new_len) {
 741                 int retval;
 742
 743                 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
 744                                   &uf_unmap, true);
 745                 if (retval < 0 && old_len != new_len) {
 746                         ret = retval;
 747                         goto out;
 748                 /* Returning 1 indicates mmap_lock is downgraded to read. */
 749                 } else if (retval == 1)
 750                         downgraded = true;
 751                 ret = addr;
 752                 goto out;
 753         }
 754
 755         /*
 756          * Ok, we need to grow..
 757          */
 758         vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
 759         if (IS_ERR(vma)) {
 760                 ret = PTR_ERR(vma);
 761                 goto out;
 762         }
 763
 764         /* old_len exactly to the end of the area..
 765          */
 766         if (old_len == vma->vm_end - addr) {
 767                 /* can we just expand the current mapping? */
 768                 if (vma_expandable(vma, new_len - old_len)) {
 769                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 770
 771                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 772                                        vma->vm_pgoff, NULL)) {
 773                                 ret = -ENOMEM;
 774                                 goto out;
 775                         }
 776
 777                         vm_stat_account(mm, vma->vm_flags, pages);
 778                         if (vma->vm_flags & VM_LOCKED) {
 779                                 mm->locked_vm += pages;
 780                                 locked = true;
 781                                 new_addr = addr;
 782                         }
 783                         ret = addr;
 784                         goto out;
 785                 }
 786         }
 787
 788         /*
 789          * We weren't able to just expand or shrink the area,
 790          * we need to create a new one and move it..
 791          */
 792         ret = -ENOMEM;
 793         if (flags & MREMAP_MAYMOVE) {
 794                 unsigned long map_flags = 0;
 795                 if (vma->vm_flags & VM_MAYSHARE)
 796                         map_flags |= MAP_SHARED;
 797
 798                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 799                                         vma->vm_pgoff +
 800                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 801                                         map_flags);
 802                 if (IS_ERR_VALUE(new_addr)) {
 803                         ret = new_addr;
 804                         goto out;
 805                 }
 806
 807                 ret = move_vma(vma, addr, old_len, new_len, new_addr,
 808                                &locked, flags, &uf, &uf_unmap);
 809         }
 810 out:
 811         if (offset_in_page(ret)) {
 812                 vm_unacct_memory(charged);
 813                 locked = false;
 814         }
 815         if (downgraded)
 816                 mmap_read_unlock(current->mm);
 817         else
 818                 mmap_write_unlock(current->mm);
 819         if (locked && new_len > old_len)
 820                 mm_populate(new_addr + old_len, new_len - old_len);
 821         userfaultfd_unmap_complete(mm, &uf_unmap_early);
 822         mremap_userfaultfd_complete(&uf, addr, ret, old_len);
 823         userfaultfd_unmap_complete(mm, &uf_unmap);
 824         return ret;
 825 }