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/mm_inline.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/shm.h>
  15 #include <linux/ksm.h>
  16 #include <linux/mman.h>
  17 #include <linux/swap.h>
  18 #include <linux/capability.h>
  19 #include <linux/fs.h>
  20 #include <linux/swapops.h>
  21 #include <linux/highmem.h>
  22 #include <linux/security.h>
  23 #include <linux/syscalls.h>
  24 #include <linux/mmu_notifier.h>
  25 #include <linux/uaccess.h>
  26 #include <linux/userfaultfd_k.h>
  27 #include <linux/mempolicy.h>
  28
  29 #include <asm/cacheflush.h>
  30 #include <asm/tlb.h>
  31 #include <asm/pgalloc.h>
  32
  33 #include "internal.h"
  34
  35 static pud_t *get_old_pud(struct mm_struct *mm, unsigned long addr)
  36 {
  37         pgd_t *pgd;
  38         p4d_t *p4d;
  39         pud_t *pud;
  40
  41         pgd = pgd_offset(mm, addr);
  42         if (pgd_none_or_clear_bad(pgd))
  43                 return NULL;
  44
  45         p4d = p4d_offset(pgd, addr);
  46         if (p4d_none_or_clear_bad(p4d))
  47                 return NULL;
  48
  49         pud = pud_offset(p4d, addr);
  50         if (pud_none_or_clear_bad(pud))
  51                 return NULL;
  52
  53         return pud;
  54 }
  55
  56 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  57 {
  58         pud_t *pud;
  59         pmd_t *pmd;
  60
  61         pud = get_old_pud(mm, addr);
  62         if (!pud)
  63                 return NULL;
  64
  65         pmd = pmd_offset(pud, addr);
  66         if (pmd_none(*pmd))
  67                 return NULL;
  68
  69         return pmd;
  70 }
  71
  72 static pud_t *alloc_new_pud(struct mm_struct *mm, struct vm_area_struct *vma,
  73                             unsigned long addr)
  74 {
  75         pgd_t *pgd;
  76         p4d_t *p4d;
  77
  78         pgd = pgd_offset(mm, addr);
  79         p4d = p4d_alloc(mm, pgd, addr);
  80         if (!p4d)
  81                 return NULL;
  82
  83         return pud_alloc(mm, p4d, addr);
  84 }
  85
  86 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
  87                             unsigned long addr)
  88 {
  89         pud_t *pud;
  90         pmd_t *pmd;
  91
  92         pud = alloc_new_pud(mm, vma, addr);
  93         if (!pud)
  94                 return NULL;
  95
  96         pmd = pmd_alloc(mm, pud, addr);
  97         if (!pmd)
  98                 return NULL;
  99
 100         VM_BUG_ON(pmd_trans_huge(*pmd));
 101
 102         return pmd;
 103 }
 104
 105 static void take_rmap_locks(struct vm_area_struct *vma)
 106 {
 107         if (vma->vm_file)
 108                 i_mmap_lock_write(vma->vm_file->f_mapping);
 109         if (vma->anon_vma)
 110                 anon_vma_lock_write(vma->anon_vma);
 111 }
 112
 113 static void drop_rmap_locks(struct vm_area_struct *vma)
 114 {
 115         if (vma->anon_vma)
 116                 anon_vma_unlock_write(vma->anon_vma);
 117         if (vma->vm_file)
 118                 i_mmap_unlock_write(vma->vm_file->f_mapping);
 119 }
 120
 121 static pte_t move_soft_dirty_pte(pte_t pte)
 122 {
 123         /*
 124          * Set soft dirty bit so we can notice
 125          * in userspace the ptes were moved.
 126          */
 127 #ifdef CONFIG_MEM_SOFT_DIRTY
 128         if (pte_present(pte))
 129                 pte = pte_mksoft_dirty(pte);
 130         else if (is_swap_pte(pte))
 131                 pte = pte_swp_mksoft_dirty(pte);
 132 #endif
 133         return pte;
 134 }
 135
 136 static int move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
 137                 unsigned long old_addr, unsigned long old_end,
 138                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
 139                 unsigned long new_addr, bool need_rmap_locks)
 140 {
 141         struct mm_struct *mm = vma->vm_mm;
 142         pte_t *old_pte, *new_pte, pte;
 143         spinlock_t *old_ptl, *new_ptl;
 144         bool force_flush = false;
 145         unsigned long len = old_end - old_addr;
 146         int err = 0;
 147
 148         /*
 149          * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
 150          * locks to ensure that rmap will always observe either the old or the
 151          * new ptes. This is the easiest way to avoid races with
 152          * truncate_pagecache(), page migration, etc...
 153          *
 154          * When need_rmap_locks is false, we use other ways to avoid
 155          * such races:
 156          *
 157          * - During exec() shift_arg_pages(), we use a specially tagged vma
 158          *   which rmap call sites look for using vma_is_temporary_stack().
 159          *
 160          * - During mremap(), new_vma is often known to be placed after vma
 161          *   in rmap traversal order. This ensures rmap will always observe
 162          *   either the old pte, or the new pte, or both (the page table locks
 163          *   serialize access to individual ptes, but only rmap traversal
 164          *   order guarantees that we won't miss both the old and new ptes).
 165          */
 166         if (need_rmap_locks)
 167                 take_rmap_locks(vma);
 168
 169         /*
 170          * We don't have to worry about the ordering of src and dst
 171          * pte locks because exclusive mmap_lock prevents deadlock.
 172          */
 173         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 174         if (!old_pte) {
 175                 err = -EAGAIN;
 176                 goto out;
 177         }
 178         new_pte = pte_offset_map_nolock(mm, new_pmd, new_addr, &new_ptl);
 179         if (!new_pte) {
 180                 pte_unmap_unlock(old_pte, old_ptl);
 181                 err = -EAGAIN;
 182                 goto out;
 183         }
 184         if (new_ptl != old_ptl)
 185                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 186         flush_tlb_batched_pending(vma->vm_mm);
 187         arch_enter_lazy_mmu_mode();
 188
 189         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 190                                    new_pte++, new_addr += PAGE_SIZE) {
 191                 if (pte_none(ptep_get(old_pte)))
 192                         continue;
 193
 194                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
 195                 /*
 196                  * If we are remapping a valid PTE, make sure
 197                  * to flush TLB before we drop the PTL for the
 198                  * PTE.
 199                  *
 200                  * NOTE! Both old and new PTL matter: the old one
 201                  * for racing with page_mkclean(), the new one to
 202                  * make sure the physical page stays valid until
 203                  * the TLB entry for the old mapping has been
 204                  * flushed.
 205                  */
 206                 if (pte_present(pte))
 207                         force_flush = true;
 208                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 209                 pte = move_soft_dirty_pte(pte);
 210                 set_pte_at(mm, new_addr, new_pte, pte);
 211         }
 212
 213         arch_leave_lazy_mmu_mode();
 214         if (force_flush)
 215                 flush_tlb_range(vma, old_end - len, old_end);
 216         if (new_ptl != old_ptl)
 217                 spin_unlock(new_ptl);
 218         pte_unmap(new_pte - 1);
 219         pte_unmap_unlock(old_pte - 1, old_ptl);
 220 out:
 221         if (need_rmap_locks)
 222                 drop_rmap_locks(vma);
 223         return err;
 224 }
 225
 226 #ifndef arch_supports_page_table_move
 227 #define arch_supports_page_table_move arch_supports_page_table_move
 228 static inline bool arch_supports_page_table_move(void)
 229 {
 230         return IS_ENABLED(CONFIG_HAVE_MOVE_PMD) ||
 231                 IS_ENABLED(CONFIG_HAVE_MOVE_PUD);
 232 }
 233 #endif
 234
 235 #ifdef CONFIG_HAVE_MOVE_PMD
 236 static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
 237                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
 238 {
 239         spinlock_t *old_ptl, *new_ptl;
 240         struct mm_struct *mm = vma->vm_mm;
 241         pmd_t pmd;
 242
 243         if (!arch_supports_page_table_move())
 244                 return false;
 245         /*
 246          * The destination pmd shouldn't be established, free_pgtables()
 247          * should have released it.
 248          *
 249          * However, there's a case during execve() where we use mremap
 250          * to move the initial stack, and in that case the target area
 251          * may overlap the source area (always moving down).
 252          *
 253          * If everything is PMD-aligned, that works fine, as moving
 254          * each pmd down will clear the source pmd. But if we first
 255          * have a few 4kB-only pages that get moved down, and then
 256          * hit the "now the rest is PMD-aligned, let's do everything
 257          * one pmd at a time", we will still have the old (now empty
 258          * of any 4kB pages, but still there) PMD in the page table
 259          * tree.
 260          *
 261          * Warn on it once - because we really should try to figure
 262          * out how to do this better - but then say "I won't move
 263          * this pmd".
 264          *
 265          * One alternative might be to just unmap the target pmd at
 266          * this point, and verify that it really is empty. We'll see.
 267          */
 268         if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
 269                 return false;
 270
 271         /*
 272          * We don't have to worry about the ordering of src and dst
 273          * ptlocks because exclusive mmap_lock prevents deadlock.
 274          */
 275         old_ptl = pmd_lock(vma->vm_mm, old_pmd);
 276         new_ptl = pmd_lockptr(mm, new_pmd);
 277         if (new_ptl != old_ptl)
 278                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 279
 280         /* Clear the pmd */
 281         pmd = *old_pmd;
 282         pmd_clear(old_pmd);
 283
 284         VM_BUG_ON(!pmd_none(*new_pmd));
 285
 286         pmd_populate(mm, new_pmd, pmd_pgtable(pmd));
 287         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
 288         if (new_ptl != old_ptl)
 289                 spin_unlock(new_ptl);
 290         spin_unlock(old_ptl);
 291
 292         return true;
 293 }
 294 #else
 295 static inline bool move_normal_pmd(struct vm_area_struct *vma,
 296                 unsigned long old_addr, unsigned long new_addr, pmd_t *old_pmd,
 297                 pmd_t *new_pmd)
 298 {
 299         return false;
 300 }
 301 #endif
 302
 303 #if CONFIG_PGTABLE_LEVELS > 2 && defined(CONFIG_HAVE_MOVE_PUD)
 304 static bool move_normal_pud(struct vm_area_struct *vma, unsigned long old_addr,
 305                   unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 306 {
 307         spinlock_t *old_ptl, *new_ptl;
 308         struct mm_struct *mm = vma->vm_mm;
 309         pud_t pud;
 310
 311         if (!arch_supports_page_table_move())
 312                 return false;
 313         /*
 314          * The destination pud shouldn't be established, free_pgtables()
 315          * should have released it.
 316          */
 317         if (WARN_ON_ONCE(!pud_none(*new_pud)))
 318                 return false;
 319
 320         /*
 321          * We don't have to worry about the ordering of src and dst
 322          * ptlocks because exclusive mmap_lock prevents deadlock.
 323          */
 324         old_ptl = pud_lock(vma->vm_mm, old_pud);
 325         new_ptl = pud_lockptr(mm, new_pud);
 326         if (new_ptl != old_ptl)
 327                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 328
 329         /* Clear the pud */
 330         pud = *old_pud;
 331         pud_clear(old_pud);
 332
 333         VM_BUG_ON(!pud_none(*new_pud));
 334
 335         pud_populate(mm, new_pud, pud_pgtable(pud));
 336         flush_tlb_range(vma, old_addr, old_addr + PUD_SIZE);
 337         if (new_ptl != old_ptl)
 338                 spin_unlock(new_ptl);
 339         spin_unlock(old_ptl);
 340
 341         return true;
 342 }
 343 #else
 344 static inline bool move_normal_pud(struct vm_area_struct *vma,
 345                 unsigned long old_addr, unsigned long new_addr, pud_t *old_pud,
 346                 pud_t *new_pud)
 347 {
 348         return false;
 349 }
 350 #endif
 351
 352 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 353 static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
 354                           unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 355 {
 356         spinlock_t *old_ptl, *new_ptl;
 357         struct mm_struct *mm = vma->vm_mm;
 358         pud_t pud;
 359
 360         /*
 361          * The destination pud shouldn't be established, free_pgtables()
 362          * should have released it.
 363          */
 364         if (WARN_ON_ONCE(!pud_none(*new_pud)))
 365                 return false;
 366
 367         /*
 368          * We don't have to worry about the ordering of src and dst
 369          * ptlocks because exclusive mmap_lock prevents deadlock.
 370          */
 371         old_ptl = pud_lock(vma->vm_mm, old_pud);
 372         new_ptl = pud_lockptr(mm, new_pud);
 373         if (new_ptl != old_ptl)
 374                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 375
 376         /* Clear the pud */
 377         pud = *old_pud;
 378         pud_clear(old_pud);
 379
 380         VM_BUG_ON(!pud_none(*new_pud));
 381
 382         /* Set the new pud */
 383         /* mark soft_ditry when we add pud level soft dirty support */
 384         set_pud_at(mm, new_addr, new_pud, pud);
 385         flush_pud_tlb_range(vma, old_addr, old_addr + HPAGE_PUD_SIZE);
 386         if (new_ptl != old_ptl)
 387                 spin_unlock(new_ptl);
 388         spin_unlock(old_ptl);
 389
 390         return true;
 391 }
 392 #else
 393 static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
 394                           unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 395 {
 396         WARN_ON_ONCE(1);
 397         return false;
 398
 399 }
 400 #endif
 401
 402 enum pgt_entry {
 403         NORMAL_PMD,
 404         HPAGE_PMD,
 405         NORMAL_PUD,
 406         HPAGE_PUD,
 407 };
 408
 409 /*
 410  * Returns an extent of the corresponding size for the pgt_entry specified if
 411  * valid. Else returns a smaller extent bounded by the end of the source and
 412  * destination pgt_entry.
 413  */
 414 static __always_inline unsigned long get_extent(enum pgt_entry entry,
 415                         unsigned long old_addr, unsigned long old_end,
 416                         unsigned long new_addr)
 417 {
 418         unsigned long next, extent, mask, size;
 419
 420         switch (entry) {
 421         case HPAGE_PMD:
 422         case NORMAL_PMD:
 423                 mask = PMD_MASK;
 424                 size = PMD_SIZE;
 425                 break;
 426         case HPAGE_PUD:
 427         case NORMAL_PUD:
 428                 mask = PUD_MASK;
 429                 size = PUD_SIZE;
 430                 break;
 431         default:
 432                 BUILD_BUG();
 433                 break;
 434         }
 435
 436         next = (old_addr + size) & mask;
 437         /* even if next overflowed, extent below will be ok */
 438         extent = next - old_addr;
 439         if (extent > old_end - old_addr)
 440                 extent = old_end - old_addr;
 441         next = (new_addr + size) & mask;
 442         if (extent > next - new_addr)
 443                 extent = next - new_addr;
 444         return extent;
 445 }
 446
 447 /*
 448  * Attempts to speedup the move by moving entry at the level corresponding to
 449  * pgt_entry. Returns true if the move was successful, else false.
 450  */
 451 static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
 452                         unsigned long old_addr, unsigned long new_addr,
 453                         void *old_entry, void *new_entry, bool need_rmap_locks)
 454 {
 455         bool moved = false;
 456
 457         /* See comment in move_ptes() */
 458         if (need_rmap_locks)
 459                 take_rmap_locks(vma);
 460
 461         switch (entry) {
 462         case NORMAL_PMD:
 463                 moved = move_normal_pmd(vma, old_addr, new_addr, old_entry,
 464                                         new_entry);
 465                 break;
 466         case NORMAL_PUD:
 467                 moved = move_normal_pud(vma, old_addr, new_addr, old_entry,
 468                                         new_entry);
 469                 break;
 470         case HPAGE_PMD:
 471                 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 472                         move_huge_pmd(vma, old_addr, new_addr, old_entry,
 473                                       new_entry);
 474                 break;
 475         case HPAGE_PUD:
 476                 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 477                         move_huge_pud(vma, old_addr, new_addr, old_entry,
 478                                       new_entry);
 479                 break;
 480
 481         default:
 482                 WARN_ON_ONCE(1);
 483                 break;
 484         }
 485
 486         if (need_rmap_locks)
 487                 drop_rmap_locks(vma);
 488
 489         return moved;
 490 }
 491
 492 unsigned long move_page_tables(struct vm_area_struct *vma,
 493                 unsigned long old_addr, struct vm_area_struct *new_vma,
 494                 unsigned long new_addr, unsigned long len,
 495                 bool need_rmap_locks)
 496 {
 497         unsigned long extent, old_end;
 498         struct mmu_notifier_range range;
 499         pmd_t *old_pmd, *new_pmd;
 500         pud_t *old_pud, *new_pud;
 501
 502         if (!len)
 503                 return 0;
 504
 505         old_end = old_addr + len;
 506
 507         if (is_vm_hugetlb_page(vma))
 508                 return move_hugetlb_page_tables(vma, new_vma, old_addr,
 509                                                 new_addr, len);
 510
 511         flush_cache_range(vma, old_addr, old_end);
 512         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma->vm_mm,
 513                                 old_addr, old_end);
 514         mmu_notifier_invalidate_range_start(&range);
 515
 516         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 517                 cond_resched();
 518                 /*
 519                  * If extent is PUD-sized try to speed up the move by moving at the
 520                  * PUD level if possible.
 521                  */
 522                 extent = get_extent(NORMAL_PUD, old_addr, old_end, new_addr);
 523
 524                 old_pud = get_old_pud(vma->vm_mm, old_addr);
 525                 if (!old_pud)
 526                         continue;
 527                 new_pud = alloc_new_pud(vma->vm_mm, vma, new_addr);
 528                 if (!new_pud)
 529                         break;
 530                 if (pud_trans_huge(*old_pud) || pud_devmap(*old_pud)) {
 531                         if (extent == HPAGE_PUD_SIZE) {
 532                                 move_pgt_entry(HPAGE_PUD, vma, old_addr, new_addr,
 533                                                old_pud, new_pud, need_rmap_locks);
 534                                 /* We ignore and continue on error? */
 535                                 continue;
 536                         }
 537                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
 538
 539                         if (move_pgt_entry(NORMAL_PUD, vma, old_addr, new_addr,
 540                                            old_pud, new_pud, true))
 541                                 continue;
 542                 }
 543
 544                 extent = get_extent(NORMAL_PMD, old_addr, old_end, new_addr);
 545                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 546                 if (!old_pmd)
 547                         continue;
 548                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 549                 if (!new_pmd)
 550                         break;
 551 again:
 552                 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) ||
 553                     pmd_devmap(*old_pmd)) {
 554                         if (extent == HPAGE_PMD_SIZE &&
 555                             move_pgt_entry(HPAGE_PMD, vma, old_addr, new_addr,
 556                                            old_pmd, new_pmd, need_rmap_locks))
 557                                 continue;
 558                         split_huge_pmd(vma, old_pmd, old_addr);
 559                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
 560                            extent == PMD_SIZE) {
 561                         /*
 562                          * If the extent is PMD-sized, try to speed the move by
 563                          * moving at the PMD level if possible.
 564                          */
 565                         if (move_pgt_entry(NORMAL_PMD, vma, old_addr, new_addr,
 566                                            old_pmd, new_pmd, true))
 567                                 continue;
 568                 }
 569                 if (pmd_none(*old_pmd))
 570                         continue;
 571                 if (pte_alloc(new_vma->vm_mm, new_pmd))
 572                         break;
 573                 if (move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 574                               new_vma, new_pmd, new_addr, need_rmap_locks) < 0)
 575                         goto again;
 576         }
 577
 578         mmu_notifier_invalidate_range_end(&range);
 579
 580         return len + old_addr - old_end;        /* how much done */
 581 }
 582
 583 static unsigned long move_vma(struct vm_area_struct *vma,
 584                 unsigned long old_addr, unsigned long old_len,
 585                 unsigned long new_len, unsigned long new_addr,
 586                 bool *locked, unsigned long flags,
 587                 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
 588 {
 589         long to_account = new_len - old_len;
 590         struct mm_struct *mm = vma->vm_mm;
 591         struct vm_area_struct *new_vma;
 592         unsigned long vm_flags = vma->vm_flags;
 593         unsigned long new_pgoff;
 594         unsigned long moved_len;
 595         unsigned long account_start = 0;
 596         unsigned long account_end = 0;
 597         unsigned long hiwater_vm;
 598         int err = 0;
 599         bool need_rmap_locks;
 600         struct vma_iterator vmi;
 601
 602         /*
 603          * We'd prefer to avoid failure later on in do_munmap:
 604          * which may split one vma into three before unmapping.
 605          */
 606         if (mm->map_count >= sysctl_max_map_count - 3)
 607                 return -ENOMEM;
 608
 609         if (unlikely(flags & MREMAP_DONTUNMAP))
 610                 to_account = new_len;
 611
 612         if (vma->vm_ops && vma->vm_ops->may_split) {
 613                 if (vma->vm_start != old_addr)
 614                         err = vma->vm_ops->may_split(vma, old_addr);
 615                 if (!err && vma->vm_end != old_addr + old_len)
 616                         err = vma->vm_ops->may_split(vma, old_addr + old_len);
 617                 if (err)
 618                         return err;
 619         }
 620
 621         /*
 622          * Advise KSM to break any KSM pages in the area to be moved:
 623          * it would be confusing if they were to turn up at the new
 624          * location, where they happen to coincide with different KSM
 625          * pages recently unmapped.  But leave vma->vm_flags as it was,
 626          * so KSM can come around to merge on vma and new_vma afterwards.
 627          */
 628         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 629                                                 MADV_UNMERGEABLE, &vm_flags);
 630         if (err)
 631                 return err;
 632
 633         if (vm_flags & VM_ACCOUNT) {
 634                 if (security_vm_enough_memory_mm(mm, to_account >> PAGE_SHIFT))
 635                         return -ENOMEM;
 636         }
 637
 638         vma_start_write(vma);
 639         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 640         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 641                            &need_rmap_locks);
 642         if (!new_vma) {
 643                 if (vm_flags & VM_ACCOUNT)
 644                         vm_unacct_memory(to_account >> PAGE_SHIFT);
 645                 return -ENOMEM;
 646         }
 647
 648         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 649                                      need_rmap_locks);
 650         if (moved_len < old_len) {
 651                 err = -ENOMEM;
 652         } else if (vma->vm_ops && vma->vm_ops->mremap) {
 653                 err = vma->vm_ops->mremap(new_vma);
 654         }
 655
 656         if (unlikely(err)) {
 657                 /*
 658                  * On error, move entries back from new area to old,
 659                  * which will succeed since page tables still there,
 660                  * and then proceed to unmap new area instead of old.
 661                  */
 662                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 663                                  true);
 664                 vma = new_vma;
 665                 old_len = new_len;
 666                 old_addr = new_addr;
 667                 new_addr = err;
 668         } else {
 669                 mremap_userfaultfd_prep(new_vma, uf);
 670         }
 671
 672         if (is_vm_hugetlb_page(vma)) {
 673                 clear_vma_resv_huge_pages(vma);
 674         }
 675
 676         /* Conceal VM_ACCOUNT so old reservation is not undone */
 677         if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP)) {
 678                 vm_flags_clear(vma, VM_ACCOUNT);
 679                 if (vma->vm_start < old_addr)
 680                         account_start = vma->vm_start;
 681                 if (vma->vm_end > old_addr + old_len)
 682                         account_end = vma->vm_end;
 683         }
 684
 685         /*
 686          * If we failed to move page tables we still do total_vm increment
 687          * since do_munmap() will decrement it by old_len == new_len.
 688          *
 689          * Since total_vm is about to be raised artificially high for a
 690          * moment, we need to restore high watermark afterwards: if stats
 691          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 692          * If this were a serious issue, we'd add a flag to do_munmap().
 693          */
 694         hiwater_vm = mm->hiwater_vm;
 695         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 696
 697         /* Tell pfnmap has moved from this vma */
 698         if (unlikely(vma->vm_flags & VM_PFNMAP))
 699                 untrack_pfn_clear(vma);
 700
 701         if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
 702                 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
 703                 vm_flags_clear(vma, VM_LOCKED_MASK);
 704
 705                 /*
 706                  * anon_vma links of the old vma is no longer needed after its page
 707                  * table has been moved.
 708                  */
 709                 if (new_vma != vma && vma->vm_start == old_addr &&
 710                         vma->vm_end == (old_addr + old_len))
 711                         unlink_anon_vmas(vma);
 712
 713                 /* Because we won't unmap we don't need to touch locked_vm */
 714                 return new_addr;
 715         }
 716
 717         vma_iter_init(&vmi, mm, old_addr);
 718         if (!do_vmi_munmap(&vmi, mm, old_addr, old_len, uf_unmap, false)) {
 719                 /* OOM: unable to split vma, just get accounts right */
 720                 if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP))
 721                         vm_acct_memory(old_len >> PAGE_SHIFT);
 722                 account_start = account_end = 0;
 723         }
 724
 725         if (vm_flags & VM_LOCKED) {
 726                 mm->locked_vm += new_len >> PAGE_SHIFT;
 727                 *locked = true;
 728         }
 729
 730         mm->hiwater_vm = hiwater_vm;
 731
 732         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 733         if (account_start) {
 734                 vma = vma_prev(&vmi);
 735                 vm_flags_set(vma, VM_ACCOUNT);
 736         }
 737
 738         if (account_end) {
 739                 vma = vma_next(&vmi);
 740                 vm_flags_set(vma, VM_ACCOUNT);
 741         }
 742
 743         return new_addr;
 744 }
 745
 746 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 747         unsigned long old_len, unsigned long new_len, unsigned long flags)
 748 {
 749         struct mm_struct *mm = current->mm;
 750         struct vm_area_struct *vma;
 751         unsigned long pgoff;
 752
 753         vma = vma_lookup(mm, addr);
 754         if (!vma)
 755                 return ERR_PTR(-EFAULT);
 756
 757         /*
 758          * !old_len is a special case where an attempt is made to 'duplicate'
 759          * a mapping.  This makes no sense for private mappings as it will
 760          * instead create a fresh/new mapping unrelated to the original.  This
 761          * is contrary to the basic idea of mremap which creates new mappings
 762          * based on the original.  There are no known use cases for this
 763          * behavior.  As a result, fail such attempts.
 764          */
 765         if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
 766                 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap.  This is not supported.\n", current->comm, current->pid);
 767                 return ERR_PTR(-EINVAL);
 768         }
 769
 770         if ((flags & MREMAP_DONTUNMAP) &&
 771                         (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)))
 772                 return ERR_PTR(-EINVAL);
 773
 774         /* We can't remap across vm area boundaries */
 775         if (old_len > vma->vm_end - addr)
 776                 return ERR_PTR(-EFAULT);
 777
 778         if (new_len == old_len)
 779                 return vma;
 780
 781         /* Need to be careful about a growing mapping */
 782         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 783         pgoff += vma->vm_pgoff;
 784         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 785                 return ERR_PTR(-EINVAL);
 786
 787         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 788                 return ERR_PTR(-EFAULT);
 789
 790         if (!mlock_future_ok(mm, vma->vm_flags, new_len - old_len))
 791                 return ERR_PTR(-EAGAIN);
 792
 793         if (!may_expand_vm(mm, vma->vm_flags,
 794                                 (new_len - old_len) >> PAGE_SHIFT))
 795                 return ERR_PTR(-ENOMEM);
 796
 797         return vma;
 798 }
 799
 800 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 801                 unsigned long new_addr, unsigned long new_len, bool *locked,
 802                 unsigned long flags, struct vm_userfaultfd_ctx *uf,
 803                 struct list_head *uf_unmap_early,
 804                 struct list_head *uf_unmap)
 805 {
 806         struct mm_struct *mm = current->mm;
 807         struct vm_area_struct *vma;
 808         unsigned long ret = -EINVAL;
 809         unsigned long map_flags = 0;
 810
 811         if (offset_in_page(new_addr))
 812                 goto out;
 813
 814         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 815                 goto out;
 816
 817         /* Ensure the old/new locations do not overlap */
 818         if (addr + old_len > new_addr && new_addr + new_len > addr)
 819                 goto out;
 820
 821         /*
 822          * move_vma() need us to stay 4 maps below the threshold, otherwise
 823          * it will bail out at the very beginning.
 824          * That is a problem if we have already unmaped the regions here
 825          * (new_addr, and old_addr), because userspace will not know the
 826          * state of the vma's after it gets -ENOMEM.
 827          * So, to avoid such scenario we can pre-compute if the whole
 828          * operation has high chances to success map-wise.
 829          * Worst-scenario case is when both vma's (new_addr and old_addr) get
 830          * split in 3 before unmapping it.
 831          * That means 2 more maps (1 for each) to the ones we already hold.
 832          * Check whether current map count plus 2 still leads us to 4 maps below
 833          * the threshold, otherwise return -ENOMEM here to be more safe.
 834          */
 835         if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
 836                 return -ENOMEM;
 837
 838         if (flags & MREMAP_FIXED) {
 839                 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
 840                 if (ret)
 841                         goto out;
 842         }
 843
 844         if (old_len > new_len) {
 845                 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
 846                 if (ret)
 847                         goto out;
 848                 old_len = new_len;
 849         }
 850
 851         vma = vma_to_resize(addr, old_len, new_len, flags);
 852         if (IS_ERR(vma)) {
 853                 ret = PTR_ERR(vma);
 854                 goto out;
 855         }
 856
 857         /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
 858         if (flags & MREMAP_DONTUNMAP &&
 859                 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
 860                 ret = -ENOMEM;
 861                 goto out;
 862         }
 863
 864         if (flags & MREMAP_FIXED)
 865                 map_flags |= MAP_FIXED;
 866
 867         if (vma->vm_flags & VM_MAYSHARE)
 868                 map_flags |= MAP_SHARED;
 869
 870         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 871                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 872                                 map_flags);
 873         if (IS_ERR_VALUE(ret))
 874                 goto out;
 875
 876         /* We got a new mapping */
 877         if (!(flags & MREMAP_FIXED))
 878                 new_addr = ret;
 879
 880         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
 881                        uf_unmap);
 882
 883 out:
 884         return ret;
 885 }
 886
 887 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 888 {
 889         unsigned long end = vma->vm_end + delta;
 890
 891         if (end < vma->vm_end) /* overflow */
 892                 return 0;
 893         if (find_vma_intersection(vma->vm_mm, vma->vm_end, end))
 894                 return 0;
 895         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 896                               0, MAP_FIXED) & ~PAGE_MASK)
 897                 return 0;
 898         return 1;
 899 }
 900
 901 /*
 902  * Expand (or shrink) an existing mapping, potentially moving it at the
 903  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 904  *
 905  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 906  * This option implies MREMAP_MAYMOVE.
 907  */
 908 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 909                 unsigned long, new_len, unsigned long, flags,
 910                 unsigned long, new_addr)
 911 {
 912         struct mm_struct *mm = current->mm;
 913         struct vm_area_struct *vma;
 914         unsigned long ret = -EINVAL;
 915         bool locked = false;
 916         struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
 917         LIST_HEAD(uf_unmap_early);
 918         LIST_HEAD(uf_unmap);
 919
 920         /*
 921          * There is a deliberate asymmetry here: we strip the pointer tag
 922          * from the old address but leave the new address alone. This is
 923          * for consistency with mmap(), where we prevent the creation of
 924          * aliasing mappings in userspace by leaving the tag bits of the
 925          * mapping address intact. A non-zero tag will cause the subsequent
 926          * range checks to reject the address as invalid.
 927          *
 928          * See Documentation/arch/arm64/tagged-address-abi.rst for more
 929          * information.
 930          */
 931         addr = untagged_addr(addr);
 932
 933         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
 934                 return ret;
 935
 936         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 937                 return ret;
 938
 939         /*
 940          * MREMAP_DONTUNMAP is always a move and it does not allow resizing
 941          * in the process.
 942          */
 943         if (flags & MREMAP_DONTUNMAP &&
 944                         (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
 945                 return ret;
 946
 947
 948         if (offset_in_page(addr))
 949                 return ret;
 950
 951         old_len = PAGE_ALIGN(old_len);
 952         new_len = PAGE_ALIGN(new_len);
 953
 954         /*
 955          * We allow a zero old-len as a special case
 956          * for DOS-emu "duplicate shm area" thing. But
 957          * a zero new-len is nonsensical.
 958          */
 959         if (!new_len)
 960                 return ret;
 961
 962         if (mmap_write_lock_killable(current->mm))
 963                 return -EINTR;
 964         vma = vma_lookup(mm, addr);
 965         if (!vma) {
 966                 ret = -EFAULT;
 967                 goto out;
 968         }
 969
 970         if (is_vm_hugetlb_page(vma)) {
 971                 struct hstate *h __maybe_unused = hstate_vma(vma);
 972
 973                 old_len = ALIGN(old_len, huge_page_size(h));
 974                 new_len = ALIGN(new_len, huge_page_size(h));
 975
 976                 /* addrs must be huge page aligned */
 977                 if (addr & ~huge_page_mask(h))
 978                         goto out;
 979                 if (new_addr & ~huge_page_mask(h))
 980                         goto out;
 981
 982                 /*
 983                  * Don't allow remap expansion, because the underlying hugetlb
 984                  * reservation is not yet capable to handle split reservation.
 985                  */
 986                 if (new_len > old_len)
 987                         goto out;
 988         }
 989
 990         if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
 991                 ret = mremap_to(addr, old_len, new_addr, new_len,
 992                                 &locked, flags, &uf, &uf_unmap_early,
 993                                 &uf_unmap);
 994                 goto out;
 995         }
 996
 997         /*
 998          * Always allow a shrinking remap: that just unmaps
 999          * the unnecessary pages..
1000          * do_vmi_munmap does all the needed commit accounting, and
1001          * unlocks the mmap_lock if so directed.
1002          */
1003         if (old_len >= new_len) {
1004                 VMA_ITERATOR(vmi, mm, addr + new_len);
1005
1006                 if (old_len == new_len) {
1007                         ret = addr;
1008                         goto out;
1009                 }
1010
1011                 ret = do_vmi_munmap(&vmi, mm, addr + new_len, old_len - new_len,
1012                                     &uf_unmap, true);
1013                 if (ret)
1014                         goto out;
1015
1016                 ret = addr;
1017                 goto out_unlocked;
1018         }
1019
1020         /*
1021          * Ok, we need to grow..
1022          */
1023         vma = vma_to_resize(addr, old_len, new_len, flags);
1024         if (IS_ERR(vma)) {
1025                 ret = PTR_ERR(vma);
1026                 goto out;
1027         }
1028
1029         /* old_len exactly to the end of the area..
1030          */
1031         if (old_len == vma->vm_end - addr) {
1032                 /* can we just expand the current mapping? */
1033                 if (vma_expandable(vma, new_len - old_len)) {
1034                         long pages = (new_len - old_len) >> PAGE_SHIFT;
1035                         unsigned long extension_start = addr + old_len;
1036                         unsigned long extension_end = addr + new_len;
1037                         pgoff_t extension_pgoff = vma->vm_pgoff +
1038                                 ((extension_start - vma->vm_start) >> PAGE_SHIFT);
1039                         VMA_ITERATOR(vmi, mm, extension_start);
1040
1041                         if (vma->vm_flags & VM_ACCOUNT) {
1042                                 if (security_vm_enough_memory_mm(mm, pages)) {
1043                                         ret = -ENOMEM;
1044                                         goto out;
1045                                 }
1046                         }
1047
1048                         /*
1049                          * Function vma_merge() is called on the extension we
1050                          * are adding to the already existing vma, vma_merge()
1051                          * will merge this extension with the already existing
1052                          * vma (expand operation itself) and possibly also with
1053                          * the next vma if it becomes adjacent to the expanded
1054                          * vma and  otherwise compatible.
1055                          */
1056                         vma = vma_merge(&vmi, mm, vma, extension_start,
1057                                 extension_end, vma->vm_flags, vma->anon_vma,
1058                                 vma->vm_file, extension_pgoff, vma_policy(vma),
1059                                 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
1060                         if (!vma) {
1061                                 vm_unacct_memory(pages);
1062                                 ret = -ENOMEM;
1063                                 goto out;
1064                         }
1065
1066                         vm_stat_account(mm, vma->vm_flags, pages);
1067                         if (vma->vm_flags & VM_LOCKED) {
1068                                 mm->locked_vm += pages;
1069                                 locked = true;
1070                                 new_addr = addr;
1071                         }
1072                         ret = addr;
1073                         goto out;
1074                 }
1075         }
1076
1077         /*
1078          * We weren't able to just expand or shrink the area,
1079          * we need to create a new one and move it..
1080          */
1081         ret = -ENOMEM;
1082         if (flags & MREMAP_MAYMOVE) {
1083                 unsigned long map_flags = 0;
1084                 if (vma->vm_flags & VM_MAYSHARE)
1085                         map_flags |= MAP_SHARED;
1086
1087                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
1088                                         vma->vm_pgoff +
1089                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
1090                                         map_flags);
1091                 if (IS_ERR_VALUE(new_addr)) {
1092                         ret = new_addr;
1093                         goto out;
1094                 }
1095
1096                 ret = move_vma(vma, addr, old_len, new_len, new_addr,
1097                                &locked, flags, &uf, &uf_unmap);
1098         }
1099 out:
1100         if (offset_in_page(ret))
1101                 locked = false;
1102         mmap_write_unlock(current->mm);
1103         if (locked && new_len > old_len)
1104                 mm_populate(new_addr + old_len, new_len - old_len);
1105 out_unlocked:
1106         userfaultfd_unmap_complete(mm, &uf_unmap_early);
1107         mremap_userfaultfd_complete(&uf, addr, ret, old_len);
1108         userfaultfd_unmap_complete(mm, &uf_unmap);
1109         return ret;
1110 }