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