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   <[email protected]>
   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         pmd_t dummy_pmdval;
 144         spinlock_t *old_ptl, *new_ptl;
 145         bool force_flush = false;
 146         unsigned long len = old_end - old_addr;
 147         int err = 0;
 148
 149         /*
 150          * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
 151          * locks to ensure that rmap will always observe either the old or the
 152          * new ptes. This is the easiest way to avoid races with
 153          * truncate_pagecache(), page migration, etc...
 154          *
 155          * When need_rmap_locks is false, we use other ways to avoid
 156          * such races:
 157          *
 158          * - During exec() shift_arg_pages(), we use a specially tagged vma
 159          *   which rmap call sites look for using vma_is_temporary_stack().
 160          *
 161          * - During mremap(), new_vma is often known to be placed after vma
 162          *   in rmap traversal order. This ensures rmap will always observe
 163          *   either the old pte, or the new pte, or both (the page table locks
 164          *   serialize access to individual ptes, but only rmap traversal
 165          *   order guarantees that we won't miss both the old and new ptes).
 166          */
 167         if (need_rmap_locks)
 168                 take_rmap_locks(vma);
 169
 170         /*
 171          * We don't have to worry about the ordering of src and dst
 172          * pte locks because exclusive mmap_lock prevents deadlock.
 173          */
 174         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 175         if (!old_pte) {
 176                 err = -EAGAIN;
 177                 goto out;
 178         }
 179         /*
 180          * Now new_pte is none, so hpage_collapse_scan_file() path can not find
 181          * this by traversing file->f_mapping, so there is no concurrency with
 182          * retract_page_tables(). In addition, we already hold the exclusive
 183          * mmap_lock, so this new_pte page is stable, so there is no need to get
 184          * pmdval and do pmd_same() check.
 185          */
 186         new_pte = pte_offset_map_rw_nolock(mm, new_pmd, new_addr, &dummy_pmdval,
 187                                            &new_ptl);
 188         if (!new_pte) {
 189                 pte_unmap_unlock(old_pte, old_ptl);
 190                 err = -EAGAIN;
 191                 goto out;
 192         }
 193         if (new_ptl != old_ptl)
 194                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 195         flush_tlb_batched_pending(vma->vm_mm);
 196         arch_enter_lazy_mmu_mode();
 197
 198         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 199                                    new_pte++, new_addr += PAGE_SIZE) {
 200                 if (pte_none(ptep_get(old_pte)))
 201                         continue;
 202
 203                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
 204                 /*
 205                  * If we are remapping a valid PTE, make sure
 206                  * to flush TLB before we drop the PTL for the
 207                  * PTE.
 208                  *
 209                  * NOTE! Both old and new PTL matter: the old one
 210                  * for racing with folio_mkclean(), the new one to
 211                  * make sure the physical page stays valid until
 212                  * the TLB entry for the old mapping has been
 213                  * flushed.
 214                  */
 215                 if (pte_present(pte))
 216                         force_flush = true;
 217                 pte = move_pte(pte, old_addr, new_addr);
 218                 pte = move_soft_dirty_pte(pte);
 219                 set_pte_at(mm, new_addr, new_pte, pte);
 220         }
 221
 222         arch_leave_lazy_mmu_mode();
 223         if (force_flush)
 224                 flush_tlb_range(vma, old_end - len, old_end);
 225         if (new_ptl != old_ptl)
 226                 spin_unlock(new_ptl);
 227         pte_unmap(new_pte - 1);
 228         pte_unmap_unlock(old_pte - 1, old_ptl);
 229 out:
 230         if (need_rmap_locks)
 231                 drop_rmap_locks(vma);
 232         return err;
 233 }
 234
 235 #ifndef arch_supports_page_table_move
 236 #define arch_supports_page_table_move arch_supports_page_table_move
 237 static inline bool arch_supports_page_table_move(void)
 238 {
 239         return IS_ENABLED(CONFIG_HAVE_MOVE_PMD) ||
 240                 IS_ENABLED(CONFIG_HAVE_MOVE_PUD);
 241 }
 242 #endif
 243
 244 #ifdef CONFIG_HAVE_MOVE_PMD
 245 static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
 246                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
 247 {
 248         spinlock_t *old_ptl, *new_ptl;
 249         struct mm_struct *mm = vma->vm_mm;
 250         bool res = false;
 251         pmd_t pmd;
 252
 253         if (!arch_supports_page_table_move())
 254                 return false;
 255         /*
 256          * The destination pmd shouldn't be established, free_pgtables()
 257          * should have released it.
 258          *
 259          * However, there's a case during execve() where we use mremap
 260          * to move the initial stack, and in that case the target area
 261          * may overlap the source area (always moving down).
 262          *
 263          * If everything is PMD-aligned, that works fine, as moving
 264          * each pmd down will clear the source pmd. But if we first
 265          * have a few 4kB-only pages that get moved down, and then
 266          * hit the "now the rest is PMD-aligned, let's do everything
 267          * one pmd at a time", we will still have the old (now empty
 268          * of any 4kB pages, but still there) PMD in the page table
 269          * tree.
 270          *
 271          * Warn on it once - because we really should try to figure
 272          * out how to do this better - but then say "I won't move
 273          * this pmd".
 274          *
 275          * One alternative might be to just unmap the target pmd at
 276          * this point, and verify that it really is empty. We'll see.
 277          */
 278         if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
 279                 return false;
 280
 281         /*
 282          * We don't have to worry about the ordering of src and dst
 283          * ptlocks because exclusive mmap_lock prevents deadlock.
 284          */
 285         old_ptl = pmd_lock(vma->vm_mm, old_pmd);
 286         new_ptl = pmd_lockptr(mm, new_pmd);
 287         if (new_ptl != old_ptl)
 288                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 289
 290         pmd = *old_pmd;
 291
 292         /* Racing with collapse? */
 293         if (unlikely(!pmd_present(pmd) || pmd_leaf(pmd)))
 294                 goto out_unlock;
 295         /* Clear the pmd */
 296         pmd_clear(old_pmd);
 297         res = true;
 298
 299         VM_BUG_ON(!pmd_none(*new_pmd));
 300
 301         pmd_populate(mm, new_pmd, pmd_pgtable(pmd));
 302         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
 303 out_unlock:
 304         if (new_ptl != old_ptl)
 305                 spin_unlock(new_ptl);
 306         spin_unlock(old_ptl);
 307
 308         return res;
 309 }
 310 #else
 311 static inline bool move_normal_pmd(struct vm_area_struct *vma,
 312                 unsigned long old_addr, unsigned long new_addr, pmd_t *old_pmd,
 313                 pmd_t *new_pmd)
 314 {
 315         return false;
 316 }
 317 #endif
 318
 319 #if CONFIG_PGTABLE_LEVELS > 2 && defined(CONFIG_HAVE_MOVE_PUD)
 320 static bool move_normal_pud(struct vm_area_struct *vma, unsigned long old_addr,
 321                   unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 322 {
 323         spinlock_t *old_ptl, *new_ptl;
 324         struct mm_struct *mm = vma->vm_mm;
 325         pud_t pud;
 326
 327         if (!arch_supports_page_table_move())
 328                 return false;
 329         /*
 330          * The destination pud shouldn't be established, free_pgtables()
 331          * should have released it.
 332          */
 333         if (WARN_ON_ONCE(!pud_none(*new_pud)))
 334                 return false;
 335
 336         /*
 337          * We don't have to worry about the ordering of src and dst
 338          * ptlocks because exclusive mmap_lock prevents deadlock.
 339          */
 340         old_ptl = pud_lock(vma->vm_mm, old_pud);
 341         new_ptl = pud_lockptr(mm, new_pud);
 342         if (new_ptl != old_ptl)
 343                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 344
 345         /* Clear the pud */
 346         pud = *old_pud;
 347         pud_clear(old_pud);
 348
 349         VM_BUG_ON(!pud_none(*new_pud));
 350
 351         pud_populate(mm, new_pud, pud_pgtable(pud));
 352         flush_tlb_range(vma, old_addr, old_addr + PUD_SIZE);
 353         if (new_ptl != old_ptl)
 354                 spin_unlock(new_ptl);
 355         spin_unlock(old_ptl);
 356
 357         return true;
 358 }
 359 #else
 360 static inline bool move_normal_pud(struct vm_area_struct *vma,
 361                 unsigned long old_addr, unsigned long new_addr, pud_t *old_pud,
 362                 pud_t *new_pud)
 363 {
 364         return false;
 365 }
 366 #endif
 367
 368 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
 369 static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
 370                           unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 371 {
 372         spinlock_t *old_ptl, *new_ptl;
 373         struct mm_struct *mm = vma->vm_mm;
 374         pud_t pud;
 375
 376         /*
 377          * The destination pud shouldn't be established, free_pgtables()
 378          * should have released it.
 379          */
 380         if (WARN_ON_ONCE(!pud_none(*new_pud)))
 381                 return false;
 382
 383         /*
 384          * We don't have to worry about the ordering of src and dst
 385          * ptlocks because exclusive mmap_lock prevents deadlock.
 386          */
 387         old_ptl = pud_lock(vma->vm_mm, old_pud);
 388         new_ptl = pud_lockptr(mm, new_pud);
 389         if (new_ptl != old_ptl)
 390                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 391
 392         /* Clear the pud */
 393         pud = *old_pud;
 394         pud_clear(old_pud);
 395
 396         VM_BUG_ON(!pud_none(*new_pud));
 397
 398         /* Set the new pud */
 399         /* mark soft_ditry when we add pud level soft dirty support */
 400         set_pud_at(mm, new_addr, new_pud, pud);
 401         flush_pud_tlb_range(vma, old_addr, old_addr + HPAGE_PUD_SIZE);
 402         if (new_ptl != old_ptl)
 403                 spin_unlock(new_ptl);
 404         spin_unlock(old_ptl);
 405
 406         return true;
 407 }
 408 #else
 409 static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
 410                           unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 411 {
 412         WARN_ON_ONCE(1);
 413         return false;
 414
 415 }
 416 #endif
 417
 418 enum pgt_entry {
 419         NORMAL_PMD,
 420         HPAGE_PMD,
 421         NORMAL_PUD,
 422         HPAGE_PUD,
 423 };
 424
 425 /*
 426  * Returns an extent of the corresponding size for the pgt_entry specified if
 427  * valid. Else returns a smaller extent bounded by the end of the source and
 428  * destination pgt_entry.
 429  */
 430 static __always_inline unsigned long get_extent(enum pgt_entry entry,
 431                         unsigned long old_addr, unsigned long old_end,
 432                         unsigned long new_addr)
 433 {
 434         unsigned long next, extent, mask, size;
 435
 436         switch (entry) {
 437         case HPAGE_PMD:
 438         case NORMAL_PMD:
 439                 mask = PMD_MASK;
 440                 size = PMD_SIZE;
 441                 break;
 442         case HPAGE_PUD:
 443         case NORMAL_PUD:
 444                 mask = PUD_MASK;
 445                 size = PUD_SIZE;
 446                 break;
 447         default:
 448                 BUILD_BUG();
 449                 break;
 450         }
 451
 452         next = (old_addr + size) & mask;
 453         /* even if next overflowed, extent below will be ok */
 454         extent = next - old_addr;
 455         if (extent > old_end - old_addr)
 456                 extent = old_end - old_addr;
 457         next = (new_addr + size) & mask;
 458         if (extent > next - new_addr)
 459                 extent = next - new_addr;
 460         return extent;
 461 }
 462
 463 /*
 464  * Attempts to speedup the move by moving entry at the level corresponding to
 465  * pgt_entry. Returns true if the move was successful, else false.
 466  */
 467 static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
 468                         unsigned long old_addr, unsigned long new_addr,
 469                         void *old_entry, void *new_entry, bool need_rmap_locks)
 470 {
 471         bool moved = false;
 472
 473         /* See comment in move_ptes() */
 474         if (need_rmap_locks)
 475                 take_rmap_locks(vma);
 476
 477         switch (entry) {
 478         case NORMAL_PMD:
 479                 moved = move_normal_pmd(vma, old_addr, new_addr, old_entry,
 480                                         new_entry);
 481                 break;
 482         case NORMAL_PUD:
 483                 moved = move_normal_pud(vma, old_addr, new_addr, old_entry,
 484                                         new_entry);
 485                 break;
 486         case HPAGE_PMD:
 487                 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 488                         move_huge_pmd(vma, old_addr, new_addr, old_entry,
 489                                       new_entry);
 490                 break;
 491         case HPAGE_PUD:
 492                 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 493                         move_huge_pud(vma, old_addr, new_addr, old_entry,
 494                                       new_entry);
 495                 break;
 496
 497         default:
 498                 WARN_ON_ONCE(1);
 499                 break;
 500         }
 501
 502         if (need_rmap_locks)
 503                 drop_rmap_locks(vma);
 504
 505         return moved;
 506 }
 507
 508 /*
 509  * A helper to check if aligning down is OK. The aligned address should fall
 510  * on *no mapping*. For the stack moving down, that's a special move within
 511  * the VMA that is created to span the source and destination of the move,
 512  * so we make an exception for it.
 513  */
 514 static bool can_align_down(struct vm_area_struct *vma, unsigned long addr_to_align,
 515                             unsigned long mask, bool for_stack)
 516 {
 517         unsigned long addr_masked = addr_to_align & mask;
 518
 519         /*
 520          * If @addr_to_align of either source or destination is not the beginning
 521          * of the corresponding VMA, we can't align down or we will destroy part
 522          * of the current mapping.
 523          */
 524         if (!for_stack && vma->vm_start != addr_to_align)
 525                 return false;
 526
 527         /* In the stack case we explicitly permit in-VMA alignment. */
 528         if (for_stack && addr_masked >= vma->vm_start)
 529                 return true;
 530
 531         /*
 532          * Make sure the realignment doesn't cause the address to fall on an
 533          * existing mapping.
 534          */
 535         return find_vma_intersection(vma->vm_mm, addr_masked, vma->vm_start) == NULL;
 536 }
 537
 538 /* Opportunistically realign to specified boundary for faster copy. */
 539 static void try_realign_addr(unsigned long *old_addr, struct vm_area_struct *old_vma,
 540                              unsigned long *new_addr, struct vm_area_struct *new_vma,
 541                              unsigned long mask, bool for_stack)
 542 {
 543         /* Skip if the addresses are already aligned. */
 544         if ((*old_addr & ~mask) == 0)
 545                 return;
 546
 547         /* Only realign if the new and old addresses are mutually aligned. */
 548         if ((*old_addr & ~mask) != (*new_addr & ~mask))
 549                 return;
 550
 551         /* Ensure realignment doesn't cause overlap with existing mappings. */
 552         if (!can_align_down(old_vma, *old_addr, mask, for_stack) ||
 553             !can_align_down(new_vma, *new_addr, mask, for_stack))
 554                 return;
 555
 556         *old_addr = *old_addr & mask;
 557         *new_addr = *new_addr & mask;
 558 }
 559
 560 unsigned long move_page_tables(struct vm_area_struct *vma,
 561                 unsigned long old_addr, struct vm_area_struct *new_vma,
 562                 unsigned long new_addr, unsigned long len,
 563                 bool need_rmap_locks, bool for_stack)
 564 {
 565         unsigned long extent, old_end;
 566         struct mmu_notifier_range range;
 567         pmd_t *old_pmd, *new_pmd;
 568         pud_t *old_pud, *new_pud;
 569
 570         if (!len)
 571                 return 0;
 572
 573         old_end = old_addr + len;
 574
 575         if (is_vm_hugetlb_page(vma))
 576                 return move_hugetlb_page_tables(vma, new_vma, old_addr,
 577                                                 new_addr, len);
 578
 579         /*
 580          * If possible, realign addresses to PMD boundary for faster copy.
 581          * Only realign if the mremap copying hits a PMD boundary.
 582          */
 583         if (len >= PMD_SIZE - (old_addr & ~PMD_MASK))
 584                 try_realign_addr(&old_addr, vma, &new_addr, new_vma, PMD_MASK,
 585                                  for_stack);
 586
 587         flush_cache_range(vma, old_addr, old_end);
 588         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma->vm_mm,
 589                                 old_addr, old_end);
 590         mmu_notifier_invalidate_range_start(&range);
 591
 592         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 593                 cond_resched();
 594                 /*
 595                  * If extent is PUD-sized try to speed up the move by moving at the
 596                  * PUD level if possible.
 597                  */
 598                 extent = get_extent(NORMAL_PUD, old_addr, old_end, new_addr);
 599
 600                 old_pud = get_old_pud(vma->vm_mm, old_addr);
 601                 if (!old_pud)
 602                         continue;
 603                 new_pud = alloc_new_pud(vma->vm_mm, vma, new_addr);
 604                 if (!new_pud)
 605                         break;
 606                 if (pud_trans_huge(*old_pud) || pud_devmap(*old_pud)) {
 607                         if (extent == HPAGE_PUD_SIZE) {
 608                                 move_pgt_entry(HPAGE_PUD, vma, old_addr, new_addr,
 609                                                old_pud, new_pud, need_rmap_locks);
 610                                 /* We ignore and continue on error? */
 611                                 continue;
 612                         }
 613                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
 614
 615                         if (move_pgt_entry(NORMAL_PUD, vma, old_addr, new_addr,
 616                                            old_pud, new_pud, true))
 617                                 continue;
 618                 }
 619
 620                 extent = get_extent(NORMAL_PMD, old_addr, old_end, new_addr);
 621                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 622                 if (!old_pmd)
 623                         continue;
 624                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 625                 if (!new_pmd)
 626                         break;
 627 again:
 628                 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) ||
 629                     pmd_devmap(*old_pmd)) {
 630                         if (extent == HPAGE_PMD_SIZE &&
 631                             move_pgt_entry(HPAGE_PMD, vma, old_addr, new_addr,
 632                                            old_pmd, new_pmd, need_rmap_locks))
 633                                 continue;
 634                         split_huge_pmd(vma, old_pmd, old_addr);
 635                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
 636                            extent == PMD_SIZE) {
 637                         /*
 638                          * If the extent is PMD-sized, try to speed the move by
 639                          * moving at the PMD level if possible.
 640                          */
 641                         if (move_pgt_entry(NORMAL_PMD, vma, old_addr, new_addr,
 642                                            old_pmd, new_pmd, true))
 643                                 continue;
 644                 }
 645                 if (pmd_none(*old_pmd))
 646                         continue;
 647                 if (pte_alloc(new_vma->vm_mm, new_pmd))
 648                         break;
 649                 if (move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 650                               new_vma, new_pmd, new_addr, need_rmap_locks) < 0)
 651                         goto again;
 652         }
 653
 654         mmu_notifier_invalidate_range_end(&range);
 655
 656         /*
 657          * Prevent negative return values when {old,new}_addr was realigned
 658          * but we broke out of the above loop for the first PMD itself.
 659          */
 660         if (old_addr < old_end - len)
 661                 return 0;
 662
 663         return len + old_addr - old_end;        /* how much done */
 664 }
 665
 666 static unsigned long move_vma(struct vm_area_struct *vma,
 667                 unsigned long old_addr, unsigned long old_len,
 668                 unsigned long new_len, unsigned long new_addr,
 669                 bool *locked, unsigned long flags,
 670                 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
 671 {
 672         long to_account = new_len - old_len;
 673         struct mm_struct *mm = vma->vm_mm;
 674         struct vm_area_struct *new_vma;
 675         unsigned long vm_flags = vma->vm_flags;
 676         unsigned long new_pgoff;
 677         unsigned long moved_len;
 678         unsigned long account_start = 0;
 679         unsigned long account_end = 0;
 680         unsigned long hiwater_vm;
 681         int err = 0;
 682         bool need_rmap_locks;
 683         struct vma_iterator vmi;
 684
 685         /*
 686          * We'd prefer to avoid failure later on in do_munmap:
 687          * which may split one vma into three before unmapping.
 688          */
 689         if (mm->map_count >= sysctl_max_map_count - 3)
 690                 return -ENOMEM;
 691
 692         if (unlikely(flags & MREMAP_DONTUNMAP))
 693                 to_account = new_len;
 694
 695         if (vma->vm_ops && vma->vm_ops->may_split) {
 696                 if (vma->vm_start != old_addr)
 697                         err = vma->vm_ops->may_split(vma, old_addr);
 698                 if (!err && vma->vm_end != old_addr + old_len)
 699                         err = vma->vm_ops->may_split(vma, old_addr + old_len);
 700                 if (err)
 701                         return err;
 702         }
 703
 704         /*
 705          * Advise KSM to break any KSM pages in the area to be moved:
 706          * it would be confusing if they were to turn up at the new
 707          * location, where they happen to coincide with different KSM
 708          * pages recently unmapped.  But leave vma->vm_flags as it was,
 709          * so KSM can come around to merge on vma and new_vma afterwards.
 710          */
 711         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 712                                                 MADV_UNMERGEABLE, &vm_flags);
 713         if (err)
 714                 return err;
 715
 716         if (vm_flags & VM_ACCOUNT) {
 717                 if (security_vm_enough_memory_mm(mm, to_account >> PAGE_SHIFT))
 718                         return -ENOMEM;
 719         }
 720
 721         vma_start_write(vma);
 722         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 723         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 724                            &need_rmap_locks);
 725         if (!new_vma) {
 726                 if (vm_flags & VM_ACCOUNT)
 727                         vm_unacct_memory(to_account >> PAGE_SHIFT);
 728                 return -ENOMEM;
 729         }
 730
 731         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 732                                      need_rmap_locks, false);
 733         if (moved_len < old_len) {
 734                 err = -ENOMEM;
 735         } else if (vma->vm_ops && vma->vm_ops->mremap) {
 736                 err = vma->vm_ops->mremap(new_vma);
 737         }
 738
 739         if (unlikely(err)) {
 740                 /*
 741                  * On error, move entries back from new area to old,
 742                  * which will succeed since page tables still there,
 743                  * and then proceed to unmap new area instead of old.
 744                  */
 745                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 746                                  true, false);
 747                 vma = new_vma;
 748                 old_len = new_len;
 749                 old_addr = new_addr;
 750                 new_addr = err;
 751         } else {
 752                 mremap_userfaultfd_prep(new_vma, uf);
 753         }
 754
 755         if (is_vm_hugetlb_page(vma)) {
 756                 clear_vma_resv_huge_pages(vma);
 757         }
 758
 759         /* Conceal VM_ACCOUNT so old reservation is not undone */
 760         if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP)) {
 761                 vm_flags_clear(vma, VM_ACCOUNT);
 762                 if (vma->vm_start < old_addr)
 763                         account_start = vma->vm_start;
 764                 if (vma->vm_end > old_addr + old_len)
 765                         account_end = vma->vm_end;
 766         }
 767
 768         /*
 769          * If we failed to move page tables we still do total_vm increment
 770          * since do_munmap() will decrement it by old_len == new_len.
 771          *
 772          * Since total_vm is about to be raised artificially high for a
 773          * moment, we need to restore high watermark afterwards: if stats
 774          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 775          * If this were a serious issue, we'd add a flag to do_munmap().
 776          */
 777         hiwater_vm = mm->hiwater_vm;
 778         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 779
 780         /* Tell pfnmap has moved from this vma */
 781         if (unlikely(vma->vm_flags & VM_PFNMAP))
 782                 untrack_pfn_clear(vma);
 783
 784         if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
 785                 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
 786                 vm_flags_clear(vma, VM_LOCKED_MASK);
 787
 788                 /*
 789                  * anon_vma links of the old vma is no longer needed after its page
 790                  * table has been moved.
 791                  */
 792                 if (new_vma != vma && vma->vm_start == old_addr &&
 793                         vma->vm_end == (old_addr + old_len))
 794                         unlink_anon_vmas(vma);
 795
 796                 /* Because we won't unmap we don't need to touch locked_vm */
 797                 return new_addr;
 798         }
 799
 800         vma_iter_init(&vmi, mm, old_addr);
 801         if (do_vmi_munmap(&vmi, mm, old_addr, old_len, uf_unmap, false) < 0) {
 802                 /* OOM: unable to split vma, just get accounts right */
 803                 if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP))
 804                         vm_acct_memory(old_len >> PAGE_SHIFT);
 805                 account_start = account_end = 0;
 806         }
 807
 808         if (vm_flags & VM_LOCKED) {
 809                 mm->locked_vm += new_len >> PAGE_SHIFT;
 810                 *locked = true;
 811         }
 812
 813         mm->hiwater_vm = hiwater_vm;
 814
 815         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 816         if (account_start) {
 817                 vma = vma_prev(&vmi);
 818                 vm_flags_set(vma, VM_ACCOUNT);
 819         }
 820
 821         if (account_end) {
 822                 vma = vma_next(&vmi);
 823                 vm_flags_set(vma, VM_ACCOUNT);
 824         }
 825
 826         return new_addr;
 827 }
 828
 829 /*
 830  * resize_is_valid() - Ensure the vma can be resized to the new length at the give
 831  * address.
 832  *
 833  * @vma: The vma to resize
 834  * @addr: The old address
 835  * @old_len: The current size
 836  * @new_len: The desired size
 837  * @flags: The vma flags
 838  *
 839  * Return 0 on success, error otherwise.
 840  */
 841 static int resize_is_valid(struct vm_area_struct *vma, unsigned long addr,
 842         unsigned long old_len, unsigned long new_len, unsigned long flags)
 843 {
 844         struct mm_struct *mm = current->mm;
 845         unsigned long pgoff;
 846
 847         /*
 848          * !old_len is a special case where an attempt is made to 'duplicate'
 849          * a mapping.  This makes no sense for private mappings as it will
 850          * instead create a fresh/new mapping unrelated to the original.  This
 851          * is contrary to the basic idea of mremap which creates new mappings
 852          * based on the original.  There are no known use cases for this
 853          * behavior.  As a result, fail such attempts.
 854          */
 855         if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
 856                 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap.  This is not supported.\n", current->comm, current->pid);
 857                 return -EINVAL;
 858         }
 859
 860         if ((flags & MREMAP_DONTUNMAP) &&
 861                         (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)))
 862                 return -EINVAL;
 863
 864         /* We can't remap across vm area boundaries */
 865         if (old_len > vma->vm_end - addr)
 866                 return -EFAULT;
 867
 868         if (new_len == old_len)
 869                 return 0;
 870
 871         /* Need to be careful about a growing mapping */
 872         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 873         pgoff += vma->vm_pgoff;
 874         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 875                 return -EINVAL;
 876
 877         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 878                 return -EFAULT;
 879
 880         if (!mlock_future_ok(mm, vma->vm_flags, new_len - old_len))
 881                 return -EAGAIN;
 882
 883         if (!may_expand_vm(mm, vma->vm_flags,
 884                                 (new_len - old_len) >> PAGE_SHIFT))
 885                 return -ENOMEM;
 886
 887         return 0;
 888 }
 889
 890 /*
 891  * mremap_to() - remap a vma to a new location
 892  * @addr: The old address
 893  * @old_len: The old size
 894  * @new_addr: The target address
 895  * @new_len: The new size
 896  * @locked: If the returned vma is locked (VM_LOCKED)
 897  * @flags: the mremap flags
 898  * @uf: The mremap userfaultfd context
 899  * @uf_unmap_early: The userfaultfd unmap early context
 900  * @uf_unmap: The userfaultfd unmap context
 901  *
 902  * Returns: The new address of the vma or an error.
 903  */
 904 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 905                 unsigned long new_addr, unsigned long new_len, bool *locked,
 906                 unsigned long flags, struct vm_userfaultfd_ctx *uf,
 907                 struct list_head *uf_unmap_early,
 908                 struct list_head *uf_unmap)
 909 {
 910         struct mm_struct *mm = current->mm;
 911         struct vm_area_struct *vma;
 912         unsigned long ret;
 913         unsigned long map_flags = 0;
 914
 915         if (offset_in_page(new_addr))
 916                 return -EINVAL;
 917
 918         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 919                 return -EINVAL;
 920
 921         /* Ensure the old/new locations do not overlap */
 922         if (addr + old_len > new_addr && new_addr + new_len > addr)
 923                 return -EINVAL;
 924
 925         /*
 926          * move_vma() need us to stay 4 maps below the threshold, otherwise
 927          * it will bail out at the very beginning.
 928          * That is a problem if we have already unmaped the regions here
 929          * (new_addr, and old_addr), because userspace will not know the
 930          * state of the vma's after it gets -ENOMEM.
 931          * So, to avoid such scenario we can pre-compute if the whole
 932          * operation has high chances to success map-wise.
 933          * Worst-scenario case is when both vma's (new_addr and old_addr) get
 934          * split in 3 before unmapping it.
 935          * That means 2 more maps (1 for each) to the ones we already hold.
 936          * Check whether current map count plus 2 still leads us to 4 maps below
 937          * the threshold, otherwise return -ENOMEM here to be more safe.
 938          */
 939         if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
 940                 return -ENOMEM;
 941
 942         if (flags & MREMAP_FIXED) {
 943                 /*
 944                  * In mremap_to().
 945                  * VMA is moved to dst address, and munmap dst first.
 946                  * do_munmap will check if dst is sealed.
 947                  */
 948                 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
 949                 if (ret)
 950                         return ret;
 951         }
 952
 953         if (old_len > new_len) {
 954                 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
 955                 if (ret)
 956                         return ret;
 957                 old_len = new_len;
 958         }
 959
 960         vma = vma_lookup(mm, addr);
 961         if (!vma)
 962                 return -EFAULT;
 963
 964         ret = resize_is_valid(vma, addr, old_len, new_len, flags);
 965         if (ret)
 966                 return ret;
 967
 968         /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
 969         if (flags & MREMAP_DONTUNMAP &&
 970                 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
 971                 return -ENOMEM;
 972         }
 973
 974         if (flags & MREMAP_FIXED)
 975                 map_flags |= MAP_FIXED;
 976
 977         if (vma->vm_flags & VM_MAYSHARE)
 978                 map_flags |= MAP_SHARED;
 979
 980         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 981                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 982                                 map_flags);
 983         if (IS_ERR_VALUE(ret))
 984                 return ret;
 985
 986         /* We got a new mapping */
 987         if (!(flags & MREMAP_FIXED))
 988                 new_addr = ret;
 989
 990         return move_vma(vma, addr, old_len, new_len, new_addr, locked, flags,
 991                         uf, uf_unmap);
 992 }
 993
 994 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 995 {
 996         unsigned long end = vma->vm_end + delta;
 997
 998         if (end < vma->vm_end) /* overflow */
 999                 return 0;
1000         if (find_vma_intersection(vma->vm_mm, vma->vm_end, end))
1001                 return 0;
1002         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
1003                               0, MAP_FIXED) & ~PAGE_MASK)
1004                 return 0;
1005         return 1;
1006 }
1007
1008 /*
1009  * Expand (or shrink) an existing mapping, potentially moving it at the
1010  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
1011  *
1012  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
1013  * This option implies MREMAP_MAYMOVE.
1014  */
1015 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
1016                 unsigned long, new_len, unsigned long, flags,
1017                 unsigned long, new_addr)
1018 {
1019         struct mm_struct *mm = current->mm;
1020         struct vm_area_struct *vma;
1021         unsigned long ret = -EINVAL;
1022         bool locked = false;
1023         struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
1024         LIST_HEAD(uf_unmap_early);
1025         LIST_HEAD(uf_unmap);
1026
1027         /*
1028          * There is a deliberate asymmetry here: we strip the pointer tag
1029          * from the old address but leave the new address alone. This is
1030          * for consistency with mmap(), where we prevent the creation of
1031          * aliasing mappings in userspace by leaving the tag bits of the
1032          * mapping address intact. A non-zero tag will cause the subsequent
1033          * range checks to reject the address as invalid.
1034          *
1035          * See Documentation/arch/arm64/tagged-address-abi.rst for more
1036          * information.
1037          */
1038         addr = untagged_addr(addr);
1039
1040         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
1041                 return ret;
1042
1043         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
1044                 return ret;
1045
1046         /*
1047          * MREMAP_DONTUNMAP is always a move and it does not allow resizing
1048          * in the process.
1049          */
1050         if (flags & MREMAP_DONTUNMAP &&
1051                         (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
1052                 return ret;
1053
1054
1055         if (offset_in_page(addr))
1056                 return ret;
1057
1058         old_len = PAGE_ALIGN(old_len);
1059         new_len = PAGE_ALIGN(new_len);
1060
1061         /*
1062          * We allow a zero old-len as a special case
1063          * for DOS-emu "duplicate shm area" thing. But
1064          * a zero new-len is nonsensical.
1065          */
1066         if (!new_len)
1067                 return ret;
1068
1069         if (mmap_write_lock_killable(current->mm))
1070                 return -EINTR;
1071         vma = vma_lookup(mm, addr);
1072         if (!vma) {
1073                 ret = -EFAULT;
1074                 goto out;
1075         }
1076
1077         /* Don't allow remapping vmas when they have already been sealed */
1078         if (!can_modify_vma(vma)) {
1079                 ret = -EPERM;
1080                 goto out;
1081         }
1082
1083         if (is_vm_hugetlb_page(vma)) {
1084                 struct hstate *h __maybe_unused = hstate_vma(vma);
1085
1086                 old_len = ALIGN(old_len, huge_page_size(h));
1087                 new_len = ALIGN(new_len, huge_page_size(h));
1088
1089                 /* addrs must be huge page aligned */
1090                 if (addr & ~huge_page_mask(h))
1091                         goto out;
1092                 if (new_addr & ~huge_page_mask(h))
1093                         goto out;
1094
1095                 /*
1096                  * Don't allow remap expansion, because the underlying hugetlb
1097                  * reservation is not yet capable to handle split reservation.
1098                  */
1099                 if (new_len > old_len)
1100                         goto out;
1101         }
1102
1103         if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
1104                 ret = mremap_to(addr, old_len, new_addr, new_len,
1105                                 &locked, flags, &uf, &uf_unmap_early,
1106                                 &uf_unmap);
1107                 goto out;
1108         }
1109
1110         /*
1111          * Always allow a shrinking remap: that just unmaps
1112          * the unnecessary pages..
1113          * do_vmi_munmap does all the needed commit accounting, and
1114          * unlocks the mmap_lock if so directed.
1115          */
1116         if (old_len >= new_len) {
1117                 VMA_ITERATOR(vmi, mm, addr + new_len);
1118
1119                 if (old_len == new_len) {
1120                         ret = addr;
1121                         goto out;
1122                 }
1123
1124                 ret = do_vmi_munmap(&vmi, mm, addr + new_len, old_len - new_len,
1125                                     &uf_unmap, true);
1126                 if (ret)
1127                         goto out;
1128
1129                 ret = addr;
1130                 goto out_unlocked;
1131         }
1132
1133         /*
1134          * Ok, we need to grow..
1135          */
1136         ret = resize_is_valid(vma, addr, old_len, new_len, flags);
1137         if (ret)
1138                 goto out;
1139
1140         /* old_len exactly to the end of the area..
1141          */
1142         if (old_len == vma->vm_end - addr) {
1143                 unsigned long delta = new_len - old_len;
1144
1145                 /* can we just expand the current mapping? */
1146                 if (vma_expandable(vma, delta)) {
1147                         long pages = delta >> PAGE_SHIFT;
1148                         VMA_ITERATOR(vmi, mm, vma->vm_end);
1149                         long charged = 0;
1150
1151                         if (vma->vm_flags & VM_ACCOUNT) {
1152                                 if (security_vm_enough_memory_mm(mm, pages)) {
1153                                         ret = -ENOMEM;
1154                                         goto out;
1155                                 }
1156                                 charged = pages;
1157                         }
1158
1159                         /*
1160                          * Function vma_merge_extend() is called on the
1161                          * extension we are adding to the already existing vma,
1162                          * vma_merge_extend() will merge this extension with the
1163                          * already existing vma (expand operation itself) and
1164                          * possibly also with the next vma if it becomes
1165                          * adjacent to the expanded vma and otherwise
1166                          * compatible.
1167                          */
1168                         vma = vma_merge_extend(&vmi, vma, delta);
1169                         if (!vma) {
1170                                 vm_unacct_memory(charged);
1171                                 ret = -ENOMEM;
1172                                 goto out;
1173                         }
1174
1175                         vm_stat_account(mm, vma->vm_flags, pages);
1176                         if (vma->vm_flags & VM_LOCKED) {
1177                                 mm->locked_vm += pages;
1178                                 locked = true;
1179                                 new_addr = addr;
1180                         }
1181                         ret = addr;
1182                         goto out;
1183                 }
1184         }
1185
1186         /*
1187          * We weren't able to just expand or shrink the area,
1188          * we need to create a new one and move it..
1189          */
1190         ret = -ENOMEM;
1191         if (flags & MREMAP_MAYMOVE) {
1192                 unsigned long map_flags = 0;
1193                 if (vma->vm_flags & VM_MAYSHARE)
1194                         map_flags |= MAP_SHARED;
1195
1196                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
1197                                         vma->vm_pgoff +
1198                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
1199                                         map_flags);
1200                 if (IS_ERR_VALUE(new_addr)) {
1201                         ret = new_addr;
1202                         goto out;
1203                 }
1204
1205                 ret = move_vma(vma, addr, old_len, new_len, new_addr,
1206                                &locked, flags, &uf, &uf_unmap);
1207         }
1208 out:
1209         if (offset_in_page(ret))
1210                 locked = false;
1211         mmap_write_unlock(current->mm);
1212         if (locked && new_len > old_len)
1213                 mm_populate(new_addr + old_len, new_len - old_len);
1214 out_unlocked:
1215         userfaultfd_unmap_complete(mm, &uf_unmap_early);
1216         mremap_userfaultfd_complete(&uf, addr, ret, old_len);
1217         userfaultfd_unmap_complete(mm, &uf_unmap);
1218         return ret;
1219 }