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