mm/mremap.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *      mm/mremap.c
   4  *
   5  *      (C) Copyright 1996 Linus Torvalds
   6  *
   7  *      Address space accounting code   <alan@lxorguk.ukuu.org.uk>
   8  *      (C) Copyright 2002 Red Hat Inc, All Rights Reserved
   9  */
  10
  11 #include <linux/mm.h>
  12 #include <linux/hugetlb.h>
  13 #include <linux/shm.h>
  14 #include <linux/ksm.h>
  15 #include <linux/mman.h>
  16 #include <linux/swap.h>
  17 #include <linux/capability.h>
  18 #include <linux/fs.h>
  19 #include <linux/swapops.h>
  20 #include <linux/highmem.h>
  21 #include <linux/security.h>
  22 #include <linux/syscalls.h>
  23 #include <linux/mmu_notifier.h>
  24 #include <linux/uaccess.h>
  25 #include <linux/userfaultfd_k.h>
  26
  27 #include <asm/cacheflush.h>
  28 #include <asm/tlb.h>
  29 #include <asm/pgalloc.h>
  30
  31 #include "internal.h"
  32
  33 static pud_t *get_old_pud(struct mm_struct *mm, unsigned long addr)
  34 {
  35         pgd_t *pgd;
  36         p4d_t *p4d;
  37         pud_t *pud;
  38
  39         pgd = pgd_offset(mm, addr);
  40         if (pgd_none_or_clear_bad(pgd))
  41                 return NULL;
  42
  43         p4d = p4d_offset(pgd, addr);
  44         if (p4d_none_or_clear_bad(p4d))
  45                 return NULL;
  46
  47         pud = pud_offset(p4d, addr);
  48         if (pud_none_or_clear_bad(pud))
  49                 return NULL;
  50
  51         return pud;
  52 }
  53
  54 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  55 {
  56         pud_t *pud;
  57         pmd_t *pmd;
  58
  59         pud = get_old_pud(mm, addr);
  60         if (!pud)
  61                 return NULL;
  62
  63         pmd = pmd_offset(pud, addr);
  64         if (pmd_none(*pmd))
  65                 return NULL;
  66
  67         return pmd;
  68 }
  69
  70 static pud_t *alloc_new_pud(struct mm_struct *mm, struct vm_area_struct *vma,
  71                             unsigned long addr)
  72 {
  73         pgd_t *pgd;
  74         p4d_t *p4d;
  75
  76         pgd = pgd_offset(mm, addr);
  77         p4d = p4d_alloc(mm, pgd, addr);
  78         if (!p4d)
  79                 return NULL;
  80
  81         return pud_alloc(mm, p4d, addr);
  82 }
  83
  84 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
  85                             unsigned long addr)
  86 {
  87         pud_t *pud;
  88         pmd_t *pmd;
  89
  90         pud = alloc_new_pud(mm, vma, addr);
  91         if (!pud)
  92                 return NULL;
  93
  94         pmd = pmd_alloc(mm, pud, addr);
  95         if (!pmd)
  96                 return NULL;
  97
  98         VM_BUG_ON(pmd_trans_huge(*pmd));
  99
 100         return pmd;
 101 }
 102
 103 static void take_rmap_locks(struct vm_area_struct *vma)
 104 {
 105         if (vma->vm_file)
 106                 i_mmap_lock_write(vma->vm_file->f_mapping);
 107         if (vma->anon_vma)
 108                 anon_vma_lock_write(vma->anon_vma);
 109 }
 110
 111 static void drop_rmap_locks(struct vm_area_struct *vma)
 112 {
 113         if (vma->anon_vma)
 114                 anon_vma_unlock_write(vma->anon_vma);
 115         if (vma->vm_file)
 116                 i_mmap_unlock_write(vma->vm_file->f_mapping);
 117 }
 118
 119 static pte_t move_soft_dirty_pte(pte_t pte)
 120 {
 121         /*
 122          * Set soft dirty bit so we can notice
 123          * in userspace the ptes were moved.
 124          */
 125 #ifdef CONFIG_MEM_SOFT_DIRTY
 126         if (pte_present(pte))
 127                 pte = pte_mksoft_dirty(pte);
 128         else if (is_swap_pte(pte))
 129                 pte = pte_swp_mksoft_dirty(pte);
 130 #endif
 131         return pte;
 132 }
 133
 134 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
 135                 unsigned long old_addr, unsigned long old_end,
 136                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
 137                 unsigned long new_addr, bool need_rmap_locks)
 138 {
 139         struct mm_struct *mm = vma->vm_mm;
 140         pte_t *old_pte, *new_pte, pte;
 141         spinlock_t *old_ptl, *new_ptl;
 142         bool force_flush = false;
 143         unsigned long len = old_end - old_addr;
 144
 145         /*
 146          * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
 147          * locks to ensure that rmap will always observe either the old or the
 148          * new ptes. This is the easiest way to avoid races with
 149          * truncate_pagecache(), page migration, etc...
 150          *
 151          * When need_rmap_locks is false, we use other ways to avoid
 152          * such races:
 153          *
 154          * - During exec() shift_arg_pages(), we use a specially tagged vma
 155          *   which rmap call sites look for using vma_is_temporary_stack().
 156          *
 157          * - During mremap(), new_vma is often known to be placed after vma
 158          *   in rmap traversal order. This ensures rmap will always observe
 159          *   either the old pte, or the new pte, or both (the page table locks
 160          *   serialize access to individual ptes, but only rmap traversal
 161          *   order guarantees that we won't miss both the old and new ptes).
 162          */
 163         if (need_rmap_locks)
 164                 take_rmap_locks(vma);
 165
 166         /*
 167          * We don't have to worry about the ordering of src and dst
 168          * pte locks because exclusive mmap_lock prevents deadlock.
 169          */
 170         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 171         new_pte = pte_offset_map(new_pmd, new_addr);
 172         new_ptl = pte_lockptr(mm, new_pmd);
 173         if (new_ptl != old_ptl)
 174                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 175         flush_tlb_batched_pending(vma->vm_mm);
 176         arch_enter_lazy_mmu_mode();
 177
 178         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 179                                    new_pte++, new_addr += PAGE_SIZE) {
 180                 if (pte_none(*old_pte))
 181                         continue;
 182
 183                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
 184                 /*
 185                  * If we are remapping a valid PTE, make sure
 186                  * to flush TLB before we drop the PTL for the
 187                  * PTE.
 188                  *
 189                  * NOTE! Both old and new PTL matter: the old one
 190                  * for racing with page_mkclean(), the new one to
 191                  * make sure the physical page stays valid until
 192                  * the TLB entry for the old mapping has been
 193                  * flushed.
 194                  */
 195                 if (pte_present(pte))
 196                         force_flush = true;
 197                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 198                 pte = move_soft_dirty_pte(pte);
 199                 set_pte_at(mm, new_addr, new_pte, pte);
 200         }
 201
 202         arch_leave_lazy_mmu_mode();
 203         if (force_flush)
 204                 flush_tlb_range(vma, old_end - len, old_end);
 205         if (new_ptl != old_ptl)
 206                 spin_unlock(new_ptl);
 207         pte_unmap(new_pte - 1);
 208         pte_unmap_unlock(old_pte - 1, old_ptl);
 209         if (need_rmap_locks)
 210                 drop_rmap_locks(vma);
 211 }
 212
 213 #ifndef arch_supports_page_table_move
 214 #define arch_supports_page_table_move arch_supports_page_table_move
 215 static inline bool arch_supports_page_table_move(void)
 216 {
 217         return IS_ENABLED(CONFIG_HAVE_MOVE_PMD) ||
 218                 IS_ENABLED(CONFIG_HAVE_MOVE_PUD);
 219 }
 220 #endif
 221
 222 #ifdef CONFIG_HAVE_MOVE_PMD
 223 static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
 224                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
 225 {
 226         spinlock_t *old_ptl, *new_ptl;
 227         struct mm_struct *mm = vma->vm_mm;
 228         pmd_t pmd;
 229
 230         if (!arch_supports_page_table_move())
 231                 return false;
 232         /*
 233          * The destination pmd shouldn't be established, free_pgtables()
 234          * should have released it.
 235          *
 236          * However, there's a case during execve() where we use mremap
 237          * to move the initial stack, and in that case the target area
 238          * may overlap the source area (always moving down).
 239          *
 240          * If everything is PMD-aligned, that works fine, as moving
 241          * each pmd down will clear the source pmd. But if we first
 242          * have a few 4kB-only pages that get moved down, and then
 243          * hit the "now the rest is PMD-aligned, let's do everything
 244          * one pmd at a time", we will still have the old (now empty
 245          * of any 4kB pages, but still there) PMD in the page table
 246          * tree.
 247          *
 248          * Warn on it once - because we really should try to figure
 249          * out how to do this better - but then say "I won't move
 250          * this pmd".
 251          *
 252          * One alternative might be to just unmap the target pmd at
 253          * this point, and verify that it really is empty. We'll see.
 254          */
 255         if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
 256                 return false;
 257
 258         /*
 259          * We don't have to worry about the ordering of src and dst
 260          * ptlocks because exclusive mmap_lock prevents deadlock.
 261          */
 262         old_ptl = pmd_lock(vma->vm_mm, old_pmd);
 263         new_ptl = pmd_lockptr(mm, new_pmd);
 264         if (new_ptl != old_ptl)
 265                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 266
 267         /* Clear the pmd */
 268         pmd = *old_pmd;
 269         pmd_clear(old_pmd);
 270
 271         VM_BUG_ON(!pmd_none(*new_pmd));
 272
 273         pmd_populate(mm, new_pmd, pmd_pgtable(pmd));
 274         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
 275         if (new_ptl != old_ptl)
 276                 spin_unlock(new_ptl);
 277         spin_unlock(old_ptl);
 278
 279         return true;
 280 }
 281 #else
 282 static inline bool move_normal_pmd(struct vm_area_struct *vma,
 283                 unsigned long old_addr, unsigned long new_addr, pmd_t *old_pmd,
 284                 pmd_t *new_pmd)
 285 {
 286         return false;
 287 }
 288 #endif
 289
 290 #if CONFIG_PGTABLE_LEVELS > 2 && defined(CONFIG_HAVE_MOVE_PUD)
 291 static bool move_normal_pud(struct vm_area_struct *vma, unsigned long old_addr,
 292                   unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 293 {
 294         spinlock_t *old_ptl, *new_ptl;
 295         struct mm_struct *mm = vma->vm_mm;
 296         pud_t pud;
 297
 298         if (!arch_supports_page_table_move())
 299                 return false;
 300         /*
 301          * The destination pud shouldn't be established, free_pgtables()
 302          * should have released it.
 303          */
 304         if (WARN_ON_ONCE(!pud_none(*new_pud)))
 305                 return false;
 306
 307         /*
 308          * We don't have to worry about the ordering of src and dst
 309          * ptlocks because exclusive mmap_lock prevents deadlock.
 310          */
 311         old_ptl = pud_lock(vma->vm_mm, old_pud);
 312         new_ptl = pud_lockptr(mm, new_pud);
 313         if (new_ptl != old_ptl)
 314                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 315
 316         /* Clear the pud */
 317         pud = *old_pud;
 318         pud_clear(old_pud);
 319
 320         VM_BUG_ON(!pud_none(*new_pud));
 321
 322         pud_populate(mm, new_pud, pud_pgtable(pud));
 323         flush_tlb_range(vma, old_addr, old_addr + PUD_SIZE);
 324         if (new_ptl != old_ptl)
 325                 spin_unlock(new_ptl);
 326         spin_unlock(old_ptl);
 327
 328         return true;
 329 }
 330 #else
 331 static inline bool move_normal_pud(struct vm_area_struct *vma,
 332                 unsigned long old_addr, unsigned long new_addr, pud_t *old_pud,
 333                 pud_t *new_pud)
 334 {
 335         return false;
 336 }
 337 #endif
 338
 339 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 340 static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
 341                           unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 342 {
 343         spinlock_t *old_ptl, *new_ptl;
 344         struct mm_struct *mm = vma->vm_mm;
 345         pud_t pud;
 346
 347         /*
 348          * The destination pud shouldn't be established, free_pgtables()
 349          * should have released it.
 350          */
 351         if (WARN_ON_ONCE(!pud_none(*new_pud)))
 352                 return false;
 353
 354         /*
 355          * We don't have to worry about the ordering of src and dst
 356          * ptlocks because exclusive mmap_lock prevents deadlock.
 357          */
 358         old_ptl = pud_lock(vma->vm_mm, old_pud);
 359         new_ptl = pud_lockptr(mm, new_pud);
 360         if (new_ptl != old_ptl)
 361                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 362
 363         /* Clear the pud */
 364         pud = *old_pud;
 365         pud_clear(old_pud);
 366
 367         VM_BUG_ON(!pud_none(*new_pud));
 368
 369         /* Set the new pud */
 370         /* mark soft_ditry when we add pud level soft dirty support */
 371         set_pud_at(mm, new_addr, new_pud, pud);
 372         flush_pud_tlb_range(vma, old_addr, old_addr + HPAGE_PUD_SIZE);
 373         if (new_ptl != old_ptl)
 374                 spin_unlock(new_ptl);
 375         spin_unlock(old_ptl);
 376
 377         return true;
 378 }
 379 #else
 380 static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
 381                           unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 382 {
 383         WARN_ON_ONCE(1);
 384         return false;
 385
 386 }
 387 #endif
 388
 389 enum pgt_entry {
 390         NORMAL_PMD,
 391         HPAGE_PMD,
 392         NORMAL_PUD,
 393         HPAGE_PUD,
 394 };
 395
 396 /*
 397  * Returns an extent of the corresponding size for the pgt_entry specified if
 398  * valid. Else returns a smaller extent bounded by the end of the source and
 399  * destination pgt_entry.
 400  */
 401 static __always_inline unsigned long get_extent(enum pgt_entry entry,
 402                         unsigned long old_addr, unsigned long old_end,
 403                         unsigned long new_addr)
 404 {
 405         unsigned long next, extent, mask, size;
 406
 407         switch (entry) {
 408         case HPAGE_PMD:
 409         case NORMAL_PMD:
 410                 mask = PMD_MASK;
 411                 size = PMD_SIZE;
 412                 break;
 413         case HPAGE_PUD:
 414         case NORMAL_PUD:
 415                 mask = PUD_MASK;
 416                 size = PUD_SIZE;
 417                 break;
 418         default:
 419                 BUILD_BUG();
 420                 break;
 421         }
 422
 423         next = (old_addr + size) & mask;
 424         /* even if next overflowed, extent below will be ok */
 425         extent = next - old_addr;
 426         if (extent > old_end - old_addr)
 427                 extent = old_end - old_addr;
 428         next = (new_addr + size) & mask;
 429         if (extent > next - new_addr)
 430                 extent = next - new_addr;
 431         return extent;
 432 }
 433
 434 /*
 435  * Attempts to speedup the move by moving entry at the level corresponding to
 436  * pgt_entry. Returns true if the move was successful, else false.
 437  */
 438 static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
 439                         unsigned long old_addr, unsigned long new_addr,
 440                         void *old_entry, void *new_entry, bool need_rmap_locks)
 441 {
 442         bool moved = false;
 443
 444         /* See comment in move_ptes() */
 445         if (need_rmap_locks)
 446                 take_rmap_locks(vma);
 447
 448         switch (entry) {
 449         case NORMAL_PMD:
 450                 moved = move_normal_pmd(vma, old_addr, new_addr, old_entry,
 451                                         new_entry);
 452                 break;
 453         case NORMAL_PUD:
 454                 moved = move_normal_pud(vma, old_addr, new_addr, old_entry,
 455                                         new_entry);
 456                 break;
 457         case HPAGE_PMD:
 458                 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 459                         move_huge_pmd(vma, old_addr, new_addr, old_entry,
 460                                       new_entry);
 461                 break;
 462         case HPAGE_PUD:
 463                 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 464                         move_huge_pud(vma, old_addr, new_addr, old_entry,
 465                                       new_entry);
 466                 break;
 467
 468         default:
 469                 WARN_ON_ONCE(1);
 470                 break;
 471         }
 472
 473         if (need_rmap_locks)
 474                 drop_rmap_locks(vma);
 475
 476         return moved;
 477 }
 478
 479 unsigned long move_page_tables(struct vm_area_struct *vma,
 480                 unsigned long old_addr, struct vm_area_struct *new_vma,
 481                 unsigned long new_addr, unsigned long len,
 482                 bool need_rmap_locks)
 483 {
 484         unsigned long extent, old_end;
 485         struct mmu_notifier_range range;
 486         pmd_t *old_pmd, *new_pmd;
 487         pud_t *old_pud, *new_pud;
 488
 489         if (!len)
 490                 return 0;
 491
 492         old_end = old_addr + len;
 493         flush_cache_range(vma, old_addr, old_end);
 494
 495         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
 496                                 old_addr, old_end);
 497         mmu_notifier_invalidate_range_start(&range);
 498
 499         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 500                 cond_resched();
 501                 /*
 502                  * If extent is PUD-sized try to speed up the move by moving at the
 503                  * PUD level if possible.
 504                  */
 505                 extent = get_extent(NORMAL_PUD, old_addr, old_end, new_addr);
 506
 507                 old_pud = get_old_pud(vma->vm_mm, old_addr);
 508                 if (!old_pud)
 509                         continue;
 510                 new_pud = alloc_new_pud(vma->vm_mm, vma, new_addr);
 511                 if (!new_pud)
 512                         break;
 513                 if (pud_trans_huge(*old_pud) || pud_devmap(*old_pud)) {
 514                         if (extent == HPAGE_PUD_SIZE) {
 515                                 move_pgt_entry(HPAGE_PUD, vma, old_addr, new_addr,
 516                                                old_pud, new_pud, need_rmap_locks);
 517                                 /* We ignore and continue on error? */
 518                                 continue;
 519                         }
 520                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
 521
 522                         if (move_pgt_entry(NORMAL_PUD, vma, old_addr, new_addr,
 523                                            old_pud, new_pud, true))
 524                                 continue;
 525                 }
 526
 527                 extent = get_extent(NORMAL_PMD, old_addr, old_end, new_addr);
 528                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 529                 if (!old_pmd)
 530                         continue;
 531                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 532                 if (!new_pmd)
 533                         break;
 534                 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) ||
 535                     pmd_devmap(*old_pmd)) {
 536                         if (extent == HPAGE_PMD_SIZE &&
 537                             move_pgt_entry(HPAGE_PMD, vma, old_addr, new_addr,
 538                                            old_pmd, new_pmd, need_rmap_locks))
 539                                 continue;
 540                         split_huge_pmd(vma, old_pmd, old_addr);
 541                         if (pmd_trans_unstable(old_pmd))
 542                                 continue;
 543                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
 544                            extent == PMD_SIZE) {
 545                         /*
 546                          * If the extent is PMD-sized, try to speed the move by
 547                          * moving at the PMD level if possible.
 548                          */
 549                         if (move_pgt_entry(NORMAL_PMD, vma, old_addr, new_addr,
 550                                            old_pmd, new_pmd, true))
 551                                 continue;
 552                 }
 553
 554                 if (pte_alloc(new_vma->vm_mm, new_pmd))
 555                         break;
 556                 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
 557                           new_pmd, new_addr, need_rmap_locks);
 558         }
 559
 560         mmu_notifier_invalidate_range_end(&range);
 561
 562         return len + old_addr - old_end;        /* how much done */
 563 }
 564
 565 static unsigned long move_vma(struct vm_area_struct *vma,
 566                 unsigned long old_addr, unsigned long old_len,
 567                 unsigned long new_len, unsigned long new_addr,
 568                 bool *locked, unsigned long flags,
 569                 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
 570 {
 571         struct mm_struct *mm = vma->vm_mm;
 572         struct vm_area_struct *new_vma;
 573         unsigned long vm_flags = vma->vm_flags;
 574         unsigned long new_pgoff;
 575         unsigned long moved_len;
 576         unsigned long excess = 0;
 577         unsigned long hiwater_vm;
 578         int split = 0;
 579         int err = 0;
 580         bool need_rmap_locks;
 581
 582         /*
 583          * We'd prefer to avoid failure later on in do_munmap:
 584          * which may split one vma into three before unmapping.
 585          */
 586         if (mm->map_count >= sysctl_max_map_count - 3)
 587                 return -ENOMEM;
 588
 589         if (vma->vm_ops && vma->vm_ops->may_split) {
 590                 if (vma->vm_start != old_addr)
 591                         err = vma->vm_ops->may_split(vma, old_addr);
 592                 if (!err && vma->vm_end != old_addr + old_len)
 593                         err = vma->vm_ops->may_split(vma, old_addr + old_len);
 594                 if (err)
 595                         return err;
 596         }
 597
 598         /*
 599          * Advise KSM to break any KSM pages in the area to be moved:
 600          * it would be confusing if they were to turn up at the new
 601          * location, where they happen to coincide with different KSM
 602          * pages recently unmapped.  But leave vma->vm_flags as it was,
 603          * so KSM can come around to merge on vma and new_vma afterwards.
 604          */
 605         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 606                                                 MADV_UNMERGEABLE, &vm_flags);
 607         if (err)
 608                 return err;
 609
 610         if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT)) {
 611                 if (security_vm_enough_memory_mm(mm, new_len >> PAGE_SHIFT))
 612                         return -ENOMEM;
 613         }
 614
 615         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 616         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 617                            &need_rmap_locks);
 618         if (!new_vma) {
 619                 if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT))
 620                         vm_unacct_memory(new_len >> PAGE_SHIFT);
 621                 return -ENOMEM;
 622         }
 623
 624         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 625                                      need_rmap_locks);
 626         if (moved_len < old_len) {
 627                 err = -ENOMEM;
 628         } else if (vma->vm_ops && vma->vm_ops->mremap) {
 629                 err = vma->vm_ops->mremap(new_vma);
 630         }
 631
 632         if (unlikely(err)) {
 633                 /*
 634                  * On error, move entries back from new area to old,
 635                  * which will succeed since page tables still there,
 636                  * and then proceed to unmap new area instead of old.
 637                  */
 638                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 639                                  true);
 640                 vma = new_vma;
 641                 old_len = new_len;
 642                 old_addr = new_addr;
 643                 new_addr = err;
 644         } else {
 645                 mremap_userfaultfd_prep(new_vma, uf);
 646         }
 647
 648         /* Conceal VM_ACCOUNT so old reservation is not undone */
 649         if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP)) {
 650                 vma->vm_flags &= ~VM_ACCOUNT;
 651                 excess = vma->vm_end - vma->vm_start - old_len;
 652                 if (old_addr > vma->vm_start &&
 653                     old_addr + old_len < vma->vm_end)
 654                         split = 1;
 655         }
 656
 657         /*
 658          * If we failed to move page tables we still do total_vm increment
 659          * since do_munmap() will decrement it by old_len == new_len.
 660          *
 661          * Since total_vm is about to be raised artificially high for a
 662          * moment, we need to restore high watermark afterwards: if stats
 663          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 664          * If this were a serious issue, we'd add a flag to do_munmap().
 665          */
 666         hiwater_vm = mm->hiwater_vm;
 667         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 668
 669         /* Tell pfnmap has moved from this vma */
 670         if (unlikely(vma->vm_flags & VM_PFNMAP))
 671                 untrack_pfn_moved(vma);
 672
 673         if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
 674                 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
 675                 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
 676
 677                 /*
 678                  * anon_vma links of the old vma is no longer needed after its page
 679                  * table has been moved.
 680                  */
 681                 if (new_vma != vma && vma->vm_start == old_addr &&
 682                         vma->vm_end == (old_addr + old_len))
 683                         unlink_anon_vmas(vma);
 684
 685                 /* Because we won't unmap we don't need to touch locked_vm */
 686                 return new_addr;
 687         }
 688
 689         if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
 690                 /* OOM: unable to split vma, just get accounts right */
 691                 if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP))
 692                         vm_acct_memory(old_len >> PAGE_SHIFT);
 693                 excess = 0;
 694         }
 695
 696         if (vm_flags & VM_LOCKED) {
 697                 mm->locked_vm += new_len >> PAGE_SHIFT;
 698                 *locked = true;
 699         }
 700
 701         mm->hiwater_vm = hiwater_vm;
 702
 703         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 704         if (excess) {
 705                 vma->vm_flags |= VM_ACCOUNT;
 706                 if (split)
 707                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 708         }
 709
 710         return new_addr;
 711 }
 712
 713 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 714         unsigned long old_len, unsigned long new_len, unsigned long flags,
 715         unsigned long *p)
 716 {
 717         struct mm_struct *mm = current->mm;
 718         struct vm_area_struct *vma;
 719         unsigned long pgoff;
 720
 721         vma = vma_lookup(mm, addr);
 722         if (!vma)
 723                 return ERR_PTR(-EFAULT);
 724
 725         /*
 726          * !old_len is a special case where an attempt is made to 'duplicate'
 727          * a mapping.  This makes no sense for private mappings as it will
 728          * instead create a fresh/new mapping unrelated to the original.  This
 729          * is contrary to the basic idea of mremap which creates new mappings
 730          * based on the original.  There are no known use cases for this
 731          * behavior.  As a result, fail such attempts.
 732          */
 733         if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
 734                 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap.  This is not supported.\n", current->comm, current->pid);
 735                 return ERR_PTR(-EINVAL);
 736         }
 737
 738         if ((flags & MREMAP_DONTUNMAP) &&
 739                         (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)))
 740                 return ERR_PTR(-EINVAL);
 741
 742         if (is_vm_hugetlb_page(vma))
 743                 return ERR_PTR(-EINVAL);
 744
 745         /* We can't remap across vm area boundaries */
 746         if (old_len > vma->vm_end - addr)
 747                 return ERR_PTR(-EFAULT);
 748
 749         if (new_len == old_len)
 750                 return vma;
 751
 752         /* Need to be careful about a growing mapping */
 753         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 754         pgoff += vma->vm_pgoff;
 755         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 756                 return ERR_PTR(-EINVAL);
 757
 758         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 759                 return ERR_PTR(-EFAULT);
 760
 761         if (vma->vm_flags & VM_LOCKED) {
 762                 unsigned long locked, lock_limit;
 763                 locked = mm->locked_vm << PAGE_SHIFT;
 764                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 765                 locked += new_len - old_len;
 766                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 767                         return ERR_PTR(-EAGAIN);
 768         }
 769
 770         if (!may_expand_vm(mm, vma->vm_flags,
 771                                 (new_len - old_len) >> PAGE_SHIFT))
 772                 return ERR_PTR(-ENOMEM);
 773
 774         if (vma->vm_flags & VM_ACCOUNT) {
 775                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 776                 if (security_vm_enough_memory_mm(mm, charged))
 777                         return ERR_PTR(-ENOMEM);
 778                 *p = charged;
 779         }
 780
 781         return vma;
 782 }
 783
 784 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 785                 unsigned long new_addr, unsigned long new_len, bool *locked,
 786                 unsigned long flags, struct vm_userfaultfd_ctx *uf,
 787                 struct list_head *uf_unmap_early,
 788                 struct list_head *uf_unmap)
 789 {
 790         struct mm_struct *mm = current->mm;
 791         struct vm_area_struct *vma;
 792         unsigned long ret = -EINVAL;
 793         unsigned long charged = 0;
 794         unsigned long map_flags = 0;
 795
 796         if (offset_in_page(new_addr))
 797                 goto out;
 798
 799         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 800                 goto out;
 801
 802         /* Ensure the old/new locations do not overlap */
 803         if (addr + old_len > new_addr && new_addr + new_len > addr)
 804                 goto out;
 805
 806         /*
 807          * move_vma() need us to stay 4 maps below the threshold, otherwise
 808          * it will bail out at the very beginning.
 809          * That is a problem if we have already unmaped the regions here
 810          * (new_addr, and old_addr), because userspace will not know the
 811          * state of the vma's after it gets -ENOMEM.
 812          * So, to avoid such scenario we can pre-compute if the whole
 813          * operation has high chances to success map-wise.
 814          * Worst-scenario case is when both vma's (new_addr and old_addr) get
 815          * split in 3 before unmapping it.
 816          * That means 2 more maps (1 for each) to the ones we already hold.
 817          * Check whether current map count plus 2 still leads us to 4 maps below
 818          * the threshold, otherwise return -ENOMEM here to be more safe.
 819          */
 820         if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
 821                 return -ENOMEM;
 822
 823         if (flags & MREMAP_FIXED) {
 824                 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
 825                 if (ret)
 826                         goto out;
 827         }
 828
 829         if (old_len >= new_len) {
 830                 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
 831                 if (ret && old_len != new_len)
 832                         goto out;
 833                 old_len = new_len;
 834         }
 835
 836         vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
 837         if (IS_ERR(vma)) {
 838                 ret = PTR_ERR(vma);
 839                 goto out;
 840         }
 841
 842         /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
 843         if (flags & MREMAP_DONTUNMAP &&
 844                 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
 845                 ret = -ENOMEM;
 846                 goto out;
 847         }
 848
 849         if (flags & MREMAP_FIXED)
 850                 map_flags |= MAP_FIXED;
 851
 852         if (vma->vm_flags & VM_MAYSHARE)
 853                 map_flags |= MAP_SHARED;
 854
 855         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 856                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 857                                 map_flags);
 858         if (IS_ERR_VALUE(ret))
 859                 goto out1;
 860
 861         /* We got a new mapping */
 862         if (!(flags & MREMAP_FIXED))
 863                 new_addr = ret;
 864
 865         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
 866                        uf_unmap);
 867
 868         if (!(offset_in_page(ret)))
 869                 goto out;
 870
 871 out1:
 872         vm_unacct_memory(charged);
 873
 874 out:
 875         return ret;
 876 }
 877
 878 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 879 {
 880         unsigned long end = vma->vm_end + delta;
 881         if (end < vma->vm_end) /* overflow */
 882                 return 0;
 883         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 884                 return 0;
 885         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 886                               0, MAP_FIXED) & ~PAGE_MASK)
 887                 return 0;
 888         return 1;
 889 }
 890
 891 /*
 892  * Expand (or shrink) an existing mapping, potentially moving it at the
 893  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 894  *
 895  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 896  * This option implies MREMAP_MAYMOVE.
 897  */
 898 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 899                 unsigned long, new_len, unsigned long, flags,
 900                 unsigned long, new_addr)
 901 {
 902         struct mm_struct *mm = current->mm;
 903         struct vm_area_struct *vma;
 904         unsigned long ret = -EINVAL;
 905         unsigned long charged = 0;
 906         bool locked = false;
 907         bool downgraded = false;
 908         struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
 909         LIST_HEAD(uf_unmap_early);
 910         LIST_HEAD(uf_unmap);
 911
 912         /*
 913          * There is a deliberate asymmetry here: we strip the pointer tag
 914          * from the old address but leave the new address alone. This is
 915          * for consistency with mmap(), where we prevent the creation of
 916          * aliasing mappings in userspace by leaving the tag bits of the
 917          * mapping address intact. A non-zero tag will cause the subsequent
 918          * range checks to reject the address as invalid.
 919          *
 920          * See Documentation/arm64/tagged-address-abi.rst for more information.
 921          */
 922         addr = untagged_addr(addr);
 923
 924         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
 925                 return ret;
 926
 927         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 928                 return ret;
 929
 930         /*
 931          * MREMAP_DONTUNMAP is always a move and it does not allow resizing
 932          * in the process.
 933          */
 934         if (flags & MREMAP_DONTUNMAP &&
 935                         (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
 936                 return ret;
 937
 938
 939         if (offset_in_page(addr))
 940                 return ret;
 941
 942         old_len = PAGE_ALIGN(old_len);
 943         new_len = PAGE_ALIGN(new_len);
 944
 945         /*
 946          * We allow a zero old-len as a special case
 947          * for DOS-emu "duplicate shm area" thing. But
 948          * a zero new-len is nonsensical.
 949          */
 950         if (!new_len)
 951                 return ret;
 952
 953         if (mmap_write_lock_killable(current->mm))
 954                 return -EINTR;
 955
 956         if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
 957                 ret = mremap_to(addr, old_len, new_addr, new_len,
 958                                 &locked, flags, &uf, &uf_unmap_early,
 959                                 &uf_unmap);
 960                 goto out;
 961         }
 962
 963         /*
 964          * Always allow a shrinking remap: that just unmaps
 965          * the unnecessary pages..
 966          * __do_munmap does all the needed commit accounting, and
 967          * downgrades mmap_lock to read if so directed.
 968          */
 969         if (old_len >= new_len) {
 970                 int retval;
 971
 972                 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
 973                                   &uf_unmap, true);
 974                 if (retval < 0 && old_len != new_len) {
 975                         ret = retval;
 976                         goto out;
 977                 /* Returning 1 indicates mmap_lock is downgraded to read. */
 978                 } else if (retval == 1)
 979                         downgraded = true;
 980                 ret = addr;
 981                 goto out;
 982         }
 983
 984         /*
 985          * Ok, we need to grow..
 986          */
 987         vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
 988         if (IS_ERR(vma)) {
 989                 ret = PTR_ERR(vma);
 990                 goto out;
 991         }
 992
 993         /* old_len exactly to the end of the area..
 994          */
 995         if (old_len == vma->vm_end - addr) {
 996                 /* can we just expand the current mapping? */
 997                 if (vma_expandable(vma, new_len - old_len)) {
 998                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 999
1000                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
1001                                        vma->vm_pgoff, NULL)) {
1002                                 ret = -ENOMEM;
1003                                 goto out;
1004                         }
1005
1006                         vm_stat_account(mm, vma->vm_flags, pages);
1007                         if (vma->vm_flags & VM_LOCKED) {
1008                                 mm->locked_vm += pages;
1009                                 locked = true;
1010                                 new_addr = addr;
1011                         }
1012                         ret = addr;
1013                         goto out;
1014                 }
1015         }
1016
1017         /*
1018          * We weren't able to just expand or shrink the area,
1019          * we need to create a new one and move it..
1020          */
1021         ret = -ENOMEM;
1022         if (flags & MREMAP_MAYMOVE) {
1023                 unsigned long map_flags = 0;
1024                 if (vma->vm_flags & VM_MAYSHARE)
1025                         map_flags |= MAP_SHARED;
1026
1027                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
1028                                         vma->vm_pgoff +
1029                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
1030                                         map_flags);
1031                 if (IS_ERR_VALUE(new_addr)) {
1032                         ret = new_addr;
1033                         goto out;
1034                 }
1035
1036                 ret = move_vma(vma, addr, old_len, new_len, new_addr,
1037                                &locked, flags, &uf, &uf_unmap);
1038         }
1039 out:
1040         if (offset_in_page(ret)) {
1041                 vm_unacct_memory(charged);
1042                 locked = false;
1043         }
1044         if (downgraded)
1045                 mmap_read_unlock(current->mm);
1046         else
1047                 mmap_write_unlock(current->mm);
1048         if (locked && new_len > old_len)
1049                 mm_populate(new_addr + old_len, new_len - old_len);
1050         userfaultfd_unmap_complete(mm, &uf_unmap_early);
1051         mremap_userfaultfd_complete(&uf, addr, ret, old_len);
1052         userfaultfd_unmap_complete(mm, &uf_unmap);
1053         return ret;
1054 }