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         old_end = old_addr + len;
 490         flush_cache_range(vma, old_addr, old_end);
 491
 492         if (is_vm_hugetlb_page(vma))
 493                 return move_hugetlb_page_tables(vma, new_vma, old_addr,
 494                                                 new_addr, len);
 495
 496         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
 497                                 old_addr, old_end);
 498         mmu_notifier_invalidate_range_start(&range);
 499
 500         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 501                 cond_resched();
 502                 /*
 503                  * If extent is PUD-sized try to speed up the move by moving at the
 504                  * PUD level if possible.
 505                  */
 506                 extent = get_extent(NORMAL_PUD, old_addr, old_end, new_addr);
 507
 508                 old_pud = get_old_pud(vma->vm_mm, old_addr);
 509                 if (!old_pud)
 510                         continue;
 511                 new_pud = alloc_new_pud(vma->vm_mm, vma, new_addr);
 512                 if (!new_pud)
 513                         break;
 514                 if (pud_trans_huge(*old_pud) || pud_devmap(*old_pud)) {
 515                         if (extent == HPAGE_PUD_SIZE) {
 516                                 move_pgt_entry(HPAGE_PUD, vma, old_addr, new_addr,
 517                                                old_pud, new_pud, need_rmap_locks);
 518                                 /* We ignore and continue on error? */
 519                                 continue;
 520                         }
 521                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
 522
 523                         if (move_pgt_entry(NORMAL_PUD, vma, old_addr, new_addr,
 524                                            old_pud, new_pud, true))
 525                                 continue;
 526                 }
 527
 528                 extent = get_extent(NORMAL_PMD, old_addr, old_end, new_addr);
 529                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 530                 if (!old_pmd)
 531                         continue;
 532                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 533                 if (!new_pmd)
 534                         break;
 535                 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) ||
 536                     pmd_devmap(*old_pmd)) {
 537                         if (extent == HPAGE_PMD_SIZE &&
 538                             move_pgt_entry(HPAGE_PMD, vma, old_addr, new_addr,
 539                                            old_pmd, new_pmd, need_rmap_locks))
 540                                 continue;
 541                         split_huge_pmd(vma, old_pmd, old_addr);
 542                         if (pmd_trans_unstable(old_pmd))
 543                                 continue;
 544                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
 545                            extent == PMD_SIZE) {
 546                         /*
 547                          * If the extent is PMD-sized, try to speed the move by
 548                          * moving at the PMD level if possible.
 549                          */
 550                         if (move_pgt_entry(NORMAL_PMD, vma, old_addr, new_addr,
 551                                            old_pmd, new_pmd, true))
 552                                 continue;
 553                 }
 554
 555                 if (pte_alloc(new_vma->vm_mm, new_pmd))
 556                         break;
 557                 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
 558                           new_pmd, new_addr, need_rmap_locks);
 559         }
 560
 561         mmu_notifier_invalidate_range_end(&range);
 562
 563         return len + old_addr - old_end;        /* how much done */
 564 }
 565
 566 static unsigned long move_vma(struct vm_area_struct *vma,
 567                 unsigned long old_addr, unsigned long old_len,
 568                 unsigned long new_len, unsigned long new_addr,
 569                 bool *locked, unsigned long flags,
 570                 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
 571 {
 572         long to_account = new_len - old_len;
 573         struct mm_struct *mm = vma->vm_mm;
 574         struct vm_area_struct *new_vma;
 575         unsigned long vm_flags = vma->vm_flags;
 576         unsigned long new_pgoff;
 577         unsigned long moved_len;
 578         unsigned long excess = 0;
 579         unsigned long hiwater_vm;
 580         int split = 0;
 581         int err = 0;
 582         bool need_rmap_locks;
 583
 584         /*
 585          * We'd prefer to avoid failure later on in do_munmap:
 586          * which may split one vma into three before unmapping.
 587          */
 588         if (mm->map_count >= sysctl_max_map_count - 3)
 589                 return -ENOMEM;
 590
 591         if (unlikely(flags & MREMAP_DONTUNMAP))
 592                 to_account = new_len;
 593
 594         if (vma->vm_ops && vma->vm_ops->may_split) {
 595                 if (vma->vm_start != old_addr)
 596                         err = vma->vm_ops->may_split(vma, old_addr);
 597                 if (!err && vma->vm_end != old_addr + old_len)
 598                         err = vma->vm_ops->may_split(vma, old_addr + old_len);
 599                 if (err)
 600                         return err;
 601         }
 602
 603         /*
 604          * Advise KSM to break any KSM pages in the area to be moved:
 605          * it would be confusing if they were to turn up at the new
 606          * location, where they happen to coincide with different KSM
 607          * pages recently unmapped.  But leave vma->vm_flags as it was,
 608          * so KSM can come around to merge on vma and new_vma afterwards.
 609          */
 610         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 611                                                 MADV_UNMERGEABLE, &vm_flags);
 612         if (err)
 613                 return err;
 614
 615         if (vm_flags & VM_ACCOUNT) {
 616                 if (security_vm_enough_memory_mm(mm, to_account >> PAGE_SHIFT))
 617                         return -ENOMEM;
 618         }
 619
 620         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 621         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 622                            &need_rmap_locks);
 623         if (!new_vma) {
 624                 if (vm_flags & VM_ACCOUNT)
 625                         vm_unacct_memory(to_account >> PAGE_SHIFT);
 626                 return -ENOMEM;
 627         }
 628
 629         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 630                                      need_rmap_locks);
 631         if (moved_len < old_len) {
 632                 err = -ENOMEM;
 633         } else if (vma->vm_ops && vma->vm_ops->mremap) {
 634                 err = vma->vm_ops->mremap(new_vma);
 635         }
 636
 637         if (unlikely(err)) {
 638                 /*
 639                  * On error, move entries back from new area to old,
 640                  * which will succeed since page tables still there,
 641                  * and then proceed to unmap new area instead of old.
 642                  */
 643                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 644                                  true);
 645                 vma = new_vma;
 646                 old_len = new_len;
 647                 old_addr = new_addr;
 648                 new_addr = err;
 649         } else {
 650                 mremap_userfaultfd_prep(new_vma, uf);
 651         }
 652
 653         if (is_vm_hugetlb_page(vma)) {
 654                 clear_vma_resv_huge_pages(vma);
 655         }
 656
 657         /* Conceal VM_ACCOUNT so old reservation is not undone */
 658         if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP)) {
 659                 vma->vm_flags &= ~VM_ACCOUNT;
 660                 excess = vma->vm_end - vma->vm_start - old_len;
 661                 if (old_addr > vma->vm_start &&
 662                     old_addr + old_len < vma->vm_end)
 663                         split = 1;
 664         }
 665
 666         /*
 667          * If we failed to move page tables we still do total_vm increment
 668          * since do_munmap() will decrement it by old_len == new_len.
 669          *
 670          * Since total_vm is about to be raised artificially high for a
 671          * moment, we need to restore high watermark afterwards: if stats
 672          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 673          * If this were a serious issue, we'd add a flag to do_munmap().
 674          */
 675         hiwater_vm = mm->hiwater_vm;
 676         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 677
 678         /* Tell pfnmap has moved from this vma */
 679         if (unlikely(vma->vm_flags & VM_PFNMAP))
 680                 untrack_pfn_moved(vma);
 681
 682         if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
 683                 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
 684                 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
 685
 686                 /*
 687                  * anon_vma links of the old vma is no longer needed after its page
 688                  * table has been moved.
 689                  */
 690                 if (new_vma != vma && vma->vm_start == old_addr &&
 691                         vma->vm_end == (old_addr + old_len))
 692                         unlink_anon_vmas(vma);
 693
 694                 /* Because we won't unmap we don't need to touch locked_vm */
 695                 return new_addr;
 696         }
 697
 698         if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
 699                 /* OOM: unable to split vma, just get accounts right */
 700                 if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP))
 701                         vm_acct_memory(old_len >> PAGE_SHIFT);
 702                 excess = 0;
 703         }
 704
 705         if (vm_flags & VM_LOCKED) {
 706                 mm->locked_vm += new_len >> PAGE_SHIFT;
 707                 *locked = true;
 708         }
 709
 710         mm->hiwater_vm = hiwater_vm;
 711
 712         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 713         if (excess) {
 714                 vma->vm_flags |= VM_ACCOUNT;
 715                 if (split)
 716                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 717         }
 718
 719         return new_addr;
 720 }
 721
 722 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 723         unsigned long old_len, unsigned long new_len, unsigned long flags)
 724 {
 725         struct mm_struct *mm = current->mm;
 726         struct vm_area_struct *vma;
 727         unsigned long pgoff;
 728
 729         vma = vma_lookup(mm, addr);
 730         if (!vma)
 731                 return ERR_PTR(-EFAULT);
 732
 733         /*
 734          * !old_len is a special case where an attempt is made to 'duplicate'
 735          * a mapping.  This makes no sense for private mappings as it will
 736          * instead create a fresh/new mapping unrelated to the original.  This
 737          * is contrary to the basic idea of mremap which creates new mappings
 738          * based on the original.  There are no known use cases for this
 739          * behavior.  As a result, fail such attempts.
 740          */
 741         if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
 742                 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap.  This is not supported.\n", current->comm, current->pid);
 743                 return ERR_PTR(-EINVAL);
 744         }
 745
 746         if ((flags & MREMAP_DONTUNMAP) &&
 747                         (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)))
 748                 return ERR_PTR(-EINVAL);
 749
 750         /* We can't remap across vm area boundaries */
 751         if (old_len > vma->vm_end - addr)
 752                 return ERR_PTR(-EFAULT);
 753
 754         if (new_len == old_len)
 755                 return vma;
 756
 757         /* Need to be careful about a growing mapping */
 758         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 759         pgoff += vma->vm_pgoff;
 760         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 761                 return ERR_PTR(-EINVAL);
 762
 763         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 764                 return ERR_PTR(-EFAULT);
 765
 766         if (vma->vm_flags & VM_LOCKED) {
 767                 unsigned long locked, lock_limit;
 768                 locked = mm->locked_vm << PAGE_SHIFT;
 769                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 770                 locked += new_len - old_len;
 771                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 772                         return ERR_PTR(-EAGAIN);
 773         }
 774
 775         if (!may_expand_vm(mm, vma->vm_flags,
 776                                 (new_len - old_len) >> PAGE_SHIFT))
 777                 return ERR_PTR(-ENOMEM);
 778
 779         return vma;
 780 }
 781
 782 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 783                 unsigned long new_addr, unsigned long new_len, bool *locked,
 784                 unsigned long flags, struct vm_userfaultfd_ctx *uf,
 785                 struct list_head *uf_unmap_early,
 786                 struct list_head *uf_unmap)
 787 {
 788         struct mm_struct *mm = current->mm;
 789         struct vm_area_struct *vma;
 790         unsigned long ret = -EINVAL;
 791         unsigned long map_flags = 0;
 792
 793         if (offset_in_page(new_addr))
 794                 goto out;
 795
 796         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 797                 goto out;
 798
 799         /* Ensure the old/new locations do not overlap */
 800         if (addr + old_len > new_addr && new_addr + new_len > addr)
 801                 goto out;
 802
 803         /*
 804          * move_vma() need us to stay 4 maps below the threshold, otherwise
 805          * it will bail out at the very beginning.
 806          * That is a problem if we have already unmaped the regions here
 807          * (new_addr, and old_addr), because userspace will not know the
 808          * state of the vma's after it gets -ENOMEM.
 809          * So, to avoid such scenario we can pre-compute if the whole
 810          * operation has high chances to success map-wise.
 811          * Worst-scenario case is when both vma's (new_addr and old_addr) get
 812          * split in 3 before unmapping it.
 813          * That means 2 more maps (1 for each) to the ones we already hold.
 814          * Check whether current map count plus 2 still leads us to 4 maps below
 815          * the threshold, otherwise return -ENOMEM here to be more safe.
 816          */
 817         if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
 818                 return -ENOMEM;
 819
 820         if (flags & MREMAP_FIXED) {
 821                 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
 822                 if (ret)
 823                         goto out;
 824         }
 825
 826         if (old_len >= new_len) {
 827                 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
 828                 if (ret && old_len != new_len)
 829                         goto out;
 830                 old_len = new_len;
 831         }
 832
 833         vma = vma_to_resize(addr, old_len, new_len, flags);
 834         if (IS_ERR(vma)) {
 835                 ret = PTR_ERR(vma);
 836                 goto out;
 837         }
 838
 839         /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
 840         if (flags & MREMAP_DONTUNMAP &&
 841                 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
 842                 ret = -ENOMEM;
 843                 goto out;
 844         }
 845
 846         if (flags & MREMAP_FIXED)
 847                 map_flags |= MAP_FIXED;
 848
 849         if (vma->vm_flags & VM_MAYSHARE)
 850                 map_flags |= MAP_SHARED;
 851
 852         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 853                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 854                                 map_flags);
 855         if (IS_ERR_VALUE(ret))
 856                 goto out;
 857
 858         /* We got a new mapping */
 859         if (!(flags & MREMAP_FIXED))
 860                 new_addr = ret;
 861
 862         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
 863                        uf_unmap);
 864
 865 out:
 866         return ret;
 867 }
 868
 869 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 870 {
 871         unsigned long end = vma->vm_end + delta;
 872         if (end < vma->vm_end) /* overflow */
 873                 return 0;
 874         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 875                 return 0;
 876         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 877                               0, MAP_FIXED) & ~PAGE_MASK)
 878                 return 0;
 879         return 1;
 880 }
 881
 882 /*
 883  * Expand (or shrink) an existing mapping, potentially moving it at the
 884  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 885  *
 886  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 887  * This option implies MREMAP_MAYMOVE.
 888  */
 889 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 890                 unsigned long, new_len, unsigned long, flags,
 891                 unsigned long, new_addr)
 892 {
 893         struct mm_struct *mm = current->mm;
 894         struct vm_area_struct *vma;
 895         unsigned long ret = -EINVAL;
 896         bool locked = false;
 897         bool downgraded = false;
 898         struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
 899         LIST_HEAD(uf_unmap_early);
 900         LIST_HEAD(uf_unmap);
 901
 902         /*
 903          * There is a deliberate asymmetry here: we strip the pointer tag
 904          * from the old address but leave the new address alone. This is
 905          * for consistency with mmap(), where we prevent the creation of
 906          * aliasing mappings in userspace by leaving the tag bits of the
 907          * mapping address intact. A non-zero tag will cause the subsequent
 908          * range checks to reject the address as invalid.
 909          *
 910          * See Documentation/arm64/tagged-address-abi.rst for more information.
 911          */
 912         addr = untagged_addr(addr);
 913
 914         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
 915                 return ret;
 916
 917         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 918                 return ret;
 919
 920         /*
 921          * MREMAP_DONTUNMAP is always a move and it does not allow resizing
 922          * in the process.
 923          */
 924         if (flags & MREMAP_DONTUNMAP &&
 925                         (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
 926                 return ret;
 927
 928
 929         if (offset_in_page(addr))
 930                 return ret;
 931
 932         old_len = PAGE_ALIGN(old_len);
 933         new_len = PAGE_ALIGN(new_len);
 934
 935         /*
 936          * We allow a zero old-len as a special case
 937          * for DOS-emu "duplicate shm area" thing. But
 938          * a zero new-len is nonsensical.
 939          */
 940         if (!new_len)
 941                 return ret;
 942
 943         if (mmap_write_lock_killable(current->mm))
 944                 return -EINTR;
 945         vma = vma_lookup(mm, addr);
 946         if (!vma) {
 947                 ret = EFAULT;
 948                 goto out;
 949         }
 950
 951         if (is_vm_hugetlb_page(vma)) {
 952                 struct hstate *h __maybe_unused = hstate_vma(vma);
 953
 954                 old_len = ALIGN(old_len, huge_page_size(h));
 955                 new_len = ALIGN(new_len, huge_page_size(h));
 956
 957                 /* addrs must be huge page aligned */
 958                 if (addr & ~huge_page_mask(h))
 959                         goto out;
 960                 if (new_addr & ~huge_page_mask(h))
 961                         goto out;
 962
 963                 /*
 964                  * Don't allow remap expansion, because the underlying hugetlb
 965                  * reservation is not yet capable to handle split reservation.
 966                  */
 967                 if (new_len > old_len)
 968                         goto out;
 969         }
 970
 971         if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
 972                 ret = mremap_to(addr, old_len, new_addr, new_len,
 973                                 &locked, flags, &uf, &uf_unmap_early,
 974                                 &uf_unmap);
 975                 goto out;
 976         }
 977
 978         /*
 979          * Always allow a shrinking remap: that just unmaps
 980          * the unnecessary pages..
 981          * __do_munmap does all the needed commit accounting, and
 982          * downgrades mmap_lock to read if so directed.
 983          */
 984         if (old_len >= new_len) {
 985                 int retval;
 986
 987                 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
 988                                   &uf_unmap, true);
 989                 if (retval < 0 && old_len != new_len) {
 990                         ret = retval;
 991                         goto out;
 992                 /* Returning 1 indicates mmap_lock is downgraded to read. */
 993                 } else if (retval == 1)
 994                         downgraded = true;
 995                 ret = addr;
 996                 goto out;
 997         }
 998
 999         /*
1000          * Ok, we need to grow..
1001          */
1002         vma = vma_to_resize(addr, old_len, new_len, flags);
1003         if (IS_ERR(vma)) {
1004                 ret = PTR_ERR(vma);
1005                 goto out;
1006         }
1007
1008         /* old_len exactly to the end of the area..
1009          */
1010         if (old_len == vma->vm_end - addr) {
1011                 /* can we just expand the current mapping? */
1012                 if (vma_expandable(vma, new_len - old_len)) {
1013                         long pages = (new_len - old_len) >> PAGE_SHIFT;
1014
1015                         if (vma->vm_flags & VM_ACCOUNT) {
1016                                 if (security_vm_enough_memory_mm(mm, pages)) {
1017                                         ret = -ENOMEM;
1018                                         goto out;
1019                                 }
1020                         }
1021
1022                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
1023                                        vma->vm_pgoff, NULL)) {
1024                                 vm_unacct_memory(pages);
1025                                 ret = -ENOMEM;
1026                                 goto out;
1027                         }
1028
1029                         vm_stat_account(mm, vma->vm_flags, pages);
1030                         if (vma->vm_flags & VM_LOCKED) {
1031                                 mm->locked_vm += pages;
1032                                 locked = true;
1033                                 new_addr = addr;
1034                         }
1035                         ret = addr;
1036                         goto out;
1037                 }
1038         }
1039
1040         /*
1041          * We weren't able to just expand or shrink the area,
1042          * we need to create a new one and move it..
1043          */
1044         ret = -ENOMEM;
1045         if (flags & MREMAP_MAYMOVE) {
1046                 unsigned long map_flags = 0;
1047                 if (vma->vm_flags & VM_MAYSHARE)
1048                         map_flags |= MAP_SHARED;
1049
1050                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
1051                                         vma->vm_pgoff +
1052                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
1053                                         map_flags);
1054                 if (IS_ERR_VALUE(new_addr)) {
1055                         ret = new_addr;
1056                         goto out;
1057                 }
1058
1059                 ret = move_vma(vma, addr, old_len, new_len, new_addr,
1060                                &locked, flags, &uf, &uf_unmap);
1061         }
1062 out:
1063         if (offset_in_page(ret))
1064                 locked = false;
1065         if (downgraded)
1066                 mmap_read_unlock(current->mm);
1067         else
1068                 mmap_write_unlock(current->mm);
1069         if (locked && new_len > old_len)
1070                 mm_populate(new_addr + old_len, new_len - old_len);
1071         userfaultfd_unmap_complete(mm, &uf_unmap_early);
1072         mremap_userfaultfd_complete(&uf, addr, ret, old_len);
1073         userfaultfd_unmap_complete(mm, &uf_unmap);
1074         return ret;
1075 }