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/tlbflush.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 #ifdef CONFIG_HAVE_MOVE_PMD
 214 static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
 215                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
 216 {
 217         spinlock_t *old_ptl, *new_ptl;
 218         struct mm_struct *mm = vma->vm_mm;
 219         pmd_t pmd;
 220
 221         /*
 222          * The destination pmd shouldn't be established, free_pgtables()
 223          * should have released it.
 224          *
 225          * However, there's a case during execve() where we use mremap
 226          * to move the initial stack, and in that case the target area
 227          * may overlap the source area (always moving down).
 228          *
 229          * If everything is PMD-aligned, that works fine, as moving
 230          * each pmd down will clear the source pmd. But if we first
 231          * have a few 4kB-only pages that get moved down, and then
 232          * hit the "now the rest is PMD-aligned, let's do everything
 233          * one pmd at a time", we will still have the old (now empty
 234          * of any 4kB pages, but still there) PMD in the page table
 235          * tree.
 236          *
 237          * Warn on it once - because we really should try to figure
 238          * out how to do this better - but then say "I won't move
 239          * this pmd".
 240          *
 241          * One alternative might be to just unmap the target pmd at
 242          * this point, and verify that it really is empty. We'll see.
 243          */
 244         if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
 245                 return false;
 246
 247         /*
 248          * We don't have to worry about the ordering of src and dst
 249          * ptlocks because exclusive mmap_lock prevents deadlock.
 250          */
 251         old_ptl = pmd_lock(vma->vm_mm, old_pmd);
 252         new_ptl = pmd_lockptr(mm, new_pmd);
 253         if (new_ptl != old_ptl)
 254                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 255
 256         /* Clear the pmd */
 257         pmd = *old_pmd;
 258         pmd_clear(old_pmd);
 259
 260         VM_BUG_ON(!pmd_none(*new_pmd));
 261
 262         pmd_populate(mm, new_pmd, pmd_pgtable(pmd));
 263         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
 264         if (new_ptl != old_ptl)
 265                 spin_unlock(new_ptl);
 266         spin_unlock(old_ptl);
 267
 268         return true;
 269 }
 270 #else
 271 static inline bool move_normal_pmd(struct vm_area_struct *vma,
 272                 unsigned long old_addr, unsigned long new_addr, pmd_t *old_pmd,
 273                 pmd_t *new_pmd)
 274 {
 275         return false;
 276 }
 277 #endif
 278
 279 #if CONFIG_PGTABLE_LEVELS > 2 && defined(CONFIG_HAVE_MOVE_PUD)
 280 static bool move_normal_pud(struct vm_area_struct *vma, unsigned long old_addr,
 281                   unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 282 {
 283         spinlock_t *old_ptl, *new_ptl;
 284         struct mm_struct *mm = vma->vm_mm;
 285         pud_t pud;
 286
 287         /*
 288          * The destination pud shouldn't be established, free_pgtables()
 289          * should have released it.
 290          */
 291         if (WARN_ON_ONCE(!pud_none(*new_pud)))
 292                 return false;
 293
 294         /*
 295          * We don't have to worry about the ordering of src and dst
 296          * ptlocks because exclusive mmap_lock prevents deadlock.
 297          */
 298         old_ptl = pud_lock(vma->vm_mm, old_pud);
 299         new_ptl = pud_lockptr(mm, new_pud);
 300         if (new_ptl != old_ptl)
 301                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 302
 303         /* Clear the pud */
 304         pud = *old_pud;
 305         pud_clear(old_pud);
 306
 307         VM_BUG_ON(!pud_none(*new_pud));
 308
 309         pud_populate(mm, new_pud, pud_pgtable(pud));
 310         flush_tlb_range(vma, old_addr, old_addr + PUD_SIZE);
 311         if (new_ptl != old_ptl)
 312                 spin_unlock(new_ptl);
 313         spin_unlock(old_ptl);
 314
 315         return true;
 316 }
 317 #else
 318 static inline bool move_normal_pud(struct vm_area_struct *vma,
 319                 unsigned long old_addr, unsigned long new_addr, pud_t *old_pud,
 320                 pud_t *new_pud)
 321 {
 322         return false;
 323 }
 324 #endif
 325
 326 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 327 static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
 328                           unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 329 {
 330         spinlock_t *old_ptl, *new_ptl;
 331         struct mm_struct *mm = vma->vm_mm;
 332         pud_t pud;
 333
 334         /*
 335          * The destination pud shouldn't be established, free_pgtables()
 336          * should have released it.
 337          */
 338         if (WARN_ON_ONCE(!pud_none(*new_pud)))
 339                 return false;
 340
 341         /*
 342          * We don't have to worry about the ordering of src and dst
 343          * ptlocks because exclusive mmap_lock prevents deadlock.
 344          */
 345         old_ptl = pud_lock(vma->vm_mm, old_pud);
 346         new_ptl = pud_lockptr(mm, new_pud);
 347         if (new_ptl != old_ptl)
 348                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 349
 350         /* Clear the pud */
 351         pud = *old_pud;
 352         pud_clear(old_pud);
 353
 354         VM_BUG_ON(!pud_none(*new_pud));
 355
 356         /* Set the new pud */
 357         /* mark soft_ditry when we add pud level soft dirty support */
 358         set_pud_at(mm, new_addr, new_pud, pud);
 359         flush_pud_tlb_range(vma, old_addr, old_addr + HPAGE_PUD_SIZE);
 360         if (new_ptl != old_ptl)
 361                 spin_unlock(new_ptl);
 362         spin_unlock(old_ptl);
 363
 364         return true;
 365 }
 366 #else
 367 static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
 368                           unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 369 {
 370         WARN_ON_ONCE(1);
 371         return false;
 372
 373 }
 374 #endif
 375
 376 enum pgt_entry {
 377         NORMAL_PMD,
 378         HPAGE_PMD,
 379         NORMAL_PUD,
 380         HPAGE_PUD,
 381 };
 382
 383 /*
 384  * Returns an extent of the corresponding size for the pgt_entry specified if
 385  * valid. Else returns a smaller extent bounded by the end of the source and
 386  * destination pgt_entry.
 387  */
 388 static __always_inline unsigned long get_extent(enum pgt_entry entry,
 389                         unsigned long old_addr, unsigned long old_end,
 390                         unsigned long new_addr)
 391 {
 392         unsigned long next, extent, mask, size;
 393
 394         switch (entry) {
 395         case HPAGE_PMD:
 396         case NORMAL_PMD:
 397                 mask = PMD_MASK;
 398                 size = PMD_SIZE;
 399                 break;
 400         case HPAGE_PUD:
 401         case NORMAL_PUD:
 402                 mask = PUD_MASK;
 403                 size = PUD_SIZE;
 404                 break;
 405         default:
 406                 BUILD_BUG();
 407                 break;
 408         }
 409
 410         next = (old_addr + size) & mask;
 411         /* even if next overflowed, extent below will be ok */
 412         extent = next - old_addr;
 413         if (extent > old_end - old_addr)
 414                 extent = old_end - old_addr;
 415         next = (new_addr + size) & mask;
 416         if (extent > next - new_addr)
 417                 extent = next - new_addr;
 418         return extent;
 419 }
 420
 421 /*
 422  * Attempts to speedup the move by moving entry at the level corresponding to
 423  * pgt_entry. Returns true if the move was successful, else false.
 424  */
 425 static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
 426                         unsigned long old_addr, unsigned long new_addr,
 427                         void *old_entry, void *new_entry, bool need_rmap_locks)
 428 {
 429         bool moved = false;
 430
 431         /* See comment in move_ptes() */
 432         if (need_rmap_locks)
 433                 take_rmap_locks(vma);
 434
 435         switch (entry) {
 436         case NORMAL_PMD:
 437                 moved = move_normal_pmd(vma, old_addr, new_addr, old_entry,
 438                                         new_entry);
 439                 break;
 440         case NORMAL_PUD:
 441                 moved = move_normal_pud(vma, old_addr, new_addr, old_entry,
 442                                         new_entry);
 443                 break;
 444         case HPAGE_PMD:
 445                 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 446                         move_huge_pmd(vma, old_addr, new_addr, old_entry,
 447                                       new_entry);
 448                 break;
 449         case HPAGE_PUD:
 450                 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 451                         move_huge_pud(vma, old_addr, new_addr, old_entry,
 452                                       new_entry);
 453                 break;
 454
 455         default:
 456                 WARN_ON_ONCE(1);
 457                 break;
 458         }
 459
 460         if (need_rmap_locks)
 461                 drop_rmap_locks(vma);
 462
 463         return moved;
 464 }
 465
 466 unsigned long move_page_tables(struct vm_area_struct *vma,
 467                 unsigned long old_addr, struct vm_area_struct *new_vma,
 468                 unsigned long new_addr, unsigned long len,
 469                 bool need_rmap_locks)
 470 {
 471         unsigned long extent, old_end;
 472         struct mmu_notifier_range range;
 473         pmd_t *old_pmd, *new_pmd;
 474         pud_t *old_pud, *new_pud;
 475
 476         old_end = old_addr + len;
 477         flush_cache_range(vma, old_addr, old_end);
 478
 479         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
 480                                 old_addr, old_end);
 481         mmu_notifier_invalidate_range_start(&range);
 482
 483         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 484                 cond_resched();
 485                 /*
 486                  * If extent is PUD-sized try to speed up the move by moving at the
 487                  * PUD level if possible.
 488                  */
 489                 extent = get_extent(NORMAL_PUD, old_addr, old_end, new_addr);
 490
 491                 old_pud = get_old_pud(vma->vm_mm, old_addr);
 492                 if (!old_pud)
 493                         continue;
 494                 new_pud = alloc_new_pud(vma->vm_mm, vma, new_addr);
 495                 if (!new_pud)
 496                         break;
 497                 if (pud_trans_huge(*old_pud) || pud_devmap(*old_pud)) {
 498                         if (extent == HPAGE_PUD_SIZE) {
 499                                 move_pgt_entry(HPAGE_PUD, vma, old_addr, new_addr,
 500                                                old_pud, new_pud, need_rmap_locks);
 501                                 /* We ignore and continue on error? */
 502                                 continue;
 503                         }
 504                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
 505
 506                         if (move_pgt_entry(NORMAL_PUD, vma, old_addr, new_addr,
 507                                            old_pud, new_pud, true))
 508                                 continue;
 509                 }
 510
 511                 extent = get_extent(NORMAL_PMD, old_addr, old_end, new_addr);
 512                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 513                 if (!old_pmd)
 514                         continue;
 515                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 516                 if (!new_pmd)
 517                         break;
 518                 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) ||
 519                     pmd_devmap(*old_pmd)) {
 520                         if (extent == HPAGE_PMD_SIZE &&
 521                             move_pgt_entry(HPAGE_PMD, vma, old_addr, new_addr,
 522                                            old_pmd, new_pmd, need_rmap_locks))
 523                                 continue;
 524                         split_huge_pmd(vma, old_pmd, old_addr);
 525                         if (pmd_trans_unstable(old_pmd))
 526                                 continue;
 527                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
 528                            extent == PMD_SIZE) {
 529                         /*
 530                          * If the extent is PMD-sized, try to speed the move by
 531                          * moving at the PMD level if possible.
 532                          */
 533                         if (move_pgt_entry(NORMAL_PMD, vma, old_addr, new_addr,
 534                                            old_pmd, new_pmd, true))
 535                                 continue;
 536                 }
 537
 538                 if (pte_alloc(new_vma->vm_mm, new_pmd))
 539                         break;
 540                 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
 541                           new_pmd, new_addr, need_rmap_locks);
 542         }
 543
 544         mmu_notifier_invalidate_range_end(&range);
 545
 546         return len + old_addr - old_end;        /* how much done */
 547 }
 548
 549 static unsigned long move_vma(struct vm_area_struct *vma,
 550                 unsigned long old_addr, unsigned long old_len,
 551                 unsigned long new_len, unsigned long new_addr,
 552                 bool *locked, unsigned long flags,
 553                 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
 554 {
 555         struct mm_struct *mm = vma->vm_mm;
 556         struct vm_area_struct *new_vma;
 557         unsigned long vm_flags = vma->vm_flags;
 558         unsigned long new_pgoff;
 559         unsigned long moved_len;
 560         unsigned long excess = 0;
 561         unsigned long hiwater_vm;
 562         int split = 0;
 563         int err = 0;
 564         bool need_rmap_locks;
 565
 566         /*
 567          * We'd prefer to avoid failure later on in do_munmap:
 568          * which may split one vma into three before unmapping.
 569          */
 570         if (mm->map_count >= sysctl_max_map_count - 3)
 571                 return -ENOMEM;
 572
 573         if (vma->vm_ops && vma->vm_ops->may_split) {
 574                 if (vma->vm_start != old_addr)
 575                         err = vma->vm_ops->may_split(vma, old_addr);
 576                 if (!err && vma->vm_end != old_addr + old_len)
 577                         err = vma->vm_ops->may_split(vma, old_addr + old_len);
 578                 if (err)
 579                         return err;
 580         }
 581
 582         /*
 583          * Advise KSM to break any KSM pages in the area to be moved:
 584          * it would be confusing if they were to turn up at the new
 585          * location, where they happen to coincide with different KSM
 586          * pages recently unmapped.  But leave vma->vm_flags as it was,
 587          * so KSM can come around to merge on vma and new_vma afterwards.
 588          */
 589         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 590                                                 MADV_UNMERGEABLE, &vm_flags);
 591         if (err)
 592                 return err;
 593
 594         if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT)) {
 595                 if (security_vm_enough_memory_mm(mm, new_len >> PAGE_SHIFT))
 596                         return -ENOMEM;
 597         }
 598
 599         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 600         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 601                            &need_rmap_locks);
 602         if (!new_vma) {
 603                 if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT))
 604                         vm_unacct_memory(new_len >> PAGE_SHIFT);
 605                 return -ENOMEM;
 606         }
 607
 608         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 609                                      need_rmap_locks);
 610         if (moved_len < old_len) {
 611                 err = -ENOMEM;
 612         } else if (vma->vm_ops && vma->vm_ops->mremap) {
 613                 err = vma->vm_ops->mremap(new_vma);
 614         }
 615
 616         if (unlikely(err)) {
 617                 /*
 618                  * On error, move entries back from new area to old,
 619                  * which will succeed since page tables still there,
 620                  * and then proceed to unmap new area instead of old.
 621                  */
 622                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 623                                  true);
 624                 vma = new_vma;
 625                 old_len = new_len;
 626                 old_addr = new_addr;
 627                 new_addr = err;
 628         } else {
 629                 mremap_userfaultfd_prep(new_vma, uf);
 630         }
 631
 632         /* Conceal VM_ACCOUNT so old reservation is not undone */
 633         if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP)) {
 634                 vma->vm_flags &= ~VM_ACCOUNT;
 635                 excess = vma->vm_end - vma->vm_start - old_len;
 636                 if (old_addr > vma->vm_start &&
 637                     old_addr + old_len < vma->vm_end)
 638                         split = 1;
 639         }
 640
 641         /*
 642          * If we failed to move page tables we still do total_vm increment
 643          * since do_munmap() will decrement it by old_len == new_len.
 644          *
 645          * Since total_vm is about to be raised artificially high for a
 646          * moment, we need to restore high watermark afterwards: if stats
 647          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 648          * If this were a serious issue, we'd add a flag to do_munmap().
 649          */
 650         hiwater_vm = mm->hiwater_vm;
 651         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 652
 653         /* Tell pfnmap has moved from this vma */
 654         if (unlikely(vma->vm_flags & VM_PFNMAP))
 655                 untrack_pfn_moved(vma);
 656
 657         if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
 658                 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
 659                 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
 660
 661                 /*
 662                  * anon_vma links of the old vma is no longer needed after its page
 663                  * table has been moved.
 664                  */
 665                 if (new_vma != vma && vma->vm_start == old_addr &&
 666                         vma->vm_end == (old_addr + old_len))
 667                         unlink_anon_vmas(vma);
 668
 669                 /* Because we won't unmap we don't need to touch locked_vm */
 670                 return new_addr;
 671         }
 672
 673         if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
 674                 /* OOM: unable to split vma, just get accounts right */
 675                 if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP))
 676                         vm_acct_memory(new_len >> PAGE_SHIFT);
 677                 excess = 0;
 678         }
 679
 680         if (vm_flags & VM_LOCKED) {
 681                 mm->locked_vm += new_len >> PAGE_SHIFT;
 682                 *locked = true;
 683         }
 684
 685         mm->hiwater_vm = hiwater_vm;
 686
 687         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 688         if (excess) {
 689                 vma->vm_flags |= VM_ACCOUNT;
 690                 if (split)
 691                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 692         }
 693
 694         return new_addr;
 695 }
 696
 697 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 698         unsigned long old_len, unsigned long new_len, unsigned long flags,
 699         unsigned long *p)
 700 {
 701         struct mm_struct *mm = current->mm;
 702         struct vm_area_struct *vma;
 703         unsigned long pgoff;
 704
 705         vma = vma_lookup(mm, addr);
 706         if (!vma)
 707                 return ERR_PTR(-EFAULT);
 708
 709         /*
 710          * !old_len is a special case where an attempt is made to 'duplicate'
 711          * a mapping.  This makes no sense for private mappings as it will
 712          * instead create a fresh/new mapping unrelated to the original.  This
 713          * is contrary to the basic idea of mremap which creates new mappings
 714          * based on the original.  There are no known use cases for this
 715          * behavior.  As a result, fail such attempts.
 716          */
 717         if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
 718                 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap.  This is not supported.\n", current->comm, current->pid);
 719                 return ERR_PTR(-EINVAL);
 720         }
 721
 722         if ((flags & MREMAP_DONTUNMAP) &&
 723                         (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)))
 724                 return ERR_PTR(-EINVAL);
 725
 726         if (is_vm_hugetlb_page(vma))
 727                 return ERR_PTR(-EINVAL);
 728
 729         /* We can't remap across vm area boundaries */
 730         if (old_len > vma->vm_end - addr)
 731                 return ERR_PTR(-EFAULT);
 732
 733         if (new_len == old_len)
 734                 return vma;
 735
 736         /* Need to be careful about a growing mapping */
 737         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 738         pgoff += vma->vm_pgoff;
 739         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 740                 return ERR_PTR(-EINVAL);
 741
 742         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 743                 return ERR_PTR(-EFAULT);
 744
 745         if (vma->vm_flags & VM_LOCKED) {
 746                 unsigned long locked, lock_limit;
 747                 locked = mm->locked_vm << PAGE_SHIFT;
 748                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 749                 locked += new_len - old_len;
 750                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 751                         return ERR_PTR(-EAGAIN);
 752         }
 753
 754         if (!may_expand_vm(mm, vma->vm_flags,
 755                                 (new_len - old_len) >> PAGE_SHIFT))
 756                 return ERR_PTR(-ENOMEM);
 757
 758         if (vma->vm_flags & VM_ACCOUNT) {
 759                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 760                 if (security_vm_enough_memory_mm(mm, charged))
 761                         return ERR_PTR(-ENOMEM);
 762                 *p = charged;
 763         }
 764
 765         return vma;
 766 }
 767
 768 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 769                 unsigned long new_addr, unsigned long new_len, bool *locked,
 770                 unsigned long flags, struct vm_userfaultfd_ctx *uf,
 771                 struct list_head *uf_unmap_early,
 772                 struct list_head *uf_unmap)
 773 {
 774         struct mm_struct *mm = current->mm;
 775         struct vm_area_struct *vma;
 776         unsigned long ret = -EINVAL;
 777         unsigned long charged = 0;
 778         unsigned long map_flags = 0;
 779
 780         if (offset_in_page(new_addr))
 781                 goto out;
 782
 783         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 784                 goto out;
 785
 786         /* Ensure the old/new locations do not overlap */
 787         if (addr + old_len > new_addr && new_addr + new_len > addr)
 788                 goto out;
 789
 790         /*
 791          * move_vma() need us to stay 4 maps below the threshold, otherwise
 792          * it will bail out at the very beginning.
 793          * That is a problem if we have already unmaped the regions here
 794          * (new_addr, and old_addr), because userspace will not know the
 795          * state of the vma's after it gets -ENOMEM.
 796          * So, to avoid such scenario we can pre-compute if the whole
 797          * operation has high chances to success map-wise.
 798          * Worst-scenario case is when both vma's (new_addr and old_addr) get
 799          * split in 3 before unmapping it.
 800          * That means 2 more maps (1 for each) to the ones we already hold.
 801          * Check whether current map count plus 2 still leads us to 4 maps below
 802          * the threshold, otherwise return -ENOMEM here to be more safe.
 803          */
 804         if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
 805                 return -ENOMEM;
 806
 807         if (flags & MREMAP_FIXED) {
 808                 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
 809                 if (ret)
 810                         goto out;
 811         }
 812
 813         if (old_len >= new_len) {
 814                 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
 815                 if (ret && old_len != new_len)
 816                         goto out;
 817                 old_len = new_len;
 818         }
 819
 820         vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
 821         if (IS_ERR(vma)) {
 822                 ret = PTR_ERR(vma);
 823                 goto out;
 824         }
 825
 826         /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
 827         if (flags & MREMAP_DONTUNMAP &&
 828                 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
 829                 ret = -ENOMEM;
 830                 goto out;
 831         }
 832
 833         if (flags & MREMAP_FIXED)
 834                 map_flags |= MAP_FIXED;
 835
 836         if (vma->vm_flags & VM_MAYSHARE)
 837                 map_flags |= MAP_SHARED;
 838
 839         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 840                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 841                                 map_flags);
 842         if (IS_ERR_VALUE(ret))
 843                 goto out1;
 844
 845         /* We got a new mapping */
 846         if (!(flags & MREMAP_FIXED))
 847                 new_addr = ret;
 848
 849         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
 850                        uf_unmap);
 851
 852         if (!(offset_in_page(ret)))
 853                 goto out;
 854
 855 out1:
 856         vm_unacct_memory(charged);
 857
 858 out:
 859         return ret;
 860 }
 861
 862 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 863 {
 864         unsigned long end = vma->vm_end + delta;
 865         if (end < vma->vm_end) /* overflow */
 866                 return 0;
 867         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 868                 return 0;
 869         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 870                               0, MAP_FIXED) & ~PAGE_MASK)
 871                 return 0;
 872         return 1;
 873 }
 874
 875 /*
 876  * Expand (or shrink) an existing mapping, potentially moving it at the
 877  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 878  *
 879  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 880  * This option implies MREMAP_MAYMOVE.
 881  */
 882 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 883                 unsigned long, new_len, unsigned long, flags,
 884                 unsigned long, new_addr)
 885 {
 886         struct mm_struct *mm = current->mm;
 887         struct vm_area_struct *vma;
 888         unsigned long ret = -EINVAL;
 889         unsigned long charged = 0;
 890         bool locked = false;
 891         bool downgraded = false;
 892         struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
 893         LIST_HEAD(uf_unmap_early);
 894         LIST_HEAD(uf_unmap);
 895
 896         /*
 897          * There is a deliberate asymmetry here: we strip the pointer tag
 898          * from the old address but leave the new address alone. This is
 899          * for consistency with mmap(), where we prevent the creation of
 900          * aliasing mappings in userspace by leaving the tag bits of the
 901          * mapping address intact. A non-zero tag will cause the subsequent
 902          * range checks to reject the address as invalid.
 903          *
 904          * See Documentation/arm64/tagged-address-abi.rst for more information.
 905          */
 906         addr = untagged_addr(addr);
 907
 908         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
 909                 return ret;
 910
 911         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 912                 return ret;
 913
 914         /*
 915          * MREMAP_DONTUNMAP is always a move and it does not allow resizing
 916          * in the process.
 917          */
 918         if (flags & MREMAP_DONTUNMAP &&
 919                         (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
 920                 return ret;
 921
 922
 923         if (offset_in_page(addr))
 924                 return ret;
 925
 926         old_len = PAGE_ALIGN(old_len);
 927         new_len = PAGE_ALIGN(new_len);
 928
 929         /*
 930          * We allow a zero old-len as a special case
 931          * for DOS-emu "duplicate shm area" thing. But
 932          * a zero new-len is nonsensical.
 933          */
 934         if (!new_len)
 935                 return ret;
 936
 937         if (mmap_write_lock_killable(current->mm))
 938                 return -EINTR;
 939
 940         if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
 941                 ret = mremap_to(addr, old_len, new_addr, new_len,
 942                                 &locked, flags, &uf, &uf_unmap_early,
 943                                 &uf_unmap);
 944                 goto out;
 945         }
 946
 947         /*
 948          * Always allow a shrinking remap: that just unmaps
 949          * the unnecessary pages..
 950          * __do_munmap does all the needed commit accounting, and
 951          * downgrades mmap_lock to read if so directed.
 952          */
 953         if (old_len >= new_len) {
 954                 int retval;
 955
 956                 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
 957                                   &uf_unmap, true);
 958                 if (retval < 0 && old_len != new_len) {
 959                         ret = retval;
 960                         goto out;
 961                 /* Returning 1 indicates mmap_lock is downgraded to read. */
 962                 } else if (retval == 1)
 963                         downgraded = true;
 964                 ret = addr;
 965                 goto out;
 966         }
 967
 968         /*
 969          * Ok, we need to grow..
 970          */
 971         vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
 972         if (IS_ERR(vma)) {
 973                 ret = PTR_ERR(vma);
 974                 goto out;
 975         }
 976
 977         /* old_len exactly to the end of the area..
 978          */
 979         if (old_len == vma->vm_end - addr) {
 980                 /* can we just expand the current mapping? */
 981                 if (vma_expandable(vma, new_len - old_len)) {
 982                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 983
 984                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 985                                        vma->vm_pgoff, NULL)) {
 986                                 ret = -ENOMEM;
 987                                 goto out;
 988                         }
 989
 990                         vm_stat_account(mm, vma->vm_flags, pages);
 991                         if (vma->vm_flags & VM_LOCKED) {
 992                                 mm->locked_vm += pages;
 993                                 locked = true;
 994                                 new_addr = addr;
 995                         }
 996                         ret = addr;
 997                         goto out;
 998                 }
 999         }
1000
1001         /*
1002          * We weren't able to just expand or shrink the area,
1003          * we need to create a new one and move it..
1004          */
1005         ret = -ENOMEM;
1006         if (flags & MREMAP_MAYMOVE) {
1007                 unsigned long map_flags = 0;
1008                 if (vma->vm_flags & VM_MAYSHARE)
1009                         map_flags |= MAP_SHARED;
1010
1011                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
1012                                         vma->vm_pgoff +
1013                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
1014                                         map_flags);
1015                 if (IS_ERR_VALUE(new_addr)) {
1016                         ret = new_addr;
1017                         goto out;
1018                 }
1019
1020                 ret = move_vma(vma, addr, old_len, new_len, new_addr,
1021                                &locked, flags, &uf, &uf_unmap);
1022         }
1023 out:
1024         if (offset_in_page(ret)) {
1025                 vm_unacct_memory(charged);
1026                 locked = false;
1027         }
1028         if (downgraded)
1029                 mmap_read_unlock(current->mm);
1030         else
1031                 mmap_write_unlock(current->mm);
1032         if (locked && new_len > old_len)
1033                 mm_populate(new_addr + old_len, new_len - old_len);
1034         userfaultfd_unmap_complete(mm, &uf_unmap_early);
1035         mremap_userfaultfd_complete(&uf, addr, ret, old_len);
1036         userfaultfd_unmap_complete(mm, &uf_unmap);
1037         return ret;
1038 }