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/mm-arch-hooks.h>
  26 #include <linux/userfaultfd_k.h>
  27
  28 #include <asm/cacheflush.h>
  29 #include <asm/tlbflush.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         /* Set the new pmd */
 263         set_pmd_at(mm, new_addr, new_pmd, pmd);
 264         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
 265         if (new_ptl != old_ptl)
 266                 spin_unlock(new_ptl);
 267         spin_unlock(old_ptl);
 268
 269         return true;
 270 }
 271 #else
 272 static inline bool move_normal_pmd(struct vm_area_struct *vma,
 273                 unsigned long old_addr, unsigned long new_addr, pmd_t *old_pmd,
 274                 pmd_t *new_pmd)
 275 {
 276         return false;
 277 }
 278 #endif
 279
 280 #ifdef CONFIG_HAVE_MOVE_PUD
 281 static bool move_normal_pud(struct vm_area_struct *vma, unsigned long old_addr,
 282                   unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 283 {
 284         spinlock_t *old_ptl, *new_ptl;
 285         struct mm_struct *mm = vma->vm_mm;
 286         pud_t pud;
 287
 288         /*
 289          * The destination pud shouldn't be established, free_pgtables()
 290          * should have released it.
 291          */
 292         if (WARN_ON_ONCE(!pud_none(*new_pud)))
 293                 return false;
 294
 295         /*
 296          * We don't have to worry about the ordering of src and dst
 297          * ptlocks because exclusive mmap_lock prevents deadlock.
 298          */
 299         old_ptl = pud_lock(vma->vm_mm, old_pud);
 300         new_ptl = pud_lockptr(mm, new_pud);
 301         if (new_ptl != old_ptl)
 302                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 303
 304         /* Clear the pud */
 305         pud = *old_pud;
 306         pud_clear(old_pud);
 307
 308         VM_BUG_ON(!pud_none(*new_pud));
 309
 310         /* Set the new pud */
 311         set_pud_at(mm, new_addr, new_pud, pud);
 312         flush_tlb_range(vma, old_addr, old_addr + PUD_SIZE);
 313         if (new_ptl != old_ptl)
 314                 spin_unlock(new_ptl);
 315         spin_unlock(old_ptl);
 316
 317         return true;
 318 }
 319 #else
 320 static inline bool move_normal_pud(struct vm_area_struct *vma,
 321                 unsigned long old_addr, unsigned long new_addr, pud_t *old_pud,
 322                 pud_t *new_pud)
 323 {
 324         return false;
 325 }
 326 #endif
 327
 328 enum pgt_entry {
 329         NORMAL_PMD,
 330         HPAGE_PMD,
 331         NORMAL_PUD,
 332 };
 333
 334 /*
 335  * Returns an extent of the corresponding size for the pgt_entry specified if
 336  * valid. Else returns a smaller extent bounded by the end of the source and
 337  * destination pgt_entry.
 338  */
 339 static unsigned long get_extent(enum pgt_entry entry, unsigned long old_addr,
 340                         unsigned long old_end, unsigned long new_addr)
 341 {
 342         unsigned long next, extent, mask, size;
 343
 344         switch (entry) {
 345         case HPAGE_PMD:
 346         case NORMAL_PMD:
 347                 mask = PMD_MASK;
 348                 size = PMD_SIZE;
 349                 break;
 350         case NORMAL_PUD:
 351                 mask = PUD_MASK;
 352                 size = PUD_SIZE;
 353                 break;
 354         default:
 355                 BUILD_BUG();
 356                 break;
 357         }
 358
 359         next = (old_addr + size) & mask;
 360         /* even if next overflowed, extent below will be ok */
 361         extent = (next > old_end) ? old_end - old_addr : next - old_addr;
 362         next = (new_addr + size) & mask;
 363         if (extent > next - new_addr)
 364                 extent = next - new_addr;
 365         return extent;
 366 }
 367
 368 /*
 369  * Attempts to speedup the move by moving entry at the level corresponding to
 370  * pgt_entry. Returns true if the move was successful, else false.
 371  */
 372 static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
 373                         unsigned long old_addr, unsigned long new_addr,
 374                         void *old_entry, void *new_entry, bool need_rmap_locks)
 375 {
 376         bool moved = false;
 377
 378         /* See comment in move_ptes() */
 379         if (need_rmap_locks)
 380                 take_rmap_locks(vma);
 381
 382         switch (entry) {
 383         case NORMAL_PMD:
 384                 moved = move_normal_pmd(vma, old_addr, new_addr, old_entry,
 385                                         new_entry);
 386                 break;
 387         case NORMAL_PUD:
 388                 moved = move_normal_pud(vma, old_addr, new_addr, old_entry,
 389                                         new_entry);
 390                 break;
 391         case HPAGE_PMD:
 392                 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 393                         move_huge_pmd(vma, old_addr, new_addr, old_entry,
 394                                       new_entry);
 395                 break;
 396         default:
 397                 WARN_ON_ONCE(1);
 398                 break;
 399         }
 400
 401         if (need_rmap_locks)
 402                 drop_rmap_locks(vma);
 403
 404         return moved;
 405 }
 406
 407 unsigned long move_page_tables(struct vm_area_struct *vma,
 408                 unsigned long old_addr, struct vm_area_struct *new_vma,
 409                 unsigned long new_addr, unsigned long len,
 410                 bool need_rmap_locks)
 411 {
 412         unsigned long extent, old_end;
 413         struct mmu_notifier_range range;
 414         pmd_t *old_pmd, *new_pmd;
 415
 416         old_end = old_addr + len;
 417         flush_cache_range(vma, old_addr, old_end);
 418
 419         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
 420                                 old_addr, old_end);
 421         mmu_notifier_invalidate_range_start(&range);
 422
 423         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 424                 cond_resched();
 425                 /*
 426                  * If extent is PUD-sized try to speed up the move by moving at the
 427                  * PUD level if possible.
 428                  */
 429                 extent = get_extent(NORMAL_PUD, old_addr, old_end, new_addr);
 430                 if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
 431                         pud_t *old_pud, *new_pud;
 432
 433                         old_pud = get_old_pud(vma->vm_mm, old_addr);
 434                         if (!old_pud)
 435                                 continue;
 436                         new_pud = alloc_new_pud(vma->vm_mm, vma, new_addr);
 437                         if (!new_pud)
 438                                 break;
 439                         if (move_pgt_entry(NORMAL_PUD, vma, old_addr, new_addr,
 440                                            old_pud, new_pud, need_rmap_locks))
 441                                 continue;
 442                 }
 443
 444                 extent = get_extent(NORMAL_PMD, old_addr, old_end, new_addr);
 445                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 446                 if (!old_pmd)
 447                         continue;
 448                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 449                 if (!new_pmd)
 450                         break;
 451                 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) ||
 452                     pmd_devmap(*old_pmd)) {
 453                         if (extent == HPAGE_PMD_SIZE &&
 454                             move_pgt_entry(HPAGE_PMD, vma, old_addr, new_addr,
 455                                            old_pmd, new_pmd, need_rmap_locks))
 456                                 continue;
 457                         split_huge_pmd(vma, old_pmd, old_addr);
 458                         if (pmd_trans_unstable(old_pmd))
 459                                 continue;
 460                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
 461                            extent == PMD_SIZE) {
 462                         /*
 463                          * If the extent is PMD-sized, try to speed the move by
 464                          * moving at the PMD level if possible.
 465                          */
 466                         if (move_pgt_entry(NORMAL_PMD, vma, old_addr, new_addr,
 467                                            old_pmd, new_pmd, need_rmap_locks))
 468                                 continue;
 469                 }
 470
 471                 if (pte_alloc(new_vma->vm_mm, new_pmd))
 472                         break;
 473                 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
 474                           new_pmd, new_addr, need_rmap_locks);
 475         }
 476
 477         mmu_notifier_invalidate_range_end(&range);
 478
 479         return len + old_addr - old_end;        /* how much done */
 480 }
 481
 482 static unsigned long move_vma(struct vm_area_struct *vma,
 483                 unsigned long old_addr, unsigned long old_len,
 484                 unsigned long new_len, unsigned long new_addr,
 485                 bool *locked, unsigned long flags,
 486                 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
 487 {
 488         struct mm_struct *mm = vma->vm_mm;
 489         struct vm_area_struct *new_vma;
 490         unsigned long vm_flags = vma->vm_flags;
 491         unsigned long new_pgoff;
 492         unsigned long moved_len;
 493         unsigned long excess = 0;
 494         unsigned long hiwater_vm;
 495         int split = 0;
 496         int err = 0;
 497         bool need_rmap_locks;
 498
 499         /*
 500          * We'd prefer to avoid failure later on in do_munmap:
 501          * which may split one vma into three before unmapping.
 502          */
 503         if (mm->map_count >= sysctl_max_map_count - 3)
 504                 return -ENOMEM;
 505
 506         if (vma->vm_ops && vma->vm_ops->may_split) {
 507                 if (vma->vm_start != old_addr)
 508                         err = vma->vm_ops->may_split(vma, old_addr);
 509                 if (!err && vma->vm_end != old_addr + old_len)
 510                         err = vma->vm_ops->may_split(vma, old_addr + old_len);
 511                 if (err)
 512                         return err;
 513         }
 514
 515         /*
 516          * Advise KSM to break any KSM pages in the area to be moved:
 517          * it would be confusing if they were to turn up at the new
 518          * location, where they happen to coincide with different KSM
 519          * pages recently unmapped.  But leave vma->vm_flags as it was,
 520          * so KSM can come around to merge on vma and new_vma afterwards.
 521          */
 522         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 523                                                 MADV_UNMERGEABLE, &vm_flags);
 524         if (err)
 525                 return err;
 526
 527         if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT)) {
 528                 if (security_vm_enough_memory_mm(mm, new_len >> PAGE_SHIFT))
 529                         return -ENOMEM;
 530         }
 531
 532         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 533         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 534                            &need_rmap_locks);
 535         if (!new_vma) {
 536                 if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT))
 537                         vm_unacct_memory(new_len >> PAGE_SHIFT);
 538                 return -ENOMEM;
 539         }
 540
 541         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 542                                      need_rmap_locks);
 543         if (moved_len < old_len) {
 544                 err = -ENOMEM;
 545         } else if (vma->vm_ops && vma->vm_ops->mremap) {
 546                 err = vma->vm_ops->mremap(new_vma, flags);
 547         }
 548
 549         if (unlikely(err)) {
 550                 /*
 551                  * On error, move entries back from new area to old,
 552                  * which will succeed since page tables still there,
 553                  * and then proceed to unmap new area instead of old.
 554                  */
 555                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 556                                  true);
 557                 vma = new_vma;
 558                 old_len = new_len;
 559                 old_addr = new_addr;
 560                 new_addr = err;
 561         } else {
 562                 mremap_userfaultfd_prep(new_vma, uf);
 563                 arch_remap(mm, old_addr, old_addr + old_len,
 564                            new_addr, new_addr + new_len);
 565         }
 566
 567         /* Conceal VM_ACCOUNT so old reservation is not undone */
 568         if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP)) {
 569                 vma->vm_flags &= ~VM_ACCOUNT;
 570                 excess = vma->vm_end - vma->vm_start - old_len;
 571                 if (old_addr > vma->vm_start &&
 572                     old_addr + old_len < vma->vm_end)
 573                         split = 1;
 574         }
 575
 576         /*
 577          * If we failed to move page tables we still do total_vm increment
 578          * since do_munmap() will decrement it by old_len == new_len.
 579          *
 580          * Since total_vm is about to be raised artificially high for a
 581          * moment, we need to restore high watermark afterwards: if stats
 582          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 583          * If this were a serious issue, we'd add a flag to do_munmap().
 584          */
 585         hiwater_vm = mm->hiwater_vm;
 586         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 587
 588         /* Tell pfnmap has moved from this vma */
 589         if (unlikely(vma->vm_flags & VM_PFNMAP))
 590                 untrack_pfn_moved(vma);
 591
 592         if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
 593                 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
 594                 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
 595
 596                 /* Because we won't unmap we don't need to touch locked_vm */
 597                 return new_addr;
 598         }
 599
 600         if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
 601                 /* OOM: unable to split vma, just get accounts right */
 602                 if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP))
 603                         vm_acct_memory(new_len >> PAGE_SHIFT);
 604                 excess = 0;
 605         }
 606
 607         if (vm_flags & VM_LOCKED) {
 608                 mm->locked_vm += new_len >> PAGE_SHIFT;
 609                 *locked = true;
 610         }
 611
 612         mm->hiwater_vm = hiwater_vm;
 613
 614         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 615         if (excess) {
 616                 vma->vm_flags |= VM_ACCOUNT;
 617                 if (split)
 618                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 619         }
 620
 621         return new_addr;
 622 }
 623
 624 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 625         unsigned long old_len, unsigned long new_len, unsigned long flags,
 626         unsigned long *p)
 627 {
 628         struct mm_struct *mm = current->mm;
 629         struct vm_area_struct *vma = find_vma(mm, addr);
 630         unsigned long pgoff;
 631
 632         if (!vma || vma->vm_start > addr)
 633                 return ERR_PTR(-EFAULT);
 634
 635         /*
 636          * !old_len is a special case where an attempt is made to 'duplicate'
 637          * a mapping.  This makes no sense for private mappings as it will
 638          * instead create a fresh/new mapping unrelated to the original.  This
 639          * is contrary to the basic idea of mremap which creates new mappings
 640          * based on the original.  There are no known use cases for this
 641          * behavior.  As a result, fail such attempts.
 642          */
 643         if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
 644                 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap.  This is not supported.\n", current->comm, current->pid);
 645                 return ERR_PTR(-EINVAL);
 646         }
 647
 648         if (flags & MREMAP_DONTUNMAP && (!vma_is_anonymous(vma) ||
 649                         vma->vm_flags & VM_SHARED))
 650                 return ERR_PTR(-EINVAL);
 651
 652         if (is_vm_hugetlb_page(vma))
 653                 return ERR_PTR(-EINVAL);
 654
 655         /* We can't remap across vm area boundaries */
 656         if (old_len > vma->vm_end - addr)
 657                 return ERR_PTR(-EFAULT);
 658
 659         if (new_len == old_len)
 660                 return vma;
 661
 662         /* Need to be careful about a growing mapping */
 663         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 664         pgoff += vma->vm_pgoff;
 665         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 666                 return ERR_PTR(-EINVAL);
 667
 668         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 669                 return ERR_PTR(-EFAULT);
 670
 671         if (vma->vm_flags & VM_LOCKED) {
 672                 unsigned long locked, lock_limit;
 673                 locked = mm->locked_vm << PAGE_SHIFT;
 674                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 675                 locked += new_len - old_len;
 676                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 677                         return ERR_PTR(-EAGAIN);
 678         }
 679
 680         if (!may_expand_vm(mm, vma->vm_flags,
 681                                 (new_len - old_len) >> PAGE_SHIFT))
 682                 return ERR_PTR(-ENOMEM);
 683
 684         if (vma->vm_flags & VM_ACCOUNT) {
 685                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 686                 if (security_vm_enough_memory_mm(mm, charged))
 687                         return ERR_PTR(-ENOMEM);
 688                 *p = charged;
 689         }
 690
 691         return vma;
 692 }
 693
 694 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 695                 unsigned long new_addr, unsigned long new_len, bool *locked,
 696                 unsigned long flags, struct vm_userfaultfd_ctx *uf,
 697                 struct list_head *uf_unmap_early,
 698                 struct list_head *uf_unmap)
 699 {
 700         struct mm_struct *mm = current->mm;
 701         struct vm_area_struct *vma;
 702         unsigned long ret = -EINVAL;
 703         unsigned long charged = 0;
 704         unsigned long map_flags = 0;
 705
 706         if (offset_in_page(new_addr))
 707                 goto out;
 708
 709         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 710                 goto out;
 711
 712         /* Ensure the old/new locations do not overlap */
 713         if (addr + old_len > new_addr && new_addr + new_len > addr)
 714                 goto out;
 715
 716         /*
 717          * move_vma() need us to stay 4 maps below the threshold, otherwise
 718          * it will bail out at the very beginning.
 719          * That is a problem if we have already unmaped the regions here
 720          * (new_addr, and old_addr), because userspace will not know the
 721          * state of the vma's after it gets -ENOMEM.
 722          * So, to avoid such scenario we can pre-compute if the whole
 723          * operation has high chances to success map-wise.
 724          * Worst-scenario case is when both vma's (new_addr and old_addr) get
 725          * split in 3 before unmaping it.
 726          * That means 2 more maps (1 for each) to the ones we already hold.
 727          * Check whether current map count plus 2 still leads us to 4 maps below
 728          * the threshold, otherwise return -ENOMEM here to be more safe.
 729          */
 730         if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
 731                 return -ENOMEM;
 732
 733         if (flags & MREMAP_FIXED) {
 734                 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
 735                 if (ret)
 736                         goto out;
 737         }
 738
 739         if (old_len >= new_len) {
 740                 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
 741                 if (ret && old_len != new_len)
 742                         goto out;
 743                 old_len = new_len;
 744         }
 745
 746         vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
 747         if (IS_ERR(vma)) {
 748                 ret = PTR_ERR(vma);
 749                 goto out;
 750         }
 751
 752         /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
 753         if (flags & MREMAP_DONTUNMAP &&
 754                 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
 755                 ret = -ENOMEM;
 756                 goto out;
 757         }
 758
 759         if (flags & MREMAP_FIXED)
 760                 map_flags |= MAP_FIXED;
 761
 762         if (vma->vm_flags & VM_MAYSHARE)
 763                 map_flags |= MAP_SHARED;
 764
 765         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 766                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 767                                 map_flags);
 768         if (IS_ERR_VALUE(ret))
 769                 goto out1;
 770
 771         /* We got a new mapping */
 772         if (!(flags & MREMAP_FIXED))
 773                 new_addr = ret;
 774
 775         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
 776                        uf_unmap);
 777
 778         if (!(offset_in_page(ret)))
 779                 goto out;
 780
 781 out1:
 782         vm_unacct_memory(charged);
 783
 784 out:
 785         return ret;
 786 }
 787
 788 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 789 {
 790         unsigned long end = vma->vm_end + delta;
 791         if (end < vma->vm_end) /* overflow */
 792                 return 0;
 793         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 794                 return 0;
 795         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 796                               0, MAP_FIXED) & ~PAGE_MASK)
 797                 return 0;
 798         return 1;
 799 }
 800
 801 /*
 802  * Expand (or shrink) an existing mapping, potentially moving it at the
 803  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 804  *
 805  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 806  * This option implies MREMAP_MAYMOVE.
 807  */
 808 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 809                 unsigned long, new_len, unsigned long, flags,
 810                 unsigned long, new_addr)
 811 {
 812         struct mm_struct *mm = current->mm;
 813         struct vm_area_struct *vma;
 814         unsigned long ret = -EINVAL;
 815         unsigned long charged = 0;
 816         bool locked = false;
 817         bool downgraded = false;
 818         struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
 819         LIST_HEAD(uf_unmap_early);
 820         LIST_HEAD(uf_unmap);
 821
 822         /*
 823          * There is a deliberate asymmetry here: we strip the pointer tag
 824          * from the old address but leave the new address alone. This is
 825          * for consistency with mmap(), where we prevent the creation of
 826          * aliasing mappings in userspace by leaving the tag bits of the
 827          * mapping address intact. A non-zero tag will cause the subsequent
 828          * range checks to reject the address as invalid.
 829          *
 830          * See Documentation/arm64/tagged-address-abi.rst for more information.
 831          */
 832         addr = untagged_addr(addr);
 833
 834         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
 835                 return ret;
 836
 837         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 838                 return ret;
 839
 840         /*
 841          * MREMAP_DONTUNMAP is always a move and it does not allow resizing
 842          * in the process.
 843          */
 844         if (flags & MREMAP_DONTUNMAP &&
 845                         (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
 846                 return ret;
 847
 848
 849         if (offset_in_page(addr))
 850                 return ret;
 851
 852         old_len = PAGE_ALIGN(old_len);
 853         new_len = PAGE_ALIGN(new_len);
 854
 855         /*
 856          * We allow a zero old-len as a special case
 857          * for DOS-emu "duplicate shm area" thing. But
 858          * a zero new-len is nonsensical.
 859          */
 860         if (!new_len)
 861                 return ret;
 862
 863         if (mmap_write_lock_killable(current->mm))
 864                 return -EINTR;
 865
 866         if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
 867                 ret = mremap_to(addr, old_len, new_addr, new_len,
 868                                 &locked, flags, &uf, &uf_unmap_early,
 869                                 &uf_unmap);
 870                 goto out;
 871         }
 872
 873         /*
 874          * Always allow a shrinking remap: that just unmaps
 875          * the unnecessary pages..
 876          * __do_munmap does all the needed commit accounting, and
 877          * downgrades mmap_lock to read if so directed.
 878          */
 879         if (old_len >= new_len) {
 880                 int retval;
 881
 882                 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
 883                                   &uf_unmap, true);
 884                 if (retval < 0 && old_len != new_len) {
 885                         ret = retval;
 886                         goto out;
 887                 /* Returning 1 indicates mmap_lock is downgraded to read. */
 888                 } else if (retval == 1)
 889                         downgraded = true;
 890                 ret = addr;
 891                 goto out;
 892         }
 893
 894         /*
 895          * Ok, we need to grow..
 896          */
 897         vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
 898         if (IS_ERR(vma)) {
 899                 ret = PTR_ERR(vma);
 900                 goto out;
 901         }
 902
 903         /* old_len exactly to the end of the area..
 904          */
 905         if (old_len == vma->vm_end - addr) {
 906                 /* can we just expand the current mapping? */
 907                 if (vma_expandable(vma, new_len - old_len)) {
 908                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 909
 910                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 911                                        vma->vm_pgoff, NULL)) {
 912                                 ret = -ENOMEM;
 913                                 goto out;
 914                         }
 915
 916                         vm_stat_account(mm, vma->vm_flags, pages);
 917                         if (vma->vm_flags & VM_LOCKED) {
 918                                 mm->locked_vm += pages;
 919                                 locked = true;
 920                                 new_addr = addr;
 921                         }
 922                         ret = addr;
 923                         goto out;
 924                 }
 925         }
 926
 927         /*
 928          * We weren't able to just expand or shrink the area,
 929          * we need to create a new one and move it..
 930          */
 931         ret = -ENOMEM;
 932         if (flags & MREMAP_MAYMOVE) {
 933                 unsigned long map_flags = 0;
 934                 if (vma->vm_flags & VM_MAYSHARE)
 935                         map_flags |= MAP_SHARED;
 936
 937                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 938                                         vma->vm_pgoff +
 939                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 940                                         map_flags);
 941                 if (IS_ERR_VALUE(new_addr)) {
 942                         ret = new_addr;
 943                         goto out;
 944                 }
 945
 946                 ret = move_vma(vma, addr, old_len, new_len, new_addr,
 947                                &locked, flags, &uf, &uf_unmap);
 948         }
 949 out:
 950         if (offset_in_page(ret)) {
 951                 vm_unacct_memory(charged);
 952                 locked = false;
 953         }
 954         if (downgraded)
 955                 mmap_read_unlock(current->mm);
 956         else
 957                 mmap_write_unlock(current->mm);
 958         if (locked && new_len > old_len)
 959                 mm_populate(new_addr + old_len, new_len - old_len);
 960         userfaultfd_unmap_complete(mm, &uf_unmap_early);
 961         mremap_userfaultfd_complete(&uf, addr, ret, old_len);
 962         userfaultfd_unmap_complete(mm, &uf_unmap);
 963         return ret;
 964 }