Merge tag 'mips_5.12' of git://git.kernel.org/pub/scm/linux/kernel/git/mips/linux
[linux-2.6-microblaze.git] / 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 __always_inline unsigned long get_extent(enum pgt_entry entry,
340                         unsigned long old_addr, unsigned long old_end,
341                         unsigned long new_addr)
342 {
343         unsigned long next, extent, mask, size;
344
345         switch (entry) {
346         case HPAGE_PMD:
347         case NORMAL_PMD:
348                 mask = PMD_MASK;
349                 size = PMD_SIZE;
350                 break;
351         case NORMAL_PUD:
352                 mask = PUD_MASK;
353                 size = PUD_SIZE;
354                 break;
355         default:
356                 BUILD_BUG();
357                 break;
358         }
359
360         next = (old_addr + size) & mask;
361         /* even if next overflowed, extent below will be ok */
362         extent = next - old_addr;
363         if (extent > old_end - old_addr)
364                 extent = old_end - old_addr;
365         next = (new_addr + size) & mask;
366         if (extent > next - new_addr)
367                 extent = next - new_addr;
368         return extent;
369 }
370
371 /*
372  * Attempts to speedup the move by moving entry at the level corresponding to
373  * pgt_entry. Returns true if the move was successful, else false.
374  */
375 static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
376                         unsigned long old_addr, unsigned long new_addr,
377                         void *old_entry, void *new_entry, bool need_rmap_locks)
378 {
379         bool moved = false;
380
381         /* See comment in move_ptes() */
382         if (need_rmap_locks)
383                 take_rmap_locks(vma);
384
385         switch (entry) {
386         case NORMAL_PMD:
387                 moved = move_normal_pmd(vma, old_addr, new_addr, old_entry,
388                                         new_entry);
389                 break;
390         case NORMAL_PUD:
391                 moved = move_normal_pud(vma, old_addr, new_addr, old_entry,
392                                         new_entry);
393                 break;
394         case HPAGE_PMD:
395                 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
396                         move_huge_pmd(vma, old_addr, new_addr, old_entry,
397                                       new_entry);
398                 break;
399         default:
400                 WARN_ON_ONCE(1);
401                 break;
402         }
403
404         if (need_rmap_locks)
405                 drop_rmap_locks(vma);
406
407         return moved;
408 }
409
410 unsigned long move_page_tables(struct vm_area_struct *vma,
411                 unsigned long old_addr, struct vm_area_struct *new_vma,
412                 unsigned long new_addr, unsigned long len,
413                 bool need_rmap_locks)
414 {
415         unsigned long extent, old_end;
416         struct mmu_notifier_range range;
417         pmd_t *old_pmd, *new_pmd;
418
419         old_end = old_addr + len;
420         flush_cache_range(vma, old_addr, old_end);
421
422         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
423                                 old_addr, old_end);
424         mmu_notifier_invalidate_range_start(&range);
425
426         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
427                 cond_resched();
428                 /*
429                  * If extent is PUD-sized try to speed up the move by moving at the
430                  * PUD level if possible.
431                  */
432                 extent = get_extent(NORMAL_PUD, old_addr, old_end, new_addr);
433                 if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
434                         pud_t *old_pud, *new_pud;
435
436                         old_pud = get_old_pud(vma->vm_mm, old_addr);
437                         if (!old_pud)
438                                 continue;
439                         new_pud = alloc_new_pud(vma->vm_mm, vma, new_addr);
440                         if (!new_pud)
441                                 break;
442                         if (move_pgt_entry(NORMAL_PUD, vma, old_addr, new_addr,
443                                            old_pud, new_pud, need_rmap_locks))
444                                 continue;
445                 }
446
447                 extent = get_extent(NORMAL_PMD, old_addr, old_end, new_addr);
448                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
449                 if (!old_pmd)
450                         continue;
451                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
452                 if (!new_pmd)
453                         break;
454                 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) ||
455                     pmd_devmap(*old_pmd)) {
456                         if (extent == HPAGE_PMD_SIZE &&
457                             move_pgt_entry(HPAGE_PMD, vma, old_addr, new_addr,
458                                            old_pmd, new_pmd, need_rmap_locks))
459                                 continue;
460                         split_huge_pmd(vma, old_pmd, old_addr);
461                         if (pmd_trans_unstable(old_pmd))
462                                 continue;
463                 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
464                            extent == PMD_SIZE) {
465                         /*
466                          * If the extent is PMD-sized, try to speed the move by
467                          * moving at the PMD level if possible.
468                          */
469                         if (move_pgt_entry(NORMAL_PMD, vma, old_addr, new_addr,
470                                            old_pmd, new_pmd, need_rmap_locks))
471                                 continue;
472                 }
473
474                 if (pte_alloc(new_vma->vm_mm, new_pmd))
475                         break;
476                 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
477                           new_pmd, new_addr, need_rmap_locks);
478         }
479
480         mmu_notifier_invalidate_range_end(&range);
481
482         return len + old_addr - old_end;        /* how much done */
483 }
484
485 static unsigned long move_vma(struct vm_area_struct *vma,
486                 unsigned long old_addr, unsigned long old_len,
487                 unsigned long new_len, unsigned long new_addr,
488                 bool *locked, unsigned long flags,
489                 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
490 {
491         struct mm_struct *mm = vma->vm_mm;
492         struct vm_area_struct *new_vma;
493         unsigned long vm_flags = vma->vm_flags;
494         unsigned long new_pgoff;
495         unsigned long moved_len;
496         unsigned long excess = 0;
497         unsigned long hiwater_vm;
498         int split = 0;
499         int err = 0;
500         bool need_rmap_locks;
501
502         /*
503          * We'd prefer to avoid failure later on in do_munmap:
504          * which may split one vma into three before unmapping.
505          */
506         if (mm->map_count >= sysctl_max_map_count - 3)
507                 return -ENOMEM;
508
509         if (vma->vm_ops && vma->vm_ops->may_split) {
510                 if (vma->vm_start != old_addr)
511                         err = vma->vm_ops->may_split(vma, old_addr);
512                 if (!err && vma->vm_end != old_addr + old_len)
513                         err = vma->vm_ops->may_split(vma, old_addr + old_len);
514                 if (err)
515                         return err;
516         }
517
518         /*
519          * Advise KSM to break any KSM pages in the area to be moved:
520          * it would be confusing if they were to turn up at the new
521          * location, where they happen to coincide with different KSM
522          * pages recently unmapped.  But leave vma->vm_flags as it was,
523          * so KSM can come around to merge on vma and new_vma afterwards.
524          */
525         err = ksm_madvise(vma, old_addr, old_addr + old_len,
526                                                 MADV_UNMERGEABLE, &vm_flags);
527         if (err)
528                 return err;
529
530         if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT)) {
531                 if (security_vm_enough_memory_mm(mm, new_len >> PAGE_SHIFT))
532                         return -ENOMEM;
533         }
534
535         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
536         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
537                            &need_rmap_locks);
538         if (!new_vma) {
539                 if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT))
540                         vm_unacct_memory(new_len >> PAGE_SHIFT);
541                 return -ENOMEM;
542         }
543
544         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
545                                      need_rmap_locks);
546         if (moved_len < old_len) {
547                 err = -ENOMEM;
548         } else if (vma->vm_ops && vma->vm_ops->mremap) {
549                 err = vma->vm_ops->mremap(new_vma, flags);
550         }
551
552         if (unlikely(err)) {
553                 /*
554                  * On error, move entries back from new area to old,
555                  * which will succeed since page tables still there,
556                  * and then proceed to unmap new area instead of old.
557                  */
558                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
559                                  true);
560                 vma = new_vma;
561                 old_len = new_len;
562                 old_addr = new_addr;
563                 new_addr = err;
564         } else {
565                 mremap_userfaultfd_prep(new_vma, uf);
566                 arch_remap(mm, old_addr, old_addr + old_len,
567                            new_addr, new_addr + new_len);
568         }
569
570         /* Conceal VM_ACCOUNT so old reservation is not undone */
571         if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP)) {
572                 vma->vm_flags &= ~VM_ACCOUNT;
573                 excess = vma->vm_end - vma->vm_start - old_len;
574                 if (old_addr > vma->vm_start &&
575                     old_addr + old_len < vma->vm_end)
576                         split = 1;
577         }
578
579         /*
580          * If we failed to move page tables we still do total_vm increment
581          * since do_munmap() will decrement it by old_len == new_len.
582          *
583          * Since total_vm is about to be raised artificially high for a
584          * moment, we need to restore high watermark afterwards: if stats
585          * are taken meanwhile, total_vm and hiwater_vm appear too high.
586          * If this were a serious issue, we'd add a flag to do_munmap().
587          */
588         hiwater_vm = mm->hiwater_vm;
589         vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
590
591         /* Tell pfnmap has moved from this vma */
592         if (unlikely(vma->vm_flags & VM_PFNMAP))
593                 untrack_pfn_moved(vma);
594
595         if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
596                 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
597                 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
598
599                 /* Because we won't unmap we don't need to touch locked_vm */
600                 return new_addr;
601         }
602
603         if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
604                 /* OOM: unable to split vma, just get accounts right */
605                 if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP))
606                         vm_acct_memory(new_len >> PAGE_SHIFT);
607                 excess = 0;
608         }
609
610         if (vm_flags & VM_LOCKED) {
611                 mm->locked_vm += new_len >> PAGE_SHIFT;
612                 *locked = true;
613         }
614
615         mm->hiwater_vm = hiwater_vm;
616
617         /* Restore VM_ACCOUNT if one or two pieces of vma left */
618         if (excess) {
619                 vma->vm_flags |= VM_ACCOUNT;
620                 if (split)
621                         vma->vm_next->vm_flags |= VM_ACCOUNT;
622         }
623
624         return new_addr;
625 }
626
627 static struct vm_area_struct *vma_to_resize(unsigned long addr,
628         unsigned long old_len, unsigned long new_len, unsigned long flags,
629         unsigned long *p)
630 {
631         struct mm_struct *mm = current->mm;
632         struct vm_area_struct *vma = find_vma(mm, addr);
633         unsigned long pgoff;
634
635         if (!vma || vma->vm_start > addr)
636                 return ERR_PTR(-EFAULT);
637
638         /*
639          * !old_len is a special case where an attempt is made to 'duplicate'
640          * a mapping.  This makes no sense for private mappings as it will
641          * instead create a fresh/new mapping unrelated to the original.  This
642          * is contrary to the basic idea of mremap which creates new mappings
643          * based on the original.  There are no known use cases for this
644          * behavior.  As a result, fail such attempts.
645          */
646         if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
647                 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap.  This is not supported.\n", current->comm, current->pid);
648                 return ERR_PTR(-EINVAL);
649         }
650
651         if (flags & MREMAP_DONTUNMAP && (!vma_is_anonymous(vma) ||
652                         vma->vm_flags & VM_SHARED))
653                 return ERR_PTR(-EINVAL);
654
655         if (is_vm_hugetlb_page(vma))
656                 return ERR_PTR(-EINVAL);
657
658         /* We can't remap across vm area boundaries */
659         if (old_len > vma->vm_end - addr)
660                 return ERR_PTR(-EFAULT);
661
662         if (new_len == old_len)
663                 return vma;
664
665         /* Need to be careful about a growing mapping */
666         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
667         pgoff += vma->vm_pgoff;
668         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
669                 return ERR_PTR(-EINVAL);
670
671         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
672                 return ERR_PTR(-EFAULT);
673
674         if (vma->vm_flags & VM_LOCKED) {
675                 unsigned long locked, lock_limit;
676                 locked = mm->locked_vm << PAGE_SHIFT;
677                 lock_limit = rlimit(RLIMIT_MEMLOCK);
678                 locked += new_len - old_len;
679                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
680                         return ERR_PTR(-EAGAIN);
681         }
682
683         if (!may_expand_vm(mm, vma->vm_flags,
684                                 (new_len - old_len) >> PAGE_SHIFT))
685                 return ERR_PTR(-ENOMEM);
686
687         if (vma->vm_flags & VM_ACCOUNT) {
688                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
689                 if (security_vm_enough_memory_mm(mm, charged))
690                         return ERR_PTR(-ENOMEM);
691                 *p = charged;
692         }
693
694         return vma;
695 }
696
697 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
698                 unsigned long new_addr, unsigned long new_len, bool *locked,
699                 unsigned long flags, struct vm_userfaultfd_ctx *uf,
700                 struct list_head *uf_unmap_early,
701                 struct list_head *uf_unmap)
702 {
703         struct mm_struct *mm = current->mm;
704         struct vm_area_struct *vma;
705         unsigned long ret = -EINVAL;
706         unsigned long charged = 0;
707         unsigned long map_flags = 0;
708
709         if (offset_in_page(new_addr))
710                 goto out;
711
712         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
713                 goto out;
714
715         /* Ensure the old/new locations do not overlap */
716         if (addr + old_len > new_addr && new_addr + new_len > addr)
717                 goto out;
718
719         /*
720          * move_vma() need us to stay 4 maps below the threshold, otherwise
721          * it will bail out at the very beginning.
722          * That is a problem if we have already unmaped the regions here
723          * (new_addr, and old_addr), because userspace will not know the
724          * state of the vma's after it gets -ENOMEM.
725          * So, to avoid such scenario we can pre-compute if the whole
726          * operation has high chances to success map-wise.
727          * Worst-scenario case is when both vma's (new_addr and old_addr) get
728          * split in 3 before unmaping it.
729          * That means 2 more maps (1 for each) to the ones we already hold.
730          * Check whether current map count plus 2 still leads us to 4 maps below
731          * the threshold, otherwise return -ENOMEM here to be more safe.
732          */
733         if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
734                 return -ENOMEM;
735
736         if (flags & MREMAP_FIXED) {
737                 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
738                 if (ret)
739                         goto out;
740         }
741
742         if (old_len >= new_len) {
743                 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
744                 if (ret && old_len != new_len)
745                         goto out;
746                 old_len = new_len;
747         }
748
749         vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
750         if (IS_ERR(vma)) {
751                 ret = PTR_ERR(vma);
752                 goto out;
753         }
754
755         /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
756         if (flags & MREMAP_DONTUNMAP &&
757                 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
758                 ret = -ENOMEM;
759                 goto out;
760         }
761
762         if (flags & MREMAP_FIXED)
763                 map_flags |= MAP_FIXED;
764
765         if (vma->vm_flags & VM_MAYSHARE)
766                 map_flags |= MAP_SHARED;
767
768         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
769                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
770                                 map_flags);
771         if (IS_ERR_VALUE(ret))
772                 goto out1;
773
774         /* We got a new mapping */
775         if (!(flags & MREMAP_FIXED))
776                 new_addr = ret;
777
778         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
779                        uf_unmap);
780
781         if (!(offset_in_page(ret)))
782                 goto out;
783
784 out1:
785         vm_unacct_memory(charged);
786
787 out:
788         return ret;
789 }
790
791 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
792 {
793         unsigned long end = vma->vm_end + delta;
794         if (end < vma->vm_end) /* overflow */
795                 return 0;
796         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
797                 return 0;
798         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
799                               0, MAP_FIXED) & ~PAGE_MASK)
800                 return 0;
801         return 1;
802 }
803
804 /*
805  * Expand (or shrink) an existing mapping, potentially moving it at the
806  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
807  *
808  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
809  * This option implies MREMAP_MAYMOVE.
810  */
811 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
812                 unsigned long, new_len, unsigned long, flags,
813                 unsigned long, new_addr)
814 {
815         struct mm_struct *mm = current->mm;
816         struct vm_area_struct *vma;
817         unsigned long ret = -EINVAL;
818         unsigned long charged = 0;
819         bool locked = false;
820         bool downgraded = false;
821         struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
822         LIST_HEAD(uf_unmap_early);
823         LIST_HEAD(uf_unmap);
824
825         /*
826          * There is a deliberate asymmetry here: we strip the pointer tag
827          * from the old address but leave the new address alone. This is
828          * for consistency with mmap(), where we prevent the creation of
829          * aliasing mappings in userspace by leaving the tag bits of the
830          * mapping address intact. A non-zero tag will cause the subsequent
831          * range checks to reject the address as invalid.
832          *
833          * See Documentation/arm64/tagged-address-abi.rst for more information.
834          */
835         addr = untagged_addr(addr);
836
837         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
838                 return ret;
839
840         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
841                 return ret;
842
843         /*
844          * MREMAP_DONTUNMAP is always a move and it does not allow resizing
845          * in the process.
846          */
847         if (flags & MREMAP_DONTUNMAP &&
848                         (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
849                 return ret;
850
851
852         if (offset_in_page(addr))
853                 return ret;
854
855         old_len = PAGE_ALIGN(old_len);
856         new_len = PAGE_ALIGN(new_len);
857
858         /*
859          * We allow a zero old-len as a special case
860          * for DOS-emu "duplicate shm area" thing. But
861          * a zero new-len is nonsensical.
862          */
863         if (!new_len)
864                 return ret;
865
866         if (mmap_write_lock_killable(current->mm))
867                 return -EINTR;
868
869         if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
870                 ret = mremap_to(addr, old_len, new_addr, new_len,
871                                 &locked, flags, &uf, &uf_unmap_early,
872                                 &uf_unmap);
873                 goto out;
874         }
875
876         /*
877          * Always allow a shrinking remap: that just unmaps
878          * the unnecessary pages..
879          * __do_munmap does all the needed commit accounting, and
880          * downgrades mmap_lock to read if so directed.
881          */
882         if (old_len >= new_len) {
883                 int retval;
884
885                 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
886                                   &uf_unmap, true);
887                 if (retval < 0 && old_len != new_len) {
888                         ret = retval;
889                         goto out;
890                 /* Returning 1 indicates mmap_lock is downgraded to read. */
891                 } else if (retval == 1)
892                         downgraded = true;
893                 ret = addr;
894                 goto out;
895         }
896
897         /*
898          * Ok, we need to grow..
899          */
900         vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
901         if (IS_ERR(vma)) {
902                 ret = PTR_ERR(vma);
903                 goto out;
904         }
905
906         /* old_len exactly to the end of the area..
907          */
908         if (old_len == vma->vm_end - addr) {
909                 /* can we just expand the current mapping? */
910                 if (vma_expandable(vma, new_len - old_len)) {
911                         int pages = (new_len - old_len) >> PAGE_SHIFT;
912
913                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
914                                        vma->vm_pgoff, NULL)) {
915                                 ret = -ENOMEM;
916                                 goto out;
917                         }
918
919                         vm_stat_account(mm, vma->vm_flags, pages);
920                         if (vma->vm_flags & VM_LOCKED) {
921                                 mm->locked_vm += pages;
922                                 locked = true;
923                                 new_addr = addr;
924                         }
925                         ret = addr;
926                         goto out;
927                 }
928         }
929
930         /*
931          * We weren't able to just expand or shrink the area,
932          * we need to create a new one and move it..
933          */
934         ret = -ENOMEM;
935         if (flags & MREMAP_MAYMOVE) {
936                 unsigned long map_flags = 0;
937                 if (vma->vm_flags & VM_MAYSHARE)
938                         map_flags |= MAP_SHARED;
939
940                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
941                                         vma->vm_pgoff +
942                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
943                                         map_flags);
944                 if (IS_ERR_VALUE(new_addr)) {
945                         ret = new_addr;
946                         goto out;
947                 }
948
949                 ret = move_vma(vma, addr, old_len, new_len, new_addr,
950                                &locked, flags, &uf, &uf_unmap);
951         }
952 out:
953         if (offset_in_page(ret)) {
954                 vm_unacct_memory(charged);
955                 locked = false;
956         }
957         if (downgraded)
958                 mmap_read_unlock(current->mm);
959         else
960                 mmap_write_unlock(current->mm);
961         if (locked && new_len > old_len)
962                 mm_populate(new_addr + old_len, new_len - old_len);
963         userfaultfd_unmap_complete(mm, &uf_unmap_early);
964         mremap_userfaultfd_complete(&uf, addr, ret, old_len);
965         userfaultfd_unmap_complete(mm, &uf_unmap);
966         return ret;
967 }