4 * (C) Copyright 1996 Linus Torvalds
6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
7 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
11 #include <linux/hugetlb.h>
12 #include <linux/shm.h>
13 #include <linux/ksm.h>
14 #include <linux/mman.h>
15 #include <linux/swap.h>
16 #include <linux/capability.h>
18 #include <linux/swapops.h>
19 #include <linux/highmem.h>
20 #include <linux/security.h>
21 #include <linux/syscalls.h>
22 #include <linux/mmu_notifier.h>
23 #include <linux/uaccess.h>
24 #include <linux/mm-arch-hooks.h>
25 #include <linux/userfaultfd_k.h>
27 #include <asm/cacheflush.h>
28 #include <asm/tlbflush.h>
32 static pmd_t
*get_old_pmd(struct mm_struct
*mm
, unsigned long addr
)
39 pgd
= pgd_offset(mm
, addr
);
40 if (pgd_none_or_clear_bad(pgd
))
43 p4d
= p4d_offset(pgd
, addr
);
44 if (p4d_none_or_clear_bad(p4d
))
47 pud
= pud_offset(p4d
, addr
);
48 if (pud_none_or_clear_bad(pud
))
51 pmd
= pmd_offset(pud
, addr
);
58 static pmd_t
*alloc_new_pmd(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
66 pgd
= pgd_offset(mm
, addr
);
67 p4d
= p4d_alloc(mm
, pgd
, addr
);
70 pud
= pud_alloc(mm
, p4d
, addr
);
74 pmd
= pmd_alloc(mm
, pud
, addr
);
78 VM_BUG_ON(pmd_trans_huge(*pmd
));
83 static void take_rmap_locks(struct vm_area_struct
*vma
)
86 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
88 anon_vma_lock_write(vma
->anon_vma
);
91 static void drop_rmap_locks(struct vm_area_struct
*vma
)
94 anon_vma_unlock_write(vma
->anon_vma
);
96 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
99 static pte_t
move_soft_dirty_pte(pte_t pte
)
102 * Set soft dirty bit so we can notice
103 * in userspace the ptes were moved.
105 #ifdef CONFIG_MEM_SOFT_DIRTY
106 if (pte_present(pte
))
107 pte
= pte_mksoft_dirty(pte
);
108 else if (is_swap_pte(pte
))
109 pte
= pte_swp_mksoft_dirty(pte
);
114 static void move_ptes(struct vm_area_struct
*vma
, pmd_t
*old_pmd
,
115 unsigned long old_addr
, unsigned long old_end
,
116 struct vm_area_struct
*new_vma
, pmd_t
*new_pmd
,
117 unsigned long new_addr
, bool need_rmap_locks
, bool *need_flush
)
119 struct mm_struct
*mm
= vma
->vm_mm
;
120 pte_t
*old_pte
, *new_pte
, pte
;
121 spinlock_t
*old_ptl
, *new_ptl
;
122 bool force_flush
= false;
123 unsigned long len
= old_end
- old_addr
;
126 * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
127 * locks to ensure that rmap will always observe either the old or the
128 * new ptes. This is the easiest way to avoid races with
129 * truncate_pagecache(), page migration, etc...
131 * When need_rmap_locks is false, we use other ways to avoid
134 * - During exec() shift_arg_pages(), we use a specially tagged vma
135 * which rmap call sites look for using is_vma_temporary_stack().
137 * - During mremap(), new_vma is often known to be placed after vma
138 * in rmap traversal order. This ensures rmap will always observe
139 * either the old pte, or the new pte, or both (the page table locks
140 * serialize access to individual ptes, but only rmap traversal
141 * order guarantees that we won't miss both the old and new ptes).
144 take_rmap_locks(vma
);
147 * We don't have to worry about the ordering of src and dst
148 * pte locks because exclusive mmap_sem prevents deadlock.
150 old_pte
= pte_offset_map_lock(mm
, old_pmd
, old_addr
, &old_ptl
);
151 new_pte
= pte_offset_map(new_pmd
, new_addr
);
152 new_ptl
= pte_lockptr(mm
, new_pmd
);
153 if (new_ptl
!= old_ptl
)
154 spin_lock_nested(new_ptl
, SINGLE_DEPTH_NESTING
);
155 flush_tlb_batched_pending(vma
->vm_mm
);
156 arch_enter_lazy_mmu_mode();
158 for (; old_addr
< old_end
; old_pte
++, old_addr
+= PAGE_SIZE
,
159 new_pte
++, new_addr
+= PAGE_SIZE
) {
160 if (pte_none(*old_pte
))
163 pte
= ptep_get_and_clear(mm
, old_addr
, old_pte
);
165 * If we are remapping a dirty PTE, make sure
166 * to flush TLB before we drop the PTL for the
167 * old PTE or we may race with page_mkclean().
169 * This check has to be done after we removed the
170 * old PTE from page tables or another thread may
171 * dirty it after the check and before the removal.
173 if (pte_present(pte
) && pte_dirty(pte
))
175 pte
= move_pte(pte
, new_vma
->vm_page_prot
, old_addr
, new_addr
);
176 pte
= move_soft_dirty_pte(pte
);
177 set_pte_at(mm
, new_addr
, new_pte
, pte
);
180 arch_leave_lazy_mmu_mode();
181 if (new_ptl
!= old_ptl
)
182 spin_unlock(new_ptl
);
183 pte_unmap(new_pte
- 1);
185 flush_tlb_range(vma
, old_end
- len
, old_end
);
188 pte_unmap_unlock(old_pte
- 1, old_ptl
);
190 drop_rmap_locks(vma
);
193 #define LATENCY_LIMIT (64 * PAGE_SIZE)
195 unsigned long move_page_tables(struct vm_area_struct
*vma
,
196 unsigned long old_addr
, struct vm_area_struct
*new_vma
,
197 unsigned long new_addr
, unsigned long len
,
198 bool need_rmap_locks
)
200 unsigned long extent
, next
, old_end
;
201 pmd_t
*old_pmd
, *new_pmd
;
202 bool need_flush
= false;
203 unsigned long mmun_start
; /* For mmu_notifiers */
204 unsigned long mmun_end
; /* For mmu_notifiers */
206 old_end
= old_addr
+ len
;
207 flush_cache_range(vma
, old_addr
, old_end
);
209 mmun_start
= old_addr
;
211 mmu_notifier_invalidate_range_start(vma
->vm_mm
, mmun_start
, mmun_end
);
213 for (; old_addr
< old_end
; old_addr
+= extent
, new_addr
+= extent
) {
215 next
= (old_addr
+ PMD_SIZE
) & PMD_MASK
;
216 /* even if next overflowed, extent below will be ok */
217 extent
= next
- old_addr
;
218 if (extent
> old_end
- old_addr
)
219 extent
= old_end
- old_addr
;
220 old_pmd
= get_old_pmd(vma
->vm_mm
, old_addr
);
223 new_pmd
= alloc_new_pmd(vma
->vm_mm
, vma
, new_addr
);
226 if (pmd_trans_huge(*old_pmd
)) {
227 if (extent
== HPAGE_PMD_SIZE
) {
229 /* See comment in move_ptes() */
231 take_rmap_locks(vma
);
232 moved
= move_huge_pmd(vma
, old_addr
, new_addr
,
233 old_end
, old_pmd
, new_pmd
,
236 drop_rmap_locks(vma
);
240 split_huge_pmd(vma
, old_pmd
, old_addr
);
241 if (pmd_trans_unstable(old_pmd
))
244 if (pte_alloc(new_vma
->vm_mm
, new_pmd
, new_addr
))
246 next
= (new_addr
+ PMD_SIZE
) & PMD_MASK
;
247 if (extent
> next
- new_addr
)
248 extent
= next
- new_addr
;
249 if (extent
> LATENCY_LIMIT
)
250 extent
= LATENCY_LIMIT
;
251 move_ptes(vma
, old_pmd
, old_addr
, old_addr
+ extent
, new_vma
,
252 new_pmd
, new_addr
, need_rmap_locks
, &need_flush
);
255 flush_tlb_range(vma
, old_end
-len
, old_addr
);
257 mmu_notifier_invalidate_range_end(vma
->vm_mm
, mmun_start
, mmun_end
);
259 return len
+ old_addr
- old_end
; /* how much done */
262 static unsigned long move_vma(struct vm_area_struct
*vma
,
263 unsigned long old_addr
, unsigned long old_len
,
264 unsigned long new_len
, unsigned long new_addr
,
265 bool *locked
, struct vm_userfaultfd_ctx
*uf
,
266 struct list_head
*uf_unmap
)
268 struct mm_struct
*mm
= vma
->vm_mm
;
269 struct vm_area_struct
*new_vma
;
270 unsigned long vm_flags
= vma
->vm_flags
;
271 unsigned long new_pgoff
;
272 unsigned long moved_len
;
273 unsigned long excess
= 0;
274 unsigned long hiwater_vm
;
277 bool need_rmap_locks
;
280 * We'd prefer to avoid failure later on in do_munmap:
281 * which may split one vma into three before unmapping.
283 if (mm
->map_count
>= sysctl_max_map_count
- 3)
287 * Advise KSM to break any KSM pages in the area to be moved:
288 * it would be confusing if they were to turn up at the new
289 * location, where they happen to coincide with different KSM
290 * pages recently unmapped. But leave vma->vm_flags as it was,
291 * so KSM can come around to merge on vma and new_vma afterwards.
293 err
= ksm_madvise(vma
, old_addr
, old_addr
+ old_len
,
294 MADV_UNMERGEABLE
, &vm_flags
);
298 new_pgoff
= vma
->vm_pgoff
+ ((old_addr
- vma
->vm_start
) >> PAGE_SHIFT
);
299 new_vma
= copy_vma(&vma
, new_addr
, new_len
, new_pgoff
,
304 moved_len
= move_page_tables(vma
, old_addr
, new_vma
, new_addr
, old_len
,
306 if (moved_len
< old_len
) {
308 } else if (vma
->vm_ops
&& vma
->vm_ops
->mremap
) {
309 err
= vma
->vm_ops
->mremap(new_vma
);
314 * On error, move entries back from new area to old,
315 * which will succeed since page tables still there,
316 * and then proceed to unmap new area instead of old.
318 move_page_tables(new_vma
, new_addr
, vma
, old_addr
, moved_len
,
325 mremap_userfaultfd_prep(new_vma
, uf
);
326 arch_remap(mm
, old_addr
, old_addr
+ old_len
,
327 new_addr
, new_addr
+ new_len
);
330 /* Conceal VM_ACCOUNT so old reservation is not undone */
331 if (vm_flags
& VM_ACCOUNT
) {
332 vma
->vm_flags
&= ~VM_ACCOUNT
;
333 excess
= vma
->vm_end
- vma
->vm_start
- old_len
;
334 if (old_addr
> vma
->vm_start
&&
335 old_addr
+ old_len
< vma
->vm_end
)
340 * If we failed to move page tables we still do total_vm increment
341 * since do_munmap() will decrement it by old_len == new_len.
343 * Since total_vm is about to be raised artificially high for a
344 * moment, we need to restore high watermark afterwards: if stats
345 * are taken meanwhile, total_vm and hiwater_vm appear too high.
346 * If this were a serious issue, we'd add a flag to do_munmap().
348 hiwater_vm
= mm
->hiwater_vm
;
349 vm_stat_account(mm
, vma
->vm_flags
, new_len
>> PAGE_SHIFT
);
351 /* Tell pfnmap has moved from this vma */
352 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
353 untrack_pfn_moved(vma
);
355 if (do_munmap(mm
, old_addr
, old_len
, uf_unmap
) < 0) {
356 /* OOM: unable to split vma, just get accounts right */
357 vm_unacct_memory(excess
>> PAGE_SHIFT
);
360 mm
->hiwater_vm
= hiwater_vm
;
362 /* Restore VM_ACCOUNT if one or two pieces of vma left */
364 vma
->vm_flags
|= VM_ACCOUNT
;
366 vma
->vm_next
->vm_flags
|= VM_ACCOUNT
;
369 if (vm_flags
& VM_LOCKED
) {
370 mm
->locked_vm
+= new_len
>> PAGE_SHIFT
;
377 static struct vm_area_struct
*vma_to_resize(unsigned long addr
,
378 unsigned long old_len
, unsigned long new_len
, unsigned long *p
)
380 struct mm_struct
*mm
= current
->mm
;
381 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
384 if (!vma
|| vma
->vm_start
> addr
)
385 return ERR_PTR(-EFAULT
);
387 if (is_vm_hugetlb_page(vma
))
388 return ERR_PTR(-EINVAL
);
390 /* We can't remap across vm area boundaries */
391 if (old_len
> vma
->vm_end
- addr
)
392 return ERR_PTR(-EFAULT
);
394 if (new_len
== old_len
)
397 /* Need to be careful about a growing mapping */
398 pgoff
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
399 pgoff
+= vma
->vm_pgoff
;
400 if (pgoff
+ (new_len
>> PAGE_SHIFT
) < pgoff
)
401 return ERR_PTR(-EINVAL
);
403 if (vma
->vm_flags
& (VM_DONTEXPAND
| VM_PFNMAP
))
404 return ERR_PTR(-EFAULT
);
406 if (vma
->vm_flags
& VM_LOCKED
) {
407 unsigned long locked
, lock_limit
;
408 locked
= mm
->locked_vm
<< PAGE_SHIFT
;
409 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
410 locked
+= new_len
- old_len
;
411 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
412 return ERR_PTR(-EAGAIN
);
415 if (!may_expand_vm(mm
, vma
->vm_flags
,
416 (new_len
- old_len
) >> PAGE_SHIFT
))
417 return ERR_PTR(-ENOMEM
);
419 if (vma
->vm_flags
& VM_ACCOUNT
) {
420 unsigned long charged
= (new_len
- old_len
) >> PAGE_SHIFT
;
421 if (security_vm_enough_memory_mm(mm
, charged
))
422 return ERR_PTR(-ENOMEM
);
429 static unsigned long mremap_to(unsigned long addr
, unsigned long old_len
,
430 unsigned long new_addr
, unsigned long new_len
, bool *locked
,
431 struct vm_userfaultfd_ctx
*uf
,
432 struct list_head
*uf_unmap_early
,
433 struct list_head
*uf_unmap
)
435 struct mm_struct
*mm
= current
->mm
;
436 struct vm_area_struct
*vma
;
437 unsigned long ret
= -EINVAL
;
438 unsigned long charged
= 0;
439 unsigned long map_flags
;
441 if (offset_in_page(new_addr
))
444 if (new_len
> TASK_SIZE
|| new_addr
> TASK_SIZE
- new_len
)
447 /* Ensure the old/new locations do not overlap */
448 if (addr
+ old_len
> new_addr
&& new_addr
+ new_len
> addr
)
451 ret
= do_munmap(mm
, new_addr
, new_len
, uf_unmap_early
);
455 if (old_len
>= new_len
) {
456 ret
= do_munmap(mm
, addr
+new_len
, old_len
- new_len
, uf_unmap
);
457 if (ret
&& old_len
!= new_len
)
462 vma
= vma_to_resize(addr
, old_len
, new_len
, &charged
);
468 map_flags
= MAP_FIXED
;
469 if (vma
->vm_flags
& VM_MAYSHARE
)
470 map_flags
|= MAP_SHARED
;
472 ret
= get_unmapped_area(vma
->vm_file
, new_addr
, new_len
, vma
->vm_pgoff
+
473 ((addr
- vma
->vm_start
) >> PAGE_SHIFT
),
475 if (offset_in_page(ret
))
478 ret
= move_vma(vma
, addr
, old_len
, new_len
, new_addr
, locked
, uf
,
480 if (!(offset_in_page(ret
)))
483 vm_unacct_memory(charged
);
489 static int vma_expandable(struct vm_area_struct
*vma
, unsigned long delta
)
491 unsigned long end
= vma
->vm_end
+ delta
;
492 if (end
< vma
->vm_end
) /* overflow */
494 if (vma
->vm_next
&& vma
->vm_next
->vm_start
< end
) /* intersection */
496 if (get_unmapped_area(NULL
, vma
->vm_start
, end
- vma
->vm_start
,
497 0, MAP_FIXED
) & ~PAGE_MASK
)
503 * Expand (or shrink) an existing mapping, potentially moving it at the
504 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
506 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
507 * This option implies MREMAP_MAYMOVE.
509 SYSCALL_DEFINE5(mremap
, unsigned long, addr
, unsigned long, old_len
,
510 unsigned long, new_len
, unsigned long, flags
,
511 unsigned long, new_addr
)
513 struct mm_struct
*mm
= current
->mm
;
514 struct vm_area_struct
*vma
;
515 unsigned long ret
= -EINVAL
;
516 unsigned long charged
= 0;
518 struct vm_userfaultfd_ctx uf
= NULL_VM_UFFD_CTX
;
519 LIST_HEAD(uf_unmap_early
);
522 if (flags
& ~(MREMAP_FIXED
| MREMAP_MAYMOVE
))
525 if (flags
& MREMAP_FIXED
&& !(flags
& MREMAP_MAYMOVE
))
528 if (offset_in_page(addr
))
531 old_len
= PAGE_ALIGN(old_len
);
532 new_len
= PAGE_ALIGN(new_len
);
535 * We allow a zero old-len as a special case
536 * for DOS-emu "duplicate shm area" thing. But
537 * a zero new-len is nonsensical.
542 if (down_write_killable(¤t
->mm
->mmap_sem
))
545 if (flags
& MREMAP_FIXED
) {
546 ret
= mremap_to(addr
, old_len
, new_addr
, new_len
,
547 &locked
, &uf
, &uf_unmap_early
, &uf_unmap
);
552 * Always allow a shrinking remap: that just unmaps
553 * the unnecessary pages..
554 * do_munmap does all the needed commit accounting
556 if (old_len
>= new_len
) {
557 ret
= do_munmap(mm
, addr
+new_len
, old_len
- new_len
, &uf_unmap
);
558 if (ret
&& old_len
!= new_len
)
565 * Ok, we need to grow..
567 vma
= vma_to_resize(addr
, old_len
, new_len
, &charged
);
573 /* old_len exactly to the end of the area..
575 if (old_len
== vma
->vm_end
- addr
) {
576 /* can we just expand the current mapping? */
577 if (vma_expandable(vma
, new_len
- old_len
)) {
578 int pages
= (new_len
- old_len
) >> PAGE_SHIFT
;
580 if (vma_adjust(vma
, vma
->vm_start
, addr
+ new_len
,
581 vma
->vm_pgoff
, NULL
)) {
586 vm_stat_account(mm
, vma
->vm_flags
, pages
);
587 if (vma
->vm_flags
& VM_LOCKED
) {
588 mm
->locked_vm
+= pages
;
598 * We weren't able to just expand or shrink the area,
599 * we need to create a new one and move it..
602 if (flags
& MREMAP_MAYMOVE
) {
603 unsigned long map_flags
= 0;
604 if (vma
->vm_flags
& VM_MAYSHARE
)
605 map_flags
|= MAP_SHARED
;
607 new_addr
= get_unmapped_area(vma
->vm_file
, 0, new_len
,
609 ((addr
- vma
->vm_start
) >> PAGE_SHIFT
),
611 if (offset_in_page(new_addr
)) {
616 ret
= move_vma(vma
, addr
, old_len
, new_len
, new_addr
,
617 &locked
, &uf
, &uf_unmap
);
620 if (offset_in_page(ret
)) {
621 vm_unacct_memory(charged
);
624 up_write(¤t
->mm
->mmap_sem
);
625 if (locked
&& new_len
> old_len
)
626 mm_populate(new_addr
+ old_len
, new_len
- old_len
);
627 userfaultfd_unmap_complete(mm
, &uf_unmap_early
);
628 mremap_userfaultfd_complete(&uf
, addr
, new_addr
, old_len
);
629 userfaultfd_unmap_complete(mm
, &uf_unmap
);