1 // SPDX-License-Identifier: GPL-2.0-only
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/backing-dev.h>
16 #include <linux/mm_inline.h>
17 #include <linux/vmacache.h>
18 #include <linux/shm.h>
19 #include <linux/mman.h>
20 #include <linux/pagemap.h>
21 #include <linux/swap.h>
22 #include <linux/syscalls.h>
23 #include <linux/capability.h>
24 #include <linux/init.h>
25 #include <linux/file.h>
27 #include <linux/personality.h>
28 #include <linux/security.h>
29 #include <linux/hugetlb.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/profile.h>
32 #include <linux/export.h>
33 #include <linux/mount.h>
34 #include <linux/mempolicy.h>
35 #include <linux/rmap.h>
36 #include <linux/mmu_notifier.h>
37 #include <linux/mmdebug.h>
38 #include <linux/perf_event.h>
39 #include <linux/audit.h>
40 #include <linux/khugepaged.h>
41 #include <linux/uprobes.h>
42 #include <linux/rbtree_augmented.h>
43 #include <linux/notifier.h>
44 #include <linux/memory.h>
45 #include <linux/printk.h>
46 #include <linux/userfaultfd_k.h>
47 #include <linux/moduleparam.h>
48 #include <linux/pkeys.h>
49 #include <linux/oom.h>
50 #include <linux/sched/mm.h>
52 #include <linux/uaccess.h>
53 #include <asm/cacheflush.h>
55 #include <asm/mmu_context.h>
57 #define CREATE_TRACE_POINTS
58 #include <trace/events/mmap.h>
62 #ifndef arch_mmap_check
63 #define arch_mmap_check(addr, len, flags) (0)
66 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
67 const int mmap_rnd_bits_min
= CONFIG_ARCH_MMAP_RND_BITS_MIN
;
68 const int mmap_rnd_bits_max
= CONFIG_ARCH_MMAP_RND_BITS_MAX
;
69 int mmap_rnd_bits __read_mostly
= CONFIG_ARCH_MMAP_RND_BITS
;
71 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
72 const int mmap_rnd_compat_bits_min
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN
;
73 const int mmap_rnd_compat_bits_max
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX
;
74 int mmap_rnd_compat_bits __read_mostly
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS
;
77 static bool ignore_rlimit_data
;
78 core_param(ignore_rlimit_data
, ignore_rlimit_data
, bool, 0644);
80 static void unmap_region(struct mm_struct
*mm
,
81 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
82 unsigned long start
, unsigned long end
);
84 static pgprot_t
vm_pgprot_modify(pgprot_t oldprot
, unsigned long vm_flags
)
86 return pgprot_modify(oldprot
, vm_get_page_prot(vm_flags
));
89 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
90 void vma_set_page_prot(struct vm_area_struct
*vma
)
92 unsigned long vm_flags
= vma
->vm_flags
;
93 pgprot_t vm_page_prot
;
95 vm_page_prot
= vm_pgprot_modify(vma
->vm_page_prot
, vm_flags
);
96 if (vma_wants_writenotify(vma
, vm_page_prot
)) {
97 vm_flags
&= ~VM_SHARED
;
98 vm_page_prot
= vm_pgprot_modify(vm_page_prot
, vm_flags
);
100 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
101 WRITE_ONCE(vma
->vm_page_prot
, vm_page_prot
);
105 * Requires inode->i_mapping->i_mmap_rwsem
107 static void __remove_shared_vm_struct(struct vm_area_struct
*vma
,
108 struct file
*file
, struct address_space
*mapping
)
110 if (vma
->vm_flags
& VM_SHARED
)
111 mapping_unmap_writable(mapping
);
113 flush_dcache_mmap_lock(mapping
);
114 vma_interval_tree_remove(vma
, &mapping
->i_mmap
);
115 flush_dcache_mmap_unlock(mapping
);
119 * Unlink a file-based vm structure from its interval tree, to hide
120 * vma from rmap and vmtruncate before freeing its page tables.
122 void unlink_file_vma(struct vm_area_struct
*vma
)
124 struct file
*file
= vma
->vm_file
;
127 struct address_space
*mapping
= file
->f_mapping
;
128 i_mmap_lock_write(mapping
);
129 __remove_shared_vm_struct(vma
, file
, mapping
);
130 i_mmap_unlock_write(mapping
);
135 * Close a vm structure and free it, returning the next.
137 static struct vm_area_struct
*remove_vma(struct vm_area_struct
*vma
)
139 struct vm_area_struct
*next
= vma
->vm_next
;
142 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
143 vma
->vm_ops
->close(vma
);
146 mpol_put(vma_policy(vma
));
151 static int do_brk_flags(unsigned long addr
, unsigned long request
, unsigned long flags
,
152 struct list_head
*uf
);
153 SYSCALL_DEFINE1(brk
, unsigned long, brk
)
155 unsigned long newbrk
, oldbrk
, origbrk
;
156 struct mm_struct
*mm
= current
->mm
;
157 struct vm_area_struct
*next
;
158 unsigned long min_brk
;
160 bool downgraded
= false;
163 if (mmap_write_lock_killable(mm
))
168 #ifdef CONFIG_COMPAT_BRK
170 * CONFIG_COMPAT_BRK can still be overridden by setting
171 * randomize_va_space to 2, which will still cause mm->start_brk
172 * to be arbitrarily shifted
174 if (current
->brk_randomized
)
175 min_brk
= mm
->start_brk
;
177 min_brk
= mm
->end_data
;
179 min_brk
= mm
->start_brk
;
185 * Check against rlimit here. If this check is done later after the test
186 * of oldbrk with newbrk then it can escape the test and let the data
187 * segment grow beyond its set limit the in case where the limit is
188 * not page aligned -Ram Gupta
190 if (check_data_rlimit(rlimit(RLIMIT_DATA
), brk
, mm
->start_brk
,
191 mm
->end_data
, mm
->start_data
))
194 newbrk
= PAGE_ALIGN(brk
);
195 oldbrk
= PAGE_ALIGN(mm
->brk
);
196 if (oldbrk
== newbrk
) {
202 * Always allow shrinking brk.
203 * __do_munmap() may downgrade mmap_lock to read.
205 if (brk
<= mm
->brk
) {
209 * mm->brk must to be protected by write mmap_lock so update it
210 * before downgrading mmap_lock. When __do_munmap() fails,
211 * mm->brk will be restored from origbrk.
214 ret
= __do_munmap(mm
, newbrk
, oldbrk
-newbrk
, &uf
, true);
218 } else if (ret
== 1) {
224 /* Check against existing mmap mappings. */
225 next
= find_vma(mm
, oldbrk
);
226 if (next
&& newbrk
+ PAGE_SIZE
> vm_start_gap(next
))
229 /* Ok, looks good - let it rip. */
230 if (do_brk_flags(oldbrk
, newbrk
-oldbrk
, 0, &uf
) < 0)
235 populate
= newbrk
> oldbrk
&& (mm
->def_flags
& VM_LOCKED
) != 0;
237 mmap_read_unlock(mm
);
239 mmap_write_unlock(mm
);
240 userfaultfd_unmap_complete(mm
, &uf
);
242 mm_populate(oldbrk
, newbrk
- oldbrk
);
246 mmap_write_unlock(mm
);
250 static inline unsigned long vma_compute_gap(struct vm_area_struct
*vma
)
252 unsigned long gap
, prev_end
;
255 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
256 * allow two stack_guard_gaps between them here, and when choosing
257 * an unmapped area; whereas when expanding we only require one.
258 * That's a little inconsistent, but keeps the code here simpler.
260 gap
= vm_start_gap(vma
);
262 prev_end
= vm_end_gap(vma
->vm_prev
);
271 #ifdef CONFIG_DEBUG_VM_RB
272 static unsigned long vma_compute_subtree_gap(struct vm_area_struct
*vma
)
274 unsigned long max
= vma_compute_gap(vma
), subtree_gap
;
275 if (vma
->vm_rb
.rb_left
) {
276 subtree_gap
= rb_entry(vma
->vm_rb
.rb_left
,
277 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
278 if (subtree_gap
> max
)
281 if (vma
->vm_rb
.rb_right
) {
282 subtree_gap
= rb_entry(vma
->vm_rb
.rb_right
,
283 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
284 if (subtree_gap
> max
)
290 static int browse_rb(struct mm_struct
*mm
)
292 struct rb_root
*root
= &mm
->mm_rb
;
293 int i
= 0, j
, bug
= 0;
294 struct rb_node
*nd
, *pn
= NULL
;
295 unsigned long prev
= 0, pend
= 0;
297 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
298 struct vm_area_struct
*vma
;
299 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
300 if (vma
->vm_start
< prev
) {
301 pr_emerg("vm_start %lx < prev %lx\n",
302 vma
->vm_start
, prev
);
305 if (vma
->vm_start
< pend
) {
306 pr_emerg("vm_start %lx < pend %lx\n",
307 vma
->vm_start
, pend
);
310 if (vma
->vm_start
> vma
->vm_end
) {
311 pr_emerg("vm_start %lx > vm_end %lx\n",
312 vma
->vm_start
, vma
->vm_end
);
315 spin_lock(&mm
->page_table_lock
);
316 if (vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
)) {
317 pr_emerg("free gap %lx, correct %lx\n",
319 vma_compute_subtree_gap(vma
));
322 spin_unlock(&mm
->page_table_lock
);
325 prev
= vma
->vm_start
;
329 for (nd
= pn
; nd
; nd
= rb_prev(nd
))
332 pr_emerg("backwards %d, forwards %d\n", j
, i
);
338 static void validate_mm_rb(struct rb_root
*root
, struct vm_area_struct
*ignore
)
342 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
343 struct vm_area_struct
*vma
;
344 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
345 VM_BUG_ON_VMA(vma
!= ignore
&&
346 vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
),
351 static void validate_mm(struct mm_struct
*mm
)
355 unsigned long highest_address
= 0;
356 struct vm_area_struct
*vma
= mm
->mmap
;
359 struct anon_vma
*anon_vma
= vma
->anon_vma
;
360 struct anon_vma_chain
*avc
;
363 anon_vma_lock_read(anon_vma
);
364 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
365 anon_vma_interval_tree_verify(avc
);
366 anon_vma_unlock_read(anon_vma
);
369 highest_address
= vm_end_gap(vma
);
373 if (i
!= mm
->map_count
) {
374 pr_emerg("map_count %d vm_next %d\n", mm
->map_count
, i
);
377 if (highest_address
!= mm
->highest_vm_end
) {
378 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
379 mm
->highest_vm_end
, highest_address
);
383 if (i
!= mm
->map_count
) {
385 pr_emerg("map_count %d rb %d\n", mm
->map_count
, i
);
388 VM_BUG_ON_MM(bug
, mm
);
391 #define validate_mm_rb(root, ignore) do { } while (0)
392 #define validate_mm(mm) do { } while (0)
395 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks
,
396 struct vm_area_struct
, vm_rb
,
397 unsigned long, rb_subtree_gap
, vma_compute_gap
)
400 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
401 * vma->vm_prev->vm_end values changed, without modifying the vma's position
404 static void vma_gap_update(struct vm_area_struct
*vma
)
407 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
408 * a callback function that does exactly what we want.
410 vma_gap_callbacks_propagate(&vma
->vm_rb
, NULL
);
413 static inline void vma_rb_insert(struct vm_area_struct
*vma
,
414 struct rb_root
*root
)
416 /* All rb_subtree_gap values must be consistent prior to insertion */
417 validate_mm_rb(root
, NULL
);
419 rb_insert_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
422 static void __vma_rb_erase(struct vm_area_struct
*vma
, struct rb_root
*root
)
425 * Note rb_erase_augmented is a fairly large inline function,
426 * so make sure we instantiate it only once with our desired
427 * augmented rbtree callbacks.
429 rb_erase_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
432 static __always_inline
void vma_rb_erase_ignore(struct vm_area_struct
*vma
,
433 struct rb_root
*root
,
434 struct vm_area_struct
*ignore
)
437 * All rb_subtree_gap values must be consistent prior to erase,
438 * with the possible exception of
440 * a. the "next" vma being erased if next->vm_start was reduced in
441 * __vma_adjust() -> __vma_unlink()
442 * b. the vma being erased in detach_vmas_to_be_unmapped() ->
445 validate_mm_rb(root
, ignore
);
447 __vma_rb_erase(vma
, root
);
450 static __always_inline
void vma_rb_erase(struct vm_area_struct
*vma
,
451 struct rb_root
*root
)
453 vma_rb_erase_ignore(vma
, root
, vma
);
457 * vma has some anon_vma assigned, and is already inserted on that
458 * anon_vma's interval trees.
460 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
461 * vma must be removed from the anon_vma's interval trees using
462 * anon_vma_interval_tree_pre_update_vma().
464 * After the update, the vma will be reinserted using
465 * anon_vma_interval_tree_post_update_vma().
467 * The entire update must be protected by exclusive mmap_lock and by
468 * the root anon_vma's mutex.
471 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct
*vma
)
473 struct anon_vma_chain
*avc
;
475 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
476 anon_vma_interval_tree_remove(avc
, &avc
->anon_vma
->rb_root
);
480 anon_vma_interval_tree_post_update_vma(struct vm_area_struct
*vma
)
482 struct anon_vma_chain
*avc
;
484 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
485 anon_vma_interval_tree_insert(avc
, &avc
->anon_vma
->rb_root
);
488 static int find_vma_links(struct mm_struct
*mm
, unsigned long addr
,
489 unsigned long end
, struct vm_area_struct
**pprev
,
490 struct rb_node
***rb_link
, struct rb_node
**rb_parent
)
492 struct rb_node
**__rb_link
, *__rb_parent
, *rb_prev
;
494 mmap_assert_locked(mm
);
495 __rb_link
= &mm
->mm_rb
.rb_node
;
496 rb_prev
= __rb_parent
= NULL
;
499 struct vm_area_struct
*vma_tmp
;
501 __rb_parent
= *__rb_link
;
502 vma_tmp
= rb_entry(__rb_parent
, struct vm_area_struct
, vm_rb
);
504 if (vma_tmp
->vm_end
> addr
) {
505 /* Fail if an existing vma overlaps the area */
506 if (vma_tmp
->vm_start
< end
)
508 __rb_link
= &__rb_parent
->rb_left
;
510 rb_prev
= __rb_parent
;
511 __rb_link
= &__rb_parent
->rb_right
;
517 *pprev
= rb_entry(rb_prev
, struct vm_area_struct
, vm_rb
);
518 *rb_link
= __rb_link
;
519 *rb_parent
= __rb_parent
;
524 * vma_next() - Get the next VMA.
525 * @mm: The mm_struct.
526 * @vma: The current vma.
528 * If @vma is NULL, return the first vma in the mm.
530 * Returns: The next VMA after @vma.
532 static inline struct vm_area_struct
*vma_next(struct mm_struct
*mm
,
533 struct vm_area_struct
*vma
)
542 * munmap_vma_range() - munmap VMAs that overlap a range.
544 * @start: The start of the range.
545 * @len: The length of the range.
546 * @pprev: pointer to the pointer that will be set to previous vm_area_struct
547 * @rb_link: the rb_node
548 * @rb_parent: the parent rb_node
550 * Find all the vm_area_struct that overlap from @start to
551 * @end and munmap them. Set @pprev to the previous vm_area_struct.
553 * Returns: -ENOMEM on munmap failure or 0 on success.
556 munmap_vma_range(struct mm_struct
*mm
, unsigned long start
, unsigned long len
,
557 struct vm_area_struct
**pprev
, struct rb_node
***link
,
558 struct rb_node
**parent
, struct list_head
*uf
)
561 while (find_vma_links(mm
, start
, start
+ len
, pprev
, link
, parent
))
562 if (do_munmap(mm
, start
, len
, uf
))
567 static unsigned long count_vma_pages_range(struct mm_struct
*mm
,
568 unsigned long addr
, unsigned long end
)
570 unsigned long nr_pages
= 0;
571 struct vm_area_struct
*vma
;
573 /* Find first overlapping mapping */
574 vma
= find_vma_intersection(mm
, addr
, end
);
578 nr_pages
= (min(end
, vma
->vm_end
) -
579 max(addr
, vma
->vm_start
)) >> PAGE_SHIFT
;
581 /* Iterate over the rest of the overlaps */
582 for (vma
= vma
->vm_next
; vma
; vma
= vma
->vm_next
) {
583 unsigned long overlap_len
;
585 if (vma
->vm_start
> end
)
588 overlap_len
= min(end
, vma
->vm_end
) - vma
->vm_start
;
589 nr_pages
+= overlap_len
>> PAGE_SHIFT
;
595 void __vma_link_rb(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
596 struct rb_node
**rb_link
, struct rb_node
*rb_parent
)
598 /* Update tracking information for the gap following the new vma. */
600 vma_gap_update(vma
->vm_next
);
602 mm
->highest_vm_end
= vm_end_gap(vma
);
605 * vma->vm_prev wasn't known when we followed the rbtree to find the
606 * correct insertion point for that vma. As a result, we could not
607 * update the vma vm_rb parents rb_subtree_gap values on the way down.
608 * So, we first insert the vma with a zero rb_subtree_gap value
609 * (to be consistent with what we did on the way down), and then
610 * immediately update the gap to the correct value. Finally we
611 * rebalance the rbtree after all augmented values have been set.
613 rb_link_node(&vma
->vm_rb
, rb_parent
, rb_link
);
614 vma
->rb_subtree_gap
= 0;
616 vma_rb_insert(vma
, &mm
->mm_rb
);
619 static void __vma_link_file(struct vm_area_struct
*vma
)
625 struct address_space
*mapping
= file
->f_mapping
;
627 if (vma
->vm_flags
& VM_SHARED
)
628 mapping_allow_writable(mapping
);
630 flush_dcache_mmap_lock(mapping
);
631 vma_interval_tree_insert(vma
, &mapping
->i_mmap
);
632 flush_dcache_mmap_unlock(mapping
);
637 __vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
638 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
639 struct rb_node
*rb_parent
)
641 __vma_link_list(mm
, vma
, prev
);
642 __vma_link_rb(mm
, vma
, rb_link
, rb_parent
);
645 static void vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
646 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
647 struct rb_node
*rb_parent
)
649 struct address_space
*mapping
= NULL
;
652 mapping
= vma
->vm_file
->f_mapping
;
653 i_mmap_lock_write(mapping
);
656 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
657 __vma_link_file(vma
);
660 i_mmap_unlock_write(mapping
);
667 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
668 * mm's list and rbtree. It has already been inserted into the interval tree.
670 static void __insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
672 struct vm_area_struct
*prev
;
673 struct rb_node
**rb_link
, *rb_parent
;
675 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
676 &prev
, &rb_link
, &rb_parent
))
678 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
682 static __always_inline
void __vma_unlink(struct mm_struct
*mm
,
683 struct vm_area_struct
*vma
,
684 struct vm_area_struct
*ignore
)
686 vma_rb_erase_ignore(vma
, &mm
->mm_rb
, ignore
);
687 __vma_unlink_list(mm
, vma
);
689 vmacache_invalidate(mm
);
693 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
694 * is already present in an i_mmap tree without adjusting the tree.
695 * The following helper function should be used when such adjustments
696 * are necessary. The "insert" vma (if any) is to be inserted
697 * before we drop the necessary locks.
699 int __vma_adjust(struct vm_area_struct
*vma
, unsigned long start
,
700 unsigned long end
, pgoff_t pgoff
, struct vm_area_struct
*insert
,
701 struct vm_area_struct
*expand
)
703 struct mm_struct
*mm
= vma
->vm_mm
;
704 struct vm_area_struct
*next
= vma
->vm_next
, *orig_vma
= vma
;
705 struct address_space
*mapping
= NULL
;
706 struct rb_root_cached
*root
= NULL
;
707 struct anon_vma
*anon_vma
= NULL
;
708 struct file
*file
= vma
->vm_file
;
709 bool start_changed
= false, end_changed
= false;
710 long adjust_next
= 0;
713 if (next
&& !insert
) {
714 struct vm_area_struct
*exporter
= NULL
, *importer
= NULL
;
716 if (end
>= next
->vm_end
) {
718 * vma expands, overlapping all the next, and
719 * perhaps the one after too (mprotect case 6).
720 * The only other cases that gets here are
721 * case 1, case 7 and case 8.
723 if (next
== expand
) {
725 * The only case where we don't expand "vma"
726 * and we expand "next" instead is case 8.
728 VM_WARN_ON(end
!= next
->vm_end
);
730 * remove_next == 3 means we're
731 * removing "vma" and that to do so we
732 * swapped "vma" and "next".
735 VM_WARN_ON(file
!= next
->vm_file
);
738 VM_WARN_ON(expand
!= vma
);
740 * case 1, 6, 7, remove_next == 2 is case 6,
741 * remove_next == 1 is case 1 or 7.
743 remove_next
= 1 + (end
> next
->vm_end
);
744 VM_WARN_ON(remove_next
== 2 &&
745 end
!= next
->vm_next
->vm_end
);
746 /* trim end to next, for case 6 first pass */
754 * If next doesn't have anon_vma, import from vma after
755 * next, if the vma overlaps with it.
757 if (remove_next
== 2 && !next
->anon_vma
)
758 exporter
= next
->vm_next
;
760 } else if (end
> next
->vm_start
) {
762 * vma expands, overlapping part of the next:
763 * mprotect case 5 shifting the boundary up.
765 adjust_next
= (end
- next
->vm_start
);
768 VM_WARN_ON(expand
!= importer
);
769 } else if (end
< vma
->vm_end
) {
771 * vma shrinks, and !insert tells it's not
772 * split_vma inserting another: so it must be
773 * mprotect case 4 shifting the boundary down.
775 adjust_next
= -(vma
->vm_end
- end
);
778 VM_WARN_ON(expand
!= importer
);
782 * Easily overlooked: when mprotect shifts the boundary,
783 * make sure the expanding vma has anon_vma set if the
784 * shrinking vma had, to cover any anon pages imported.
786 if (exporter
&& exporter
->anon_vma
&& !importer
->anon_vma
) {
789 importer
->anon_vma
= exporter
->anon_vma
;
790 error
= anon_vma_clone(importer
, exporter
);
796 vma_adjust_trans_huge(orig_vma
, start
, end
, adjust_next
);
799 mapping
= file
->f_mapping
;
800 root
= &mapping
->i_mmap
;
801 uprobe_munmap(vma
, vma
->vm_start
, vma
->vm_end
);
804 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
806 i_mmap_lock_write(mapping
);
809 * Put into interval tree now, so instantiated pages
810 * are visible to arm/parisc __flush_dcache_page
811 * throughout; but we cannot insert into address
812 * space until vma start or end is updated.
814 __vma_link_file(insert
);
818 anon_vma
= vma
->anon_vma
;
819 if (!anon_vma
&& adjust_next
)
820 anon_vma
= next
->anon_vma
;
822 VM_WARN_ON(adjust_next
&& next
->anon_vma
&&
823 anon_vma
!= next
->anon_vma
);
824 anon_vma_lock_write(anon_vma
);
825 anon_vma_interval_tree_pre_update_vma(vma
);
827 anon_vma_interval_tree_pre_update_vma(next
);
831 flush_dcache_mmap_lock(mapping
);
832 vma_interval_tree_remove(vma
, root
);
834 vma_interval_tree_remove(next
, root
);
837 if (start
!= vma
->vm_start
) {
838 vma
->vm_start
= start
;
839 start_changed
= true;
841 if (end
!= vma
->vm_end
) {
845 vma
->vm_pgoff
= pgoff
;
847 next
->vm_start
+= adjust_next
;
848 next
->vm_pgoff
+= adjust_next
>> PAGE_SHIFT
;
853 vma_interval_tree_insert(next
, root
);
854 vma_interval_tree_insert(vma
, root
);
855 flush_dcache_mmap_unlock(mapping
);
860 * vma_merge has merged next into vma, and needs
861 * us to remove next before dropping the locks.
863 if (remove_next
!= 3)
864 __vma_unlink(mm
, next
, next
);
867 * vma is not before next if they've been
870 * pre-swap() next->vm_start was reduced so
871 * tell validate_mm_rb to ignore pre-swap()
872 * "next" (which is stored in post-swap()
875 __vma_unlink(mm
, next
, vma
);
877 __remove_shared_vm_struct(next
, file
, mapping
);
880 * split_vma has split insert from vma, and needs
881 * us to insert it before dropping the locks
882 * (it may either follow vma or precede it).
884 __insert_vm_struct(mm
, insert
);
890 mm
->highest_vm_end
= vm_end_gap(vma
);
891 else if (!adjust_next
)
892 vma_gap_update(next
);
897 anon_vma_interval_tree_post_update_vma(vma
);
899 anon_vma_interval_tree_post_update_vma(next
);
900 anon_vma_unlock_write(anon_vma
);
904 i_mmap_unlock_write(mapping
);
913 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
917 anon_vma_merge(vma
, next
);
919 mpol_put(vma_policy(next
));
922 * In mprotect's case 6 (see comments on vma_merge),
923 * we must remove another next too. It would clutter
924 * up the code too much to do both in one go.
926 if (remove_next
!= 3) {
928 * If "next" was removed and vma->vm_end was
929 * expanded (up) over it, in turn
930 * "next->vm_prev->vm_end" changed and the
931 * "vma->vm_next" gap must be updated.
936 * For the scope of the comment "next" and
937 * "vma" considered pre-swap(): if "vma" was
938 * removed, next->vm_start was expanded (down)
939 * over it and the "next" gap must be updated.
940 * Because of the swap() the post-swap() "vma"
941 * actually points to pre-swap() "next"
942 * (post-swap() "next" as opposed is now a
947 if (remove_next
== 2) {
953 vma_gap_update(next
);
956 * If remove_next == 2 we obviously can't
959 * If remove_next == 3 we can't reach this
960 * path because pre-swap() next is always not
961 * NULL. pre-swap() "next" is not being
962 * removed and its next->vm_end is not altered
963 * (and furthermore "end" already matches
964 * next->vm_end in remove_next == 3).
966 * We reach this only in the remove_next == 1
967 * case if the "next" vma that was removed was
968 * the highest vma of the mm. However in such
969 * case next->vm_end == "end" and the extended
970 * "vma" has vma->vm_end == next->vm_end so
971 * mm->highest_vm_end doesn't need any update
972 * in remove_next == 1 case.
974 VM_WARN_ON(mm
->highest_vm_end
!= vm_end_gap(vma
));
986 * If the vma has a ->close operation then the driver probably needs to release
987 * per-vma resources, so we don't attempt to merge those.
989 static inline int is_mergeable_vma(struct vm_area_struct
*vma
,
990 struct file
*file
, unsigned long vm_flags
,
991 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
,
992 struct anon_vma_name
*anon_name
)
995 * VM_SOFTDIRTY should not prevent from VMA merging, if we
996 * match the flags but dirty bit -- the caller should mark
997 * merged VMA as dirty. If dirty bit won't be excluded from
998 * comparison, we increase pressure on the memory system forcing
999 * the kernel to generate new VMAs when old one could be
1002 if ((vma
->vm_flags
^ vm_flags
) & ~VM_SOFTDIRTY
)
1004 if (vma
->vm_file
!= file
)
1006 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
1008 if (!is_mergeable_vm_userfaultfd_ctx(vma
, vm_userfaultfd_ctx
))
1010 if (!anon_vma_name_eq(anon_vma_name(vma
), anon_name
))
1015 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
1016 struct anon_vma
*anon_vma2
,
1017 struct vm_area_struct
*vma
)
1020 * The list_is_singular() test is to avoid merging VMA cloned from
1021 * parents. This can improve scalability caused by anon_vma lock.
1023 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
1024 list_is_singular(&vma
->anon_vma_chain
)))
1026 return anon_vma1
== anon_vma2
;
1030 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1031 * in front of (at a lower virtual address and file offset than) the vma.
1033 * We cannot merge two vmas if they have differently assigned (non-NULL)
1034 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1036 * We don't check here for the merged mmap wrapping around the end of pagecache
1037 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1038 * wrap, nor mmaps which cover the final page at index -1UL.
1041 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
1042 struct anon_vma
*anon_vma
, struct file
*file
,
1044 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
,
1045 struct anon_vma_name
*anon_name
)
1047 if (is_mergeable_vma(vma
, file
, vm_flags
, vm_userfaultfd_ctx
, anon_name
) &&
1048 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
1049 if (vma
->vm_pgoff
== vm_pgoff
)
1056 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1057 * beyond (at a higher virtual address and file offset than) the vma.
1059 * We cannot merge two vmas if they have differently assigned (non-NULL)
1060 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1063 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
1064 struct anon_vma
*anon_vma
, struct file
*file
,
1066 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
,
1067 struct anon_vma_name
*anon_name
)
1069 if (is_mergeable_vma(vma
, file
, vm_flags
, vm_userfaultfd_ctx
, anon_name
) &&
1070 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
1072 vm_pglen
= vma_pages(vma
);
1073 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
1080 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
1081 * figure out whether that can be merged with its predecessor or its
1082 * successor. Or both (it neatly fills a hole).
1084 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1085 * certain not to be mapped by the time vma_merge is called; but when
1086 * called for mprotect, it is certain to be already mapped (either at
1087 * an offset within prev, or at the start of next), and the flags of
1088 * this area are about to be changed to vm_flags - and the no-change
1089 * case has already been eliminated.
1091 * The following mprotect cases have to be considered, where AAAA is
1092 * the area passed down from mprotect_fixup, never extending beyond one
1093 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1096 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
1097 * cannot merge might become might become
1098 * PPNNNNNNNNNN PPPPPPPPPPNN
1099 * mmap, brk or case 4 below case 5 below
1102 * PPPP NNNN PPPPNNNNXXXX
1103 * might become might become
1104 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
1105 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
1106 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
1108 * It is important for case 8 that the vma NNNN overlapping the
1109 * region AAAA is never going to extended over XXXX. Instead XXXX must
1110 * be extended in region AAAA and NNNN must be removed. This way in
1111 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1112 * rmap_locks, the properties of the merged vma will be already
1113 * correct for the whole merged range. Some of those properties like
1114 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1115 * be correct for the whole merged range immediately after the
1116 * rmap_locks are released. Otherwise if XXXX would be removed and
1117 * NNNN would be extended over the XXXX range, remove_migration_ptes
1118 * or other rmap walkers (if working on addresses beyond the "end"
1119 * parameter) may establish ptes with the wrong permissions of NNNN
1120 * instead of the right permissions of XXXX.
1122 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
1123 struct vm_area_struct
*prev
, unsigned long addr
,
1124 unsigned long end
, unsigned long vm_flags
,
1125 struct anon_vma
*anon_vma
, struct file
*file
,
1126 pgoff_t pgoff
, struct mempolicy
*policy
,
1127 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
,
1128 struct anon_vma_name
*anon_name
)
1130 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
1131 struct vm_area_struct
*area
, *next
;
1135 * We later require that vma->vm_flags == vm_flags,
1136 * so this tests vma->vm_flags & VM_SPECIAL, too.
1138 if (vm_flags
& VM_SPECIAL
)
1141 next
= vma_next(mm
, prev
);
1143 if (area
&& area
->vm_end
== end
) /* cases 6, 7, 8 */
1144 next
= next
->vm_next
;
1146 /* verify some invariant that must be enforced by the caller */
1147 VM_WARN_ON(prev
&& addr
<= prev
->vm_start
);
1148 VM_WARN_ON(area
&& end
> area
->vm_end
);
1149 VM_WARN_ON(addr
>= end
);
1152 * Can it merge with the predecessor?
1154 if (prev
&& prev
->vm_end
== addr
&&
1155 mpol_equal(vma_policy(prev
), policy
) &&
1156 can_vma_merge_after(prev
, vm_flags
,
1157 anon_vma
, file
, pgoff
,
1158 vm_userfaultfd_ctx
, anon_name
)) {
1160 * OK, it can. Can we now merge in the successor as well?
1162 if (next
&& end
== next
->vm_start
&&
1163 mpol_equal(policy
, vma_policy(next
)) &&
1164 can_vma_merge_before(next
, vm_flags
,
1167 vm_userfaultfd_ctx
, anon_name
) &&
1168 is_mergeable_anon_vma(prev
->anon_vma
,
1169 next
->anon_vma
, NULL
)) {
1171 err
= __vma_adjust(prev
, prev
->vm_start
,
1172 next
->vm_end
, prev
->vm_pgoff
, NULL
,
1174 } else /* cases 2, 5, 7 */
1175 err
= __vma_adjust(prev
, prev
->vm_start
,
1176 end
, prev
->vm_pgoff
, NULL
, prev
);
1179 khugepaged_enter_vma(prev
, vm_flags
);
1184 * Can this new request be merged in front of next?
1186 if (next
&& end
== next
->vm_start
&&
1187 mpol_equal(policy
, vma_policy(next
)) &&
1188 can_vma_merge_before(next
, vm_flags
,
1189 anon_vma
, file
, pgoff
+pglen
,
1190 vm_userfaultfd_ctx
, anon_name
)) {
1191 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
1192 err
= __vma_adjust(prev
, prev
->vm_start
,
1193 addr
, prev
->vm_pgoff
, NULL
, next
);
1194 else { /* cases 3, 8 */
1195 err
= __vma_adjust(area
, addr
, next
->vm_end
,
1196 next
->vm_pgoff
- pglen
, NULL
, next
);
1198 * In case 3 area is already equal to next and
1199 * this is a noop, but in case 8 "area" has
1200 * been removed and next was expanded over it.
1206 khugepaged_enter_vma(area
, vm_flags
);
1214 * Rough compatibility check to quickly see if it's even worth looking
1215 * at sharing an anon_vma.
1217 * They need to have the same vm_file, and the flags can only differ
1218 * in things that mprotect may change.
1220 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1221 * we can merge the two vma's. For example, we refuse to merge a vma if
1222 * there is a vm_ops->close() function, because that indicates that the
1223 * driver is doing some kind of reference counting. But that doesn't
1224 * really matter for the anon_vma sharing case.
1226 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1228 return a
->vm_end
== b
->vm_start
&&
1229 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
1230 a
->vm_file
== b
->vm_file
&&
1231 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_ACCESS_FLAGS
| VM_SOFTDIRTY
)) &&
1232 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
1236 * Do some basic sanity checking to see if we can re-use the anon_vma
1237 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1238 * the same as 'old', the other will be the new one that is trying
1239 * to share the anon_vma.
1241 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1242 * the anon_vma of 'old' is concurrently in the process of being set up
1243 * by another page fault trying to merge _that_. But that's ok: if it
1244 * is being set up, that automatically means that it will be a singleton
1245 * acceptable for merging, so we can do all of this optimistically. But
1246 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1248 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1249 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1250 * is to return an anon_vma that is "complex" due to having gone through
1253 * We also make sure that the two vma's are compatible (adjacent,
1254 * and with the same memory policies). That's all stable, even with just
1255 * a read lock on the mmap_lock.
1257 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1259 if (anon_vma_compatible(a
, b
)) {
1260 struct anon_vma
*anon_vma
= READ_ONCE(old
->anon_vma
);
1262 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
1269 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1270 * neighbouring vmas for a suitable anon_vma, before it goes off
1271 * to allocate a new anon_vma. It checks because a repetitive
1272 * sequence of mprotects and faults may otherwise lead to distinct
1273 * anon_vmas being allocated, preventing vma merge in subsequent
1276 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
1278 struct anon_vma
*anon_vma
= NULL
;
1280 /* Try next first. */
1282 anon_vma
= reusable_anon_vma(vma
->vm_next
, vma
, vma
->vm_next
);
1287 /* Try prev next. */
1289 anon_vma
= reusable_anon_vma(vma
->vm_prev
, vma
->vm_prev
, vma
);
1292 * We might reach here with anon_vma == NULL if we can't find
1293 * any reusable anon_vma.
1294 * There's no absolute need to look only at touching neighbours:
1295 * we could search further afield for "compatible" anon_vmas.
1296 * But it would probably just be a waste of time searching,
1297 * or lead to too many vmas hanging off the same anon_vma.
1298 * We're trying to allow mprotect remerging later on,
1299 * not trying to minimize memory used for anon_vmas.
1305 * If a hint addr is less than mmap_min_addr change hint to be as
1306 * low as possible but still greater than mmap_min_addr
1308 static inline unsigned long round_hint_to_min(unsigned long hint
)
1311 if (((void *)hint
!= NULL
) &&
1312 (hint
< mmap_min_addr
))
1313 return PAGE_ALIGN(mmap_min_addr
);
1317 int mlock_future_check(struct mm_struct
*mm
, unsigned long flags
,
1320 unsigned long locked
, lock_limit
;
1322 /* mlock MCL_FUTURE? */
1323 if (flags
& VM_LOCKED
) {
1324 locked
= len
>> PAGE_SHIFT
;
1325 locked
+= mm
->locked_vm
;
1326 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1327 lock_limit
>>= PAGE_SHIFT
;
1328 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1334 static inline u64
file_mmap_size_max(struct file
*file
, struct inode
*inode
)
1336 if (S_ISREG(inode
->i_mode
))
1337 return MAX_LFS_FILESIZE
;
1339 if (S_ISBLK(inode
->i_mode
))
1340 return MAX_LFS_FILESIZE
;
1342 if (S_ISSOCK(inode
->i_mode
))
1343 return MAX_LFS_FILESIZE
;
1345 /* Special "we do even unsigned file positions" case */
1346 if (file
->f_mode
& FMODE_UNSIGNED_OFFSET
)
1349 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1353 static inline bool file_mmap_ok(struct file
*file
, struct inode
*inode
,
1354 unsigned long pgoff
, unsigned long len
)
1356 u64 maxsize
= file_mmap_size_max(file
, inode
);
1358 if (maxsize
&& len
> maxsize
)
1361 if (pgoff
> maxsize
>> PAGE_SHIFT
)
1367 * The caller must write-lock current->mm->mmap_lock.
1369 unsigned long do_mmap(struct file
*file
, unsigned long addr
,
1370 unsigned long len
, unsigned long prot
,
1371 unsigned long flags
, unsigned long pgoff
,
1372 unsigned long *populate
, struct list_head
*uf
)
1374 struct mm_struct
*mm
= current
->mm
;
1375 vm_flags_t vm_flags
;
1384 * Does the application expect PROT_READ to imply PROT_EXEC?
1386 * (the exception is when the underlying filesystem is noexec
1387 * mounted, in which case we dont add PROT_EXEC.)
1389 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1390 if (!(file
&& path_noexec(&file
->f_path
)))
1393 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1394 if (flags
& MAP_FIXED_NOREPLACE
)
1397 if (!(flags
& MAP_FIXED
))
1398 addr
= round_hint_to_min(addr
);
1400 /* Careful about overflows.. */
1401 len
= PAGE_ALIGN(len
);
1405 /* offset overflow? */
1406 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1409 /* Too many mappings? */
1410 if (mm
->map_count
> sysctl_max_map_count
)
1413 /* Obtain the address to map to. we verify (or select) it and ensure
1414 * that it represents a valid section of the address space.
1416 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1417 if (IS_ERR_VALUE(addr
))
1420 if (flags
& MAP_FIXED_NOREPLACE
) {
1421 if (find_vma_intersection(mm
, addr
, addr
+ len
))
1425 if (prot
== PROT_EXEC
) {
1426 pkey
= execute_only_pkey(mm
);
1431 /* Do simple checking here so the lower-level routines won't have
1432 * to. we assume access permissions have been handled by the open
1433 * of the memory object, so we don't do any here.
1435 vm_flags
= calc_vm_prot_bits(prot
, pkey
) | calc_vm_flag_bits(flags
) |
1436 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1438 if (flags
& MAP_LOCKED
)
1439 if (!can_do_mlock())
1442 if (mlock_future_check(mm
, vm_flags
, len
))
1446 struct inode
*inode
= file_inode(file
);
1447 unsigned long flags_mask
;
1449 if (!file_mmap_ok(file
, inode
, pgoff
, len
))
1452 flags_mask
= LEGACY_MAP_MASK
| file
->f_op
->mmap_supported_flags
;
1454 switch (flags
& MAP_TYPE
) {
1457 * Force use of MAP_SHARED_VALIDATE with non-legacy
1458 * flags. E.g. MAP_SYNC is dangerous to use with
1459 * MAP_SHARED as you don't know which consistency model
1460 * you will get. We silently ignore unsupported flags
1461 * with MAP_SHARED to preserve backward compatibility.
1463 flags
&= LEGACY_MAP_MASK
;
1465 case MAP_SHARED_VALIDATE
:
1466 if (flags
& ~flags_mask
)
1468 if (prot
& PROT_WRITE
) {
1469 if (!(file
->f_mode
& FMODE_WRITE
))
1471 if (IS_SWAPFILE(file
->f_mapping
->host
))
1476 * Make sure we don't allow writing to an append-only
1479 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1482 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1483 if (!(file
->f_mode
& FMODE_WRITE
))
1484 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1487 if (!(file
->f_mode
& FMODE_READ
))
1489 if (path_noexec(&file
->f_path
)) {
1490 if (vm_flags
& VM_EXEC
)
1492 vm_flags
&= ~VM_MAYEXEC
;
1495 if (!file
->f_op
->mmap
)
1497 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1505 switch (flags
& MAP_TYPE
) {
1507 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1513 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1517 * Set pgoff according to addr for anon_vma.
1519 pgoff
= addr
>> PAGE_SHIFT
;
1527 * Set 'VM_NORESERVE' if we should not account for the
1528 * memory use of this mapping.
1530 if (flags
& MAP_NORESERVE
) {
1531 /* We honor MAP_NORESERVE if allowed to overcommit */
1532 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1533 vm_flags
|= VM_NORESERVE
;
1535 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1536 if (file
&& is_file_hugepages(file
))
1537 vm_flags
|= VM_NORESERVE
;
1540 addr
= mmap_region(file
, addr
, len
, vm_flags
, pgoff
, uf
);
1541 if (!IS_ERR_VALUE(addr
) &&
1542 ((vm_flags
& VM_LOCKED
) ||
1543 (flags
& (MAP_POPULATE
| MAP_NONBLOCK
)) == MAP_POPULATE
))
1548 unsigned long ksys_mmap_pgoff(unsigned long addr
, unsigned long len
,
1549 unsigned long prot
, unsigned long flags
,
1550 unsigned long fd
, unsigned long pgoff
)
1552 struct file
*file
= NULL
;
1553 unsigned long retval
;
1555 if (!(flags
& MAP_ANONYMOUS
)) {
1556 audit_mmap_fd(fd
, flags
);
1560 if (is_file_hugepages(file
)) {
1561 len
= ALIGN(len
, huge_page_size(hstate_file(file
)));
1562 } else if (unlikely(flags
& MAP_HUGETLB
)) {
1566 } else if (flags
& MAP_HUGETLB
) {
1569 hs
= hstate_sizelog((flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1573 len
= ALIGN(len
, huge_page_size(hs
));
1575 * VM_NORESERVE is used because the reservations will be
1576 * taken when vm_ops->mmap() is called
1578 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, len
,
1580 HUGETLB_ANONHUGE_INODE
,
1581 (flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1583 return PTR_ERR(file
);
1586 retval
= vm_mmap_pgoff(file
, addr
, len
, prot
, flags
, pgoff
);
1593 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1594 unsigned long, prot
, unsigned long, flags
,
1595 unsigned long, fd
, unsigned long, pgoff
)
1597 return ksys_mmap_pgoff(addr
, len
, prot
, flags
, fd
, pgoff
);
1600 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1601 struct mmap_arg_struct
{
1605 unsigned long flags
;
1607 unsigned long offset
;
1610 SYSCALL_DEFINE1(old_mmap
, struct mmap_arg_struct __user
*, arg
)
1612 struct mmap_arg_struct a
;
1614 if (copy_from_user(&a
, arg
, sizeof(a
)))
1616 if (offset_in_page(a
.offset
))
1619 return ksys_mmap_pgoff(a
.addr
, a
.len
, a
.prot
, a
.flags
, a
.fd
,
1620 a
.offset
>> PAGE_SHIFT
);
1622 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1625 * Some shared mappings will want the pages marked read-only
1626 * to track write events. If so, we'll downgrade vm_page_prot
1627 * to the private version (using protection_map[] without the
1630 int vma_wants_writenotify(struct vm_area_struct
*vma
, pgprot_t vm_page_prot
)
1632 vm_flags_t vm_flags
= vma
->vm_flags
;
1633 const struct vm_operations_struct
*vm_ops
= vma
->vm_ops
;
1635 /* If it was private or non-writable, the write bit is already clear */
1636 if ((vm_flags
& (VM_WRITE
|VM_SHARED
)) != ((VM_WRITE
|VM_SHARED
)))
1639 /* The backer wishes to know when pages are first written to? */
1640 if (vm_ops
&& (vm_ops
->page_mkwrite
|| vm_ops
->pfn_mkwrite
))
1643 /* The open routine did something to the protections that pgprot_modify
1644 * won't preserve? */
1645 if (pgprot_val(vm_page_prot
) !=
1646 pgprot_val(vm_pgprot_modify(vm_page_prot
, vm_flags
)))
1650 * Do we need to track softdirty? hugetlb does not support softdirty
1653 if (vma_soft_dirty_enabled(vma
) && !is_vm_hugetlb_page(vma
))
1656 /* Specialty mapping? */
1657 if (vm_flags
& VM_PFNMAP
)
1660 /* Can the mapping track the dirty pages? */
1661 return vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
1662 mapping_can_writeback(vma
->vm_file
->f_mapping
);
1666 * We account for memory if it's a private writeable mapping,
1667 * not hugepages and VM_NORESERVE wasn't set.
1669 static inline int accountable_mapping(struct file
*file
, vm_flags_t vm_flags
)
1672 * hugetlb has its own accounting separate from the core VM
1673 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1675 if (file
&& is_file_hugepages(file
))
1678 return (vm_flags
& (VM_NORESERVE
| VM_SHARED
| VM_WRITE
)) == VM_WRITE
;
1681 unsigned long mmap_region(struct file
*file
, unsigned long addr
,
1682 unsigned long len
, vm_flags_t vm_flags
, unsigned long pgoff
,
1683 struct list_head
*uf
)
1685 struct mm_struct
*mm
= current
->mm
;
1686 struct vm_area_struct
*vma
, *prev
, *merge
;
1688 struct rb_node
**rb_link
, *rb_parent
;
1689 unsigned long charged
= 0;
1691 /* Check against address space limit. */
1692 if (!may_expand_vm(mm
, vm_flags
, len
>> PAGE_SHIFT
)) {
1693 unsigned long nr_pages
;
1696 * MAP_FIXED may remove pages of mappings that intersects with
1697 * requested mapping. Account for the pages it would unmap.
1699 nr_pages
= count_vma_pages_range(mm
, addr
, addr
+ len
);
1701 if (!may_expand_vm(mm
, vm_flags
,
1702 (len
>> PAGE_SHIFT
) - nr_pages
))
1706 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1707 if (munmap_vma_range(mm
, addr
, len
, &prev
, &rb_link
, &rb_parent
, uf
))
1710 * Private writable mapping: check memory availability
1712 if (accountable_mapping(file
, vm_flags
)) {
1713 charged
= len
>> PAGE_SHIFT
;
1714 if (security_vm_enough_memory_mm(mm
, charged
))
1716 vm_flags
|= VM_ACCOUNT
;
1720 * Can we just expand an old mapping?
1722 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
,
1723 NULL
, file
, pgoff
, NULL
, NULL_VM_UFFD_CTX
, NULL
);
1728 * Determine the object being mapped and call the appropriate
1729 * specific mapper. the address has already been validated, but
1730 * not unmapped, but the maps are removed from the list.
1732 vma
= vm_area_alloc(mm
);
1738 vma
->vm_start
= addr
;
1739 vma
->vm_end
= addr
+ len
;
1740 vma
->vm_flags
= vm_flags
;
1741 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1742 vma
->vm_pgoff
= pgoff
;
1745 if (vm_flags
& VM_SHARED
) {
1746 error
= mapping_map_writable(file
->f_mapping
);
1751 vma
->vm_file
= get_file(file
);
1752 error
= call_mmap(file
, vma
);
1754 goto unmap_and_free_vma
;
1756 /* Can addr have changed??
1758 * Answer: Yes, several device drivers can do it in their
1759 * f_op->mmap method. -DaveM
1760 * Bug: If addr is changed, prev, rb_link, rb_parent should
1761 * be updated for vma_link()
1763 WARN_ON_ONCE(addr
!= vma
->vm_start
);
1765 addr
= vma
->vm_start
;
1767 /* If vm_flags changed after call_mmap(), we should try merge vma again
1768 * as we may succeed this time.
1770 if (unlikely(vm_flags
!= vma
->vm_flags
&& prev
)) {
1771 merge
= vma_merge(mm
, prev
, vma
->vm_start
, vma
->vm_end
, vma
->vm_flags
,
1772 NULL
, vma
->vm_file
, vma
->vm_pgoff
, NULL
, NULL_VM_UFFD_CTX
, NULL
);
1774 /* ->mmap() can change vma->vm_file and fput the original file. So
1775 * fput the vma->vm_file here or we would add an extra fput for file
1776 * and cause general protection fault ultimately.
1781 /* Update vm_flags to pick up the change. */
1782 vm_flags
= vma
->vm_flags
;
1783 goto unmap_writable
;
1787 vm_flags
= vma
->vm_flags
;
1788 } else if (vm_flags
& VM_SHARED
) {
1789 error
= shmem_zero_setup(vma
);
1793 vma_set_anonymous(vma
);
1796 /* Allow architectures to sanity-check the vm_flags */
1797 if (!arch_validate_flags(vma
->vm_flags
)) {
1800 goto unmap_and_free_vma
;
1805 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1808 * vma_merge() calls khugepaged_enter_vma() either, the below
1809 * call covers the non-merge case.
1811 khugepaged_enter_vma(vma
, vma
->vm_flags
);
1813 /* Once vma denies write, undo our temporary denial count */
1815 if (file
&& vm_flags
& VM_SHARED
)
1816 mapping_unmap_writable(file
->f_mapping
);
1817 file
= vma
->vm_file
;
1819 perf_event_mmap(vma
);
1821 vm_stat_account(mm
, vm_flags
, len
>> PAGE_SHIFT
);
1822 if (vm_flags
& VM_LOCKED
) {
1823 if ((vm_flags
& VM_SPECIAL
) || vma_is_dax(vma
) ||
1824 is_vm_hugetlb_page(vma
) ||
1825 vma
== get_gate_vma(current
->mm
))
1826 vma
->vm_flags
&= VM_LOCKED_CLEAR_MASK
;
1828 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1835 * New (or expanded) vma always get soft dirty status.
1836 * Otherwise user-space soft-dirty page tracker won't
1837 * be able to distinguish situation when vma area unmapped,
1838 * then new mapped in-place (which must be aimed as
1839 * a completely new data area).
1841 vma
->vm_flags
|= VM_SOFTDIRTY
;
1843 vma_set_page_prot(vma
);
1849 vma
->vm_file
= NULL
;
1851 /* Undo any partial mapping done by a device driver. */
1852 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1853 if (vm_flags
& VM_SHARED
)
1854 mapping_unmap_writable(file
->f_mapping
);
1859 vm_unacct_memory(charged
);
1863 static unsigned long unmapped_area(struct vm_unmapped_area_info
*info
)
1866 * We implement the search by looking for an rbtree node that
1867 * immediately follows a suitable gap. That is,
1868 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1869 * - gap_end = vma->vm_start >= info->low_limit + length;
1870 * - gap_end - gap_start >= length
1873 struct mm_struct
*mm
= current
->mm
;
1874 struct vm_area_struct
*vma
;
1875 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1877 /* Adjust search length to account for worst case alignment overhead */
1878 length
= info
->length
+ info
->align_mask
;
1879 if (length
< info
->length
)
1882 /* Adjust search limits by the desired length */
1883 if (info
->high_limit
< length
)
1885 high_limit
= info
->high_limit
- length
;
1887 if (info
->low_limit
> high_limit
)
1889 low_limit
= info
->low_limit
+ length
;
1891 /* Check if rbtree root looks promising */
1892 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1894 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1895 if (vma
->rb_subtree_gap
< length
)
1899 /* Visit left subtree if it looks promising */
1900 gap_end
= vm_start_gap(vma
);
1901 if (gap_end
>= low_limit
&& vma
->vm_rb
.rb_left
) {
1902 struct vm_area_struct
*left
=
1903 rb_entry(vma
->vm_rb
.rb_left
,
1904 struct vm_area_struct
, vm_rb
);
1905 if (left
->rb_subtree_gap
>= length
) {
1911 gap_start
= vma
->vm_prev
? vm_end_gap(vma
->vm_prev
) : 0;
1913 /* Check if current node has a suitable gap */
1914 if (gap_start
> high_limit
)
1916 if (gap_end
>= low_limit
&&
1917 gap_end
> gap_start
&& gap_end
- gap_start
>= length
)
1920 /* Visit right subtree if it looks promising */
1921 if (vma
->vm_rb
.rb_right
) {
1922 struct vm_area_struct
*right
=
1923 rb_entry(vma
->vm_rb
.rb_right
,
1924 struct vm_area_struct
, vm_rb
);
1925 if (right
->rb_subtree_gap
>= length
) {
1931 /* Go back up the rbtree to find next candidate node */
1933 struct rb_node
*prev
= &vma
->vm_rb
;
1934 if (!rb_parent(prev
))
1936 vma
= rb_entry(rb_parent(prev
),
1937 struct vm_area_struct
, vm_rb
);
1938 if (prev
== vma
->vm_rb
.rb_left
) {
1939 gap_start
= vm_end_gap(vma
->vm_prev
);
1940 gap_end
= vm_start_gap(vma
);
1947 /* Check highest gap, which does not precede any rbtree node */
1948 gap_start
= mm
->highest_vm_end
;
1949 gap_end
= ULONG_MAX
; /* Only for VM_BUG_ON below */
1950 if (gap_start
> high_limit
)
1954 /* We found a suitable gap. Clip it with the original low_limit. */
1955 if (gap_start
< info
->low_limit
)
1956 gap_start
= info
->low_limit
;
1958 /* Adjust gap address to the desired alignment */
1959 gap_start
+= (info
->align_offset
- gap_start
) & info
->align_mask
;
1961 VM_BUG_ON(gap_start
+ info
->length
> info
->high_limit
);
1962 VM_BUG_ON(gap_start
+ info
->length
> gap_end
);
1966 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info
*info
)
1968 struct mm_struct
*mm
= current
->mm
;
1969 struct vm_area_struct
*vma
;
1970 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1972 /* Adjust search length to account for worst case alignment overhead */
1973 length
= info
->length
+ info
->align_mask
;
1974 if (length
< info
->length
)
1978 * Adjust search limits by the desired length.
1979 * See implementation comment at top of unmapped_area().
1981 gap_end
= info
->high_limit
;
1982 if (gap_end
< length
)
1984 high_limit
= gap_end
- length
;
1986 if (info
->low_limit
> high_limit
)
1988 low_limit
= info
->low_limit
+ length
;
1990 /* Check highest gap, which does not precede any rbtree node */
1991 gap_start
= mm
->highest_vm_end
;
1992 if (gap_start
<= high_limit
)
1995 /* Check if rbtree root looks promising */
1996 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1998 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1999 if (vma
->rb_subtree_gap
< length
)
2003 /* Visit right subtree if it looks promising */
2004 gap_start
= vma
->vm_prev
? vm_end_gap(vma
->vm_prev
) : 0;
2005 if (gap_start
<= high_limit
&& vma
->vm_rb
.rb_right
) {
2006 struct vm_area_struct
*right
=
2007 rb_entry(vma
->vm_rb
.rb_right
,
2008 struct vm_area_struct
, vm_rb
);
2009 if (right
->rb_subtree_gap
>= length
) {
2016 /* Check if current node has a suitable gap */
2017 gap_end
= vm_start_gap(vma
);
2018 if (gap_end
< low_limit
)
2020 if (gap_start
<= high_limit
&&
2021 gap_end
> gap_start
&& gap_end
- gap_start
>= length
)
2024 /* Visit left subtree if it looks promising */
2025 if (vma
->vm_rb
.rb_left
) {
2026 struct vm_area_struct
*left
=
2027 rb_entry(vma
->vm_rb
.rb_left
,
2028 struct vm_area_struct
, vm_rb
);
2029 if (left
->rb_subtree_gap
>= length
) {
2035 /* Go back up the rbtree to find next candidate node */
2037 struct rb_node
*prev
= &vma
->vm_rb
;
2038 if (!rb_parent(prev
))
2040 vma
= rb_entry(rb_parent(prev
),
2041 struct vm_area_struct
, vm_rb
);
2042 if (prev
== vma
->vm_rb
.rb_right
) {
2043 gap_start
= vma
->vm_prev
?
2044 vm_end_gap(vma
->vm_prev
) : 0;
2051 /* We found a suitable gap. Clip it with the original high_limit. */
2052 if (gap_end
> info
->high_limit
)
2053 gap_end
= info
->high_limit
;
2056 /* Compute highest gap address at the desired alignment */
2057 gap_end
-= info
->length
;
2058 gap_end
-= (gap_end
- info
->align_offset
) & info
->align_mask
;
2060 VM_BUG_ON(gap_end
< info
->low_limit
);
2061 VM_BUG_ON(gap_end
< gap_start
);
2066 * Search for an unmapped address range.
2068 * We are looking for a range that:
2069 * - does not intersect with any VMA;
2070 * - is contained within the [low_limit, high_limit) interval;
2071 * - is at least the desired size.
2072 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2074 unsigned long vm_unmapped_area(struct vm_unmapped_area_info
*info
)
2078 if (info
->flags
& VM_UNMAPPED_AREA_TOPDOWN
)
2079 addr
= unmapped_area_topdown(info
);
2081 addr
= unmapped_area(info
);
2083 trace_vm_unmapped_area(addr
, info
);
2087 /* Get an address range which is currently unmapped.
2088 * For shmat() with addr=0.
2090 * Ugly calling convention alert:
2091 * Return value with the low bits set means error value,
2093 * if (ret & ~PAGE_MASK)
2096 * This function "knows" that -ENOMEM has the bits set.
2099 generic_get_unmapped_area(struct file
*filp
, unsigned long addr
,
2100 unsigned long len
, unsigned long pgoff
,
2101 unsigned long flags
)
2103 struct mm_struct
*mm
= current
->mm
;
2104 struct vm_area_struct
*vma
, *prev
;
2105 struct vm_unmapped_area_info info
;
2106 const unsigned long mmap_end
= arch_get_mmap_end(addr
, len
, flags
);
2108 if (len
> mmap_end
- mmap_min_addr
)
2111 if (flags
& MAP_FIXED
)
2115 addr
= PAGE_ALIGN(addr
);
2116 vma
= find_vma_prev(mm
, addr
, &prev
);
2117 if (mmap_end
- len
>= addr
&& addr
>= mmap_min_addr
&&
2118 (!vma
|| addr
+ len
<= vm_start_gap(vma
)) &&
2119 (!prev
|| addr
>= vm_end_gap(prev
)))
2125 info
.low_limit
= mm
->mmap_base
;
2126 info
.high_limit
= mmap_end
;
2127 info
.align_mask
= 0;
2128 info
.align_offset
= 0;
2129 return vm_unmapped_area(&info
);
2132 #ifndef HAVE_ARCH_UNMAPPED_AREA
2134 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
2135 unsigned long len
, unsigned long pgoff
,
2136 unsigned long flags
)
2138 return generic_get_unmapped_area(filp
, addr
, len
, pgoff
, flags
);
2143 * This mmap-allocator allocates new areas top-down from below the
2144 * stack's low limit (the base):
2147 generic_get_unmapped_area_topdown(struct file
*filp
, unsigned long addr
,
2148 unsigned long len
, unsigned long pgoff
,
2149 unsigned long flags
)
2151 struct vm_area_struct
*vma
, *prev
;
2152 struct mm_struct
*mm
= current
->mm
;
2153 struct vm_unmapped_area_info info
;
2154 const unsigned long mmap_end
= arch_get_mmap_end(addr
, len
, flags
);
2156 /* requested length too big for entire address space */
2157 if (len
> mmap_end
- mmap_min_addr
)
2160 if (flags
& MAP_FIXED
)
2163 /* requesting a specific address */
2165 addr
= PAGE_ALIGN(addr
);
2166 vma
= find_vma_prev(mm
, addr
, &prev
);
2167 if (mmap_end
- len
>= addr
&& addr
>= mmap_min_addr
&&
2168 (!vma
|| addr
+ len
<= vm_start_gap(vma
)) &&
2169 (!prev
|| addr
>= vm_end_gap(prev
)))
2173 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
2175 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
2176 info
.high_limit
= arch_get_mmap_base(addr
, mm
->mmap_base
);
2177 info
.align_mask
= 0;
2178 info
.align_offset
= 0;
2179 addr
= vm_unmapped_area(&info
);
2182 * A failed mmap() very likely causes application failure,
2183 * so fall back to the bottom-up function here. This scenario
2184 * can happen with large stack limits and large mmap()
2187 if (offset_in_page(addr
)) {
2188 VM_BUG_ON(addr
!= -ENOMEM
);
2190 info
.low_limit
= TASK_UNMAPPED_BASE
;
2191 info
.high_limit
= mmap_end
;
2192 addr
= vm_unmapped_area(&info
);
2198 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2200 arch_get_unmapped_area_topdown(struct file
*filp
, unsigned long addr
,
2201 unsigned long len
, unsigned long pgoff
,
2202 unsigned long flags
)
2204 return generic_get_unmapped_area_topdown(filp
, addr
, len
, pgoff
, flags
);
2209 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
2210 unsigned long pgoff
, unsigned long flags
)
2212 unsigned long (*get_area
)(struct file
*, unsigned long,
2213 unsigned long, unsigned long, unsigned long);
2215 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
2219 /* Careful about overflows.. */
2220 if (len
> TASK_SIZE
)
2223 get_area
= current
->mm
->get_unmapped_area
;
2225 if (file
->f_op
->get_unmapped_area
)
2226 get_area
= file
->f_op
->get_unmapped_area
;
2227 } else if (flags
& MAP_SHARED
) {
2229 * mmap_region() will call shmem_zero_setup() to create a file,
2230 * so use shmem's get_unmapped_area in case it can be huge.
2231 * do_mmap() will clear pgoff, so match alignment.
2234 get_area
= shmem_get_unmapped_area
;
2237 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
2238 if (IS_ERR_VALUE(addr
))
2241 if (addr
> TASK_SIZE
- len
)
2243 if (offset_in_page(addr
))
2246 error
= security_mmap_addr(addr
);
2247 return error
? error
: addr
;
2250 EXPORT_SYMBOL(get_unmapped_area
);
2252 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2253 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
2255 struct rb_node
*rb_node
;
2256 struct vm_area_struct
*vma
;
2258 mmap_assert_locked(mm
);
2259 /* Check the cache first. */
2260 vma
= vmacache_find(mm
, addr
);
2264 rb_node
= mm
->mm_rb
.rb_node
;
2267 struct vm_area_struct
*tmp
;
2269 tmp
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2271 if (tmp
->vm_end
> addr
) {
2273 if (tmp
->vm_start
<= addr
)
2275 rb_node
= rb_node
->rb_left
;
2277 rb_node
= rb_node
->rb_right
;
2281 vmacache_update(addr
, vma
);
2285 EXPORT_SYMBOL(find_vma
);
2288 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2290 struct vm_area_struct
*
2291 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
2292 struct vm_area_struct
**pprev
)
2294 struct vm_area_struct
*vma
;
2296 vma
= find_vma(mm
, addr
);
2298 *pprev
= vma
->vm_prev
;
2300 struct rb_node
*rb_node
= rb_last(&mm
->mm_rb
);
2302 *pprev
= rb_node
? rb_entry(rb_node
, struct vm_area_struct
, vm_rb
) : NULL
;
2308 * Verify that the stack growth is acceptable and
2309 * update accounting. This is shared with both the
2310 * grow-up and grow-down cases.
2312 static int acct_stack_growth(struct vm_area_struct
*vma
,
2313 unsigned long size
, unsigned long grow
)
2315 struct mm_struct
*mm
= vma
->vm_mm
;
2316 unsigned long new_start
;
2318 /* address space limit tests */
2319 if (!may_expand_vm(mm
, vma
->vm_flags
, grow
))
2322 /* Stack limit test */
2323 if (size
> rlimit(RLIMIT_STACK
))
2326 /* mlock limit tests */
2327 if (mlock_future_check(mm
, vma
->vm_flags
, grow
<< PAGE_SHIFT
))
2330 /* Check to ensure the stack will not grow into a hugetlb-only region */
2331 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
2333 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
2337 * Overcommit.. This must be the final test, as it will
2338 * update security statistics.
2340 if (security_vm_enough_memory_mm(mm
, grow
))
2346 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2348 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2349 * vma is the last one with address > vma->vm_end. Have to extend vma.
2351 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
)
2353 struct mm_struct
*mm
= vma
->vm_mm
;
2354 struct vm_area_struct
*next
;
2355 unsigned long gap_addr
;
2358 if (!(vma
->vm_flags
& VM_GROWSUP
))
2361 /* Guard against exceeding limits of the address space. */
2362 address
&= PAGE_MASK
;
2363 if (address
>= (TASK_SIZE
& PAGE_MASK
))
2365 address
+= PAGE_SIZE
;
2367 /* Enforce stack_guard_gap */
2368 gap_addr
= address
+ stack_guard_gap
;
2370 /* Guard against overflow */
2371 if (gap_addr
< address
|| gap_addr
> TASK_SIZE
)
2372 gap_addr
= TASK_SIZE
;
2374 next
= vma
->vm_next
;
2375 if (next
&& next
->vm_start
< gap_addr
&& vma_is_accessible(next
)) {
2376 if (!(next
->vm_flags
& VM_GROWSUP
))
2378 /* Check that both stack segments have the same anon_vma? */
2381 /* We must make sure the anon_vma is allocated. */
2382 if (unlikely(anon_vma_prepare(vma
)))
2386 * vma->vm_start/vm_end cannot change under us because the caller
2387 * is required to hold the mmap_lock in read mode. We need the
2388 * anon_vma lock to serialize against concurrent expand_stacks.
2390 anon_vma_lock_write(vma
->anon_vma
);
2392 /* Somebody else might have raced and expanded it already */
2393 if (address
> vma
->vm_end
) {
2394 unsigned long size
, grow
;
2396 size
= address
- vma
->vm_start
;
2397 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
2400 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
2401 error
= acct_stack_growth(vma
, size
, grow
);
2404 * vma_gap_update() doesn't support concurrent
2405 * updates, but we only hold a shared mmap_lock
2406 * lock here, so we need to protect against
2407 * concurrent vma expansions.
2408 * anon_vma_lock_write() doesn't help here, as
2409 * we don't guarantee that all growable vmas
2410 * in a mm share the same root anon vma.
2411 * So, we reuse mm->page_table_lock to guard
2412 * against concurrent vma expansions.
2414 spin_lock(&mm
->page_table_lock
);
2415 if (vma
->vm_flags
& VM_LOCKED
)
2416 mm
->locked_vm
+= grow
;
2417 vm_stat_account(mm
, vma
->vm_flags
, grow
);
2418 anon_vma_interval_tree_pre_update_vma(vma
);
2419 vma
->vm_end
= address
;
2420 anon_vma_interval_tree_post_update_vma(vma
);
2422 vma_gap_update(vma
->vm_next
);
2424 mm
->highest_vm_end
= vm_end_gap(vma
);
2425 spin_unlock(&mm
->page_table_lock
);
2427 perf_event_mmap(vma
);
2431 anon_vma_unlock_write(vma
->anon_vma
);
2432 khugepaged_enter_vma(vma
, vma
->vm_flags
);
2436 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2439 * vma is the first one with address < vma->vm_start. Have to extend vma.
2441 int expand_downwards(struct vm_area_struct
*vma
,
2442 unsigned long address
)
2444 struct mm_struct
*mm
= vma
->vm_mm
;
2445 struct vm_area_struct
*prev
;
2448 address
&= PAGE_MASK
;
2449 if (address
< mmap_min_addr
)
2452 /* Enforce stack_guard_gap */
2453 prev
= vma
->vm_prev
;
2454 /* Check that both stack segments have the same anon_vma? */
2455 if (prev
&& !(prev
->vm_flags
& VM_GROWSDOWN
) &&
2456 vma_is_accessible(prev
)) {
2457 if (address
- prev
->vm_end
< stack_guard_gap
)
2461 /* We must make sure the anon_vma is allocated. */
2462 if (unlikely(anon_vma_prepare(vma
)))
2466 * vma->vm_start/vm_end cannot change under us because the caller
2467 * is required to hold the mmap_lock in read mode. We need the
2468 * anon_vma lock to serialize against concurrent expand_stacks.
2470 anon_vma_lock_write(vma
->anon_vma
);
2472 /* Somebody else might have raced and expanded it already */
2473 if (address
< vma
->vm_start
) {
2474 unsigned long size
, grow
;
2476 size
= vma
->vm_end
- address
;
2477 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
2480 if (grow
<= vma
->vm_pgoff
) {
2481 error
= acct_stack_growth(vma
, size
, grow
);
2484 * vma_gap_update() doesn't support concurrent
2485 * updates, but we only hold a shared mmap_lock
2486 * lock here, so we need to protect against
2487 * concurrent vma expansions.
2488 * anon_vma_lock_write() doesn't help here, as
2489 * we don't guarantee that all growable vmas
2490 * in a mm share the same root anon vma.
2491 * So, we reuse mm->page_table_lock to guard
2492 * against concurrent vma expansions.
2494 spin_lock(&mm
->page_table_lock
);
2495 if (vma
->vm_flags
& VM_LOCKED
)
2496 mm
->locked_vm
+= grow
;
2497 vm_stat_account(mm
, vma
->vm_flags
, grow
);
2498 anon_vma_interval_tree_pre_update_vma(vma
);
2499 vma
->vm_start
= address
;
2500 vma
->vm_pgoff
-= grow
;
2501 anon_vma_interval_tree_post_update_vma(vma
);
2502 vma_gap_update(vma
);
2503 spin_unlock(&mm
->page_table_lock
);
2505 perf_event_mmap(vma
);
2509 anon_vma_unlock_write(vma
->anon_vma
);
2510 khugepaged_enter_vma(vma
, vma
->vm_flags
);
2515 /* enforced gap between the expanding stack and other mappings. */
2516 unsigned long stack_guard_gap
= 256UL<<PAGE_SHIFT
;
2518 static int __init
cmdline_parse_stack_guard_gap(char *p
)
2523 val
= simple_strtoul(p
, &endptr
, 10);
2525 stack_guard_gap
= val
<< PAGE_SHIFT
;
2529 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap
);
2531 #ifdef CONFIG_STACK_GROWSUP
2532 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2534 return expand_upwards(vma
, address
);
2537 struct vm_area_struct
*
2538 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2540 struct vm_area_struct
*vma
, *prev
;
2543 vma
= find_vma_prev(mm
, addr
, &prev
);
2544 if (vma
&& (vma
->vm_start
<= addr
))
2546 if (!prev
|| expand_stack(prev
, addr
))
2548 if (prev
->vm_flags
& VM_LOCKED
)
2549 populate_vma_page_range(prev
, addr
, prev
->vm_end
, NULL
);
2553 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2555 return expand_downwards(vma
, address
);
2558 struct vm_area_struct
*
2559 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2561 struct vm_area_struct
*vma
;
2562 unsigned long start
;
2565 vma
= find_vma(mm
, addr
);
2568 if (vma
->vm_start
<= addr
)
2570 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2572 start
= vma
->vm_start
;
2573 if (expand_stack(vma
, addr
))
2575 if (vma
->vm_flags
& VM_LOCKED
)
2576 populate_vma_page_range(vma
, addr
, start
, NULL
);
2581 EXPORT_SYMBOL_GPL(find_extend_vma
);
2584 * Ok - we have the memory areas we should free on the vma list,
2585 * so release them, and do the vma updates.
2587 * Called with the mm semaphore held.
2589 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2591 unsigned long nr_accounted
= 0;
2593 /* Update high watermark before we lower total_vm */
2594 update_hiwater_vm(mm
);
2596 long nrpages
= vma_pages(vma
);
2598 if (vma
->vm_flags
& VM_ACCOUNT
)
2599 nr_accounted
+= nrpages
;
2600 vm_stat_account(mm
, vma
->vm_flags
, -nrpages
);
2601 vma
= remove_vma(vma
);
2603 vm_unacct_memory(nr_accounted
);
2608 * Get rid of page table information in the indicated region.
2610 * Called with the mm semaphore held.
2612 static void unmap_region(struct mm_struct
*mm
,
2613 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
2614 unsigned long start
, unsigned long end
)
2616 struct vm_area_struct
*next
= vma_next(mm
, prev
);
2617 struct mmu_gather tlb
;
2620 tlb_gather_mmu(&tlb
, mm
);
2621 update_hiwater_rss(mm
);
2622 unmap_vmas(&tlb
, vma
, start
, end
);
2623 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
2624 next
? next
->vm_start
: USER_PGTABLES_CEILING
);
2625 tlb_finish_mmu(&tlb
);
2629 * Create a list of vma's touched by the unmap, removing them from the mm's
2630 * vma list as we go..
2633 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2634 struct vm_area_struct
*prev
, unsigned long end
)
2636 struct vm_area_struct
**insertion_point
;
2637 struct vm_area_struct
*tail_vma
= NULL
;
2639 insertion_point
= (prev
? &prev
->vm_next
: &mm
->mmap
);
2640 vma
->vm_prev
= NULL
;
2642 vma_rb_erase(vma
, &mm
->mm_rb
);
2643 if (vma
->vm_flags
& VM_LOCKED
)
2644 mm
->locked_vm
-= vma_pages(vma
);
2648 } while (vma
&& vma
->vm_start
< end
);
2649 *insertion_point
= vma
;
2651 vma
->vm_prev
= prev
;
2652 vma_gap_update(vma
);
2654 mm
->highest_vm_end
= prev
? vm_end_gap(prev
) : 0;
2655 tail_vma
->vm_next
= NULL
;
2657 /* Kill the cache */
2658 vmacache_invalidate(mm
);
2661 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2662 * VM_GROWSUP VMA. Such VMAs can change their size under
2663 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2665 if (vma
&& (vma
->vm_flags
& VM_GROWSDOWN
))
2667 if (prev
&& (prev
->vm_flags
& VM_GROWSUP
))
2673 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2674 * has already been checked or doesn't make sense to fail.
2676 int __split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2677 unsigned long addr
, int new_below
)
2679 struct vm_area_struct
*new;
2682 if (vma
->vm_ops
&& vma
->vm_ops
->may_split
) {
2683 err
= vma
->vm_ops
->may_split(vma
, addr
);
2688 new = vm_area_dup(vma
);
2695 new->vm_start
= addr
;
2696 new->vm_pgoff
+= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
);
2699 err
= vma_dup_policy(vma
, new);
2703 err
= anon_vma_clone(new, vma
);
2708 get_file(new->vm_file
);
2710 if (new->vm_ops
&& new->vm_ops
->open
)
2711 new->vm_ops
->open(new);
2714 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2715 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2717 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2723 /* Clean everything up if vma_adjust failed. */
2724 if (new->vm_ops
&& new->vm_ops
->close
)
2725 new->vm_ops
->close(new);
2728 unlink_anon_vmas(new);
2730 mpol_put(vma_policy(new));
2737 * Split a vma into two pieces at address 'addr', a new vma is allocated
2738 * either for the first part or the tail.
2740 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2741 unsigned long addr
, int new_below
)
2743 if (mm
->map_count
>= sysctl_max_map_count
)
2746 return __split_vma(mm
, vma
, addr
, new_below
);
2749 /* Munmap is split into 2 main parts -- this part which finds
2750 * what needs doing, and the areas themselves, which do the
2751 * work. This now handles partial unmappings.
2752 * Jeremy Fitzhardinge <jeremy@goop.org>
2754 int __do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
,
2755 struct list_head
*uf
, bool downgrade
)
2758 struct vm_area_struct
*vma
, *prev
, *last
;
2760 if ((offset_in_page(start
)) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2763 len
= PAGE_ALIGN(len
);
2769 * arch_unmap() might do unmaps itself. It must be called
2770 * and finish any rbtree manipulation before this code
2771 * runs and also starts to manipulate the rbtree.
2773 arch_unmap(mm
, start
, end
);
2775 /* Find the first overlapping VMA where start < vma->vm_end */
2776 vma
= find_vma_intersection(mm
, start
, end
);
2779 prev
= vma
->vm_prev
;
2782 * If we need to split any vma, do it now to save pain later.
2784 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2785 * unmapped vm_area_struct will remain in use: so lower split_vma
2786 * places tmp vma above, and higher split_vma places tmp vma below.
2788 if (start
> vma
->vm_start
) {
2792 * Make sure that map_count on return from munmap() will
2793 * not exceed its limit; but let map_count go just above
2794 * its limit temporarily, to help free resources as expected.
2796 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2799 error
= __split_vma(mm
, vma
, start
, 0);
2805 /* Does it split the last one? */
2806 last
= find_vma(mm
, end
);
2807 if (last
&& end
> last
->vm_start
) {
2808 int error
= __split_vma(mm
, last
, end
, 1);
2812 vma
= vma_next(mm
, prev
);
2816 * If userfaultfd_unmap_prep returns an error the vmas
2817 * will remain split, but userland will get a
2818 * highly unexpected error anyway. This is no
2819 * different than the case where the first of the two
2820 * __split_vma fails, but we don't undo the first
2821 * split, despite we could. This is unlikely enough
2822 * failure that it's not worth optimizing it for.
2824 int error
= userfaultfd_unmap_prep(vma
, start
, end
, uf
);
2829 /* Detach vmas from rbtree */
2830 if (!detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
))
2834 mmap_write_downgrade(mm
);
2836 unmap_region(mm
, vma
, prev
, start
, end
);
2838 /* Fix up all other VM information */
2839 remove_vma_list(mm
, vma
);
2841 return downgrade
? 1 : 0;
2844 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
,
2845 struct list_head
*uf
)
2847 return __do_munmap(mm
, start
, len
, uf
, false);
2850 static int __vm_munmap(unsigned long start
, size_t len
, bool downgrade
)
2853 struct mm_struct
*mm
= current
->mm
;
2856 if (mmap_write_lock_killable(mm
))
2859 ret
= __do_munmap(mm
, start
, len
, &uf
, downgrade
);
2861 * Returning 1 indicates mmap_lock is downgraded.
2862 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2863 * it to 0 before return.
2866 mmap_read_unlock(mm
);
2869 mmap_write_unlock(mm
);
2871 userfaultfd_unmap_complete(mm
, &uf
);
2875 int vm_munmap(unsigned long start
, size_t len
)
2877 return __vm_munmap(start
, len
, false);
2879 EXPORT_SYMBOL(vm_munmap
);
2881 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2883 addr
= untagged_addr(addr
);
2884 return __vm_munmap(addr
, len
, true);
2889 * Emulation of deprecated remap_file_pages() syscall.
2891 SYSCALL_DEFINE5(remap_file_pages
, unsigned long, start
, unsigned long, size
,
2892 unsigned long, prot
, unsigned long, pgoff
, unsigned long, flags
)
2895 struct mm_struct
*mm
= current
->mm
;
2896 struct vm_area_struct
*vma
;
2897 unsigned long populate
= 0;
2898 unsigned long ret
= -EINVAL
;
2901 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
2902 current
->comm
, current
->pid
);
2906 start
= start
& PAGE_MASK
;
2907 size
= size
& PAGE_MASK
;
2909 if (start
+ size
<= start
)
2912 /* Does pgoff wrap? */
2913 if (pgoff
+ (size
>> PAGE_SHIFT
) < pgoff
)
2916 if (mmap_write_lock_killable(mm
))
2919 vma
= vma_lookup(mm
, start
);
2921 if (!vma
|| !(vma
->vm_flags
& VM_SHARED
))
2924 if (start
+ size
> vma
->vm_end
) {
2925 struct vm_area_struct
*next
;
2927 for (next
= vma
->vm_next
; next
; next
= next
->vm_next
) {
2928 /* hole between vmas ? */
2929 if (next
->vm_start
!= next
->vm_prev
->vm_end
)
2932 if (next
->vm_file
!= vma
->vm_file
)
2935 if (next
->vm_flags
!= vma
->vm_flags
)
2938 if (start
+ size
<= next
->vm_end
)
2946 prot
|= vma
->vm_flags
& VM_READ
? PROT_READ
: 0;
2947 prot
|= vma
->vm_flags
& VM_WRITE
? PROT_WRITE
: 0;
2948 prot
|= vma
->vm_flags
& VM_EXEC
? PROT_EXEC
: 0;
2950 flags
&= MAP_NONBLOCK
;
2951 flags
|= MAP_SHARED
| MAP_FIXED
| MAP_POPULATE
;
2952 if (vma
->vm_flags
& VM_LOCKED
)
2953 flags
|= MAP_LOCKED
;
2955 file
= get_file(vma
->vm_file
);
2956 ret
= do_mmap(vma
->vm_file
, start
, size
,
2957 prot
, flags
, pgoff
, &populate
, NULL
);
2960 mmap_write_unlock(mm
);
2962 mm_populate(ret
, populate
);
2963 if (!IS_ERR_VALUE(ret
))
2969 * this is really a simplified "do_mmap". it only handles
2970 * anonymous maps. eventually we may be able to do some
2971 * brk-specific accounting here.
2973 static int do_brk_flags(unsigned long addr
, unsigned long len
, unsigned long flags
, struct list_head
*uf
)
2975 struct mm_struct
*mm
= current
->mm
;
2976 struct vm_area_struct
*vma
, *prev
;
2977 struct rb_node
**rb_link
, *rb_parent
;
2978 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2980 unsigned long mapped_addr
;
2982 /* Until we need other flags, refuse anything except VM_EXEC. */
2983 if ((flags
& (~VM_EXEC
)) != 0)
2985 flags
|= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2987 mapped_addr
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2988 if (IS_ERR_VALUE(mapped_addr
))
2991 error
= mlock_future_check(mm
, mm
->def_flags
, len
);
2995 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
2996 if (munmap_vma_range(mm
, addr
, len
, &prev
, &rb_link
, &rb_parent
, uf
))
2999 /* Check against address space limits *after* clearing old maps... */
3000 if (!may_expand_vm(mm
, flags
, len
>> PAGE_SHIFT
))
3003 if (mm
->map_count
> sysctl_max_map_count
)
3006 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
3009 /* Can we just expand an old private anonymous mapping? */
3010 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
3011 NULL
, NULL
, pgoff
, NULL
, NULL_VM_UFFD_CTX
, NULL
);
3016 * create a vma struct for an anonymous mapping
3018 vma
= vm_area_alloc(mm
);
3020 vm_unacct_memory(len
>> PAGE_SHIFT
);
3024 vma_set_anonymous(vma
);
3025 vma
->vm_start
= addr
;
3026 vma
->vm_end
= addr
+ len
;
3027 vma
->vm_pgoff
= pgoff
;
3028 vma
->vm_flags
= flags
;
3029 vma
->vm_page_prot
= vm_get_page_prot(flags
);
3030 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
3032 perf_event_mmap(vma
);
3033 mm
->total_vm
+= len
>> PAGE_SHIFT
;
3034 mm
->data_vm
+= len
>> PAGE_SHIFT
;
3035 if (flags
& VM_LOCKED
)
3036 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
3037 vma
->vm_flags
|= VM_SOFTDIRTY
;
3041 int vm_brk_flags(unsigned long addr
, unsigned long request
, unsigned long flags
)
3043 struct mm_struct
*mm
= current
->mm
;
3049 len
= PAGE_ALIGN(request
);
3055 if (mmap_write_lock_killable(mm
))
3058 ret
= do_brk_flags(addr
, len
, flags
, &uf
);
3059 populate
= ((mm
->def_flags
& VM_LOCKED
) != 0);
3060 mmap_write_unlock(mm
);
3061 userfaultfd_unmap_complete(mm
, &uf
);
3062 if (populate
&& !ret
)
3063 mm_populate(addr
, len
);
3066 EXPORT_SYMBOL(vm_brk_flags
);
3068 int vm_brk(unsigned long addr
, unsigned long len
)
3070 return vm_brk_flags(addr
, len
, 0);
3072 EXPORT_SYMBOL(vm_brk
);
3074 /* Release all mmaps. */
3075 void exit_mmap(struct mm_struct
*mm
)
3077 struct mmu_gather tlb
;
3078 struct vm_area_struct
*vma
;
3079 unsigned long nr_accounted
= 0;
3081 /* mm's last user has gone, and its about to be pulled down */
3082 mmu_notifier_release(mm
);
3084 if (unlikely(mm_is_oom_victim(mm
))) {
3086 * Manually reap the mm to free as much memory as possible.
3087 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3088 * this mm from further consideration. Taking mm->mmap_lock for
3089 * write after setting MMF_OOM_SKIP will guarantee that the oom
3090 * reaper will not run on this mm again after mmap_lock is
3093 * Nothing can be holding mm->mmap_lock here and the above call
3094 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3095 * __oom_reap_task_mm() will not block.
3097 (void)__oom_reap_task_mm(mm
);
3098 set_bit(MMF_OOM_SKIP
, &mm
->flags
);
3101 mmap_write_lock(mm
);
3106 /* Can happen if dup_mmap() received an OOM */
3107 mmap_write_unlock(mm
);
3113 tlb_gather_mmu_fullmm(&tlb
, mm
);
3114 /* update_hiwater_rss(mm) here? but nobody should be looking */
3115 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3116 unmap_vmas(&tlb
, vma
, 0, -1);
3117 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, USER_PGTABLES_CEILING
);
3118 tlb_finish_mmu(&tlb
);
3120 /* Walk the list again, actually closing and freeing it. */
3122 if (vma
->vm_flags
& VM_ACCOUNT
)
3123 nr_accounted
+= vma_pages(vma
);
3124 vma
= remove_vma(vma
);
3128 mmap_write_unlock(mm
);
3129 vm_unacct_memory(nr_accounted
);
3132 /* Insert vm structure into process list sorted by address
3133 * and into the inode's i_mmap tree. If vm_file is non-NULL
3134 * then i_mmap_rwsem is taken here.
3136 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
3138 struct vm_area_struct
*prev
;
3139 struct rb_node
**rb_link
, *rb_parent
;
3141 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
3142 &prev
, &rb_link
, &rb_parent
))
3144 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
3145 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
3149 * The vm_pgoff of a purely anonymous vma should be irrelevant
3150 * until its first write fault, when page's anon_vma and index
3151 * are set. But now set the vm_pgoff it will almost certainly
3152 * end up with (unless mremap moves it elsewhere before that
3153 * first wfault), so /proc/pid/maps tells a consistent story.
3155 * By setting it to reflect the virtual start address of the
3156 * vma, merges and splits can happen in a seamless way, just
3157 * using the existing file pgoff checks and manipulations.
3158 * Similarly in do_mmap and in do_brk_flags.
3160 if (vma_is_anonymous(vma
)) {
3161 BUG_ON(vma
->anon_vma
);
3162 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
3165 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
3170 * Copy the vma structure to a new location in the same mm,
3171 * prior to moving page table entries, to effect an mremap move.
3173 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
3174 unsigned long addr
, unsigned long len
, pgoff_t pgoff
,
3175 bool *need_rmap_locks
)
3177 struct vm_area_struct
*vma
= *vmap
;
3178 unsigned long vma_start
= vma
->vm_start
;
3179 struct mm_struct
*mm
= vma
->vm_mm
;
3180 struct vm_area_struct
*new_vma
, *prev
;
3181 struct rb_node
**rb_link
, *rb_parent
;
3182 bool faulted_in_anon_vma
= true;
3185 * If anonymous vma has not yet been faulted, update new pgoff
3186 * to match new location, to increase its chance of merging.
3188 if (unlikely(vma_is_anonymous(vma
) && !vma
->anon_vma
)) {
3189 pgoff
= addr
>> PAGE_SHIFT
;
3190 faulted_in_anon_vma
= false;
3193 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
3194 return NULL
; /* should never get here */
3195 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
3196 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
),
3197 vma
->vm_userfaultfd_ctx
, anon_vma_name(vma
));
3200 * Source vma may have been merged into new_vma
3202 if (unlikely(vma_start
>= new_vma
->vm_start
&&
3203 vma_start
< new_vma
->vm_end
)) {
3205 * The only way we can get a vma_merge with
3206 * self during an mremap is if the vma hasn't
3207 * been faulted in yet and we were allowed to
3208 * reset the dst vma->vm_pgoff to the
3209 * destination address of the mremap to allow
3210 * the merge to happen. mremap must change the
3211 * vm_pgoff linearity between src and dst vmas
3212 * (in turn preventing a vma_merge) to be
3213 * safe. It is only safe to keep the vm_pgoff
3214 * linear if there are no pages mapped yet.
3216 VM_BUG_ON_VMA(faulted_in_anon_vma
, new_vma
);
3217 *vmap
= vma
= new_vma
;
3219 *need_rmap_locks
= (new_vma
->vm_pgoff
<= vma
->vm_pgoff
);
3221 new_vma
= vm_area_dup(vma
);
3224 new_vma
->vm_start
= addr
;
3225 new_vma
->vm_end
= addr
+ len
;
3226 new_vma
->vm_pgoff
= pgoff
;
3227 if (vma_dup_policy(vma
, new_vma
))
3229 if (anon_vma_clone(new_vma
, vma
))
3230 goto out_free_mempol
;
3231 if (new_vma
->vm_file
)
3232 get_file(new_vma
->vm_file
);
3233 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
3234 new_vma
->vm_ops
->open(new_vma
);
3235 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
3236 *need_rmap_locks
= false;
3241 mpol_put(vma_policy(new_vma
));
3243 vm_area_free(new_vma
);
3249 * Return true if the calling process may expand its vm space by the passed
3252 bool may_expand_vm(struct mm_struct
*mm
, vm_flags_t flags
, unsigned long npages
)
3254 if (mm
->total_vm
+ npages
> rlimit(RLIMIT_AS
) >> PAGE_SHIFT
)
3257 if (is_data_mapping(flags
) &&
3258 mm
->data_vm
+ npages
> rlimit(RLIMIT_DATA
) >> PAGE_SHIFT
) {
3259 /* Workaround for Valgrind */
3260 if (rlimit(RLIMIT_DATA
) == 0 &&
3261 mm
->data_vm
+ npages
<= rlimit_max(RLIMIT_DATA
) >> PAGE_SHIFT
)
3264 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3265 current
->comm
, current
->pid
,
3266 (mm
->data_vm
+ npages
) << PAGE_SHIFT
,
3267 rlimit(RLIMIT_DATA
),
3268 ignore_rlimit_data
? "" : " or use boot option ignore_rlimit_data");
3270 if (!ignore_rlimit_data
)
3277 void vm_stat_account(struct mm_struct
*mm
, vm_flags_t flags
, long npages
)
3279 WRITE_ONCE(mm
->total_vm
, READ_ONCE(mm
->total_vm
)+npages
);
3281 if (is_exec_mapping(flags
))
3282 mm
->exec_vm
+= npages
;
3283 else if (is_stack_mapping(flags
))
3284 mm
->stack_vm
+= npages
;
3285 else if (is_data_mapping(flags
))
3286 mm
->data_vm
+= npages
;
3289 static vm_fault_t
special_mapping_fault(struct vm_fault
*vmf
);
3292 * Having a close hook prevents vma merging regardless of flags.
3294 static void special_mapping_close(struct vm_area_struct
*vma
)
3298 static const char *special_mapping_name(struct vm_area_struct
*vma
)
3300 return ((struct vm_special_mapping
*)vma
->vm_private_data
)->name
;
3303 static int special_mapping_mremap(struct vm_area_struct
*new_vma
)
3305 struct vm_special_mapping
*sm
= new_vma
->vm_private_data
;
3307 if (WARN_ON_ONCE(current
->mm
!= new_vma
->vm_mm
))
3311 return sm
->mremap(sm
, new_vma
);
3316 static int special_mapping_split(struct vm_area_struct
*vma
, unsigned long addr
)
3319 * Forbid splitting special mappings - kernel has expectations over
3320 * the number of pages in mapping. Together with VM_DONTEXPAND
3321 * the size of vma should stay the same over the special mapping's
3327 static const struct vm_operations_struct special_mapping_vmops
= {
3328 .close
= special_mapping_close
,
3329 .fault
= special_mapping_fault
,
3330 .mremap
= special_mapping_mremap
,
3331 .name
= special_mapping_name
,
3332 /* vDSO code relies that VVAR can't be accessed remotely */
3334 .may_split
= special_mapping_split
,
3337 static const struct vm_operations_struct legacy_special_mapping_vmops
= {
3338 .close
= special_mapping_close
,
3339 .fault
= special_mapping_fault
,
3342 static vm_fault_t
special_mapping_fault(struct vm_fault
*vmf
)
3344 struct vm_area_struct
*vma
= vmf
->vma
;
3346 struct page
**pages
;
3348 if (vma
->vm_ops
== &legacy_special_mapping_vmops
) {
3349 pages
= vma
->vm_private_data
;
3351 struct vm_special_mapping
*sm
= vma
->vm_private_data
;
3354 return sm
->fault(sm
, vmf
->vma
, vmf
);
3359 for (pgoff
= vmf
->pgoff
; pgoff
&& *pages
; ++pages
)
3363 struct page
*page
= *pages
;
3369 return VM_FAULT_SIGBUS
;
3372 static struct vm_area_struct
*__install_special_mapping(
3373 struct mm_struct
*mm
,
3374 unsigned long addr
, unsigned long len
,
3375 unsigned long vm_flags
, void *priv
,
3376 const struct vm_operations_struct
*ops
)
3379 struct vm_area_struct
*vma
;
3381 vma
= vm_area_alloc(mm
);
3382 if (unlikely(vma
== NULL
))
3383 return ERR_PTR(-ENOMEM
);
3385 vma
->vm_start
= addr
;
3386 vma
->vm_end
= addr
+ len
;
3388 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
| VM_SOFTDIRTY
;
3389 vma
->vm_flags
&= VM_LOCKED_CLEAR_MASK
;
3390 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
3393 vma
->vm_private_data
= priv
;
3395 ret
= insert_vm_struct(mm
, vma
);
3399 vm_stat_account(mm
, vma
->vm_flags
, len
>> PAGE_SHIFT
);
3401 perf_event_mmap(vma
);
3407 return ERR_PTR(ret
);
3410 bool vma_is_special_mapping(const struct vm_area_struct
*vma
,
3411 const struct vm_special_mapping
*sm
)
3413 return vma
->vm_private_data
== sm
&&
3414 (vma
->vm_ops
== &special_mapping_vmops
||
3415 vma
->vm_ops
== &legacy_special_mapping_vmops
);
3419 * Called with mm->mmap_lock held for writing.
3420 * Insert a new vma covering the given region, with the given flags.
3421 * Its pages are supplied by the given array of struct page *.
3422 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3423 * The region past the last page supplied will always produce SIGBUS.
3424 * The array pointer and the pages it points to are assumed to stay alive
3425 * for as long as this mapping might exist.
3427 struct vm_area_struct
*_install_special_mapping(
3428 struct mm_struct
*mm
,
3429 unsigned long addr
, unsigned long len
,
3430 unsigned long vm_flags
, const struct vm_special_mapping
*spec
)
3432 return __install_special_mapping(mm
, addr
, len
, vm_flags
, (void *)spec
,
3433 &special_mapping_vmops
);
3436 int install_special_mapping(struct mm_struct
*mm
,
3437 unsigned long addr
, unsigned long len
,
3438 unsigned long vm_flags
, struct page
**pages
)
3440 struct vm_area_struct
*vma
= __install_special_mapping(
3441 mm
, addr
, len
, vm_flags
, (void *)pages
,
3442 &legacy_special_mapping_vmops
);
3444 return PTR_ERR_OR_ZERO(vma
);
3447 static DEFINE_MUTEX(mm_all_locks_mutex
);
3449 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
3451 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_root
.rb_node
)) {
3453 * The LSB of head.next can't change from under us
3454 * because we hold the mm_all_locks_mutex.
3456 down_write_nest_lock(&anon_vma
->root
->rwsem
, &mm
->mmap_lock
);
3458 * We can safely modify head.next after taking the
3459 * anon_vma->root->rwsem. If some other vma in this mm shares
3460 * the same anon_vma we won't take it again.
3462 * No need of atomic instructions here, head.next
3463 * can't change from under us thanks to the
3464 * anon_vma->root->rwsem.
3466 if (__test_and_set_bit(0, (unsigned long *)
3467 &anon_vma
->root
->rb_root
.rb_root
.rb_node
))
3472 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
3474 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3476 * AS_MM_ALL_LOCKS can't change from under us because
3477 * we hold the mm_all_locks_mutex.
3479 * Operations on ->flags have to be atomic because
3480 * even if AS_MM_ALL_LOCKS is stable thanks to the
3481 * mm_all_locks_mutex, there may be other cpus
3482 * changing other bitflags in parallel to us.
3484 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
3486 down_write_nest_lock(&mapping
->i_mmap_rwsem
, &mm
->mmap_lock
);
3491 * This operation locks against the VM for all pte/vma/mm related
3492 * operations that could ever happen on a certain mm. This includes
3493 * vmtruncate, try_to_unmap, and all page faults.
3495 * The caller must take the mmap_lock in write mode before calling
3496 * mm_take_all_locks(). The caller isn't allowed to release the
3497 * mmap_lock until mm_drop_all_locks() returns.
3499 * mmap_lock in write mode is required in order to block all operations
3500 * that could modify pagetables and free pages without need of
3501 * altering the vma layout. It's also needed in write mode to avoid new
3502 * anon_vmas to be associated with existing vmas.
3504 * A single task can't take more than one mm_take_all_locks() in a row
3505 * or it would deadlock.
3507 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3508 * mapping->flags avoid to take the same lock twice, if more than one
3509 * vma in this mm is backed by the same anon_vma or address_space.
3511 * We take locks in following order, accordingly to comment at beginning
3513 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3515 * - all i_mmap_rwsem locks;
3516 * - all anon_vma->rwseml
3518 * We can take all locks within these types randomly because the VM code
3519 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3520 * mm_all_locks_mutex.
3522 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3523 * that may have to take thousand of locks.
3525 * mm_take_all_locks() can fail if it's interrupted by signals.
3527 int mm_take_all_locks(struct mm_struct
*mm
)
3529 struct vm_area_struct
*vma
;
3530 struct anon_vma_chain
*avc
;
3532 mmap_assert_write_locked(mm
);
3534 mutex_lock(&mm_all_locks_mutex
);
3536 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3537 if (signal_pending(current
))
3539 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
3540 is_vm_hugetlb_page(vma
))
3541 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3544 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3545 if (signal_pending(current
))
3547 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
3548 !is_vm_hugetlb_page(vma
))
3549 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3552 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3553 if (signal_pending(current
))
3556 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3557 vm_lock_anon_vma(mm
, avc
->anon_vma
);
3563 mm_drop_all_locks(mm
);
3567 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
3569 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_root
.rb_node
)) {
3571 * The LSB of head.next can't change to 0 from under
3572 * us because we hold the mm_all_locks_mutex.
3574 * We must however clear the bitflag before unlocking
3575 * the vma so the users using the anon_vma->rb_root will
3576 * never see our bitflag.
3578 * No need of atomic instructions here, head.next
3579 * can't change from under us until we release the
3580 * anon_vma->root->rwsem.
3582 if (!__test_and_clear_bit(0, (unsigned long *)
3583 &anon_vma
->root
->rb_root
.rb_root
.rb_node
))
3585 anon_vma_unlock_write(anon_vma
);
3589 static void vm_unlock_mapping(struct address_space
*mapping
)
3591 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3593 * AS_MM_ALL_LOCKS can't change to 0 from under us
3594 * because we hold the mm_all_locks_mutex.
3596 i_mmap_unlock_write(mapping
);
3597 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
3604 * The mmap_lock cannot be released by the caller until
3605 * mm_drop_all_locks() returns.
3607 void mm_drop_all_locks(struct mm_struct
*mm
)
3609 struct vm_area_struct
*vma
;
3610 struct anon_vma_chain
*avc
;
3612 mmap_assert_write_locked(mm
);
3613 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
3615 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3617 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3618 vm_unlock_anon_vma(avc
->anon_vma
);
3619 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3620 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
3623 mutex_unlock(&mm_all_locks_mutex
);
3627 * initialise the percpu counter for VM
3629 void __init
mmap_init(void)
3633 ret
= percpu_counter_init(&vm_committed_as
, 0, GFP_KERNEL
);
3638 * Initialise sysctl_user_reserve_kbytes.
3640 * This is intended to prevent a user from starting a single memory hogging
3641 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3644 * The default value is min(3% of free memory, 128MB)
3645 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3647 static int init_user_reserve(void)
3649 unsigned long free_kbytes
;
3651 free_kbytes
= global_zone_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3653 sysctl_user_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 17);
3656 subsys_initcall(init_user_reserve
);
3659 * Initialise sysctl_admin_reserve_kbytes.
3661 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3662 * to log in and kill a memory hogging process.
3664 * Systems with more than 256MB will reserve 8MB, enough to recover
3665 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3666 * only reserve 3% of free pages by default.
3668 static int init_admin_reserve(void)
3670 unsigned long free_kbytes
;
3672 free_kbytes
= global_zone_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3674 sysctl_admin_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 13);
3677 subsys_initcall(init_admin_reserve
);
3680 * Reinititalise user and admin reserves if memory is added or removed.
3682 * The default user reserve max is 128MB, and the default max for the
3683 * admin reserve is 8MB. These are usually, but not always, enough to
3684 * enable recovery from a memory hogging process using login/sshd, a shell,
3685 * and tools like top. It may make sense to increase or even disable the
3686 * reserve depending on the existence of swap or variations in the recovery
3687 * tools. So, the admin may have changed them.
3689 * If memory is added and the reserves have been eliminated or increased above
3690 * the default max, then we'll trust the admin.
3692 * If memory is removed and there isn't enough free memory, then we
3693 * need to reset the reserves.
3695 * Otherwise keep the reserve set by the admin.
3697 static int reserve_mem_notifier(struct notifier_block
*nb
,
3698 unsigned long action
, void *data
)
3700 unsigned long tmp
, free_kbytes
;
3704 /* Default max is 128MB. Leave alone if modified by operator. */
3705 tmp
= sysctl_user_reserve_kbytes
;
3706 if (0 < tmp
&& tmp
< (1UL << 17))
3707 init_user_reserve();
3709 /* Default max is 8MB. Leave alone if modified by operator. */
3710 tmp
= sysctl_admin_reserve_kbytes
;
3711 if (0 < tmp
&& tmp
< (1UL << 13))
3712 init_admin_reserve();
3716 free_kbytes
= global_zone_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3718 if (sysctl_user_reserve_kbytes
> free_kbytes
) {
3719 init_user_reserve();
3720 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3721 sysctl_user_reserve_kbytes
);
3724 if (sysctl_admin_reserve_kbytes
> free_kbytes
) {
3725 init_admin_reserve();
3726 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3727 sysctl_admin_reserve_kbytes
);
3736 static struct notifier_block reserve_mem_nb
= {
3737 .notifier_call
= reserve_mem_notifier
,
3740 static int __meminit
init_reserve_notifier(void)
3742 if (register_hotmemory_notifier(&reserve_mem_nb
))
3743 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3747 subsys_initcall(init_reserve_notifier
);