6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/shmem_fs.h>
29 #include <linux/profile.h>
30 #include <linux/export.h>
31 #include <linux/mount.h>
32 #include <linux/mempolicy.h>
33 #include <linux/rmap.h>
34 #include <linux/mmu_notifier.h>
35 #include <linux/mmdebug.h>
36 #include <linux/perf_event.h>
37 #include <linux/audit.h>
38 #include <linux/khugepaged.h>
39 #include <linux/uprobes.h>
40 #include <linux/rbtree_augmented.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/moduleparam.h>
46 #include <linux/pkeys.h>
48 #include <asm/uaccess.h>
49 #include <asm/cacheflush.h>
51 #include <asm/mmu_context.h>
55 #ifndef arch_mmap_check
56 #define arch_mmap_check(addr, len, flags) (0)
59 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
60 const int mmap_rnd_bits_min
= CONFIG_ARCH_MMAP_RND_BITS_MIN
;
61 const int mmap_rnd_bits_max
= CONFIG_ARCH_MMAP_RND_BITS_MAX
;
62 int mmap_rnd_bits __read_mostly
= CONFIG_ARCH_MMAP_RND_BITS
;
64 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
65 const int mmap_rnd_compat_bits_min
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN
;
66 const int mmap_rnd_compat_bits_max
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX
;
67 int mmap_rnd_compat_bits __read_mostly
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS
;
70 static bool ignore_rlimit_data
;
71 core_param(ignore_rlimit_data
, ignore_rlimit_data
, bool, 0644);
73 static void unmap_region(struct mm_struct
*mm
,
74 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
75 unsigned long start
, unsigned long end
);
77 /* description of effects of mapping type and prot in current implementation.
78 * this is due to the limited x86 page protection hardware. The expected
79 * behavior is in parens:
82 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
83 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
84 * w: (no) no w: (no) no w: (yes) yes w: (no) no
85 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
87 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
88 * w: (no) no w: (no) no w: (copy) copy w: (no) no
89 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
91 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
97 pgprot_t protection_map
[16] = {
98 __P000
, __P001
, __P010
, __P011
, __P100
, __P101
, __P110
, __P111
,
99 __S000
, __S001
, __S010
, __S011
, __S100
, __S101
, __S110
, __S111
102 pgprot_t
vm_get_page_prot(unsigned long vm_flags
)
104 return __pgprot(pgprot_val(protection_map
[vm_flags
&
105 (VM_READ
|VM_WRITE
|VM_EXEC
|VM_SHARED
)]) |
106 pgprot_val(arch_vm_get_page_prot(vm_flags
)));
108 EXPORT_SYMBOL(vm_get_page_prot
);
110 static pgprot_t
vm_pgprot_modify(pgprot_t oldprot
, unsigned long vm_flags
)
112 return pgprot_modify(oldprot
, vm_get_page_prot(vm_flags
));
115 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
116 void vma_set_page_prot(struct vm_area_struct
*vma
)
118 unsigned long vm_flags
= vma
->vm_flags
;
120 vma
->vm_page_prot
= vm_pgprot_modify(vma
->vm_page_prot
, vm_flags
);
121 if (vma_wants_writenotify(vma
)) {
122 vm_flags
&= ~VM_SHARED
;
123 vma
->vm_page_prot
= vm_pgprot_modify(vma
->vm_page_prot
,
129 * Requires inode->i_mapping->i_mmap_rwsem
131 static void __remove_shared_vm_struct(struct vm_area_struct
*vma
,
132 struct file
*file
, struct address_space
*mapping
)
134 if (vma
->vm_flags
& VM_DENYWRITE
)
135 atomic_inc(&file_inode(file
)->i_writecount
);
136 if (vma
->vm_flags
& VM_SHARED
)
137 mapping_unmap_writable(mapping
);
139 flush_dcache_mmap_lock(mapping
);
140 vma_interval_tree_remove(vma
, &mapping
->i_mmap
);
141 flush_dcache_mmap_unlock(mapping
);
145 * Unlink a file-based vm structure from its interval tree, to hide
146 * vma from rmap and vmtruncate before freeing its page tables.
148 void unlink_file_vma(struct vm_area_struct
*vma
)
150 struct file
*file
= vma
->vm_file
;
153 struct address_space
*mapping
= file
->f_mapping
;
154 i_mmap_lock_write(mapping
);
155 __remove_shared_vm_struct(vma
, file
, mapping
);
156 i_mmap_unlock_write(mapping
);
161 * Close a vm structure and free it, returning the next.
163 static struct vm_area_struct
*remove_vma(struct vm_area_struct
*vma
)
165 struct vm_area_struct
*next
= vma
->vm_next
;
168 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
169 vma
->vm_ops
->close(vma
);
172 mpol_put(vma_policy(vma
));
173 kmem_cache_free(vm_area_cachep
, vma
);
177 static int do_brk(unsigned long addr
, unsigned long len
);
179 SYSCALL_DEFINE1(brk
, unsigned long, brk
)
181 unsigned long retval
;
182 unsigned long newbrk
, oldbrk
;
183 struct mm_struct
*mm
= current
->mm
;
184 unsigned long min_brk
;
187 if (down_write_killable(&mm
->mmap_sem
))
190 #ifdef CONFIG_COMPAT_BRK
192 * CONFIG_COMPAT_BRK can still be overridden by setting
193 * randomize_va_space to 2, which will still cause mm->start_brk
194 * to be arbitrarily shifted
196 if (current
->brk_randomized
)
197 min_brk
= mm
->start_brk
;
199 min_brk
= mm
->end_data
;
201 min_brk
= mm
->start_brk
;
207 * Check against rlimit here. If this check is done later after the test
208 * of oldbrk with newbrk then it can escape the test and let the data
209 * segment grow beyond its set limit the in case where the limit is
210 * not page aligned -Ram Gupta
212 if (check_data_rlimit(rlimit(RLIMIT_DATA
), brk
, mm
->start_brk
,
213 mm
->end_data
, mm
->start_data
))
216 newbrk
= PAGE_ALIGN(brk
);
217 oldbrk
= PAGE_ALIGN(mm
->brk
);
218 if (oldbrk
== newbrk
)
221 /* Always allow shrinking brk. */
222 if (brk
<= mm
->brk
) {
223 if (!do_munmap(mm
, newbrk
, oldbrk
-newbrk
))
228 /* Check against existing mmap mappings. */
229 if (find_vma_intersection(mm
, oldbrk
, newbrk
+PAGE_SIZE
))
232 /* Ok, looks good - let it rip. */
233 if (do_brk(oldbrk
, newbrk
-oldbrk
) < 0)
238 populate
= newbrk
> oldbrk
&& (mm
->def_flags
& VM_LOCKED
) != 0;
239 up_write(&mm
->mmap_sem
);
241 mm_populate(oldbrk
, newbrk
- oldbrk
);
246 up_write(&mm
->mmap_sem
);
250 static long vma_compute_subtree_gap(struct vm_area_struct
*vma
)
252 unsigned long max
, subtree_gap
;
255 max
-= vma
->vm_prev
->vm_end
;
256 if (vma
->vm_rb
.rb_left
) {
257 subtree_gap
= rb_entry(vma
->vm_rb
.rb_left
,
258 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
259 if (subtree_gap
> max
)
262 if (vma
->vm_rb
.rb_right
) {
263 subtree_gap
= rb_entry(vma
->vm_rb
.rb_right
,
264 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
265 if (subtree_gap
> max
)
271 #ifdef CONFIG_DEBUG_VM_RB
272 static int browse_rb(struct mm_struct
*mm
)
274 struct rb_root
*root
= &mm
->mm_rb
;
275 int i
= 0, j
, bug
= 0;
276 struct rb_node
*nd
, *pn
= NULL
;
277 unsigned long prev
= 0, pend
= 0;
279 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
280 struct vm_area_struct
*vma
;
281 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
282 if (vma
->vm_start
< prev
) {
283 pr_emerg("vm_start %lx < prev %lx\n",
284 vma
->vm_start
, prev
);
287 if (vma
->vm_start
< pend
) {
288 pr_emerg("vm_start %lx < pend %lx\n",
289 vma
->vm_start
, pend
);
292 if (vma
->vm_start
> vma
->vm_end
) {
293 pr_emerg("vm_start %lx > vm_end %lx\n",
294 vma
->vm_start
, vma
->vm_end
);
297 spin_lock(&mm
->page_table_lock
);
298 if (vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
)) {
299 pr_emerg("free gap %lx, correct %lx\n",
301 vma_compute_subtree_gap(vma
));
304 spin_unlock(&mm
->page_table_lock
);
307 prev
= vma
->vm_start
;
311 for (nd
= pn
; nd
; nd
= rb_prev(nd
))
314 pr_emerg("backwards %d, forwards %d\n", j
, i
);
320 static void validate_mm_rb(struct rb_root
*root
, struct vm_area_struct
*ignore
)
324 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
325 struct vm_area_struct
*vma
;
326 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
327 VM_BUG_ON_VMA(vma
!= ignore
&&
328 vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
),
333 static void validate_mm(struct mm_struct
*mm
)
337 unsigned long highest_address
= 0;
338 struct vm_area_struct
*vma
= mm
->mmap
;
341 struct anon_vma
*anon_vma
= vma
->anon_vma
;
342 struct anon_vma_chain
*avc
;
345 anon_vma_lock_read(anon_vma
);
346 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
347 anon_vma_interval_tree_verify(avc
);
348 anon_vma_unlock_read(anon_vma
);
351 highest_address
= vma
->vm_end
;
355 if (i
!= mm
->map_count
) {
356 pr_emerg("map_count %d vm_next %d\n", mm
->map_count
, i
);
359 if (highest_address
!= mm
->highest_vm_end
) {
360 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
361 mm
->highest_vm_end
, highest_address
);
365 if (i
!= mm
->map_count
) {
367 pr_emerg("map_count %d rb %d\n", mm
->map_count
, i
);
370 VM_BUG_ON_MM(bug
, mm
);
373 #define validate_mm_rb(root, ignore) do { } while (0)
374 #define validate_mm(mm) do { } while (0)
377 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks
, struct vm_area_struct
, vm_rb
,
378 unsigned long, rb_subtree_gap
, vma_compute_subtree_gap
)
381 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
382 * vma->vm_prev->vm_end values changed, without modifying the vma's position
385 static void vma_gap_update(struct vm_area_struct
*vma
)
388 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
389 * function that does exacltly what we want.
391 vma_gap_callbacks_propagate(&vma
->vm_rb
, NULL
);
394 static inline void vma_rb_insert(struct vm_area_struct
*vma
,
395 struct rb_root
*root
)
397 /* All rb_subtree_gap values must be consistent prior to insertion */
398 validate_mm_rb(root
, NULL
);
400 rb_insert_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
403 static void vma_rb_erase(struct vm_area_struct
*vma
, struct rb_root
*root
)
406 * All rb_subtree_gap values must be consistent prior to erase,
407 * with the possible exception of the vma being erased.
409 validate_mm_rb(root
, vma
);
412 * Note rb_erase_augmented is a fairly large inline function,
413 * so make sure we instantiate it only once with our desired
414 * augmented rbtree callbacks.
416 rb_erase_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
420 * vma has some anon_vma assigned, and is already inserted on that
421 * anon_vma's interval trees.
423 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
424 * vma must be removed from the anon_vma's interval trees using
425 * anon_vma_interval_tree_pre_update_vma().
427 * After the update, the vma will be reinserted using
428 * anon_vma_interval_tree_post_update_vma().
430 * The entire update must be protected by exclusive mmap_sem and by
431 * the root anon_vma's mutex.
434 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct
*vma
)
436 struct anon_vma_chain
*avc
;
438 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
439 anon_vma_interval_tree_remove(avc
, &avc
->anon_vma
->rb_root
);
443 anon_vma_interval_tree_post_update_vma(struct vm_area_struct
*vma
)
445 struct anon_vma_chain
*avc
;
447 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
448 anon_vma_interval_tree_insert(avc
, &avc
->anon_vma
->rb_root
);
451 static int find_vma_links(struct mm_struct
*mm
, unsigned long addr
,
452 unsigned long end
, struct vm_area_struct
**pprev
,
453 struct rb_node
***rb_link
, struct rb_node
**rb_parent
)
455 struct rb_node
**__rb_link
, *__rb_parent
, *rb_prev
;
457 __rb_link
= &mm
->mm_rb
.rb_node
;
458 rb_prev
= __rb_parent
= NULL
;
461 struct vm_area_struct
*vma_tmp
;
463 __rb_parent
= *__rb_link
;
464 vma_tmp
= rb_entry(__rb_parent
, struct vm_area_struct
, vm_rb
);
466 if (vma_tmp
->vm_end
> addr
) {
467 /* Fail if an existing vma overlaps the area */
468 if (vma_tmp
->vm_start
< end
)
470 __rb_link
= &__rb_parent
->rb_left
;
472 rb_prev
= __rb_parent
;
473 __rb_link
= &__rb_parent
->rb_right
;
479 *pprev
= rb_entry(rb_prev
, struct vm_area_struct
, vm_rb
);
480 *rb_link
= __rb_link
;
481 *rb_parent
= __rb_parent
;
485 static unsigned long count_vma_pages_range(struct mm_struct
*mm
,
486 unsigned long addr
, unsigned long end
)
488 unsigned long nr_pages
= 0;
489 struct vm_area_struct
*vma
;
491 /* Find first overlaping mapping */
492 vma
= find_vma_intersection(mm
, addr
, end
);
496 nr_pages
= (min(end
, vma
->vm_end
) -
497 max(addr
, vma
->vm_start
)) >> PAGE_SHIFT
;
499 /* Iterate over the rest of the overlaps */
500 for (vma
= vma
->vm_next
; vma
; vma
= vma
->vm_next
) {
501 unsigned long overlap_len
;
503 if (vma
->vm_start
> end
)
506 overlap_len
= min(end
, vma
->vm_end
) - vma
->vm_start
;
507 nr_pages
+= overlap_len
>> PAGE_SHIFT
;
513 void __vma_link_rb(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
514 struct rb_node
**rb_link
, struct rb_node
*rb_parent
)
516 /* Update tracking information for the gap following the new vma. */
518 vma_gap_update(vma
->vm_next
);
520 mm
->highest_vm_end
= vma
->vm_end
;
523 * vma->vm_prev wasn't known when we followed the rbtree to find the
524 * correct insertion point for that vma. As a result, we could not
525 * update the vma vm_rb parents rb_subtree_gap values on the way down.
526 * So, we first insert the vma with a zero rb_subtree_gap value
527 * (to be consistent with what we did on the way down), and then
528 * immediately update the gap to the correct value. Finally we
529 * rebalance the rbtree after all augmented values have been set.
531 rb_link_node(&vma
->vm_rb
, rb_parent
, rb_link
);
532 vma
->rb_subtree_gap
= 0;
534 vma_rb_insert(vma
, &mm
->mm_rb
);
537 static void __vma_link_file(struct vm_area_struct
*vma
)
543 struct address_space
*mapping
= file
->f_mapping
;
545 if (vma
->vm_flags
& VM_DENYWRITE
)
546 atomic_dec(&file_inode(file
)->i_writecount
);
547 if (vma
->vm_flags
& VM_SHARED
)
548 atomic_inc(&mapping
->i_mmap_writable
);
550 flush_dcache_mmap_lock(mapping
);
551 vma_interval_tree_insert(vma
, &mapping
->i_mmap
);
552 flush_dcache_mmap_unlock(mapping
);
557 __vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
558 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
559 struct rb_node
*rb_parent
)
561 __vma_link_list(mm
, vma
, prev
, rb_parent
);
562 __vma_link_rb(mm
, vma
, rb_link
, rb_parent
);
565 static void vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
566 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
567 struct rb_node
*rb_parent
)
569 struct address_space
*mapping
= NULL
;
572 mapping
= vma
->vm_file
->f_mapping
;
573 i_mmap_lock_write(mapping
);
576 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
577 __vma_link_file(vma
);
580 i_mmap_unlock_write(mapping
);
587 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
588 * mm's list and rbtree. It has already been inserted into the interval tree.
590 static void __insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
592 struct vm_area_struct
*prev
;
593 struct rb_node
**rb_link
, *rb_parent
;
595 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
596 &prev
, &rb_link
, &rb_parent
))
598 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
603 __vma_unlink(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
604 struct vm_area_struct
*prev
)
606 struct vm_area_struct
*next
;
608 vma_rb_erase(vma
, &mm
->mm_rb
);
609 prev
->vm_next
= next
= vma
->vm_next
;
611 next
->vm_prev
= prev
;
614 vmacache_invalidate(mm
);
618 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
619 * is already present in an i_mmap tree without adjusting the tree.
620 * The following helper function should be used when such adjustments
621 * are necessary. The "insert" vma (if any) is to be inserted
622 * before we drop the necessary locks.
624 int vma_adjust(struct vm_area_struct
*vma
, unsigned long start
,
625 unsigned long end
, pgoff_t pgoff
, struct vm_area_struct
*insert
)
627 struct mm_struct
*mm
= vma
->vm_mm
;
628 struct vm_area_struct
*next
= vma
->vm_next
;
629 struct address_space
*mapping
= NULL
;
630 struct rb_root
*root
= NULL
;
631 struct anon_vma
*anon_vma
= NULL
;
632 struct file
*file
= vma
->vm_file
;
633 bool start_changed
= false, end_changed
= false;
634 long adjust_next
= 0;
637 if (next
&& !insert
) {
638 struct vm_area_struct
*exporter
= NULL
, *importer
= NULL
;
640 if (end
>= next
->vm_end
) {
642 * vma expands, overlapping all the next, and
643 * perhaps the one after too (mprotect case 6).
645 remove_next
= 1 + (end
> next
->vm_end
);
651 * If next doesn't have anon_vma, import from vma after
652 * next, if the vma overlaps with it.
654 if (remove_next
== 2 && next
&& !next
->anon_vma
)
655 exporter
= next
->vm_next
;
657 } else if (end
> next
->vm_start
) {
659 * vma expands, overlapping part of the next:
660 * mprotect case 5 shifting the boundary up.
662 adjust_next
= (end
- next
->vm_start
) >> PAGE_SHIFT
;
665 } else if (end
< vma
->vm_end
) {
667 * vma shrinks, and !insert tells it's not
668 * split_vma inserting another: so it must be
669 * mprotect case 4 shifting the boundary down.
671 adjust_next
= -((vma
->vm_end
- end
) >> PAGE_SHIFT
);
677 * Easily overlooked: when mprotect shifts the boundary,
678 * make sure the expanding vma has anon_vma set if the
679 * shrinking vma had, to cover any anon pages imported.
681 if (exporter
&& exporter
->anon_vma
&& !importer
->anon_vma
) {
684 importer
->anon_vma
= exporter
->anon_vma
;
685 error
= anon_vma_clone(importer
, exporter
);
691 vma_adjust_trans_huge(vma
, start
, end
, adjust_next
);
694 mapping
= file
->f_mapping
;
695 root
= &mapping
->i_mmap
;
696 uprobe_munmap(vma
, vma
->vm_start
, vma
->vm_end
);
699 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
701 i_mmap_lock_write(mapping
);
704 * Put into interval tree now, so instantiated pages
705 * are visible to arm/parisc __flush_dcache_page
706 * throughout; but we cannot insert into address
707 * space until vma start or end is updated.
709 __vma_link_file(insert
);
713 anon_vma
= vma
->anon_vma
;
714 if (!anon_vma
&& adjust_next
)
715 anon_vma
= next
->anon_vma
;
717 VM_BUG_ON_VMA(adjust_next
&& next
->anon_vma
&&
718 anon_vma
!= next
->anon_vma
, next
);
719 anon_vma_lock_write(anon_vma
);
720 anon_vma_interval_tree_pre_update_vma(vma
);
722 anon_vma_interval_tree_pre_update_vma(next
);
726 flush_dcache_mmap_lock(mapping
);
727 vma_interval_tree_remove(vma
, root
);
729 vma_interval_tree_remove(next
, root
);
732 if (start
!= vma
->vm_start
) {
733 vma
->vm_start
= start
;
734 start_changed
= true;
736 if (end
!= vma
->vm_end
) {
740 vma
->vm_pgoff
= pgoff
;
742 next
->vm_start
+= adjust_next
<< PAGE_SHIFT
;
743 next
->vm_pgoff
+= adjust_next
;
748 vma_interval_tree_insert(next
, root
);
749 vma_interval_tree_insert(vma
, root
);
750 flush_dcache_mmap_unlock(mapping
);
755 * vma_merge has merged next into vma, and needs
756 * us to remove next before dropping the locks.
758 __vma_unlink(mm
, next
, vma
);
760 __remove_shared_vm_struct(next
, file
, mapping
);
763 * split_vma has split insert from vma, and needs
764 * us to insert it before dropping the locks
765 * (it may either follow vma or precede it).
767 __insert_vm_struct(mm
, insert
);
773 mm
->highest_vm_end
= end
;
774 else if (!adjust_next
)
775 vma_gap_update(next
);
780 anon_vma_interval_tree_post_update_vma(vma
);
782 anon_vma_interval_tree_post_update_vma(next
);
783 anon_vma_unlock_write(anon_vma
);
786 i_mmap_unlock_write(mapping
);
797 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
801 anon_vma_merge(vma
, next
);
803 mpol_put(vma_policy(next
));
804 kmem_cache_free(vm_area_cachep
, next
);
806 * In mprotect's case 6 (see comments on vma_merge),
807 * we must remove another next too. It would clutter
808 * up the code too much to do both in one go.
811 if (remove_next
== 2) {
817 vma_gap_update(next
);
819 mm
->highest_vm_end
= end
;
830 * If the vma has a ->close operation then the driver probably needs to release
831 * per-vma resources, so we don't attempt to merge those.
833 static inline int is_mergeable_vma(struct vm_area_struct
*vma
,
834 struct file
*file
, unsigned long vm_flags
,
835 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
838 * VM_SOFTDIRTY should not prevent from VMA merging, if we
839 * match the flags but dirty bit -- the caller should mark
840 * merged VMA as dirty. If dirty bit won't be excluded from
841 * comparison, we increase pressue on the memory system forcing
842 * the kernel to generate new VMAs when old one could be
845 if ((vma
->vm_flags
^ vm_flags
) & ~VM_SOFTDIRTY
)
847 if (vma
->vm_file
!= file
)
849 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
851 if (!is_mergeable_vm_userfaultfd_ctx(vma
, vm_userfaultfd_ctx
))
856 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
857 struct anon_vma
*anon_vma2
,
858 struct vm_area_struct
*vma
)
861 * The list_is_singular() test is to avoid merging VMA cloned from
862 * parents. This can improve scalability caused by anon_vma lock.
864 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
865 list_is_singular(&vma
->anon_vma_chain
)))
867 return anon_vma1
== anon_vma2
;
871 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
872 * in front of (at a lower virtual address and file offset than) the vma.
874 * We cannot merge two vmas if they have differently assigned (non-NULL)
875 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
877 * We don't check here for the merged mmap wrapping around the end of pagecache
878 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
879 * wrap, nor mmaps which cover the final page at index -1UL.
882 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
883 struct anon_vma
*anon_vma
, struct file
*file
,
885 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
887 if (is_mergeable_vma(vma
, file
, vm_flags
, vm_userfaultfd_ctx
) &&
888 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
889 if (vma
->vm_pgoff
== vm_pgoff
)
896 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
897 * beyond (at a higher virtual address and file offset than) the vma.
899 * We cannot merge two vmas if they have differently assigned (non-NULL)
900 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
903 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
904 struct anon_vma
*anon_vma
, struct file
*file
,
906 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
908 if (is_mergeable_vma(vma
, file
, vm_flags
, vm_userfaultfd_ctx
) &&
909 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
911 vm_pglen
= vma_pages(vma
);
912 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
919 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
920 * whether that can be merged with its predecessor or its successor.
921 * Or both (it neatly fills a hole).
923 * In most cases - when called for mmap, brk or mremap - [addr,end) is
924 * certain not to be mapped by the time vma_merge is called; but when
925 * called for mprotect, it is certain to be already mapped (either at
926 * an offset within prev, or at the start of next), and the flags of
927 * this area are about to be changed to vm_flags - and the no-change
928 * case has already been eliminated.
930 * The following mprotect cases have to be considered, where AAAA is
931 * the area passed down from mprotect_fixup, never extending beyond one
932 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
934 * AAAA AAAA AAAA AAAA
935 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
936 * cannot merge might become might become might become
937 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
938 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
939 * mremap move: PPPPNNNNNNNN 8
941 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
942 * might become case 1 below case 2 below case 3 below
944 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
945 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
947 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
948 struct vm_area_struct
*prev
, unsigned long addr
,
949 unsigned long end
, unsigned long vm_flags
,
950 struct anon_vma
*anon_vma
, struct file
*file
,
951 pgoff_t pgoff
, struct mempolicy
*policy
,
952 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
954 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
955 struct vm_area_struct
*area
, *next
;
959 * We later require that vma->vm_flags == vm_flags,
960 * so this tests vma->vm_flags & VM_SPECIAL, too.
962 if (vm_flags
& VM_SPECIAL
)
966 next
= prev
->vm_next
;
970 if (next
&& next
->vm_end
== end
) /* cases 6, 7, 8 */
971 next
= next
->vm_next
;
974 * Can it merge with the predecessor?
976 if (prev
&& prev
->vm_end
== addr
&&
977 mpol_equal(vma_policy(prev
), policy
) &&
978 can_vma_merge_after(prev
, vm_flags
,
979 anon_vma
, file
, pgoff
,
980 vm_userfaultfd_ctx
)) {
982 * OK, it can. Can we now merge in the successor as well?
984 if (next
&& end
== next
->vm_start
&&
985 mpol_equal(policy
, vma_policy(next
)) &&
986 can_vma_merge_before(next
, vm_flags
,
989 vm_userfaultfd_ctx
) &&
990 is_mergeable_anon_vma(prev
->anon_vma
,
991 next
->anon_vma
, NULL
)) {
993 err
= vma_adjust(prev
, prev
->vm_start
,
994 next
->vm_end
, prev
->vm_pgoff
, NULL
);
995 } else /* cases 2, 5, 7 */
996 err
= vma_adjust(prev
, prev
->vm_start
,
997 end
, prev
->vm_pgoff
, NULL
);
1000 khugepaged_enter_vma_merge(prev
, vm_flags
);
1005 * Can this new request be merged in front of next?
1007 if (next
&& end
== next
->vm_start
&&
1008 mpol_equal(policy
, vma_policy(next
)) &&
1009 can_vma_merge_before(next
, vm_flags
,
1010 anon_vma
, file
, pgoff
+pglen
,
1011 vm_userfaultfd_ctx
)) {
1012 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
1013 err
= vma_adjust(prev
, prev
->vm_start
,
1014 addr
, prev
->vm_pgoff
, NULL
);
1015 else /* cases 3, 8 */
1016 err
= vma_adjust(area
, addr
, next
->vm_end
,
1017 next
->vm_pgoff
- pglen
, NULL
);
1020 khugepaged_enter_vma_merge(area
, vm_flags
);
1028 * Rough compatbility check to quickly see if it's even worth looking
1029 * at sharing an anon_vma.
1031 * They need to have the same vm_file, and the flags can only differ
1032 * in things that mprotect may change.
1034 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1035 * we can merge the two vma's. For example, we refuse to merge a vma if
1036 * there is a vm_ops->close() function, because that indicates that the
1037 * driver is doing some kind of reference counting. But that doesn't
1038 * really matter for the anon_vma sharing case.
1040 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1042 return a
->vm_end
== b
->vm_start
&&
1043 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
1044 a
->vm_file
== b
->vm_file
&&
1045 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_READ
|VM_WRITE
|VM_EXEC
|VM_SOFTDIRTY
)) &&
1046 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
1050 * Do some basic sanity checking to see if we can re-use the anon_vma
1051 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1052 * the same as 'old', the other will be the new one that is trying
1053 * to share the anon_vma.
1055 * NOTE! This runs with mm_sem held for reading, so it is possible that
1056 * the anon_vma of 'old' is concurrently in the process of being set up
1057 * by another page fault trying to merge _that_. But that's ok: if it
1058 * is being set up, that automatically means that it will be a singleton
1059 * acceptable for merging, so we can do all of this optimistically. But
1060 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1062 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1063 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1064 * is to return an anon_vma that is "complex" due to having gone through
1067 * We also make sure that the two vma's are compatible (adjacent,
1068 * and with the same memory policies). That's all stable, even with just
1069 * a read lock on the mm_sem.
1071 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1073 if (anon_vma_compatible(a
, b
)) {
1074 struct anon_vma
*anon_vma
= READ_ONCE(old
->anon_vma
);
1076 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
1083 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1084 * neighbouring vmas for a suitable anon_vma, before it goes off
1085 * to allocate a new anon_vma. It checks because a repetitive
1086 * sequence of mprotects and faults may otherwise lead to distinct
1087 * anon_vmas being allocated, preventing vma merge in subsequent
1090 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
1092 struct anon_vma
*anon_vma
;
1093 struct vm_area_struct
*near
;
1095 near
= vma
->vm_next
;
1099 anon_vma
= reusable_anon_vma(near
, vma
, near
);
1103 near
= vma
->vm_prev
;
1107 anon_vma
= reusable_anon_vma(near
, near
, vma
);
1112 * There's no absolute need to look only at touching neighbours:
1113 * we could search further afield for "compatible" anon_vmas.
1114 * But it would probably just be a waste of time searching,
1115 * or lead to too many vmas hanging off the same anon_vma.
1116 * We're trying to allow mprotect remerging later on,
1117 * not trying to minimize memory used for anon_vmas.
1123 * If a hint addr is less than mmap_min_addr change hint to be as
1124 * low as possible but still greater than mmap_min_addr
1126 static inline unsigned long round_hint_to_min(unsigned long hint
)
1129 if (((void *)hint
!= NULL
) &&
1130 (hint
< mmap_min_addr
))
1131 return PAGE_ALIGN(mmap_min_addr
);
1135 static inline int mlock_future_check(struct mm_struct
*mm
,
1136 unsigned long flags
,
1139 unsigned long locked
, lock_limit
;
1141 /* mlock MCL_FUTURE? */
1142 if (flags
& VM_LOCKED
) {
1143 locked
= len
>> PAGE_SHIFT
;
1144 locked
+= mm
->locked_vm
;
1145 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1146 lock_limit
>>= PAGE_SHIFT
;
1147 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1154 * The caller must hold down_write(¤t->mm->mmap_sem).
1156 unsigned long do_mmap(struct file
*file
, unsigned long addr
,
1157 unsigned long len
, unsigned long prot
,
1158 unsigned long flags
, vm_flags_t vm_flags
,
1159 unsigned long pgoff
, unsigned long *populate
)
1161 struct mm_struct
*mm
= current
->mm
;
1170 * Does the application expect PROT_READ to imply PROT_EXEC?
1172 * (the exception is when the underlying filesystem is noexec
1173 * mounted, in which case we dont add PROT_EXEC.)
1175 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1176 if (!(file
&& path_noexec(&file
->f_path
)))
1179 if (!(flags
& MAP_FIXED
))
1180 addr
= round_hint_to_min(addr
);
1182 /* Careful about overflows.. */
1183 len
= PAGE_ALIGN(len
);
1187 /* offset overflow? */
1188 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1191 /* Too many mappings? */
1192 if (mm
->map_count
> sysctl_max_map_count
)
1195 /* Obtain the address to map to. we verify (or select) it and ensure
1196 * that it represents a valid section of the address space.
1198 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1199 if (offset_in_page(addr
))
1202 if (prot
== PROT_EXEC
) {
1203 pkey
= execute_only_pkey(mm
);
1208 /* Do simple checking here so the lower-level routines won't have
1209 * to. we assume access permissions have been handled by the open
1210 * of the memory object, so we don't do any here.
1212 vm_flags
|= calc_vm_prot_bits(prot
, pkey
) | calc_vm_flag_bits(flags
) |
1213 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1215 if (flags
& MAP_LOCKED
)
1216 if (!can_do_mlock())
1219 if (mlock_future_check(mm
, vm_flags
, len
))
1223 struct inode
*inode
= file_inode(file
);
1225 switch (flags
& MAP_TYPE
) {
1227 if ((prot
&PROT_WRITE
) && !(file
->f_mode
&FMODE_WRITE
))
1231 * Make sure we don't allow writing to an append-only
1234 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1238 * Make sure there are no mandatory locks on the file.
1240 if (locks_verify_locked(file
))
1243 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1244 if (!(file
->f_mode
& FMODE_WRITE
))
1245 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1249 if (!(file
->f_mode
& FMODE_READ
))
1251 if (path_noexec(&file
->f_path
)) {
1252 if (vm_flags
& VM_EXEC
)
1254 vm_flags
&= ~VM_MAYEXEC
;
1257 if (!file
->f_op
->mmap
)
1259 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1267 switch (flags
& MAP_TYPE
) {
1269 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1275 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1279 * Set pgoff according to addr for anon_vma.
1281 pgoff
= addr
>> PAGE_SHIFT
;
1289 * Set 'VM_NORESERVE' if we should not account for the
1290 * memory use of this mapping.
1292 if (flags
& MAP_NORESERVE
) {
1293 /* We honor MAP_NORESERVE if allowed to overcommit */
1294 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1295 vm_flags
|= VM_NORESERVE
;
1297 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1298 if (file
&& is_file_hugepages(file
))
1299 vm_flags
|= VM_NORESERVE
;
1302 addr
= mmap_region(file
, addr
, len
, vm_flags
, pgoff
);
1303 if (!IS_ERR_VALUE(addr
) &&
1304 ((vm_flags
& VM_LOCKED
) ||
1305 (flags
& (MAP_POPULATE
| MAP_NONBLOCK
)) == MAP_POPULATE
))
1310 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1311 unsigned long, prot
, unsigned long, flags
,
1312 unsigned long, fd
, unsigned long, pgoff
)
1314 struct file
*file
= NULL
;
1315 unsigned long retval
;
1317 if (!(flags
& MAP_ANONYMOUS
)) {
1318 audit_mmap_fd(fd
, flags
);
1322 if (is_file_hugepages(file
))
1323 len
= ALIGN(len
, huge_page_size(hstate_file(file
)));
1325 if (unlikely(flags
& MAP_HUGETLB
&& !is_file_hugepages(file
)))
1327 } else if (flags
& MAP_HUGETLB
) {
1328 struct user_struct
*user
= NULL
;
1331 hs
= hstate_sizelog((flags
>> MAP_HUGE_SHIFT
) & SHM_HUGE_MASK
);
1335 len
= ALIGN(len
, huge_page_size(hs
));
1337 * VM_NORESERVE is used because the reservations will be
1338 * taken when vm_ops->mmap() is called
1339 * A dummy user value is used because we are not locking
1340 * memory so no accounting is necessary
1342 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, len
,
1344 &user
, HUGETLB_ANONHUGE_INODE
,
1345 (flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1347 return PTR_ERR(file
);
1350 flags
&= ~(MAP_EXECUTABLE
| MAP_DENYWRITE
);
1352 retval
= vm_mmap_pgoff(file
, addr
, len
, prot
, flags
, pgoff
);
1359 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1360 struct mmap_arg_struct
{
1364 unsigned long flags
;
1366 unsigned long offset
;
1369 SYSCALL_DEFINE1(old_mmap
, struct mmap_arg_struct __user
*, arg
)
1371 struct mmap_arg_struct a
;
1373 if (copy_from_user(&a
, arg
, sizeof(a
)))
1375 if (offset_in_page(a
.offset
))
1378 return sys_mmap_pgoff(a
.addr
, a
.len
, a
.prot
, a
.flags
, a
.fd
,
1379 a
.offset
>> PAGE_SHIFT
);
1381 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1384 * Some shared mappigns will want the pages marked read-only
1385 * to track write events. If so, we'll downgrade vm_page_prot
1386 * to the private version (using protection_map[] without the
1389 int vma_wants_writenotify(struct vm_area_struct
*vma
)
1391 vm_flags_t vm_flags
= vma
->vm_flags
;
1392 const struct vm_operations_struct
*vm_ops
= vma
->vm_ops
;
1394 /* If it was private or non-writable, the write bit is already clear */
1395 if ((vm_flags
& (VM_WRITE
|VM_SHARED
)) != ((VM_WRITE
|VM_SHARED
)))
1398 /* The backer wishes to know when pages are first written to? */
1399 if (vm_ops
&& (vm_ops
->page_mkwrite
|| vm_ops
->pfn_mkwrite
))
1402 /* The open routine did something to the protections that pgprot_modify
1403 * won't preserve? */
1404 if (pgprot_val(vma
->vm_page_prot
) !=
1405 pgprot_val(vm_pgprot_modify(vma
->vm_page_prot
, vm_flags
)))
1408 /* Do we need to track softdirty? */
1409 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY
) && !(vm_flags
& VM_SOFTDIRTY
))
1412 /* Specialty mapping? */
1413 if (vm_flags
& VM_PFNMAP
)
1416 /* Can the mapping track the dirty pages? */
1417 return vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
1418 mapping_cap_account_dirty(vma
->vm_file
->f_mapping
);
1422 * We account for memory if it's a private writeable mapping,
1423 * not hugepages and VM_NORESERVE wasn't set.
1425 static inline int accountable_mapping(struct file
*file
, vm_flags_t vm_flags
)
1428 * hugetlb has its own accounting separate from the core VM
1429 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1431 if (file
&& is_file_hugepages(file
))
1434 return (vm_flags
& (VM_NORESERVE
| VM_SHARED
| VM_WRITE
)) == VM_WRITE
;
1437 unsigned long mmap_region(struct file
*file
, unsigned long addr
,
1438 unsigned long len
, vm_flags_t vm_flags
, unsigned long pgoff
)
1440 struct mm_struct
*mm
= current
->mm
;
1441 struct vm_area_struct
*vma
, *prev
;
1443 struct rb_node
**rb_link
, *rb_parent
;
1444 unsigned long charged
= 0;
1446 /* Check against address space limit. */
1447 if (!may_expand_vm(mm
, vm_flags
, len
>> PAGE_SHIFT
)) {
1448 unsigned long nr_pages
;
1451 * MAP_FIXED may remove pages of mappings that intersects with
1452 * requested mapping. Account for the pages it would unmap.
1454 nr_pages
= count_vma_pages_range(mm
, addr
, addr
+ len
);
1456 if (!may_expand_vm(mm
, vm_flags
,
1457 (len
>> PAGE_SHIFT
) - nr_pages
))
1461 /* Clear old maps */
1462 while (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
,
1464 if (do_munmap(mm
, addr
, len
))
1469 * Private writable mapping: check memory availability
1471 if (accountable_mapping(file
, vm_flags
)) {
1472 charged
= len
>> PAGE_SHIFT
;
1473 if (security_vm_enough_memory_mm(mm
, charged
))
1475 vm_flags
|= VM_ACCOUNT
;
1479 * Can we just expand an old mapping?
1481 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
,
1482 NULL
, file
, pgoff
, NULL
, NULL_VM_UFFD_CTX
);
1487 * Determine the object being mapped and call the appropriate
1488 * specific mapper. the address has already been validated, but
1489 * not unmapped, but the maps are removed from the list.
1491 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
1498 vma
->vm_start
= addr
;
1499 vma
->vm_end
= addr
+ len
;
1500 vma
->vm_flags
= vm_flags
;
1501 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1502 vma
->vm_pgoff
= pgoff
;
1503 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
1506 if (vm_flags
& VM_DENYWRITE
) {
1507 error
= deny_write_access(file
);
1511 if (vm_flags
& VM_SHARED
) {
1512 error
= mapping_map_writable(file
->f_mapping
);
1514 goto allow_write_and_free_vma
;
1517 /* ->mmap() can change vma->vm_file, but must guarantee that
1518 * vma_link() below can deny write-access if VM_DENYWRITE is set
1519 * and map writably if VM_SHARED is set. This usually means the
1520 * new file must not have been exposed to user-space, yet.
1522 vma
->vm_file
= get_file(file
);
1523 error
= file
->f_op
->mmap(file
, vma
);
1525 goto unmap_and_free_vma
;
1527 /* Can addr have changed??
1529 * Answer: Yes, several device drivers can do it in their
1530 * f_op->mmap method. -DaveM
1531 * Bug: If addr is changed, prev, rb_link, rb_parent should
1532 * be updated for vma_link()
1534 WARN_ON_ONCE(addr
!= vma
->vm_start
);
1536 addr
= vma
->vm_start
;
1537 vm_flags
= vma
->vm_flags
;
1538 } else if (vm_flags
& VM_SHARED
) {
1539 error
= shmem_zero_setup(vma
);
1544 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1545 /* Once vma denies write, undo our temporary denial count */
1547 if (vm_flags
& VM_SHARED
)
1548 mapping_unmap_writable(file
->f_mapping
);
1549 if (vm_flags
& VM_DENYWRITE
)
1550 allow_write_access(file
);
1552 file
= vma
->vm_file
;
1554 perf_event_mmap(vma
);
1556 vm_stat_account(mm
, vm_flags
, len
>> PAGE_SHIFT
);
1557 if (vm_flags
& VM_LOCKED
) {
1558 if (!((vm_flags
& VM_SPECIAL
) || is_vm_hugetlb_page(vma
) ||
1559 vma
== get_gate_vma(current
->mm
)))
1560 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1562 vma
->vm_flags
&= VM_LOCKED_CLEAR_MASK
;
1569 * New (or expanded) vma always get soft dirty status.
1570 * Otherwise user-space soft-dirty page tracker won't
1571 * be able to distinguish situation when vma area unmapped,
1572 * then new mapped in-place (which must be aimed as
1573 * a completely new data area).
1575 vma
->vm_flags
|= VM_SOFTDIRTY
;
1577 vma_set_page_prot(vma
);
1582 vma
->vm_file
= NULL
;
1585 /* Undo any partial mapping done by a device driver. */
1586 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1588 if (vm_flags
& VM_SHARED
)
1589 mapping_unmap_writable(file
->f_mapping
);
1590 allow_write_and_free_vma
:
1591 if (vm_flags
& VM_DENYWRITE
)
1592 allow_write_access(file
);
1594 kmem_cache_free(vm_area_cachep
, vma
);
1597 vm_unacct_memory(charged
);
1601 unsigned long unmapped_area(struct vm_unmapped_area_info
*info
)
1604 * We implement the search by looking for an rbtree node that
1605 * immediately follows a suitable gap. That is,
1606 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1607 * - gap_end = vma->vm_start >= info->low_limit + length;
1608 * - gap_end - gap_start >= length
1611 struct mm_struct
*mm
= current
->mm
;
1612 struct vm_area_struct
*vma
;
1613 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1615 /* Adjust search length to account for worst case alignment overhead */
1616 length
= info
->length
+ info
->align_mask
;
1617 if (length
< info
->length
)
1620 /* Adjust search limits by the desired length */
1621 if (info
->high_limit
< length
)
1623 high_limit
= info
->high_limit
- length
;
1625 if (info
->low_limit
> high_limit
)
1627 low_limit
= info
->low_limit
+ length
;
1629 /* Check if rbtree root looks promising */
1630 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1632 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1633 if (vma
->rb_subtree_gap
< length
)
1637 /* Visit left subtree if it looks promising */
1638 gap_end
= vma
->vm_start
;
1639 if (gap_end
>= low_limit
&& vma
->vm_rb
.rb_left
) {
1640 struct vm_area_struct
*left
=
1641 rb_entry(vma
->vm_rb
.rb_left
,
1642 struct vm_area_struct
, vm_rb
);
1643 if (left
->rb_subtree_gap
>= length
) {
1649 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1651 /* Check if current node has a suitable gap */
1652 if (gap_start
> high_limit
)
1654 if (gap_end
>= low_limit
&& gap_end
- gap_start
>= length
)
1657 /* Visit right subtree if it looks promising */
1658 if (vma
->vm_rb
.rb_right
) {
1659 struct vm_area_struct
*right
=
1660 rb_entry(vma
->vm_rb
.rb_right
,
1661 struct vm_area_struct
, vm_rb
);
1662 if (right
->rb_subtree_gap
>= length
) {
1668 /* Go back up the rbtree to find next candidate node */
1670 struct rb_node
*prev
= &vma
->vm_rb
;
1671 if (!rb_parent(prev
))
1673 vma
= rb_entry(rb_parent(prev
),
1674 struct vm_area_struct
, vm_rb
);
1675 if (prev
== vma
->vm_rb
.rb_left
) {
1676 gap_start
= vma
->vm_prev
->vm_end
;
1677 gap_end
= vma
->vm_start
;
1684 /* Check highest gap, which does not precede any rbtree node */
1685 gap_start
= mm
->highest_vm_end
;
1686 gap_end
= ULONG_MAX
; /* Only for VM_BUG_ON below */
1687 if (gap_start
> high_limit
)
1691 /* We found a suitable gap. Clip it with the original low_limit. */
1692 if (gap_start
< info
->low_limit
)
1693 gap_start
= info
->low_limit
;
1695 /* Adjust gap address to the desired alignment */
1696 gap_start
+= (info
->align_offset
- gap_start
) & info
->align_mask
;
1698 VM_BUG_ON(gap_start
+ info
->length
> info
->high_limit
);
1699 VM_BUG_ON(gap_start
+ info
->length
> gap_end
);
1703 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info
*info
)
1705 struct mm_struct
*mm
= current
->mm
;
1706 struct vm_area_struct
*vma
;
1707 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1709 /* Adjust search length to account for worst case alignment overhead */
1710 length
= info
->length
+ info
->align_mask
;
1711 if (length
< info
->length
)
1715 * Adjust search limits by the desired length.
1716 * See implementation comment at top of unmapped_area().
1718 gap_end
= info
->high_limit
;
1719 if (gap_end
< length
)
1721 high_limit
= gap_end
- length
;
1723 if (info
->low_limit
> high_limit
)
1725 low_limit
= info
->low_limit
+ length
;
1727 /* Check highest gap, which does not precede any rbtree node */
1728 gap_start
= mm
->highest_vm_end
;
1729 if (gap_start
<= high_limit
)
1732 /* Check if rbtree root looks promising */
1733 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1735 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1736 if (vma
->rb_subtree_gap
< length
)
1740 /* Visit right subtree if it looks promising */
1741 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1742 if (gap_start
<= high_limit
&& vma
->vm_rb
.rb_right
) {
1743 struct vm_area_struct
*right
=
1744 rb_entry(vma
->vm_rb
.rb_right
,
1745 struct vm_area_struct
, vm_rb
);
1746 if (right
->rb_subtree_gap
>= length
) {
1753 /* Check if current node has a suitable gap */
1754 gap_end
= vma
->vm_start
;
1755 if (gap_end
< low_limit
)
1757 if (gap_start
<= high_limit
&& gap_end
- gap_start
>= length
)
1760 /* Visit left subtree if it looks promising */
1761 if (vma
->vm_rb
.rb_left
) {
1762 struct vm_area_struct
*left
=
1763 rb_entry(vma
->vm_rb
.rb_left
,
1764 struct vm_area_struct
, vm_rb
);
1765 if (left
->rb_subtree_gap
>= length
) {
1771 /* Go back up the rbtree to find next candidate node */
1773 struct rb_node
*prev
= &vma
->vm_rb
;
1774 if (!rb_parent(prev
))
1776 vma
= rb_entry(rb_parent(prev
),
1777 struct vm_area_struct
, vm_rb
);
1778 if (prev
== vma
->vm_rb
.rb_right
) {
1779 gap_start
= vma
->vm_prev
?
1780 vma
->vm_prev
->vm_end
: 0;
1787 /* We found a suitable gap. Clip it with the original high_limit. */
1788 if (gap_end
> info
->high_limit
)
1789 gap_end
= info
->high_limit
;
1792 /* Compute highest gap address at the desired alignment */
1793 gap_end
-= info
->length
;
1794 gap_end
-= (gap_end
- info
->align_offset
) & info
->align_mask
;
1796 VM_BUG_ON(gap_end
< info
->low_limit
);
1797 VM_BUG_ON(gap_end
< gap_start
);
1801 /* Get an address range which is currently unmapped.
1802 * For shmat() with addr=0.
1804 * Ugly calling convention alert:
1805 * Return value with the low bits set means error value,
1807 * if (ret & ~PAGE_MASK)
1810 * This function "knows" that -ENOMEM has the bits set.
1812 #ifndef HAVE_ARCH_UNMAPPED_AREA
1814 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
1815 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
1817 struct mm_struct
*mm
= current
->mm
;
1818 struct vm_area_struct
*vma
;
1819 struct vm_unmapped_area_info info
;
1821 if (len
> TASK_SIZE
- mmap_min_addr
)
1824 if (flags
& MAP_FIXED
)
1828 addr
= PAGE_ALIGN(addr
);
1829 vma
= find_vma(mm
, addr
);
1830 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1831 (!vma
|| addr
+ len
<= vma
->vm_start
))
1837 info
.low_limit
= mm
->mmap_base
;
1838 info
.high_limit
= TASK_SIZE
;
1839 info
.align_mask
= 0;
1840 return vm_unmapped_area(&info
);
1845 * This mmap-allocator allocates new areas top-down from below the
1846 * stack's low limit (the base):
1848 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1850 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
1851 const unsigned long len
, const unsigned long pgoff
,
1852 const unsigned long flags
)
1854 struct vm_area_struct
*vma
;
1855 struct mm_struct
*mm
= current
->mm
;
1856 unsigned long addr
= addr0
;
1857 struct vm_unmapped_area_info info
;
1859 /* requested length too big for entire address space */
1860 if (len
> TASK_SIZE
- mmap_min_addr
)
1863 if (flags
& MAP_FIXED
)
1866 /* requesting a specific address */
1868 addr
= PAGE_ALIGN(addr
);
1869 vma
= find_vma(mm
, addr
);
1870 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1871 (!vma
|| addr
+ len
<= vma
->vm_start
))
1875 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
1877 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
1878 info
.high_limit
= mm
->mmap_base
;
1879 info
.align_mask
= 0;
1880 addr
= vm_unmapped_area(&info
);
1883 * A failed mmap() very likely causes application failure,
1884 * so fall back to the bottom-up function here. This scenario
1885 * can happen with large stack limits and large mmap()
1888 if (offset_in_page(addr
)) {
1889 VM_BUG_ON(addr
!= -ENOMEM
);
1891 info
.low_limit
= TASK_UNMAPPED_BASE
;
1892 info
.high_limit
= TASK_SIZE
;
1893 addr
= vm_unmapped_area(&info
);
1901 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
1902 unsigned long pgoff
, unsigned long flags
)
1904 unsigned long (*get_area
)(struct file
*, unsigned long,
1905 unsigned long, unsigned long, unsigned long);
1907 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
1911 /* Careful about overflows.. */
1912 if (len
> TASK_SIZE
)
1915 get_area
= current
->mm
->get_unmapped_area
;
1917 if (file
->f_op
->get_unmapped_area
)
1918 get_area
= file
->f_op
->get_unmapped_area
;
1919 } else if (flags
& MAP_SHARED
) {
1921 * mmap_region() will call shmem_zero_setup() to create a file,
1922 * so use shmem's get_unmapped_area in case it can be huge.
1923 * do_mmap_pgoff() will clear pgoff, so match alignment.
1926 get_area
= shmem_get_unmapped_area
;
1929 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
1930 if (IS_ERR_VALUE(addr
))
1933 if (addr
> TASK_SIZE
- len
)
1935 if (offset_in_page(addr
))
1938 error
= security_mmap_addr(addr
);
1939 return error
? error
: addr
;
1942 EXPORT_SYMBOL(get_unmapped_area
);
1944 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1945 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
1947 struct rb_node
*rb_node
;
1948 struct vm_area_struct
*vma
;
1950 /* Check the cache first. */
1951 vma
= vmacache_find(mm
, addr
);
1955 rb_node
= mm
->mm_rb
.rb_node
;
1958 struct vm_area_struct
*tmp
;
1960 tmp
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
1962 if (tmp
->vm_end
> addr
) {
1964 if (tmp
->vm_start
<= addr
)
1966 rb_node
= rb_node
->rb_left
;
1968 rb_node
= rb_node
->rb_right
;
1972 vmacache_update(addr
, vma
);
1976 EXPORT_SYMBOL(find_vma
);
1979 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1981 struct vm_area_struct
*
1982 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
1983 struct vm_area_struct
**pprev
)
1985 struct vm_area_struct
*vma
;
1987 vma
= find_vma(mm
, addr
);
1989 *pprev
= vma
->vm_prev
;
1991 struct rb_node
*rb_node
= mm
->mm_rb
.rb_node
;
1994 *pprev
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
1995 rb_node
= rb_node
->rb_right
;
2002 * Verify that the stack growth is acceptable and
2003 * update accounting. This is shared with both the
2004 * grow-up and grow-down cases.
2006 static int acct_stack_growth(struct vm_area_struct
*vma
, unsigned long size
, unsigned long grow
)
2008 struct mm_struct
*mm
= vma
->vm_mm
;
2009 struct rlimit
*rlim
= current
->signal
->rlim
;
2010 unsigned long new_start
, actual_size
;
2012 /* address space limit tests */
2013 if (!may_expand_vm(mm
, vma
->vm_flags
, grow
))
2016 /* Stack limit test */
2018 if (size
&& (vma
->vm_flags
& (VM_GROWSUP
| VM_GROWSDOWN
)))
2019 actual_size
-= PAGE_SIZE
;
2020 if (actual_size
> READ_ONCE(rlim
[RLIMIT_STACK
].rlim_cur
))
2023 /* mlock limit tests */
2024 if (vma
->vm_flags
& VM_LOCKED
) {
2025 unsigned long locked
;
2026 unsigned long limit
;
2027 locked
= mm
->locked_vm
+ grow
;
2028 limit
= READ_ONCE(rlim
[RLIMIT_MEMLOCK
].rlim_cur
);
2029 limit
>>= PAGE_SHIFT
;
2030 if (locked
> limit
&& !capable(CAP_IPC_LOCK
))
2034 /* Check to ensure the stack will not grow into a hugetlb-only region */
2035 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
2037 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
2041 * Overcommit.. This must be the final test, as it will
2042 * update security statistics.
2044 if (security_vm_enough_memory_mm(mm
, grow
))
2050 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2052 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2053 * vma is the last one with address > vma->vm_end. Have to extend vma.
2055 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
)
2057 struct mm_struct
*mm
= vma
->vm_mm
;
2060 if (!(vma
->vm_flags
& VM_GROWSUP
))
2063 /* Guard against wrapping around to address 0. */
2064 if (address
< PAGE_ALIGN(address
+4))
2065 address
= PAGE_ALIGN(address
+4);
2069 /* We must make sure the anon_vma is allocated. */
2070 if (unlikely(anon_vma_prepare(vma
)))
2074 * vma->vm_start/vm_end cannot change under us because the caller
2075 * is required to hold the mmap_sem in read mode. We need the
2076 * anon_vma lock to serialize against concurrent expand_stacks.
2078 anon_vma_lock_write(vma
->anon_vma
);
2080 /* Somebody else might have raced and expanded it already */
2081 if (address
> vma
->vm_end
) {
2082 unsigned long size
, grow
;
2084 size
= address
- vma
->vm_start
;
2085 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
2088 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
2089 error
= acct_stack_growth(vma
, size
, grow
);
2092 * vma_gap_update() doesn't support concurrent
2093 * updates, but we only hold a shared mmap_sem
2094 * lock here, so we need to protect against
2095 * concurrent vma expansions.
2096 * anon_vma_lock_write() doesn't help here, as
2097 * we don't guarantee that all growable vmas
2098 * in a mm share the same root anon vma.
2099 * So, we reuse mm->page_table_lock to guard
2100 * against concurrent vma expansions.
2102 spin_lock(&mm
->page_table_lock
);
2103 if (vma
->vm_flags
& VM_LOCKED
)
2104 mm
->locked_vm
+= grow
;
2105 vm_stat_account(mm
, vma
->vm_flags
, grow
);
2106 anon_vma_interval_tree_pre_update_vma(vma
);
2107 vma
->vm_end
= address
;
2108 anon_vma_interval_tree_post_update_vma(vma
);
2110 vma_gap_update(vma
->vm_next
);
2112 mm
->highest_vm_end
= address
;
2113 spin_unlock(&mm
->page_table_lock
);
2115 perf_event_mmap(vma
);
2119 anon_vma_unlock_write(vma
->anon_vma
);
2120 khugepaged_enter_vma_merge(vma
, vma
->vm_flags
);
2124 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2127 * vma is the first one with address < vma->vm_start. Have to extend vma.
2129 int expand_downwards(struct vm_area_struct
*vma
,
2130 unsigned long address
)
2132 struct mm_struct
*mm
= vma
->vm_mm
;
2135 address
&= PAGE_MASK
;
2136 error
= security_mmap_addr(address
);
2140 /* We must make sure the anon_vma is allocated. */
2141 if (unlikely(anon_vma_prepare(vma
)))
2145 * vma->vm_start/vm_end cannot change under us because the caller
2146 * is required to hold the mmap_sem in read mode. We need the
2147 * anon_vma lock to serialize against concurrent expand_stacks.
2149 anon_vma_lock_write(vma
->anon_vma
);
2151 /* Somebody else might have raced and expanded it already */
2152 if (address
< vma
->vm_start
) {
2153 unsigned long size
, grow
;
2155 size
= vma
->vm_end
- address
;
2156 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
2159 if (grow
<= vma
->vm_pgoff
) {
2160 error
= acct_stack_growth(vma
, size
, grow
);
2163 * vma_gap_update() doesn't support concurrent
2164 * updates, but we only hold a shared mmap_sem
2165 * lock here, so we need to protect against
2166 * concurrent vma expansions.
2167 * anon_vma_lock_write() doesn't help here, as
2168 * we don't guarantee that all growable vmas
2169 * in a mm share the same root anon vma.
2170 * So, we reuse mm->page_table_lock to guard
2171 * against concurrent vma expansions.
2173 spin_lock(&mm
->page_table_lock
);
2174 if (vma
->vm_flags
& VM_LOCKED
)
2175 mm
->locked_vm
+= grow
;
2176 vm_stat_account(mm
, vma
->vm_flags
, grow
);
2177 anon_vma_interval_tree_pre_update_vma(vma
);
2178 vma
->vm_start
= address
;
2179 vma
->vm_pgoff
-= grow
;
2180 anon_vma_interval_tree_post_update_vma(vma
);
2181 vma_gap_update(vma
);
2182 spin_unlock(&mm
->page_table_lock
);
2184 perf_event_mmap(vma
);
2188 anon_vma_unlock_write(vma
->anon_vma
);
2189 khugepaged_enter_vma_merge(vma
, vma
->vm_flags
);
2195 * Note how expand_stack() refuses to expand the stack all the way to
2196 * abut the next virtual mapping, *unless* that mapping itself is also
2197 * a stack mapping. We want to leave room for a guard page, after all
2198 * (the guard page itself is not added here, that is done by the
2199 * actual page faulting logic)
2201 * This matches the behavior of the guard page logic (see mm/memory.c:
2202 * check_stack_guard_page()), which only allows the guard page to be
2203 * removed under these circumstances.
2205 #ifdef CONFIG_STACK_GROWSUP
2206 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2208 struct vm_area_struct
*next
;
2210 address
&= PAGE_MASK
;
2211 next
= vma
->vm_next
;
2212 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
) {
2213 if (!(next
->vm_flags
& VM_GROWSUP
))
2216 return expand_upwards(vma
, address
);
2219 struct vm_area_struct
*
2220 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2222 struct vm_area_struct
*vma
, *prev
;
2225 vma
= find_vma_prev(mm
, addr
, &prev
);
2226 if (vma
&& (vma
->vm_start
<= addr
))
2228 if (!prev
|| expand_stack(prev
, addr
))
2230 if (prev
->vm_flags
& VM_LOCKED
)
2231 populate_vma_page_range(prev
, addr
, prev
->vm_end
, NULL
);
2235 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2237 struct vm_area_struct
*prev
;
2239 address
&= PAGE_MASK
;
2240 prev
= vma
->vm_prev
;
2241 if (prev
&& prev
->vm_end
== address
) {
2242 if (!(prev
->vm_flags
& VM_GROWSDOWN
))
2245 return expand_downwards(vma
, address
);
2248 struct vm_area_struct
*
2249 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2251 struct vm_area_struct
*vma
;
2252 unsigned long start
;
2255 vma
= find_vma(mm
, addr
);
2258 if (vma
->vm_start
<= addr
)
2260 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2262 start
= vma
->vm_start
;
2263 if (expand_stack(vma
, addr
))
2265 if (vma
->vm_flags
& VM_LOCKED
)
2266 populate_vma_page_range(vma
, addr
, start
, NULL
);
2271 EXPORT_SYMBOL_GPL(find_extend_vma
);
2274 * Ok - we have the memory areas we should free on the vma list,
2275 * so release them, and do the vma updates.
2277 * Called with the mm semaphore held.
2279 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2281 unsigned long nr_accounted
= 0;
2283 /* Update high watermark before we lower total_vm */
2284 update_hiwater_vm(mm
);
2286 long nrpages
= vma_pages(vma
);
2288 if (vma
->vm_flags
& VM_ACCOUNT
)
2289 nr_accounted
+= nrpages
;
2290 vm_stat_account(mm
, vma
->vm_flags
, -nrpages
);
2291 vma
= remove_vma(vma
);
2293 vm_unacct_memory(nr_accounted
);
2298 * Get rid of page table information in the indicated region.
2300 * Called with the mm semaphore held.
2302 static void unmap_region(struct mm_struct
*mm
,
2303 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
2304 unsigned long start
, unsigned long end
)
2306 struct vm_area_struct
*next
= prev
? prev
->vm_next
: mm
->mmap
;
2307 struct mmu_gather tlb
;
2310 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2311 update_hiwater_rss(mm
);
2312 unmap_vmas(&tlb
, vma
, start
, end
);
2313 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
2314 next
? next
->vm_start
: USER_PGTABLES_CEILING
);
2315 tlb_finish_mmu(&tlb
, start
, end
);
2319 * Create a list of vma's touched by the unmap, removing them from the mm's
2320 * vma list as we go..
2323 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2324 struct vm_area_struct
*prev
, unsigned long end
)
2326 struct vm_area_struct
**insertion_point
;
2327 struct vm_area_struct
*tail_vma
= NULL
;
2329 insertion_point
= (prev
? &prev
->vm_next
: &mm
->mmap
);
2330 vma
->vm_prev
= NULL
;
2332 vma_rb_erase(vma
, &mm
->mm_rb
);
2336 } while (vma
&& vma
->vm_start
< end
);
2337 *insertion_point
= vma
;
2339 vma
->vm_prev
= prev
;
2340 vma_gap_update(vma
);
2342 mm
->highest_vm_end
= prev
? prev
->vm_end
: 0;
2343 tail_vma
->vm_next
= NULL
;
2345 /* Kill the cache */
2346 vmacache_invalidate(mm
);
2350 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2351 * munmap path where it doesn't make sense to fail.
2353 static int __split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2354 unsigned long addr
, int new_below
)
2356 struct vm_area_struct
*new;
2359 if (is_vm_hugetlb_page(vma
) && (addr
&
2360 ~(huge_page_mask(hstate_vma(vma
)))))
2363 new = kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2367 /* most fields are the same, copy all, and then fixup */
2370 INIT_LIST_HEAD(&new->anon_vma_chain
);
2375 new->vm_start
= addr
;
2376 new->vm_pgoff
+= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
);
2379 err
= vma_dup_policy(vma
, new);
2383 err
= anon_vma_clone(new, vma
);
2388 get_file(new->vm_file
);
2390 if (new->vm_ops
&& new->vm_ops
->open
)
2391 new->vm_ops
->open(new);
2394 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2395 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2397 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2403 /* Clean everything up if vma_adjust failed. */
2404 if (new->vm_ops
&& new->vm_ops
->close
)
2405 new->vm_ops
->close(new);
2408 unlink_anon_vmas(new);
2410 mpol_put(vma_policy(new));
2412 kmem_cache_free(vm_area_cachep
, new);
2417 * Split a vma into two pieces at address 'addr', a new vma is allocated
2418 * either for the first part or the tail.
2420 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2421 unsigned long addr
, int new_below
)
2423 if (mm
->map_count
>= sysctl_max_map_count
)
2426 return __split_vma(mm
, vma
, addr
, new_below
);
2429 /* Munmap is split into 2 main parts -- this part which finds
2430 * what needs doing, and the areas themselves, which do the
2431 * work. This now handles partial unmappings.
2432 * Jeremy Fitzhardinge <jeremy@goop.org>
2434 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
)
2437 struct vm_area_struct
*vma
, *prev
, *last
;
2439 if ((offset_in_page(start
)) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2442 len
= PAGE_ALIGN(len
);
2446 /* Find the first overlapping VMA */
2447 vma
= find_vma(mm
, start
);
2450 prev
= vma
->vm_prev
;
2451 /* we have start < vma->vm_end */
2453 /* if it doesn't overlap, we have nothing.. */
2455 if (vma
->vm_start
>= end
)
2459 * If we need to split any vma, do it now to save pain later.
2461 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2462 * unmapped vm_area_struct will remain in use: so lower split_vma
2463 * places tmp vma above, and higher split_vma places tmp vma below.
2465 if (start
> vma
->vm_start
) {
2469 * Make sure that map_count on return from munmap() will
2470 * not exceed its limit; but let map_count go just above
2471 * its limit temporarily, to help free resources as expected.
2473 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2476 error
= __split_vma(mm
, vma
, start
, 0);
2482 /* Does it split the last one? */
2483 last
= find_vma(mm
, end
);
2484 if (last
&& end
> last
->vm_start
) {
2485 int error
= __split_vma(mm
, last
, end
, 1);
2489 vma
= prev
? prev
->vm_next
: mm
->mmap
;
2492 * unlock any mlock()ed ranges before detaching vmas
2494 if (mm
->locked_vm
) {
2495 struct vm_area_struct
*tmp
= vma
;
2496 while (tmp
&& tmp
->vm_start
< end
) {
2497 if (tmp
->vm_flags
& VM_LOCKED
) {
2498 mm
->locked_vm
-= vma_pages(tmp
);
2499 munlock_vma_pages_all(tmp
);
2506 * Remove the vma's, and unmap the actual pages
2508 detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
);
2509 unmap_region(mm
, vma
, prev
, start
, end
);
2511 arch_unmap(mm
, vma
, start
, end
);
2513 /* Fix up all other VM information */
2514 remove_vma_list(mm
, vma
);
2519 int vm_munmap(unsigned long start
, size_t len
)
2522 struct mm_struct
*mm
= current
->mm
;
2524 if (down_write_killable(&mm
->mmap_sem
))
2527 ret
= do_munmap(mm
, start
, len
);
2528 up_write(&mm
->mmap_sem
);
2531 EXPORT_SYMBOL(vm_munmap
);
2533 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2536 struct mm_struct
*mm
= current
->mm
;
2538 profile_munmap(addr
);
2539 if (down_write_killable(&mm
->mmap_sem
))
2541 ret
= do_munmap(mm
, addr
, len
);
2542 up_write(&mm
->mmap_sem
);
2548 * Emulation of deprecated remap_file_pages() syscall.
2550 SYSCALL_DEFINE5(remap_file_pages
, unsigned long, start
, unsigned long, size
,
2551 unsigned long, prot
, unsigned long, pgoff
, unsigned long, flags
)
2554 struct mm_struct
*mm
= current
->mm
;
2555 struct vm_area_struct
*vma
;
2556 unsigned long populate
= 0;
2557 unsigned long ret
= -EINVAL
;
2560 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2561 current
->comm
, current
->pid
);
2565 start
= start
& PAGE_MASK
;
2566 size
= size
& PAGE_MASK
;
2568 if (start
+ size
<= start
)
2571 /* Does pgoff wrap? */
2572 if (pgoff
+ (size
>> PAGE_SHIFT
) < pgoff
)
2575 if (down_write_killable(&mm
->mmap_sem
))
2578 vma
= find_vma(mm
, start
);
2580 if (!vma
|| !(vma
->vm_flags
& VM_SHARED
))
2583 if (start
< vma
->vm_start
)
2586 if (start
+ size
> vma
->vm_end
) {
2587 struct vm_area_struct
*next
;
2589 for (next
= vma
->vm_next
; next
; next
= next
->vm_next
) {
2590 /* hole between vmas ? */
2591 if (next
->vm_start
!= next
->vm_prev
->vm_end
)
2594 if (next
->vm_file
!= vma
->vm_file
)
2597 if (next
->vm_flags
!= vma
->vm_flags
)
2600 if (start
+ size
<= next
->vm_end
)
2608 prot
|= vma
->vm_flags
& VM_READ
? PROT_READ
: 0;
2609 prot
|= vma
->vm_flags
& VM_WRITE
? PROT_WRITE
: 0;
2610 prot
|= vma
->vm_flags
& VM_EXEC
? PROT_EXEC
: 0;
2612 flags
&= MAP_NONBLOCK
;
2613 flags
|= MAP_SHARED
| MAP_FIXED
| MAP_POPULATE
;
2614 if (vma
->vm_flags
& VM_LOCKED
) {
2615 struct vm_area_struct
*tmp
;
2616 flags
|= MAP_LOCKED
;
2618 /* drop PG_Mlocked flag for over-mapped range */
2619 for (tmp
= vma
; tmp
->vm_start
>= start
+ size
;
2620 tmp
= tmp
->vm_next
) {
2622 * Split pmd and munlock page on the border
2625 vma_adjust_trans_huge(tmp
, start
, start
+ size
, 0);
2627 munlock_vma_pages_range(tmp
,
2628 max(tmp
->vm_start
, start
),
2629 min(tmp
->vm_end
, start
+ size
));
2633 file
= get_file(vma
->vm_file
);
2634 ret
= do_mmap_pgoff(vma
->vm_file
, start
, size
,
2635 prot
, flags
, pgoff
, &populate
);
2638 up_write(&mm
->mmap_sem
);
2640 mm_populate(ret
, populate
);
2641 if (!IS_ERR_VALUE(ret
))
2646 static inline void verify_mm_writelocked(struct mm_struct
*mm
)
2648 #ifdef CONFIG_DEBUG_VM
2649 if (unlikely(down_read_trylock(&mm
->mmap_sem
))) {
2651 up_read(&mm
->mmap_sem
);
2657 * this is really a simplified "do_mmap". it only handles
2658 * anonymous maps. eventually we may be able to do some
2659 * brk-specific accounting here.
2661 static int do_brk(unsigned long addr
, unsigned long request
)
2663 struct mm_struct
*mm
= current
->mm
;
2664 struct vm_area_struct
*vma
, *prev
;
2665 unsigned long flags
, len
;
2666 struct rb_node
**rb_link
, *rb_parent
;
2667 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2670 len
= PAGE_ALIGN(request
);
2676 flags
= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2678 error
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2679 if (offset_in_page(error
))
2682 error
= mlock_future_check(mm
, mm
->def_flags
, len
);
2687 * mm->mmap_sem is required to protect against another thread
2688 * changing the mappings in case we sleep.
2690 verify_mm_writelocked(mm
);
2693 * Clear old maps. this also does some error checking for us
2695 while (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
,
2697 if (do_munmap(mm
, addr
, len
))
2701 /* Check against address space limits *after* clearing old maps... */
2702 if (!may_expand_vm(mm
, flags
, len
>> PAGE_SHIFT
))
2705 if (mm
->map_count
> sysctl_max_map_count
)
2708 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
2711 /* Can we just expand an old private anonymous mapping? */
2712 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
2713 NULL
, NULL
, pgoff
, NULL
, NULL_VM_UFFD_CTX
);
2718 * create a vma struct for an anonymous mapping
2720 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2722 vm_unacct_memory(len
>> PAGE_SHIFT
);
2726 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2728 vma
->vm_start
= addr
;
2729 vma
->vm_end
= addr
+ len
;
2730 vma
->vm_pgoff
= pgoff
;
2731 vma
->vm_flags
= flags
;
2732 vma
->vm_page_prot
= vm_get_page_prot(flags
);
2733 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2735 perf_event_mmap(vma
);
2736 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2737 mm
->data_vm
+= len
>> PAGE_SHIFT
;
2738 if (flags
& VM_LOCKED
)
2739 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
2740 vma
->vm_flags
|= VM_SOFTDIRTY
;
2744 int vm_brk(unsigned long addr
, unsigned long len
)
2746 struct mm_struct
*mm
= current
->mm
;
2750 if (down_write_killable(&mm
->mmap_sem
))
2753 ret
= do_brk(addr
, len
);
2754 populate
= ((mm
->def_flags
& VM_LOCKED
) != 0);
2755 up_write(&mm
->mmap_sem
);
2756 if (populate
&& !ret
)
2757 mm_populate(addr
, len
);
2760 EXPORT_SYMBOL(vm_brk
);
2762 /* Release all mmaps. */
2763 void exit_mmap(struct mm_struct
*mm
)
2765 struct mmu_gather tlb
;
2766 struct vm_area_struct
*vma
;
2767 unsigned long nr_accounted
= 0;
2769 /* mm's last user has gone, and its about to be pulled down */
2770 mmu_notifier_release(mm
);
2772 if (mm
->locked_vm
) {
2775 if (vma
->vm_flags
& VM_LOCKED
)
2776 munlock_vma_pages_all(vma
);
2784 if (!vma
) /* Can happen if dup_mmap() received an OOM */
2789 tlb_gather_mmu(&tlb
, mm
, 0, -1);
2790 /* update_hiwater_rss(mm) here? but nobody should be looking */
2791 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2792 unmap_vmas(&tlb
, vma
, 0, -1);
2794 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, USER_PGTABLES_CEILING
);
2795 tlb_finish_mmu(&tlb
, 0, -1);
2798 * Walk the list again, actually closing and freeing it,
2799 * with preemption enabled, without holding any MM locks.
2802 if (vma
->vm_flags
& VM_ACCOUNT
)
2803 nr_accounted
+= vma_pages(vma
);
2804 vma
= remove_vma(vma
);
2806 vm_unacct_memory(nr_accounted
);
2809 /* Insert vm structure into process list sorted by address
2810 * and into the inode's i_mmap tree. If vm_file is non-NULL
2811 * then i_mmap_rwsem is taken here.
2813 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2815 struct vm_area_struct
*prev
;
2816 struct rb_node
**rb_link
, *rb_parent
;
2818 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
2819 &prev
, &rb_link
, &rb_parent
))
2821 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
2822 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
2826 * The vm_pgoff of a purely anonymous vma should be irrelevant
2827 * until its first write fault, when page's anon_vma and index
2828 * are set. But now set the vm_pgoff it will almost certainly
2829 * end up with (unless mremap moves it elsewhere before that
2830 * first wfault), so /proc/pid/maps tells a consistent story.
2832 * By setting it to reflect the virtual start address of the
2833 * vma, merges and splits can happen in a seamless way, just
2834 * using the existing file pgoff checks and manipulations.
2835 * Similarly in do_mmap_pgoff and in do_brk.
2837 if (vma_is_anonymous(vma
)) {
2838 BUG_ON(vma
->anon_vma
);
2839 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2842 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2847 * Copy the vma structure to a new location in the same mm,
2848 * prior to moving page table entries, to effect an mremap move.
2850 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
2851 unsigned long addr
, unsigned long len
, pgoff_t pgoff
,
2852 bool *need_rmap_locks
)
2854 struct vm_area_struct
*vma
= *vmap
;
2855 unsigned long vma_start
= vma
->vm_start
;
2856 struct mm_struct
*mm
= vma
->vm_mm
;
2857 struct vm_area_struct
*new_vma
, *prev
;
2858 struct rb_node
**rb_link
, *rb_parent
;
2859 bool faulted_in_anon_vma
= true;
2862 * If anonymous vma has not yet been faulted, update new pgoff
2863 * to match new location, to increase its chance of merging.
2865 if (unlikely(vma_is_anonymous(vma
) && !vma
->anon_vma
)) {
2866 pgoff
= addr
>> PAGE_SHIFT
;
2867 faulted_in_anon_vma
= false;
2870 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
2871 return NULL
; /* should never get here */
2872 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
2873 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
),
2874 vma
->vm_userfaultfd_ctx
);
2877 * Source vma may have been merged into new_vma
2879 if (unlikely(vma_start
>= new_vma
->vm_start
&&
2880 vma_start
< new_vma
->vm_end
)) {
2882 * The only way we can get a vma_merge with
2883 * self during an mremap is if the vma hasn't
2884 * been faulted in yet and we were allowed to
2885 * reset the dst vma->vm_pgoff to the
2886 * destination address of the mremap to allow
2887 * the merge to happen. mremap must change the
2888 * vm_pgoff linearity between src and dst vmas
2889 * (in turn preventing a vma_merge) to be
2890 * safe. It is only safe to keep the vm_pgoff
2891 * linear if there are no pages mapped yet.
2893 VM_BUG_ON_VMA(faulted_in_anon_vma
, new_vma
);
2894 *vmap
= vma
= new_vma
;
2896 *need_rmap_locks
= (new_vma
->vm_pgoff
<= vma
->vm_pgoff
);
2898 new_vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2902 new_vma
->vm_start
= addr
;
2903 new_vma
->vm_end
= addr
+ len
;
2904 new_vma
->vm_pgoff
= pgoff
;
2905 if (vma_dup_policy(vma
, new_vma
))
2907 INIT_LIST_HEAD(&new_vma
->anon_vma_chain
);
2908 if (anon_vma_clone(new_vma
, vma
))
2909 goto out_free_mempol
;
2910 if (new_vma
->vm_file
)
2911 get_file(new_vma
->vm_file
);
2912 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
2913 new_vma
->vm_ops
->open(new_vma
);
2914 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
2915 *need_rmap_locks
= false;
2920 mpol_put(vma_policy(new_vma
));
2922 kmem_cache_free(vm_area_cachep
, new_vma
);
2928 * Return true if the calling process may expand its vm space by the passed
2931 bool may_expand_vm(struct mm_struct
*mm
, vm_flags_t flags
, unsigned long npages
)
2933 if (mm
->total_vm
+ npages
> rlimit(RLIMIT_AS
) >> PAGE_SHIFT
)
2936 if (is_data_mapping(flags
) &&
2937 mm
->data_vm
+ npages
> rlimit(RLIMIT_DATA
) >> PAGE_SHIFT
) {
2938 /* Workaround for Valgrind */
2939 if (rlimit(RLIMIT_DATA
) == 0 &&
2940 mm
->data_vm
+ npages
<= rlimit_max(RLIMIT_DATA
) >> PAGE_SHIFT
)
2942 if (!ignore_rlimit_data
) {
2943 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits or use boot option ignore_rlimit_data.\n",
2944 current
->comm
, current
->pid
,
2945 (mm
->data_vm
+ npages
) << PAGE_SHIFT
,
2946 rlimit(RLIMIT_DATA
));
2954 void vm_stat_account(struct mm_struct
*mm
, vm_flags_t flags
, long npages
)
2956 mm
->total_vm
+= npages
;
2958 if (is_exec_mapping(flags
))
2959 mm
->exec_vm
+= npages
;
2960 else if (is_stack_mapping(flags
))
2961 mm
->stack_vm
+= npages
;
2962 else if (is_data_mapping(flags
))
2963 mm
->data_vm
+= npages
;
2966 static int special_mapping_fault(struct vm_area_struct
*vma
,
2967 struct vm_fault
*vmf
);
2970 * Having a close hook prevents vma merging regardless of flags.
2972 static void special_mapping_close(struct vm_area_struct
*vma
)
2976 static const char *special_mapping_name(struct vm_area_struct
*vma
)
2978 return ((struct vm_special_mapping
*)vma
->vm_private_data
)->name
;
2981 static int special_mapping_mremap(struct vm_area_struct
*new_vma
)
2983 struct vm_special_mapping
*sm
= new_vma
->vm_private_data
;
2986 return sm
->mremap(sm
, new_vma
);
2990 static const struct vm_operations_struct special_mapping_vmops
= {
2991 .close
= special_mapping_close
,
2992 .fault
= special_mapping_fault
,
2993 .mremap
= special_mapping_mremap
,
2994 .name
= special_mapping_name
,
2997 static const struct vm_operations_struct legacy_special_mapping_vmops
= {
2998 .close
= special_mapping_close
,
2999 .fault
= special_mapping_fault
,
3002 static int special_mapping_fault(struct vm_area_struct
*vma
,
3003 struct vm_fault
*vmf
)
3006 struct page
**pages
;
3008 if (vma
->vm_ops
== &legacy_special_mapping_vmops
) {
3009 pages
= vma
->vm_private_data
;
3011 struct vm_special_mapping
*sm
= vma
->vm_private_data
;
3014 return sm
->fault(sm
, vma
, vmf
);
3019 for (pgoff
= vmf
->pgoff
; pgoff
&& *pages
; ++pages
)
3023 struct page
*page
= *pages
;
3029 return VM_FAULT_SIGBUS
;
3032 static struct vm_area_struct
*__install_special_mapping(
3033 struct mm_struct
*mm
,
3034 unsigned long addr
, unsigned long len
,
3035 unsigned long vm_flags
, void *priv
,
3036 const struct vm_operations_struct
*ops
)
3039 struct vm_area_struct
*vma
;
3041 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
3042 if (unlikely(vma
== NULL
))
3043 return ERR_PTR(-ENOMEM
);
3045 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
3047 vma
->vm_start
= addr
;
3048 vma
->vm_end
= addr
+ len
;
3050 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
| VM_SOFTDIRTY
;
3051 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
3054 vma
->vm_private_data
= priv
;
3056 ret
= insert_vm_struct(mm
, vma
);
3060 vm_stat_account(mm
, vma
->vm_flags
, len
>> PAGE_SHIFT
);
3062 perf_event_mmap(vma
);
3067 kmem_cache_free(vm_area_cachep
, vma
);
3068 return ERR_PTR(ret
);
3071 bool vma_is_special_mapping(const struct vm_area_struct
*vma
,
3072 const struct vm_special_mapping
*sm
)
3074 return vma
->vm_private_data
== sm
&&
3075 (vma
->vm_ops
== &special_mapping_vmops
||
3076 vma
->vm_ops
== &legacy_special_mapping_vmops
);
3080 * Called with mm->mmap_sem held for writing.
3081 * Insert a new vma covering the given region, with the given flags.
3082 * Its pages are supplied by the given array of struct page *.
3083 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3084 * The region past the last page supplied will always produce SIGBUS.
3085 * The array pointer and the pages it points to are assumed to stay alive
3086 * for as long as this mapping might exist.
3088 struct vm_area_struct
*_install_special_mapping(
3089 struct mm_struct
*mm
,
3090 unsigned long addr
, unsigned long len
,
3091 unsigned long vm_flags
, const struct vm_special_mapping
*spec
)
3093 return __install_special_mapping(mm
, addr
, len
, vm_flags
, (void *)spec
,
3094 &special_mapping_vmops
);
3097 int install_special_mapping(struct mm_struct
*mm
,
3098 unsigned long addr
, unsigned long len
,
3099 unsigned long vm_flags
, struct page
**pages
)
3101 struct vm_area_struct
*vma
= __install_special_mapping(
3102 mm
, addr
, len
, vm_flags
, (void *)pages
,
3103 &legacy_special_mapping_vmops
);
3105 return PTR_ERR_OR_ZERO(vma
);
3108 static DEFINE_MUTEX(mm_all_locks_mutex
);
3110 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
3112 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3114 * The LSB of head.next can't change from under us
3115 * because we hold the mm_all_locks_mutex.
3117 down_write_nest_lock(&anon_vma
->root
->rwsem
, &mm
->mmap_sem
);
3119 * We can safely modify head.next after taking the
3120 * anon_vma->root->rwsem. If some other vma in this mm shares
3121 * the same anon_vma we won't take it again.
3123 * No need of atomic instructions here, head.next
3124 * can't change from under us thanks to the
3125 * anon_vma->root->rwsem.
3127 if (__test_and_set_bit(0, (unsigned long *)
3128 &anon_vma
->root
->rb_root
.rb_node
))
3133 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
3135 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3137 * AS_MM_ALL_LOCKS can't change from under us because
3138 * we hold the mm_all_locks_mutex.
3140 * Operations on ->flags have to be atomic because
3141 * even if AS_MM_ALL_LOCKS is stable thanks to the
3142 * mm_all_locks_mutex, there may be other cpus
3143 * changing other bitflags in parallel to us.
3145 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
3147 down_write_nest_lock(&mapping
->i_mmap_rwsem
, &mm
->mmap_sem
);
3152 * This operation locks against the VM for all pte/vma/mm related
3153 * operations that could ever happen on a certain mm. This includes
3154 * vmtruncate, try_to_unmap, and all page faults.
3156 * The caller must take the mmap_sem in write mode before calling
3157 * mm_take_all_locks(). The caller isn't allowed to release the
3158 * mmap_sem until mm_drop_all_locks() returns.
3160 * mmap_sem in write mode is required in order to block all operations
3161 * that could modify pagetables and free pages without need of
3162 * altering the vma layout. It's also needed in write mode to avoid new
3163 * anon_vmas to be associated with existing vmas.
3165 * A single task can't take more than one mm_take_all_locks() in a row
3166 * or it would deadlock.
3168 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3169 * mapping->flags avoid to take the same lock twice, if more than one
3170 * vma in this mm is backed by the same anon_vma or address_space.
3172 * We take locks in following order, accordingly to comment at beginning
3174 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3176 * - all i_mmap_rwsem locks;
3177 * - all anon_vma->rwseml
3179 * We can take all locks within these types randomly because the VM code
3180 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3181 * mm_all_locks_mutex.
3183 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3184 * that may have to take thousand of locks.
3186 * mm_take_all_locks() can fail if it's interrupted by signals.
3188 int mm_take_all_locks(struct mm_struct
*mm
)
3190 struct vm_area_struct
*vma
;
3191 struct anon_vma_chain
*avc
;
3193 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3195 mutex_lock(&mm_all_locks_mutex
);
3197 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3198 if (signal_pending(current
))
3200 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
3201 is_vm_hugetlb_page(vma
))
3202 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3205 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3206 if (signal_pending(current
))
3208 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
3209 !is_vm_hugetlb_page(vma
))
3210 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3213 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3214 if (signal_pending(current
))
3217 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3218 vm_lock_anon_vma(mm
, avc
->anon_vma
);
3224 mm_drop_all_locks(mm
);
3228 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
3230 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3232 * The LSB of head.next can't change to 0 from under
3233 * us because we hold the mm_all_locks_mutex.
3235 * We must however clear the bitflag before unlocking
3236 * the vma so the users using the anon_vma->rb_root will
3237 * never see our bitflag.
3239 * No need of atomic instructions here, head.next
3240 * can't change from under us until we release the
3241 * anon_vma->root->rwsem.
3243 if (!__test_and_clear_bit(0, (unsigned long *)
3244 &anon_vma
->root
->rb_root
.rb_node
))
3246 anon_vma_unlock_write(anon_vma
);
3250 static void vm_unlock_mapping(struct address_space
*mapping
)
3252 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3254 * AS_MM_ALL_LOCKS can't change to 0 from under us
3255 * because we hold the mm_all_locks_mutex.
3257 i_mmap_unlock_write(mapping
);
3258 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
3265 * The mmap_sem cannot be released by the caller until
3266 * mm_drop_all_locks() returns.
3268 void mm_drop_all_locks(struct mm_struct
*mm
)
3270 struct vm_area_struct
*vma
;
3271 struct anon_vma_chain
*avc
;
3273 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3274 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
3276 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3278 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3279 vm_unlock_anon_vma(avc
->anon_vma
);
3280 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3281 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
3284 mutex_unlock(&mm_all_locks_mutex
);
3288 * initialise the VMA slab
3290 void __init
mmap_init(void)
3294 ret
= percpu_counter_init(&vm_committed_as
, 0, GFP_KERNEL
);
3299 * Initialise sysctl_user_reserve_kbytes.
3301 * This is intended to prevent a user from starting a single memory hogging
3302 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3305 * The default value is min(3% of free memory, 128MB)
3306 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3308 static int init_user_reserve(void)
3310 unsigned long free_kbytes
;
3312 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3314 sysctl_user_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 17);
3317 subsys_initcall(init_user_reserve
);
3320 * Initialise sysctl_admin_reserve_kbytes.
3322 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3323 * to log in and kill a memory hogging process.
3325 * Systems with more than 256MB will reserve 8MB, enough to recover
3326 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3327 * only reserve 3% of free pages by default.
3329 static int init_admin_reserve(void)
3331 unsigned long free_kbytes
;
3333 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3335 sysctl_admin_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 13);
3338 subsys_initcall(init_admin_reserve
);
3341 * Reinititalise user and admin reserves if memory is added or removed.
3343 * The default user reserve max is 128MB, and the default max for the
3344 * admin reserve is 8MB. These are usually, but not always, enough to
3345 * enable recovery from a memory hogging process using login/sshd, a shell,
3346 * and tools like top. It may make sense to increase or even disable the
3347 * reserve depending on the existence of swap or variations in the recovery
3348 * tools. So, the admin may have changed them.
3350 * If memory is added and the reserves have been eliminated or increased above
3351 * the default max, then we'll trust the admin.
3353 * If memory is removed and there isn't enough free memory, then we
3354 * need to reset the reserves.
3356 * Otherwise keep the reserve set by the admin.
3358 static int reserve_mem_notifier(struct notifier_block
*nb
,
3359 unsigned long action
, void *data
)
3361 unsigned long tmp
, free_kbytes
;
3365 /* Default max is 128MB. Leave alone if modified by operator. */
3366 tmp
= sysctl_user_reserve_kbytes
;
3367 if (0 < tmp
&& tmp
< (1UL << 17))
3368 init_user_reserve();
3370 /* Default max is 8MB. Leave alone if modified by operator. */
3371 tmp
= sysctl_admin_reserve_kbytes
;
3372 if (0 < tmp
&& tmp
< (1UL << 13))
3373 init_admin_reserve();
3377 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3379 if (sysctl_user_reserve_kbytes
> free_kbytes
) {
3380 init_user_reserve();
3381 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3382 sysctl_user_reserve_kbytes
);
3385 if (sysctl_admin_reserve_kbytes
> free_kbytes
) {
3386 init_admin_reserve();
3387 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3388 sysctl_admin_reserve_kbytes
);
3397 static struct notifier_block reserve_mem_nb
= {
3398 .notifier_call
= reserve_mem_notifier
,
3401 static int __meminit
init_reserve_notifier(void)
3403 if (register_hotmemory_notifier(&reserve_mem_nb
))
3404 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3408 subsys_initcall(init_reserve_notifier
);