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/profile.h>
29 #include <linux/export.h>
30 #include <linux/mount.h>
31 #include <linux/mempolicy.h>
32 #include <linux/rmap.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/mmdebug.h>
35 #include <linux/perf_event.h>
36 #include <linux/audit.h>
37 #include <linux/khugepaged.h>
38 #include <linux/uprobes.h>
39 #include <linux/rbtree_augmented.h>
40 #include <linux/notifier.h>
41 #include <linux/memory.h>
42 #include <linux/printk.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/moduleparam.h>
46 #include <asm/uaccess.h>
47 #include <asm/cacheflush.h>
49 #include <asm/mmu_context.h>
53 #ifndef arch_mmap_check
54 #define arch_mmap_check(addr, len, flags) (0)
57 #ifndef arch_rebalance_pgtables
58 #define arch_rebalance_pgtables(addr, len) (addr)
61 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
62 const int mmap_rnd_bits_min
= CONFIG_ARCH_MMAP_RND_BITS_MIN
;
63 const int mmap_rnd_bits_max
= CONFIG_ARCH_MMAP_RND_BITS_MAX
;
64 int mmap_rnd_bits __read_mostly
= CONFIG_ARCH_MMAP_RND_BITS
;
66 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
67 const int mmap_rnd_compat_bits_min
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN
;
68 const int mmap_rnd_compat_bits_max
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX
;
69 int mmap_rnd_compat_bits __read_mostly
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS
;
72 static bool ignore_rlimit_data
= true;
73 core_param(ignore_rlimit_data
, ignore_rlimit_data
, bool, 0644);
75 static void unmap_region(struct mm_struct
*mm
,
76 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
77 unsigned long start
, unsigned long end
);
79 /* description of effects of mapping type and prot in current implementation.
80 * this is due to the limited x86 page protection hardware. The expected
81 * behavior is in parens:
84 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
85 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
86 * w: (no) no w: (no) no w: (yes) yes w: (no) no
87 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
89 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
90 * w: (no) no w: (no) no w: (copy) copy w: (no) no
91 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
94 pgprot_t protection_map
[16] = {
95 __P000
, __P001
, __P010
, __P011
, __P100
, __P101
, __P110
, __P111
,
96 __S000
, __S001
, __S010
, __S011
, __S100
, __S101
, __S110
, __S111
99 pgprot_t
vm_get_page_prot(unsigned long vm_flags
)
101 return __pgprot(pgprot_val(protection_map
[vm_flags
&
102 (VM_READ
|VM_WRITE
|VM_EXEC
|VM_SHARED
)]) |
103 pgprot_val(arch_vm_get_page_prot(vm_flags
)));
105 EXPORT_SYMBOL(vm_get_page_prot
);
107 static pgprot_t
vm_pgprot_modify(pgprot_t oldprot
, unsigned long vm_flags
)
109 return pgprot_modify(oldprot
, vm_get_page_prot(vm_flags
));
112 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
113 void vma_set_page_prot(struct vm_area_struct
*vma
)
115 unsigned long vm_flags
= vma
->vm_flags
;
117 vma
->vm_page_prot
= vm_pgprot_modify(vma
->vm_page_prot
, vm_flags
);
118 if (vma_wants_writenotify(vma
)) {
119 vm_flags
&= ~VM_SHARED
;
120 vma
->vm_page_prot
= vm_pgprot_modify(vma
->vm_page_prot
,
126 * Requires inode->i_mapping->i_mmap_rwsem
128 static void __remove_shared_vm_struct(struct vm_area_struct
*vma
,
129 struct file
*file
, struct address_space
*mapping
)
131 if (vma
->vm_flags
& VM_DENYWRITE
)
132 atomic_inc(&file_inode(file
)->i_writecount
);
133 if (vma
->vm_flags
& VM_SHARED
)
134 mapping_unmap_writable(mapping
);
136 flush_dcache_mmap_lock(mapping
);
137 vma_interval_tree_remove(vma
, &mapping
->i_mmap
);
138 flush_dcache_mmap_unlock(mapping
);
142 * Unlink a file-based vm structure from its interval tree, to hide
143 * vma from rmap and vmtruncate before freeing its page tables.
145 void unlink_file_vma(struct vm_area_struct
*vma
)
147 struct file
*file
= vma
->vm_file
;
150 struct address_space
*mapping
= file
->f_mapping
;
151 i_mmap_lock_write(mapping
);
152 __remove_shared_vm_struct(vma
, file
, mapping
);
153 i_mmap_unlock_write(mapping
);
158 * Close a vm structure and free it, returning the next.
160 static struct vm_area_struct
*remove_vma(struct vm_area_struct
*vma
)
162 struct vm_area_struct
*next
= vma
->vm_next
;
165 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
166 vma
->vm_ops
->close(vma
);
169 mpol_put(vma_policy(vma
));
170 kmem_cache_free(vm_area_cachep
, vma
);
174 static unsigned long do_brk(unsigned long addr
, unsigned long len
);
176 SYSCALL_DEFINE1(brk
, unsigned long, brk
)
178 unsigned long retval
;
179 unsigned long newbrk
, oldbrk
;
180 struct mm_struct
*mm
= current
->mm
;
181 unsigned long min_brk
;
184 down_write(&mm
->mmap_sem
);
186 #ifdef CONFIG_COMPAT_BRK
188 * CONFIG_COMPAT_BRK can still be overridden by setting
189 * randomize_va_space to 2, which will still cause mm->start_brk
190 * to be arbitrarily shifted
192 if (current
->brk_randomized
)
193 min_brk
= mm
->start_brk
;
195 min_brk
= mm
->end_data
;
197 min_brk
= mm
->start_brk
;
203 * Check against rlimit here. If this check is done later after the test
204 * of oldbrk with newbrk then it can escape the test and let the data
205 * segment grow beyond its set limit the in case where the limit is
206 * not page aligned -Ram Gupta
208 if (check_data_rlimit(rlimit(RLIMIT_DATA
), brk
, mm
->start_brk
,
209 mm
->end_data
, mm
->start_data
))
212 newbrk
= PAGE_ALIGN(brk
);
213 oldbrk
= PAGE_ALIGN(mm
->brk
);
214 if (oldbrk
== newbrk
)
217 /* Always allow shrinking brk. */
218 if (brk
<= mm
->brk
) {
219 if (!do_munmap(mm
, newbrk
, oldbrk
-newbrk
))
224 /* Check against existing mmap mappings. */
225 if (find_vma_intersection(mm
, oldbrk
, newbrk
+PAGE_SIZE
))
228 /* Ok, looks good - let it rip. */
229 if (do_brk(oldbrk
, newbrk
-oldbrk
) != oldbrk
)
234 populate
= newbrk
> oldbrk
&& (mm
->def_flags
& VM_LOCKED
) != 0;
235 up_write(&mm
->mmap_sem
);
237 mm_populate(oldbrk
, newbrk
- oldbrk
);
242 up_write(&mm
->mmap_sem
);
246 static long vma_compute_subtree_gap(struct vm_area_struct
*vma
)
248 unsigned long max
, subtree_gap
;
251 max
-= vma
->vm_prev
->vm_end
;
252 if (vma
->vm_rb
.rb_left
) {
253 subtree_gap
= rb_entry(vma
->vm_rb
.rb_left
,
254 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
255 if (subtree_gap
> max
)
258 if (vma
->vm_rb
.rb_right
) {
259 subtree_gap
= rb_entry(vma
->vm_rb
.rb_right
,
260 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
261 if (subtree_gap
> max
)
267 #ifdef CONFIG_DEBUG_VM_RB
268 static int browse_rb(struct mm_struct
*mm
)
270 struct rb_root
*root
= &mm
->mm_rb
;
271 int i
= 0, j
, bug
= 0;
272 struct rb_node
*nd
, *pn
= NULL
;
273 unsigned long prev
= 0, pend
= 0;
275 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
276 struct vm_area_struct
*vma
;
277 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
278 if (vma
->vm_start
< prev
) {
279 pr_emerg("vm_start %lx < prev %lx\n",
280 vma
->vm_start
, prev
);
283 if (vma
->vm_start
< pend
) {
284 pr_emerg("vm_start %lx < pend %lx\n",
285 vma
->vm_start
, pend
);
288 if (vma
->vm_start
> vma
->vm_end
) {
289 pr_emerg("vm_start %lx > vm_end %lx\n",
290 vma
->vm_start
, vma
->vm_end
);
293 spin_lock(&mm
->page_table_lock
);
294 if (vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
)) {
295 pr_emerg("free gap %lx, correct %lx\n",
297 vma_compute_subtree_gap(vma
));
300 spin_unlock(&mm
->page_table_lock
);
303 prev
= vma
->vm_start
;
307 for (nd
= pn
; nd
; nd
= rb_prev(nd
))
310 pr_emerg("backwards %d, forwards %d\n", j
, i
);
316 static void validate_mm_rb(struct rb_root
*root
, struct vm_area_struct
*ignore
)
320 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
321 struct vm_area_struct
*vma
;
322 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
323 VM_BUG_ON_VMA(vma
!= ignore
&&
324 vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
),
329 static void validate_mm(struct mm_struct
*mm
)
333 unsigned long highest_address
= 0;
334 struct vm_area_struct
*vma
= mm
->mmap
;
337 struct anon_vma
*anon_vma
= vma
->anon_vma
;
338 struct anon_vma_chain
*avc
;
341 anon_vma_lock_read(anon_vma
);
342 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
343 anon_vma_interval_tree_verify(avc
);
344 anon_vma_unlock_read(anon_vma
);
347 highest_address
= vma
->vm_end
;
351 if (i
!= mm
->map_count
) {
352 pr_emerg("map_count %d vm_next %d\n", mm
->map_count
, i
);
355 if (highest_address
!= mm
->highest_vm_end
) {
356 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
357 mm
->highest_vm_end
, highest_address
);
361 if (i
!= mm
->map_count
) {
363 pr_emerg("map_count %d rb %d\n", mm
->map_count
, i
);
366 VM_BUG_ON_MM(bug
, mm
);
369 #define validate_mm_rb(root, ignore) do { } while (0)
370 #define validate_mm(mm) do { } while (0)
373 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks
, struct vm_area_struct
, vm_rb
,
374 unsigned long, rb_subtree_gap
, vma_compute_subtree_gap
)
377 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
378 * vma->vm_prev->vm_end values changed, without modifying the vma's position
381 static void vma_gap_update(struct vm_area_struct
*vma
)
384 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
385 * function that does exacltly what we want.
387 vma_gap_callbacks_propagate(&vma
->vm_rb
, NULL
);
390 static inline void vma_rb_insert(struct vm_area_struct
*vma
,
391 struct rb_root
*root
)
393 /* All rb_subtree_gap values must be consistent prior to insertion */
394 validate_mm_rb(root
, NULL
);
396 rb_insert_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
399 static void vma_rb_erase(struct vm_area_struct
*vma
, struct rb_root
*root
)
402 * All rb_subtree_gap values must be consistent prior to erase,
403 * with the possible exception of the vma being erased.
405 validate_mm_rb(root
, vma
);
408 * Note rb_erase_augmented is a fairly large inline function,
409 * so make sure we instantiate it only once with our desired
410 * augmented rbtree callbacks.
412 rb_erase_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
416 * vma has some anon_vma assigned, and is already inserted on that
417 * anon_vma's interval trees.
419 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
420 * vma must be removed from the anon_vma's interval trees using
421 * anon_vma_interval_tree_pre_update_vma().
423 * After the update, the vma will be reinserted using
424 * anon_vma_interval_tree_post_update_vma().
426 * The entire update must be protected by exclusive mmap_sem and by
427 * the root anon_vma's mutex.
430 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct
*vma
)
432 struct anon_vma_chain
*avc
;
434 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
435 anon_vma_interval_tree_remove(avc
, &avc
->anon_vma
->rb_root
);
439 anon_vma_interval_tree_post_update_vma(struct vm_area_struct
*vma
)
441 struct anon_vma_chain
*avc
;
443 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
444 anon_vma_interval_tree_insert(avc
, &avc
->anon_vma
->rb_root
);
447 static int find_vma_links(struct mm_struct
*mm
, unsigned long addr
,
448 unsigned long end
, struct vm_area_struct
**pprev
,
449 struct rb_node
***rb_link
, struct rb_node
**rb_parent
)
451 struct rb_node
**__rb_link
, *__rb_parent
, *rb_prev
;
453 __rb_link
= &mm
->mm_rb
.rb_node
;
454 rb_prev
= __rb_parent
= NULL
;
457 struct vm_area_struct
*vma_tmp
;
459 __rb_parent
= *__rb_link
;
460 vma_tmp
= rb_entry(__rb_parent
, struct vm_area_struct
, vm_rb
);
462 if (vma_tmp
->vm_end
> addr
) {
463 /* Fail if an existing vma overlaps the area */
464 if (vma_tmp
->vm_start
< end
)
466 __rb_link
= &__rb_parent
->rb_left
;
468 rb_prev
= __rb_parent
;
469 __rb_link
= &__rb_parent
->rb_right
;
475 *pprev
= rb_entry(rb_prev
, struct vm_area_struct
, vm_rb
);
476 *rb_link
= __rb_link
;
477 *rb_parent
= __rb_parent
;
481 static unsigned long count_vma_pages_range(struct mm_struct
*mm
,
482 unsigned long addr
, unsigned long end
)
484 unsigned long nr_pages
= 0;
485 struct vm_area_struct
*vma
;
487 /* Find first overlaping mapping */
488 vma
= find_vma_intersection(mm
, addr
, end
);
492 nr_pages
= (min(end
, vma
->vm_end
) -
493 max(addr
, vma
->vm_start
)) >> PAGE_SHIFT
;
495 /* Iterate over the rest of the overlaps */
496 for (vma
= vma
->vm_next
; vma
; vma
= vma
->vm_next
) {
497 unsigned long overlap_len
;
499 if (vma
->vm_start
> end
)
502 overlap_len
= min(end
, vma
->vm_end
) - vma
->vm_start
;
503 nr_pages
+= overlap_len
>> PAGE_SHIFT
;
509 void __vma_link_rb(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
510 struct rb_node
**rb_link
, struct rb_node
*rb_parent
)
512 /* Update tracking information for the gap following the new vma. */
514 vma_gap_update(vma
->vm_next
);
516 mm
->highest_vm_end
= vma
->vm_end
;
519 * vma->vm_prev wasn't known when we followed the rbtree to find the
520 * correct insertion point for that vma. As a result, we could not
521 * update the vma vm_rb parents rb_subtree_gap values on the way down.
522 * So, we first insert the vma with a zero rb_subtree_gap value
523 * (to be consistent with what we did on the way down), and then
524 * immediately update the gap to the correct value. Finally we
525 * rebalance the rbtree after all augmented values have been set.
527 rb_link_node(&vma
->vm_rb
, rb_parent
, rb_link
);
528 vma
->rb_subtree_gap
= 0;
530 vma_rb_insert(vma
, &mm
->mm_rb
);
533 static void __vma_link_file(struct vm_area_struct
*vma
)
539 struct address_space
*mapping
= file
->f_mapping
;
541 if (vma
->vm_flags
& VM_DENYWRITE
)
542 atomic_dec(&file_inode(file
)->i_writecount
);
543 if (vma
->vm_flags
& VM_SHARED
)
544 atomic_inc(&mapping
->i_mmap_writable
);
546 flush_dcache_mmap_lock(mapping
);
547 vma_interval_tree_insert(vma
, &mapping
->i_mmap
);
548 flush_dcache_mmap_unlock(mapping
);
553 __vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
554 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
555 struct rb_node
*rb_parent
)
557 __vma_link_list(mm
, vma
, prev
, rb_parent
);
558 __vma_link_rb(mm
, vma
, rb_link
, rb_parent
);
561 static void vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
562 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
563 struct rb_node
*rb_parent
)
565 struct address_space
*mapping
= NULL
;
568 mapping
= vma
->vm_file
->f_mapping
;
569 i_mmap_lock_write(mapping
);
572 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
573 __vma_link_file(vma
);
576 i_mmap_unlock_write(mapping
);
583 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
584 * mm's list and rbtree. It has already been inserted into the interval tree.
586 static void __insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
588 struct vm_area_struct
*prev
;
589 struct rb_node
**rb_link
, *rb_parent
;
591 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
592 &prev
, &rb_link
, &rb_parent
))
594 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
599 __vma_unlink(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
600 struct vm_area_struct
*prev
)
602 struct vm_area_struct
*next
;
604 vma_rb_erase(vma
, &mm
->mm_rb
);
605 prev
->vm_next
= next
= vma
->vm_next
;
607 next
->vm_prev
= prev
;
610 vmacache_invalidate(mm
);
614 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
615 * is already present in an i_mmap tree without adjusting the tree.
616 * The following helper function should be used when such adjustments
617 * are necessary. The "insert" vma (if any) is to be inserted
618 * before we drop the necessary locks.
620 int vma_adjust(struct vm_area_struct
*vma
, unsigned long start
,
621 unsigned long end
, pgoff_t pgoff
, struct vm_area_struct
*insert
)
623 struct mm_struct
*mm
= vma
->vm_mm
;
624 struct vm_area_struct
*next
= vma
->vm_next
;
625 struct vm_area_struct
*importer
= NULL
;
626 struct address_space
*mapping
= NULL
;
627 struct rb_root
*root
= NULL
;
628 struct anon_vma
*anon_vma
= NULL
;
629 struct file
*file
= vma
->vm_file
;
630 bool start_changed
= false, end_changed
= false;
631 long adjust_next
= 0;
634 if (next
&& !insert
) {
635 struct vm_area_struct
*exporter
= NULL
;
637 if (end
>= next
->vm_end
) {
639 * vma expands, overlapping all the next, and
640 * perhaps the one after too (mprotect case 6).
642 again
: remove_next
= 1 + (end
> next
->vm_end
);
646 } else if (end
> next
->vm_start
) {
648 * vma expands, overlapping part of the next:
649 * mprotect case 5 shifting the boundary up.
651 adjust_next
= (end
- next
->vm_start
) >> PAGE_SHIFT
;
654 } else if (end
< vma
->vm_end
) {
656 * vma shrinks, and !insert tells it's not
657 * split_vma inserting another: so it must be
658 * mprotect case 4 shifting the boundary down.
660 adjust_next
= -((vma
->vm_end
- end
) >> PAGE_SHIFT
);
666 * Easily overlooked: when mprotect shifts the boundary,
667 * make sure the expanding vma has anon_vma set if the
668 * shrinking vma had, to cover any anon pages imported.
670 if (exporter
&& exporter
->anon_vma
&& !importer
->anon_vma
) {
673 importer
->anon_vma
= exporter
->anon_vma
;
674 error
= anon_vma_clone(importer
, exporter
);
681 mapping
= file
->f_mapping
;
682 root
= &mapping
->i_mmap
;
683 uprobe_munmap(vma
, vma
->vm_start
, vma
->vm_end
);
686 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
688 i_mmap_lock_write(mapping
);
691 * Put into interval tree now, so instantiated pages
692 * are visible to arm/parisc __flush_dcache_page
693 * throughout; but we cannot insert into address
694 * space until vma start or end is updated.
696 __vma_link_file(insert
);
700 vma_adjust_trans_huge(vma
, start
, end
, adjust_next
);
702 anon_vma
= vma
->anon_vma
;
703 if (!anon_vma
&& adjust_next
)
704 anon_vma
= next
->anon_vma
;
706 VM_BUG_ON_VMA(adjust_next
&& next
->anon_vma
&&
707 anon_vma
!= next
->anon_vma
, next
);
708 anon_vma_lock_write(anon_vma
);
709 anon_vma_interval_tree_pre_update_vma(vma
);
711 anon_vma_interval_tree_pre_update_vma(next
);
715 flush_dcache_mmap_lock(mapping
);
716 vma_interval_tree_remove(vma
, root
);
718 vma_interval_tree_remove(next
, root
);
721 if (start
!= vma
->vm_start
) {
722 vma
->vm_start
= start
;
723 start_changed
= true;
725 if (end
!= vma
->vm_end
) {
729 vma
->vm_pgoff
= pgoff
;
731 next
->vm_start
+= adjust_next
<< PAGE_SHIFT
;
732 next
->vm_pgoff
+= adjust_next
;
737 vma_interval_tree_insert(next
, root
);
738 vma_interval_tree_insert(vma
, root
);
739 flush_dcache_mmap_unlock(mapping
);
744 * vma_merge has merged next into vma, and needs
745 * us to remove next before dropping the locks.
747 __vma_unlink(mm
, next
, vma
);
749 __remove_shared_vm_struct(next
, file
, mapping
);
752 * split_vma has split insert from vma, and needs
753 * us to insert it before dropping the locks
754 * (it may either follow vma or precede it).
756 __insert_vm_struct(mm
, insert
);
762 mm
->highest_vm_end
= end
;
763 else if (!adjust_next
)
764 vma_gap_update(next
);
769 anon_vma_interval_tree_post_update_vma(vma
);
771 anon_vma_interval_tree_post_update_vma(next
);
772 anon_vma_unlock_write(anon_vma
);
775 i_mmap_unlock_write(mapping
);
786 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
790 anon_vma_merge(vma
, next
);
792 mpol_put(vma_policy(next
));
793 kmem_cache_free(vm_area_cachep
, next
);
795 * In mprotect's case 6 (see comments on vma_merge),
796 * we must remove another next too. It would clutter
797 * up the code too much to do both in one go.
800 if (remove_next
== 2)
803 vma_gap_update(next
);
805 mm
->highest_vm_end
= end
;
816 * If the vma has a ->close operation then the driver probably needs to release
817 * per-vma resources, so we don't attempt to merge those.
819 static inline int is_mergeable_vma(struct vm_area_struct
*vma
,
820 struct file
*file
, unsigned long vm_flags
,
821 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
824 * VM_SOFTDIRTY should not prevent from VMA merging, if we
825 * match the flags but dirty bit -- the caller should mark
826 * merged VMA as dirty. If dirty bit won't be excluded from
827 * comparison, we increase pressue on the memory system forcing
828 * the kernel to generate new VMAs when old one could be
831 if ((vma
->vm_flags
^ vm_flags
) & ~VM_SOFTDIRTY
)
833 if (vma
->vm_file
!= file
)
835 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
837 if (!is_mergeable_vm_userfaultfd_ctx(vma
, vm_userfaultfd_ctx
))
842 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
843 struct anon_vma
*anon_vma2
,
844 struct vm_area_struct
*vma
)
847 * The list_is_singular() test is to avoid merging VMA cloned from
848 * parents. This can improve scalability caused by anon_vma lock.
850 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
851 list_is_singular(&vma
->anon_vma_chain
)))
853 return anon_vma1
== anon_vma2
;
857 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
858 * in front of (at a lower virtual address and file offset than) the vma.
860 * We cannot merge two vmas if they have differently assigned (non-NULL)
861 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
863 * We don't check here for the merged mmap wrapping around the end of pagecache
864 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
865 * wrap, nor mmaps which cover the final page at index -1UL.
868 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
869 struct anon_vma
*anon_vma
, struct file
*file
,
871 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
873 if (is_mergeable_vma(vma
, file
, vm_flags
, vm_userfaultfd_ctx
) &&
874 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
875 if (vma
->vm_pgoff
== vm_pgoff
)
882 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
883 * beyond (at a higher virtual address and file offset than) the vma.
885 * We cannot merge two vmas if they have differently assigned (non-NULL)
886 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
889 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
890 struct anon_vma
*anon_vma
, struct file
*file
,
892 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
894 if (is_mergeable_vma(vma
, file
, vm_flags
, vm_userfaultfd_ctx
) &&
895 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
897 vm_pglen
= vma_pages(vma
);
898 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
905 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
906 * whether that can be merged with its predecessor or its successor.
907 * Or both (it neatly fills a hole).
909 * In most cases - when called for mmap, brk or mremap - [addr,end) is
910 * certain not to be mapped by the time vma_merge is called; but when
911 * called for mprotect, it is certain to be already mapped (either at
912 * an offset within prev, or at the start of next), and the flags of
913 * this area are about to be changed to vm_flags - and the no-change
914 * case has already been eliminated.
916 * The following mprotect cases have to be considered, where AAAA is
917 * the area passed down from mprotect_fixup, never extending beyond one
918 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
920 * AAAA AAAA AAAA AAAA
921 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
922 * cannot merge might become might become might become
923 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
924 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
925 * mremap move: PPPPNNNNNNNN 8
927 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
928 * might become case 1 below case 2 below case 3 below
930 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
931 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
933 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
934 struct vm_area_struct
*prev
, unsigned long addr
,
935 unsigned long end
, unsigned long vm_flags
,
936 struct anon_vma
*anon_vma
, struct file
*file
,
937 pgoff_t pgoff
, struct mempolicy
*policy
,
938 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
940 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
941 struct vm_area_struct
*area
, *next
;
945 * We later require that vma->vm_flags == vm_flags,
946 * so this tests vma->vm_flags & VM_SPECIAL, too.
948 if (vm_flags
& VM_SPECIAL
)
952 next
= prev
->vm_next
;
956 if (next
&& next
->vm_end
== end
) /* cases 6, 7, 8 */
957 next
= next
->vm_next
;
960 * Can it merge with the predecessor?
962 if (prev
&& prev
->vm_end
== addr
&&
963 mpol_equal(vma_policy(prev
), policy
) &&
964 can_vma_merge_after(prev
, vm_flags
,
965 anon_vma
, file
, pgoff
,
966 vm_userfaultfd_ctx
)) {
968 * OK, it can. Can we now merge in the successor as well?
970 if (next
&& end
== next
->vm_start
&&
971 mpol_equal(policy
, vma_policy(next
)) &&
972 can_vma_merge_before(next
, vm_flags
,
975 vm_userfaultfd_ctx
) &&
976 is_mergeable_anon_vma(prev
->anon_vma
,
977 next
->anon_vma
, NULL
)) {
979 err
= vma_adjust(prev
, prev
->vm_start
,
980 next
->vm_end
, prev
->vm_pgoff
, NULL
);
981 } else /* cases 2, 5, 7 */
982 err
= vma_adjust(prev
, prev
->vm_start
,
983 end
, prev
->vm_pgoff
, NULL
);
986 khugepaged_enter_vma_merge(prev
, vm_flags
);
991 * Can this new request be merged in front of next?
993 if (next
&& end
== next
->vm_start
&&
994 mpol_equal(policy
, vma_policy(next
)) &&
995 can_vma_merge_before(next
, vm_flags
,
996 anon_vma
, file
, pgoff
+pglen
,
997 vm_userfaultfd_ctx
)) {
998 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
999 err
= vma_adjust(prev
, prev
->vm_start
,
1000 addr
, prev
->vm_pgoff
, NULL
);
1001 else /* cases 3, 8 */
1002 err
= vma_adjust(area
, addr
, next
->vm_end
,
1003 next
->vm_pgoff
- pglen
, NULL
);
1006 khugepaged_enter_vma_merge(area
, vm_flags
);
1014 * Rough compatbility check to quickly see if it's even worth looking
1015 * at sharing an anon_vma.
1017 * They need to have the same vm_file, and the flags can only differ
1018 * in things that mprotect may change.
1020 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1021 * we can merge the two vma's. For example, we refuse to merge a vma if
1022 * there is a vm_ops->close() function, because that indicates that the
1023 * driver is doing some kind of reference counting. But that doesn't
1024 * really matter for the anon_vma sharing case.
1026 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1028 return a
->vm_end
== b
->vm_start
&&
1029 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
1030 a
->vm_file
== b
->vm_file
&&
1031 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_READ
|VM_WRITE
|VM_EXEC
|VM_SOFTDIRTY
)) &&
1032 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
1036 * Do some basic sanity checking to see if we can re-use the anon_vma
1037 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1038 * the same as 'old', the other will be the new one that is trying
1039 * to share the anon_vma.
1041 * NOTE! This runs with mm_sem held for reading, so it is possible that
1042 * the anon_vma of 'old' is concurrently in the process of being set up
1043 * by another page fault trying to merge _that_. But that's ok: if it
1044 * is being set up, that automatically means that it will be a singleton
1045 * acceptable for merging, so we can do all of this optimistically. But
1046 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1048 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1049 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1050 * is to return an anon_vma that is "complex" due to having gone through
1053 * We also make sure that the two vma's are compatible (adjacent,
1054 * and with the same memory policies). That's all stable, even with just
1055 * a read lock on the mm_sem.
1057 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1059 if (anon_vma_compatible(a
, b
)) {
1060 struct anon_vma
*anon_vma
= READ_ONCE(old
->anon_vma
);
1062 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
1069 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1070 * neighbouring vmas for a suitable anon_vma, before it goes off
1071 * to allocate a new anon_vma. It checks because a repetitive
1072 * sequence of mprotects and faults may otherwise lead to distinct
1073 * anon_vmas being allocated, preventing vma merge in subsequent
1076 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
1078 struct anon_vma
*anon_vma
;
1079 struct vm_area_struct
*near
;
1081 near
= vma
->vm_next
;
1085 anon_vma
= reusable_anon_vma(near
, vma
, near
);
1089 near
= vma
->vm_prev
;
1093 anon_vma
= reusable_anon_vma(near
, near
, vma
);
1098 * There's no absolute need to look only at touching neighbours:
1099 * we could search further afield for "compatible" anon_vmas.
1100 * But it would probably just be a waste of time searching,
1101 * or lead to too many vmas hanging off the same anon_vma.
1102 * We're trying to allow mprotect remerging later on,
1103 * not trying to minimize memory used for anon_vmas.
1109 * If a hint addr is less than mmap_min_addr change hint to be as
1110 * low as possible but still greater than mmap_min_addr
1112 static inline unsigned long round_hint_to_min(unsigned long hint
)
1115 if (((void *)hint
!= NULL
) &&
1116 (hint
< mmap_min_addr
))
1117 return PAGE_ALIGN(mmap_min_addr
);
1121 static inline int mlock_future_check(struct mm_struct
*mm
,
1122 unsigned long flags
,
1125 unsigned long locked
, lock_limit
;
1127 /* mlock MCL_FUTURE? */
1128 if (flags
& VM_LOCKED
) {
1129 locked
= len
>> PAGE_SHIFT
;
1130 locked
+= mm
->locked_vm
;
1131 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1132 lock_limit
>>= PAGE_SHIFT
;
1133 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1140 * The caller must hold down_write(¤t->mm->mmap_sem).
1142 unsigned long do_mmap(struct file
*file
, unsigned long addr
,
1143 unsigned long len
, unsigned long prot
,
1144 unsigned long flags
, vm_flags_t vm_flags
,
1145 unsigned long pgoff
, unsigned long *populate
)
1147 struct mm_struct
*mm
= current
->mm
;
1155 * Does the application expect PROT_READ to imply PROT_EXEC?
1157 * (the exception is when the underlying filesystem is noexec
1158 * mounted, in which case we dont add PROT_EXEC.)
1160 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1161 if (!(file
&& path_noexec(&file
->f_path
)))
1164 if (!(flags
& MAP_FIXED
))
1165 addr
= round_hint_to_min(addr
);
1167 /* Careful about overflows.. */
1168 len
= PAGE_ALIGN(len
);
1172 /* offset overflow? */
1173 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1176 /* Too many mappings? */
1177 if (mm
->map_count
> sysctl_max_map_count
)
1180 /* Obtain the address to map to. we verify (or select) it and ensure
1181 * that it represents a valid section of the address space.
1183 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1184 if (offset_in_page(addr
))
1187 /* Do simple checking here so the lower-level routines won't have
1188 * to. we assume access permissions have been handled by the open
1189 * of the memory object, so we don't do any here.
1191 vm_flags
|= calc_vm_prot_bits(prot
) | calc_vm_flag_bits(flags
) |
1192 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1194 if (flags
& MAP_LOCKED
)
1195 if (!can_do_mlock())
1198 if (mlock_future_check(mm
, vm_flags
, len
))
1202 struct inode
*inode
= file_inode(file
);
1204 switch (flags
& MAP_TYPE
) {
1206 if ((prot
&PROT_WRITE
) && !(file
->f_mode
&FMODE_WRITE
))
1210 * Make sure we don't allow writing to an append-only
1213 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1217 * Make sure there are no mandatory locks on the file.
1219 if (locks_verify_locked(file
))
1222 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1223 if (!(file
->f_mode
& FMODE_WRITE
))
1224 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1228 if (!(file
->f_mode
& FMODE_READ
))
1230 if (path_noexec(&file
->f_path
)) {
1231 if (vm_flags
& VM_EXEC
)
1233 vm_flags
&= ~VM_MAYEXEC
;
1236 if (!file
->f_op
->mmap
)
1238 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1246 switch (flags
& MAP_TYPE
) {
1248 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1254 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1258 * Set pgoff according to addr for anon_vma.
1260 pgoff
= addr
>> PAGE_SHIFT
;
1268 * Set 'VM_NORESERVE' if we should not account for the
1269 * memory use of this mapping.
1271 if (flags
& MAP_NORESERVE
) {
1272 /* We honor MAP_NORESERVE if allowed to overcommit */
1273 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1274 vm_flags
|= VM_NORESERVE
;
1276 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1277 if (file
&& is_file_hugepages(file
))
1278 vm_flags
|= VM_NORESERVE
;
1281 addr
= mmap_region(file
, addr
, len
, vm_flags
, pgoff
);
1282 if (!IS_ERR_VALUE(addr
) &&
1283 ((vm_flags
& VM_LOCKED
) ||
1284 (flags
& (MAP_POPULATE
| MAP_NONBLOCK
)) == MAP_POPULATE
))
1289 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1290 unsigned long, prot
, unsigned long, flags
,
1291 unsigned long, fd
, unsigned long, pgoff
)
1293 struct file
*file
= NULL
;
1294 unsigned long retval
;
1296 if (!(flags
& MAP_ANONYMOUS
)) {
1297 audit_mmap_fd(fd
, flags
);
1301 if (is_file_hugepages(file
))
1302 len
= ALIGN(len
, huge_page_size(hstate_file(file
)));
1304 if (unlikely(flags
& MAP_HUGETLB
&& !is_file_hugepages(file
)))
1306 } else if (flags
& MAP_HUGETLB
) {
1307 struct user_struct
*user
= NULL
;
1310 hs
= hstate_sizelog((flags
>> MAP_HUGE_SHIFT
) & SHM_HUGE_MASK
);
1314 len
= ALIGN(len
, huge_page_size(hs
));
1316 * VM_NORESERVE is used because the reservations will be
1317 * taken when vm_ops->mmap() is called
1318 * A dummy user value is used because we are not locking
1319 * memory so no accounting is necessary
1321 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, len
,
1323 &user
, HUGETLB_ANONHUGE_INODE
,
1324 (flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1326 return PTR_ERR(file
);
1329 flags
&= ~(MAP_EXECUTABLE
| MAP_DENYWRITE
);
1331 retval
= vm_mmap_pgoff(file
, addr
, len
, prot
, flags
, pgoff
);
1338 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1339 struct mmap_arg_struct
{
1343 unsigned long flags
;
1345 unsigned long offset
;
1348 SYSCALL_DEFINE1(old_mmap
, struct mmap_arg_struct __user
*, arg
)
1350 struct mmap_arg_struct a
;
1352 if (copy_from_user(&a
, arg
, sizeof(a
)))
1354 if (offset_in_page(a
.offset
))
1357 return sys_mmap_pgoff(a
.addr
, a
.len
, a
.prot
, a
.flags
, a
.fd
,
1358 a
.offset
>> PAGE_SHIFT
);
1360 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1363 * Some shared mappigns will want the pages marked read-only
1364 * to track write events. If so, we'll downgrade vm_page_prot
1365 * to the private version (using protection_map[] without the
1368 int vma_wants_writenotify(struct vm_area_struct
*vma
)
1370 vm_flags_t vm_flags
= vma
->vm_flags
;
1371 const struct vm_operations_struct
*vm_ops
= vma
->vm_ops
;
1373 /* If it was private or non-writable, the write bit is already clear */
1374 if ((vm_flags
& (VM_WRITE
|VM_SHARED
)) != ((VM_WRITE
|VM_SHARED
)))
1377 /* The backer wishes to know when pages are first written to? */
1378 if (vm_ops
&& (vm_ops
->page_mkwrite
|| vm_ops
->pfn_mkwrite
))
1381 /* The open routine did something to the protections that pgprot_modify
1382 * won't preserve? */
1383 if (pgprot_val(vma
->vm_page_prot
) !=
1384 pgprot_val(vm_pgprot_modify(vma
->vm_page_prot
, vm_flags
)))
1387 /* Do we need to track softdirty? */
1388 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY
) && !(vm_flags
& VM_SOFTDIRTY
))
1391 /* Specialty mapping? */
1392 if (vm_flags
& VM_PFNMAP
)
1395 /* Can the mapping track the dirty pages? */
1396 return vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
1397 mapping_cap_account_dirty(vma
->vm_file
->f_mapping
);
1401 * We account for memory if it's a private writeable mapping,
1402 * not hugepages and VM_NORESERVE wasn't set.
1404 static inline int accountable_mapping(struct file
*file
, vm_flags_t vm_flags
)
1407 * hugetlb has its own accounting separate from the core VM
1408 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1410 if (file
&& is_file_hugepages(file
))
1413 return (vm_flags
& (VM_NORESERVE
| VM_SHARED
| VM_WRITE
)) == VM_WRITE
;
1416 unsigned long mmap_region(struct file
*file
, unsigned long addr
,
1417 unsigned long len
, vm_flags_t vm_flags
, unsigned long pgoff
)
1419 struct mm_struct
*mm
= current
->mm
;
1420 struct vm_area_struct
*vma
, *prev
;
1422 struct rb_node
**rb_link
, *rb_parent
;
1423 unsigned long charged
= 0;
1425 /* Check against address space limit. */
1426 if (!may_expand_vm(mm
, vm_flags
, len
>> PAGE_SHIFT
)) {
1427 unsigned long nr_pages
;
1430 * MAP_FIXED may remove pages of mappings that intersects with
1431 * requested mapping. Account for the pages it would unmap.
1433 nr_pages
= count_vma_pages_range(mm
, addr
, addr
+ len
);
1435 if (!may_expand_vm(mm
, vm_flags
,
1436 (len
>> PAGE_SHIFT
) - nr_pages
))
1440 /* Clear old maps */
1441 while (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
,
1443 if (do_munmap(mm
, addr
, len
))
1448 * Private writable mapping: check memory availability
1450 if (accountable_mapping(file
, vm_flags
)) {
1451 charged
= len
>> PAGE_SHIFT
;
1452 if (security_vm_enough_memory_mm(mm
, charged
))
1454 vm_flags
|= VM_ACCOUNT
;
1458 * Can we just expand an old mapping?
1460 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
,
1461 NULL
, file
, pgoff
, NULL
, NULL_VM_UFFD_CTX
);
1466 * Determine the object being mapped and call the appropriate
1467 * specific mapper. the address has already been validated, but
1468 * not unmapped, but the maps are removed from the list.
1470 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
1477 vma
->vm_start
= addr
;
1478 vma
->vm_end
= addr
+ len
;
1479 vma
->vm_flags
= vm_flags
;
1480 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1481 vma
->vm_pgoff
= pgoff
;
1482 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
1485 if (vm_flags
& VM_DENYWRITE
) {
1486 error
= deny_write_access(file
);
1490 if (vm_flags
& VM_SHARED
) {
1491 error
= mapping_map_writable(file
->f_mapping
);
1493 goto allow_write_and_free_vma
;
1496 /* ->mmap() can change vma->vm_file, but must guarantee that
1497 * vma_link() below can deny write-access if VM_DENYWRITE is set
1498 * and map writably if VM_SHARED is set. This usually means the
1499 * new file must not have been exposed to user-space, yet.
1501 vma
->vm_file
= get_file(file
);
1502 error
= file
->f_op
->mmap(file
, vma
);
1504 goto unmap_and_free_vma
;
1506 /* Can addr have changed??
1508 * Answer: Yes, several device drivers can do it in their
1509 * f_op->mmap method. -DaveM
1510 * Bug: If addr is changed, prev, rb_link, rb_parent should
1511 * be updated for vma_link()
1513 WARN_ON_ONCE(addr
!= vma
->vm_start
);
1515 addr
= vma
->vm_start
;
1516 vm_flags
= vma
->vm_flags
;
1517 } else if (vm_flags
& VM_SHARED
) {
1518 error
= shmem_zero_setup(vma
);
1523 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1524 /* Once vma denies write, undo our temporary denial count */
1526 if (vm_flags
& VM_SHARED
)
1527 mapping_unmap_writable(file
->f_mapping
);
1528 if (vm_flags
& VM_DENYWRITE
)
1529 allow_write_access(file
);
1531 file
= vma
->vm_file
;
1533 perf_event_mmap(vma
);
1535 vm_stat_account(mm
, vm_flags
, len
>> PAGE_SHIFT
);
1536 if (vm_flags
& VM_LOCKED
) {
1537 if (!((vm_flags
& VM_SPECIAL
) || is_vm_hugetlb_page(vma
) ||
1538 vma
== get_gate_vma(current
->mm
)))
1539 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1541 vma
->vm_flags
&= VM_LOCKED_CLEAR_MASK
;
1548 * New (or expanded) vma always get soft dirty status.
1549 * Otherwise user-space soft-dirty page tracker won't
1550 * be able to distinguish situation when vma area unmapped,
1551 * then new mapped in-place (which must be aimed as
1552 * a completely new data area).
1554 vma
->vm_flags
|= VM_SOFTDIRTY
;
1556 vma_set_page_prot(vma
);
1561 vma
->vm_file
= NULL
;
1564 /* Undo any partial mapping done by a device driver. */
1565 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1567 if (vm_flags
& VM_SHARED
)
1568 mapping_unmap_writable(file
->f_mapping
);
1569 allow_write_and_free_vma
:
1570 if (vm_flags
& VM_DENYWRITE
)
1571 allow_write_access(file
);
1573 kmem_cache_free(vm_area_cachep
, vma
);
1576 vm_unacct_memory(charged
);
1580 unsigned long unmapped_area(struct vm_unmapped_area_info
*info
)
1583 * We implement the search by looking for an rbtree node that
1584 * immediately follows a suitable gap. That is,
1585 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1586 * - gap_end = vma->vm_start >= info->low_limit + length;
1587 * - gap_end - gap_start >= length
1590 struct mm_struct
*mm
= current
->mm
;
1591 struct vm_area_struct
*vma
;
1592 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1594 /* Adjust search length to account for worst case alignment overhead */
1595 length
= info
->length
+ info
->align_mask
;
1596 if (length
< info
->length
)
1599 /* Adjust search limits by the desired length */
1600 if (info
->high_limit
< length
)
1602 high_limit
= info
->high_limit
- length
;
1604 if (info
->low_limit
> high_limit
)
1606 low_limit
= info
->low_limit
+ length
;
1608 /* Check if rbtree root looks promising */
1609 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1611 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1612 if (vma
->rb_subtree_gap
< length
)
1616 /* Visit left subtree if it looks promising */
1617 gap_end
= vma
->vm_start
;
1618 if (gap_end
>= low_limit
&& vma
->vm_rb
.rb_left
) {
1619 struct vm_area_struct
*left
=
1620 rb_entry(vma
->vm_rb
.rb_left
,
1621 struct vm_area_struct
, vm_rb
);
1622 if (left
->rb_subtree_gap
>= length
) {
1628 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1630 /* Check if current node has a suitable gap */
1631 if (gap_start
> high_limit
)
1633 if (gap_end
>= low_limit
&& gap_end
- gap_start
>= length
)
1636 /* Visit right subtree if it looks promising */
1637 if (vma
->vm_rb
.rb_right
) {
1638 struct vm_area_struct
*right
=
1639 rb_entry(vma
->vm_rb
.rb_right
,
1640 struct vm_area_struct
, vm_rb
);
1641 if (right
->rb_subtree_gap
>= length
) {
1647 /* Go back up the rbtree to find next candidate node */
1649 struct rb_node
*prev
= &vma
->vm_rb
;
1650 if (!rb_parent(prev
))
1652 vma
= rb_entry(rb_parent(prev
),
1653 struct vm_area_struct
, vm_rb
);
1654 if (prev
== vma
->vm_rb
.rb_left
) {
1655 gap_start
= vma
->vm_prev
->vm_end
;
1656 gap_end
= vma
->vm_start
;
1663 /* Check highest gap, which does not precede any rbtree node */
1664 gap_start
= mm
->highest_vm_end
;
1665 gap_end
= ULONG_MAX
; /* Only for VM_BUG_ON below */
1666 if (gap_start
> high_limit
)
1670 /* We found a suitable gap. Clip it with the original low_limit. */
1671 if (gap_start
< info
->low_limit
)
1672 gap_start
= info
->low_limit
;
1674 /* Adjust gap address to the desired alignment */
1675 gap_start
+= (info
->align_offset
- gap_start
) & info
->align_mask
;
1677 VM_BUG_ON(gap_start
+ info
->length
> info
->high_limit
);
1678 VM_BUG_ON(gap_start
+ info
->length
> gap_end
);
1682 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info
*info
)
1684 struct mm_struct
*mm
= current
->mm
;
1685 struct vm_area_struct
*vma
;
1686 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1688 /* Adjust search length to account for worst case alignment overhead */
1689 length
= info
->length
+ info
->align_mask
;
1690 if (length
< info
->length
)
1694 * Adjust search limits by the desired length.
1695 * See implementation comment at top of unmapped_area().
1697 gap_end
= info
->high_limit
;
1698 if (gap_end
< length
)
1700 high_limit
= gap_end
- length
;
1702 if (info
->low_limit
> high_limit
)
1704 low_limit
= info
->low_limit
+ length
;
1706 /* Check highest gap, which does not precede any rbtree node */
1707 gap_start
= mm
->highest_vm_end
;
1708 if (gap_start
<= high_limit
)
1711 /* Check if rbtree root looks promising */
1712 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1714 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1715 if (vma
->rb_subtree_gap
< length
)
1719 /* Visit right subtree if it looks promising */
1720 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1721 if (gap_start
<= high_limit
&& vma
->vm_rb
.rb_right
) {
1722 struct vm_area_struct
*right
=
1723 rb_entry(vma
->vm_rb
.rb_right
,
1724 struct vm_area_struct
, vm_rb
);
1725 if (right
->rb_subtree_gap
>= length
) {
1732 /* Check if current node has a suitable gap */
1733 gap_end
= vma
->vm_start
;
1734 if (gap_end
< low_limit
)
1736 if (gap_start
<= high_limit
&& gap_end
- gap_start
>= length
)
1739 /* Visit left subtree if it looks promising */
1740 if (vma
->vm_rb
.rb_left
) {
1741 struct vm_area_struct
*left
=
1742 rb_entry(vma
->vm_rb
.rb_left
,
1743 struct vm_area_struct
, vm_rb
);
1744 if (left
->rb_subtree_gap
>= length
) {
1750 /* Go back up the rbtree to find next candidate node */
1752 struct rb_node
*prev
= &vma
->vm_rb
;
1753 if (!rb_parent(prev
))
1755 vma
= rb_entry(rb_parent(prev
),
1756 struct vm_area_struct
, vm_rb
);
1757 if (prev
== vma
->vm_rb
.rb_right
) {
1758 gap_start
= vma
->vm_prev
?
1759 vma
->vm_prev
->vm_end
: 0;
1766 /* We found a suitable gap. Clip it with the original high_limit. */
1767 if (gap_end
> info
->high_limit
)
1768 gap_end
= info
->high_limit
;
1771 /* Compute highest gap address at the desired alignment */
1772 gap_end
-= info
->length
;
1773 gap_end
-= (gap_end
- info
->align_offset
) & info
->align_mask
;
1775 VM_BUG_ON(gap_end
< info
->low_limit
);
1776 VM_BUG_ON(gap_end
< gap_start
);
1780 /* Get an address range which is currently unmapped.
1781 * For shmat() with addr=0.
1783 * Ugly calling convention alert:
1784 * Return value with the low bits set means error value,
1786 * if (ret & ~PAGE_MASK)
1789 * This function "knows" that -ENOMEM has the bits set.
1791 #ifndef HAVE_ARCH_UNMAPPED_AREA
1793 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
1794 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
1796 struct mm_struct
*mm
= current
->mm
;
1797 struct vm_area_struct
*vma
;
1798 struct vm_unmapped_area_info info
;
1800 if (len
> TASK_SIZE
- mmap_min_addr
)
1803 if (flags
& MAP_FIXED
)
1807 addr
= PAGE_ALIGN(addr
);
1808 vma
= find_vma(mm
, addr
);
1809 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1810 (!vma
|| addr
+ len
<= vma
->vm_start
))
1816 info
.low_limit
= mm
->mmap_base
;
1817 info
.high_limit
= TASK_SIZE
;
1818 info
.align_mask
= 0;
1819 return vm_unmapped_area(&info
);
1824 * This mmap-allocator allocates new areas top-down from below the
1825 * stack's low limit (the base):
1827 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1829 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
1830 const unsigned long len
, const unsigned long pgoff
,
1831 const unsigned long flags
)
1833 struct vm_area_struct
*vma
;
1834 struct mm_struct
*mm
= current
->mm
;
1835 unsigned long addr
= addr0
;
1836 struct vm_unmapped_area_info info
;
1838 /* requested length too big for entire address space */
1839 if (len
> TASK_SIZE
- mmap_min_addr
)
1842 if (flags
& MAP_FIXED
)
1845 /* requesting a specific address */
1847 addr
= PAGE_ALIGN(addr
);
1848 vma
= find_vma(mm
, addr
);
1849 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1850 (!vma
|| addr
+ len
<= vma
->vm_start
))
1854 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
1856 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
1857 info
.high_limit
= mm
->mmap_base
;
1858 info
.align_mask
= 0;
1859 addr
= vm_unmapped_area(&info
);
1862 * A failed mmap() very likely causes application failure,
1863 * so fall back to the bottom-up function here. This scenario
1864 * can happen with large stack limits and large mmap()
1867 if (offset_in_page(addr
)) {
1868 VM_BUG_ON(addr
!= -ENOMEM
);
1870 info
.low_limit
= TASK_UNMAPPED_BASE
;
1871 info
.high_limit
= TASK_SIZE
;
1872 addr
= vm_unmapped_area(&info
);
1880 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
1881 unsigned long pgoff
, unsigned long flags
)
1883 unsigned long (*get_area
)(struct file
*, unsigned long,
1884 unsigned long, unsigned long, unsigned long);
1886 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
1890 /* Careful about overflows.. */
1891 if (len
> TASK_SIZE
)
1894 get_area
= current
->mm
->get_unmapped_area
;
1895 if (file
&& file
->f_op
->get_unmapped_area
)
1896 get_area
= file
->f_op
->get_unmapped_area
;
1897 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
1898 if (IS_ERR_VALUE(addr
))
1901 if (addr
> TASK_SIZE
- len
)
1903 if (offset_in_page(addr
))
1906 addr
= arch_rebalance_pgtables(addr
, len
);
1907 error
= security_mmap_addr(addr
);
1908 return error
? error
: addr
;
1911 EXPORT_SYMBOL(get_unmapped_area
);
1913 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1914 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
1916 struct rb_node
*rb_node
;
1917 struct vm_area_struct
*vma
;
1919 /* Check the cache first. */
1920 vma
= vmacache_find(mm
, addr
);
1924 rb_node
= mm
->mm_rb
.rb_node
;
1927 struct vm_area_struct
*tmp
;
1929 tmp
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
1931 if (tmp
->vm_end
> addr
) {
1933 if (tmp
->vm_start
<= addr
)
1935 rb_node
= rb_node
->rb_left
;
1937 rb_node
= rb_node
->rb_right
;
1941 vmacache_update(addr
, vma
);
1945 EXPORT_SYMBOL(find_vma
);
1948 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1950 struct vm_area_struct
*
1951 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
1952 struct vm_area_struct
**pprev
)
1954 struct vm_area_struct
*vma
;
1956 vma
= find_vma(mm
, addr
);
1958 *pprev
= vma
->vm_prev
;
1960 struct rb_node
*rb_node
= mm
->mm_rb
.rb_node
;
1963 *pprev
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
1964 rb_node
= rb_node
->rb_right
;
1971 * Verify that the stack growth is acceptable and
1972 * update accounting. This is shared with both the
1973 * grow-up and grow-down cases.
1975 static int acct_stack_growth(struct vm_area_struct
*vma
, unsigned long size
, unsigned long grow
)
1977 struct mm_struct
*mm
= vma
->vm_mm
;
1978 struct rlimit
*rlim
= current
->signal
->rlim
;
1979 unsigned long new_start
, actual_size
;
1981 /* address space limit tests */
1982 if (!may_expand_vm(mm
, vma
->vm_flags
, grow
))
1985 /* Stack limit test */
1987 if (size
&& (vma
->vm_flags
& (VM_GROWSUP
| VM_GROWSDOWN
)))
1988 actual_size
-= PAGE_SIZE
;
1989 if (actual_size
> READ_ONCE(rlim
[RLIMIT_STACK
].rlim_cur
))
1992 /* mlock limit tests */
1993 if (vma
->vm_flags
& VM_LOCKED
) {
1994 unsigned long locked
;
1995 unsigned long limit
;
1996 locked
= mm
->locked_vm
+ grow
;
1997 limit
= READ_ONCE(rlim
[RLIMIT_MEMLOCK
].rlim_cur
);
1998 limit
>>= PAGE_SHIFT
;
1999 if (locked
> limit
&& !capable(CAP_IPC_LOCK
))
2003 /* Check to ensure the stack will not grow into a hugetlb-only region */
2004 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
2006 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
2010 * Overcommit.. This must be the final test, as it will
2011 * update security statistics.
2013 if (security_vm_enough_memory_mm(mm
, grow
))
2019 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2021 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2022 * vma is the last one with address > vma->vm_end. Have to extend vma.
2024 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
)
2026 struct mm_struct
*mm
= vma
->vm_mm
;
2029 if (!(vma
->vm_flags
& VM_GROWSUP
))
2032 /* Guard against wrapping around to address 0. */
2033 if (address
< PAGE_ALIGN(address
+4))
2034 address
= PAGE_ALIGN(address
+4);
2038 /* We must make sure the anon_vma is allocated. */
2039 if (unlikely(anon_vma_prepare(vma
)))
2043 * vma->vm_start/vm_end cannot change under us because the caller
2044 * is required to hold the mmap_sem in read mode. We need the
2045 * anon_vma lock to serialize against concurrent expand_stacks.
2047 anon_vma_lock_write(vma
->anon_vma
);
2049 /* Somebody else might have raced and expanded it already */
2050 if (address
> vma
->vm_end
) {
2051 unsigned long size
, grow
;
2053 size
= address
- vma
->vm_start
;
2054 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
2057 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
2058 error
= acct_stack_growth(vma
, size
, grow
);
2061 * vma_gap_update() doesn't support concurrent
2062 * updates, but we only hold a shared mmap_sem
2063 * lock here, so we need to protect against
2064 * concurrent vma expansions.
2065 * anon_vma_lock_write() doesn't help here, as
2066 * we don't guarantee that all growable vmas
2067 * in a mm share the same root anon vma.
2068 * So, we reuse mm->page_table_lock to guard
2069 * against concurrent vma expansions.
2071 spin_lock(&mm
->page_table_lock
);
2072 if (vma
->vm_flags
& VM_LOCKED
)
2073 mm
->locked_vm
+= grow
;
2074 vm_stat_account(mm
, vma
->vm_flags
, grow
);
2075 anon_vma_interval_tree_pre_update_vma(vma
);
2076 vma
->vm_end
= address
;
2077 anon_vma_interval_tree_post_update_vma(vma
);
2079 vma_gap_update(vma
->vm_next
);
2081 mm
->highest_vm_end
= address
;
2082 spin_unlock(&mm
->page_table_lock
);
2084 perf_event_mmap(vma
);
2088 anon_vma_unlock_write(vma
->anon_vma
);
2089 khugepaged_enter_vma_merge(vma
, vma
->vm_flags
);
2093 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2096 * vma is the first one with address < vma->vm_start. Have to extend vma.
2098 int expand_downwards(struct vm_area_struct
*vma
,
2099 unsigned long address
)
2101 struct mm_struct
*mm
= vma
->vm_mm
;
2104 address
&= PAGE_MASK
;
2105 error
= security_mmap_addr(address
);
2109 /* We must make sure the anon_vma is allocated. */
2110 if (unlikely(anon_vma_prepare(vma
)))
2114 * vma->vm_start/vm_end cannot change under us because the caller
2115 * is required to hold the mmap_sem in read mode. We need the
2116 * anon_vma lock to serialize against concurrent expand_stacks.
2118 anon_vma_lock_write(vma
->anon_vma
);
2120 /* Somebody else might have raced and expanded it already */
2121 if (address
< vma
->vm_start
) {
2122 unsigned long size
, grow
;
2124 size
= vma
->vm_end
- address
;
2125 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
2128 if (grow
<= vma
->vm_pgoff
) {
2129 error
= acct_stack_growth(vma
, size
, grow
);
2132 * vma_gap_update() doesn't support concurrent
2133 * updates, but we only hold a shared mmap_sem
2134 * lock here, so we need to protect against
2135 * concurrent vma expansions.
2136 * anon_vma_lock_write() doesn't help here, as
2137 * we don't guarantee that all growable vmas
2138 * in a mm share the same root anon vma.
2139 * So, we reuse mm->page_table_lock to guard
2140 * against concurrent vma expansions.
2142 spin_lock(&mm
->page_table_lock
);
2143 if (vma
->vm_flags
& VM_LOCKED
)
2144 mm
->locked_vm
+= grow
;
2145 vm_stat_account(mm
, vma
->vm_flags
, grow
);
2146 anon_vma_interval_tree_pre_update_vma(vma
);
2147 vma
->vm_start
= address
;
2148 vma
->vm_pgoff
-= grow
;
2149 anon_vma_interval_tree_post_update_vma(vma
);
2150 vma_gap_update(vma
);
2151 spin_unlock(&mm
->page_table_lock
);
2153 perf_event_mmap(vma
);
2157 anon_vma_unlock_write(vma
->anon_vma
);
2158 khugepaged_enter_vma_merge(vma
, vma
->vm_flags
);
2164 * Note how expand_stack() refuses to expand the stack all the way to
2165 * abut the next virtual mapping, *unless* that mapping itself is also
2166 * a stack mapping. We want to leave room for a guard page, after all
2167 * (the guard page itself is not added here, that is done by the
2168 * actual page faulting logic)
2170 * This matches the behavior of the guard page logic (see mm/memory.c:
2171 * check_stack_guard_page()), which only allows the guard page to be
2172 * removed under these circumstances.
2174 #ifdef CONFIG_STACK_GROWSUP
2175 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2177 struct vm_area_struct
*next
;
2179 address
&= PAGE_MASK
;
2180 next
= vma
->vm_next
;
2181 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
) {
2182 if (!(next
->vm_flags
& VM_GROWSUP
))
2185 return expand_upwards(vma
, address
);
2188 struct vm_area_struct
*
2189 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2191 struct vm_area_struct
*vma
, *prev
;
2194 vma
= find_vma_prev(mm
, addr
, &prev
);
2195 if (vma
&& (vma
->vm_start
<= addr
))
2197 if (!prev
|| expand_stack(prev
, addr
))
2199 if (prev
->vm_flags
& VM_LOCKED
)
2200 populate_vma_page_range(prev
, addr
, prev
->vm_end
, NULL
);
2204 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2206 struct vm_area_struct
*prev
;
2208 address
&= PAGE_MASK
;
2209 prev
= vma
->vm_prev
;
2210 if (prev
&& prev
->vm_end
== address
) {
2211 if (!(prev
->vm_flags
& VM_GROWSDOWN
))
2214 return expand_downwards(vma
, address
);
2217 struct vm_area_struct
*
2218 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2220 struct vm_area_struct
*vma
;
2221 unsigned long start
;
2224 vma
= find_vma(mm
, addr
);
2227 if (vma
->vm_start
<= addr
)
2229 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2231 start
= vma
->vm_start
;
2232 if (expand_stack(vma
, addr
))
2234 if (vma
->vm_flags
& VM_LOCKED
)
2235 populate_vma_page_range(vma
, addr
, start
, NULL
);
2240 EXPORT_SYMBOL_GPL(find_extend_vma
);
2243 * Ok - we have the memory areas we should free on the vma list,
2244 * so release them, and do the vma updates.
2246 * Called with the mm semaphore held.
2248 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2250 unsigned long nr_accounted
= 0;
2252 /* Update high watermark before we lower total_vm */
2253 update_hiwater_vm(mm
);
2255 long nrpages
= vma_pages(vma
);
2257 if (vma
->vm_flags
& VM_ACCOUNT
)
2258 nr_accounted
+= nrpages
;
2259 vm_stat_account(mm
, vma
->vm_flags
, -nrpages
);
2260 vma
= remove_vma(vma
);
2262 vm_unacct_memory(nr_accounted
);
2267 * Get rid of page table information in the indicated region.
2269 * Called with the mm semaphore held.
2271 static void unmap_region(struct mm_struct
*mm
,
2272 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
2273 unsigned long start
, unsigned long end
)
2275 struct vm_area_struct
*next
= prev
? prev
->vm_next
: mm
->mmap
;
2276 struct mmu_gather tlb
;
2279 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2280 update_hiwater_rss(mm
);
2281 unmap_vmas(&tlb
, vma
, start
, end
);
2282 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
2283 next
? next
->vm_start
: USER_PGTABLES_CEILING
);
2284 tlb_finish_mmu(&tlb
, start
, end
);
2288 * Create a list of vma's touched by the unmap, removing them from the mm's
2289 * vma list as we go..
2292 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2293 struct vm_area_struct
*prev
, unsigned long end
)
2295 struct vm_area_struct
**insertion_point
;
2296 struct vm_area_struct
*tail_vma
= NULL
;
2298 insertion_point
= (prev
? &prev
->vm_next
: &mm
->mmap
);
2299 vma
->vm_prev
= NULL
;
2301 vma_rb_erase(vma
, &mm
->mm_rb
);
2305 } while (vma
&& vma
->vm_start
< end
);
2306 *insertion_point
= vma
;
2308 vma
->vm_prev
= prev
;
2309 vma_gap_update(vma
);
2311 mm
->highest_vm_end
= prev
? prev
->vm_end
: 0;
2312 tail_vma
->vm_next
= NULL
;
2314 /* Kill the cache */
2315 vmacache_invalidate(mm
);
2319 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2320 * munmap path where it doesn't make sense to fail.
2322 static int __split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2323 unsigned long addr
, int new_below
)
2325 struct vm_area_struct
*new;
2328 if (is_vm_hugetlb_page(vma
) && (addr
&
2329 ~(huge_page_mask(hstate_vma(vma
)))))
2332 new = kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2336 /* most fields are the same, copy all, and then fixup */
2339 INIT_LIST_HEAD(&new->anon_vma_chain
);
2344 new->vm_start
= addr
;
2345 new->vm_pgoff
+= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
);
2348 err
= vma_dup_policy(vma
, new);
2352 err
= anon_vma_clone(new, vma
);
2357 get_file(new->vm_file
);
2359 if (new->vm_ops
&& new->vm_ops
->open
)
2360 new->vm_ops
->open(new);
2363 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2364 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2366 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2372 /* Clean everything up if vma_adjust failed. */
2373 if (new->vm_ops
&& new->vm_ops
->close
)
2374 new->vm_ops
->close(new);
2377 unlink_anon_vmas(new);
2379 mpol_put(vma_policy(new));
2381 kmem_cache_free(vm_area_cachep
, new);
2386 * Split a vma into two pieces at address 'addr', a new vma is allocated
2387 * either for the first part or the tail.
2389 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2390 unsigned long addr
, int new_below
)
2392 if (mm
->map_count
>= sysctl_max_map_count
)
2395 return __split_vma(mm
, vma
, addr
, new_below
);
2398 /* Munmap is split into 2 main parts -- this part which finds
2399 * what needs doing, and the areas themselves, which do the
2400 * work. This now handles partial unmappings.
2401 * Jeremy Fitzhardinge <jeremy@goop.org>
2403 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
)
2406 struct vm_area_struct
*vma
, *prev
, *last
;
2408 if ((offset_in_page(start
)) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2411 len
= PAGE_ALIGN(len
);
2415 /* Find the first overlapping VMA */
2416 vma
= find_vma(mm
, start
);
2419 prev
= vma
->vm_prev
;
2420 /* we have start < vma->vm_end */
2422 /* if it doesn't overlap, we have nothing.. */
2424 if (vma
->vm_start
>= end
)
2428 * If we need to split any vma, do it now to save pain later.
2430 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2431 * unmapped vm_area_struct will remain in use: so lower split_vma
2432 * places tmp vma above, and higher split_vma places tmp vma below.
2434 if (start
> vma
->vm_start
) {
2438 * Make sure that map_count on return from munmap() will
2439 * not exceed its limit; but let map_count go just above
2440 * its limit temporarily, to help free resources as expected.
2442 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2445 error
= __split_vma(mm
, vma
, start
, 0);
2451 /* Does it split the last one? */
2452 last
= find_vma(mm
, end
);
2453 if (last
&& end
> last
->vm_start
) {
2454 int error
= __split_vma(mm
, last
, end
, 1);
2458 vma
= prev
? prev
->vm_next
: mm
->mmap
;
2461 * unlock any mlock()ed ranges before detaching vmas
2463 if (mm
->locked_vm
) {
2464 struct vm_area_struct
*tmp
= vma
;
2465 while (tmp
&& tmp
->vm_start
< end
) {
2466 if (tmp
->vm_flags
& VM_LOCKED
) {
2467 mm
->locked_vm
-= vma_pages(tmp
);
2468 munlock_vma_pages_all(tmp
);
2475 * Remove the vma's, and unmap the actual pages
2477 detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
);
2478 unmap_region(mm
, vma
, prev
, start
, end
);
2480 arch_unmap(mm
, vma
, start
, end
);
2482 /* Fix up all other VM information */
2483 remove_vma_list(mm
, vma
);
2488 int vm_munmap(unsigned long start
, size_t len
)
2491 struct mm_struct
*mm
= current
->mm
;
2493 down_write(&mm
->mmap_sem
);
2494 ret
= do_munmap(mm
, start
, len
);
2495 up_write(&mm
->mmap_sem
);
2498 EXPORT_SYMBOL(vm_munmap
);
2500 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2502 profile_munmap(addr
);
2503 return vm_munmap(addr
, len
);
2508 * Emulation of deprecated remap_file_pages() syscall.
2510 SYSCALL_DEFINE5(remap_file_pages
, unsigned long, start
, unsigned long, size
,
2511 unsigned long, prot
, unsigned long, pgoff
, unsigned long, flags
)
2514 struct mm_struct
*mm
= current
->mm
;
2515 struct vm_area_struct
*vma
;
2516 unsigned long populate
= 0;
2517 unsigned long ret
= -EINVAL
;
2520 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2521 current
->comm
, current
->pid
);
2525 start
= start
& PAGE_MASK
;
2526 size
= size
& PAGE_MASK
;
2528 if (start
+ size
<= start
)
2531 /* Does pgoff wrap? */
2532 if (pgoff
+ (size
>> PAGE_SHIFT
) < pgoff
)
2535 down_write(&mm
->mmap_sem
);
2536 vma
= find_vma(mm
, start
);
2538 if (!vma
|| !(vma
->vm_flags
& VM_SHARED
))
2541 if (start
< vma
->vm_start
)
2544 if (start
+ size
> vma
->vm_end
) {
2545 struct vm_area_struct
*next
;
2547 for (next
= vma
->vm_next
; next
; next
= next
->vm_next
) {
2548 /* hole between vmas ? */
2549 if (next
->vm_start
!= next
->vm_prev
->vm_end
)
2552 if (next
->vm_file
!= vma
->vm_file
)
2555 if (next
->vm_flags
!= vma
->vm_flags
)
2558 if (start
+ size
<= next
->vm_end
)
2566 prot
|= vma
->vm_flags
& VM_READ
? PROT_READ
: 0;
2567 prot
|= vma
->vm_flags
& VM_WRITE
? PROT_WRITE
: 0;
2568 prot
|= vma
->vm_flags
& VM_EXEC
? PROT_EXEC
: 0;
2570 flags
&= MAP_NONBLOCK
;
2571 flags
|= MAP_SHARED
| MAP_FIXED
| MAP_POPULATE
;
2572 if (vma
->vm_flags
& VM_LOCKED
) {
2573 struct vm_area_struct
*tmp
;
2574 flags
|= MAP_LOCKED
;
2576 /* drop PG_Mlocked flag for over-mapped range */
2577 for (tmp
= vma
; tmp
->vm_start
>= start
+ size
;
2578 tmp
= tmp
->vm_next
) {
2579 munlock_vma_pages_range(tmp
,
2580 max(tmp
->vm_start
, start
),
2581 min(tmp
->vm_end
, start
+ size
));
2585 file
= get_file(vma
->vm_file
);
2586 ret
= do_mmap_pgoff(vma
->vm_file
, start
, size
,
2587 prot
, flags
, pgoff
, &populate
);
2590 up_write(&mm
->mmap_sem
);
2592 mm_populate(ret
, populate
);
2593 if (!IS_ERR_VALUE(ret
))
2598 static inline void verify_mm_writelocked(struct mm_struct
*mm
)
2600 #ifdef CONFIG_DEBUG_VM
2601 if (unlikely(down_read_trylock(&mm
->mmap_sem
))) {
2603 up_read(&mm
->mmap_sem
);
2609 * this is really a simplified "do_mmap". it only handles
2610 * anonymous maps. eventually we may be able to do some
2611 * brk-specific accounting here.
2613 static unsigned long do_brk(unsigned long addr
, unsigned long len
)
2615 struct mm_struct
*mm
= current
->mm
;
2616 struct vm_area_struct
*vma
, *prev
;
2617 unsigned long flags
;
2618 struct rb_node
**rb_link
, *rb_parent
;
2619 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2622 len
= PAGE_ALIGN(len
);
2626 flags
= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2628 error
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2629 if (offset_in_page(error
))
2632 error
= mlock_future_check(mm
, mm
->def_flags
, len
);
2637 * mm->mmap_sem is required to protect against another thread
2638 * changing the mappings in case we sleep.
2640 verify_mm_writelocked(mm
);
2643 * Clear old maps. this also does some error checking for us
2645 while (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
,
2647 if (do_munmap(mm
, addr
, len
))
2651 /* Check against address space limits *after* clearing old maps... */
2652 if (!may_expand_vm(mm
, flags
, len
>> PAGE_SHIFT
))
2655 if (mm
->map_count
> sysctl_max_map_count
)
2658 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
2661 /* Can we just expand an old private anonymous mapping? */
2662 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
2663 NULL
, NULL
, pgoff
, NULL
, NULL_VM_UFFD_CTX
);
2668 * create a vma struct for an anonymous mapping
2670 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2672 vm_unacct_memory(len
>> PAGE_SHIFT
);
2676 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2678 vma
->vm_start
= addr
;
2679 vma
->vm_end
= addr
+ len
;
2680 vma
->vm_pgoff
= pgoff
;
2681 vma
->vm_flags
= flags
;
2682 vma
->vm_page_prot
= vm_get_page_prot(flags
);
2683 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2685 perf_event_mmap(vma
);
2686 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2687 mm
->data_vm
+= len
>> PAGE_SHIFT
;
2688 if (flags
& VM_LOCKED
)
2689 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
2690 vma
->vm_flags
|= VM_SOFTDIRTY
;
2694 unsigned long vm_brk(unsigned long addr
, unsigned long len
)
2696 struct mm_struct
*mm
= current
->mm
;
2700 down_write(&mm
->mmap_sem
);
2701 ret
= do_brk(addr
, len
);
2702 populate
= ((mm
->def_flags
& VM_LOCKED
) != 0);
2703 up_write(&mm
->mmap_sem
);
2705 mm_populate(addr
, len
);
2708 EXPORT_SYMBOL(vm_brk
);
2710 /* Release all mmaps. */
2711 void exit_mmap(struct mm_struct
*mm
)
2713 struct mmu_gather tlb
;
2714 struct vm_area_struct
*vma
;
2715 unsigned long nr_accounted
= 0;
2717 /* mm's last user has gone, and its about to be pulled down */
2718 mmu_notifier_release(mm
);
2720 if (mm
->locked_vm
) {
2723 if (vma
->vm_flags
& VM_LOCKED
)
2724 munlock_vma_pages_all(vma
);
2732 if (!vma
) /* Can happen if dup_mmap() received an OOM */
2737 tlb_gather_mmu(&tlb
, mm
, 0, -1);
2738 /* update_hiwater_rss(mm) here? but nobody should be looking */
2739 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2740 unmap_vmas(&tlb
, vma
, 0, -1);
2742 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, USER_PGTABLES_CEILING
);
2743 tlb_finish_mmu(&tlb
, 0, -1);
2746 * Walk the list again, actually closing and freeing it,
2747 * with preemption enabled, without holding any MM locks.
2750 if (vma
->vm_flags
& VM_ACCOUNT
)
2751 nr_accounted
+= vma_pages(vma
);
2752 vma
= remove_vma(vma
);
2754 vm_unacct_memory(nr_accounted
);
2757 /* Insert vm structure into process list sorted by address
2758 * and into the inode's i_mmap tree. If vm_file is non-NULL
2759 * then i_mmap_rwsem is taken here.
2761 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2763 struct vm_area_struct
*prev
;
2764 struct rb_node
**rb_link
, *rb_parent
;
2766 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
2767 &prev
, &rb_link
, &rb_parent
))
2769 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
2770 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
2774 * The vm_pgoff of a purely anonymous vma should be irrelevant
2775 * until its first write fault, when page's anon_vma and index
2776 * are set. But now set the vm_pgoff it will almost certainly
2777 * end up with (unless mremap moves it elsewhere before that
2778 * first wfault), so /proc/pid/maps tells a consistent story.
2780 * By setting it to reflect the virtual start address of the
2781 * vma, merges and splits can happen in a seamless way, just
2782 * using the existing file pgoff checks and manipulations.
2783 * Similarly in do_mmap_pgoff and in do_brk.
2785 if (vma_is_anonymous(vma
)) {
2786 BUG_ON(vma
->anon_vma
);
2787 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2790 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2795 * Copy the vma structure to a new location in the same mm,
2796 * prior to moving page table entries, to effect an mremap move.
2798 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
2799 unsigned long addr
, unsigned long len
, pgoff_t pgoff
,
2800 bool *need_rmap_locks
)
2802 struct vm_area_struct
*vma
= *vmap
;
2803 unsigned long vma_start
= vma
->vm_start
;
2804 struct mm_struct
*mm
= vma
->vm_mm
;
2805 struct vm_area_struct
*new_vma
, *prev
;
2806 struct rb_node
**rb_link
, *rb_parent
;
2807 bool faulted_in_anon_vma
= true;
2810 * If anonymous vma has not yet been faulted, update new pgoff
2811 * to match new location, to increase its chance of merging.
2813 if (unlikely(vma_is_anonymous(vma
) && !vma
->anon_vma
)) {
2814 pgoff
= addr
>> PAGE_SHIFT
;
2815 faulted_in_anon_vma
= false;
2818 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
2819 return NULL
; /* should never get here */
2820 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
2821 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
),
2822 vma
->vm_userfaultfd_ctx
);
2825 * Source vma may have been merged into new_vma
2827 if (unlikely(vma_start
>= new_vma
->vm_start
&&
2828 vma_start
< new_vma
->vm_end
)) {
2830 * The only way we can get a vma_merge with
2831 * self during an mremap is if the vma hasn't
2832 * been faulted in yet and we were allowed to
2833 * reset the dst vma->vm_pgoff to the
2834 * destination address of the mremap to allow
2835 * the merge to happen. mremap must change the
2836 * vm_pgoff linearity between src and dst vmas
2837 * (in turn preventing a vma_merge) to be
2838 * safe. It is only safe to keep the vm_pgoff
2839 * linear if there are no pages mapped yet.
2841 VM_BUG_ON_VMA(faulted_in_anon_vma
, new_vma
);
2842 *vmap
= vma
= new_vma
;
2844 *need_rmap_locks
= (new_vma
->vm_pgoff
<= vma
->vm_pgoff
);
2846 new_vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2850 new_vma
->vm_start
= addr
;
2851 new_vma
->vm_end
= addr
+ len
;
2852 new_vma
->vm_pgoff
= pgoff
;
2853 if (vma_dup_policy(vma
, new_vma
))
2855 INIT_LIST_HEAD(&new_vma
->anon_vma_chain
);
2856 if (anon_vma_clone(new_vma
, vma
))
2857 goto out_free_mempol
;
2858 if (new_vma
->vm_file
)
2859 get_file(new_vma
->vm_file
);
2860 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
2861 new_vma
->vm_ops
->open(new_vma
);
2862 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
2863 *need_rmap_locks
= false;
2868 mpol_put(vma_policy(new_vma
));
2870 kmem_cache_free(vm_area_cachep
, new_vma
);
2876 * Return true if the calling process may expand its vm space by the passed
2879 bool may_expand_vm(struct mm_struct
*mm
, vm_flags_t flags
, unsigned long npages
)
2881 if (mm
->total_vm
+ npages
> rlimit(RLIMIT_AS
) >> PAGE_SHIFT
)
2884 if (is_data_mapping(flags
) &&
2885 mm
->data_vm
+ npages
> rlimit(RLIMIT_DATA
) >> PAGE_SHIFT
) {
2886 if (ignore_rlimit_data
)
2887 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Will be forbidden soon.\n",
2888 current
->comm
, current
->pid
,
2889 (mm
->data_vm
+ npages
) << PAGE_SHIFT
,
2890 rlimit(RLIMIT_DATA
));
2898 void vm_stat_account(struct mm_struct
*mm
, vm_flags_t flags
, long npages
)
2900 mm
->total_vm
+= npages
;
2902 if (is_exec_mapping(flags
))
2903 mm
->exec_vm
+= npages
;
2904 else if (is_stack_mapping(flags
))
2905 mm
->stack_vm
+= npages
;
2906 else if (is_data_mapping(flags
))
2907 mm
->data_vm
+= npages
;
2910 static int special_mapping_fault(struct vm_area_struct
*vma
,
2911 struct vm_fault
*vmf
);
2914 * Having a close hook prevents vma merging regardless of flags.
2916 static void special_mapping_close(struct vm_area_struct
*vma
)
2920 static const char *special_mapping_name(struct vm_area_struct
*vma
)
2922 return ((struct vm_special_mapping
*)vma
->vm_private_data
)->name
;
2925 static const struct vm_operations_struct special_mapping_vmops
= {
2926 .close
= special_mapping_close
,
2927 .fault
= special_mapping_fault
,
2928 .name
= special_mapping_name
,
2931 static const struct vm_operations_struct legacy_special_mapping_vmops
= {
2932 .close
= special_mapping_close
,
2933 .fault
= special_mapping_fault
,
2936 static int special_mapping_fault(struct vm_area_struct
*vma
,
2937 struct vm_fault
*vmf
)
2940 struct page
**pages
;
2942 if (vma
->vm_ops
== &legacy_special_mapping_vmops
) {
2943 pages
= vma
->vm_private_data
;
2945 struct vm_special_mapping
*sm
= vma
->vm_private_data
;
2948 return sm
->fault(sm
, vma
, vmf
);
2953 for (pgoff
= vmf
->pgoff
; pgoff
&& *pages
; ++pages
)
2957 struct page
*page
= *pages
;
2963 return VM_FAULT_SIGBUS
;
2966 static struct vm_area_struct
*__install_special_mapping(
2967 struct mm_struct
*mm
,
2968 unsigned long addr
, unsigned long len
,
2969 unsigned long vm_flags
, void *priv
,
2970 const struct vm_operations_struct
*ops
)
2973 struct vm_area_struct
*vma
;
2975 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2976 if (unlikely(vma
== NULL
))
2977 return ERR_PTR(-ENOMEM
);
2979 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2981 vma
->vm_start
= addr
;
2982 vma
->vm_end
= addr
+ len
;
2984 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
| VM_SOFTDIRTY
;
2985 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
2988 vma
->vm_private_data
= priv
;
2990 ret
= insert_vm_struct(mm
, vma
);
2994 vm_stat_account(mm
, vma
->vm_flags
, len
>> PAGE_SHIFT
);
2996 perf_event_mmap(vma
);
3001 kmem_cache_free(vm_area_cachep
, vma
);
3002 return ERR_PTR(ret
);
3006 * Called with mm->mmap_sem held for writing.
3007 * Insert a new vma covering the given region, with the given flags.
3008 * Its pages are supplied by the given array of struct page *.
3009 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3010 * The region past the last page supplied will always produce SIGBUS.
3011 * The array pointer and the pages it points to are assumed to stay alive
3012 * for as long as this mapping might exist.
3014 struct vm_area_struct
*_install_special_mapping(
3015 struct mm_struct
*mm
,
3016 unsigned long addr
, unsigned long len
,
3017 unsigned long vm_flags
, const struct vm_special_mapping
*spec
)
3019 return __install_special_mapping(mm
, addr
, len
, vm_flags
, (void *)spec
,
3020 &special_mapping_vmops
);
3023 int install_special_mapping(struct mm_struct
*mm
,
3024 unsigned long addr
, unsigned long len
,
3025 unsigned long vm_flags
, struct page
**pages
)
3027 struct vm_area_struct
*vma
= __install_special_mapping(
3028 mm
, addr
, len
, vm_flags
, (void *)pages
,
3029 &legacy_special_mapping_vmops
);
3031 return PTR_ERR_OR_ZERO(vma
);
3034 static DEFINE_MUTEX(mm_all_locks_mutex
);
3036 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
3038 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3040 * The LSB of head.next can't change from under us
3041 * because we hold the mm_all_locks_mutex.
3043 down_write_nest_lock(&anon_vma
->root
->rwsem
, &mm
->mmap_sem
);
3045 * We can safely modify head.next after taking the
3046 * anon_vma->root->rwsem. If some other vma in this mm shares
3047 * the same anon_vma we won't take it again.
3049 * No need of atomic instructions here, head.next
3050 * can't change from under us thanks to the
3051 * anon_vma->root->rwsem.
3053 if (__test_and_set_bit(0, (unsigned long *)
3054 &anon_vma
->root
->rb_root
.rb_node
))
3059 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
3061 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3063 * AS_MM_ALL_LOCKS can't change from under us because
3064 * we hold the mm_all_locks_mutex.
3066 * Operations on ->flags have to be atomic because
3067 * even if AS_MM_ALL_LOCKS is stable thanks to the
3068 * mm_all_locks_mutex, there may be other cpus
3069 * changing other bitflags in parallel to us.
3071 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
3073 down_write_nest_lock(&mapping
->i_mmap_rwsem
, &mm
->mmap_sem
);
3078 * This operation locks against the VM for all pte/vma/mm related
3079 * operations that could ever happen on a certain mm. This includes
3080 * vmtruncate, try_to_unmap, and all page faults.
3082 * The caller must take the mmap_sem in write mode before calling
3083 * mm_take_all_locks(). The caller isn't allowed to release the
3084 * mmap_sem until mm_drop_all_locks() returns.
3086 * mmap_sem in write mode is required in order to block all operations
3087 * that could modify pagetables and free pages without need of
3088 * altering the vma layout. It's also needed in write mode to avoid new
3089 * anon_vmas to be associated with existing vmas.
3091 * A single task can't take more than one mm_take_all_locks() in a row
3092 * or it would deadlock.
3094 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3095 * mapping->flags avoid to take the same lock twice, if more than one
3096 * vma in this mm is backed by the same anon_vma or address_space.
3098 * We take locks in following order, accordingly to comment at beginning
3100 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3102 * - all i_mmap_rwsem locks;
3103 * - all anon_vma->rwseml
3105 * We can take all locks within these types randomly because the VM code
3106 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3107 * mm_all_locks_mutex.
3109 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3110 * that may have to take thousand of locks.
3112 * mm_take_all_locks() can fail if it's interrupted by signals.
3114 int mm_take_all_locks(struct mm_struct
*mm
)
3116 struct vm_area_struct
*vma
;
3117 struct anon_vma_chain
*avc
;
3119 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3121 mutex_lock(&mm_all_locks_mutex
);
3123 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3124 if (signal_pending(current
))
3126 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
3127 is_vm_hugetlb_page(vma
))
3128 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3131 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3132 if (signal_pending(current
))
3134 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
3135 !is_vm_hugetlb_page(vma
))
3136 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3139 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3140 if (signal_pending(current
))
3143 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3144 vm_lock_anon_vma(mm
, avc
->anon_vma
);
3150 mm_drop_all_locks(mm
);
3154 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
3156 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3158 * The LSB of head.next can't change to 0 from under
3159 * us because we hold the mm_all_locks_mutex.
3161 * We must however clear the bitflag before unlocking
3162 * the vma so the users using the anon_vma->rb_root will
3163 * never see our bitflag.
3165 * No need of atomic instructions here, head.next
3166 * can't change from under us until we release the
3167 * anon_vma->root->rwsem.
3169 if (!__test_and_clear_bit(0, (unsigned long *)
3170 &anon_vma
->root
->rb_root
.rb_node
))
3172 anon_vma_unlock_write(anon_vma
);
3176 static void vm_unlock_mapping(struct address_space
*mapping
)
3178 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3180 * AS_MM_ALL_LOCKS can't change to 0 from under us
3181 * because we hold the mm_all_locks_mutex.
3183 i_mmap_unlock_write(mapping
);
3184 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
3191 * The mmap_sem cannot be released by the caller until
3192 * mm_drop_all_locks() returns.
3194 void mm_drop_all_locks(struct mm_struct
*mm
)
3196 struct vm_area_struct
*vma
;
3197 struct anon_vma_chain
*avc
;
3199 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3200 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
3202 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3204 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3205 vm_unlock_anon_vma(avc
->anon_vma
);
3206 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3207 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
3210 mutex_unlock(&mm_all_locks_mutex
);
3214 * initialise the VMA slab
3216 void __init
mmap_init(void)
3220 ret
= percpu_counter_init(&vm_committed_as
, 0, GFP_KERNEL
);
3225 * Initialise sysctl_user_reserve_kbytes.
3227 * This is intended to prevent a user from starting a single memory hogging
3228 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3231 * The default value is min(3% of free memory, 128MB)
3232 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3234 static int init_user_reserve(void)
3236 unsigned long free_kbytes
;
3238 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3240 sysctl_user_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 17);
3243 subsys_initcall(init_user_reserve
);
3246 * Initialise sysctl_admin_reserve_kbytes.
3248 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3249 * to log in and kill a memory hogging process.
3251 * Systems with more than 256MB will reserve 8MB, enough to recover
3252 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3253 * only reserve 3% of free pages by default.
3255 static int init_admin_reserve(void)
3257 unsigned long free_kbytes
;
3259 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3261 sysctl_admin_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 13);
3264 subsys_initcall(init_admin_reserve
);
3267 * Reinititalise user and admin reserves if memory is added or removed.
3269 * The default user reserve max is 128MB, and the default max for the
3270 * admin reserve is 8MB. These are usually, but not always, enough to
3271 * enable recovery from a memory hogging process using login/sshd, a shell,
3272 * and tools like top. It may make sense to increase or even disable the
3273 * reserve depending on the existence of swap or variations in the recovery
3274 * tools. So, the admin may have changed them.
3276 * If memory is added and the reserves have been eliminated or increased above
3277 * the default max, then we'll trust the admin.
3279 * If memory is removed and there isn't enough free memory, then we
3280 * need to reset the reserves.
3282 * Otherwise keep the reserve set by the admin.
3284 static int reserve_mem_notifier(struct notifier_block
*nb
,
3285 unsigned long action
, void *data
)
3287 unsigned long tmp
, free_kbytes
;
3291 /* Default max is 128MB. Leave alone if modified by operator. */
3292 tmp
= sysctl_user_reserve_kbytes
;
3293 if (0 < tmp
&& tmp
< (1UL << 17))
3294 init_user_reserve();
3296 /* Default max is 8MB. Leave alone if modified by operator. */
3297 tmp
= sysctl_admin_reserve_kbytes
;
3298 if (0 < tmp
&& tmp
< (1UL << 13))
3299 init_admin_reserve();
3303 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3305 if (sysctl_user_reserve_kbytes
> free_kbytes
) {
3306 init_user_reserve();
3307 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3308 sysctl_user_reserve_kbytes
);
3311 if (sysctl_admin_reserve_kbytes
> free_kbytes
) {
3312 init_admin_reserve();
3313 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3314 sysctl_admin_reserve_kbytes
);
3323 static struct notifier_block reserve_mem_nb
= {
3324 .notifier_call
= reserve_mem_notifier
,
3327 static int __meminit
init_reserve_notifier(void)
3329 if (register_hotmemory_notifier(&reserve_mem_nb
))
3330 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3334 subsys_initcall(init_reserve_notifier
);