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 <linux/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
;
119 pgprot_t vm_page_prot
;
121 vm_page_prot
= vm_pgprot_modify(vma
->vm_page_prot
, vm_flags
);
122 if (vma_wants_writenotify(vma
, vm_page_prot
)) {
123 vm_flags
&= ~VM_SHARED
;
124 vm_page_prot
= vm_pgprot_modify(vm_page_prot
, vm_flags
);
126 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
127 WRITE_ONCE(vma
->vm_page_prot
, vm_page_prot
);
131 * Requires inode->i_mapping->i_mmap_rwsem
133 static void __remove_shared_vm_struct(struct vm_area_struct
*vma
,
134 struct file
*file
, struct address_space
*mapping
)
136 if (vma
->vm_flags
& VM_DENYWRITE
)
137 atomic_inc(&file_inode(file
)->i_writecount
);
138 if (vma
->vm_flags
& VM_SHARED
)
139 mapping_unmap_writable(mapping
);
141 flush_dcache_mmap_lock(mapping
);
142 vma_interval_tree_remove(vma
, &mapping
->i_mmap
);
143 flush_dcache_mmap_unlock(mapping
);
147 * Unlink a file-based vm structure from its interval tree, to hide
148 * vma from rmap and vmtruncate before freeing its page tables.
150 void unlink_file_vma(struct vm_area_struct
*vma
)
152 struct file
*file
= vma
->vm_file
;
155 struct address_space
*mapping
= file
->f_mapping
;
156 i_mmap_lock_write(mapping
);
157 __remove_shared_vm_struct(vma
, file
, mapping
);
158 i_mmap_unlock_write(mapping
);
163 * Close a vm structure and free it, returning the next.
165 static struct vm_area_struct
*remove_vma(struct vm_area_struct
*vma
)
167 struct vm_area_struct
*next
= vma
->vm_next
;
170 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
171 vma
->vm_ops
->close(vma
);
174 mpol_put(vma_policy(vma
));
175 kmem_cache_free(vm_area_cachep
, vma
);
179 static int do_brk(unsigned long addr
, unsigned long len
);
181 SYSCALL_DEFINE1(brk
, unsigned long, brk
)
183 unsigned long retval
;
184 unsigned long newbrk
, oldbrk
;
185 struct mm_struct
*mm
= current
->mm
;
186 unsigned long min_brk
;
189 if (down_write_killable(&mm
->mmap_sem
))
192 #ifdef CONFIG_COMPAT_BRK
194 * CONFIG_COMPAT_BRK can still be overridden by setting
195 * randomize_va_space to 2, which will still cause mm->start_brk
196 * to be arbitrarily shifted
198 if (current
->brk_randomized
)
199 min_brk
= mm
->start_brk
;
201 min_brk
= mm
->end_data
;
203 min_brk
= mm
->start_brk
;
209 * Check against rlimit here. If this check is done later after the test
210 * of oldbrk with newbrk then it can escape the test and let the data
211 * segment grow beyond its set limit the in case where the limit is
212 * not page aligned -Ram Gupta
214 if (check_data_rlimit(rlimit(RLIMIT_DATA
), brk
, mm
->start_brk
,
215 mm
->end_data
, mm
->start_data
))
218 newbrk
= PAGE_ALIGN(brk
);
219 oldbrk
= PAGE_ALIGN(mm
->brk
);
220 if (oldbrk
== newbrk
)
223 /* Always allow shrinking brk. */
224 if (brk
<= mm
->brk
) {
225 if (!do_munmap(mm
, newbrk
, oldbrk
-newbrk
))
230 /* Check against existing mmap mappings. */
231 if (find_vma_intersection(mm
, oldbrk
, newbrk
+PAGE_SIZE
))
234 /* Ok, looks good - let it rip. */
235 if (do_brk(oldbrk
, newbrk
-oldbrk
) < 0)
240 populate
= newbrk
> oldbrk
&& (mm
->def_flags
& VM_LOCKED
) != 0;
241 up_write(&mm
->mmap_sem
);
243 mm_populate(oldbrk
, newbrk
- oldbrk
);
248 up_write(&mm
->mmap_sem
);
252 static long vma_compute_subtree_gap(struct vm_area_struct
*vma
)
254 unsigned long max
, subtree_gap
;
257 max
-= vma
->vm_prev
->vm_end
;
258 if (vma
->vm_rb
.rb_left
) {
259 subtree_gap
= rb_entry(vma
->vm_rb
.rb_left
,
260 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
261 if (subtree_gap
> max
)
264 if (vma
->vm_rb
.rb_right
) {
265 subtree_gap
= rb_entry(vma
->vm_rb
.rb_right
,
266 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
267 if (subtree_gap
> max
)
273 #ifdef CONFIG_DEBUG_VM_RB
274 static int browse_rb(struct mm_struct
*mm
)
276 struct rb_root
*root
= &mm
->mm_rb
;
277 int i
= 0, j
, bug
= 0;
278 struct rb_node
*nd
, *pn
= NULL
;
279 unsigned long prev
= 0, pend
= 0;
281 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
282 struct vm_area_struct
*vma
;
283 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
284 if (vma
->vm_start
< prev
) {
285 pr_emerg("vm_start %lx < prev %lx\n",
286 vma
->vm_start
, prev
);
289 if (vma
->vm_start
< pend
) {
290 pr_emerg("vm_start %lx < pend %lx\n",
291 vma
->vm_start
, pend
);
294 if (vma
->vm_start
> vma
->vm_end
) {
295 pr_emerg("vm_start %lx > vm_end %lx\n",
296 vma
->vm_start
, vma
->vm_end
);
299 spin_lock(&mm
->page_table_lock
);
300 if (vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
)) {
301 pr_emerg("free gap %lx, correct %lx\n",
303 vma_compute_subtree_gap(vma
));
306 spin_unlock(&mm
->page_table_lock
);
309 prev
= vma
->vm_start
;
313 for (nd
= pn
; nd
; nd
= rb_prev(nd
))
316 pr_emerg("backwards %d, forwards %d\n", j
, i
);
322 static void validate_mm_rb(struct rb_root
*root
, struct vm_area_struct
*ignore
)
326 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
327 struct vm_area_struct
*vma
;
328 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
329 VM_BUG_ON_VMA(vma
!= ignore
&&
330 vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
),
335 static void validate_mm(struct mm_struct
*mm
)
339 unsigned long highest_address
= 0;
340 struct vm_area_struct
*vma
= mm
->mmap
;
343 struct anon_vma
*anon_vma
= vma
->anon_vma
;
344 struct anon_vma_chain
*avc
;
347 anon_vma_lock_read(anon_vma
);
348 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
349 anon_vma_interval_tree_verify(avc
);
350 anon_vma_unlock_read(anon_vma
);
353 highest_address
= vma
->vm_end
;
357 if (i
!= mm
->map_count
) {
358 pr_emerg("map_count %d vm_next %d\n", mm
->map_count
, i
);
361 if (highest_address
!= mm
->highest_vm_end
) {
362 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
363 mm
->highest_vm_end
, highest_address
);
367 if (i
!= mm
->map_count
) {
369 pr_emerg("map_count %d rb %d\n", mm
->map_count
, i
);
372 VM_BUG_ON_MM(bug
, mm
);
375 #define validate_mm_rb(root, ignore) do { } while (0)
376 #define validate_mm(mm) do { } while (0)
379 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks
, struct vm_area_struct
, vm_rb
,
380 unsigned long, rb_subtree_gap
, vma_compute_subtree_gap
)
383 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
384 * vma->vm_prev->vm_end values changed, without modifying the vma's position
387 static void vma_gap_update(struct vm_area_struct
*vma
)
390 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
391 * function that does exacltly what we want.
393 vma_gap_callbacks_propagate(&vma
->vm_rb
, NULL
);
396 static inline void vma_rb_insert(struct vm_area_struct
*vma
,
397 struct rb_root
*root
)
399 /* All rb_subtree_gap values must be consistent prior to insertion */
400 validate_mm_rb(root
, NULL
);
402 rb_insert_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
405 static void __vma_rb_erase(struct vm_area_struct
*vma
, struct rb_root
*root
)
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
);
415 static __always_inline
void vma_rb_erase_ignore(struct vm_area_struct
*vma
,
416 struct rb_root
*root
,
417 struct vm_area_struct
*ignore
)
420 * All rb_subtree_gap values must be consistent prior to erase,
421 * with the possible exception of the "next" vma being erased if
422 * next->vm_start was reduced.
424 validate_mm_rb(root
, ignore
);
426 __vma_rb_erase(vma
, root
);
429 static __always_inline
void vma_rb_erase(struct vm_area_struct
*vma
,
430 struct rb_root
*root
)
433 * All rb_subtree_gap values must be consistent prior to erase,
434 * with the possible exception of the vma being erased.
436 validate_mm_rb(root
, vma
);
438 __vma_rb_erase(vma
, root
);
442 * vma has some anon_vma assigned, and is already inserted on that
443 * anon_vma's interval trees.
445 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
446 * vma must be removed from the anon_vma's interval trees using
447 * anon_vma_interval_tree_pre_update_vma().
449 * After the update, the vma will be reinserted using
450 * anon_vma_interval_tree_post_update_vma().
452 * The entire update must be protected by exclusive mmap_sem and by
453 * the root anon_vma's mutex.
456 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct
*vma
)
458 struct anon_vma_chain
*avc
;
460 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
461 anon_vma_interval_tree_remove(avc
, &avc
->anon_vma
->rb_root
);
465 anon_vma_interval_tree_post_update_vma(struct vm_area_struct
*vma
)
467 struct anon_vma_chain
*avc
;
469 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
470 anon_vma_interval_tree_insert(avc
, &avc
->anon_vma
->rb_root
);
473 static int find_vma_links(struct mm_struct
*mm
, unsigned long addr
,
474 unsigned long end
, struct vm_area_struct
**pprev
,
475 struct rb_node
***rb_link
, struct rb_node
**rb_parent
)
477 struct rb_node
**__rb_link
, *__rb_parent
, *rb_prev
;
479 __rb_link
= &mm
->mm_rb
.rb_node
;
480 rb_prev
= __rb_parent
= NULL
;
483 struct vm_area_struct
*vma_tmp
;
485 __rb_parent
= *__rb_link
;
486 vma_tmp
= rb_entry(__rb_parent
, struct vm_area_struct
, vm_rb
);
488 if (vma_tmp
->vm_end
> addr
) {
489 /* Fail if an existing vma overlaps the area */
490 if (vma_tmp
->vm_start
< end
)
492 __rb_link
= &__rb_parent
->rb_left
;
494 rb_prev
= __rb_parent
;
495 __rb_link
= &__rb_parent
->rb_right
;
501 *pprev
= rb_entry(rb_prev
, struct vm_area_struct
, vm_rb
);
502 *rb_link
= __rb_link
;
503 *rb_parent
= __rb_parent
;
507 static unsigned long count_vma_pages_range(struct mm_struct
*mm
,
508 unsigned long addr
, unsigned long end
)
510 unsigned long nr_pages
= 0;
511 struct vm_area_struct
*vma
;
513 /* Find first overlaping mapping */
514 vma
= find_vma_intersection(mm
, addr
, end
);
518 nr_pages
= (min(end
, vma
->vm_end
) -
519 max(addr
, vma
->vm_start
)) >> PAGE_SHIFT
;
521 /* Iterate over the rest of the overlaps */
522 for (vma
= vma
->vm_next
; vma
; vma
= vma
->vm_next
) {
523 unsigned long overlap_len
;
525 if (vma
->vm_start
> end
)
528 overlap_len
= min(end
, vma
->vm_end
) - vma
->vm_start
;
529 nr_pages
+= overlap_len
>> PAGE_SHIFT
;
535 void __vma_link_rb(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
536 struct rb_node
**rb_link
, struct rb_node
*rb_parent
)
538 /* Update tracking information for the gap following the new vma. */
540 vma_gap_update(vma
->vm_next
);
542 mm
->highest_vm_end
= vma
->vm_end
;
545 * vma->vm_prev wasn't known when we followed the rbtree to find the
546 * correct insertion point for that vma. As a result, we could not
547 * update the vma vm_rb parents rb_subtree_gap values on the way down.
548 * So, we first insert the vma with a zero rb_subtree_gap value
549 * (to be consistent with what we did on the way down), and then
550 * immediately update the gap to the correct value. Finally we
551 * rebalance the rbtree after all augmented values have been set.
553 rb_link_node(&vma
->vm_rb
, rb_parent
, rb_link
);
554 vma
->rb_subtree_gap
= 0;
556 vma_rb_insert(vma
, &mm
->mm_rb
);
559 static void __vma_link_file(struct vm_area_struct
*vma
)
565 struct address_space
*mapping
= file
->f_mapping
;
567 if (vma
->vm_flags
& VM_DENYWRITE
)
568 atomic_dec(&file_inode(file
)->i_writecount
);
569 if (vma
->vm_flags
& VM_SHARED
)
570 atomic_inc(&mapping
->i_mmap_writable
);
572 flush_dcache_mmap_lock(mapping
);
573 vma_interval_tree_insert(vma
, &mapping
->i_mmap
);
574 flush_dcache_mmap_unlock(mapping
);
579 __vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
580 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
581 struct rb_node
*rb_parent
)
583 __vma_link_list(mm
, vma
, prev
, rb_parent
);
584 __vma_link_rb(mm
, vma
, rb_link
, rb_parent
);
587 static void vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
588 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
589 struct rb_node
*rb_parent
)
591 struct address_space
*mapping
= NULL
;
594 mapping
= vma
->vm_file
->f_mapping
;
595 i_mmap_lock_write(mapping
);
598 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
599 __vma_link_file(vma
);
602 i_mmap_unlock_write(mapping
);
609 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
610 * mm's list and rbtree. It has already been inserted into the interval tree.
612 static void __insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
614 struct vm_area_struct
*prev
;
615 struct rb_node
**rb_link
, *rb_parent
;
617 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
618 &prev
, &rb_link
, &rb_parent
))
620 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
624 static __always_inline
void __vma_unlink_common(struct mm_struct
*mm
,
625 struct vm_area_struct
*vma
,
626 struct vm_area_struct
*prev
,
628 struct vm_area_struct
*ignore
)
630 struct vm_area_struct
*next
;
632 vma_rb_erase_ignore(vma
, &mm
->mm_rb
, ignore
);
635 prev
->vm_next
= next
;
639 prev
->vm_next
= next
;
644 next
->vm_prev
= prev
;
647 vmacache_invalidate(mm
);
650 static inline void __vma_unlink_prev(struct mm_struct
*mm
,
651 struct vm_area_struct
*vma
,
652 struct vm_area_struct
*prev
)
654 __vma_unlink_common(mm
, vma
, prev
, true, vma
);
658 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
659 * is already present in an i_mmap tree without adjusting the tree.
660 * The following helper function should be used when such adjustments
661 * are necessary. The "insert" vma (if any) is to be inserted
662 * before we drop the necessary locks.
664 int __vma_adjust(struct vm_area_struct
*vma
, unsigned long start
,
665 unsigned long end
, pgoff_t pgoff
, struct vm_area_struct
*insert
,
666 struct vm_area_struct
*expand
)
668 struct mm_struct
*mm
= vma
->vm_mm
;
669 struct vm_area_struct
*next
= vma
->vm_next
, *orig_vma
= vma
;
670 struct address_space
*mapping
= NULL
;
671 struct rb_root
*root
= NULL
;
672 struct anon_vma
*anon_vma
= NULL
;
673 struct file
*file
= vma
->vm_file
;
674 bool start_changed
= false, end_changed
= false;
675 long adjust_next
= 0;
678 if (next
&& !insert
) {
679 struct vm_area_struct
*exporter
= NULL
, *importer
= NULL
;
681 if (end
>= next
->vm_end
) {
683 * vma expands, overlapping all the next, and
684 * perhaps the one after too (mprotect case 6).
685 * The only other cases that gets here are
686 * case 1, case 7 and case 8.
688 if (next
== expand
) {
690 * The only case where we don't expand "vma"
691 * and we expand "next" instead is case 8.
693 VM_WARN_ON(end
!= next
->vm_end
);
695 * remove_next == 3 means we're
696 * removing "vma" and that to do so we
697 * swapped "vma" and "next".
700 VM_WARN_ON(file
!= next
->vm_file
);
703 VM_WARN_ON(expand
!= vma
);
705 * case 1, 6, 7, remove_next == 2 is case 6,
706 * remove_next == 1 is case 1 or 7.
708 remove_next
= 1 + (end
> next
->vm_end
);
709 VM_WARN_ON(remove_next
== 2 &&
710 end
!= next
->vm_next
->vm_end
);
711 VM_WARN_ON(remove_next
== 1 &&
712 end
!= next
->vm_end
);
713 /* trim end to next, for case 6 first pass */
721 * If next doesn't have anon_vma, import from vma after
722 * next, if the vma overlaps with it.
724 if (remove_next
== 2 && !next
->anon_vma
)
725 exporter
= next
->vm_next
;
727 } else if (end
> next
->vm_start
) {
729 * vma expands, overlapping part of the next:
730 * mprotect case 5 shifting the boundary up.
732 adjust_next
= (end
- next
->vm_start
) >> PAGE_SHIFT
;
735 VM_WARN_ON(expand
!= importer
);
736 } else if (end
< vma
->vm_end
) {
738 * vma shrinks, and !insert tells it's not
739 * split_vma inserting another: so it must be
740 * mprotect case 4 shifting the boundary down.
742 adjust_next
= -((vma
->vm_end
- end
) >> PAGE_SHIFT
);
745 VM_WARN_ON(expand
!= importer
);
749 * Easily overlooked: when mprotect shifts the boundary,
750 * make sure the expanding vma has anon_vma set if the
751 * shrinking vma had, to cover any anon pages imported.
753 if (exporter
&& exporter
->anon_vma
&& !importer
->anon_vma
) {
756 importer
->anon_vma
= exporter
->anon_vma
;
757 error
= anon_vma_clone(importer
, exporter
);
763 vma_adjust_trans_huge(orig_vma
, start
, end
, adjust_next
);
766 mapping
= file
->f_mapping
;
767 root
= &mapping
->i_mmap
;
768 uprobe_munmap(vma
, vma
->vm_start
, vma
->vm_end
);
771 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
773 i_mmap_lock_write(mapping
);
776 * Put into interval tree now, so instantiated pages
777 * are visible to arm/parisc __flush_dcache_page
778 * throughout; but we cannot insert into address
779 * space until vma start or end is updated.
781 __vma_link_file(insert
);
785 anon_vma
= vma
->anon_vma
;
786 if (!anon_vma
&& adjust_next
)
787 anon_vma
= next
->anon_vma
;
789 VM_WARN_ON(adjust_next
&& next
->anon_vma
&&
790 anon_vma
!= next
->anon_vma
);
791 anon_vma_lock_write(anon_vma
);
792 anon_vma_interval_tree_pre_update_vma(vma
);
794 anon_vma_interval_tree_pre_update_vma(next
);
798 flush_dcache_mmap_lock(mapping
);
799 vma_interval_tree_remove(vma
, root
);
801 vma_interval_tree_remove(next
, root
);
804 if (start
!= vma
->vm_start
) {
805 vma
->vm_start
= start
;
806 start_changed
= true;
808 if (end
!= vma
->vm_end
) {
812 vma
->vm_pgoff
= pgoff
;
814 next
->vm_start
+= adjust_next
<< PAGE_SHIFT
;
815 next
->vm_pgoff
+= adjust_next
;
820 vma_interval_tree_insert(next
, root
);
821 vma_interval_tree_insert(vma
, root
);
822 flush_dcache_mmap_unlock(mapping
);
827 * vma_merge has merged next into vma, and needs
828 * us to remove next before dropping the locks.
830 if (remove_next
!= 3)
831 __vma_unlink_prev(mm
, next
, vma
);
834 * vma is not before next if they've been
837 * pre-swap() next->vm_start was reduced so
838 * tell validate_mm_rb to ignore pre-swap()
839 * "next" (which is stored in post-swap()
842 __vma_unlink_common(mm
, next
, NULL
, false, vma
);
844 __remove_shared_vm_struct(next
, file
, mapping
);
847 * split_vma has split insert from vma, and needs
848 * us to insert it before dropping the locks
849 * (it may either follow vma or precede it).
851 __insert_vm_struct(mm
, insert
);
857 mm
->highest_vm_end
= end
;
858 else if (!adjust_next
)
859 vma_gap_update(next
);
864 anon_vma_interval_tree_post_update_vma(vma
);
866 anon_vma_interval_tree_post_update_vma(next
);
867 anon_vma_unlock_write(anon_vma
);
870 i_mmap_unlock_write(mapping
);
881 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
885 anon_vma_merge(vma
, next
);
887 mpol_put(vma_policy(next
));
888 kmem_cache_free(vm_area_cachep
, next
);
890 * In mprotect's case 6 (see comments on vma_merge),
891 * we must remove another next too. It would clutter
892 * up the code too much to do both in one go.
894 if (remove_next
!= 3) {
896 * If "next" was removed and vma->vm_end was
897 * expanded (up) over it, in turn
898 * "next->vm_prev->vm_end" changed and the
899 * "vma->vm_next" gap must be updated.
904 * For the scope of the comment "next" and
905 * "vma" considered pre-swap(): if "vma" was
906 * removed, next->vm_start was expanded (down)
907 * over it and the "next" gap must be updated.
908 * Because of the swap() the post-swap() "vma"
909 * actually points to pre-swap() "next"
910 * (post-swap() "next" as opposed is now a
915 if (remove_next
== 2) {
921 vma_gap_update(next
);
924 * If remove_next == 2 we obviously can't
927 * If remove_next == 3 we can't reach this
928 * path because pre-swap() next is always not
929 * NULL. pre-swap() "next" is not being
930 * removed and its next->vm_end is not altered
931 * (and furthermore "end" already matches
932 * next->vm_end in remove_next == 3).
934 * We reach this only in the remove_next == 1
935 * case if the "next" vma that was removed was
936 * the highest vma of the mm. However in such
937 * case next->vm_end == "end" and the extended
938 * "vma" has vma->vm_end == next->vm_end so
939 * mm->highest_vm_end doesn't need any update
940 * in remove_next == 1 case.
942 VM_WARN_ON(mm
->highest_vm_end
!= end
);
954 * If the vma has a ->close operation then the driver probably needs to release
955 * per-vma resources, so we don't attempt to merge those.
957 static inline int is_mergeable_vma(struct vm_area_struct
*vma
,
958 struct file
*file
, unsigned long vm_flags
,
959 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
962 * VM_SOFTDIRTY should not prevent from VMA merging, if we
963 * match the flags but dirty bit -- the caller should mark
964 * merged VMA as dirty. If dirty bit won't be excluded from
965 * comparison, we increase pressue on the memory system forcing
966 * the kernel to generate new VMAs when old one could be
969 if ((vma
->vm_flags
^ vm_flags
) & ~VM_SOFTDIRTY
)
971 if (vma
->vm_file
!= file
)
973 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
975 if (!is_mergeable_vm_userfaultfd_ctx(vma
, vm_userfaultfd_ctx
))
980 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
981 struct anon_vma
*anon_vma2
,
982 struct vm_area_struct
*vma
)
985 * The list_is_singular() test is to avoid merging VMA cloned from
986 * parents. This can improve scalability caused by anon_vma lock.
988 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
989 list_is_singular(&vma
->anon_vma_chain
)))
991 return anon_vma1
== anon_vma2
;
995 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
996 * in front of (at a lower virtual address and file offset than) the vma.
998 * We cannot merge two vmas if they have differently assigned (non-NULL)
999 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1001 * We don't check here for the merged mmap wrapping around the end of pagecache
1002 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1003 * wrap, nor mmaps which cover the final page at index -1UL.
1006 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
1007 struct anon_vma
*anon_vma
, struct file
*file
,
1009 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
1011 if (is_mergeable_vma(vma
, file
, vm_flags
, vm_userfaultfd_ctx
) &&
1012 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
1013 if (vma
->vm_pgoff
== vm_pgoff
)
1020 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1021 * beyond (at a higher virtual address and file offset than) the vma.
1023 * We cannot merge two vmas if they have differently assigned (non-NULL)
1024 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1027 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
1028 struct anon_vma
*anon_vma
, struct file
*file
,
1030 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
1032 if (is_mergeable_vma(vma
, file
, vm_flags
, vm_userfaultfd_ctx
) &&
1033 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
1035 vm_pglen
= vma_pages(vma
);
1036 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
1043 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1044 * whether that can be merged with its predecessor or its successor.
1045 * Or both (it neatly fills a hole).
1047 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1048 * certain not to be mapped by the time vma_merge is called; but when
1049 * called for mprotect, it is certain to be already mapped (either at
1050 * an offset within prev, or at the start of next), and the flags of
1051 * this area are about to be changed to vm_flags - and the no-change
1052 * case has already been eliminated.
1054 * The following mprotect cases have to be considered, where AAAA is
1055 * the area passed down from mprotect_fixup, never extending beyond one
1056 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1058 * AAAA AAAA AAAA AAAA
1059 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1060 * cannot merge might become might become might become
1061 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1062 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1063 * mremap move: PPPPXXXXXXXX 8
1065 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1066 * might become case 1 below case 2 below case 3 below
1068 * It is important for case 8 that the the vma NNNN overlapping the
1069 * region AAAA is never going to extended over XXXX. Instead XXXX must
1070 * be extended in region AAAA and NNNN must be removed. This way in
1071 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1072 * rmap_locks, the properties of the merged vma will be already
1073 * correct for the whole merged range. Some of those properties like
1074 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1075 * be correct for the whole merged range immediately after the
1076 * rmap_locks are released. Otherwise if XXXX would be removed and
1077 * NNNN would be extended over the XXXX range, remove_migration_ptes
1078 * or other rmap walkers (if working on addresses beyond the "end"
1079 * parameter) may establish ptes with the wrong permissions of NNNN
1080 * instead of the right permissions of XXXX.
1082 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
1083 struct vm_area_struct
*prev
, unsigned long addr
,
1084 unsigned long end
, unsigned long vm_flags
,
1085 struct anon_vma
*anon_vma
, struct file
*file
,
1086 pgoff_t pgoff
, struct mempolicy
*policy
,
1087 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
)
1089 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
1090 struct vm_area_struct
*area
, *next
;
1094 * We later require that vma->vm_flags == vm_flags,
1095 * so this tests vma->vm_flags & VM_SPECIAL, too.
1097 if (vm_flags
& VM_SPECIAL
)
1101 next
= prev
->vm_next
;
1105 if (area
&& area
->vm_end
== end
) /* cases 6, 7, 8 */
1106 next
= next
->vm_next
;
1108 /* verify some invariant that must be enforced by the caller */
1109 VM_WARN_ON(prev
&& addr
<= prev
->vm_start
);
1110 VM_WARN_ON(area
&& end
> area
->vm_end
);
1111 VM_WARN_ON(addr
>= end
);
1114 * Can it merge with the predecessor?
1116 if (prev
&& prev
->vm_end
== addr
&&
1117 mpol_equal(vma_policy(prev
), policy
) &&
1118 can_vma_merge_after(prev
, vm_flags
,
1119 anon_vma
, file
, pgoff
,
1120 vm_userfaultfd_ctx
)) {
1122 * OK, it can. Can we now merge in the successor as well?
1124 if (next
&& end
== next
->vm_start
&&
1125 mpol_equal(policy
, vma_policy(next
)) &&
1126 can_vma_merge_before(next
, vm_flags
,
1129 vm_userfaultfd_ctx
) &&
1130 is_mergeable_anon_vma(prev
->anon_vma
,
1131 next
->anon_vma
, NULL
)) {
1133 err
= __vma_adjust(prev
, prev
->vm_start
,
1134 next
->vm_end
, prev
->vm_pgoff
, NULL
,
1136 } else /* cases 2, 5, 7 */
1137 err
= __vma_adjust(prev
, prev
->vm_start
,
1138 end
, prev
->vm_pgoff
, NULL
, prev
);
1141 khugepaged_enter_vma_merge(prev
, vm_flags
);
1146 * Can this new request be merged in front of next?
1148 if (next
&& end
== next
->vm_start
&&
1149 mpol_equal(policy
, vma_policy(next
)) &&
1150 can_vma_merge_before(next
, vm_flags
,
1151 anon_vma
, file
, pgoff
+pglen
,
1152 vm_userfaultfd_ctx
)) {
1153 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
1154 err
= __vma_adjust(prev
, prev
->vm_start
,
1155 addr
, prev
->vm_pgoff
, NULL
, next
);
1156 else { /* cases 3, 8 */
1157 err
= __vma_adjust(area
, addr
, next
->vm_end
,
1158 next
->vm_pgoff
- pglen
, NULL
, next
);
1160 * In case 3 area is already equal to next and
1161 * this is a noop, but in case 8 "area" has
1162 * been removed and next was expanded over it.
1168 khugepaged_enter_vma_merge(area
, vm_flags
);
1176 * Rough compatbility check to quickly see if it's even worth looking
1177 * at sharing an anon_vma.
1179 * They need to have the same vm_file, and the flags can only differ
1180 * in things that mprotect may change.
1182 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1183 * we can merge the two vma's. For example, we refuse to merge a vma if
1184 * there is a vm_ops->close() function, because that indicates that the
1185 * driver is doing some kind of reference counting. But that doesn't
1186 * really matter for the anon_vma sharing case.
1188 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1190 return a
->vm_end
== b
->vm_start
&&
1191 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
1192 a
->vm_file
== b
->vm_file
&&
1193 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_READ
|VM_WRITE
|VM_EXEC
|VM_SOFTDIRTY
)) &&
1194 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
1198 * Do some basic sanity checking to see if we can re-use the anon_vma
1199 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1200 * the same as 'old', the other will be the new one that is trying
1201 * to share the anon_vma.
1203 * NOTE! This runs with mm_sem held for reading, so it is possible that
1204 * the anon_vma of 'old' is concurrently in the process of being set up
1205 * by another page fault trying to merge _that_. But that's ok: if it
1206 * is being set up, that automatically means that it will be a singleton
1207 * acceptable for merging, so we can do all of this optimistically. But
1208 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1210 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1211 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1212 * is to return an anon_vma that is "complex" due to having gone through
1215 * We also make sure that the two vma's are compatible (adjacent,
1216 * and with the same memory policies). That's all stable, even with just
1217 * a read lock on the mm_sem.
1219 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1221 if (anon_vma_compatible(a
, b
)) {
1222 struct anon_vma
*anon_vma
= READ_ONCE(old
->anon_vma
);
1224 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
1231 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1232 * neighbouring vmas for a suitable anon_vma, before it goes off
1233 * to allocate a new anon_vma. It checks because a repetitive
1234 * sequence of mprotects and faults may otherwise lead to distinct
1235 * anon_vmas being allocated, preventing vma merge in subsequent
1238 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
1240 struct anon_vma
*anon_vma
;
1241 struct vm_area_struct
*near
;
1243 near
= vma
->vm_next
;
1247 anon_vma
= reusable_anon_vma(near
, vma
, near
);
1251 near
= vma
->vm_prev
;
1255 anon_vma
= reusable_anon_vma(near
, near
, vma
);
1260 * There's no absolute need to look only at touching neighbours:
1261 * we could search further afield for "compatible" anon_vmas.
1262 * But it would probably just be a waste of time searching,
1263 * or lead to too many vmas hanging off the same anon_vma.
1264 * We're trying to allow mprotect remerging later on,
1265 * not trying to minimize memory used for anon_vmas.
1271 * If a hint addr is less than mmap_min_addr change hint to be as
1272 * low as possible but still greater than mmap_min_addr
1274 static inline unsigned long round_hint_to_min(unsigned long hint
)
1277 if (((void *)hint
!= NULL
) &&
1278 (hint
< mmap_min_addr
))
1279 return PAGE_ALIGN(mmap_min_addr
);
1283 static inline int mlock_future_check(struct mm_struct
*mm
,
1284 unsigned long flags
,
1287 unsigned long locked
, lock_limit
;
1289 /* mlock MCL_FUTURE? */
1290 if (flags
& VM_LOCKED
) {
1291 locked
= len
>> PAGE_SHIFT
;
1292 locked
+= mm
->locked_vm
;
1293 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1294 lock_limit
>>= PAGE_SHIFT
;
1295 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1302 * The caller must hold down_write(¤t->mm->mmap_sem).
1304 unsigned long do_mmap(struct file
*file
, unsigned long addr
,
1305 unsigned long len
, unsigned long prot
,
1306 unsigned long flags
, vm_flags_t vm_flags
,
1307 unsigned long pgoff
, unsigned long *populate
)
1309 struct mm_struct
*mm
= current
->mm
;
1318 * Does the application expect PROT_READ to imply PROT_EXEC?
1320 * (the exception is when the underlying filesystem is noexec
1321 * mounted, in which case we dont add PROT_EXEC.)
1323 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1324 if (!(file
&& path_noexec(&file
->f_path
)))
1327 if (!(flags
& MAP_FIXED
))
1328 addr
= round_hint_to_min(addr
);
1330 /* Careful about overflows.. */
1331 len
= PAGE_ALIGN(len
);
1335 /* offset overflow? */
1336 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1339 /* Too many mappings? */
1340 if (mm
->map_count
> sysctl_max_map_count
)
1343 /* Obtain the address to map to. we verify (or select) it and ensure
1344 * that it represents a valid section of the address space.
1346 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1347 if (offset_in_page(addr
))
1350 if (prot
== PROT_EXEC
) {
1351 pkey
= execute_only_pkey(mm
);
1356 /* Do simple checking here so the lower-level routines won't have
1357 * to. we assume access permissions have been handled by the open
1358 * of the memory object, so we don't do any here.
1360 vm_flags
|= calc_vm_prot_bits(prot
, pkey
) | calc_vm_flag_bits(flags
) |
1361 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1363 if (flags
& MAP_LOCKED
)
1364 if (!can_do_mlock())
1367 if (mlock_future_check(mm
, vm_flags
, len
))
1371 struct inode
*inode
= file_inode(file
);
1373 switch (flags
& MAP_TYPE
) {
1375 if ((prot
&PROT_WRITE
) && !(file
->f_mode
&FMODE_WRITE
))
1379 * Make sure we don't allow writing to an append-only
1382 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1386 * Make sure there are no mandatory locks on the file.
1388 if (locks_verify_locked(file
))
1391 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1392 if (!(file
->f_mode
& FMODE_WRITE
))
1393 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1397 if (!(file
->f_mode
& FMODE_READ
))
1399 if (path_noexec(&file
->f_path
)) {
1400 if (vm_flags
& VM_EXEC
)
1402 vm_flags
&= ~VM_MAYEXEC
;
1405 if (!file
->f_op
->mmap
)
1407 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1415 switch (flags
& MAP_TYPE
) {
1417 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1423 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1427 * Set pgoff according to addr for anon_vma.
1429 pgoff
= addr
>> PAGE_SHIFT
;
1437 * Set 'VM_NORESERVE' if we should not account for the
1438 * memory use of this mapping.
1440 if (flags
& MAP_NORESERVE
) {
1441 /* We honor MAP_NORESERVE if allowed to overcommit */
1442 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1443 vm_flags
|= VM_NORESERVE
;
1445 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1446 if (file
&& is_file_hugepages(file
))
1447 vm_flags
|= VM_NORESERVE
;
1450 addr
= mmap_region(file
, addr
, len
, vm_flags
, pgoff
);
1451 if (!IS_ERR_VALUE(addr
) &&
1452 ((vm_flags
& VM_LOCKED
) ||
1453 (flags
& (MAP_POPULATE
| MAP_NONBLOCK
)) == MAP_POPULATE
))
1458 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1459 unsigned long, prot
, unsigned long, flags
,
1460 unsigned long, fd
, unsigned long, pgoff
)
1462 struct file
*file
= NULL
;
1463 unsigned long retval
;
1465 if (!(flags
& MAP_ANONYMOUS
)) {
1466 audit_mmap_fd(fd
, flags
);
1470 if (is_file_hugepages(file
))
1471 len
= ALIGN(len
, huge_page_size(hstate_file(file
)));
1473 if (unlikely(flags
& MAP_HUGETLB
&& !is_file_hugepages(file
)))
1475 } else if (flags
& MAP_HUGETLB
) {
1476 struct user_struct
*user
= NULL
;
1479 hs
= hstate_sizelog((flags
>> MAP_HUGE_SHIFT
) & SHM_HUGE_MASK
);
1483 len
= ALIGN(len
, huge_page_size(hs
));
1485 * VM_NORESERVE is used because the reservations will be
1486 * taken when vm_ops->mmap() is called
1487 * A dummy user value is used because we are not locking
1488 * memory so no accounting is necessary
1490 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, len
,
1492 &user
, HUGETLB_ANONHUGE_INODE
,
1493 (flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1495 return PTR_ERR(file
);
1498 flags
&= ~(MAP_EXECUTABLE
| MAP_DENYWRITE
);
1500 retval
= vm_mmap_pgoff(file
, addr
, len
, prot
, flags
, pgoff
);
1507 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1508 struct mmap_arg_struct
{
1512 unsigned long flags
;
1514 unsigned long offset
;
1517 SYSCALL_DEFINE1(old_mmap
, struct mmap_arg_struct __user
*, arg
)
1519 struct mmap_arg_struct a
;
1521 if (copy_from_user(&a
, arg
, sizeof(a
)))
1523 if (offset_in_page(a
.offset
))
1526 return sys_mmap_pgoff(a
.addr
, a
.len
, a
.prot
, a
.flags
, a
.fd
,
1527 a
.offset
>> PAGE_SHIFT
);
1529 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1532 * Some shared mappigns will want the pages marked read-only
1533 * to track write events. If so, we'll downgrade vm_page_prot
1534 * to the private version (using protection_map[] without the
1537 int vma_wants_writenotify(struct vm_area_struct
*vma
, pgprot_t vm_page_prot
)
1539 vm_flags_t vm_flags
= vma
->vm_flags
;
1540 const struct vm_operations_struct
*vm_ops
= vma
->vm_ops
;
1542 /* If it was private or non-writable, the write bit is already clear */
1543 if ((vm_flags
& (VM_WRITE
|VM_SHARED
)) != ((VM_WRITE
|VM_SHARED
)))
1546 /* The backer wishes to know when pages are first written to? */
1547 if (vm_ops
&& (vm_ops
->page_mkwrite
|| vm_ops
->pfn_mkwrite
))
1550 /* The open routine did something to the protections that pgprot_modify
1551 * won't preserve? */
1552 if (pgprot_val(vm_page_prot
) !=
1553 pgprot_val(vm_pgprot_modify(vm_page_prot
, vm_flags
)))
1556 /* Do we need to track softdirty? */
1557 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY
) && !(vm_flags
& VM_SOFTDIRTY
))
1560 /* Specialty mapping? */
1561 if (vm_flags
& VM_PFNMAP
)
1564 /* Can the mapping track the dirty pages? */
1565 return vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
1566 mapping_cap_account_dirty(vma
->vm_file
->f_mapping
);
1570 * We account for memory if it's a private writeable mapping,
1571 * not hugepages and VM_NORESERVE wasn't set.
1573 static inline int accountable_mapping(struct file
*file
, vm_flags_t vm_flags
)
1576 * hugetlb has its own accounting separate from the core VM
1577 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1579 if (file
&& is_file_hugepages(file
))
1582 return (vm_flags
& (VM_NORESERVE
| VM_SHARED
| VM_WRITE
)) == VM_WRITE
;
1585 unsigned long mmap_region(struct file
*file
, unsigned long addr
,
1586 unsigned long len
, vm_flags_t vm_flags
, unsigned long pgoff
)
1588 struct mm_struct
*mm
= current
->mm
;
1589 struct vm_area_struct
*vma
, *prev
;
1591 struct rb_node
**rb_link
, *rb_parent
;
1592 unsigned long charged
= 0;
1594 /* Check against address space limit. */
1595 if (!may_expand_vm(mm
, vm_flags
, len
>> PAGE_SHIFT
)) {
1596 unsigned long nr_pages
;
1599 * MAP_FIXED may remove pages of mappings that intersects with
1600 * requested mapping. Account for the pages it would unmap.
1602 nr_pages
= count_vma_pages_range(mm
, addr
, addr
+ len
);
1604 if (!may_expand_vm(mm
, vm_flags
,
1605 (len
>> PAGE_SHIFT
) - nr_pages
))
1609 /* Clear old maps */
1610 while (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
,
1612 if (do_munmap(mm
, addr
, len
))
1617 * Private writable mapping: check memory availability
1619 if (accountable_mapping(file
, vm_flags
)) {
1620 charged
= len
>> PAGE_SHIFT
;
1621 if (security_vm_enough_memory_mm(mm
, charged
))
1623 vm_flags
|= VM_ACCOUNT
;
1627 * Can we just expand an old mapping?
1629 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
,
1630 NULL
, file
, pgoff
, NULL
, NULL_VM_UFFD_CTX
);
1635 * Determine the object being mapped and call the appropriate
1636 * specific mapper. the address has already been validated, but
1637 * not unmapped, but the maps are removed from the list.
1639 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
1646 vma
->vm_start
= addr
;
1647 vma
->vm_end
= addr
+ len
;
1648 vma
->vm_flags
= vm_flags
;
1649 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1650 vma
->vm_pgoff
= pgoff
;
1651 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
1654 if (vm_flags
& VM_DENYWRITE
) {
1655 error
= deny_write_access(file
);
1659 if (vm_flags
& VM_SHARED
) {
1660 error
= mapping_map_writable(file
->f_mapping
);
1662 goto allow_write_and_free_vma
;
1665 /* ->mmap() can change vma->vm_file, but must guarantee that
1666 * vma_link() below can deny write-access if VM_DENYWRITE is set
1667 * and map writably if VM_SHARED is set. This usually means the
1668 * new file must not have been exposed to user-space, yet.
1670 vma
->vm_file
= get_file(file
);
1671 error
= file
->f_op
->mmap(file
, vma
);
1673 goto unmap_and_free_vma
;
1675 /* Can addr have changed??
1677 * Answer: Yes, several device drivers can do it in their
1678 * f_op->mmap method. -DaveM
1679 * Bug: If addr is changed, prev, rb_link, rb_parent should
1680 * be updated for vma_link()
1682 WARN_ON_ONCE(addr
!= vma
->vm_start
);
1684 addr
= vma
->vm_start
;
1685 vm_flags
= vma
->vm_flags
;
1686 } else if (vm_flags
& VM_SHARED
) {
1687 error
= shmem_zero_setup(vma
);
1692 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1693 /* Once vma denies write, undo our temporary denial count */
1695 if (vm_flags
& VM_SHARED
)
1696 mapping_unmap_writable(file
->f_mapping
);
1697 if (vm_flags
& VM_DENYWRITE
)
1698 allow_write_access(file
);
1700 file
= vma
->vm_file
;
1702 perf_event_mmap(vma
);
1704 vm_stat_account(mm
, vm_flags
, len
>> PAGE_SHIFT
);
1705 if (vm_flags
& VM_LOCKED
) {
1706 if (!((vm_flags
& VM_SPECIAL
) || is_vm_hugetlb_page(vma
) ||
1707 vma
== get_gate_vma(current
->mm
)))
1708 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1710 vma
->vm_flags
&= VM_LOCKED_CLEAR_MASK
;
1717 * New (or expanded) vma always get soft dirty status.
1718 * Otherwise user-space soft-dirty page tracker won't
1719 * be able to distinguish situation when vma area unmapped,
1720 * then new mapped in-place (which must be aimed as
1721 * a completely new data area).
1723 vma
->vm_flags
|= VM_SOFTDIRTY
;
1725 vma_set_page_prot(vma
);
1731 vma
->vm_file
= NULL
;
1733 /* Undo any partial mapping done by a device driver. */
1734 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1736 if (vm_flags
& VM_SHARED
)
1737 mapping_unmap_writable(file
->f_mapping
);
1738 allow_write_and_free_vma
:
1739 if (vm_flags
& VM_DENYWRITE
)
1740 allow_write_access(file
);
1742 kmem_cache_free(vm_area_cachep
, vma
);
1745 vm_unacct_memory(charged
);
1749 unsigned long unmapped_area(struct vm_unmapped_area_info
*info
)
1752 * We implement the search by looking for an rbtree node that
1753 * immediately follows a suitable gap. That is,
1754 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1755 * - gap_end = vma->vm_start >= info->low_limit + length;
1756 * - gap_end - gap_start >= length
1759 struct mm_struct
*mm
= current
->mm
;
1760 struct vm_area_struct
*vma
;
1761 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1763 /* Adjust search length to account for worst case alignment overhead */
1764 length
= info
->length
+ info
->align_mask
;
1765 if (length
< info
->length
)
1768 /* Adjust search limits by the desired length */
1769 if (info
->high_limit
< length
)
1771 high_limit
= info
->high_limit
- length
;
1773 if (info
->low_limit
> high_limit
)
1775 low_limit
= info
->low_limit
+ length
;
1777 /* Check if rbtree root looks promising */
1778 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1780 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1781 if (vma
->rb_subtree_gap
< length
)
1785 /* Visit left subtree if it looks promising */
1786 gap_end
= vma
->vm_start
;
1787 if (gap_end
>= low_limit
&& vma
->vm_rb
.rb_left
) {
1788 struct vm_area_struct
*left
=
1789 rb_entry(vma
->vm_rb
.rb_left
,
1790 struct vm_area_struct
, vm_rb
);
1791 if (left
->rb_subtree_gap
>= length
) {
1797 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1799 /* Check if current node has a suitable gap */
1800 if (gap_start
> high_limit
)
1802 if (gap_end
>= low_limit
&& gap_end
- gap_start
>= length
)
1805 /* Visit right subtree if it looks promising */
1806 if (vma
->vm_rb
.rb_right
) {
1807 struct vm_area_struct
*right
=
1808 rb_entry(vma
->vm_rb
.rb_right
,
1809 struct vm_area_struct
, vm_rb
);
1810 if (right
->rb_subtree_gap
>= length
) {
1816 /* Go back up the rbtree to find next candidate node */
1818 struct rb_node
*prev
= &vma
->vm_rb
;
1819 if (!rb_parent(prev
))
1821 vma
= rb_entry(rb_parent(prev
),
1822 struct vm_area_struct
, vm_rb
);
1823 if (prev
== vma
->vm_rb
.rb_left
) {
1824 gap_start
= vma
->vm_prev
->vm_end
;
1825 gap_end
= vma
->vm_start
;
1832 /* Check highest gap, which does not precede any rbtree node */
1833 gap_start
= mm
->highest_vm_end
;
1834 gap_end
= ULONG_MAX
; /* Only for VM_BUG_ON below */
1835 if (gap_start
> high_limit
)
1839 /* We found a suitable gap. Clip it with the original low_limit. */
1840 if (gap_start
< info
->low_limit
)
1841 gap_start
= info
->low_limit
;
1843 /* Adjust gap address to the desired alignment */
1844 gap_start
+= (info
->align_offset
- gap_start
) & info
->align_mask
;
1846 VM_BUG_ON(gap_start
+ info
->length
> info
->high_limit
);
1847 VM_BUG_ON(gap_start
+ info
->length
> gap_end
);
1851 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info
*info
)
1853 struct mm_struct
*mm
= current
->mm
;
1854 struct vm_area_struct
*vma
;
1855 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1857 /* Adjust search length to account for worst case alignment overhead */
1858 length
= info
->length
+ info
->align_mask
;
1859 if (length
< info
->length
)
1863 * Adjust search limits by the desired length.
1864 * See implementation comment at top of unmapped_area().
1866 gap_end
= info
->high_limit
;
1867 if (gap_end
< length
)
1869 high_limit
= gap_end
- length
;
1871 if (info
->low_limit
> high_limit
)
1873 low_limit
= info
->low_limit
+ length
;
1875 /* Check highest gap, which does not precede any rbtree node */
1876 gap_start
= mm
->highest_vm_end
;
1877 if (gap_start
<= high_limit
)
1880 /* Check if rbtree root looks promising */
1881 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1883 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1884 if (vma
->rb_subtree_gap
< length
)
1888 /* Visit right subtree if it looks promising */
1889 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1890 if (gap_start
<= high_limit
&& vma
->vm_rb
.rb_right
) {
1891 struct vm_area_struct
*right
=
1892 rb_entry(vma
->vm_rb
.rb_right
,
1893 struct vm_area_struct
, vm_rb
);
1894 if (right
->rb_subtree_gap
>= length
) {
1901 /* Check if current node has a suitable gap */
1902 gap_end
= vma
->vm_start
;
1903 if (gap_end
< low_limit
)
1905 if (gap_start
<= high_limit
&& gap_end
- gap_start
>= length
)
1908 /* Visit left subtree if it looks promising */
1909 if (vma
->vm_rb
.rb_left
) {
1910 struct vm_area_struct
*left
=
1911 rb_entry(vma
->vm_rb
.rb_left
,
1912 struct vm_area_struct
, vm_rb
);
1913 if (left
->rb_subtree_gap
>= length
) {
1919 /* Go back up the rbtree to find next candidate node */
1921 struct rb_node
*prev
= &vma
->vm_rb
;
1922 if (!rb_parent(prev
))
1924 vma
= rb_entry(rb_parent(prev
),
1925 struct vm_area_struct
, vm_rb
);
1926 if (prev
== vma
->vm_rb
.rb_right
) {
1927 gap_start
= vma
->vm_prev
?
1928 vma
->vm_prev
->vm_end
: 0;
1935 /* We found a suitable gap. Clip it with the original high_limit. */
1936 if (gap_end
> info
->high_limit
)
1937 gap_end
= info
->high_limit
;
1940 /* Compute highest gap address at the desired alignment */
1941 gap_end
-= info
->length
;
1942 gap_end
-= (gap_end
- info
->align_offset
) & info
->align_mask
;
1944 VM_BUG_ON(gap_end
< info
->low_limit
);
1945 VM_BUG_ON(gap_end
< gap_start
);
1949 /* Get an address range which is currently unmapped.
1950 * For shmat() with addr=0.
1952 * Ugly calling convention alert:
1953 * Return value with the low bits set means error value,
1955 * if (ret & ~PAGE_MASK)
1958 * This function "knows" that -ENOMEM has the bits set.
1960 #ifndef HAVE_ARCH_UNMAPPED_AREA
1962 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
1963 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
1965 struct mm_struct
*mm
= current
->mm
;
1966 struct vm_area_struct
*vma
;
1967 struct vm_unmapped_area_info info
;
1969 if (len
> TASK_SIZE
- mmap_min_addr
)
1972 if (flags
& MAP_FIXED
)
1976 addr
= PAGE_ALIGN(addr
);
1977 vma
= find_vma(mm
, addr
);
1978 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1979 (!vma
|| addr
+ len
<= vma
->vm_start
))
1985 info
.low_limit
= mm
->mmap_base
;
1986 info
.high_limit
= TASK_SIZE
;
1987 info
.align_mask
= 0;
1988 return vm_unmapped_area(&info
);
1993 * This mmap-allocator allocates new areas top-down from below the
1994 * stack's low limit (the base):
1996 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1998 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
1999 const unsigned long len
, const unsigned long pgoff
,
2000 const unsigned long flags
)
2002 struct vm_area_struct
*vma
;
2003 struct mm_struct
*mm
= current
->mm
;
2004 unsigned long addr
= addr0
;
2005 struct vm_unmapped_area_info info
;
2007 /* requested length too big for entire address space */
2008 if (len
> TASK_SIZE
- mmap_min_addr
)
2011 if (flags
& MAP_FIXED
)
2014 /* requesting a specific address */
2016 addr
= PAGE_ALIGN(addr
);
2017 vma
= find_vma(mm
, addr
);
2018 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
2019 (!vma
|| addr
+ len
<= vma
->vm_start
))
2023 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
2025 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
2026 info
.high_limit
= mm
->mmap_base
;
2027 info
.align_mask
= 0;
2028 addr
= vm_unmapped_area(&info
);
2031 * A failed mmap() very likely causes application failure,
2032 * so fall back to the bottom-up function here. This scenario
2033 * can happen with large stack limits and large mmap()
2036 if (offset_in_page(addr
)) {
2037 VM_BUG_ON(addr
!= -ENOMEM
);
2039 info
.low_limit
= TASK_UNMAPPED_BASE
;
2040 info
.high_limit
= TASK_SIZE
;
2041 addr
= vm_unmapped_area(&info
);
2049 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
2050 unsigned long pgoff
, unsigned long flags
)
2052 unsigned long (*get_area
)(struct file
*, unsigned long,
2053 unsigned long, unsigned long, unsigned long);
2055 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
2059 /* Careful about overflows.. */
2060 if (len
> TASK_SIZE
)
2063 get_area
= current
->mm
->get_unmapped_area
;
2065 if (file
->f_op
->get_unmapped_area
)
2066 get_area
= file
->f_op
->get_unmapped_area
;
2067 } else if (flags
& MAP_SHARED
) {
2069 * mmap_region() will call shmem_zero_setup() to create a file,
2070 * so use shmem's get_unmapped_area in case it can be huge.
2071 * do_mmap_pgoff() will clear pgoff, so match alignment.
2074 get_area
= shmem_get_unmapped_area
;
2077 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
2078 if (IS_ERR_VALUE(addr
))
2081 if (addr
> TASK_SIZE
- len
)
2083 if (offset_in_page(addr
))
2086 error
= security_mmap_addr(addr
);
2087 return error
? error
: addr
;
2090 EXPORT_SYMBOL(get_unmapped_area
);
2092 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2093 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
2095 struct rb_node
*rb_node
;
2096 struct vm_area_struct
*vma
;
2098 /* Check the cache first. */
2099 vma
= vmacache_find(mm
, addr
);
2103 rb_node
= mm
->mm_rb
.rb_node
;
2106 struct vm_area_struct
*tmp
;
2108 tmp
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2110 if (tmp
->vm_end
> addr
) {
2112 if (tmp
->vm_start
<= addr
)
2114 rb_node
= rb_node
->rb_left
;
2116 rb_node
= rb_node
->rb_right
;
2120 vmacache_update(addr
, vma
);
2124 EXPORT_SYMBOL(find_vma
);
2127 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2129 struct vm_area_struct
*
2130 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
2131 struct vm_area_struct
**pprev
)
2133 struct vm_area_struct
*vma
;
2135 vma
= find_vma(mm
, addr
);
2137 *pprev
= vma
->vm_prev
;
2139 struct rb_node
*rb_node
= mm
->mm_rb
.rb_node
;
2142 *pprev
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2143 rb_node
= rb_node
->rb_right
;
2150 * Verify that the stack growth is acceptable and
2151 * update accounting. This is shared with both the
2152 * grow-up and grow-down cases.
2154 static int acct_stack_growth(struct vm_area_struct
*vma
, unsigned long size
, unsigned long grow
,
2157 struct mm_struct
*mm
= vma
->vm_mm
;
2158 struct rlimit
*rlim
= current
->signal
->rlim
;
2159 unsigned long new_start
, actual_size
;
2161 /* address space limit tests */
2162 if (!may_expand_vm(mm
, vma
->vm_flags
, grow
))
2165 /* Stack limit test */
2167 if (size
&& (vma
->vm_flags
& (VM_GROWSUP
| VM_GROWSDOWN
)))
2169 if (actual_size
> READ_ONCE(rlim
[RLIMIT_STACK
].rlim_cur
))
2172 /* mlock limit tests */
2173 if (vma
->vm_flags
& VM_LOCKED
) {
2174 unsigned long locked
;
2175 unsigned long limit
;
2176 locked
= mm
->locked_vm
+ grow
;
2177 limit
= READ_ONCE(rlim
[RLIMIT_MEMLOCK
].rlim_cur
);
2178 limit
>>= PAGE_SHIFT
;
2179 if (locked
> limit
&& !capable(CAP_IPC_LOCK
))
2183 /* Check to ensure the stack will not grow into a hugetlb-only region */
2184 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
2186 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
2190 * Overcommit.. This must be the final test, as it will
2191 * update security statistics.
2193 if (security_vm_enough_memory_mm(mm
, grow
))
2199 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2201 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2202 * vma is the last one with address > vma->vm_end. Have to extend vma.
2204 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
, unsigned long gap
)
2206 struct mm_struct
*mm
= vma
->vm_mm
;
2209 if (!(vma
->vm_flags
& VM_GROWSUP
))
2212 /* We must make sure the anon_vma is allocated. */
2213 if (unlikely(anon_vma_prepare(vma
)))
2217 * vma->vm_start/vm_end cannot change under us because the caller
2218 * is required to hold the mmap_sem in read mode. We need the
2219 * anon_vma lock to serialize against concurrent expand_stacks.
2221 anon_vma_lock_write(vma
->anon_vma
);
2223 /* Somebody else might have raced and expanded it already */
2224 if (address
> vma
->vm_end
) {
2225 unsigned long size
, grow
;
2227 size
= address
- vma
->vm_start
;
2228 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
2231 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
2232 error
= acct_stack_growth(vma
, size
, grow
, gap
);
2235 * vma_gap_update() doesn't support concurrent
2236 * updates, but we only hold a shared mmap_sem
2237 * lock here, so we need to protect against
2238 * concurrent vma expansions.
2239 * anon_vma_lock_write() doesn't help here, as
2240 * we don't guarantee that all growable vmas
2241 * in a mm share the same root anon vma.
2242 * So, we reuse mm->page_table_lock to guard
2243 * against concurrent vma expansions.
2245 spin_lock(&mm
->page_table_lock
);
2246 if (vma
->vm_flags
& VM_LOCKED
)
2247 mm
->locked_vm
+= grow
;
2248 vm_stat_account(mm
, vma
->vm_flags
, grow
);
2249 anon_vma_interval_tree_pre_update_vma(vma
);
2250 vma
->vm_end
= address
;
2251 anon_vma_interval_tree_post_update_vma(vma
);
2253 vma_gap_update(vma
->vm_next
);
2255 mm
->highest_vm_end
= address
;
2256 spin_unlock(&mm
->page_table_lock
);
2258 perf_event_mmap(vma
);
2262 anon_vma_unlock_write(vma
->anon_vma
);
2263 khugepaged_enter_vma_merge(vma
, vma
->vm_flags
);
2267 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2270 * vma is the first one with address < vma->vm_start. Have to extend vma.
2272 int expand_downwards(struct vm_area_struct
*vma
,
2273 unsigned long address
, unsigned long gap
)
2275 struct mm_struct
*mm
= vma
->vm_mm
;
2278 address
&= PAGE_MASK
;
2279 error
= security_mmap_addr(address
);
2283 /* We must make sure the anon_vma is allocated. */
2284 if (unlikely(anon_vma_prepare(vma
)))
2288 * vma->vm_start/vm_end cannot change under us because the caller
2289 * is required to hold the mmap_sem in read mode. We need the
2290 * anon_vma lock to serialize against concurrent expand_stacks.
2292 anon_vma_lock_write(vma
->anon_vma
);
2294 /* Somebody else might have raced and expanded it already */
2295 if (address
< vma
->vm_start
) {
2296 unsigned long size
, grow
;
2298 size
= vma
->vm_end
- address
;
2299 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
2302 if (grow
<= vma
->vm_pgoff
) {
2303 error
= acct_stack_growth(vma
, size
, grow
, gap
);
2306 * vma_gap_update() doesn't support concurrent
2307 * updates, but we only hold a shared mmap_sem
2308 * lock here, so we need to protect against
2309 * concurrent vma expansions.
2310 * anon_vma_lock_write() doesn't help here, as
2311 * we don't guarantee that all growable vmas
2312 * in a mm share the same root anon vma.
2313 * So, we reuse mm->page_table_lock to guard
2314 * against concurrent vma expansions.
2316 spin_lock(&mm
->page_table_lock
);
2317 if (vma
->vm_flags
& VM_LOCKED
)
2318 mm
->locked_vm
+= grow
;
2319 vm_stat_account(mm
, vma
->vm_flags
, grow
);
2320 anon_vma_interval_tree_pre_update_vma(vma
);
2321 vma
->vm_start
= address
;
2322 vma
->vm_pgoff
-= grow
;
2323 anon_vma_interval_tree_post_update_vma(vma
);
2324 vma_gap_update(vma
);
2325 spin_unlock(&mm
->page_table_lock
);
2327 perf_event_mmap(vma
);
2331 anon_vma_unlock_write(vma
->anon_vma
);
2332 khugepaged_enter_vma_merge(vma
, vma
->vm_flags
);
2337 /* enforced gap between the expanding stack and other mappings. */
2338 unsigned long stack_guard_gap
= 256UL<<PAGE_SHIFT
;
2341 * Note how expand_stack() refuses to expand the stack all the way to
2342 * abut the next virtual mapping, *unless* that mapping itself is also
2343 * a stack mapping. We want to leave room for a guard area, after all
2344 * (the guard page itself is not added here, that is done by the
2345 * actual page faulting logic)
2347 #ifdef CONFIG_STACK_GROWSUP
2348 unsigned long expandable_stack_area(struct vm_area_struct
*vma
,
2349 unsigned long address
, unsigned long *gap
)
2351 struct vm_area_struct
*next
= vma
->vm_next
;
2352 unsigned long guard_gap
= stack_guard_gap
;
2353 unsigned long guard_addr
;
2355 address
= ALIGN(address
, PAGE_SIZE
);;
2359 if (next
->vm_flags
& VM_GROWSUP
) {
2360 guard_gap
= min(guard_gap
, next
->vm_start
- address
);
2364 if (next
->vm_start
- address
< guard_gap
)
2367 if (TASK_SIZE
- address
< guard_gap
)
2368 guard_gap
= TASK_SIZE
- address
;
2369 guard_addr
= address
+ guard_gap
;
2375 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2379 address
= expandable_stack_area(vma
, address
, &gap
);
2380 if (IS_ERR_VALUE(address
))
2382 return expand_upwards(vma
, address
, gap
);
2385 int stack_guard_area(struct vm_area_struct
*vma
, unsigned long address
)
2387 struct vm_area_struct
*next
;
2389 if (!(vma
->vm_flags
& VM_GROWSUP
))
2393 * strictly speaking there is a guard gap between disjoint stacks
2394 * but the gap is not canonical (it might be smaller) and it is
2395 * reasonably safe to assume that we can ignore that gap for stack
2396 * POPULATE or /proc/<pid>[s]maps purposes
2398 next
= vma
->vm_next
;
2399 if (next
&& next
->vm_flags
& VM_GROWSUP
)
2402 return vma
->vm_end
- address
<= stack_guard_gap
;
2405 struct vm_area_struct
*
2406 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2408 struct vm_area_struct
*vma
, *prev
;
2411 vma
= find_vma_prev(mm
, addr
, &prev
);
2412 if (vma
&& (vma
->vm_start
<= addr
))
2414 if (!prev
|| expand_stack(prev
, addr
))
2416 if (prev
->vm_flags
& VM_LOCKED
)
2417 populate_vma_page_range(prev
, addr
, prev
->vm_end
, NULL
);
2421 unsigned long expandable_stack_area(struct vm_area_struct
*vma
,
2422 unsigned long address
, unsigned long *gap
)
2424 struct vm_area_struct
*prev
= vma
->vm_prev
;
2425 unsigned long guard_gap
= stack_guard_gap
;
2426 unsigned long guard_addr
;
2428 address
&= PAGE_MASK
;
2433 * Is there a mapping abutting this one below?
2435 * That's only ok if it's the same stack mapping
2436 * that has gotten split or there is sufficient gap
2439 if (prev
->vm_flags
& VM_GROWSDOWN
) {
2440 guard_gap
= min(guard_gap
, address
- prev
->vm_end
);
2444 if (address
- prev
->vm_end
< guard_gap
)
2448 /* make sure we won't underflow */
2449 if (address
< mmap_min_addr
)
2451 if (address
- mmap_min_addr
< guard_gap
)
2452 guard_gap
= address
- mmap_min_addr
;
2454 guard_addr
= address
- guard_gap
;
2460 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2464 address
= expandable_stack_area(vma
, address
, &gap
);
2465 if (IS_ERR_VALUE(address
))
2467 return expand_downwards(vma
, address
, gap
);
2470 int stack_guard_area(struct vm_area_struct
*vma
, unsigned long address
)
2472 struct vm_area_struct
*prev
;
2474 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2478 * strictly speaking there is a guard gap between disjoint stacks
2479 * but the gap is not canonical (it might be smaller) and it is
2480 * reasonably safe to assume that we can ignore that gap for stack
2481 * POPULATE or /proc/<pid>[s]maps purposes
2483 prev
= vma
->vm_prev
;
2484 if (prev
&& prev
->vm_flags
& VM_GROWSDOWN
)
2487 return address
- vma
->vm_start
< stack_guard_gap
;
2490 struct vm_area_struct
*
2491 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2493 struct vm_area_struct
*vma
;
2494 unsigned long start
;
2497 vma
= find_vma(mm
, addr
);
2500 if (vma
->vm_start
<= addr
)
2502 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2504 start
= vma
->vm_start
;
2505 if (expand_stack(vma
, addr
))
2507 if (vma
->vm_flags
& VM_LOCKED
)
2508 populate_vma_page_range(vma
, addr
, start
, NULL
);
2513 EXPORT_SYMBOL_GPL(find_extend_vma
);
2516 * Ok - we have the memory areas we should free on the vma list,
2517 * so release them, and do the vma updates.
2519 * Called with the mm semaphore held.
2521 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2523 unsigned long nr_accounted
= 0;
2525 /* Update high watermark before we lower total_vm */
2526 update_hiwater_vm(mm
);
2528 long nrpages
= vma_pages(vma
);
2530 if (vma
->vm_flags
& VM_ACCOUNT
)
2531 nr_accounted
+= nrpages
;
2532 vm_stat_account(mm
, vma
->vm_flags
, -nrpages
);
2533 vma
= remove_vma(vma
);
2535 vm_unacct_memory(nr_accounted
);
2540 * Get rid of page table information in the indicated region.
2542 * Called with the mm semaphore held.
2544 static void unmap_region(struct mm_struct
*mm
,
2545 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
2546 unsigned long start
, unsigned long end
)
2548 struct vm_area_struct
*next
= prev
? prev
->vm_next
: mm
->mmap
;
2549 struct mmu_gather tlb
;
2552 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2553 update_hiwater_rss(mm
);
2554 unmap_vmas(&tlb
, vma
, start
, end
);
2555 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
2556 next
? next
->vm_start
: USER_PGTABLES_CEILING
);
2557 tlb_finish_mmu(&tlb
, start
, end
);
2561 * Create a list of vma's touched by the unmap, removing them from the mm's
2562 * vma list as we go..
2565 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2566 struct vm_area_struct
*prev
, unsigned long end
)
2568 struct vm_area_struct
**insertion_point
;
2569 struct vm_area_struct
*tail_vma
= NULL
;
2571 insertion_point
= (prev
? &prev
->vm_next
: &mm
->mmap
);
2572 vma
->vm_prev
= NULL
;
2574 vma_rb_erase(vma
, &mm
->mm_rb
);
2578 } while (vma
&& vma
->vm_start
< end
);
2579 *insertion_point
= vma
;
2581 vma
->vm_prev
= prev
;
2582 vma_gap_update(vma
);
2584 mm
->highest_vm_end
= prev
? prev
->vm_end
: 0;
2585 tail_vma
->vm_next
= NULL
;
2587 /* Kill the cache */
2588 vmacache_invalidate(mm
);
2592 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2593 * munmap path where it doesn't make sense to fail.
2595 static int __split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2596 unsigned long addr
, int new_below
)
2598 struct vm_area_struct
*new;
2601 if (is_vm_hugetlb_page(vma
) && (addr
&
2602 ~(huge_page_mask(hstate_vma(vma
)))))
2605 new = kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2609 /* most fields are the same, copy all, and then fixup */
2612 INIT_LIST_HEAD(&new->anon_vma_chain
);
2617 new->vm_start
= addr
;
2618 new->vm_pgoff
+= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
);
2621 err
= vma_dup_policy(vma
, new);
2625 err
= anon_vma_clone(new, vma
);
2632 if (new->vm_ops
&& new->vm_ops
->open
)
2633 new->vm_ops
->open(new);
2636 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2637 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2639 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2645 /* Clean everything up if vma_adjust failed. */
2646 if (new->vm_ops
&& new->vm_ops
->close
)
2647 new->vm_ops
->close(new);
2650 unlink_anon_vmas(new);
2652 mpol_put(vma_policy(new));
2654 kmem_cache_free(vm_area_cachep
, new);
2659 * Split a vma into two pieces at address 'addr', a new vma is allocated
2660 * either for the first part or the tail.
2662 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2663 unsigned long addr
, int new_below
)
2665 if (mm
->map_count
>= sysctl_max_map_count
)
2668 return __split_vma(mm
, vma
, addr
, new_below
);
2671 /* Munmap is split into 2 main parts -- this part which finds
2672 * what needs doing, and the areas themselves, which do the
2673 * work. This now handles partial unmappings.
2674 * Jeremy Fitzhardinge <jeremy@goop.org>
2676 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
)
2679 struct vm_area_struct
*vma
, *prev
, *last
;
2681 if ((offset_in_page(start
)) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2684 len
= PAGE_ALIGN(len
);
2688 /* Find the first overlapping VMA */
2689 vma
= find_vma(mm
, start
);
2692 prev
= vma
->vm_prev
;
2693 /* we have start < vma->vm_end */
2695 /* if it doesn't overlap, we have nothing.. */
2697 if (vma
->vm_start
>= end
)
2701 * If we need to split any vma, do it now to save pain later.
2703 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2704 * unmapped vm_area_struct will remain in use: so lower split_vma
2705 * places tmp vma above, and higher split_vma places tmp vma below.
2707 if (start
> vma
->vm_start
) {
2711 * Make sure that map_count on return from munmap() will
2712 * not exceed its limit; but let map_count go just above
2713 * its limit temporarily, to help free resources as expected.
2715 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2718 error
= __split_vma(mm
, vma
, start
, 0);
2724 /* Does it split the last one? */
2725 last
= find_vma(mm
, end
);
2726 if (last
&& end
> last
->vm_start
) {
2727 int error
= __split_vma(mm
, last
, end
, 1);
2731 vma
= prev
? prev
->vm_next
: mm
->mmap
;
2734 * unlock any mlock()ed ranges before detaching vmas
2736 if (mm
->locked_vm
) {
2737 struct vm_area_struct
*tmp
= vma
;
2738 while (tmp
&& tmp
->vm_start
< end
) {
2739 if (tmp
->vm_flags
& VM_LOCKED
) {
2740 mm
->locked_vm
-= vma_pages(tmp
);
2741 munlock_vma_pages_all(tmp
);
2748 * Remove the vma's, and unmap the actual pages
2750 detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
);
2751 unmap_region(mm
, vma
, prev
, start
, end
);
2753 arch_unmap(mm
, vma
, start
, end
);
2755 /* Fix up all other VM information */
2756 remove_vma_list(mm
, vma
);
2761 int vm_munmap(unsigned long start
, size_t len
)
2764 struct mm_struct
*mm
= current
->mm
;
2766 if (down_write_killable(&mm
->mmap_sem
))
2769 ret
= do_munmap(mm
, start
, len
);
2770 up_write(&mm
->mmap_sem
);
2773 EXPORT_SYMBOL(vm_munmap
);
2775 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2778 struct mm_struct
*mm
= current
->mm
;
2780 profile_munmap(addr
);
2781 if (down_write_killable(&mm
->mmap_sem
))
2783 ret
= do_munmap(mm
, addr
, len
);
2784 up_write(&mm
->mmap_sem
);
2790 * Emulation of deprecated remap_file_pages() syscall.
2792 SYSCALL_DEFINE5(remap_file_pages
, unsigned long, start
, unsigned long, size
,
2793 unsigned long, prot
, unsigned long, pgoff
, unsigned long, flags
)
2796 struct mm_struct
*mm
= current
->mm
;
2797 struct vm_area_struct
*vma
;
2798 unsigned long populate
= 0;
2799 unsigned long ret
= -EINVAL
;
2800 struct file
*file
, *prfile
;
2802 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2803 current
->comm
, current
->pid
);
2807 start
= start
& PAGE_MASK
;
2808 size
= size
& PAGE_MASK
;
2810 if (start
+ size
<= start
)
2813 /* Does pgoff wrap? */
2814 if (pgoff
+ (size
>> PAGE_SHIFT
) < pgoff
)
2817 if (down_write_killable(&mm
->mmap_sem
))
2820 vma
= find_vma(mm
, start
);
2822 if (!vma
|| !(vma
->vm_flags
& VM_SHARED
))
2825 if (start
< vma
->vm_start
)
2828 if (start
+ size
> vma
->vm_end
) {
2829 struct vm_area_struct
*next
;
2831 for (next
= vma
->vm_next
; next
; next
= next
->vm_next
) {
2832 /* hole between vmas ? */
2833 if (next
->vm_start
!= next
->vm_prev
->vm_end
)
2836 if (next
->vm_file
!= vma
->vm_file
)
2839 if (next
->vm_flags
!= vma
->vm_flags
)
2842 if (start
+ size
<= next
->vm_end
)
2850 prot
|= vma
->vm_flags
& VM_READ
? PROT_READ
: 0;
2851 prot
|= vma
->vm_flags
& VM_WRITE
? PROT_WRITE
: 0;
2852 prot
|= vma
->vm_flags
& VM_EXEC
? PROT_EXEC
: 0;
2854 flags
&= MAP_NONBLOCK
;
2855 flags
|= MAP_SHARED
| MAP_FIXED
| MAP_POPULATE
;
2856 if (vma
->vm_flags
& VM_LOCKED
) {
2857 struct vm_area_struct
*tmp
;
2858 flags
|= MAP_LOCKED
;
2860 /* drop PG_Mlocked flag for over-mapped range */
2861 for (tmp
= vma
; tmp
->vm_start
>= start
+ size
;
2862 tmp
= tmp
->vm_next
) {
2864 * Split pmd and munlock page on the border
2867 vma_adjust_trans_huge(tmp
, start
, start
+ size
, 0);
2869 munlock_vma_pages_range(tmp
,
2870 max(tmp
->vm_start
, start
),
2871 min(tmp
->vm_end
, start
+ size
));
2876 file
= vma
->vm_file
;
2877 prfile
= vma
->vm_prfile
;
2878 ret
= do_mmap_pgoff(vma
->vm_file
, start
, size
,
2879 prot
, flags
, pgoff
, &populate
);
2880 if (!IS_ERR_VALUE(ret
) && file
&& prfile
) {
2881 struct vm_area_struct
*new_vma
;
2883 new_vma
= find_vma(mm
, ret
);
2884 if (!new_vma
->vm_prfile
)
2885 new_vma
->vm_prfile
= prfile
;
2890 * two fput()s instead of vma_fput(vma),
2891 * coz vma may not be available anymore.
2897 up_write(&mm
->mmap_sem
);
2899 mm_populate(ret
, populate
);
2900 if (!IS_ERR_VALUE(ret
))
2905 static inline void verify_mm_writelocked(struct mm_struct
*mm
)
2907 #ifdef CONFIG_DEBUG_VM
2908 if (unlikely(down_read_trylock(&mm
->mmap_sem
))) {
2910 up_read(&mm
->mmap_sem
);
2916 * this is really a simplified "do_mmap". it only handles
2917 * anonymous maps. eventually we may be able to do some
2918 * brk-specific accounting here.
2920 static int do_brk(unsigned long addr
, unsigned long request
)
2922 struct mm_struct
*mm
= current
->mm
;
2923 struct vm_area_struct
*vma
, *prev
;
2924 unsigned long flags
, len
;
2925 struct rb_node
**rb_link
, *rb_parent
;
2926 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2929 len
= PAGE_ALIGN(request
);
2935 flags
= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2937 error
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2938 if (offset_in_page(error
))
2941 error
= mlock_future_check(mm
, mm
->def_flags
, len
);
2946 * mm->mmap_sem is required to protect against another thread
2947 * changing the mappings in case we sleep.
2949 verify_mm_writelocked(mm
);
2952 * Clear old maps. this also does some error checking for us
2954 while (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
,
2956 if (do_munmap(mm
, addr
, len
))
2960 /* Check against address space limits *after* clearing old maps... */
2961 if (!may_expand_vm(mm
, flags
, len
>> PAGE_SHIFT
))
2964 if (mm
->map_count
> sysctl_max_map_count
)
2967 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
2970 /* Can we just expand an old private anonymous mapping? */
2971 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
2972 NULL
, NULL
, pgoff
, NULL
, NULL_VM_UFFD_CTX
);
2977 * create a vma struct for an anonymous mapping
2979 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2981 vm_unacct_memory(len
>> PAGE_SHIFT
);
2985 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2987 vma
->vm_start
= addr
;
2988 vma
->vm_end
= addr
+ len
;
2989 vma
->vm_pgoff
= pgoff
;
2990 vma
->vm_flags
= flags
;
2991 vma
->vm_page_prot
= vm_get_page_prot(flags
);
2992 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2994 perf_event_mmap(vma
);
2995 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2996 mm
->data_vm
+= len
>> PAGE_SHIFT
;
2997 if (flags
& VM_LOCKED
)
2998 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
2999 vma
->vm_flags
|= VM_SOFTDIRTY
;
3003 int vm_brk(unsigned long addr
, unsigned long len
)
3005 struct mm_struct
*mm
= current
->mm
;
3009 if (down_write_killable(&mm
->mmap_sem
))
3012 ret
= do_brk(addr
, len
);
3013 populate
= ((mm
->def_flags
& VM_LOCKED
) != 0);
3014 up_write(&mm
->mmap_sem
);
3015 if (populate
&& !ret
)
3016 mm_populate(addr
, len
);
3019 EXPORT_SYMBOL(vm_brk
);
3021 /* Release all mmaps. */
3022 void exit_mmap(struct mm_struct
*mm
)
3024 struct mmu_gather tlb
;
3025 struct vm_area_struct
*vma
;
3026 unsigned long nr_accounted
= 0;
3028 /* mm's last user has gone, and its about to be pulled down */
3029 mmu_notifier_release(mm
);
3031 if (mm
->locked_vm
) {
3034 if (vma
->vm_flags
& VM_LOCKED
)
3035 munlock_vma_pages_all(vma
);
3043 if (!vma
) /* Can happen if dup_mmap() received an OOM */
3048 tlb_gather_mmu(&tlb
, mm
, 0, -1);
3049 /* update_hiwater_rss(mm) here? but nobody should be looking */
3050 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3051 unmap_vmas(&tlb
, vma
, 0, -1);
3053 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, USER_PGTABLES_CEILING
);
3054 tlb_finish_mmu(&tlb
, 0, -1);
3057 * Walk the list again, actually closing and freeing it,
3058 * with preemption enabled, without holding any MM locks.
3061 if (vma
->vm_flags
& VM_ACCOUNT
)
3062 nr_accounted
+= vma_pages(vma
);
3063 vma
= remove_vma(vma
);
3065 vm_unacct_memory(nr_accounted
);
3068 /* Insert vm structure into process list sorted by address
3069 * and into the inode's i_mmap tree. If vm_file is non-NULL
3070 * then i_mmap_rwsem is taken here.
3072 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
3074 struct vm_area_struct
*prev
;
3075 struct rb_node
**rb_link
, *rb_parent
;
3077 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
3078 &prev
, &rb_link
, &rb_parent
))
3080 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
3081 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
3085 * The vm_pgoff of a purely anonymous vma should be irrelevant
3086 * until its first write fault, when page's anon_vma and index
3087 * are set. But now set the vm_pgoff it will almost certainly
3088 * end up with (unless mremap moves it elsewhere before that
3089 * first wfault), so /proc/pid/maps tells a consistent story.
3091 * By setting it to reflect the virtual start address of the
3092 * vma, merges and splits can happen in a seamless way, just
3093 * using the existing file pgoff checks and manipulations.
3094 * Similarly in do_mmap_pgoff and in do_brk.
3096 if (vma_is_anonymous(vma
)) {
3097 BUG_ON(vma
->anon_vma
);
3098 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
3101 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
3106 * Copy the vma structure to a new location in the same mm,
3107 * prior to moving page table entries, to effect an mremap move.
3109 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
3110 unsigned long addr
, unsigned long len
, pgoff_t pgoff
,
3111 bool *need_rmap_locks
)
3113 struct vm_area_struct
*vma
= *vmap
;
3114 unsigned long vma_start
= vma
->vm_start
;
3115 struct mm_struct
*mm
= vma
->vm_mm
;
3116 struct vm_area_struct
*new_vma
, *prev
;
3117 struct rb_node
**rb_link
, *rb_parent
;
3118 bool faulted_in_anon_vma
= true;
3121 * If anonymous vma has not yet been faulted, update new pgoff
3122 * to match new location, to increase its chance of merging.
3124 if (unlikely(vma_is_anonymous(vma
) && !vma
->anon_vma
)) {
3125 pgoff
= addr
>> PAGE_SHIFT
;
3126 faulted_in_anon_vma
= false;
3129 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
3130 return NULL
; /* should never get here */
3131 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
3132 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
),
3133 vma
->vm_userfaultfd_ctx
);
3136 * Source vma may have been merged into new_vma
3138 if (unlikely(vma_start
>= new_vma
->vm_start
&&
3139 vma_start
< new_vma
->vm_end
)) {
3141 * The only way we can get a vma_merge with
3142 * self during an mremap is if the vma hasn't
3143 * been faulted in yet and we were allowed to
3144 * reset the dst vma->vm_pgoff to the
3145 * destination address of the mremap to allow
3146 * the merge to happen. mremap must change the
3147 * vm_pgoff linearity between src and dst vmas
3148 * (in turn preventing a vma_merge) to be
3149 * safe. It is only safe to keep the vm_pgoff
3150 * linear if there are no pages mapped yet.
3152 VM_BUG_ON_VMA(faulted_in_anon_vma
, new_vma
);
3153 *vmap
= vma
= new_vma
;
3155 *need_rmap_locks
= (new_vma
->vm_pgoff
<= vma
->vm_pgoff
);
3157 new_vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
3161 new_vma
->vm_start
= addr
;
3162 new_vma
->vm_end
= addr
+ len
;
3163 new_vma
->vm_pgoff
= pgoff
;
3164 if (vma_dup_policy(vma
, new_vma
))
3166 INIT_LIST_HEAD(&new_vma
->anon_vma_chain
);
3167 if (anon_vma_clone(new_vma
, vma
))
3168 goto out_free_mempol
;
3169 if (new_vma
->vm_file
)
3170 vma_get_file(new_vma
);
3171 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
3172 new_vma
->vm_ops
->open(new_vma
);
3173 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
3174 *need_rmap_locks
= false;
3179 mpol_put(vma_policy(new_vma
));
3181 kmem_cache_free(vm_area_cachep
, new_vma
);
3187 * Return true if the calling process may expand its vm space by the passed
3190 bool may_expand_vm(struct mm_struct
*mm
, vm_flags_t flags
, unsigned long npages
)
3192 if (mm
->total_vm
+ npages
> rlimit(RLIMIT_AS
) >> PAGE_SHIFT
)
3195 if (is_data_mapping(flags
) &&
3196 mm
->data_vm
+ npages
> rlimit(RLIMIT_DATA
) >> PAGE_SHIFT
) {
3197 /* Workaround for Valgrind */
3198 if (rlimit(RLIMIT_DATA
) == 0 &&
3199 mm
->data_vm
+ npages
<= rlimit_max(RLIMIT_DATA
) >> PAGE_SHIFT
)
3201 if (!ignore_rlimit_data
) {
3202 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits or use boot option ignore_rlimit_data.\n",
3203 current
->comm
, current
->pid
,
3204 (mm
->data_vm
+ npages
) << PAGE_SHIFT
,
3205 rlimit(RLIMIT_DATA
));
3213 void vm_stat_account(struct mm_struct
*mm
, vm_flags_t flags
, long npages
)
3215 mm
->total_vm
+= npages
;
3217 if (is_exec_mapping(flags
))
3218 mm
->exec_vm
+= npages
;
3219 else if (is_stack_mapping(flags
))
3220 mm
->stack_vm
+= npages
;
3221 else if (is_data_mapping(flags
))
3222 mm
->data_vm
+= npages
;
3225 static int special_mapping_fault(struct vm_area_struct
*vma
,
3226 struct vm_fault
*vmf
);
3229 * Having a close hook prevents vma merging regardless of flags.
3231 static void special_mapping_close(struct vm_area_struct
*vma
)
3235 static const char *special_mapping_name(struct vm_area_struct
*vma
)
3237 return ((struct vm_special_mapping
*)vma
->vm_private_data
)->name
;
3240 static int special_mapping_mremap(struct vm_area_struct
*new_vma
)
3242 struct vm_special_mapping
*sm
= new_vma
->vm_private_data
;
3245 return sm
->mremap(sm
, new_vma
);
3249 static const struct vm_operations_struct special_mapping_vmops
= {
3250 .close
= special_mapping_close
,
3251 .fault
= special_mapping_fault
,
3252 .mremap
= special_mapping_mremap
,
3253 .name
= special_mapping_name
,
3256 static const struct vm_operations_struct legacy_special_mapping_vmops
= {
3257 .close
= special_mapping_close
,
3258 .fault
= special_mapping_fault
,
3261 static int special_mapping_fault(struct vm_area_struct
*vma
,
3262 struct vm_fault
*vmf
)
3265 struct page
**pages
;
3267 if (vma
->vm_ops
== &legacy_special_mapping_vmops
) {
3268 pages
= vma
->vm_private_data
;
3270 struct vm_special_mapping
*sm
= vma
->vm_private_data
;
3273 return sm
->fault(sm
, vma
, vmf
);
3278 for (pgoff
= vmf
->pgoff
; pgoff
&& *pages
; ++pages
)
3282 struct page
*page
= *pages
;
3288 return VM_FAULT_SIGBUS
;
3291 static struct vm_area_struct
*__install_special_mapping(
3292 struct mm_struct
*mm
,
3293 unsigned long addr
, unsigned long len
,
3294 unsigned long vm_flags
, void *priv
,
3295 const struct vm_operations_struct
*ops
)
3298 struct vm_area_struct
*vma
;
3300 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
3301 if (unlikely(vma
== NULL
))
3302 return ERR_PTR(-ENOMEM
);
3304 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
3306 vma
->vm_start
= addr
;
3307 vma
->vm_end
= addr
+ len
;
3309 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
| VM_SOFTDIRTY
;
3310 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
3313 vma
->vm_private_data
= priv
;
3315 ret
= insert_vm_struct(mm
, vma
);
3319 vm_stat_account(mm
, vma
->vm_flags
, len
>> PAGE_SHIFT
);
3321 perf_event_mmap(vma
);
3326 kmem_cache_free(vm_area_cachep
, vma
);
3327 return ERR_PTR(ret
);
3330 bool vma_is_special_mapping(const struct vm_area_struct
*vma
,
3331 const struct vm_special_mapping
*sm
)
3333 return vma
->vm_private_data
== sm
&&
3334 (vma
->vm_ops
== &special_mapping_vmops
||
3335 vma
->vm_ops
== &legacy_special_mapping_vmops
);
3339 * Called with mm->mmap_sem held for writing.
3340 * Insert a new vma covering the given region, with the given flags.
3341 * Its pages are supplied by the given array of struct page *.
3342 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3343 * The region past the last page supplied will always produce SIGBUS.
3344 * The array pointer and the pages it points to are assumed to stay alive
3345 * for as long as this mapping might exist.
3347 struct vm_area_struct
*_install_special_mapping(
3348 struct mm_struct
*mm
,
3349 unsigned long addr
, unsigned long len
,
3350 unsigned long vm_flags
, const struct vm_special_mapping
*spec
)
3352 return __install_special_mapping(mm
, addr
, len
, vm_flags
, (void *)spec
,
3353 &special_mapping_vmops
);
3356 int install_special_mapping(struct mm_struct
*mm
,
3357 unsigned long addr
, unsigned long len
,
3358 unsigned long vm_flags
, struct page
**pages
)
3360 struct vm_area_struct
*vma
= __install_special_mapping(
3361 mm
, addr
, len
, vm_flags
, (void *)pages
,
3362 &legacy_special_mapping_vmops
);
3364 return PTR_ERR_OR_ZERO(vma
);
3367 static DEFINE_MUTEX(mm_all_locks_mutex
);
3369 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
3371 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3373 * The LSB of head.next can't change from under us
3374 * because we hold the mm_all_locks_mutex.
3376 down_write_nest_lock(&anon_vma
->root
->rwsem
, &mm
->mmap_sem
);
3378 * We can safely modify head.next after taking the
3379 * anon_vma->root->rwsem. If some other vma in this mm shares
3380 * the same anon_vma we won't take it again.
3382 * No need of atomic instructions here, head.next
3383 * can't change from under us thanks to the
3384 * anon_vma->root->rwsem.
3386 if (__test_and_set_bit(0, (unsigned long *)
3387 &anon_vma
->root
->rb_root
.rb_node
))
3392 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
3394 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3396 * AS_MM_ALL_LOCKS can't change from under us because
3397 * we hold the mm_all_locks_mutex.
3399 * Operations on ->flags have to be atomic because
3400 * even if AS_MM_ALL_LOCKS is stable thanks to the
3401 * mm_all_locks_mutex, there may be other cpus
3402 * changing other bitflags in parallel to us.
3404 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
3406 down_write_nest_lock(&mapping
->i_mmap_rwsem
, &mm
->mmap_sem
);
3411 * This operation locks against the VM for all pte/vma/mm related
3412 * operations that could ever happen on a certain mm. This includes
3413 * vmtruncate, try_to_unmap, and all page faults.
3415 * The caller must take the mmap_sem in write mode before calling
3416 * mm_take_all_locks(). The caller isn't allowed to release the
3417 * mmap_sem until mm_drop_all_locks() returns.
3419 * mmap_sem in write mode is required in order to block all operations
3420 * that could modify pagetables and free pages without need of
3421 * altering the vma layout. It's also needed in write mode to avoid new
3422 * anon_vmas to be associated with existing vmas.
3424 * A single task can't take more than one mm_take_all_locks() in a row
3425 * or it would deadlock.
3427 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3428 * mapping->flags avoid to take the same lock twice, if more than one
3429 * vma in this mm is backed by the same anon_vma or address_space.
3431 * We take locks in following order, accordingly to comment at beginning
3433 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3435 * - all i_mmap_rwsem locks;
3436 * - all anon_vma->rwseml
3438 * We can take all locks within these types randomly because the VM code
3439 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3440 * mm_all_locks_mutex.
3442 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3443 * that may have to take thousand of locks.
3445 * mm_take_all_locks() can fail if it's interrupted by signals.
3447 int mm_take_all_locks(struct mm_struct
*mm
)
3449 struct vm_area_struct
*vma
;
3450 struct anon_vma_chain
*avc
;
3452 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3454 mutex_lock(&mm_all_locks_mutex
);
3456 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3457 if (signal_pending(current
))
3459 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
3460 is_vm_hugetlb_page(vma
))
3461 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3464 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3465 if (signal_pending(current
))
3467 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
3468 !is_vm_hugetlb_page(vma
))
3469 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3472 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3473 if (signal_pending(current
))
3476 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3477 vm_lock_anon_vma(mm
, avc
->anon_vma
);
3483 mm_drop_all_locks(mm
);
3487 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
3489 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3491 * The LSB of head.next can't change to 0 from under
3492 * us because we hold the mm_all_locks_mutex.
3494 * We must however clear the bitflag before unlocking
3495 * the vma so the users using the anon_vma->rb_root will
3496 * never see our bitflag.
3498 * No need of atomic instructions here, head.next
3499 * can't change from under us until we release the
3500 * anon_vma->root->rwsem.
3502 if (!__test_and_clear_bit(0, (unsigned long *)
3503 &anon_vma
->root
->rb_root
.rb_node
))
3505 anon_vma_unlock_write(anon_vma
);
3509 static void vm_unlock_mapping(struct address_space
*mapping
)
3511 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3513 * AS_MM_ALL_LOCKS can't change to 0 from under us
3514 * because we hold the mm_all_locks_mutex.
3516 i_mmap_unlock_write(mapping
);
3517 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
3524 * The mmap_sem cannot be released by the caller until
3525 * mm_drop_all_locks() returns.
3527 void mm_drop_all_locks(struct mm_struct
*mm
)
3529 struct vm_area_struct
*vma
;
3530 struct anon_vma_chain
*avc
;
3532 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3533 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
3535 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3537 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3538 vm_unlock_anon_vma(avc
->anon_vma
);
3539 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3540 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
3543 mutex_unlock(&mm_all_locks_mutex
);
3547 * initialise the VMA slab
3549 void __init
mmap_init(void)
3553 ret
= percpu_counter_init(&vm_committed_as
, 0, GFP_KERNEL
);
3558 * Initialise sysctl_user_reserve_kbytes.
3560 * This is intended to prevent a user from starting a single memory hogging
3561 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3564 * The default value is min(3% of free memory, 128MB)
3565 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3567 static int init_user_reserve(void)
3569 unsigned long free_kbytes
;
3571 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3573 sysctl_user_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 17);
3576 subsys_initcall(init_user_reserve
);
3579 * Initialise sysctl_admin_reserve_kbytes.
3581 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3582 * to log in and kill a memory hogging process.
3584 * Systems with more than 256MB will reserve 8MB, enough to recover
3585 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3586 * only reserve 3% of free pages by default.
3588 static int init_admin_reserve(void)
3590 unsigned long free_kbytes
;
3592 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3594 sysctl_admin_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 13);
3597 subsys_initcall(init_admin_reserve
);
3600 * Reinititalise user and admin reserves if memory is added or removed.
3602 * The default user reserve max is 128MB, and the default max for the
3603 * admin reserve is 8MB. These are usually, but not always, enough to
3604 * enable recovery from a memory hogging process using login/sshd, a shell,
3605 * and tools like top. It may make sense to increase or even disable the
3606 * reserve depending on the existence of swap or variations in the recovery
3607 * tools. So, the admin may have changed them.
3609 * If memory is added and the reserves have been eliminated or increased above
3610 * the default max, then we'll trust the admin.
3612 * If memory is removed and there isn't enough free memory, then we
3613 * need to reset the reserves.
3615 * Otherwise keep the reserve set by the admin.
3617 static int reserve_mem_notifier(struct notifier_block
*nb
,
3618 unsigned long action
, void *data
)
3620 unsigned long tmp
, free_kbytes
;
3624 /* Default max is 128MB. Leave alone if modified by operator. */
3625 tmp
= sysctl_user_reserve_kbytes
;
3626 if (0 < tmp
&& tmp
< (1UL << 17))
3627 init_user_reserve();
3629 /* Default max is 8MB. Leave alone if modified by operator. */
3630 tmp
= sysctl_admin_reserve_kbytes
;
3631 if (0 < tmp
&& tmp
< (1UL << 13))
3632 init_admin_reserve();
3636 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3638 if (sysctl_user_reserve_kbytes
> free_kbytes
) {
3639 init_user_reserve();
3640 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3641 sysctl_user_reserve_kbytes
);
3644 if (sysctl_admin_reserve_kbytes
> free_kbytes
) {
3645 init_admin_reserve();
3646 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3647 sysctl_admin_reserve_kbytes
);
3656 static struct notifier_block reserve_mem_nb
= {
3657 .notifier_call
= reserve_mem_notifier
,
3660 static int __meminit
init_reserve_notifier(void)
3662 if (register_hotmemory_notifier(&reserve_mem_nb
))
3663 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3667 subsys_initcall(init_reserve_notifier
);