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>
47 #include <linux/oom.h>
49 #include <linux/uaccess.h>
50 #include <asm/cacheflush.h>
52 #include <asm/mmu_context.h>
56 #ifndef arch_mmap_check
57 #define arch_mmap_check(addr, len, flags) (0)
60 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
61 const int mmap_rnd_bits_min
= CONFIG_ARCH_MMAP_RND_BITS_MIN
;
62 const int mmap_rnd_bits_max
= CONFIG_ARCH_MMAP_RND_BITS_MAX
;
63 int mmap_rnd_bits __read_mostly
= CONFIG_ARCH_MMAP_RND_BITS
;
65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
66 const int mmap_rnd_compat_bits_min
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN
;
67 const int mmap_rnd_compat_bits_max
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX
;
68 int mmap_rnd_compat_bits __read_mostly
= CONFIG_ARCH_MMAP_RND_COMPAT_BITS
;
71 static bool ignore_rlimit_data
;
72 core_param(ignore_rlimit_data
, ignore_rlimit_data
, bool, 0644);
74 static void unmap_region(struct mm_struct
*mm
,
75 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
76 unsigned long start
, unsigned long end
);
78 /* description of effects of mapping type and prot in current implementation.
79 * this is due to the limited x86 page protection hardware. The expected
80 * behavior is in parens:
83 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
84 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
85 * w: (no) no w: (no) no w: (yes) yes w: (no) no
86 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
88 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
89 * w: (no) no w: (no) no w: (copy) copy w: (no) no
90 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
92 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
98 pgprot_t protection_map
[16] __ro_after_init
= {
99 __P000
, __P001
, __P010
, __P011
, __P100
, __P101
, __P110
, __P111
,
100 __S000
, __S001
, __S010
, __S011
, __S100
, __S101
, __S110
, __S111
103 pgprot_t
vm_get_page_prot(unsigned long vm_flags
)
105 return __pgprot(pgprot_val(protection_map
[vm_flags
&
106 (VM_READ
|VM_WRITE
|VM_EXEC
|VM_SHARED
)]) |
107 pgprot_val(arch_vm_get_page_prot(vm_flags
)));
109 EXPORT_SYMBOL(vm_get_page_prot
);
111 static pgprot_t
vm_pgprot_modify(pgprot_t oldprot
, unsigned long vm_flags
)
113 return pgprot_modify(oldprot
, vm_get_page_prot(vm_flags
));
116 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
117 void vma_set_page_prot(struct vm_area_struct
*vma
)
119 unsigned long vm_flags
= vma
->vm_flags
;
120 pgprot_t vm_page_prot
;
122 vm_page_prot
= vm_pgprot_modify(vma
->vm_page_prot
, vm_flags
);
123 if (vma_wants_writenotify(vma
, vm_page_prot
)) {
124 vm_flags
&= ~VM_SHARED
;
125 vm_page_prot
= vm_pgprot_modify(vm_page_prot
, vm_flags
);
127 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
128 WRITE_ONCE(vma
->vm_page_prot
, vm_page_prot
);
132 * Requires inode->i_mapping->i_mmap_rwsem
134 static void __remove_shared_vm_struct(struct vm_area_struct
*vma
,
135 struct file
*file
, struct address_space
*mapping
)
137 if (vma
->vm_flags
& VM_DENYWRITE
)
138 atomic_inc(&file_inode(file
)->i_writecount
);
139 if (vma
->vm_flags
& VM_SHARED
)
140 mapping_unmap_writable(mapping
);
142 flush_dcache_mmap_lock(mapping
);
143 vma_interval_tree_remove(vma
, &mapping
->i_mmap
);
144 flush_dcache_mmap_unlock(mapping
);
148 * Unlink a file-based vm structure from its interval tree, to hide
149 * vma from rmap and vmtruncate before freeing its page tables.
151 void unlink_file_vma(struct vm_area_struct
*vma
)
153 struct file
*file
= vma
->vm_file
;
156 struct address_space
*mapping
= file
->f_mapping
;
157 i_mmap_lock_write(mapping
);
158 __remove_shared_vm_struct(vma
, file
, mapping
);
159 i_mmap_unlock_write(mapping
);
164 * Close a vm structure and free it, returning the next.
166 static struct vm_area_struct
*remove_vma(struct vm_area_struct
*vma
)
168 struct vm_area_struct
*next
= vma
->vm_next
;
171 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
172 vma
->vm_ops
->close(vma
);
175 mpol_put(vma_policy(vma
));
176 kmem_cache_free(vm_area_cachep
, vma
);
180 static int do_brk_flags(unsigned long addr
, unsigned long request
, unsigned long flags
,
181 struct list_head
*uf
);
182 SYSCALL_DEFINE1(brk
, unsigned long, brk
)
184 unsigned long retval
;
185 unsigned long newbrk
, oldbrk
;
186 struct mm_struct
*mm
= current
->mm
;
187 struct vm_area_struct
*next
;
188 unsigned long min_brk
;
192 if (down_write_killable(&mm
->mmap_sem
))
195 #ifdef CONFIG_COMPAT_BRK
197 * CONFIG_COMPAT_BRK can still be overridden by setting
198 * randomize_va_space to 2, which will still cause mm->start_brk
199 * to be arbitrarily shifted
201 if (current
->brk_randomized
)
202 min_brk
= mm
->start_brk
;
204 min_brk
= mm
->end_data
;
206 min_brk
= mm
->start_brk
;
212 * Check against rlimit here. If this check is done later after the test
213 * of oldbrk with newbrk then it can escape the test and let the data
214 * segment grow beyond its set limit the in case where the limit is
215 * not page aligned -Ram Gupta
217 if (check_data_rlimit(rlimit(RLIMIT_DATA
), brk
, mm
->start_brk
,
218 mm
->end_data
, mm
->start_data
))
221 newbrk
= PAGE_ALIGN(brk
);
222 oldbrk
= PAGE_ALIGN(mm
->brk
);
223 if (oldbrk
== newbrk
)
226 /* Always allow shrinking brk. */
227 if (brk
<= mm
->brk
) {
228 if (!do_munmap(mm
, newbrk
, oldbrk
-newbrk
, &uf
))
233 /* Check against existing mmap mappings. */
234 next
= find_vma(mm
, oldbrk
);
235 if (next
&& newbrk
+ PAGE_SIZE
> vm_start_gap(next
))
238 /* Ok, looks good - let it rip. */
239 if (do_brk_flags(oldbrk
, newbrk
-oldbrk
, 0, &uf
) < 0)
244 populate
= newbrk
> oldbrk
&& (mm
->def_flags
& VM_LOCKED
) != 0;
245 up_write(&mm
->mmap_sem
);
246 userfaultfd_unmap_complete(mm
, &uf
);
248 mm_populate(oldbrk
, newbrk
- oldbrk
);
253 up_write(&mm
->mmap_sem
);
257 static long vma_compute_subtree_gap(struct vm_area_struct
*vma
)
259 unsigned long max
, prev_end
, subtree_gap
;
262 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
263 * allow two stack_guard_gaps between them here, and when choosing
264 * an unmapped area; whereas when expanding we only require one.
265 * That's a little inconsistent, but keeps the code here simpler.
267 max
= vm_start_gap(vma
);
269 prev_end
= vm_end_gap(vma
->vm_prev
);
275 if (vma
->vm_rb
.rb_left
) {
276 subtree_gap
= rb_entry(vma
->vm_rb
.rb_left
,
277 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
278 if (subtree_gap
> max
)
281 if (vma
->vm_rb
.rb_right
) {
282 subtree_gap
= rb_entry(vma
->vm_rb
.rb_right
,
283 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
284 if (subtree_gap
> max
)
290 #ifdef CONFIG_DEBUG_VM_RB
291 static int browse_rb(struct mm_struct
*mm
)
293 struct rb_root
*root
= &mm
->mm_rb
;
294 int i
= 0, j
, bug
= 0;
295 struct rb_node
*nd
, *pn
= NULL
;
296 unsigned long prev
= 0, pend
= 0;
298 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
299 struct vm_area_struct
*vma
;
300 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
301 if (vma
->vm_start
< prev
) {
302 pr_emerg("vm_start %lx < prev %lx\n",
303 vma
->vm_start
, prev
);
306 if (vma
->vm_start
< pend
) {
307 pr_emerg("vm_start %lx < pend %lx\n",
308 vma
->vm_start
, pend
);
311 if (vma
->vm_start
> vma
->vm_end
) {
312 pr_emerg("vm_start %lx > vm_end %lx\n",
313 vma
->vm_start
, vma
->vm_end
);
316 spin_lock(&mm
->page_table_lock
);
317 if (vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
)) {
318 pr_emerg("free gap %lx, correct %lx\n",
320 vma_compute_subtree_gap(vma
));
323 spin_unlock(&mm
->page_table_lock
);
326 prev
= vma
->vm_start
;
330 for (nd
= pn
; nd
; nd
= rb_prev(nd
))
333 pr_emerg("backwards %d, forwards %d\n", j
, i
);
339 static void validate_mm_rb(struct rb_root
*root
, struct vm_area_struct
*ignore
)
343 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
344 struct vm_area_struct
*vma
;
345 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
346 VM_BUG_ON_VMA(vma
!= ignore
&&
347 vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
),
352 static void validate_mm(struct mm_struct
*mm
)
356 unsigned long highest_address
= 0;
357 struct vm_area_struct
*vma
= mm
->mmap
;
360 struct anon_vma
*anon_vma
= vma
->anon_vma
;
361 struct anon_vma_chain
*avc
;
364 anon_vma_lock_read(anon_vma
);
365 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
366 anon_vma_interval_tree_verify(avc
);
367 anon_vma_unlock_read(anon_vma
);
370 highest_address
= vm_end_gap(vma
);
374 if (i
!= mm
->map_count
) {
375 pr_emerg("map_count %d vm_next %d\n", mm
->map_count
, i
);
378 if (highest_address
!= mm
->highest_vm_end
) {
379 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
380 mm
->highest_vm_end
, highest_address
);
384 if (i
!= mm
->map_count
) {
386 pr_emerg("map_count %d rb %d\n", mm
->map_count
, i
);
389 VM_BUG_ON_MM(bug
, mm
);
392 #define validate_mm_rb(root, ignore) do { } while (0)
393 #define validate_mm(mm) do { } while (0)
396 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks
, struct vm_area_struct
, vm_rb
,
397 unsigned long, rb_subtree_gap
, vma_compute_subtree_gap
)
400 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
401 * vma->vm_prev->vm_end values changed, without modifying the vma's position
404 static void vma_gap_update(struct vm_area_struct
*vma
)
407 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
408 * function that does exacltly what we want.
410 vma_gap_callbacks_propagate(&vma
->vm_rb
, NULL
);
413 static inline void vma_rb_insert(struct vm_area_struct
*vma
,
414 struct rb_root
*root
)
416 /* All rb_subtree_gap values must be consistent prior to insertion */
417 validate_mm_rb(root
, NULL
);
419 rb_insert_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
422 static void __vma_rb_erase(struct vm_area_struct
*vma
, struct rb_root
*root
)
425 * Note rb_erase_augmented is a fairly large inline function,
426 * so make sure we instantiate it only once with our desired
427 * augmented rbtree callbacks.
429 rb_erase_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
432 static __always_inline
void vma_rb_erase_ignore(struct vm_area_struct
*vma
,
433 struct rb_root
*root
,
434 struct vm_area_struct
*ignore
)
437 * All rb_subtree_gap values must be consistent prior to erase,
438 * with the possible exception of the "next" vma being erased if
439 * next->vm_start was reduced.
441 validate_mm_rb(root
, ignore
);
443 __vma_rb_erase(vma
, root
);
446 static __always_inline
void vma_rb_erase(struct vm_area_struct
*vma
,
447 struct rb_root
*root
)
450 * All rb_subtree_gap values must be consistent prior to erase,
451 * with the possible exception of the vma being erased.
453 validate_mm_rb(root
, vma
);
455 __vma_rb_erase(vma
, root
);
459 * vma has some anon_vma assigned, and is already inserted on that
460 * anon_vma's interval trees.
462 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
463 * vma must be removed from the anon_vma's interval trees using
464 * anon_vma_interval_tree_pre_update_vma().
466 * After the update, the vma will be reinserted using
467 * anon_vma_interval_tree_post_update_vma().
469 * The entire update must be protected by exclusive mmap_sem and by
470 * the root anon_vma's mutex.
473 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct
*vma
)
475 struct anon_vma_chain
*avc
;
477 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
478 anon_vma_interval_tree_remove(avc
, &avc
->anon_vma
->rb_root
);
482 anon_vma_interval_tree_post_update_vma(struct vm_area_struct
*vma
)
484 struct anon_vma_chain
*avc
;
486 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
487 anon_vma_interval_tree_insert(avc
, &avc
->anon_vma
->rb_root
);
490 static int find_vma_links(struct mm_struct
*mm
, unsigned long addr
,
491 unsigned long end
, struct vm_area_struct
**pprev
,
492 struct rb_node
***rb_link
, struct rb_node
**rb_parent
)
494 struct rb_node
**__rb_link
, *__rb_parent
, *rb_prev
;
496 __rb_link
= &mm
->mm_rb
.rb_node
;
497 rb_prev
= __rb_parent
= NULL
;
500 struct vm_area_struct
*vma_tmp
;
502 __rb_parent
= *__rb_link
;
503 vma_tmp
= rb_entry(__rb_parent
, struct vm_area_struct
, vm_rb
);
505 if (vma_tmp
->vm_end
> addr
) {
506 /* Fail if an existing vma overlaps the area */
507 if (vma_tmp
->vm_start
< end
)
509 __rb_link
= &__rb_parent
->rb_left
;
511 rb_prev
= __rb_parent
;
512 __rb_link
= &__rb_parent
->rb_right
;
518 *pprev
= rb_entry(rb_prev
, struct vm_area_struct
, vm_rb
);
519 *rb_link
= __rb_link
;
520 *rb_parent
= __rb_parent
;
524 static unsigned long count_vma_pages_range(struct mm_struct
*mm
,
525 unsigned long addr
, unsigned long end
)
527 unsigned long nr_pages
= 0;
528 struct vm_area_struct
*vma
;
530 /* Find first overlaping mapping */
531 vma
= find_vma_intersection(mm
, addr
, end
);
535 nr_pages
= (min(end
, vma
->vm_end
) -
536 max(addr
, vma
->vm_start
)) >> PAGE_SHIFT
;
538 /* Iterate over the rest of the overlaps */
539 for (vma
= vma
->vm_next
; vma
; vma
= vma
->vm_next
) {
540 unsigned long overlap_len
;
542 if (vma
->vm_start
> end
)
545 overlap_len
= min(end
, vma
->vm_end
) - vma
->vm_start
;
546 nr_pages
+= overlap_len
>> PAGE_SHIFT
;
552 void __vma_link_rb(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
553 struct rb_node
**rb_link
, struct rb_node
*rb_parent
)
555 /* Update tracking information for the gap following the new vma. */
557 vma_gap_update(vma
->vm_next
);
559 mm
->highest_vm_end
= vm_end_gap(vma
);
562 * vma->vm_prev wasn't known when we followed the rbtree to find the
563 * correct insertion point for that vma. As a result, we could not
564 * update the vma vm_rb parents rb_subtree_gap values on the way down.
565 * So, we first insert the vma with a zero rb_subtree_gap value
566 * (to be consistent with what we did on the way down), and then
567 * immediately update the gap to the correct value. Finally we
568 * rebalance the rbtree after all augmented values have been set.
570 rb_link_node(&vma
->vm_rb
, rb_parent
, rb_link
);
571 vma
->rb_subtree_gap
= 0;
573 vma_rb_insert(vma
, &mm
->mm_rb
);
576 static void __vma_link_file(struct vm_area_struct
*vma
)
582 struct address_space
*mapping
= file
->f_mapping
;
584 if (vma
->vm_flags
& VM_DENYWRITE
)
585 atomic_dec(&file_inode(file
)->i_writecount
);
586 if (vma
->vm_flags
& VM_SHARED
)
587 atomic_inc(&mapping
->i_mmap_writable
);
589 flush_dcache_mmap_lock(mapping
);
590 vma_interval_tree_insert(vma
, &mapping
->i_mmap
);
591 flush_dcache_mmap_unlock(mapping
);
596 __vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
597 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
598 struct rb_node
*rb_parent
)
600 __vma_link_list(mm
, vma
, prev
, rb_parent
);
601 __vma_link_rb(mm
, vma
, rb_link
, rb_parent
);
604 static void vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
605 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
606 struct rb_node
*rb_parent
)
608 struct address_space
*mapping
= NULL
;
611 mapping
= vma
->vm_file
->f_mapping
;
612 i_mmap_lock_write(mapping
);
615 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
616 __vma_link_file(vma
);
619 i_mmap_unlock_write(mapping
);
626 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
627 * mm's list and rbtree. It has already been inserted into the interval tree.
629 static void __insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
631 struct vm_area_struct
*prev
;
632 struct rb_node
**rb_link
, *rb_parent
;
634 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
635 &prev
, &rb_link
, &rb_parent
))
637 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
641 static __always_inline
void __vma_unlink_common(struct mm_struct
*mm
,
642 struct vm_area_struct
*vma
,
643 struct vm_area_struct
*prev
,
645 struct vm_area_struct
*ignore
)
647 struct vm_area_struct
*next
;
649 vma_rb_erase_ignore(vma
, &mm
->mm_rb
, ignore
);
652 prev
->vm_next
= next
;
656 prev
->vm_next
= next
;
661 next
->vm_prev
= prev
;
664 vmacache_invalidate(mm
);
667 static inline void __vma_unlink_prev(struct mm_struct
*mm
,
668 struct vm_area_struct
*vma
,
669 struct vm_area_struct
*prev
)
671 __vma_unlink_common(mm
, vma
, prev
, true, vma
);
675 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
676 * is already present in an i_mmap tree without adjusting the tree.
677 * The following helper function should be used when such adjustments
678 * are necessary. The "insert" vma (if any) is to be inserted
679 * before we drop the necessary locks.
681 int __vma_adjust(struct vm_area_struct
*vma
, unsigned long start
,
682 unsigned long end
, pgoff_t pgoff
, struct vm_area_struct
*insert
,
683 struct vm_area_struct
*expand
)
685 struct mm_struct
*mm
= vma
->vm_mm
;
686 struct vm_area_struct
*next
= vma
->vm_next
, *orig_vma
= vma
;
687 struct address_space
*mapping
= NULL
;
688 struct rb_root_cached
*root
= NULL
;
689 struct anon_vma
*anon_vma
= NULL
;
690 struct file
*file
= vma
->vm_file
;
691 bool start_changed
= false, end_changed
= false;
692 long adjust_next
= 0;
695 if (next
&& !insert
) {
696 struct vm_area_struct
*exporter
= NULL
, *importer
= NULL
;
698 if (end
>= next
->vm_end
) {
700 * vma expands, overlapping all the next, and
701 * perhaps the one after too (mprotect case 6).
702 * The only other cases that gets here are
703 * case 1, case 7 and case 8.
705 if (next
== expand
) {
707 * The only case where we don't expand "vma"
708 * and we expand "next" instead is case 8.
710 VM_WARN_ON(end
!= next
->vm_end
);
712 * remove_next == 3 means we're
713 * removing "vma" and that to do so we
714 * swapped "vma" and "next".
717 VM_WARN_ON(file
!= next
->vm_file
);
720 VM_WARN_ON(expand
!= vma
);
722 * case 1, 6, 7, remove_next == 2 is case 6,
723 * remove_next == 1 is case 1 or 7.
725 remove_next
= 1 + (end
> next
->vm_end
);
726 VM_WARN_ON(remove_next
== 2 &&
727 end
!= next
->vm_next
->vm_end
);
728 VM_WARN_ON(remove_next
== 1 &&
729 end
!= next
->vm_end
);
730 /* trim end to next, for case 6 first pass */
738 * If next doesn't have anon_vma, import from vma after
739 * next, if the vma overlaps with it.
741 if (remove_next
== 2 && !next
->anon_vma
)
742 exporter
= next
->vm_next
;
744 } else if (end
> next
->vm_start
) {
746 * vma expands, overlapping part of the next:
747 * mprotect case 5 shifting the boundary up.
749 adjust_next
= (end
- next
->vm_start
) >> PAGE_SHIFT
;
752 VM_WARN_ON(expand
!= importer
);
753 } else if (end
< vma
->vm_end
) {
755 * vma shrinks, and !insert tells it's not
756 * split_vma inserting another: so it must be
757 * mprotect case 4 shifting the boundary down.
759 adjust_next
= -((vma
->vm_end
- end
) >> PAGE_SHIFT
);
762 VM_WARN_ON(expand
!= importer
);
766 * Easily overlooked: when mprotect shifts the boundary,
767 * make sure the expanding vma has anon_vma set if the
768 * shrinking vma had, to cover any anon pages imported.
770 if (exporter
&& exporter
->anon_vma
&& !importer
->anon_vma
) {
773 importer
->anon_vma
= exporter
->anon_vma
;
774 error
= anon_vma_clone(importer
, exporter
);
780 vma_adjust_trans_huge(orig_vma
, start
, end
, adjust_next
);
783 mapping
= file
->f_mapping
;
784 root
= &mapping
->i_mmap
;
785 uprobe_munmap(vma
, vma
->vm_start
, vma
->vm_end
);
788 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
790 i_mmap_lock_write(mapping
);
793 * Put into interval tree now, so instantiated pages
794 * are visible to arm/parisc __flush_dcache_page
795 * throughout; but we cannot insert into address
796 * space until vma start or end is updated.
798 __vma_link_file(insert
);
802 anon_vma
= vma
->anon_vma
;
803 if (!anon_vma
&& adjust_next
)
804 anon_vma
= next
->anon_vma
;
806 VM_WARN_ON(adjust_next
&& next
->anon_vma
&&
807 anon_vma
!= next
->anon_vma
);
808 anon_vma_lock_write(anon_vma
);
809 anon_vma_interval_tree_pre_update_vma(vma
);
811 anon_vma_interval_tree_pre_update_vma(next
);
815 flush_dcache_mmap_lock(mapping
);
816 vma_interval_tree_remove(vma
, root
);
818 vma_interval_tree_remove(next
, root
);
821 if (start
!= vma
->vm_start
) {
822 vma
->vm_start
= start
;
823 start_changed
= true;
825 if (end
!= vma
->vm_end
) {
829 vma
->vm_pgoff
= pgoff
;
831 next
->vm_start
+= adjust_next
<< PAGE_SHIFT
;
832 next
->vm_pgoff
+= adjust_next
;
837 vma_interval_tree_insert(next
, root
);
838 vma_interval_tree_insert(vma
, root
);
839 flush_dcache_mmap_unlock(mapping
);
844 * vma_merge has merged next into vma, and needs
845 * us to remove next before dropping the locks.
847 if (remove_next
!= 3)
848 __vma_unlink_prev(mm
, next
, vma
);
851 * vma is not before next if they've been
854 * pre-swap() next->vm_start was reduced so
855 * tell validate_mm_rb to ignore pre-swap()
856 * "next" (which is stored in post-swap()
859 __vma_unlink_common(mm
, next
, NULL
, false, vma
);
861 __remove_shared_vm_struct(next
, file
, mapping
);
864 * split_vma has split insert from vma, and needs
865 * us to insert it before dropping the locks
866 * (it may either follow vma or precede it).
868 __insert_vm_struct(mm
, insert
);
874 mm
->highest_vm_end
= vm_end_gap(vma
);
875 else if (!adjust_next
)
876 vma_gap_update(next
);
881 anon_vma_interval_tree_post_update_vma(vma
);
883 anon_vma_interval_tree_post_update_vma(next
);
884 anon_vma_unlock_write(anon_vma
);
887 i_mmap_unlock_write(mapping
);
898 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
902 anon_vma_merge(vma
, next
);
904 mpol_put(vma_policy(next
));
905 kmem_cache_free(vm_area_cachep
, next
);
907 * In mprotect's case 6 (see comments on vma_merge),
908 * we must remove another next too. It would clutter
909 * up the code too much to do both in one go.
911 if (remove_next
!= 3) {
913 * If "next" was removed and vma->vm_end was
914 * expanded (up) over it, in turn
915 * "next->vm_prev->vm_end" changed and the
916 * "vma->vm_next" gap must be updated.
921 * For the scope of the comment "next" and
922 * "vma" considered pre-swap(): if "vma" was
923 * removed, next->vm_start was expanded (down)
924 * over it and the "next" gap must be updated.
925 * Because of the swap() the post-swap() "vma"
926 * actually points to pre-swap() "next"
927 * (post-swap() "next" as opposed is now a
932 if (remove_next
== 2) {
938 vma_gap_update(next
);
941 * If remove_next == 2 we obviously can't
944 * If remove_next == 3 we can't reach this
945 * path because pre-swap() next is always not
946 * NULL. pre-swap() "next" is not being
947 * removed and its next->vm_end is not altered
948 * (and furthermore "end" already matches
949 * next->vm_end in remove_next == 3).
951 * We reach this only in the remove_next == 1
952 * case if the "next" vma that was removed was
953 * the highest vma of the mm. However in such
954 * case next->vm_end == "end" and the extended
955 * "vma" has vma->vm_end == next->vm_end so
956 * mm->highest_vm_end doesn't need any update
957 * in remove_next == 1 case.
959 VM_WARN_ON(mm
->highest_vm_end
!= vm_end_gap(vma
));
971 * If the vma has a ->close operation then the driver probably needs to release
972 * per-vma resources, so we don't attempt to merge those.
974 static inline int is_mergeable_vma(struct vm_area_struct
*vma
,
975 struct file
*file
, unsigned long vm_flags
,
976 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
,
977 const char __user
*anon_name
)
980 * VM_SOFTDIRTY should not prevent from VMA merging, if we
981 * match the flags but dirty bit -- the caller should mark
982 * merged VMA as dirty. If dirty bit won't be excluded from
983 * comparison, we increase pressue on the memory system forcing
984 * the kernel to generate new VMAs when old one could be
987 if ((vma
->vm_flags
^ vm_flags
) & ~VM_SOFTDIRTY
)
989 if (vma
->vm_file
!= file
)
991 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
993 if (!is_mergeable_vm_userfaultfd_ctx(vma
, vm_userfaultfd_ctx
))
995 if (vma_get_anon_name(vma
) != anon_name
)
1000 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
1001 struct anon_vma
*anon_vma2
,
1002 struct vm_area_struct
*vma
)
1005 * The list_is_singular() test is to avoid merging VMA cloned from
1006 * parents. This can improve scalability caused by anon_vma lock.
1008 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
1009 list_is_singular(&vma
->anon_vma_chain
)))
1011 return anon_vma1
== anon_vma2
;
1015 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1016 * in front of (at a lower virtual address and file offset than) the vma.
1018 * We cannot merge two vmas if they have differently assigned (non-NULL)
1019 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1021 * We don't check here for the merged mmap wrapping around the end of pagecache
1022 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1023 * wrap, nor mmaps which cover the final page at index -1UL.
1026 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
1027 struct anon_vma
*anon_vma
, struct file
*file
,
1029 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
,
1030 const char __user
*anon_name
)
1032 if (is_mergeable_vma(vma
, file
, vm_flags
, vm_userfaultfd_ctx
, anon_name
) &&
1033 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
1034 if (vma
->vm_pgoff
== vm_pgoff
)
1041 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1042 * beyond (at a higher virtual address and file offset than) the vma.
1044 * We cannot merge two vmas if they have differently assigned (non-NULL)
1045 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1048 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
1049 struct anon_vma
*anon_vma
, struct file
*file
,
1051 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
,
1052 const char __user
*anon_name
)
1054 if (is_mergeable_vma(vma
, file
, vm_flags
, vm_userfaultfd_ctx
, anon_name
) &&
1055 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
1057 vm_pglen
= vma_pages(vma
);
1058 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
1065 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
1066 * figure out whether that can be merged with its predecessor or its
1067 * successor. Or both (it neatly fills a hole).
1069 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1070 * certain not to be mapped by the time vma_merge is called; but when
1071 * called for mprotect, it is certain to be already mapped (either at
1072 * an offset within prev, or at the start of next), and the flags of
1073 * this area are about to be changed to vm_flags - and the no-change
1074 * case has already been eliminated.
1076 * The following mprotect cases have to be considered, where AAAA is
1077 * the area passed down from mprotect_fixup, never extending beyond one
1078 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1080 * AAAA AAAA AAAA AAAA
1081 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1082 * cannot merge might become might become might become
1083 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1084 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1085 * mremap move: PPPPXXXXXXXX 8
1087 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1088 * might become case 1 below case 2 below case 3 below
1090 * It is important for case 8 that the the vma NNNN overlapping the
1091 * region AAAA is never going to extended over XXXX. Instead XXXX must
1092 * be extended in region AAAA and NNNN must be removed. This way in
1093 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1094 * rmap_locks, the properties of the merged vma will be already
1095 * correct for the whole merged range. Some of those properties like
1096 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1097 * be correct for the whole merged range immediately after the
1098 * rmap_locks are released. Otherwise if XXXX would be removed and
1099 * NNNN would be extended over the XXXX range, remove_migration_ptes
1100 * or other rmap walkers (if working on addresses beyond the "end"
1101 * parameter) may establish ptes with the wrong permissions of NNNN
1102 * instead of the right permissions of XXXX.
1104 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
1105 struct vm_area_struct
*prev
, unsigned long addr
,
1106 unsigned long end
, unsigned long vm_flags
,
1107 struct anon_vma
*anon_vma
, struct file
*file
,
1108 pgoff_t pgoff
, struct mempolicy
*policy
,
1109 struct vm_userfaultfd_ctx vm_userfaultfd_ctx
,
1110 const char __user
*anon_name
)
1112 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
1113 struct vm_area_struct
*area
, *next
;
1117 * We later require that vma->vm_flags == vm_flags,
1118 * so this tests vma->vm_flags & VM_SPECIAL, too.
1120 if (vm_flags
& VM_SPECIAL
)
1124 next
= prev
->vm_next
;
1128 if (area
&& area
->vm_end
== end
) /* cases 6, 7, 8 */
1129 next
= next
->vm_next
;
1131 /* verify some invariant that must be enforced by the caller */
1132 VM_WARN_ON(prev
&& addr
<= prev
->vm_start
);
1133 VM_WARN_ON(area
&& end
> area
->vm_end
);
1134 VM_WARN_ON(addr
>= end
);
1137 * Can it merge with the predecessor?
1139 if (prev
&& prev
->vm_end
== addr
&&
1140 mpol_equal(vma_policy(prev
), policy
) &&
1141 can_vma_merge_after(prev
, vm_flags
,
1142 anon_vma
, file
, pgoff
,
1146 * OK, it can. Can we now merge in the successor as well?
1148 if (next
&& end
== next
->vm_start
&&
1149 mpol_equal(policy
, vma_policy(next
)) &&
1150 can_vma_merge_before(next
, vm_flags
,
1155 is_mergeable_anon_vma(prev
->anon_vma
,
1156 next
->anon_vma
, NULL
)) {
1158 err
= __vma_adjust(prev
, prev
->vm_start
,
1159 next
->vm_end
, prev
->vm_pgoff
, NULL
,
1161 } else /* cases 2, 5, 7 */
1162 err
= __vma_adjust(prev
, prev
->vm_start
,
1163 end
, prev
->vm_pgoff
, NULL
, prev
);
1166 khugepaged_enter_vma_merge(prev
, vm_flags
);
1171 * Can this new request be merged in front of next?
1173 if (next
&& end
== next
->vm_start
&&
1174 mpol_equal(policy
, vma_policy(next
)) &&
1175 can_vma_merge_before(next
, vm_flags
,
1176 anon_vma
, file
, pgoff
+pglen
,
1179 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
1180 err
= __vma_adjust(prev
, prev
->vm_start
,
1181 addr
, prev
->vm_pgoff
, NULL
, next
);
1182 else { /* cases 3, 8 */
1183 err
= __vma_adjust(area
, addr
, next
->vm_end
,
1184 next
->vm_pgoff
- pglen
, NULL
, next
);
1186 * In case 3 area is already equal to next and
1187 * this is a noop, but in case 8 "area" has
1188 * been removed and next was expanded over it.
1194 khugepaged_enter_vma_merge(area
, vm_flags
);
1202 * Rough compatbility check to quickly see if it's even worth looking
1203 * at sharing an anon_vma.
1205 * They need to have the same vm_file, and the flags can only differ
1206 * in things that mprotect may change.
1208 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1209 * we can merge the two vma's. For example, we refuse to merge a vma if
1210 * there is a vm_ops->close() function, because that indicates that the
1211 * driver is doing some kind of reference counting. But that doesn't
1212 * really matter for the anon_vma sharing case.
1214 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1216 return a
->vm_end
== b
->vm_start
&&
1217 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
1218 a
->vm_file
== b
->vm_file
&&
1219 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_READ
|VM_WRITE
|VM_EXEC
|VM_SOFTDIRTY
)) &&
1220 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
1224 * Do some basic sanity checking to see if we can re-use the anon_vma
1225 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1226 * the same as 'old', the other will be the new one that is trying
1227 * to share the anon_vma.
1229 * NOTE! This runs with mm_sem held for reading, so it is possible that
1230 * the anon_vma of 'old' is concurrently in the process of being set up
1231 * by another page fault trying to merge _that_. But that's ok: if it
1232 * is being set up, that automatically means that it will be a singleton
1233 * acceptable for merging, so we can do all of this optimistically. But
1234 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1236 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1237 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1238 * is to return an anon_vma that is "complex" due to having gone through
1241 * We also make sure that the two vma's are compatible (adjacent,
1242 * and with the same memory policies). That's all stable, even with just
1243 * a read lock on the mm_sem.
1245 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1247 if (anon_vma_compatible(a
, b
)) {
1248 struct anon_vma
*anon_vma
= READ_ONCE(old
->anon_vma
);
1250 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
1257 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1258 * neighbouring vmas for a suitable anon_vma, before it goes off
1259 * to allocate a new anon_vma. It checks because a repetitive
1260 * sequence of mprotects and faults may otherwise lead to distinct
1261 * anon_vmas being allocated, preventing vma merge in subsequent
1264 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
1266 struct anon_vma
*anon_vma
;
1267 struct vm_area_struct
*near
;
1269 near
= vma
->vm_next
;
1273 anon_vma
= reusable_anon_vma(near
, vma
, near
);
1277 near
= vma
->vm_prev
;
1281 anon_vma
= reusable_anon_vma(near
, near
, vma
);
1286 * There's no absolute need to look only at touching neighbours:
1287 * we could search further afield for "compatible" anon_vmas.
1288 * But it would probably just be a waste of time searching,
1289 * or lead to too many vmas hanging off the same anon_vma.
1290 * We're trying to allow mprotect remerging later on,
1291 * not trying to minimize memory used for anon_vmas.
1297 * If a hint addr is less than mmap_min_addr change hint to be as
1298 * low as possible but still greater than mmap_min_addr
1300 static inline unsigned long round_hint_to_min(unsigned long hint
)
1303 if (((void *)hint
!= NULL
) &&
1304 (hint
< mmap_min_addr
))
1305 return PAGE_ALIGN(mmap_min_addr
);
1309 static inline int mlock_future_check(struct mm_struct
*mm
,
1310 unsigned long flags
,
1313 unsigned long locked
, lock_limit
;
1315 /* mlock MCL_FUTURE? */
1316 if (flags
& VM_LOCKED
) {
1317 locked
= len
>> PAGE_SHIFT
;
1318 locked
+= mm
->locked_vm
;
1319 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1320 lock_limit
>>= PAGE_SHIFT
;
1321 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1327 static inline u64
file_mmap_size_max(struct file
*file
, struct inode
*inode
)
1329 if (S_ISREG(inode
->i_mode
))
1330 return MAX_LFS_FILESIZE
;
1332 if (S_ISBLK(inode
->i_mode
))
1333 return MAX_LFS_FILESIZE
;
1335 /* Special "we do even unsigned file positions" case */
1336 if (file
->f_mode
& FMODE_UNSIGNED_OFFSET
)
1339 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1343 static inline bool file_mmap_ok(struct file
*file
, struct inode
*inode
,
1344 unsigned long pgoff
, unsigned long len
)
1346 u64 maxsize
= file_mmap_size_max(file
, inode
);
1348 if (maxsize
&& len
> maxsize
)
1351 if (pgoff
> maxsize
>> PAGE_SHIFT
)
1357 * The caller must hold down_write(¤t->mm->mmap_sem).
1359 unsigned long do_mmap(struct file
*file
, unsigned long addr
,
1360 unsigned long len
, unsigned long prot
,
1361 unsigned long flags
, vm_flags_t vm_flags
,
1362 unsigned long pgoff
, unsigned long *populate
,
1363 struct list_head
*uf
)
1365 struct mm_struct
*mm
= current
->mm
;
1374 * Does the application expect PROT_READ to imply PROT_EXEC?
1376 * (the exception is when the underlying filesystem is noexec
1377 * mounted, in which case we dont add PROT_EXEC.)
1379 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1380 if (!(file
&& path_noexec(&file
->f_path
)))
1383 if (!(flags
& MAP_FIXED
))
1384 addr
= round_hint_to_min(addr
);
1386 /* Careful about overflows.. */
1387 len
= PAGE_ALIGN(len
);
1391 /* offset overflow? */
1392 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1395 /* Too many mappings? */
1396 if (mm
->map_count
> sysctl_max_map_count
)
1399 /* Obtain the address to map to. we verify (or select) it and ensure
1400 * that it represents a valid section of the address space.
1402 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1403 if (offset_in_page(addr
))
1406 if (prot
== PROT_EXEC
) {
1407 pkey
= execute_only_pkey(mm
);
1412 /* Do simple checking here so the lower-level routines won't have
1413 * to. we assume access permissions have been handled by the open
1414 * of the memory object, so we don't do any here.
1416 vm_flags
|= calc_vm_prot_bits(prot
, pkey
) | calc_vm_flag_bits(flags
) |
1417 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1419 if (flags
& MAP_LOCKED
)
1420 if (!can_do_mlock())
1423 if (mlock_future_check(mm
, vm_flags
, len
))
1427 struct inode
*inode
= file_inode(file
);
1429 if (!file_mmap_ok(file
, inode
, pgoff
, len
))
1432 switch (flags
& MAP_TYPE
) {
1434 if ((prot
&PROT_WRITE
) && !(file
->f_mode
&FMODE_WRITE
))
1438 * Make sure we don't allow writing to an append-only
1441 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1445 * Make sure there are no mandatory locks on the file.
1447 if (locks_verify_locked(file
))
1450 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1451 if (!(file
->f_mode
& FMODE_WRITE
))
1452 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1456 if (!(file
->f_mode
& FMODE_READ
))
1458 if (path_noexec(&file
->f_path
)) {
1459 if (vm_flags
& VM_EXEC
)
1461 vm_flags
&= ~VM_MAYEXEC
;
1464 if (!file
->f_op
->mmap
)
1466 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1474 switch (flags
& MAP_TYPE
) {
1476 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1482 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1486 * Set pgoff according to addr for anon_vma.
1488 pgoff
= addr
>> PAGE_SHIFT
;
1496 * Set 'VM_NORESERVE' if we should not account for the
1497 * memory use of this mapping.
1499 if (flags
& MAP_NORESERVE
) {
1500 /* We honor MAP_NORESERVE if allowed to overcommit */
1501 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1502 vm_flags
|= VM_NORESERVE
;
1504 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1505 if (file
&& is_file_hugepages(file
))
1506 vm_flags
|= VM_NORESERVE
;
1509 addr
= mmap_region(file
, addr
, len
, vm_flags
, pgoff
, uf
);
1510 if (!IS_ERR_VALUE(addr
) &&
1511 ((vm_flags
& VM_LOCKED
) ||
1512 (flags
& (MAP_POPULATE
| MAP_NONBLOCK
)) == MAP_POPULATE
))
1517 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1518 unsigned long, prot
, unsigned long, flags
,
1519 unsigned long, fd
, unsigned long, pgoff
)
1521 struct file
*file
= NULL
;
1522 unsigned long retval
;
1524 if (!(flags
& MAP_ANONYMOUS
)) {
1525 audit_mmap_fd(fd
, flags
);
1529 if (is_file_hugepages(file
))
1530 len
= ALIGN(len
, huge_page_size(hstate_file(file
)));
1532 if (unlikely(flags
& MAP_HUGETLB
&& !is_file_hugepages(file
)))
1534 } else if (flags
& MAP_HUGETLB
) {
1535 struct user_struct
*user
= NULL
;
1538 hs
= hstate_sizelog((flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1542 len
= ALIGN(len
, huge_page_size(hs
));
1544 * VM_NORESERVE is used because the reservations will be
1545 * taken when vm_ops->mmap() is called
1546 * A dummy user value is used because we are not locking
1547 * memory so no accounting is necessary
1549 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, len
,
1551 &user
, HUGETLB_ANONHUGE_INODE
,
1552 (flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1554 return PTR_ERR(file
);
1557 flags
&= ~(MAP_EXECUTABLE
| MAP_DENYWRITE
);
1559 retval
= vm_mmap_pgoff(file
, addr
, len
, prot
, flags
, pgoff
);
1566 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1567 struct mmap_arg_struct
{
1571 unsigned long flags
;
1573 unsigned long offset
;
1576 SYSCALL_DEFINE1(old_mmap
, struct mmap_arg_struct __user
*, arg
)
1578 struct mmap_arg_struct a
;
1580 if (copy_from_user(&a
, arg
, sizeof(a
)))
1582 if (offset_in_page(a
.offset
))
1585 return sys_mmap_pgoff(a
.addr
, a
.len
, a
.prot
, a
.flags
, a
.fd
,
1586 a
.offset
>> PAGE_SHIFT
);
1588 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1591 * Some shared mappigns will want the pages marked read-only
1592 * to track write events. If so, we'll downgrade vm_page_prot
1593 * to the private version (using protection_map[] without the
1596 int vma_wants_writenotify(struct vm_area_struct
*vma
, pgprot_t vm_page_prot
)
1598 vm_flags_t vm_flags
= vma
->vm_flags
;
1599 const struct vm_operations_struct
*vm_ops
= vma
->vm_ops
;
1601 /* If it was private or non-writable, the write bit is already clear */
1602 if ((vm_flags
& (VM_WRITE
|VM_SHARED
)) != ((VM_WRITE
|VM_SHARED
)))
1605 /* The backer wishes to know when pages are first written to? */
1606 if (vm_ops
&& (vm_ops
->page_mkwrite
|| vm_ops
->pfn_mkwrite
))
1609 /* The open routine did something to the protections that pgprot_modify
1610 * won't preserve? */
1611 if (pgprot_val(vm_page_prot
) !=
1612 pgprot_val(vm_pgprot_modify(vm_page_prot
, vm_flags
)))
1615 /* Do we need to track softdirty? */
1616 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY
) && !(vm_flags
& VM_SOFTDIRTY
))
1619 /* Specialty mapping? */
1620 if (vm_flags
& VM_PFNMAP
)
1623 /* Can the mapping track the dirty pages? */
1624 return vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
1625 mapping_cap_account_dirty(vma
->vm_file
->f_mapping
);
1629 * We account for memory if it's a private writeable mapping,
1630 * not hugepages and VM_NORESERVE wasn't set.
1632 static inline int accountable_mapping(struct file
*file
, vm_flags_t vm_flags
)
1635 * hugetlb has its own accounting separate from the core VM
1636 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1638 if (file
&& is_file_hugepages(file
))
1641 return (vm_flags
& (VM_NORESERVE
| VM_SHARED
| VM_WRITE
)) == VM_WRITE
;
1644 unsigned long mmap_region(struct file
*file
, unsigned long addr
,
1645 unsigned long len
, vm_flags_t vm_flags
, unsigned long pgoff
,
1646 struct list_head
*uf
)
1648 struct mm_struct
*mm
= current
->mm
;
1649 struct vm_area_struct
*vma
, *prev
;
1651 struct rb_node
**rb_link
, *rb_parent
;
1652 unsigned long charged
= 0;
1654 /* Check against address space limit. */
1655 if (!may_expand_vm(mm
, vm_flags
, len
>> PAGE_SHIFT
)) {
1656 unsigned long nr_pages
;
1659 * MAP_FIXED may remove pages of mappings that intersects with
1660 * requested mapping. Account for the pages it would unmap.
1662 nr_pages
= count_vma_pages_range(mm
, addr
, addr
+ len
);
1664 if (!may_expand_vm(mm
, vm_flags
,
1665 (len
>> PAGE_SHIFT
) - nr_pages
))
1669 /* Clear old maps */
1670 while (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
,
1672 if (do_munmap(mm
, addr
, len
, uf
))
1677 * Private writable mapping: check memory availability
1679 if (accountable_mapping(file
, vm_flags
)) {
1680 charged
= len
>> PAGE_SHIFT
;
1681 if (security_vm_enough_memory_mm(mm
, charged
))
1683 vm_flags
|= VM_ACCOUNT
;
1687 * Can we just expand an old mapping?
1689 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
,
1690 NULL
, file
, pgoff
, NULL
, NULL_VM_UFFD_CTX
, NULL
);
1695 * Determine the object being mapped and call the appropriate
1696 * specific mapper. the address has already been validated, but
1697 * not unmapped, but the maps are removed from the list.
1699 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
1706 vma
->vm_start
= addr
;
1707 vma
->vm_end
= addr
+ len
;
1708 vma
->vm_flags
= vm_flags
;
1709 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1710 vma
->vm_pgoff
= pgoff
;
1711 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
1714 if (vm_flags
& VM_DENYWRITE
) {
1715 error
= deny_write_access(file
);
1719 if (vm_flags
& VM_SHARED
) {
1720 error
= mapping_map_writable(file
->f_mapping
);
1722 goto allow_write_and_free_vma
;
1725 /* ->mmap() can change vma->vm_file, but must guarantee that
1726 * vma_link() below can deny write-access if VM_DENYWRITE is set
1727 * and map writably if VM_SHARED is set. This usually means the
1728 * new file must not have been exposed to user-space, yet.
1730 vma
->vm_file
= get_file(file
);
1731 error
= call_mmap(file
, vma
);
1733 goto unmap_and_free_vma
;
1735 /* Can addr have changed??
1737 * Answer: Yes, several device drivers can do it in their
1738 * f_op->mmap method. -DaveM
1739 * Bug: If addr is changed, prev, rb_link, rb_parent should
1740 * be updated for vma_link()
1742 WARN_ON_ONCE(addr
!= vma
->vm_start
);
1744 addr
= vma
->vm_start
;
1745 vm_flags
= vma
->vm_flags
;
1746 } else if (vm_flags
& VM_SHARED
) {
1747 error
= shmem_zero_setup(vma
);
1752 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1753 /* Once vma denies write, undo our temporary denial count */
1755 if (vm_flags
& VM_SHARED
)
1756 mapping_unmap_writable(file
->f_mapping
);
1757 if (vm_flags
& VM_DENYWRITE
)
1758 allow_write_access(file
);
1760 file
= vma
->vm_file
;
1762 perf_event_mmap(vma
);
1764 vm_stat_account(mm
, vm_flags
, len
>> PAGE_SHIFT
);
1765 if (vm_flags
& VM_LOCKED
) {
1766 if (!((vm_flags
& VM_SPECIAL
) || is_vm_hugetlb_page(vma
) ||
1767 vma
== get_gate_vma(current
->mm
)))
1768 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1770 vma
->vm_flags
&= VM_LOCKED_CLEAR_MASK
;
1777 * New (or expanded) vma always get soft dirty status.
1778 * Otherwise user-space soft-dirty page tracker won't
1779 * be able to distinguish situation when vma area unmapped,
1780 * then new mapped in-place (which must be aimed as
1781 * a completely new data area).
1783 vma
->vm_flags
|= VM_SOFTDIRTY
;
1785 vma_set_page_prot(vma
);
1790 vma
->vm_file
= NULL
;
1793 /* Undo any partial mapping done by a device driver. */
1794 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1796 if (vm_flags
& VM_SHARED
)
1797 mapping_unmap_writable(file
->f_mapping
);
1798 allow_write_and_free_vma
:
1799 if (vm_flags
& VM_DENYWRITE
)
1800 allow_write_access(file
);
1802 kmem_cache_free(vm_area_cachep
, vma
);
1805 vm_unacct_memory(charged
);
1809 unsigned long unmapped_area(struct vm_unmapped_area_info
*info
)
1812 * We implement the search by looking for an rbtree node that
1813 * immediately follows a suitable gap. That is,
1814 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1815 * - gap_end = vma->vm_start >= info->low_limit + length;
1816 * - gap_end - gap_start >= length
1819 struct mm_struct
*mm
= current
->mm
;
1820 struct vm_area_struct
*vma
;
1821 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1823 /* Adjust search length to account for worst case alignment overhead */
1824 length
= info
->length
+ info
->align_mask
;
1825 if (length
< info
->length
)
1828 /* Adjust search limits by the desired length */
1829 if (info
->high_limit
< length
)
1831 high_limit
= info
->high_limit
- length
;
1833 if (info
->low_limit
> high_limit
)
1835 low_limit
= info
->low_limit
+ length
;
1837 /* Check if rbtree root looks promising */
1838 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1840 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1841 if (vma
->rb_subtree_gap
< length
)
1845 /* Visit left subtree if it looks promising */
1846 gap_end
= vm_start_gap(vma
);
1847 if (gap_end
>= low_limit
&& vma
->vm_rb
.rb_left
) {
1848 struct vm_area_struct
*left
=
1849 rb_entry(vma
->vm_rb
.rb_left
,
1850 struct vm_area_struct
, vm_rb
);
1851 if (left
->rb_subtree_gap
>= length
) {
1857 gap_start
= vma
->vm_prev
? vm_end_gap(vma
->vm_prev
) : 0;
1859 /* Check if current node has a suitable gap */
1860 if (gap_start
> high_limit
)
1862 if (gap_end
>= low_limit
&&
1863 gap_end
> gap_start
&& gap_end
- gap_start
>= length
)
1866 /* Visit right subtree if it looks promising */
1867 if (vma
->vm_rb
.rb_right
) {
1868 struct vm_area_struct
*right
=
1869 rb_entry(vma
->vm_rb
.rb_right
,
1870 struct vm_area_struct
, vm_rb
);
1871 if (right
->rb_subtree_gap
>= length
) {
1877 /* Go back up the rbtree to find next candidate node */
1879 struct rb_node
*prev
= &vma
->vm_rb
;
1880 if (!rb_parent(prev
))
1882 vma
= rb_entry(rb_parent(prev
),
1883 struct vm_area_struct
, vm_rb
);
1884 if (prev
== vma
->vm_rb
.rb_left
) {
1885 gap_start
= vm_end_gap(vma
->vm_prev
);
1886 gap_end
= vm_start_gap(vma
);
1893 /* Check highest gap, which does not precede any rbtree node */
1894 gap_start
= mm
->highest_vm_end
;
1895 gap_end
= ULONG_MAX
; /* Only for VM_BUG_ON below */
1896 if (gap_start
> high_limit
)
1900 /* We found a suitable gap. Clip it with the original low_limit. */
1901 if (gap_start
< info
->low_limit
)
1902 gap_start
= info
->low_limit
;
1904 /* Adjust gap address to the desired alignment */
1905 gap_start
+= (info
->align_offset
- gap_start
) & info
->align_mask
;
1907 VM_BUG_ON(gap_start
+ info
->length
> info
->high_limit
);
1908 VM_BUG_ON(gap_start
+ info
->length
> gap_end
);
1912 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info
*info
)
1914 struct mm_struct
*mm
= current
->mm
;
1915 struct vm_area_struct
*vma
;
1916 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1918 /* Adjust search length to account for worst case alignment overhead */
1919 length
= info
->length
+ info
->align_mask
;
1920 if (length
< info
->length
)
1924 * Adjust search limits by the desired length.
1925 * See implementation comment at top of unmapped_area().
1927 gap_end
= info
->high_limit
;
1928 if (gap_end
< length
)
1930 high_limit
= gap_end
- length
;
1932 if (info
->low_limit
> high_limit
)
1934 low_limit
= info
->low_limit
+ length
;
1936 /* Check highest gap, which does not precede any rbtree node */
1937 gap_start
= mm
->highest_vm_end
;
1938 if (gap_start
<= high_limit
)
1941 /* Check if rbtree root looks promising */
1942 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1944 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1945 if (vma
->rb_subtree_gap
< length
)
1949 /* Visit right subtree if it looks promising */
1950 gap_start
= vma
->vm_prev
? vm_end_gap(vma
->vm_prev
) : 0;
1951 if (gap_start
<= high_limit
&& vma
->vm_rb
.rb_right
) {
1952 struct vm_area_struct
*right
=
1953 rb_entry(vma
->vm_rb
.rb_right
,
1954 struct vm_area_struct
, vm_rb
);
1955 if (right
->rb_subtree_gap
>= length
) {
1962 /* Check if current node has a suitable gap */
1963 gap_end
= vm_start_gap(vma
);
1964 if (gap_end
< low_limit
)
1966 if (gap_start
<= high_limit
&&
1967 gap_end
> gap_start
&& gap_end
- gap_start
>= length
)
1970 /* Visit left subtree if it looks promising */
1971 if (vma
->vm_rb
.rb_left
) {
1972 struct vm_area_struct
*left
=
1973 rb_entry(vma
->vm_rb
.rb_left
,
1974 struct vm_area_struct
, vm_rb
);
1975 if (left
->rb_subtree_gap
>= length
) {
1981 /* Go back up the rbtree to find next candidate node */
1983 struct rb_node
*prev
= &vma
->vm_rb
;
1984 if (!rb_parent(prev
))
1986 vma
= rb_entry(rb_parent(prev
),
1987 struct vm_area_struct
, vm_rb
);
1988 if (prev
== vma
->vm_rb
.rb_right
) {
1989 gap_start
= vma
->vm_prev
?
1990 vm_end_gap(vma
->vm_prev
) : 0;
1997 /* We found a suitable gap. Clip it with the original high_limit. */
1998 if (gap_end
> info
->high_limit
)
1999 gap_end
= info
->high_limit
;
2002 /* Compute highest gap address at the desired alignment */
2003 gap_end
-= info
->length
;
2004 gap_end
-= (gap_end
- info
->align_offset
) & info
->align_mask
;
2006 VM_BUG_ON(gap_end
< info
->low_limit
);
2007 VM_BUG_ON(gap_end
< gap_start
);
2011 /* Get an address range which is currently unmapped.
2012 * For shmat() with addr=0.
2014 * Ugly calling convention alert:
2015 * Return value with the low bits set means error value,
2017 * if (ret & ~PAGE_MASK)
2020 * This function "knows" that -ENOMEM has the bits set.
2022 #ifndef HAVE_ARCH_UNMAPPED_AREA
2024 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
2025 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
2027 struct mm_struct
*mm
= current
->mm
;
2028 struct vm_area_struct
*vma
, *prev
;
2029 struct vm_unmapped_area_info info
;
2031 if (len
> TASK_SIZE
- mmap_min_addr
)
2034 if (flags
& MAP_FIXED
)
2038 addr
= PAGE_ALIGN(addr
);
2039 vma
= find_vma_prev(mm
, addr
, &prev
);
2040 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
2041 (!vma
|| addr
+ len
<= vm_start_gap(vma
)) &&
2042 (!prev
|| addr
>= vm_end_gap(prev
)))
2048 info
.low_limit
= mm
->mmap_base
;
2049 info
.high_limit
= TASK_SIZE
;
2050 info
.align_mask
= 0;
2051 return vm_unmapped_area(&info
);
2056 * This mmap-allocator allocates new areas top-down from below the
2057 * stack's low limit (the base):
2059 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2061 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
2062 const unsigned long len
, const unsigned long pgoff
,
2063 const unsigned long flags
)
2065 struct vm_area_struct
*vma
, *prev
;
2066 struct mm_struct
*mm
= current
->mm
;
2067 unsigned long addr
= addr0
;
2068 struct vm_unmapped_area_info info
;
2070 /* requested length too big for entire address space */
2071 if (len
> TASK_SIZE
- mmap_min_addr
)
2074 if (flags
& MAP_FIXED
)
2077 /* requesting a specific address */
2079 addr
= PAGE_ALIGN(addr
);
2080 vma
= find_vma_prev(mm
, addr
, &prev
);
2081 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
2082 (!vma
|| addr
+ len
<= vm_start_gap(vma
)) &&
2083 (!prev
|| addr
>= vm_end_gap(prev
)))
2087 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
2089 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
2090 info
.high_limit
= mm
->mmap_base
;
2091 info
.align_mask
= 0;
2092 addr
= vm_unmapped_area(&info
);
2095 * A failed mmap() very likely causes application failure,
2096 * so fall back to the bottom-up function here. This scenario
2097 * can happen with large stack limits and large mmap()
2100 if (offset_in_page(addr
)) {
2101 VM_BUG_ON(addr
!= -ENOMEM
);
2103 info
.low_limit
= TASK_UNMAPPED_BASE
;
2104 info
.high_limit
= TASK_SIZE
;
2105 addr
= vm_unmapped_area(&info
);
2113 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
2114 unsigned long pgoff
, unsigned long flags
)
2116 unsigned long (*get_area
)(struct file
*, unsigned long,
2117 unsigned long, unsigned long, unsigned long);
2119 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
2123 /* Careful about overflows.. */
2124 if (len
> TASK_SIZE
)
2127 get_area
= current
->mm
->get_unmapped_area
;
2129 if (file
->f_op
->get_unmapped_area
)
2130 get_area
= file
->f_op
->get_unmapped_area
;
2131 } else if (flags
& MAP_SHARED
) {
2133 * mmap_region() will call shmem_zero_setup() to create a file,
2134 * so use shmem's get_unmapped_area in case it can be huge.
2135 * do_mmap_pgoff() will clear pgoff, so match alignment.
2138 get_area
= shmem_get_unmapped_area
;
2141 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
2142 if (IS_ERR_VALUE(addr
))
2145 if (addr
> TASK_SIZE
- len
)
2147 if (offset_in_page(addr
))
2150 error
= security_mmap_addr(addr
);
2151 return error
? error
: addr
;
2154 EXPORT_SYMBOL(get_unmapped_area
);
2156 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2157 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
2159 struct rb_node
*rb_node
;
2160 struct vm_area_struct
*vma
;
2162 /* Check the cache first. */
2163 vma
= vmacache_find(mm
, addr
);
2167 rb_node
= mm
->mm_rb
.rb_node
;
2170 struct vm_area_struct
*tmp
;
2172 tmp
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2174 if (tmp
->vm_end
> addr
) {
2176 if (tmp
->vm_start
<= addr
)
2178 rb_node
= rb_node
->rb_left
;
2180 rb_node
= rb_node
->rb_right
;
2184 vmacache_update(addr
, vma
);
2188 EXPORT_SYMBOL(find_vma
);
2191 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2193 struct vm_area_struct
*
2194 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
2195 struct vm_area_struct
**pprev
)
2197 struct vm_area_struct
*vma
;
2199 vma
= find_vma(mm
, addr
);
2201 *pprev
= vma
->vm_prev
;
2203 struct rb_node
*rb_node
= mm
->mm_rb
.rb_node
;
2206 *pprev
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2207 rb_node
= rb_node
->rb_right
;
2214 * Verify that the stack growth is acceptable and
2215 * update accounting. This is shared with both the
2216 * grow-up and grow-down cases.
2218 static int acct_stack_growth(struct vm_area_struct
*vma
,
2219 unsigned long size
, unsigned long grow
)
2221 struct mm_struct
*mm
= vma
->vm_mm
;
2222 unsigned long new_start
;
2224 /* address space limit tests */
2225 if (!may_expand_vm(mm
, vma
->vm_flags
, grow
))
2228 /* Stack limit test */
2229 if (size
> rlimit(RLIMIT_STACK
))
2232 /* mlock limit tests */
2233 if (vma
->vm_flags
& VM_LOCKED
) {
2234 unsigned long locked
;
2235 unsigned long limit
;
2236 locked
= mm
->locked_vm
+ grow
;
2237 limit
= rlimit(RLIMIT_MEMLOCK
);
2238 limit
>>= PAGE_SHIFT
;
2239 if (locked
> limit
&& !capable(CAP_IPC_LOCK
))
2243 /* Check to ensure the stack will not grow into a hugetlb-only region */
2244 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
2246 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
2250 * Overcommit.. This must be the final test, as it will
2251 * update security statistics.
2253 if (security_vm_enough_memory_mm(mm
, grow
))
2259 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2261 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2262 * vma is the last one with address > vma->vm_end. Have to extend vma.
2264 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
)
2266 struct mm_struct
*mm
= vma
->vm_mm
;
2267 struct vm_area_struct
*next
;
2268 unsigned long gap_addr
;
2271 if (!(vma
->vm_flags
& VM_GROWSUP
))
2274 /* Guard against exceeding limits of the address space. */
2275 address
&= PAGE_MASK
;
2276 if (address
>= (TASK_SIZE
& PAGE_MASK
))
2278 address
+= PAGE_SIZE
;
2280 /* Enforce stack_guard_gap */
2281 gap_addr
= address
+ stack_guard_gap
;
2283 /* Guard against overflow */
2284 if (gap_addr
< address
|| gap_addr
> TASK_SIZE
)
2285 gap_addr
= TASK_SIZE
;
2287 next
= vma
->vm_next
;
2288 if (next
&& next
->vm_start
< gap_addr
&&
2289 (next
->vm_flags
& (VM_WRITE
|VM_READ
|VM_EXEC
))) {
2290 if (!(next
->vm_flags
& VM_GROWSUP
))
2292 /* Check that both stack segments have the same anon_vma? */
2295 /* We must make sure the anon_vma is allocated. */
2296 if (unlikely(anon_vma_prepare(vma
)))
2300 * vma->vm_start/vm_end cannot change under us because the caller
2301 * is required to hold the mmap_sem in read mode. We need the
2302 * anon_vma lock to serialize against concurrent expand_stacks.
2304 anon_vma_lock_write(vma
->anon_vma
);
2306 /* Somebody else might have raced and expanded it already */
2307 if (address
> vma
->vm_end
) {
2308 unsigned long size
, grow
;
2310 size
= address
- vma
->vm_start
;
2311 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
2314 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
2315 error
= acct_stack_growth(vma
, size
, grow
);
2318 * vma_gap_update() doesn't support concurrent
2319 * updates, but we only hold a shared mmap_sem
2320 * lock here, so we need to protect against
2321 * concurrent vma expansions.
2322 * anon_vma_lock_write() doesn't help here, as
2323 * we don't guarantee that all growable vmas
2324 * in a mm share the same root anon vma.
2325 * So, we reuse mm->page_table_lock to guard
2326 * against concurrent vma expansions.
2328 spin_lock(&mm
->page_table_lock
);
2329 if (vma
->vm_flags
& VM_LOCKED
)
2330 mm
->locked_vm
+= grow
;
2331 vm_stat_account(mm
, vma
->vm_flags
, grow
);
2332 anon_vma_interval_tree_pre_update_vma(vma
);
2333 vma
->vm_end
= address
;
2334 anon_vma_interval_tree_post_update_vma(vma
);
2336 vma_gap_update(vma
->vm_next
);
2338 mm
->highest_vm_end
= vm_end_gap(vma
);
2339 spin_unlock(&mm
->page_table_lock
);
2341 perf_event_mmap(vma
);
2345 anon_vma_unlock_write(vma
->anon_vma
);
2346 khugepaged_enter_vma_merge(vma
, vma
->vm_flags
);
2350 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2353 * vma is the first one with address < vma->vm_start. Have to extend vma.
2355 int expand_downwards(struct vm_area_struct
*vma
,
2356 unsigned long address
)
2358 struct mm_struct
*mm
= vma
->vm_mm
;
2359 struct vm_area_struct
*prev
;
2362 address
&= PAGE_MASK
;
2363 error
= security_mmap_addr(address
);
2367 /* Enforce stack_guard_gap */
2368 prev
= vma
->vm_prev
;
2369 /* Check that both stack segments have the same anon_vma? */
2370 if (prev
&& !(prev
->vm_flags
& VM_GROWSDOWN
) &&
2371 (prev
->vm_flags
& (VM_WRITE
|VM_READ
|VM_EXEC
))) {
2372 if (address
- prev
->vm_end
< stack_guard_gap
)
2376 /* We must make sure the anon_vma is allocated. */
2377 if (unlikely(anon_vma_prepare(vma
)))
2381 * vma->vm_start/vm_end cannot change under us because the caller
2382 * is required to hold the mmap_sem in read mode. We need the
2383 * anon_vma lock to serialize against concurrent expand_stacks.
2385 anon_vma_lock_write(vma
->anon_vma
);
2387 /* Somebody else might have raced and expanded it already */
2388 if (address
< vma
->vm_start
) {
2389 unsigned long size
, grow
;
2391 size
= vma
->vm_end
- address
;
2392 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
2395 if (grow
<= vma
->vm_pgoff
) {
2396 error
= acct_stack_growth(vma
, size
, grow
);
2399 * vma_gap_update() doesn't support concurrent
2400 * updates, but we only hold a shared mmap_sem
2401 * lock here, so we need to protect against
2402 * concurrent vma expansions.
2403 * anon_vma_lock_write() doesn't help here, as
2404 * we don't guarantee that all growable vmas
2405 * in a mm share the same root anon vma.
2406 * So, we reuse mm->page_table_lock to guard
2407 * against concurrent vma expansions.
2409 spin_lock(&mm
->page_table_lock
);
2410 if (vma
->vm_flags
& VM_LOCKED
)
2411 mm
->locked_vm
+= grow
;
2412 vm_stat_account(mm
, vma
->vm_flags
, grow
);
2413 anon_vma_interval_tree_pre_update_vma(vma
);
2414 vma
->vm_start
= address
;
2415 vma
->vm_pgoff
-= grow
;
2416 anon_vma_interval_tree_post_update_vma(vma
);
2417 vma_gap_update(vma
);
2418 spin_unlock(&mm
->page_table_lock
);
2420 perf_event_mmap(vma
);
2424 anon_vma_unlock_write(vma
->anon_vma
);
2425 khugepaged_enter_vma_merge(vma
, vma
->vm_flags
);
2430 /* enforced gap between the expanding stack and other mappings. */
2431 unsigned long stack_guard_gap
= 256UL<<PAGE_SHIFT
;
2433 static int __init
cmdline_parse_stack_guard_gap(char *p
)
2438 val
= simple_strtoul(p
, &endptr
, 10);
2440 stack_guard_gap
= val
<< PAGE_SHIFT
;
2444 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap
);
2446 #ifdef CONFIG_STACK_GROWSUP
2447 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2449 return expand_upwards(vma
, address
);
2452 struct vm_area_struct
*
2453 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2455 struct vm_area_struct
*vma
, *prev
;
2458 vma
= find_vma_prev(mm
, addr
, &prev
);
2459 if (vma
&& (vma
->vm_start
<= addr
))
2461 if (!prev
|| expand_stack(prev
, addr
))
2463 if (prev
->vm_flags
& VM_LOCKED
)
2464 populate_vma_page_range(prev
, addr
, prev
->vm_end
, NULL
);
2468 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2470 return expand_downwards(vma
, address
);
2473 struct vm_area_struct
*
2474 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2476 struct vm_area_struct
*vma
;
2477 unsigned long start
;
2480 vma
= find_vma(mm
, addr
);
2483 if (vma
->vm_start
<= addr
)
2485 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2487 start
= vma
->vm_start
;
2488 if (expand_stack(vma
, addr
))
2490 if (vma
->vm_flags
& VM_LOCKED
)
2491 populate_vma_page_range(vma
, addr
, start
, NULL
);
2496 EXPORT_SYMBOL_GPL(find_extend_vma
);
2499 * Ok - we have the memory areas we should free on the vma list,
2500 * so release them, and do the vma updates.
2502 * Called with the mm semaphore held.
2504 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2506 unsigned long nr_accounted
= 0;
2508 /* Update high watermark before we lower total_vm */
2509 update_hiwater_vm(mm
);
2511 long nrpages
= vma_pages(vma
);
2513 if (vma
->vm_flags
& VM_ACCOUNT
)
2514 nr_accounted
+= nrpages
;
2515 vm_stat_account(mm
, vma
->vm_flags
, -nrpages
);
2516 vma
= remove_vma(vma
);
2518 vm_unacct_memory(nr_accounted
);
2523 * Get rid of page table information in the indicated region.
2525 * Called with the mm semaphore held.
2527 static void unmap_region(struct mm_struct
*mm
,
2528 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
2529 unsigned long start
, unsigned long end
)
2531 struct vm_area_struct
*next
= prev
? prev
->vm_next
: mm
->mmap
;
2532 struct mmu_gather tlb
;
2535 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2536 update_hiwater_rss(mm
);
2537 unmap_vmas(&tlb
, vma
, start
, end
);
2538 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
2539 next
? next
->vm_start
: USER_PGTABLES_CEILING
);
2540 tlb_finish_mmu(&tlb
, start
, end
);
2544 * Create a list of vma's touched by the unmap, removing them from the mm's
2545 * vma list as we go..
2548 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2549 struct vm_area_struct
*prev
, unsigned long end
)
2551 struct vm_area_struct
**insertion_point
;
2552 struct vm_area_struct
*tail_vma
= NULL
;
2554 insertion_point
= (prev
? &prev
->vm_next
: &mm
->mmap
);
2555 vma
->vm_prev
= NULL
;
2557 vma_rb_erase(vma
, &mm
->mm_rb
);
2561 } while (vma
&& vma
->vm_start
< end
);
2562 *insertion_point
= vma
;
2564 vma
->vm_prev
= prev
;
2565 vma_gap_update(vma
);
2567 mm
->highest_vm_end
= prev
? vm_end_gap(prev
) : 0;
2568 tail_vma
->vm_next
= NULL
;
2570 /* Kill the cache */
2571 vmacache_invalidate(mm
);
2575 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2576 * has already been checked or doesn't make sense to fail.
2578 int __split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2579 unsigned long addr
, int new_below
)
2581 struct vm_area_struct
*new;
2584 if (vma
->vm_ops
&& vma
->vm_ops
->split
) {
2585 err
= vma
->vm_ops
->split(vma
, addr
);
2590 new = kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2594 /* most fields are the same, copy all, and then fixup */
2597 INIT_LIST_HEAD(&new->anon_vma_chain
);
2602 new->vm_start
= addr
;
2603 new->vm_pgoff
+= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
);
2606 err
= vma_dup_policy(vma
, new);
2610 err
= anon_vma_clone(new, vma
);
2615 get_file(new->vm_file
);
2617 if (new->vm_ops
&& new->vm_ops
->open
)
2618 new->vm_ops
->open(new);
2621 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2622 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2624 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2630 /* Clean everything up if vma_adjust failed. */
2631 if (new->vm_ops
&& new->vm_ops
->close
)
2632 new->vm_ops
->close(new);
2635 unlink_anon_vmas(new);
2637 mpol_put(vma_policy(new));
2639 kmem_cache_free(vm_area_cachep
, new);
2644 * Split a vma into two pieces at address 'addr', a new vma is allocated
2645 * either for the first part or the tail.
2647 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2648 unsigned long addr
, int new_below
)
2650 if (mm
->map_count
>= sysctl_max_map_count
)
2653 return __split_vma(mm
, vma
, addr
, new_below
);
2656 /* Munmap is split into 2 main parts -- this part which finds
2657 * what needs doing, and the areas themselves, which do the
2658 * work. This now handles partial unmappings.
2659 * Jeremy Fitzhardinge <jeremy@goop.org>
2661 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
,
2662 struct list_head
*uf
)
2665 struct vm_area_struct
*vma
, *prev
, *last
;
2667 if ((offset_in_page(start
)) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2670 len
= PAGE_ALIGN(len
);
2674 /* Find the first overlapping VMA */
2675 vma
= find_vma(mm
, start
);
2678 prev
= vma
->vm_prev
;
2679 /* we have start < vma->vm_end */
2681 /* if it doesn't overlap, we have nothing.. */
2683 if (vma
->vm_start
>= end
)
2687 * If we need to split any vma, do it now to save pain later.
2689 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2690 * unmapped vm_area_struct will remain in use: so lower split_vma
2691 * places tmp vma above, and higher split_vma places tmp vma below.
2693 if (start
> vma
->vm_start
) {
2697 * Make sure that map_count on return from munmap() will
2698 * not exceed its limit; but let map_count go just above
2699 * its limit temporarily, to help free resources as expected.
2701 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2704 error
= __split_vma(mm
, vma
, start
, 0);
2710 /* Does it split the last one? */
2711 last
= find_vma(mm
, end
);
2712 if (last
&& end
> last
->vm_start
) {
2713 int error
= __split_vma(mm
, last
, end
, 1);
2717 vma
= prev
? prev
->vm_next
: mm
->mmap
;
2721 * If userfaultfd_unmap_prep returns an error the vmas
2722 * will remain splitted, but userland will get a
2723 * highly unexpected error anyway. This is no
2724 * different than the case where the first of the two
2725 * __split_vma fails, but we don't undo the first
2726 * split, despite we could. This is unlikely enough
2727 * failure that it's not worth optimizing it for.
2729 int error
= userfaultfd_unmap_prep(vma
, start
, end
, uf
);
2735 * unlock any mlock()ed ranges before detaching vmas
2737 if (mm
->locked_vm
) {
2738 struct vm_area_struct
*tmp
= vma
;
2739 while (tmp
&& tmp
->vm_start
< end
) {
2740 if (tmp
->vm_flags
& VM_LOCKED
) {
2741 mm
->locked_vm
-= vma_pages(tmp
);
2742 munlock_vma_pages_all(tmp
);
2749 * Remove the vma's, and unmap the actual pages
2751 detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
);
2752 unmap_region(mm
, vma
, prev
, start
, end
);
2754 arch_unmap(mm
, vma
, start
, end
);
2756 /* Fix up all other VM information */
2757 remove_vma_list(mm
, vma
);
2761 EXPORT_SYMBOL(do_munmap
);
2763 int vm_munmap(unsigned long start
, size_t len
)
2766 struct mm_struct
*mm
= current
->mm
;
2769 if (down_write_killable(&mm
->mmap_sem
))
2772 ret
= do_munmap(mm
, start
, len
, &uf
);
2773 up_write(&mm
->mmap_sem
);
2774 userfaultfd_unmap_complete(mm
, &uf
);
2777 EXPORT_SYMBOL(vm_munmap
);
2779 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2781 profile_munmap(addr
);
2782 return vm_munmap(addr
, len
);
2787 * Emulation of deprecated remap_file_pages() syscall.
2789 SYSCALL_DEFINE5(remap_file_pages
, unsigned long, start
, unsigned long, size
,
2790 unsigned long, prot
, unsigned long, pgoff
, unsigned long, flags
)
2793 struct mm_struct
*mm
= current
->mm
;
2794 struct vm_area_struct
*vma
;
2795 unsigned long populate
= 0;
2796 unsigned long ret
= -EINVAL
;
2799 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2800 current
->comm
, current
->pid
);
2804 start
= start
& PAGE_MASK
;
2805 size
= size
& PAGE_MASK
;
2807 if (start
+ size
<= start
)
2810 /* Does pgoff wrap? */
2811 if (pgoff
+ (size
>> PAGE_SHIFT
) < pgoff
)
2814 if (down_write_killable(&mm
->mmap_sem
))
2817 vma
= find_vma(mm
, start
);
2819 if (!vma
|| !(vma
->vm_flags
& VM_SHARED
))
2822 if (start
< vma
->vm_start
)
2825 if (start
+ size
> vma
->vm_end
) {
2826 struct vm_area_struct
*next
;
2828 for (next
= vma
->vm_next
; next
; next
= next
->vm_next
) {
2829 /* hole between vmas ? */
2830 if (next
->vm_start
!= next
->vm_prev
->vm_end
)
2833 if (next
->vm_file
!= vma
->vm_file
)
2836 if (next
->vm_flags
!= vma
->vm_flags
)
2839 if (start
+ size
<= next
->vm_end
)
2847 prot
|= vma
->vm_flags
& VM_READ
? PROT_READ
: 0;
2848 prot
|= vma
->vm_flags
& VM_WRITE
? PROT_WRITE
: 0;
2849 prot
|= vma
->vm_flags
& VM_EXEC
? PROT_EXEC
: 0;
2851 flags
&= MAP_NONBLOCK
;
2852 flags
|= MAP_SHARED
| MAP_FIXED
| MAP_POPULATE
;
2853 if (vma
->vm_flags
& VM_LOCKED
) {
2854 struct vm_area_struct
*tmp
;
2855 flags
|= MAP_LOCKED
;
2857 /* drop PG_Mlocked flag for over-mapped range */
2858 for (tmp
= vma
; tmp
->vm_start
>= start
+ size
;
2859 tmp
= tmp
->vm_next
) {
2861 * Split pmd and munlock page on the border
2864 vma_adjust_trans_huge(tmp
, start
, start
+ size
, 0);
2866 munlock_vma_pages_range(tmp
,
2867 max(tmp
->vm_start
, start
),
2868 min(tmp
->vm_end
, start
+ size
));
2872 file
= get_file(vma
->vm_file
);
2873 ret
= do_mmap_pgoff(vma
->vm_file
, start
, size
,
2874 prot
, flags
, pgoff
, &populate
, NULL
);
2877 up_write(&mm
->mmap_sem
);
2879 mm_populate(ret
, populate
);
2880 if (!IS_ERR_VALUE(ret
))
2885 static inline void verify_mm_writelocked(struct mm_struct
*mm
)
2887 #ifdef CONFIG_DEBUG_VM
2888 if (unlikely(down_read_trylock(&mm
->mmap_sem
))) {
2890 up_read(&mm
->mmap_sem
);
2896 * this is really a simplified "do_mmap". it only handles
2897 * anonymous maps. eventually we may be able to do some
2898 * brk-specific accounting here.
2900 static int do_brk_flags(unsigned long addr
, unsigned long len
, unsigned long flags
, struct list_head
*uf
)
2902 struct mm_struct
*mm
= current
->mm
;
2903 struct vm_area_struct
*vma
, *prev
;
2904 struct rb_node
**rb_link
, *rb_parent
;
2905 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2908 /* Until we need other flags, refuse anything except VM_EXEC. */
2909 if ((flags
& (~VM_EXEC
)) != 0)
2911 flags
|= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2913 error
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2914 if (offset_in_page(error
))
2917 error
= mlock_future_check(mm
, mm
->def_flags
, len
);
2922 * mm->mmap_sem is required to protect against another thread
2923 * changing the mappings in case we sleep.
2925 verify_mm_writelocked(mm
);
2928 * Clear old maps. this also does some error checking for us
2930 while (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
,
2932 if (do_munmap(mm
, addr
, len
, uf
))
2936 /* Check against address space limits *after* clearing old maps... */
2937 if (!may_expand_vm(mm
, flags
, len
>> PAGE_SHIFT
))
2940 if (mm
->map_count
> sysctl_max_map_count
)
2943 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
2946 /* Can we just expand an old private anonymous mapping? */
2947 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
2948 NULL
, NULL
, pgoff
, NULL
, NULL_VM_UFFD_CTX
, NULL
);
2953 * create a vma struct for an anonymous mapping
2955 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2957 vm_unacct_memory(len
>> PAGE_SHIFT
);
2961 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2963 vma
->vm_start
= addr
;
2964 vma
->vm_end
= addr
+ len
;
2965 vma
->vm_pgoff
= pgoff
;
2966 vma
->vm_flags
= flags
;
2967 vma
->vm_page_prot
= vm_get_page_prot(flags
);
2968 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2970 perf_event_mmap(vma
);
2971 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2972 mm
->data_vm
+= len
>> PAGE_SHIFT
;
2973 if (flags
& VM_LOCKED
)
2974 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
2975 vma
->vm_flags
|= VM_SOFTDIRTY
;
2979 int vm_brk_flags(unsigned long addr
, unsigned long request
, unsigned long flags
)
2981 struct mm_struct
*mm
= current
->mm
;
2987 len
= PAGE_ALIGN(request
);
2993 if (down_write_killable(&mm
->mmap_sem
))
2996 ret
= do_brk_flags(addr
, len
, flags
, &uf
);
2997 populate
= ((mm
->def_flags
& VM_LOCKED
) != 0);
2998 up_write(&mm
->mmap_sem
);
2999 userfaultfd_unmap_complete(mm
, &uf
);
3000 if (populate
&& !ret
)
3001 mm_populate(addr
, len
);
3004 EXPORT_SYMBOL(vm_brk_flags
);
3006 int vm_brk(unsigned long addr
, unsigned long len
)
3008 return vm_brk_flags(addr
, len
, 0);
3010 EXPORT_SYMBOL(vm_brk
);
3012 /* Release all mmaps. */
3013 void exit_mmap(struct mm_struct
*mm
)
3015 struct mmu_gather tlb
;
3016 struct vm_area_struct
*vma
;
3017 unsigned long nr_accounted
= 0;
3019 /* mm's last user has gone, and its about to be pulled down */
3020 mmu_notifier_release(mm
);
3022 if (unlikely(mm_is_oom_victim(mm
))) {
3024 * Manually reap the mm to free as much memory as possible.
3025 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3026 * this mm from further consideration. Taking mm->mmap_sem for
3027 * write after setting MMF_OOM_SKIP will guarantee that the oom
3028 * reaper will not run on this mm again after mmap_sem is
3031 * Nothing can be holding mm->mmap_sem here and the above call
3032 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3033 * __oom_reap_task_mm() will not block.
3035 * This needs to be done before calling munlock_vma_pages_all(),
3036 * which clears VM_LOCKED, otherwise the oom reaper cannot
3039 mutex_lock(&oom_lock
);
3040 __oom_reap_task_mm(mm
);
3041 mutex_unlock(&oom_lock
);
3043 set_bit(MMF_OOM_SKIP
, &mm
->flags
);
3044 down_write(&mm
->mmap_sem
);
3045 up_write(&mm
->mmap_sem
);
3048 if (mm
->locked_vm
) {
3051 if (vma
->vm_flags
& VM_LOCKED
)
3052 munlock_vma_pages_all(vma
);
3060 if (!vma
) /* Can happen if dup_mmap() received an OOM */
3065 tlb_gather_mmu(&tlb
, mm
, 0, -1);
3066 /* update_hiwater_rss(mm) here? but nobody should be looking */
3067 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3068 unmap_vmas(&tlb
, vma
, 0, -1);
3069 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, USER_PGTABLES_CEILING
);
3070 tlb_finish_mmu(&tlb
, 0, -1);
3073 * Walk the list again, actually closing and freeing it,
3074 * with preemption enabled, without holding any MM locks.
3077 if (vma
->vm_flags
& VM_ACCOUNT
)
3078 nr_accounted
+= vma_pages(vma
);
3079 vma
= remove_vma(vma
);
3081 vm_unacct_memory(nr_accounted
);
3084 /* Insert vm structure into process list sorted by address
3085 * and into the inode's i_mmap tree. If vm_file is non-NULL
3086 * then i_mmap_rwsem is taken here.
3088 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
3090 struct vm_area_struct
*prev
;
3091 struct rb_node
**rb_link
, *rb_parent
;
3093 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
3094 &prev
, &rb_link
, &rb_parent
))
3096 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
3097 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
3101 * The vm_pgoff of a purely anonymous vma should be irrelevant
3102 * until its first write fault, when page's anon_vma and index
3103 * are set. But now set the vm_pgoff it will almost certainly
3104 * end up with (unless mremap moves it elsewhere before that
3105 * first wfault), so /proc/pid/maps tells a consistent story.
3107 * By setting it to reflect the virtual start address of the
3108 * vma, merges and splits can happen in a seamless way, just
3109 * using the existing file pgoff checks and manipulations.
3110 * Similarly in do_mmap_pgoff and in do_brk.
3112 if (vma_is_anonymous(vma
)) {
3113 BUG_ON(vma
->anon_vma
);
3114 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
3117 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
3122 * Copy the vma structure to a new location in the same mm,
3123 * prior to moving page table entries, to effect an mremap move.
3125 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
3126 unsigned long addr
, unsigned long len
, pgoff_t pgoff
,
3127 bool *need_rmap_locks
)
3129 struct vm_area_struct
*vma
= *vmap
;
3130 unsigned long vma_start
= vma
->vm_start
;
3131 struct mm_struct
*mm
= vma
->vm_mm
;
3132 struct vm_area_struct
*new_vma
, *prev
;
3133 struct rb_node
**rb_link
, *rb_parent
;
3134 bool faulted_in_anon_vma
= true;
3137 * If anonymous vma has not yet been faulted, update new pgoff
3138 * to match new location, to increase its chance of merging.
3140 if (unlikely(vma_is_anonymous(vma
) && !vma
->anon_vma
)) {
3141 pgoff
= addr
>> PAGE_SHIFT
;
3142 faulted_in_anon_vma
= false;
3145 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
3146 return NULL
; /* should never get here */
3147 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
3148 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
),
3149 vma
->vm_userfaultfd_ctx
, vma_get_anon_name(vma
));
3152 * Source vma may have been merged into new_vma
3154 if (unlikely(vma_start
>= new_vma
->vm_start
&&
3155 vma_start
< new_vma
->vm_end
)) {
3157 * The only way we can get a vma_merge with
3158 * self during an mremap is if the vma hasn't
3159 * been faulted in yet and we were allowed to
3160 * reset the dst vma->vm_pgoff to the
3161 * destination address of the mremap to allow
3162 * the merge to happen. mremap must change the
3163 * vm_pgoff linearity between src and dst vmas
3164 * (in turn preventing a vma_merge) to be
3165 * safe. It is only safe to keep the vm_pgoff
3166 * linear if there are no pages mapped yet.
3168 VM_BUG_ON_VMA(faulted_in_anon_vma
, new_vma
);
3169 *vmap
= vma
= new_vma
;
3171 *need_rmap_locks
= (new_vma
->vm_pgoff
<= vma
->vm_pgoff
);
3173 new_vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
3177 new_vma
->vm_start
= addr
;
3178 new_vma
->vm_end
= addr
+ len
;
3179 new_vma
->vm_pgoff
= pgoff
;
3180 if (vma_dup_policy(vma
, new_vma
))
3182 INIT_LIST_HEAD(&new_vma
->anon_vma_chain
);
3183 if (anon_vma_clone(new_vma
, vma
))
3184 goto out_free_mempol
;
3185 if (new_vma
->vm_file
)
3186 get_file(new_vma
->vm_file
);
3187 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
3188 new_vma
->vm_ops
->open(new_vma
);
3189 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
3190 *need_rmap_locks
= false;
3195 mpol_put(vma_policy(new_vma
));
3197 kmem_cache_free(vm_area_cachep
, new_vma
);
3203 * Return true if the calling process may expand its vm space by the passed
3206 bool may_expand_vm(struct mm_struct
*mm
, vm_flags_t flags
, unsigned long npages
)
3208 if (mm
->total_vm
+ npages
> rlimit(RLIMIT_AS
) >> PAGE_SHIFT
)
3211 if (is_data_mapping(flags
) &&
3212 mm
->data_vm
+ npages
> rlimit(RLIMIT_DATA
) >> PAGE_SHIFT
) {
3213 /* Workaround for Valgrind */
3214 if (rlimit(RLIMIT_DATA
) == 0 &&
3215 mm
->data_vm
+ npages
<= rlimit_max(RLIMIT_DATA
) >> PAGE_SHIFT
)
3217 if (!ignore_rlimit_data
) {
3218 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits or use boot option ignore_rlimit_data.\n",
3219 current
->comm
, current
->pid
,
3220 (mm
->data_vm
+ npages
) << PAGE_SHIFT
,
3221 rlimit(RLIMIT_DATA
));
3229 void vm_stat_account(struct mm_struct
*mm
, vm_flags_t flags
, long npages
)
3231 mm
->total_vm
+= npages
;
3233 if (is_exec_mapping(flags
))
3234 mm
->exec_vm
+= npages
;
3235 else if (is_stack_mapping(flags
))
3236 mm
->stack_vm
+= npages
;
3237 else if (is_data_mapping(flags
))
3238 mm
->data_vm
+= npages
;
3241 static int special_mapping_fault(struct vm_fault
*vmf
);
3244 * Having a close hook prevents vma merging regardless of flags.
3246 static void special_mapping_close(struct vm_area_struct
*vma
)
3250 static const char *special_mapping_name(struct vm_area_struct
*vma
)
3252 return ((struct vm_special_mapping
*)vma
->vm_private_data
)->name
;
3255 static int special_mapping_mremap(struct vm_area_struct
*new_vma
)
3257 struct vm_special_mapping
*sm
= new_vma
->vm_private_data
;
3259 if (WARN_ON_ONCE(current
->mm
!= new_vma
->vm_mm
))
3263 return sm
->mremap(sm
, new_vma
);
3268 static const struct vm_operations_struct special_mapping_vmops
= {
3269 .close
= special_mapping_close
,
3270 .fault
= special_mapping_fault
,
3271 .mremap
= special_mapping_mremap
,
3272 .name
= special_mapping_name
,
3275 static const struct vm_operations_struct legacy_special_mapping_vmops
= {
3276 .close
= special_mapping_close
,
3277 .fault
= special_mapping_fault
,
3280 static int special_mapping_fault(struct vm_fault
*vmf
)
3282 struct vm_area_struct
*vma
= vmf
->vma
;
3284 struct page
**pages
;
3286 if (vma
->vm_ops
== &legacy_special_mapping_vmops
) {
3287 pages
= vma
->vm_private_data
;
3289 struct vm_special_mapping
*sm
= vma
->vm_private_data
;
3292 return sm
->fault(sm
, vmf
->vma
, vmf
);
3297 for (pgoff
= vmf
->pgoff
; pgoff
&& *pages
; ++pages
)
3301 struct page
*page
= *pages
;
3307 return VM_FAULT_SIGBUS
;
3310 static struct vm_area_struct
*__install_special_mapping(
3311 struct mm_struct
*mm
,
3312 unsigned long addr
, unsigned long len
,
3313 unsigned long vm_flags
, void *priv
,
3314 const struct vm_operations_struct
*ops
)
3317 struct vm_area_struct
*vma
;
3319 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
3320 if (unlikely(vma
== NULL
))
3321 return ERR_PTR(-ENOMEM
);
3323 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
3325 vma
->vm_start
= addr
;
3326 vma
->vm_end
= addr
+ len
;
3328 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
| VM_SOFTDIRTY
;
3329 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
3332 vma
->vm_private_data
= priv
;
3334 ret
= insert_vm_struct(mm
, vma
);
3338 vm_stat_account(mm
, vma
->vm_flags
, len
>> PAGE_SHIFT
);
3340 perf_event_mmap(vma
);
3345 kmem_cache_free(vm_area_cachep
, vma
);
3346 return ERR_PTR(ret
);
3349 bool vma_is_special_mapping(const struct vm_area_struct
*vma
,
3350 const struct vm_special_mapping
*sm
)
3352 return vma
->vm_private_data
== sm
&&
3353 (vma
->vm_ops
== &special_mapping_vmops
||
3354 vma
->vm_ops
== &legacy_special_mapping_vmops
);
3358 * Called with mm->mmap_sem held for writing.
3359 * Insert a new vma covering the given region, with the given flags.
3360 * Its pages are supplied by the given array of struct page *.
3361 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3362 * The region past the last page supplied will always produce SIGBUS.
3363 * The array pointer and the pages it points to are assumed to stay alive
3364 * for as long as this mapping might exist.
3366 struct vm_area_struct
*_install_special_mapping(
3367 struct mm_struct
*mm
,
3368 unsigned long addr
, unsigned long len
,
3369 unsigned long vm_flags
, const struct vm_special_mapping
*spec
)
3371 return __install_special_mapping(mm
, addr
, len
, vm_flags
, (void *)spec
,
3372 &special_mapping_vmops
);
3375 int install_special_mapping(struct mm_struct
*mm
,
3376 unsigned long addr
, unsigned long len
,
3377 unsigned long vm_flags
, struct page
**pages
)
3379 struct vm_area_struct
*vma
= __install_special_mapping(
3380 mm
, addr
, len
, vm_flags
, (void *)pages
,
3381 &legacy_special_mapping_vmops
);
3383 return PTR_ERR_OR_ZERO(vma
);
3386 static DEFINE_MUTEX(mm_all_locks_mutex
);
3388 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
3390 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_root
.rb_node
)) {
3392 * The LSB of head.next can't change from under us
3393 * because we hold the mm_all_locks_mutex.
3395 down_write_nest_lock(&anon_vma
->root
->rwsem
, &mm
->mmap_sem
);
3397 * We can safely modify head.next after taking the
3398 * anon_vma->root->rwsem. If some other vma in this mm shares
3399 * the same anon_vma we won't take it again.
3401 * No need of atomic instructions here, head.next
3402 * can't change from under us thanks to the
3403 * anon_vma->root->rwsem.
3405 if (__test_and_set_bit(0, (unsigned long *)
3406 &anon_vma
->root
->rb_root
.rb_root
.rb_node
))
3411 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
3413 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3415 * AS_MM_ALL_LOCKS can't change from under us because
3416 * we hold the mm_all_locks_mutex.
3418 * Operations on ->flags have to be atomic because
3419 * even if AS_MM_ALL_LOCKS is stable thanks to the
3420 * mm_all_locks_mutex, there may be other cpus
3421 * changing other bitflags in parallel to us.
3423 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
3425 down_write_nest_lock(&mapping
->i_mmap_rwsem
, &mm
->mmap_sem
);
3430 * This operation locks against the VM for all pte/vma/mm related
3431 * operations that could ever happen on a certain mm. This includes
3432 * vmtruncate, try_to_unmap, and all page faults.
3434 * The caller must take the mmap_sem in write mode before calling
3435 * mm_take_all_locks(). The caller isn't allowed to release the
3436 * mmap_sem until mm_drop_all_locks() returns.
3438 * mmap_sem in write mode is required in order to block all operations
3439 * that could modify pagetables and free pages without need of
3440 * altering the vma layout. It's also needed in write mode to avoid new
3441 * anon_vmas to be associated with existing vmas.
3443 * A single task can't take more than one mm_take_all_locks() in a row
3444 * or it would deadlock.
3446 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3447 * mapping->flags avoid to take the same lock twice, if more than one
3448 * vma in this mm is backed by the same anon_vma or address_space.
3450 * We take locks in following order, accordingly to comment at beginning
3452 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3454 * - all i_mmap_rwsem locks;
3455 * - all anon_vma->rwseml
3457 * We can take all locks within these types randomly because the VM code
3458 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3459 * mm_all_locks_mutex.
3461 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3462 * that may have to take thousand of locks.
3464 * mm_take_all_locks() can fail if it's interrupted by signals.
3466 int mm_take_all_locks(struct mm_struct
*mm
)
3468 struct vm_area_struct
*vma
;
3469 struct anon_vma_chain
*avc
;
3471 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3473 mutex_lock(&mm_all_locks_mutex
);
3475 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3476 if (signal_pending(current
))
3478 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
3479 is_vm_hugetlb_page(vma
))
3480 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3483 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3484 if (signal_pending(current
))
3486 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
3487 !is_vm_hugetlb_page(vma
))
3488 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3491 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3492 if (signal_pending(current
))
3495 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3496 vm_lock_anon_vma(mm
, avc
->anon_vma
);
3502 mm_drop_all_locks(mm
);
3506 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
3508 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_root
.rb_node
)) {
3510 * The LSB of head.next can't change to 0 from under
3511 * us because we hold the mm_all_locks_mutex.
3513 * We must however clear the bitflag before unlocking
3514 * the vma so the users using the anon_vma->rb_root will
3515 * never see our bitflag.
3517 * No need of atomic instructions here, head.next
3518 * can't change from under us until we release the
3519 * anon_vma->root->rwsem.
3521 if (!__test_and_clear_bit(0, (unsigned long *)
3522 &anon_vma
->root
->rb_root
.rb_root
.rb_node
))
3524 anon_vma_unlock_write(anon_vma
);
3528 static void vm_unlock_mapping(struct address_space
*mapping
)
3530 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3532 * AS_MM_ALL_LOCKS can't change to 0 from under us
3533 * because we hold the mm_all_locks_mutex.
3535 i_mmap_unlock_write(mapping
);
3536 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
3543 * The mmap_sem cannot be released by the caller until
3544 * mm_drop_all_locks() returns.
3546 void mm_drop_all_locks(struct mm_struct
*mm
)
3548 struct vm_area_struct
*vma
;
3549 struct anon_vma_chain
*avc
;
3551 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3552 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
3554 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3556 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3557 vm_unlock_anon_vma(avc
->anon_vma
);
3558 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3559 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
3562 mutex_unlock(&mm_all_locks_mutex
);
3566 * initialise the percpu counter for VM
3568 void __init
mmap_init(void)
3572 ret
= percpu_counter_init(&vm_committed_as
, 0, GFP_KERNEL
);
3577 * Initialise sysctl_user_reserve_kbytes.
3579 * This is intended to prevent a user from starting a single memory hogging
3580 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3583 * The default value is min(3% of free memory, 128MB)
3584 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3586 static int init_user_reserve(void)
3588 unsigned long free_kbytes
;
3590 free_kbytes
= global_zone_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3592 sysctl_user_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 17);
3595 subsys_initcall(init_user_reserve
);
3598 * Initialise sysctl_admin_reserve_kbytes.
3600 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3601 * to log in and kill a memory hogging process.
3603 * Systems with more than 256MB will reserve 8MB, enough to recover
3604 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3605 * only reserve 3% of free pages by default.
3607 static int init_admin_reserve(void)
3609 unsigned long free_kbytes
;
3611 free_kbytes
= global_zone_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3613 sysctl_admin_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 13);
3616 subsys_initcall(init_admin_reserve
);
3619 * Reinititalise user and admin reserves if memory is added or removed.
3621 * The default user reserve max is 128MB, and the default max for the
3622 * admin reserve is 8MB. These are usually, but not always, enough to
3623 * enable recovery from a memory hogging process using login/sshd, a shell,
3624 * and tools like top. It may make sense to increase or even disable the
3625 * reserve depending on the existence of swap or variations in the recovery
3626 * tools. So, the admin may have changed them.
3628 * If memory is added and the reserves have been eliminated or increased above
3629 * the default max, then we'll trust the admin.
3631 * If memory is removed and there isn't enough free memory, then we
3632 * need to reset the reserves.
3634 * Otherwise keep the reserve set by the admin.
3636 static int reserve_mem_notifier(struct notifier_block
*nb
,
3637 unsigned long action
, void *data
)
3639 unsigned long tmp
, free_kbytes
;
3643 /* Default max is 128MB. Leave alone if modified by operator. */
3644 tmp
= sysctl_user_reserve_kbytes
;
3645 if (0 < tmp
&& tmp
< (1UL << 17))
3646 init_user_reserve();
3648 /* Default max is 8MB. Leave alone if modified by operator. */
3649 tmp
= sysctl_admin_reserve_kbytes
;
3650 if (0 < tmp
&& tmp
< (1UL << 13))
3651 init_admin_reserve();
3655 free_kbytes
= global_zone_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3657 if (sysctl_user_reserve_kbytes
> free_kbytes
) {
3658 init_user_reserve();
3659 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3660 sysctl_user_reserve_kbytes
);
3663 if (sysctl_admin_reserve_kbytes
> free_kbytes
) {
3664 init_admin_reserve();
3665 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3666 sysctl_admin_reserve_kbytes
);
3675 static struct notifier_block reserve_mem_nb
= {
3676 .notifier_call
= reserve_mem_notifier
,
3679 static int __meminit
init_reserve_notifier(void)
3681 if (register_hotmemory_notifier(&reserve_mem_nb
))
3682 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3686 subsys_initcall(init_reserve_notifier
);