6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #include <linux/kernel.h>
10 #include <linux/slab.h>
11 #include <linux/backing-dev.h>
13 #include <linux/vmacache.h>
14 #include <linux/shm.h>
15 #include <linux/mman.h>
16 #include <linux/pagemap.h>
17 #include <linux/swap.h>
18 #include <linux/syscalls.h>
19 #include <linux/capability.h>
20 #include <linux/init.h>
21 #include <linux/file.h>
23 #include <linux/personality.h>
24 #include <linux/security.h>
25 #include <linux/hugetlb.h>
26 #include <linux/profile.h>
27 #include <linux/export.h>
28 #include <linux/mount.h>
29 #include <linux/mempolicy.h>
30 #include <linux/rmap.h>
31 #include <linux/mmu_notifier.h>
32 #include <linux/perf_event.h>
33 #include <linux/audit.h>
34 #include <linux/khugepaged.h>
35 #include <linux/uprobes.h>
36 #include <linux/rbtree_augmented.h>
37 #include <linux/sched/sysctl.h>
38 #include <linux/notifier.h>
39 #include <linux/memory.h>
41 #include <asm/uaccess.h>
42 #include <asm/cacheflush.h>
44 #include <asm/mmu_context.h>
48 #ifndef arch_mmap_check
49 #define arch_mmap_check(addr, len, flags) (0)
52 #ifndef arch_rebalance_pgtables
53 #define arch_rebalance_pgtables(addr, len) (addr)
56 static void unmap_region(struct mm_struct
*mm
,
57 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
58 unsigned long start
, unsigned long end
);
60 /* description of effects of mapping type and prot in current implementation.
61 * this is due to the limited x86 page protection hardware. The expected
62 * behavior is in parens:
65 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
66 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
67 * w: (no) no w: (no) no w: (yes) yes w: (no) no
68 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
70 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
71 * w: (no) no w: (no) no w: (copy) copy w: (no) no
72 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
75 pgprot_t protection_map
[16] = {
76 __P000
, __P001
, __P010
, __P011
, __P100
, __P101
, __P110
, __P111
,
77 __S000
, __S001
, __S010
, __S011
, __S100
, __S101
, __S110
, __S111
80 pgprot_t
vm_get_page_prot(unsigned long vm_flags
)
82 return __pgprot(pgprot_val(protection_map
[vm_flags
&
83 (VM_READ
|VM_WRITE
|VM_EXEC
|VM_SHARED
)]) |
84 pgprot_val(arch_vm_get_page_prot(vm_flags
)));
86 EXPORT_SYMBOL(vm_get_page_prot
);
88 int sysctl_overcommit_memory __read_mostly
= OVERCOMMIT_GUESS
; /* heuristic overcommit */
89 int sysctl_overcommit_ratio __read_mostly
= 50; /* default is 50% */
90 unsigned long sysctl_overcommit_kbytes __read_mostly
;
91 int sysctl_max_map_count __read_mostly
= DEFAULT_MAX_MAP_COUNT
;
92 unsigned long sysctl_user_reserve_kbytes __read_mostly
= 1UL << 17; /* 128MB */
93 unsigned long sysctl_admin_reserve_kbytes __read_mostly
= 1UL << 13; /* 8MB */
95 * Make sure vm_committed_as in one cacheline and not cacheline shared with
96 * other variables. It can be updated by several CPUs frequently.
98 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp
;
101 * The global memory commitment made in the system can be a metric
102 * that can be used to drive ballooning decisions when Linux is hosted
103 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
104 * balancing memory across competing virtual machines that are hosted.
105 * Several metrics drive this policy engine including the guest reported
108 unsigned long vm_memory_committed(void)
110 return percpu_counter_read_positive(&vm_committed_as
);
112 EXPORT_SYMBOL_GPL(vm_memory_committed
);
115 * Check that a process has enough memory to allocate a new virtual
116 * mapping. 0 means there is enough memory for the allocation to
117 * succeed and -ENOMEM implies there is not.
119 * We currently support three overcommit policies, which are set via the
120 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
122 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
123 * Additional code 2002 Jul 20 by Robert Love.
125 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
127 * Note this is a helper function intended to be used by LSMs which
128 * wish to use this logic.
130 int __vm_enough_memory(struct mm_struct
*mm
, long pages
, int cap_sys_admin
)
132 unsigned long free
, allowed
, reserve
;
134 vm_acct_memory(pages
);
137 * Sometimes we want to use more memory than we have
139 if (sysctl_overcommit_memory
== OVERCOMMIT_ALWAYS
)
142 if (sysctl_overcommit_memory
== OVERCOMMIT_GUESS
) {
143 free
= global_page_state(NR_FREE_PAGES
);
144 free
+= global_page_state(NR_FILE_PAGES
);
147 * shmem pages shouldn't be counted as free in this
148 * case, they can't be purged, only swapped out, and
149 * that won't affect the overall amount of available
150 * memory in the system.
152 free
-= global_page_state(NR_SHMEM
);
154 free
+= get_nr_swap_pages();
157 * Any slabs which are created with the
158 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
159 * which are reclaimable, under pressure. The dentry
160 * cache and most inode caches should fall into this
162 free
+= global_page_state(NR_SLAB_RECLAIMABLE
);
165 * Leave reserved pages. The pages are not for anonymous pages.
167 if (free
<= totalreserve_pages
)
170 free
-= totalreserve_pages
;
173 * Reserve some for root
176 free
-= sysctl_admin_reserve_kbytes
>> (PAGE_SHIFT
- 10);
184 allowed
= vm_commit_limit();
186 * Reserve some for root
189 allowed
-= sysctl_admin_reserve_kbytes
>> (PAGE_SHIFT
- 10);
192 * Don't let a single process grow so big a user can't recover
195 reserve
= sysctl_user_reserve_kbytes
>> (PAGE_SHIFT
- 10);
196 allowed
-= min(mm
->total_vm
/ 32, reserve
);
199 if (percpu_counter_read_positive(&vm_committed_as
) < allowed
)
202 vm_unacct_memory(pages
);
208 * Requires inode->i_mapping->i_mmap_mutex
210 static void __remove_shared_vm_struct(struct vm_area_struct
*vma
,
211 struct file
*file
, struct address_space
*mapping
)
213 if (vma
->vm_flags
& VM_DENYWRITE
)
214 atomic_inc(&file_inode(file
)->i_writecount
);
215 if (vma
->vm_flags
& VM_SHARED
)
216 mapping
->i_mmap_writable
--;
218 flush_dcache_mmap_lock(mapping
);
219 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
220 list_del_init(&vma
->shared
.nonlinear
);
222 vma_interval_tree_remove(vma
, &mapping
->i_mmap
);
223 flush_dcache_mmap_unlock(mapping
);
227 * Unlink a file-based vm structure from its interval tree, to hide
228 * vma from rmap and vmtruncate before freeing its page tables.
230 void unlink_file_vma(struct vm_area_struct
*vma
)
232 struct file
*file
= vma
->vm_file
;
235 struct address_space
*mapping
= file
->f_mapping
;
236 mutex_lock(&mapping
->i_mmap_mutex
);
237 __remove_shared_vm_struct(vma
, file
, mapping
);
238 mutex_unlock(&mapping
->i_mmap_mutex
);
243 * Close a vm structure and free it, returning the next.
245 static struct vm_area_struct
*remove_vma(struct vm_area_struct
*vma
)
247 struct vm_area_struct
*next
= vma
->vm_next
;
250 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
251 vma
->vm_ops
->close(vma
);
254 mpol_put(vma_policy(vma
));
255 kmem_cache_free(vm_area_cachep
, vma
);
259 static unsigned long do_brk(unsigned long addr
, unsigned long len
);
261 SYSCALL_DEFINE1(brk
, unsigned long, brk
)
263 unsigned long rlim
, retval
;
264 unsigned long newbrk
, oldbrk
;
265 struct mm_struct
*mm
= current
->mm
;
266 unsigned long min_brk
;
269 down_write(&mm
->mmap_sem
);
271 #ifdef CONFIG_COMPAT_BRK
273 * CONFIG_COMPAT_BRK can still be overridden by setting
274 * randomize_va_space to 2, which will still cause mm->start_brk
275 * to be arbitrarily shifted
277 if (current
->brk_randomized
)
278 min_brk
= mm
->start_brk
;
280 min_brk
= mm
->end_data
;
282 min_brk
= mm
->start_brk
;
288 * Check against rlimit here. If this check is done later after the test
289 * of oldbrk with newbrk then it can escape the test and let the data
290 * segment grow beyond its set limit the in case where the limit is
291 * not page aligned -Ram Gupta
293 rlim
= rlimit(RLIMIT_DATA
);
294 if (rlim
< RLIM_INFINITY
&& (brk
- mm
->start_brk
) +
295 (mm
->end_data
- mm
->start_data
) > rlim
)
298 newbrk
= PAGE_ALIGN(brk
);
299 oldbrk
= PAGE_ALIGN(mm
->brk
);
300 if (oldbrk
== newbrk
)
303 /* Always allow shrinking brk. */
304 if (brk
<= mm
->brk
) {
305 if (!do_munmap(mm
, newbrk
, oldbrk
-newbrk
))
310 /* Check against existing mmap mappings. */
311 if (find_vma_intersection(mm
, oldbrk
, newbrk
+PAGE_SIZE
))
314 /* Ok, looks good - let it rip. */
315 if (do_brk(oldbrk
, newbrk
-oldbrk
) != oldbrk
)
320 populate
= newbrk
> oldbrk
&& (mm
->def_flags
& VM_LOCKED
) != 0;
321 up_write(&mm
->mmap_sem
);
323 mm_populate(oldbrk
, newbrk
- oldbrk
);
328 up_write(&mm
->mmap_sem
);
332 static long vma_compute_subtree_gap(struct vm_area_struct
*vma
)
334 unsigned long max
, subtree_gap
;
337 max
-= vma
->vm_prev
->vm_end
;
338 if (vma
->vm_rb
.rb_left
) {
339 subtree_gap
= rb_entry(vma
->vm_rb
.rb_left
,
340 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
341 if (subtree_gap
> max
)
344 if (vma
->vm_rb
.rb_right
) {
345 subtree_gap
= rb_entry(vma
->vm_rb
.rb_right
,
346 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
347 if (subtree_gap
> max
)
353 #ifdef CONFIG_DEBUG_VM_RB
354 static int browse_rb(struct rb_root
*root
)
356 int i
= 0, j
, bug
= 0;
357 struct rb_node
*nd
, *pn
= NULL
;
358 unsigned long prev
= 0, pend
= 0;
360 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
361 struct vm_area_struct
*vma
;
362 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
363 if (vma
->vm_start
< prev
) {
364 printk("vm_start %lx prev %lx\n", vma
->vm_start
, prev
);
367 if (vma
->vm_start
< pend
) {
368 printk("vm_start %lx pend %lx\n", vma
->vm_start
, pend
);
371 if (vma
->vm_start
> vma
->vm_end
) {
372 printk("vm_end %lx < vm_start %lx\n",
373 vma
->vm_end
, vma
->vm_start
);
376 if (vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
)) {
377 printk("free gap %lx, correct %lx\n",
379 vma_compute_subtree_gap(vma
));
384 prev
= vma
->vm_start
;
388 for (nd
= pn
; nd
; nd
= rb_prev(nd
))
391 printk("backwards %d, forwards %d\n", j
, i
);
397 static void validate_mm_rb(struct rb_root
*root
, struct vm_area_struct
*ignore
)
401 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
402 struct vm_area_struct
*vma
;
403 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
404 BUG_ON(vma
!= ignore
&&
405 vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
));
409 static void validate_mm(struct mm_struct
*mm
)
413 unsigned long highest_address
= 0;
414 struct vm_area_struct
*vma
= mm
->mmap
;
416 struct anon_vma_chain
*avc
;
417 vma_lock_anon_vma(vma
);
418 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
419 anon_vma_interval_tree_verify(avc
);
420 vma_unlock_anon_vma(vma
);
421 highest_address
= vma
->vm_end
;
425 if (i
!= mm
->map_count
) {
426 printk("map_count %d vm_next %d\n", mm
->map_count
, i
);
429 if (highest_address
!= mm
->highest_vm_end
) {
430 printk("mm->highest_vm_end %lx, found %lx\n",
431 mm
->highest_vm_end
, highest_address
);
434 i
= browse_rb(&mm
->mm_rb
);
435 if (i
!= mm
->map_count
) {
436 printk("map_count %d rb %d\n", mm
->map_count
, i
);
442 #define validate_mm_rb(root, ignore) do { } while (0)
443 #define validate_mm(mm) do { } while (0)
446 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks
, struct vm_area_struct
, vm_rb
,
447 unsigned long, rb_subtree_gap
, vma_compute_subtree_gap
)
450 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
451 * vma->vm_prev->vm_end values changed, without modifying the vma's position
454 static void vma_gap_update(struct vm_area_struct
*vma
)
457 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
458 * function that does exacltly what we want.
460 vma_gap_callbacks_propagate(&vma
->vm_rb
, NULL
);
463 static inline void vma_rb_insert(struct vm_area_struct
*vma
,
464 struct rb_root
*root
)
466 /* All rb_subtree_gap values must be consistent prior to insertion */
467 validate_mm_rb(root
, NULL
);
469 rb_insert_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
472 static void vma_rb_erase(struct vm_area_struct
*vma
, struct rb_root
*root
)
475 * All rb_subtree_gap values must be consistent prior to erase,
476 * with the possible exception of the vma being erased.
478 validate_mm_rb(root
, vma
);
481 * Note rb_erase_augmented is a fairly large inline function,
482 * so make sure we instantiate it only once with our desired
483 * augmented rbtree callbacks.
485 rb_erase_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
489 * vma has some anon_vma assigned, and is already inserted on that
490 * anon_vma's interval trees.
492 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
493 * vma must be removed from the anon_vma's interval trees using
494 * anon_vma_interval_tree_pre_update_vma().
496 * After the update, the vma will be reinserted using
497 * anon_vma_interval_tree_post_update_vma().
499 * The entire update must be protected by exclusive mmap_sem and by
500 * the root anon_vma's mutex.
503 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct
*vma
)
505 struct anon_vma_chain
*avc
;
507 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
508 anon_vma_interval_tree_remove(avc
, &avc
->anon_vma
->rb_root
);
512 anon_vma_interval_tree_post_update_vma(struct vm_area_struct
*vma
)
514 struct anon_vma_chain
*avc
;
516 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
517 anon_vma_interval_tree_insert(avc
, &avc
->anon_vma
->rb_root
);
520 static int find_vma_links(struct mm_struct
*mm
, unsigned long addr
,
521 unsigned long end
, struct vm_area_struct
**pprev
,
522 struct rb_node
***rb_link
, struct rb_node
**rb_parent
)
524 struct rb_node
**__rb_link
, *__rb_parent
, *rb_prev
;
526 __rb_link
= &mm
->mm_rb
.rb_node
;
527 rb_prev
= __rb_parent
= NULL
;
530 struct vm_area_struct
*vma_tmp
;
532 __rb_parent
= *__rb_link
;
533 vma_tmp
= rb_entry(__rb_parent
, struct vm_area_struct
, vm_rb
);
535 if (vma_tmp
->vm_end
> addr
) {
536 /* Fail if an existing vma overlaps the area */
537 if (vma_tmp
->vm_start
< end
)
539 __rb_link
= &__rb_parent
->rb_left
;
541 rb_prev
= __rb_parent
;
542 __rb_link
= &__rb_parent
->rb_right
;
548 *pprev
= rb_entry(rb_prev
, struct vm_area_struct
, vm_rb
);
549 *rb_link
= __rb_link
;
550 *rb_parent
= __rb_parent
;
554 static unsigned long count_vma_pages_range(struct mm_struct
*mm
,
555 unsigned long addr
, unsigned long end
)
557 unsigned long nr_pages
= 0;
558 struct vm_area_struct
*vma
;
560 /* Find first overlaping mapping */
561 vma
= find_vma_intersection(mm
, addr
, end
);
565 nr_pages
= (min(end
, vma
->vm_end
) -
566 max(addr
, vma
->vm_start
)) >> PAGE_SHIFT
;
568 /* Iterate over the rest of the overlaps */
569 for (vma
= vma
->vm_next
; vma
; vma
= vma
->vm_next
) {
570 unsigned long overlap_len
;
572 if (vma
->vm_start
> end
)
575 overlap_len
= min(end
, vma
->vm_end
) - vma
->vm_start
;
576 nr_pages
+= overlap_len
>> PAGE_SHIFT
;
582 void __vma_link_rb(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
583 struct rb_node
**rb_link
, struct rb_node
*rb_parent
)
585 /* Update tracking information for the gap following the new vma. */
587 vma_gap_update(vma
->vm_next
);
589 mm
->highest_vm_end
= vma
->vm_end
;
592 * vma->vm_prev wasn't known when we followed the rbtree to find the
593 * correct insertion point for that vma. As a result, we could not
594 * update the vma vm_rb parents rb_subtree_gap values on the way down.
595 * So, we first insert the vma with a zero rb_subtree_gap value
596 * (to be consistent with what we did on the way down), and then
597 * immediately update the gap to the correct value. Finally we
598 * rebalance the rbtree after all augmented values have been set.
600 rb_link_node(&vma
->vm_rb
, rb_parent
, rb_link
);
601 vma
->rb_subtree_gap
= 0;
603 vma_rb_insert(vma
, &mm
->mm_rb
);
606 static void __vma_link_file(struct vm_area_struct
*vma
)
612 struct address_space
*mapping
= file
->f_mapping
;
614 if (vma
->vm_flags
& VM_DENYWRITE
)
615 atomic_dec(&file_inode(file
)->i_writecount
);
616 if (vma
->vm_flags
& VM_SHARED
)
617 mapping
->i_mmap_writable
++;
619 flush_dcache_mmap_lock(mapping
);
620 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
621 vma_nonlinear_insert(vma
, &mapping
->i_mmap_nonlinear
);
623 vma_interval_tree_insert(vma
, &mapping
->i_mmap
);
624 flush_dcache_mmap_unlock(mapping
);
629 __vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
630 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
631 struct rb_node
*rb_parent
)
633 __vma_link_list(mm
, vma
, prev
, rb_parent
);
634 __vma_link_rb(mm
, vma
, rb_link
, rb_parent
);
637 static void vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
638 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
639 struct rb_node
*rb_parent
)
641 struct address_space
*mapping
= NULL
;
644 mapping
= vma
->vm_file
->f_mapping
;
647 mutex_lock(&mapping
->i_mmap_mutex
);
649 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
650 __vma_link_file(vma
);
653 mutex_unlock(&mapping
->i_mmap_mutex
);
660 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
661 * mm's list and rbtree. It has already been inserted into the interval tree.
663 static void __insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
665 struct vm_area_struct
*prev
;
666 struct rb_node
**rb_link
, *rb_parent
;
668 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
669 &prev
, &rb_link
, &rb_parent
))
671 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
676 __vma_unlink(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
677 struct vm_area_struct
*prev
)
679 struct vm_area_struct
*next
;
681 vma_rb_erase(vma
, &mm
->mm_rb
);
682 prev
->vm_next
= next
= vma
->vm_next
;
684 next
->vm_prev
= prev
;
687 vmacache_invalidate(mm
);
691 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
692 * is already present in an i_mmap tree without adjusting the tree.
693 * The following helper function should be used when such adjustments
694 * are necessary. The "insert" vma (if any) is to be inserted
695 * before we drop the necessary locks.
697 int vma_adjust(struct vm_area_struct
*vma
, unsigned long start
,
698 unsigned long end
, pgoff_t pgoff
, struct vm_area_struct
*insert
)
700 struct mm_struct
*mm
= vma
->vm_mm
;
701 struct vm_area_struct
*next
= vma
->vm_next
;
702 struct vm_area_struct
*importer
= NULL
;
703 struct address_space
*mapping
= NULL
;
704 struct rb_root
*root
= NULL
;
705 struct anon_vma
*anon_vma
= NULL
;
706 struct file
*file
= vma
->vm_file
;
707 bool start_changed
= false, end_changed
= false;
708 long adjust_next
= 0;
711 if (next
&& !insert
) {
712 struct vm_area_struct
*exporter
= NULL
;
714 if (end
>= next
->vm_end
) {
716 * vma expands, overlapping all the next, and
717 * perhaps the one after too (mprotect case 6).
719 again
: remove_next
= 1 + (end
> next
->vm_end
);
723 } else if (end
> next
->vm_start
) {
725 * vma expands, overlapping part of the next:
726 * mprotect case 5 shifting the boundary up.
728 adjust_next
= (end
- next
->vm_start
) >> PAGE_SHIFT
;
731 } else if (end
< vma
->vm_end
) {
733 * vma shrinks, and !insert tells it's not
734 * split_vma inserting another: so it must be
735 * mprotect case 4 shifting the boundary down.
737 adjust_next
= - ((vma
->vm_end
- end
) >> PAGE_SHIFT
);
743 * Easily overlooked: when mprotect shifts the boundary,
744 * make sure the expanding vma has anon_vma set if the
745 * shrinking vma had, to cover any anon pages imported.
747 if (exporter
&& exporter
->anon_vma
&& !importer
->anon_vma
) {
748 if (anon_vma_clone(importer
, exporter
))
750 importer
->anon_vma
= exporter
->anon_vma
;
755 mapping
= file
->f_mapping
;
756 if (!(vma
->vm_flags
& VM_NONLINEAR
)) {
757 root
= &mapping
->i_mmap
;
758 uprobe_munmap(vma
, vma
->vm_start
, vma
->vm_end
);
761 uprobe_munmap(next
, next
->vm_start
,
765 mutex_lock(&mapping
->i_mmap_mutex
);
768 * Put into interval tree now, so instantiated pages
769 * are visible to arm/parisc __flush_dcache_page
770 * throughout; but we cannot insert into address
771 * space until vma start or end is updated.
773 __vma_link_file(insert
);
777 vma_adjust_trans_huge(vma
, start
, end
, adjust_next
);
779 anon_vma
= vma
->anon_vma
;
780 if (!anon_vma
&& adjust_next
)
781 anon_vma
= next
->anon_vma
;
783 VM_BUG_ON(adjust_next
&& next
->anon_vma
&&
784 anon_vma
!= next
->anon_vma
);
785 anon_vma_lock_write(anon_vma
);
786 anon_vma_interval_tree_pre_update_vma(vma
);
788 anon_vma_interval_tree_pre_update_vma(next
);
792 flush_dcache_mmap_lock(mapping
);
793 vma_interval_tree_remove(vma
, root
);
795 vma_interval_tree_remove(next
, root
);
798 if (start
!= vma
->vm_start
) {
799 vma
->vm_start
= start
;
800 start_changed
= true;
802 if (end
!= vma
->vm_end
) {
806 vma
->vm_pgoff
= pgoff
;
808 next
->vm_start
+= adjust_next
<< PAGE_SHIFT
;
809 next
->vm_pgoff
+= adjust_next
;
814 vma_interval_tree_insert(next
, root
);
815 vma_interval_tree_insert(vma
, root
);
816 flush_dcache_mmap_unlock(mapping
);
821 * vma_merge has merged next into vma, and needs
822 * us to remove next before dropping the locks.
824 __vma_unlink(mm
, next
, vma
);
826 __remove_shared_vm_struct(next
, file
, mapping
);
829 * split_vma has split insert from vma, and needs
830 * us to insert it before dropping the locks
831 * (it may either follow vma or precede it).
833 __insert_vm_struct(mm
, insert
);
839 mm
->highest_vm_end
= end
;
840 else if (!adjust_next
)
841 vma_gap_update(next
);
846 anon_vma_interval_tree_post_update_vma(vma
);
848 anon_vma_interval_tree_post_update_vma(next
);
849 anon_vma_unlock_write(anon_vma
);
852 mutex_unlock(&mapping
->i_mmap_mutex
);
863 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
867 anon_vma_merge(vma
, next
);
869 mpol_put(vma_policy(next
));
870 kmem_cache_free(vm_area_cachep
, next
);
872 * In mprotect's case 6 (see comments on vma_merge),
873 * we must remove another next too. It would clutter
874 * up the code too much to do both in one go.
877 if (remove_next
== 2)
880 vma_gap_update(next
);
882 mm
->highest_vm_end
= end
;
893 * If the vma has a ->close operation then the driver probably needs to release
894 * per-vma resources, so we don't attempt to merge those.
896 static inline int is_mergeable_vma(struct vm_area_struct
*vma
,
897 struct file
*file
, unsigned long vm_flags
)
900 * VM_SOFTDIRTY should not prevent from VMA merging, if we
901 * match the flags but dirty bit -- the caller should mark
902 * merged VMA as dirty. If dirty bit won't be excluded from
903 * comparison, we increase pressue on the memory system forcing
904 * the kernel to generate new VMAs when old one could be
907 if ((vma
->vm_flags
^ vm_flags
) & ~VM_SOFTDIRTY
)
909 if (vma
->vm_file
!= file
)
911 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
916 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
917 struct anon_vma
*anon_vma2
,
918 struct vm_area_struct
*vma
)
921 * The list_is_singular() test is to avoid merging VMA cloned from
922 * parents. This can improve scalability caused by anon_vma lock.
924 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
925 list_is_singular(&vma
->anon_vma_chain
)))
927 return anon_vma1
== anon_vma2
;
931 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
932 * in front of (at a lower virtual address and file offset than) the vma.
934 * We cannot merge two vmas if they have differently assigned (non-NULL)
935 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
937 * We don't check here for the merged mmap wrapping around the end of pagecache
938 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
939 * wrap, nor mmaps which cover the final page at index -1UL.
942 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
943 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
945 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
946 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
947 if (vma
->vm_pgoff
== vm_pgoff
)
954 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
955 * beyond (at a higher virtual address and file offset than) the vma.
957 * We cannot merge two vmas if they have differently assigned (non-NULL)
958 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
961 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
962 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
964 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
965 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
967 vm_pglen
= vma_pages(vma
);
968 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
975 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
976 * whether that can be merged with its predecessor or its successor.
977 * Or both (it neatly fills a hole).
979 * In most cases - when called for mmap, brk or mremap - [addr,end) is
980 * certain not to be mapped by the time vma_merge is called; but when
981 * called for mprotect, it is certain to be already mapped (either at
982 * an offset within prev, or at the start of next), and the flags of
983 * this area are about to be changed to vm_flags - and the no-change
984 * case has already been eliminated.
986 * The following mprotect cases have to be considered, where AAAA is
987 * the area passed down from mprotect_fixup, never extending beyond one
988 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
990 * AAAA AAAA AAAA AAAA
991 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
992 * cannot merge might become might become might become
993 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
994 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
995 * mremap move: PPPPNNNNNNNN 8
997 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
998 * might become case 1 below case 2 below case 3 below
1000 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1001 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1003 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
1004 struct vm_area_struct
*prev
, unsigned long addr
,
1005 unsigned long end
, unsigned long vm_flags
,
1006 struct anon_vma
*anon_vma
, struct file
*file
,
1007 pgoff_t pgoff
, struct mempolicy
*policy
)
1009 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
1010 struct vm_area_struct
*area
, *next
;
1014 * We later require that vma->vm_flags == vm_flags,
1015 * so this tests vma->vm_flags & VM_SPECIAL, too.
1017 if (vm_flags
& VM_SPECIAL
)
1021 next
= prev
->vm_next
;
1025 if (next
&& next
->vm_end
== end
) /* cases 6, 7, 8 */
1026 next
= next
->vm_next
;
1029 * Can it merge with the predecessor?
1031 if (prev
&& prev
->vm_end
== addr
&&
1032 mpol_equal(vma_policy(prev
), policy
) &&
1033 can_vma_merge_after(prev
, vm_flags
,
1034 anon_vma
, file
, pgoff
)) {
1036 * OK, it can. Can we now merge in the successor as well?
1038 if (next
&& end
== next
->vm_start
&&
1039 mpol_equal(policy
, vma_policy(next
)) &&
1040 can_vma_merge_before(next
, vm_flags
,
1041 anon_vma
, file
, pgoff
+pglen
) &&
1042 is_mergeable_anon_vma(prev
->anon_vma
,
1043 next
->anon_vma
, NULL
)) {
1045 err
= vma_adjust(prev
, prev
->vm_start
,
1046 next
->vm_end
, prev
->vm_pgoff
, NULL
);
1047 } else /* cases 2, 5, 7 */
1048 err
= vma_adjust(prev
, prev
->vm_start
,
1049 end
, prev
->vm_pgoff
, NULL
);
1052 khugepaged_enter_vma_merge(prev
);
1057 * Can this new request be merged in front of next?
1059 if (next
&& end
== next
->vm_start
&&
1060 mpol_equal(policy
, vma_policy(next
)) &&
1061 can_vma_merge_before(next
, vm_flags
,
1062 anon_vma
, file
, pgoff
+pglen
)) {
1063 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
1064 err
= vma_adjust(prev
, prev
->vm_start
,
1065 addr
, prev
->vm_pgoff
, NULL
);
1066 else /* cases 3, 8 */
1067 err
= vma_adjust(area
, addr
, next
->vm_end
,
1068 next
->vm_pgoff
- pglen
, NULL
);
1071 khugepaged_enter_vma_merge(area
);
1079 * Rough compatbility check to quickly see if it's even worth looking
1080 * at sharing an anon_vma.
1082 * They need to have the same vm_file, and the flags can only differ
1083 * in things that mprotect may change.
1085 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1086 * we can merge the two vma's. For example, we refuse to merge a vma if
1087 * there is a vm_ops->close() function, because that indicates that the
1088 * driver is doing some kind of reference counting. But that doesn't
1089 * really matter for the anon_vma sharing case.
1091 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1093 return a
->vm_end
== b
->vm_start
&&
1094 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
1095 a
->vm_file
== b
->vm_file
&&
1096 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_READ
|VM_WRITE
|VM_EXEC
|VM_SOFTDIRTY
)) &&
1097 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
1101 * Do some basic sanity checking to see if we can re-use the anon_vma
1102 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1103 * the same as 'old', the other will be the new one that is trying
1104 * to share the anon_vma.
1106 * NOTE! This runs with mm_sem held for reading, so it is possible that
1107 * the anon_vma of 'old' is concurrently in the process of being set up
1108 * by another page fault trying to merge _that_. But that's ok: if it
1109 * is being set up, that automatically means that it will be a singleton
1110 * acceptable for merging, so we can do all of this optimistically. But
1111 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1113 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1114 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1115 * is to return an anon_vma that is "complex" due to having gone through
1118 * We also make sure that the two vma's are compatible (adjacent,
1119 * and with the same memory policies). That's all stable, even with just
1120 * a read lock on the mm_sem.
1122 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1124 if (anon_vma_compatible(a
, b
)) {
1125 struct anon_vma
*anon_vma
= ACCESS_ONCE(old
->anon_vma
);
1127 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
1134 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1135 * neighbouring vmas for a suitable anon_vma, before it goes off
1136 * to allocate a new anon_vma. It checks because a repetitive
1137 * sequence of mprotects and faults may otherwise lead to distinct
1138 * anon_vmas being allocated, preventing vma merge in subsequent
1141 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
1143 struct anon_vma
*anon_vma
;
1144 struct vm_area_struct
*near
;
1146 near
= vma
->vm_next
;
1150 anon_vma
= reusable_anon_vma(near
, vma
, near
);
1154 near
= vma
->vm_prev
;
1158 anon_vma
= reusable_anon_vma(near
, near
, vma
);
1163 * There's no absolute need to look only at touching neighbours:
1164 * we could search further afield for "compatible" anon_vmas.
1165 * But it would probably just be a waste of time searching,
1166 * or lead to too many vmas hanging off the same anon_vma.
1167 * We're trying to allow mprotect remerging later on,
1168 * not trying to minimize memory used for anon_vmas.
1173 #ifdef CONFIG_PROC_FS
1174 void vm_stat_account(struct mm_struct
*mm
, unsigned long flags
,
1175 struct file
*file
, long pages
)
1177 const unsigned long stack_flags
1178 = VM_STACK_FLAGS
& (VM_GROWSUP
|VM_GROWSDOWN
);
1180 mm
->total_vm
+= pages
;
1183 mm
->shared_vm
+= pages
;
1184 if ((flags
& (VM_EXEC
|VM_WRITE
)) == VM_EXEC
)
1185 mm
->exec_vm
+= pages
;
1186 } else if (flags
& stack_flags
)
1187 mm
->stack_vm
+= pages
;
1189 #endif /* CONFIG_PROC_FS */
1192 * If a hint addr is less than mmap_min_addr change hint to be as
1193 * low as possible but still greater than mmap_min_addr
1195 static inline unsigned long round_hint_to_min(unsigned long hint
)
1198 if (((void *)hint
!= NULL
) &&
1199 (hint
< mmap_min_addr
))
1200 return PAGE_ALIGN(mmap_min_addr
);
1204 static inline int mlock_future_check(struct mm_struct
*mm
,
1205 unsigned long flags
,
1208 unsigned long locked
, lock_limit
;
1210 /* mlock MCL_FUTURE? */
1211 if (flags
& VM_LOCKED
) {
1212 locked
= len
>> PAGE_SHIFT
;
1213 locked
+= mm
->locked_vm
;
1214 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1215 lock_limit
>>= PAGE_SHIFT
;
1216 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1223 * The caller must hold down_write(¤t->mm->mmap_sem).
1226 unsigned long do_mmap_pgoff(struct file
*file
, unsigned long addr
,
1227 unsigned long len
, unsigned long prot
,
1228 unsigned long flags
, unsigned long pgoff
,
1229 unsigned long *populate
)
1231 struct mm_struct
* mm
= current
->mm
;
1232 vm_flags_t vm_flags
;
1237 * Does the application expect PROT_READ to imply PROT_EXEC?
1239 * (the exception is when the underlying filesystem is noexec
1240 * mounted, in which case we dont add PROT_EXEC.)
1242 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1243 if (!(file
&& (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
)))
1249 if (!(flags
& MAP_FIXED
))
1250 addr
= round_hint_to_min(addr
);
1252 /* Careful about overflows.. */
1253 len
= PAGE_ALIGN(len
);
1257 /* offset overflow? */
1258 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1261 /* Too many mappings? */
1262 if (mm
->map_count
> sysctl_max_map_count
)
1265 /* Obtain the address to map to. we verify (or select) it and ensure
1266 * that it represents a valid section of the address space.
1268 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1269 if (addr
& ~PAGE_MASK
)
1272 /* Do simple checking here so the lower-level routines won't have
1273 * to. we assume access permissions have been handled by the open
1274 * of the memory object, so we don't do any here.
1276 vm_flags
= calc_vm_prot_bits(prot
) | calc_vm_flag_bits(flags
) |
1277 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1279 if (flags
& MAP_LOCKED
)
1280 if (!can_do_mlock())
1283 if (mlock_future_check(mm
, vm_flags
, len
))
1287 struct inode
*inode
= file_inode(file
);
1289 switch (flags
& MAP_TYPE
) {
1291 if ((prot
&PROT_WRITE
) && !(file
->f_mode
&FMODE_WRITE
))
1295 * Make sure we don't allow writing to an append-only
1298 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1302 * Make sure there are no mandatory locks on the file.
1304 if (locks_verify_locked(file
))
1307 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1308 if (!(file
->f_mode
& FMODE_WRITE
))
1309 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1313 if (!(file
->f_mode
& FMODE_READ
))
1315 if (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
) {
1316 if (vm_flags
& VM_EXEC
)
1318 vm_flags
&= ~VM_MAYEXEC
;
1321 if (!file
->f_op
->mmap
)
1323 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1331 switch (flags
& MAP_TYPE
) {
1333 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1339 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1343 * Set pgoff according to addr for anon_vma.
1345 pgoff
= addr
>> PAGE_SHIFT
;
1353 * Set 'VM_NORESERVE' if we should not account for the
1354 * memory use of this mapping.
1356 if (flags
& MAP_NORESERVE
) {
1357 /* We honor MAP_NORESERVE if allowed to overcommit */
1358 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1359 vm_flags
|= VM_NORESERVE
;
1361 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1362 if (file
&& is_file_hugepages(file
))
1363 vm_flags
|= VM_NORESERVE
;
1366 addr
= mmap_region(file
, addr
, len
, vm_flags
, pgoff
);
1367 if (!IS_ERR_VALUE(addr
) &&
1368 ((vm_flags
& VM_LOCKED
) ||
1369 (flags
& (MAP_POPULATE
| MAP_NONBLOCK
)) == MAP_POPULATE
))
1374 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1375 unsigned long, prot
, unsigned long, flags
,
1376 unsigned long, fd
, unsigned long, pgoff
)
1378 struct file
*file
= NULL
;
1379 unsigned long retval
= -EBADF
;
1381 if (!(flags
& MAP_ANONYMOUS
)) {
1382 audit_mmap_fd(fd
, flags
);
1386 if (is_file_hugepages(file
))
1387 len
= ALIGN(len
, huge_page_size(hstate_file(file
)));
1389 if (unlikely(flags
& MAP_HUGETLB
&& !is_file_hugepages(file
)))
1391 } else if (flags
& MAP_HUGETLB
) {
1392 struct user_struct
*user
= NULL
;
1395 hs
= hstate_sizelog((flags
>> MAP_HUGE_SHIFT
) & SHM_HUGE_MASK
);
1399 len
= ALIGN(len
, huge_page_size(hs
));
1401 * VM_NORESERVE is used because the reservations will be
1402 * taken when vm_ops->mmap() is called
1403 * A dummy user value is used because we are not locking
1404 * memory so no accounting is necessary
1406 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, len
,
1408 &user
, HUGETLB_ANONHUGE_INODE
,
1409 (flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1411 return PTR_ERR(file
);
1414 flags
&= ~(MAP_EXECUTABLE
| MAP_DENYWRITE
);
1416 retval
= vm_mmap_pgoff(file
, addr
, len
, prot
, flags
, pgoff
);
1424 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1425 struct mmap_arg_struct
{
1429 unsigned long flags
;
1431 unsigned long offset
;
1434 SYSCALL_DEFINE1(old_mmap
, struct mmap_arg_struct __user
*, arg
)
1436 struct mmap_arg_struct a
;
1438 if (copy_from_user(&a
, arg
, sizeof(a
)))
1440 if (a
.offset
& ~PAGE_MASK
)
1443 return sys_mmap_pgoff(a
.addr
, a
.len
, a
.prot
, a
.flags
, a
.fd
,
1444 a
.offset
>> PAGE_SHIFT
);
1446 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1449 * Some shared mappigns will want the pages marked read-only
1450 * to track write events. If so, we'll downgrade vm_page_prot
1451 * to the private version (using protection_map[] without the
1454 int vma_wants_writenotify(struct vm_area_struct
*vma
)
1456 vm_flags_t vm_flags
= vma
->vm_flags
;
1458 /* If it was private or non-writable, the write bit is already clear */
1459 if ((vm_flags
& (VM_WRITE
|VM_SHARED
)) != ((VM_WRITE
|VM_SHARED
)))
1462 /* The backer wishes to know when pages are first written to? */
1463 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
)
1466 /* The open routine did something to the protections already? */
1467 if (pgprot_val(vma
->vm_page_prot
) !=
1468 pgprot_val(vm_get_page_prot(vm_flags
)))
1471 /* Specialty mapping? */
1472 if (vm_flags
& VM_PFNMAP
)
1475 /* Can the mapping track the dirty pages? */
1476 return vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
1477 mapping_cap_account_dirty(vma
->vm_file
->f_mapping
);
1481 * We account for memory if it's a private writeable mapping,
1482 * not hugepages and VM_NORESERVE wasn't set.
1484 static inline int accountable_mapping(struct file
*file
, vm_flags_t vm_flags
)
1487 * hugetlb has its own accounting separate from the core VM
1488 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1490 if (file
&& is_file_hugepages(file
))
1493 return (vm_flags
& (VM_NORESERVE
| VM_SHARED
| VM_WRITE
)) == VM_WRITE
;
1496 unsigned long mmap_region(struct file
*file
, unsigned long addr
,
1497 unsigned long len
, vm_flags_t vm_flags
, unsigned long pgoff
)
1499 struct mm_struct
*mm
= current
->mm
;
1500 struct vm_area_struct
*vma
, *prev
;
1502 struct rb_node
**rb_link
, *rb_parent
;
1503 unsigned long charged
= 0;
1505 /* Check against address space limit. */
1506 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
)) {
1507 unsigned long nr_pages
;
1510 * MAP_FIXED may remove pages of mappings that intersects with
1511 * requested mapping. Account for the pages it would unmap.
1513 if (!(vm_flags
& MAP_FIXED
))
1516 nr_pages
= count_vma_pages_range(mm
, addr
, addr
+ len
);
1518 if (!may_expand_vm(mm
, (len
>> PAGE_SHIFT
) - nr_pages
))
1522 /* Clear old maps */
1525 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
1526 if (do_munmap(mm
, addr
, len
))
1532 * Private writable mapping: check memory availability
1534 if (accountable_mapping(file
, vm_flags
)) {
1535 charged
= len
>> PAGE_SHIFT
;
1536 if (security_vm_enough_memory_mm(mm
, charged
))
1538 vm_flags
|= VM_ACCOUNT
;
1542 * Can we just expand an old mapping?
1544 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
, NULL
, file
, pgoff
, NULL
);
1549 * Determine the object being mapped and call the appropriate
1550 * specific mapper. the address has already been validated, but
1551 * not unmapped, but the maps are removed from the list.
1553 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
1560 vma
->vm_start
= addr
;
1561 vma
->vm_end
= addr
+ len
;
1562 vma
->vm_flags
= vm_flags
;
1563 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1564 vma
->vm_pgoff
= pgoff
;
1565 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
1568 if (vm_flags
& VM_DENYWRITE
) {
1569 error
= deny_write_access(file
);
1573 vma
->vm_file
= get_file(file
);
1574 error
= file
->f_op
->mmap(file
, vma
);
1576 goto unmap_and_free_vma
;
1578 /* Can addr have changed??
1580 * Answer: Yes, several device drivers can do it in their
1581 * f_op->mmap method. -DaveM
1582 * Bug: If addr is changed, prev, rb_link, rb_parent should
1583 * be updated for vma_link()
1585 WARN_ON_ONCE(addr
!= vma
->vm_start
);
1587 addr
= vma
->vm_start
;
1588 vm_flags
= vma
->vm_flags
;
1589 } else if (vm_flags
& VM_SHARED
) {
1590 error
= shmem_zero_setup(vma
);
1595 if (vma_wants_writenotify(vma
)) {
1596 pgprot_t pprot
= vma
->vm_page_prot
;
1598 /* Can vma->vm_page_prot have changed??
1600 * Answer: Yes, drivers may have changed it in their
1601 * f_op->mmap method.
1603 * Ensures that vmas marked as uncached stay that way.
1605 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
& ~VM_SHARED
);
1606 if (pgprot_val(pprot
) == pgprot_val(pgprot_noncached(pprot
)))
1607 vma
->vm_page_prot
= pgprot_noncached(vma
->vm_page_prot
);
1610 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1611 /* Once vma denies write, undo our temporary denial count */
1612 if (vm_flags
& VM_DENYWRITE
)
1613 allow_write_access(file
);
1614 file
= vma
->vm_file
;
1616 perf_event_mmap(vma
);
1618 vm_stat_account(mm
, vm_flags
, file
, len
>> PAGE_SHIFT
);
1619 if (vm_flags
& VM_LOCKED
) {
1620 if (!((vm_flags
& VM_SPECIAL
) || is_vm_hugetlb_page(vma
) ||
1621 vma
== get_gate_vma(current
->mm
)))
1622 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1624 vma
->vm_flags
&= ~VM_LOCKED
;
1631 * New (or expanded) vma always get soft dirty status.
1632 * Otherwise user-space soft-dirty page tracker won't
1633 * be able to distinguish situation when vma area unmapped,
1634 * then new mapped in-place (which must be aimed as
1635 * a completely new data area).
1637 vma
->vm_flags
|= VM_SOFTDIRTY
;
1642 if (vm_flags
& VM_DENYWRITE
)
1643 allow_write_access(file
);
1644 vma
->vm_file
= NULL
;
1647 /* Undo any partial mapping done by a device driver. */
1648 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1651 kmem_cache_free(vm_area_cachep
, vma
);
1654 vm_unacct_memory(charged
);
1658 unsigned long unmapped_area(struct vm_unmapped_area_info
*info
)
1661 * We implement the search by looking for an rbtree node that
1662 * immediately follows a suitable gap. That is,
1663 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1664 * - gap_end = vma->vm_start >= info->low_limit + length;
1665 * - gap_end - gap_start >= length
1668 struct mm_struct
*mm
= current
->mm
;
1669 struct vm_area_struct
*vma
;
1670 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1672 /* Adjust search length to account for worst case alignment overhead */
1673 length
= info
->length
+ info
->align_mask
;
1674 if (length
< info
->length
)
1677 /* Adjust search limits by the desired length */
1678 if (info
->high_limit
< length
)
1680 high_limit
= info
->high_limit
- length
;
1682 if (info
->low_limit
> high_limit
)
1684 low_limit
= info
->low_limit
+ length
;
1686 /* Check if rbtree root looks promising */
1687 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1689 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1690 if (vma
->rb_subtree_gap
< length
)
1694 /* Visit left subtree if it looks promising */
1695 gap_end
= vma
->vm_start
;
1696 if (gap_end
>= low_limit
&& vma
->vm_rb
.rb_left
) {
1697 struct vm_area_struct
*left
=
1698 rb_entry(vma
->vm_rb
.rb_left
,
1699 struct vm_area_struct
, vm_rb
);
1700 if (left
->rb_subtree_gap
>= length
) {
1706 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1708 /* Check if current node has a suitable gap */
1709 if (gap_start
> high_limit
)
1711 if (gap_end
>= low_limit
&& gap_end
- gap_start
>= length
)
1714 /* Visit right subtree if it looks promising */
1715 if (vma
->vm_rb
.rb_right
) {
1716 struct vm_area_struct
*right
=
1717 rb_entry(vma
->vm_rb
.rb_right
,
1718 struct vm_area_struct
, vm_rb
);
1719 if (right
->rb_subtree_gap
>= length
) {
1725 /* Go back up the rbtree to find next candidate node */
1727 struct rb_node
*prev
= &vma
->vm_rb
;
1728 if (!rb_parent(prev
))
1730 vma
= rb_entry(rb_parent(prev
),
1731 struct vm_area_struct
, vm_rb
);
1732 if (prev
== vma
->vm_rb
.rb_left
) {
1733 gap_start
= vma
->vm_prev
->vm_end
;
1734 gap_end
= vma
->vm_start
;
1741 /* Check highest gap, which does not precede any rbtree node */
1742 gap_start
= mm
->highest_vm_end
;
1743 gap_end
= ULONG_MAX
; /* Only for VM_BUG_ON below */
1744 if (gap_start
> high_limit
)
1748 /* We found a suitable gap. Clip it with the original low_limit. */
1749 if (gap_start
< info
->low_limit
)
1750 gap_start
= info
->low_limit
;
1752 /* Adjust gap address to the desired alignment */
1753 gap_start
+= (info
->align_offset
- gap_start
) & info
->align_mask
;
1755 VM_BUG_ON(gap_start
+ info
->length
> info
->high_limit
);
1756 VM_BUG_ON(gap_start
+ info
->length
> gap_end
);
1760 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info
*info
)
1762 struct mm_struct
*mm
= current
->mm
;
1763 struct vm_area_struct
*vma
;
1764 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1766 /* Adjust search length to account for worst case alignment overhead */
1767 length
= info
->length
+ info
->align_mask
;
1768 if (length
< info
->length
)
1772 * Adjust search limits by the desired length.
1773 * See implementation comment at top of unmapped_area().
1775 gap_end
= info
->high_limit
;
1776 if (gap_end
< length
)
1778 high_limit
= gap_end
- length
;
1780 if (info
->low_limit
> high_limit
)
1782 low_limit
= info
->low_limit
+ length
;
1784 /* Check highest gap, which does not precede any rbtree node */
1785 gap_start
= mm
->highest_vm_end
;
1786 if (gap_start
<= high_limit
)
1789 /* Check if rbtree root looks promising */
1790 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1792 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1793 if (vma
->rb_subtree_gap
< length
)
1797 /* Visit right subtree if it looks promising */
1798 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1799 if (gap_start
<= high_limit
&& vma
->vm_rb
.rb_right
) {
1800 struct vm_area_struct
*right
=
1801 rb_entry(vma
->vm_rb
.rb_right
,
1802 struct vm_area_struct
, vm_rb
);
1803 if (right
->rb_subtree_gap
>= length
) {
1810 /* Check if current node has a suitable gap */
1811 gap_end
= vma
->vm_start
;
1812 if (gap_end
< low_limit
)
1814 if (gap_start
<= high_limit
&& gap_end
- gap_start
>= length
)
1817 /* Visit left subtree if it looks promising */
1818 if (vma
->vm_rb
.rb_left
) {
1819 struct vm_area_struct
*left
=
1820 rb_entry(vma
->vm_rb
.rb_left
,
1821 struct vm_area_struct
, vm_rb
);
1822 if (left
->rb_subtree_gap
>= length
) {
1828 /* Go back up the rbtree to find next candidate node */
1830 struct rb_node
*prev
= &vma
->vm_rb
;
1831 if (!rb_parent(prev
))
1833 vma
= rb_entry(rb_parent(prev
),
1834 struct vm_area_struct
, vm_rb
);
1835 if (prev
== vma
->vm_rb
.rb_right
) {
1836 gap_start
= vma
->vm_prev
?
1837 vma
->vm_prev
->vm_end
: 0;
1844 /* We found a suitable gap. Clip it with the original high_limit. */
1845 if (gap_end
> info
->high_limit
)
1846 gap_end
= info
->high_limit
;
1849 /* Compute highest gap address at the desired alignment */
1850 gap_end
-= info
->length
;
1851 gap_end
-= (gap_end
- info
->align_offset
) & info
->align_mask
;
1853 VM_BUG_ON(gap_end
< info
->low_limit
);
1854 VM_BUG_ON(gap_end
< gap_start
);
1858 /* Get an address range which is currently unmapped.
1859 * For shmat() with addr=0.
1861 * Ugly calling convention alert:
1862 * Return value with the low bits set means error value,
1864 * if (ret & ~PAGE_MASK)
1867 * This function "knows" that -ENOMEM has the bits set.
1869 #ifndef HAVE_ARCH_UNMAPPED_AREA
1871 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
1872 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
1874 struct mm_struct
*mm
= current
->mm
;
1875 struct vm_area_struct
*vma
;
1876 struct vm_unmapped_area_info info
;
1878 if (len
> TASK_SIZE
- mmap_min_addr
)
1881 if (flags
& MAP_FIXED
)
1885 addr
= PAGE_ALIGN(addr
);
1886 vma
= find_vma(mm
, addr
);
1887 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1888 (!vma
|| addr
+ len
<= vma
->vm_start
))
1894 info
.low_limit
= mm
->mmap_base
;
1895 info
.high_limit
= TASK_SIZE
;
1896 info
.align_mask
= 0;
1897 return vm_unmapped_area(&info
);
1902 * This mmap-allocator allocates new areas top-down from below the
1903 * stack's low limit (the base):
1905 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1907 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
1908 const unsigned long len
, const unsigned long pgoff
,
1909 const unsigned long flags
)
1911 struct vm_area_struct
*vma
;
1912 struct mm_struct
*mm
= current
->mm
;
1913 unsigned long addr
= addr0
;
1914 struct vm_unmapped_area_info info
;
1916 /* requested length too big for entire address space */
1917 if (len
> TASK_SIZE
- mmap_min_addr
)
1920 if (flags
& MAP_FIXED
)
1923 /* requesting a specific address */
1925 addr
= PAGE_ALIGN(addr
);
1926 vma
= find_vma(mm
, addr
);
1927 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1928 (!vma
|| addr
+ len
<= vma
->vm_start
))
1932 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
1934 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
1935 info
.high_limit
= mm
->mmap_base
;
1936 info
.align_mask
= 0;
1937 addr
= vm_unmapped_area(&info
);
1940 * A failed mmap() very likely causes application failure,
1941 * so fall back to the bottom-up function here. This scenario
1942 * can happen with large stack limits and large mmap()
1945 if (addr
& ~PAGE_MASK
) {
1946 VM_BUG_ON(addr
!= -ENOMEM
);
1948 info
.low_limit
= TASK_UNMAPPED_BASE
;
1949 info
.high_limit
= TASK_SIZE
;
1950 addr
= vm_unmapped_area(&info
);
1958 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
1959 unsigned long pgoff
, unsigned long flags
)
1961 unsigned long (*get_area
)(struct file
*, unsigned long,
1962 unsigned long, unsigned long, unsigned long);
1964 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
1968 /* Careful about overflows.. */
1969 if (len
> TASK_SIZE
)
1972 get_area
= current
->mm
->get_unmapped_area
;
1973 if (file
&& file
->f_op
->get_unmapped_area
)
1974 get_area
= file
->f_op
->get_unmapped_area
;
1975 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
1976 if (IS_ERR_VALUE(addr
))
1979 if (addr
> TASK_SIZE
- len
)
1981 if (addr
& ~PAGE_MASK
)
1984 addr
= arch_rebalance_pgtables(addr
, len
);
1985 error
= security_mmap_addr(addr
);
1986 return error
? error
: addr
;
1989 EXPORT_SYMBOL(get_unmapped_area
);
1991 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1992 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
1994 struct rb_node
*rb_node
;
1995 struct vm_area_struct
*vma
;
1997 /* Check the cache first. */
1998 vma
= vmacache_find(mm
, addr
);
2002 rb_node
= mm
->mm_rb
.rb_node
;
2006 struct vm_area_struct
*tmp
;
2008 tmp
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2010 if (tmp
->vm_end
> addr
) {
2012 if (tmp
->vm_start
<= addr
)
2014 rb_node
= rb_node
->rb_left
;
2016 rb_node
= rb_node
->rb_right
;
2020 vmacache_update(addr
, vma
);
2024 EXPORT_SYMBOL(find_vma
);
2027 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2029 struct vm_area_struct
*
2030 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
2031 struct vm_area_struct
**pprev
)
2033 struct vm_area_struct
*vma
;
2035 vma
= find_vma(mm
, addr
);
2037 *pprev
= vma
->vm_prev
;
2039 struct rb_node
*rb_node
= mm
->mm_rb
.rb_node
;
2042 *pprev
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2043 rb_node
= rb_node
->rb_right
;
2050 * Verify that the stack growth is acceptable and
2051 * update accounting. This is shared with both the
2052 * grow-up and grow-down cases.
2054 static int acct_stack_growth(struct vm_area_struct
*vma
, unsigned long size
, unsigned long grow
)
2056 struct mm_struct
*mm
= vma
->vm_mm
;
2057 struct rlimit
*rlim
= current
->signal
->rlim
;
2058 unsigned long new_start
;
2060 /* address space limit tests */
2061 if (!may_expand_vm(mm
, grow
))
2064 /* Stack limit test */
2065 if (size
> ACCESS_ONCE(rlim
[RLIMIT_STACK
].rlim_cur
))
2068 /* mlock limit tests */
2069 if (vma
->vm_flags
& VM_LOCKED
) {
2070 unsigned long locked
;
2071 unsigned long limit
;
2072 locked
= mm
->locked_vm
+ grow
;
2073 limit
= ACCESS_ONCE(rlim
[RLIMIT_MEMLOCK
].rlim_cur
);
2074 limit
>>= PAGE_SHIFT
;
2075 if (locked
> limit
&& !capable(CAP_IPC_LOCK
))
2079 /* Check to ensure the stack will not grow into a hugetlb-only region */
2080 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
2082 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
2086 * Overcommit.. This must be the final test, as it will
2087 * update security statistics.
2089 if (security_vm_enough_memory_mm(mm
, grow
))
2092 /* Ok, everything looks good - let it rip */
2093 if (vma
->vm_flags
& VM_LOCKED
)
2094 mm
->locked_vm
+= grow
;
2095 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, grow
);
2099 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2101 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2102 * vma is the last one with address > vma->vm_end. Have to extend vma.
2104 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
)
2108 if (!(vma
->vm_flags
& VM_GROWSUP
))
2112 * We must make sure the anon_vma is allocated
2113 * so that the anon_vma locking is not a noop.
2115 if (unlikely(anon_vma_prepare(vma
)))
2117 vma_lock_anon_vma(vma
);
2120 * vma->vm_start/vm_end cannot change under us because the caller
2121 * is required to hold the mmap_sem in read mode. We need the
2122 * anon_vma lock to serialize against concurrent expand_stacks.
2123 * Also guard against wrapping around to address 0.
2125 if (address
< PAGE_ALIGN(address
+4))
2126 address
= PAGE_ALIGN(address
+4);
2128 vma_unlock_anon_vma(vma
);
2133 /* Somebody else might have raced and expanded it already */
2134 if (address
> vma
->vm_end
) {
2135 unsigned long size
, grow
;
2137 size
= address
- vma
->vm_start
;
2138 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
2141 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
2142 error
= acct_stack_growth(vma
, size
, grow
);
2145 * vma_gap_update() doesn't support concurrent
2146 * updates, but we only hold a shared mmap_sem
2147 * lock here, so we need to protect against
2148 * concurrent vma expansions.
2149 * vma_lock_anon_vma() doesn't help here, as
2150 * we don't guarantee that all growable vmas
2151 * in a mm share the same root anon vma.
2152 * So, we reuse mm->page_table_lock to guard
2153 * against concurrent vma expansions.
2155 spin_lock(&vma
->vm_mm
->page_table_lock
);
2156 anon_vma_interval_tree_pre_update_vma(vma
);
2157 vma
->vm_end
= address
;
2158 anon_vma_interval_tree_post_update_vma(vma
);
2160 vma_gap_update(vma
->vm_next
);
2162 vma
->vm_mm
->highest_vm_end
= address
;
2163 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2165 perf_event_mmap(vma
);
2169 vma_unlock_anon_vma(vma
);
2170 khugepaged_enter_vma_merge(vma
);
2171 validate_mm(vma
->vm_mm
);
2174 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2177 * vma is the first one with address < vma->vm_start. Have to extend vma.
2179 int expand_downwards(struct vm_area_struct
*vma
,
2180 unsigned long address
)
2185 * We must make sure the anon_vma is allocated
2186 * so that the anon_vma locking is not a noop.
2188 if (unlikely(anon_vma_prepare(vma
)))
2191 address
&= PAGE_MASK
;
2192 error
= security_mmap_addr(address
);
2196 vma_lock_anon_vma(vma
);
2199 * vma->vm_start/vm_end cannot change under us because the caller
2200 * is required to hold the mmap_sem in read mode. We need the
2201 * anon_vma lock to serialize against concurrent expand_stacks.
2204 /* Somebody else might have raced and expanded it already */
2205 if (address
< vma
->vm_start
) {
2206 unsigned long size
, grow
;
2208 size
= vma
->vm_end
- address
;
2209 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
2212 if (grow
<= vma
->vm_pgoff
) {
2213 error
= acct_stack_growth(vma
, size
, grow
);
2216 * vma_gap_update() doesn't support concurrent
2217 * updates, but we only hold a shared mmap_sem
2218 * lock here, so we need to protect against
2219 * concurrent vma expansions.
2220 * vma_lock_anon_vma() doesn't help here, as
2221 * we don't guarantee that all growable vmas
2222 * in a mm share the same root anon vma.
2223 * So, we reuse mm->page_table_lock to guard
2224 * against concurrent vma expansions.
2226 spin_lock(&vma
->vm_mm
->page_table_lock
);
2227 anon_vma_interval_tree_pre_update_vma(vma
);
2228 vma
->vm_start
= address
;
2229 vma
->vm_pgoff
-= grow
;
2230 anon_vma_interval_tree_post_update_vma(vma
);
2231 vma_gap_update(vma
);
2232 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2234 perf_event_mmap(vma
);
2238 vma_unlock_anon_vma(vma
);
2239 khugepaged_enter_vma_merge(vma
);
2240 validate_mm(vma
->vm_mm
);
2245 * Note how expand_stack() refuses to expand the stack all the way to
2246 * abut the next virtual mapping, *unless* that mapping itself is also
2247 * a stack mapping. We want to leave room for a guard page, after all
2248 * (the guard page itself is not added here, that is done by the
2249 * actual page faulting logic)
2251 * This matches the behavior of the guard page logic (see mm/memory.c:
2252 * check_stack_guard_page()), which only allows the guard page to be
2253 * removed under these circumstances.
2255 #ifdef CONFIG_STACK_GROWSUP
2256 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2258 struct vm_area_struct
*next
;
2260 address
&= PAGE_MASK
;
2261 next
= vma
->vm_next
;
2262 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
) {
2263 if (!(next
->vm_flags
& VM_GROWSUP
))
2266 return expand_upwards(vma
, address
);
2269 struct vm_area_struct
*
2270 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2272 struct vm_area_struct
*vma
, *prev
;
2275 vma
= find_vma_prev(mm
, addr
, &prev
);
2276 if (vma
&& (vma
->vm_start
<= addr
))
2278 if (!prev
|| expand_stack(prev
, addr
))
2280 if (prev
->vm_flags
& VM_LOCKED
)
2281 __mlock_vma_pages_range(prev
, addr
, prev
->vm_end
, NULL
);
2285 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2287 struct vm_area_struct
*prev
;
2289 address
&= PAGE_MASK
;
2290 prev
= vma
->vm_prev
;
2291 if (prev
&& prev
->vm_end
== address
) {
2292 if (!(prev
->vm_flags
& VM_GROWSDOWN
))
2295 return expand_downwards(vma
, address
);
2298 struct vm_area_struct
*
2299 find_extend_vma(struct mm_struct
* mm
, unsigned long addr
)
2301 struct vm_area_struct
* vma
;
2302 unsigned long start
;
2305 vma
= find_vma(mm
,addr
);
2308 if (vma
->vm_start
<= addr
)
2310 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2312 start
= vma
->vm_start
;
2313 if (expand_stack(vma
, addr
))
2315 if (vma
->vm_flags
& VM_LOCKED
)
2316 __mlock_vma_pages_range(vma
, addr
, start
, NULL
);
2322 * Ok - we have the memory areas we should free on the vma list,
2323 * so release them, and do the vma updates.
2325 * Called with the mm semaphore held.
2327 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2329 unsigned long nr_accounted
= 0;
2331 /* Update high watermark before we lower total_vm */
2332 update_hiwater_vm(mm
);
2334 long nrpages
= vma_pages(vma
);
2336 if (vma
->vm_flags
& VM_ACCOUNT
)
2337 nr_accounted
+= nrpages
;
2338 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, -nrpages
);
2339 vma
= remove_vma(vma
);
2341 vm_unacct_memory(nr_accounted
);
2346 * Get rid of page table information in the indicated region.
2348 * Called with the mm semaphore held.
2350 static void unmap_region(struct mm_struct
*mm
,
2351 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
2352 unsigned long start
, unsigned long end
)
2354 struct vm_area_struct
*next
= prev
? prev
->vm_next
: mm
->mmap
;
2355 struct mmu_gather tlb
;
2358 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2359 update_hiwater_rss(mm
);
2360 unmap_vmas(&tlb
, vma
, start
, end
);
2361 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
2362 next
? next
->vm_start
: USER_PGTABLES_CEILING
);
2363 tlb_finish_mmu(&tlb
, start
, end
);
2367 * Create a list of vma's touched by the unmap, removing them from the mm's
2368 * vma list as we go..
2371 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2372 struct vm_area_struct
*prev
, unsigned long end
)
2374 struct vm_area_struct
**insertion_point
;
2375 struct vm_area_struct
*tail_vma
= NULL
;
2377 insertion_point
= (prev
? &prev
->vm_next
: &mm
->mmap
);
2378 vma
->vm_prev
= NULL
;
2380 vma_rb_erase(vma
, &mm
->mm_rb
);
2384 } while (vma
&& vma
->vm_start
< end
);
2385 *insertion_point
= vma
;
2387 vma
->vm_prev
= prev
;
2388 vma_gap_update(vma
);
2390 mm
->highest_vm_end
= prev
? prev
->vm_end
: 0;
2391 tail_vma
->vm_next
= NULL
;
2393 /* Kill the cache */
2394 vmacache_invalidate(mm
);
2398 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2399 * munmap path where it doesn't make sense to fail.
2401 static int __split_vma(struct mm_struct
* mm
, struct vm_area_struct
* vma
,
2402 unsigned long addr
, int new_below
)
2404 struct vm_area_struct
*new;
2407 if (is_vm_hugetlb_page(vma
) && (addr
&
2408 ~(huge_page_mask(hstate_vma(vma
)))))
2411 new = kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2415 /* most fields are the same, copy all, and then fixup */
2418 INIT_LIST_HEAD(&new->anon_vma_chain
);
2423 new->vm_start
= addr
;
2424 new->vm_pgoff
+= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
);
2427 err
= vma_dup_policy(vma
, new);
2431 if (anon_vma_clone(new, vma
))
2435 get_file(new->vm_file
);
2437 if (new->vm_ops
&& new->vm_ops
->open
)
2438 new->vm_ops
->open(new);
2441 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2442 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2444 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2450 /* Clean everything up if vma_adjust failed. */
2451 if (new->vm_ops
&& new->vm_ops
->close
)
2452 new->vm_ops
->close(new);
2455 unlink_anon_vmas(new);
2457 mpol_put(vma_policy(new));
2459 kmem_cache_free(vm_area_cachep
, new);
2465 * Split a vma into two pieces at address 'addr', a new vma is allocated
2466 * either for the first part or the tail.
2468 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2469 unsigned long addr
, int new_below
)
2471 if (mm
->map_count
>= sysctl_max_map_count
)
2474 return __split_vma(mm
, vma
, addr
, new_below
);
2477 /* Munmap is split into 2 main parts -- this part which finds
2478 * what needs doing, and the areas themselves, which do the
2479 * work. This now handles partial unmappings.
2480 * Jeremy Fitzhardinge <jeremy@goop.org>
2482 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
)
2485 struct vm_area_struct
*vma
, *prev
, *last
;
2487 if ((start
& ~PAGE_MASK
) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2490 if ((len
= PAGE_ALIGN(len
)) == 0)
2493 /* Find the first overlapping VMA */
2494 vma
= find_vma(mm
, start
);
2497 prev
= vma
->vm_prev
;
2498 /* we have start < vma->vm_end */
2500 /* if it doesn't overlap, we have nothing.. */
2502 if (vma
->vm_start
>= end
)
2506 * If we need to split any vma, do it now to save pain later.
2508 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2509 * unmapped vm_area_struct will remain in use: so lower split_vma
2510 * places tmp vma above, and higher split_vma places tmp vma below.
2512 if (start
> vma
->vm_start
) {
2516 * Make sure that map_count on return from munmap() will
2517 * not exceed its limit; but let map_count go just above
2518 * its limit temporarily, to help free resources as expected.
2520 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2523 error
= __split_vma(mm
, vma
, start
, 0);
2529 /* Does it split the last one? */
2530 last
= find_vma(mm
, end
);
2531 if (last
&& end
> last
->vm_start
) {
2532 int error
= __split_vma(mm
, last
, end
, 1);
2536 vma
= prev
? prev
->vm_next
: mm
->mmap
;
2539 * unlock any mlock()ed ranges before detaching vmas
2541 if (mm
->locked_vm
) {
2542 struct vm_area_struct
*tmp
= vma
;
2543 while (tmp
&& tmp
->vm_start
< end
) {
2544 if (tmp
->vm_flags
& VM_LOCKED
) {
2545 mm
->locked_vm
-= vma_pages(tmp
);
2546 munlock_vma_pages_all(tmp
);
2553 * Remove the vma's, and unmap the actual pages
2555 detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
);
2556 unmap_region(mm
, vma
, prev
, start
, end
);
2558 /* Fix up all other VM information */
2559 remove_vma_list(mm
, vma
);
2564 int vm_munmap(unsigned long start
, size_t len
)
2567 struct mm_struct
*mm
= current
->mm
;
2569 down_write(&mm
->mmap_sem
);
2570 ret
= do_munmap(mm
, start
, len
);
2571 up_write(&mm
->mmap_sem
);
2574 EXPORT_SYMBOL(vm_munmap
);
2576 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2578 profile_munmap(addr
);
2579 return vm_munmap(addr
, len
);
2582 static inline void verify_mm_writelocked(struct mm_struct
*mm
)
2584 #ifdef CONFIG_DEBUG_VM
2585 if (unlikely(down_read_trylock(&mm
->mmap_sem
))) {
2587 up_read(&mm
->mmap_sem
);
2593 * this is really a simplified "do_mmap". it only handles
2594 * anonymous maps. eventually we may be able to do some
2595 * brk-specific accounting here.
2597 static unsigned long do_brk(unsigned long addr
, unsigned long len
)
2599 struct mm_struct
* mm
= current
->mm
;
2600 struct vm_area_struct
* vma
, * prev
;
2601 unsigned long flags
;
2602 struct rb_node
** rb_link
, * rb_parent
;
2603 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2606 len
= PAGE_ALIGN(len
);
2610 flags
= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2612 error
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2613 if (error
& ~PAGE_MASK
)
2616 error
= mlock_future_check(mm
, mm
->def_flags
, len
);
2621 * mm->mmap_sem is required to protect against another thread
2622 * changing the mappings in case we sleep.
2624 verify_mm_writelocked(mm
);
2627 * Clear old maps. this also does some error checking for us
2630 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
2631 if (do_munmap(mm
, addr
, len
))
2636 /* Check against address space limits *after* clearing old maps... */
2637 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
))
2640 if (mm
->map_count
> sysctl_max_map_count
)
2643 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
2646 /* Can we just expand an old private anonymous mapping? */
2647 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
2648 NULL
, NULL
, pgoff
, NULL
);
2653 * create a vma struct for an anonymous mapping
2655 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2657 vm_unacct_memory(len
>> PAGE_SHIFT
);
2661 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2663 vma
->vm_start
= addr
;
2664 vma
->vm_end
= addr
+ len
;
2665 vma
->vm_pgoff
= pgoff
;
2666 vma
->vm_flags
= flags
;
2667 vma
->vm_page_prot
= vm_get_page_prot(flags
);
2668 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2670 perf_event_mmap(vma
);
2671 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2672 if (flags
& VM_LOCKED
)
2673 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
2674 vma
->vm_flags
|= VM_SOFTDIRTY
;
2678 unsigned long vm_brk(unsigned long addr
, unsigned long len
)
2680 struct mm_struct
*mm
= current
->mm
;
2684 down_write(&mm
->mmap_sem
);
2685 ret
= do_brk(addr
, len
);
2686 populate
= ((mm
->def_flags
& VM_LOCKED
) != 0);
2687 up_write(&mm
->mmap_sem
);
2689 mm_populate(addr
, len
);
2692 EXPORT_SYMBOL(vm_brk
);
2694 /* Release all mmaps. */
2695 void exit_mmap(struct mm_struct
*mm
)
2697 struct mmu_gather tlb
;
2698 struct vm_area_struct
*vma
;
2699 unsigned long nr_accounted
= 0;
2701 /* mm's last user has gone, and its about to be pulled down */
2702 mmu_notifier_release(mm
);
2704 if (mm
->locked_vm
) {
2707 if (vma
->vm_flags
& VM_LOCKED
)
2708 munlock_vma_pages_all(vma
);
2716 if (!vma
) /* Can happen if dup_mmap() received an OOM */
2721 tlb_gather_mmu(&tlb
, mm
, 0, -1);
2722 /* update_hiwater_rss(mm) here? but nobody should be looking */
2723 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2724 unmap_vmas(&tlb
, vma
, 0, -1);
2726 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, USER_PGTABLES_CEILING
);
2727 tlb_finish_mmu(&tlb
, 0, -1);
2730 * Walk the list again, actually closing and freeing it,
2731 * with preemption enabled, without holding any MM locks.
2734 if (vma
->vm_flags
& VM_ACCOUNT
)
2735 nr_accounted
+= vma_pages(vma
);
2736 vma
= remove_vma(vma
);
2738 vm_unacct_memory(nr_accounted
);
2740 WARN_ON(atomic_long_read(&mm
->nr_ptes
) >
2741 (FIRST_USER_ADDRESS
+PMD_SIZE
-1)>>PMD_SHIFT
);
2744 /* Insert vm structure into process list sorted by address
2745 * and into the inode's i_mmap tree. If vm_file is non-NULL
2746 * then i_mmap_mutex is taken here.
2748 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2750 struct vm_area_struct
*prev
;
2751 struct rb_node
**rb_link
, *rb_parent
;
2754 * The vm_pgoff of a purely anonymous vma should be irrelevant
2755 * until its first write fault, when page's anon_vma and index
2756 * are set. But now set the vm_pgoff it will almost certainly
2757 * end up with (unless mremap moves it elsewhere before that
2758 * first wfault), so /proc/pid/maps tells a consistent story.
2760 * By setting it to reflect the virtual start address of the
2761 * vma, merges and splits can happen in a seamless way, just
2762 * using the existing file pgoff checks and manipulations.
2763 * Similarly in do_mmap_pgoff and in do_brk.
2765 if (!vma
->vm_file
) {
2766 BUG_ON(vma
->anon_vma
);
2767 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2769 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
2770 &prev
, &rb_link
, &rb_parent
))
2772 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
2773 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
2776 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2781 * Copy the vma structure to a new location in the same mm,
2782 * prior to moving page table entries, to effect an mremap move.
2784 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
2785 unsigned long addr
, unsigned long len
, pgoff_t pgoff
,
2786 bool *need_rmap_locks
)
2788 struct vm_area_struct
*vma
= *vmap
;
2789 unsigned long vma_start
= vma
->vm_start
;
2790 struct mm_struct
*mm
= vma
->vm_mm
;
2791 struct vm_area_struct
*new_vma
, *prev
;
2792 struct rb_node
**rb_link
, *rb_parent
;
2793 bool faulted_in_anon_vma
= true;
2796 * If anonymous vma has not yet been faulted, update new pgoff
2797 * to match new location, to increase its chance of merging.
2799 if (unlikely(!vma
->vm_file
&& !vma
->anon_vma
)) {
2800 pgoff
= addr
>> PAGE_SHIFT
;
2801 faulted_in_anon_vma
= false;
2804 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
2805 return NULL
; /* should never get here */
2806 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
2807 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
));
2810 * Source vma may have been merged into new_vma
2812 if (unlikely(vma_start
>= new_vma
->vm_start
&&
2813 vma_start
< new_vma
->vm_end
)) {
2815 * The only way we can get a vma_merge with
2816 * self during an mremap is if the vma hasn't
2817 * been faulted in yet and we were allowed to
2818 * reset the dst vma->vm_pgoff to the
2819 * destination address of the mremap to allow
2820 * the merge to happen. mremap must change the
2821 * vm_pgoff linearity between src and dst vmas
2822 * (in turn preventing a vma_merge) to be
2823 * safe. It is only safe to keep the vm_pgoff
2824 * linear if there are no pages mapped yet.
2826 VM_BUG_ON(faulted_in_anon_vma
);
2827 *vmap
= vma
= new_vma
;
2829 *need_rmap_locks
= (new_vma
->vm_pgoff
<= vma
->vm_pgoff
);
2831 new_vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2834 new_vma
->vm_start
= addr
;
2835 new_vma
->vm_end
= addr
+ len
;
2836 new_vma
->vm_pgoff
= pgoff
;
2837 if (vma_dup_policy(vma
, new_vma
))
2839 INIT_LIST_HEAD(&new_vma
->anon_vma_chain
);
2840 if (anon_vma_clone(new_vma
, vma
))
2841 goto out_free_mempol
;
2842 if (new_vma
->vm_file
)
2843 get_file(new_vma
->vm_file
);
2844 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
2845 new_vma
->vm_ops
->open(new_vma
);
2846 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
2847 *need_rmap_locks
= false;
2853 mpol_put(vma_policy(new_vma
));
2855 kmem_cache_free(vm_area_cachep
, new_vma
);
2860 * Return true if the calling process may expand its vm space by the passed
2863 int may_expand_vm(struct mm_struct
*mm
, unsigned long npages
)
2865 unsigned long cur
= mm
->total_vm
; /* pages */
2868 lim
= rlimit(RLIMIT_AS
) >> PAGE_SHIFT
;
2870 if (cur
+ npages
> lim
)
2876 static int special_mapping_fault(struct vm_area_struct
*vma
,
2877 struct vm_fault
*vmf
)
2880 struct page
**pages
;
2883 * special mappings have no vm_file, and in that case, the mm
2884 * uses vm_pgoff internally. So we have to subtract it from here.
2885 * We are allowed to do this because we are the mm; do not copy
2886 * this code into drivers!
2888 pgoff
= vmf
->pgoff
- vma
->vm_pgoff
;
2890 for (pages
= vma
->vm_private_data
; pgoff
&& *pages
; ++pages
)
2894 struct page
*page
= *pages
;
2900 return VM_FAULT_SIGBUS
;
2904 * Having a close hook prevents vma merging regardless of flags.
2906 static void special_mapping_close(struct vm_area_struct
*vma
)
2910 static const struct vm_operations_struct special_mapping_vmops
= {
2911 .close
= special_mapping_close
,
2912 .fault
= special_mapping_fault
,
2916 * Called with mm->mmap_sem held for writing.
2917 * Insert a new vma covering the given region, with the given flags.
2918 * Its pages are supplied by the given array of struct page *.
2919 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2920 * The region past the last page supplied will always produce SIGBUS.
2921 * The array pointer and the pages it points to are assumed to stay alive
2922 * for as long as this mapping might exist.
2924 struct vm_area_struct
*_install_special_mapping(struct mm_struct
*mm
,
2925 unsigned long addr
, unsigned long len
,
2926 unsigned long vm_flags
, struct page
**pages
)
2929 struct vm_area_struct
*vma
;
2931 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2932 if (unlikely(vma
== NULL
))
2933 return ERR_PTR(-ENOMEM
);
2935 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2937 vma
->vm_start
= addr
;
2938 vma
->vm_end
= addr
+ len
;
2940 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
| VM_SOFTDIRTY
;
2941 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
2943 vma
->vm_ops
= &special_mapping_vmops
;
2944 vma
->vm_private_data
= pages
;
2946 ret
= insert_vm_struct(mm
, vma
);
2950 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2952 perf_event_mmap(vma
);
2957 kmem_cache_free(vm_area_cachep
, vma
);
2958 return ERR_PTR(ret
);
2961 int install_special_mapping(struct mm_struct
*mm
,
2962 unsigned long addr
, unsigned long len
,
2963 unsigned long vm_flags
, struct page
**pages
)
2965 struct vm_area_struct
*vma
= _install_special_mapping(mm
,
2966 addr
, len
, vm_flags
, pages
);
2969 return PTR_ERR(vma
);
2973 static DEFINE_MUTEX(mm_all_locks_mutex
);
2975 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
2977 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
2979 * The LSB of head.next can't change from under us
2980 * because we hold the mm_all_locks_mutex.
2982 down_write_nest_lock(&anon_vma
->root
->rwsem
, &mm
->mmap_sem
);
2984 * We can safely modify head.next after taking the
2985 * anon_vma->root->rwsem. If some other vma in this mm shares
2986 * the same anon_vma we won't take it again.
2988 * No need of atomic instructions here, head.next
2989 * can't change from under us thanks to the
2990 * anon_vma->root->rwsem.
2992 if (__test_and_set_bit(0, (unsigned long *)
2993 &anon_vma
->root
->rb_root
.rb_node
))
2998 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
3000 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3002 * AS_MM_ALL_LOCKS can't change from under us because
3003 * we hold the mm_all_locks_mutex.
3005 * Operations on ->flags have to be atomic because
3006 * even if AS_MM_ALL_LOCKS is stable thanks to the
3007 * mm_all_locks_mutex, there may be other cpus
3008 * changing other bitflags in parallel to us.
3010 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
3012 mutex_lock_nest_lock(&mapping
->i_mmap_mutex
, &mm
->mmap_sem
);
3017 * This operation locks against the VM for all pte/vma/mm related
3018 * operations that could ever happen on a certain mm. This includes
3019 * vmtruncate, try_to_unmap, and all page faults.
3021 * The caller must take the mmap_sem in write mode before calling
3022 * mm_take_all_locks(). The caller isn't allowed to release the
3023 * mmap_sem until mm_drop_all_locks() returns.
3025 * mmap_sem in write mode is required in order to block all operations
3026 * that could modify pagetables and free pages without need of
3027 * altering the vma layout (for example populate_range() with
3028 * nonlinear vmas). It's also needed in write mode to avoid new
3029 * anon_vmas to be associated with existing vmas.
3031 * A single task can't take more than one mm_take_all_locks() in a row
3032 * or it would deadlock.
3034 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3035 * mapping->flags avoid to take the same lock twice, if more than one
3036 * vma in this mm is backed by the same anon_vma or address_space.
3038 * We can take all the locks in random order because the VM code
3039 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3040 * takes more than one of them in a row. Secondly we're protected
3041 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3043 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3044 * that may have to take thousand of locks.
3046 * mm_take_all_locks() can fail if it's interrupted by signals.
3048 int mm_take_all_locks(struct mm_struct
*mm
)
3050 struct vm_area_struct
*vma
;
3051 struct anon_vma_chain
*avc
;
3053 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3055 mutex_lock(&mm_all_locks_mutex
);
3057 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3058 if (signal_pending(current
))
3060 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3061 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3064 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3065 if (signal_pending(current
))
3068 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3069 vm_lock_anon_vma(mm
, avc
->anon_vma
);
3075 mm_drop_all_locks(mm
);
3079 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
3081 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3083 * The LSB of head.next can't change to 0 from under
3084 * us because we hold the mm_all_locks_mutex.
3086 * We must however clear the bitflag before unlocking
3087 * the vma so the users using the anon_vma->rb_root will
3088 * never see our bitflag.
3090 * No need of atomic instructions here, head.next
3091 * can't change from under us until we release the
3092 * anon_vma->root->rwsem.
3094 if (!__test_and_clear_bit(0, (unsigned long *)
3095 &anon_vma
->root
->rb_root
.rb_node
))
3097 anon_vma_unlock_write(anon_vma
);
3101 static void vm_unlock_mapping(struct address_space
*mapping
)
3103 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3105 * AS_MM_ALL_LOCKS can't change to 0 from under us
3106 * because we hold the mm_all_locks_mutex.
3108 mutex_unlock(&mapping
->i_mmap_mutex
);
3109 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
3116 * The mmap_sem cannot be released by the caller until
3117 * mm_drop_all_locks() returns.
3119 void mm_drop_all_locks(struct mm_struct
*mm
)
3121 struct vm_area_struct
*vma
;
3122 struct anon_vma_chain
*avc
;
3124 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3125 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
3127 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3129 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3130 vm_unlock_anon_vma(avc
->anon_vma
);
3131 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3132 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
3135 mutex_unlock(&mm_all_locks_mutex
);
3139 * initialise the VMA slab
3141 void __init
mmap_init(void)
3145 ret
= percpu_counter_init(&vm_committed_as
, 0);
3150 * Initialise sysctl_user_reserve_kbytes.
3152 * This is intended to prevent a user from starting a single memory hogging
3153 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3156 * The default value is min(3% of free memory, 128MB)
3157 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3159 static int init_user_reserve(void)
3161 unsigned long free_kbytes
;
3163 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3165 sysctl_user_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 17);
3168 subsys_initcall(init_user_reserve
);
3171 * Initialise sysctl_admin_reserve_kbytes.
3173 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3174 * to log in and kill a memory hogging process.
3176 * Systems with more than 256MB will reserve 8MB, enough to recover
3177 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3178 * only reserve 3% of free pages by default.
3180 static int init_admin_reserve(void)
3182 unsigned long free_kbytes
;
3184 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3186 sysctl_admin_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 13);
3189 subsys_initcall(init_admin_reserve
);
3192 * Reinititalise user and admin reserves if memory is added or removed.
3194 * The default user reserve max is 128MB, and the default max for the
3195 * admin reserve is 8MB. These are usually, but not always, enough to
3196 * enable recovery from a memory hogging process using login/sshd, a shell,
3197 * and tools like top. It may make sense to increase or even disable the
3198 * reserve depending on the existence of swap or variations in the recovery
3199 * tools. So, the admin may have changed them.
3201 * If memory is added and the reserves have been eliminated or increased above
3202 * the default max, then we'll trust the admin.
3204 * If memory is removed and there isn't enough free memory, then we
3205 * need to reset the reserves.
3207 * Otherwise keep the reserve set by the admin.
3209 static int reserve_mem_notifier(struct notifier_block
*nb
,
3210 unsigned long action
, void *data
)
3212 unsigned long tmp
, free_kbytes
;
3216 /* Default max is 128MB. Leave alone if modified by operator. */
3217 tmp
= sysctl_user_reserve_kbytes
;
3218 if (0 < tmp
&& tmp
< (1UL << 17))
3219 init_user_reserve();
3221 /* Default max is 8MB. Leave alone if modified by operator. */
3222 tmp
= sysctl_admin_reserve_kbytes
;
3223 if (0 < tmp
&& tmp
< (1UL << 13))
3224 init_admin_reserve();
3228 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3230 if (sysctl_user_reserve_kbytes
> free_kbytes
) {
3231 init_user_reserve();
3232 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3233 sysctl_user_reserve_kbytes
);
3236 if (sysctl_admin_reserve_kbytes
> free_kbytes
) {
3237 init_admin_reserve();
3238 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3239 sysctl_admin_reserve_kbytes
);
3248 static struct notifier_block reserve_mem_nb
= {
3249 .notifier_call
= reserve_mem_notifier
,
3252 static int __meminit
init_reserve_notifier(void)
3254 if (register_hotmemory_notifier(&reserve_mem_nb
))
3255 printk("Failed registering memory add/remove notifier for admin reserve");
3259 subsys_initcall(init_reserve_notifier
);