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/shm.h>
14 #include <linux/mman.h>
15 #include <linux/pagemap.h>
16 #include <linux/swap.h>
17 #include <linux/syscalls.h>
18 #include <linux/capability.h>
19 #include <linux/init.h>
20 #include <linux/file.h>
22 #include <linux/personality.h>
23 #include <linux/security.h>
24 #include <linux/hugetlb.h>
25 #include <linux/profile.h>
26 #include <linux/export.h>
27 #include <linux/mount.h>
28 #include <linux/mempolicy.h>
29 #include <linux/rmap.h>
30 #include <linux/mmu_notifier.h>
31 #include <linux/perf_event.h>
32 #include <linux/audit.h>
33 #include <linux/khugepaged.h>
34 #include <linux/uprobes.h>
35 #include <linux/rbtree_augmented.h>
36 #include <linux/sched/sysctl.h>
37 #include <linux/notifier.h>
38 #include <linux/memory.h>
40 #include <asm/uaccess.h>
41 #include <asm/cacheflush.h>
43 #include <asm/mmu_context.h>
47 #ifndef arch_mmap_check
48 #define arch_mmap_check(addr, len, flags) (0)
51 #ifndef arch_rebalance_pgtables
52 #define arch_rebalance_pgtables(addr, len) (addr)
55 static void unmap_region(struct mm_struct
*mm
,
56 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
57 unsigned long start
, unsigned long end
);
59 /* description of effects of mapping type and prot in current implementation.
60 * this is due to the limited x86 page protection hardware. The expected
61 * behavior is in parens:
64 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
65 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
66 * w: (no) no w: (no) no w: (yes) yes w: (no) no
67 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
69 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
70 * w: (no) no w: (no) no w: (copy) copy w: (no) no
71 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
74 pgprot_t protection_map
[16] = {
75 __P000
, __P001
, __P010
, __P011
, __P100
, __P101
, __P110
, __P111
,
76 __S000
, __S001
, __S010
, __S011
, __S100
, __S101
, __S110
, __S111
79 pgprot_t
vm_get_page_prot(unsigned long vm_flags
)
81 return __pgprot(pgprot_val(protection_map
[vm_flags
&
82 (VM_READ
|VM_WRITE
|VM_EXEC
|VM_SHARED
)]) |
83 pgprot_val(arch_vm_get_page_prot(vm_flags
)));
85 EXPORT_SYMBOL(vm_get_page_prot
);
87 int sysctl_overcommit_memory __read_mostly
= OVERCOMMIT_GUESS
; /* heuristic overcommit */
88 int sysctl_overcommit_ratio __read_mostly
= 50; /* default is 50% */
89 unsigned long sysctl_overcommit_kbytes __read_mostly
;
90 int sysctl_max_map_count __read_mostly
= DEFAULT_MAX_MAP_COUNT
;
91 unsigned long sysctl_user_reserve_kbytes __read_mostly
= 1UL << 17; /* 128MB */
92 unsigned long sysctl_admin_reserve_kbytes __read_mostly
= 1UL << 13; /* 8MB */
94 * Make sure vm_committed_as in one cacheline and not cacheline shared with
95 * other variables. It can be updated by several CPUs frequently.
97 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp
;
100 * The global memory commitment made in the system can be a metric
101 * that can be used to drive ballooning decisions when Linux is hosted
102 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
103 * balancing memory across competing virtual machines that are hosted.
104 * Several metrics drive this policy engine including the guest reported
107 unsigned long vm_memory_committed(void)
109 return percpu_counter_read_positive(&vm_committed_as
);
111 EXPORT_SYMBOL_GPL(vm_memory_committed
);
114 * Check that a process has enough memory to allocate a new virtual
115 * mapping. 0 means there is enough memory for the allocation to
116 * succeed and -ENOMEM implies there is not.
118 * We currently support three overcommit policies, which are set via the
119 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
121 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
122 * Additional code 2002 Jul 20 by Robert Love.
124 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
126 * Note this is a helper function intended to be used by LSMs which
127 * wish to use this logic.
129 int __vm_enough_memory(struct mm_struct
*mm
, long pages
, int cap_sys_admin
)
131 unsigned long free
, allowed
, reserve
;
133 vm_acct_memory(pages
);
136 * Sometimes we want to use more memory than we have
138 if (sysctl_overcommit_memory
== OVERCOMMIT_ALWAYS
)
141 if (sysctl_overcommit_memory
== OVERCOMMIT_GUESS
) {
142 free
= global_page_state(NR_FREE_PAGES
);
143 free
+= global_page_state(NR_FILE_PAGES
);
146 * shmem pages shouldn't be counted as free in this
147 * case, they can't be purged, only swapped out, and
148 * that won't affect the overall amount of available
149 * memory in the system.
151 free
-= global_page_state(NR_SHMEM
);
153 free
+= get_nr_swap_pages();
156 * Any slabs which are created with the
157 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
158 * which are reclaimable, under pressure. The dentry
159 * cache and most inode caches should fall into this
161 free
+= global_page_state(NR_SLAB_RECLAIMABLE
);
164 * Leave reserved pages. The pages are not for anonymous pages.
166 if (free
<= totalreserve_pages
)
169 free
-= totalreserve_pages
;
172 * Reserve some for root
175 free
-= sysctl_admin_reserve_kbytes
>> (PAGE_SHIFT
- 10);
183 allowed
= vm_commit_limit();
185 * Reserve some for root
188 allowed
-= sysctl_admin_reserve_kbytes
>> (PAGE_SHIFT
- 10);
191 * Don't let a single process grow so big a user can't recover
194 reserve
= sysctl_user_reserve_kbytes
>> (PAGE_SHIFT
- 10);
195 allowed
-= min(mm
->total_vm
/ 32, reserve
);
198 if (percpu_counter_read_positive(&vm_committed_as
) < allowed
)
201 vm_unacct_memory(pages
);
207 * Requires inode->i_mapping->i_mmap_mutex
209 static void __remove_shared_vm_struct(struct vm_area_struct
*vma
,
210 struct file
*file
, struct address_space
*mapping
)
212 if (vma
->vm_flags
& VM_DENYWRITE
)
213 atomic_inc(&file_inode(file
)->i_writecount
);
214 if (vma
->vm_flags
& VM_SHARED
)
215 mapping
->i_mmap_writable
--;
217 flush_dcache_mmap_lock(mapping
);
218 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
219 list_del_init(&vma
->shared
.nonlinear
);
221 vma_interval_tree_remove(vma
, &mapping
->i_mmap
);
222 flush_dcache_mmap_unlock(mapping
);
226 * Unlink a file-based vm structure from its interval tree, to hide
227 * vma from rmap and vmtruncate before freeing its page tables.
229 void unlink_file_vma(struct vm_area_struct
*vma
)
231 struct file
*file
= vma
->vm_file
;
234 struct address_space
*mapping
= file
->f_mapping
;
235 mutex_lock(&mapping
->i_mmap_mutex
);
236 __remove_shared_vm_struct(vma
, file
, mapping
);
237 mutex_unlock(&mapping
->i_mmap_mutex
);
242 * Close a vm structure and free it, returning the next.
244 static struct vm_area_struct
*remove_vma(struct vm_area_struct
*vma
)
246 struct vm_area_struct
*next
= vma
->vm_next
;
249 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
250 vma
->vm_ops
->close(vma
);
253 mpol_put(vma_policy(vma
));
254 kmem_cache_free(vm_area_cachep
, vma
);
258 static unsigned long do_brk(unsigned long addr
, unsigned long len
);
260 SYSCALL_DEFINE1(brk
, unsigned long, brk
)
262 unsigned long rlim
, retval
;
263 unsigned long newbrk
, oldbrk
;
264 struct mm_struct
*mm
= current
->mm
;
265 unsigned long min_brk
;
268 down_write(&mm
->mmap_sem
);
270 #ifdef CONFIG_COMPAT_BRK
272 * CONFIG_COMPAT_BRK can still be overridden by setting
273 * randomize_va_space to 2, which will still cause mm->start_brk
274 * to be arbitrarily shifted
276 if (current
->brk_randomized
)
277 min_brk
= mm
->start_brk
;
279 min_brk
= mm
->end_data
;
281 min_brk
= mm
->start_brk
;
287 * Check against rlimit here. If this check is done later after the test
288 * of oldbrk with newbrk then it can escape the test and let the data
289 * segment grow beyond its set limit the in case where the limit is
290 * not page aligned -Ram Gupta
292 rlim
= rlimit(RLIMIT_DATA
);
293 if (rlim
< RLIM_INFINITY
&& (brk
- mm
->start_brk
) +
294 (mm
->end_data
- mm
->start_data
) > rlim
)
297 newbrk
= PAGE_ALIGN(brk
);
298 oldbrk
= PAGE_ALIGN(mm
->brk
);
299 if (oldbrk
== newbrk
)
302 /* Always allow shrinking brk. */
303 if (brk
<= mm
->brk
) {
304 if (!do_munmap(mm
, newbrk
, oldbrk
-newbrk
))
309 /* Check against existing mmap mappings. */
310 if (find_vma_intersection(mm
, oldbrk
, newbrk
+PAGE_SIZE
))
313 /* Ok, looks good - let it rip. */
314 if (do_brk(oldbrk
, newbrk
-oldbrk
) != oldbrk
)
319 populate
= newbrk
> oldbrk
&& (mm
->def_flags
& VM_LOCKED
) != 0;
320 up_write(&mm
->mmap_sem
);
322 mm_populate(oldbrk
, newbrk
- oldbrk
);
327 up_write(&mm
->mmap_sem
);
331 static long vma_compute_subtree_gap(struct vm_area_struct
*vma
)
333 unsigned long max
, subtree_gap
;
336 max
-= vma
->vm_prev
->vm_end
;
337 if (vma
->vm_rb
.rb_left
) {
338 subtree_gap
= rb_entry(vma
->vm_rb
.rb_left
,
339 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
340 if (subtree_gap
> max
)
343 if (vma
->vm_rb
.rb_right
) {
344 subtree_gap
= rb_entry(vma
->vm_rb
.rb_right
,
345 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
346 if (subtree_gap
> max
)
352 #ifdef CONFIG_DEBUG_VM_RB
353 static int browse_rb(struct rb_root
*root
)
355 int i
= 0, j
, bug
= 0;
356 struct rb_node
*nd
, *pn
= NULL
;
357 unsigned long prev
= 0, pend
= 0;
359 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
360 struct vm_area_struct
*vma
;
361 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
362 if (vma
->vm_start
< prev
) {
363 printk("vm_start %lx prev %lx\n", vma
->vm_start
, prev
);
366 if (vma
->vm_start
< pend
) {
367 printk("vm_start %lx pend %lx\n", vma
->vm_start
, pend
);
370 if (vma
->vm_start
> vma
->vm_end
) {
371 printk("vm_end %lx < vm_start %lx\n",
372 vma
->vm_end
, vma
->vm_start
);
375 if (vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
)) {
376 printk("free gap %lx, correct %lx\n",
378 vma_compute_subtree_gap(vma
));
383 prev
= vma
->vm_start
;
387 for (nd
= pn
; nd
; nd
= rb_prev(nd
))
390 printk("backwards %d, forwards %d\n", j
, i
);
396 static void validate_mm_rb(struct rb_root
*root
, struct vm_area_struct
*ignore
)
400 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
401 struct vm_area_struct
*vma
;
402 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
403 BUG_ON(vma
!= ignore
&&
404 vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
));
408 void validate_mm(struct mm_struct
*mm
)
412 unsigned long highest_address
= 0;
413 struct vm_area_struct
*vma
= mm
->mmap
;
415 struct anon_vma_chain
*avc
;
416 vma_lock_anon_vma(vma
);
417 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
418 anon_vma_interval_tree_verify(avc
);
419 vma_unlock_anon_vma(vma
);
420 highest_address
= vma
->vm_end
;
424 if (i
!= mm
->map_count
) {
425 printk("map_count %d vm_next %d\n", mm
->map_count
, i
);
428 if (highest_address
!= mm
->highest_vm_end
) {
429 printk("mm->highest_vm_end %lx, found %lx\n",
430 mm
->highest_vm_end
, highest_address
);
433 i
= browse_rb(&mm
->mm_rb
);
434 if (i
!= mm
->map_count
) {
435 printk("map_count %d rb %d\n", mm
->map_count
, i
);
441 #define validate_mm_rb(root, ignore) do { } while (0)
442 #define validate_mm(mm) do { } while (0)
445 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks
, struct vm_area_struct
, vm_rb
,
446 unsigned long, rb_subtree_gap
, vma_compute_subtree_gap
)
449 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
450 * vma->vm_prev->vm_end values changed, without modifying the vma's position
453 static void vma_gap_update(struct vm_area_struct
*vma
)
456 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
457 * function that does exacltly what we want.
459 vma_gap_callbacks_propagate(&vma
->vm_rb
, NULL
);
462 static inline void vma_rb_insert(struct vm_area_struct
*vma
,
463 struct rb_root
*root
)
465 /* All rb_subtree_gap values must be consistent prior to insertion */
466 validate_mm_rb(root
, NULL
);
468 rb_insert_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
471 static void vma_rb_erase(struct vm_area_struct
*vma
, struct rb_root
*root
)
474 * All rb_subtree_gap values must be consistent prior to erase,
475 * with the possible exception of the vma being erased.
477 validate_mm_rb(root
, vma
);
480 * Note rb_erase_augmented is a fairly large inline function,
481 * so make sure we instantiate it only once with our desired
482 * augmented rbtree callbacks.
484 rb_erase_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
488 * vma has some anon_vma assigned, and is already inserted on that
489 * anon_vma's interval trees.
491 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
492 * vma must be removed from the anon_vma's interval trees using
493 * anon_vma_interval_tree_pre_update_vma().
495 * After the update, the vma will be reinserted using
496 * anon_vma_interval_tree_post_update_vma().
498 * The entire update must be protected by exclusive mmap_sem and by
499 * the root anon_vma's mutex.
502 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct
*vma
)
504 struct anon_vma_chain
*avc
;
506 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
507 anon_vma_interval_tree_remove(avc
, &avc
->anon_vma
->rb_root
);
511 anon_vma_interval_tree_post_update_vma(struct vm_area_struct
*vma
)
513 struct anon_vma_chain
*avc
;
515 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
516 anon_vma_interval_tree_insert(avc
, &avc
->anon_vma
->rb_root
);
519 static int find_vma_links(struct mm_struct
*mm
, unsigned long addr
,
520 unsigned long end
, struct vm_area_struct
**pprev
,
521 struct rb_node
***rb_link
, struct rb_node
**rb_parent
)
523 struct rb_node
**__rb_link
, *__rb_parent
, *rb_prev
;
525 __rb_link
= &mm
->mm_rb
.rb_node
;
526 rb_prev
= __rb_parent
= NULL
;
529 struct vm_area_struct
*vma_tmp
;
531 __rb_parent
= *__rb_link
;
532 vma_tmp
= rb_entry(__rb_parent
, struct vm_area_struct
, vm_rb
);
534 if (vma_tmp
->vm_end
> addr
) {
535 /* Fail if an existing vma overlaps the area */
536 if (vma_tmp
->vm_start
< end
)
538 __rb_link
= &__rb_parent
->rb_left
;
540 rb_prev
= __rb_parent
;
541 __rb_link
= &__rb_parent
->rb_right
;
547 *pprev
= rb_entry(rb_prev
, struct vm_area_struct
, vm_rb
);
548 *rb_link
= __rb_link
;
549 *rb_parent
= __rb_parent
;
553 static unsigned long count_vma_pages_range(struct mm_struct
*mm
,
554 unsigned long addr
, unsigned long end
)
556 unsigned long nr_pages
= 0;
557 struct vm_area_struct
*vma
;
559 /* Find first overlaping mapping */
560 vma
= find_vma_intersection(mm
, addr
, end
);
564 nr_pages
= (min(end
, vma
->vm_end
) -
565 max(addr
, vma
->vm_start
)) >> PAGE_SHIFT
;
567 /* Iterate over the rest of the overlaps */
568 for (vma
= vma
->vm_next
; vma
; vma
= vma
->vm_next
) {
569 unsigned long overlap_len
;
571 if (vma
->vm_start
> end
)
574 overlap_len
= min(end
, vma
->vm_end
) - vma
->vm_start
;
575 nr_pages
+= overlap_len
>> PAGE_SHIFT
;
581 void __vma_link_rb(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
582 struct rb_node
**rb_link
, struct rb_node
*rb_parent
)
584 /* Update tracking information for the gap following the new vma. */
586 vma_gap_update(vma
->vm_next
);
588 mm
->highest_vm_end
= vma
->vm_end
;
591 * vma->vm_prev wasn't known when we followed the rbtree to find the
592 * correct insertion point for that vma. As a result, we could not
593 * update the vma vm_rb parents rb_subtree_gap values on the way down.
594 * So, we first insert the vma with a zero rb_subtree_gap value
595 * (to be consistent with what we did on the way down), and then
596 * immediately update the gap to the correct value. Finally we
597 * rebalance the rbtree after all augmented values have been set.
599 rb_link_node(&vma
->vm_rb
, rb_parent
, rb_link
);
600 vma
->rb_subtree_gap
= 0;
602 vma_rb_insert(vma
, &mm
->mm_rb
);
605 static void __vma_link_file(struct vm_area_struct
*vma
)
611 struct address_space
*mapping
= file
->f_mapping
;
613 if (vma
->vm_flags
& VM_DENYWRITE
)
614 atomic_dec(&file_inode(file
)->i_writecount
);
615 if (vma
->vm_flags
& VM_SHARED
)
616 mapping
->i_mmap_writable
++;
618 flush_dcache_mmap_lock(mapping
);
619 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
620 vma_nonlinear_insert(vma
, &mapping
->i_mmap_nonlinear
);
622 vma_interval_tree_insert(vma
, &mapping
->i_mmap
);
623 flush_dcache_mmap_unlock(mapping
);
628 __vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
629 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
630 struct rb_node
*rb_parent
)
632 __vma_link_list(mm
, vma
, prev
, rb_parent
);
633 __vma_link_rb(mm
, vma
, rb_link
, rb_parent
);
636 static void vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
637 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
638 struct rb_node
*rb_parent
)
640 struct address_space
*mapping
= NULL
;
643 mapping
= vma
->vm_file
->f_mapping
;
646 mutex_lock(&mapping
->i_mmap_mutex
);
648 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
649 __vma_link_file(vma
);
652 mutex_unlock(&mapping
->i_mmap_mutex
);
659 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
660 * mm's list and rbtree. It has already been inserted into the interval tree.
662 static void __insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
664 struct vm_area_struct
*prev
;
665 struct rb_node
**rb_link
, *rb_parent
;
667 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
668 &prev
, &rb_link
, &rb_parent
))
670 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
675 __vma_unlink(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
676 struct vm_area_struct
*prev
)
678 struct vm_area_struct
*next
;
680 vma_rb_erase(vma
, &mm
->mm_rb
);
681 prev
->vm_next
= next
= vma
->vm_next
;
683 next
->vm_prev
= prev
;
684 if (mm
->mmap_cache
== vma
)
685 mm
->mmap_cache
= prev
;
689 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
690 * is already present in an i_mmap tree without adjusting the tree.
691 * The following helper function should be used when such adjustments
692 * are necessary. The "insert" vma (if any) is to be inserted
693 * before we drop the necessary locks.
695 int vma_adjust(struct vm_area_struct
*vma
, unsigned long start
,
696 unsigned long end
, pgoff_t pgoff
, struct vm_area_struct
*insert
)
698 struct mm_struct
*mm
= vma
->vm_mm
;
699 struct vm_area_struct
*next
= vma
->vm_next
;
700 struct vm_area_struct
*importer
= NULL
;
701 struct address_space
*mapping
= NULL
;
702 struct rb_root
*root
= NULL
;
703 struct anon_vma
*anon_vma
= NULL
;
704 struct file
*file
= vma
->vm_file
;
705 bool start_changed
= false, end_changed
= false;
706 long adjust_next
= 0;
709 if (next
&& !insert
) {
710 struct vm_area_struct
*exporter
= NULL
;
712 if (end
>= next
->vm_end
) {
714 * vma expands, overlapping all the next, and
715 * perhaps the one after too (mprotect case 6).
717 again
: remove_next
= 1 + (end
> next
->vm_end
);
721 } else if (end
> next
->vm_start
) {
723 * vma expands, overlapping part of the next:
724 * mprotect case 5 shifting the boundary up.
726 adjust_next
= (end
- next
->vm_start
) >> PAGE_SHIFT
;
729 } else if (end
< vma
->vm_end
) {
731 * vma shrinks, and !insert tells it's not
732 * split_vma inserting another: so it must be
733 * mprotect case 4 shifting the boundary down.
735 adjust_next
= - ((vma
->vm_end
- end
) >> PAGE_SHIFT
);
741 * Easily overlooked: when mprotect shifts the boundary,
742 * make sure the expanding vma has anon_vma set if the
743 * shrinking vma had, to cover any anon pages imported.
745 if (exporter
&& exporter
->anon_vma
&& !importer
->anon_vma
) {
746 if (anon_vma_clone(importer
, exporter
))
748 importer
->anon_vma
= exporter
->anon_vma
;
753 mapping
= file
->f_mapping
;
754 if (!(vma
->vm_flags
& VM_NONLINEAR
)) {
755 root
= &mapping
->i_mmap
;
756 uprobe_munmap(vma
, vma
->vm_start
, vma
->vm_end
);
759 uprobe_munmap(next
, next
->vm_start
,
763 mutex_lock(&mapping
->i_mmap_mutex
);
766 * Put into interval tree now, so instantiated pages
767 * are visible to arm/parisc __flush_dcache_page
768 * throughout; but we cannot insert into address
769 * space until vma start or end is updated.
771 __vma_link_file(insert
);
775 vma_adjust_trans_huge(vma
, start
, end
, adjust_next
);
777 anon_vma
= vma
->anon_vma
;
778 if (!anon_vma
&& adjust_next
)
779 anon_vma
= next
->anon_vma
;
781 VM_BUG_ON(adjust_next
&& next
->anon_vma
&&
782 anon_vma
!= next
->anon_vma
);
783 anon_vma_lock_write(anon_vma
);
784 anon_vma_interval_tree_pre_update_vma(vma
);
786 anon_vma_interval_tree_pre_update_vma(next
);
790 flush_dcache_mmap_lock(mapping
);
791 vma_interval_tree_remove(vma
, root
);
793 vma_interval_tree_remove(next
, root
);
796 if (start
!= vma
->vm_start
) {
797 vma
->vm_start
= start
;
798 start_changed
= true;
800 if (end
!= vma
->vm_end
) {
804 vma
->vm_pgoff
= pgoff
;
806 next
->vm_start
+= adjust_next
<< PAGE_SHIFT
;
807 next
->vm_pgoff
+= adjust_next
;
812 vma_interval_tree_insert(next
, root
);
813 vma_interval_tree_insert(vma
, root
);
814 flush_dcache_mmap_unlock(mapping
);
819 * vma_merge has merged next into vma, and needs
820 * us to remove next before dropping the locks.
822 __vma_unlink(mm
, next
, vma
);
824 __remove_shared_vm_struct(next
, file
, mapping
);
827 * split_vma has split insert from vma, and needs
828 * us to insert it before dropping the locks
829 * (it may either follow vma or precede it).
831 __insert_vm_struct(mm
, insert
);
837 mm
->highest_vm_end
= end
;
838 else if (!adjust_next
)
839 vma_gap_update(next
);
844 anon_vma_interval_tree_post_update_vma(vma
);
846 anon_vma_interval_tree_post_update_vma(next
);
847 anon_vma_unlock_write(anon_vma
);
850 mutex_unlock(&mapping
->i_mmap_mutex
);
861 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
865 anon_vma_merge(vma
, next
);
867 mpol_put(vma_policy(next
));
868 kmem_cache_free(vm_area_cachep
, next
);
870 * In mprotect's case 6 (see comments on vma_merge),
871 * we must remove another next too. It would clutter
872 * up the code too much to do both in one go.
875 if (remove_next
== 2)
878 vma_gap_update(next
);
880 mm
->highest_vm_end
= end
;
891 * If the vma has a ->close operation then the driver probably needs to release
892 * per-vma resources, so we don't attempt to merge those.
894 static inline int is_mergeable_vma(struct vm_area_struct
*vma
,
895 struct file
*file
, unsigned long vm_flags
)
897 if (vma
->vm_flags
^ vm_flags
)
899 if (vma
->vm_file
!= file
)
901 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
906 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
907 struct anon_vma
*anon_vma2
,
908 struct vm_area_struct
*vma
)
911 * The list_is_singular() test is to avoid merging VMA cloned from
912 * parents. This can improve scalability caused by anon_vma lock.
914 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
915 list_is_singular(&vma
->anon_vma_chain
)))
917 return anon_vma1
== anon_vma2
;
921 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
922 * in front of (at a lower virtual address and file offset than) the vma.
924 * We cannot merge two vmas if they have differently assigned (non-NULL)
925 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
927 * We don't check here for the merged mmap wrapping around the end of pagecache
928 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
929 * wrap, nor mmaps which cover the final page at index -1UL.
932 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
933 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
935 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
936 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
937 if (vma
->vm_pgoff
== vm_pgoff
)
944 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
945 * beyond (at a higher virtual address and file offset than) the vma.
947 * We cannot merge two vmas if they have differently assigned (non-NULL)
948 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
951 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
952 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
954 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
955 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
957 vm_pglen
= vma_pages(vma
);
958 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
965 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
966 * whether that can be merged with its predecessor or its successor.
967 * Or both (it neatly fills a hole).
969 * In most cases - when called for mmap, brk or mremap - [addr,end) is
970 * certain not to be mapped by the time vma_merge is called; but when
971 * called for mprotect, it is certain to be already mapped (either at
972 * an offset within prev, or at the start of next), and the flags of
973 * this area are about to be changed to vm_flags - and the no-change
974 * case has already been eliminated.
976 * The following mprotect cases have to be considered, where AAAA is
977 * the area passed down from mprotect_fixup, never extending beyond one
978 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
980 * AAAA AAAA AAAA AAAA
981 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
982 * cannot merge might become might become might become
983 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
984 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
985 * mremap move: PPPPNNNNNNNN 8
987 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
988 * might become case 1 below case 2 below case 3 below
990 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
991 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
993 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
994 struct vm_area_struct
*prev
, unsigned long addr
,
995 unsigned long end
, unsigned long vm_flags
,
996 struct anon_vma
*anon_vma
, struct file
*file
,
997 pgoff_t pgoff
, struct mempolicy
*policy
)
999 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
1000 struct vm_area_struct
*area
, *next
;
1004 * We later require that vma->vm_flags == vm_flags,
1005 * so this tests vma->vm_flags & VM_SPECIAL, too.
1007 if (vm_flags
& VM_SPECIAL
)
1011 next
= prev
->vm_next
;
1015 if (next
&& next
->vm_end
== end
) /* cases 6, 7, 8 */
1016 next
= next
->vm_next
;
1019 * Can it merge with the predecessor?
1021 if (prev
&& prev
->vm_end
== addr
&&
1022 mpol_equal(vma_policy(prev
), policy
) &&
1023 can_vma_merge_after(prev
, vm_flags
,
1024 anon_vma
, file
, pgoff
)) {
1026 * OK, it can. Can we now merge in the successor as well?
1028 if (next
&& end
== next
->vm_start
&&
1029 mpol_equal(policy
, vma_policy(next
)) &&
1030 can_vma_merge_before(next
, vm_flags
,
1031 anon_vma
, file
, pgoff
+pglen
) &&
1032 is_mergeable_anon_vma(prev
->anon_vma
,
1033 next
->anon_vma
, NULL
)) {
1035 err
= vma_adjust(prev
, prev
->vm_start
,
1036 next
->vm_end
, prev
->vm_pgoff
, NULL
);
1037 } else /* cases 2, 5, 7 */
1038 err
= vma_adjust(prev
, prev
->vm_start
,
1039 end
, prev
->vm_pgoff
, NULL
);
1042 khugepaged_enter_vma_merge(prev
);
1047 * Can this new request be merged in front of next?
1049 if (next
&& end
== next
->vm_start
&&
1050 mpol_equal(policy
, vma_policy(next
)) &&
1051 can_vma_merge_before(next
, vm_flags
,
1052 anon_vma
, file
, pgoff
+pglen
)) {
1053 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
1054 err
= vma_adjust(prev
, prev
->vm_start
,
1055 addr
, prev
->vm_pgoff
, NULL
);
1056 else /* cases 3, 8 */
1057 err
= vma_adjust(area
, addr
, next
->vm_end
,
1058 next
->vm_pgoff
- pglen
, NULL
);
1061 khugepaged_enter_vma_merge(area
);
1069 * Rough compatbility check to quickly see if it's even worth looking
1070 * at sharing an anon_vma.
1072 * They need to have the same vm_file, and the flags can only differ
1073 * in things that mprotect may change.
1075 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1076 * we can merge the two vma's. For example, we refuse to merge a vma if
1077 * there is a vm_ops->close() function, because that indicates that the
1078 * driver is doing some kind of reference counting. But that doesn't
1079 * really matter for the anon_vma sharing case.
1081 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1083 return a
->vm_end
== b
->vm_start
&&
1084 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
1085 a
->vm_file
== b
->vm_file
&&
1086 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_READ
|VM_WRITE
|VM_EXEC
)) &&
1087 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
1091 * Do some basic sanity checking to see if we can re-use the anon_vma
1092 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1093 * the same as 'old', the other will be the new one that is trying
1094 * to share the anon_vma.
1096 * NOTE! This runs with mm_sem held for reading, so it is possible that
1097 * the anon_vma of 'old' is concurrently in the process of being set up
1098 * by another page fault trying to merge _that_. But that's ok: if it
1099 * is being set up, that automatically means that it will be a singleton
1100 * acceptable for merging, so we can do all of this optimistically. But
1101 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1103 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1104 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1105 * is to return an anon_vma that is "complex" due to having gone through
1108 * We also make sure that the two vma's are compatible (adjacent,
1109 * and with the same memory policies). That's all stable, even with just
1110 * a read lock on the mm_sem.
1112 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1114 if (anon_vma_compatible(a
, b
)) {
1115 struct anon_vma
*anon_vma
= ACCESS_ONCE(old
->anon_vma
);
1117 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
1124 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1125 * neighbouring vmas for a suitable anon_vma, before it goes off
1126 * to allocate a new anon_vma. It checks because a repetitive
1127 * sequence of mprotects and faults may otherwise lead to distinct
1128 * anon_vmas being allocated, preventing vma merge in subsequent
1131 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
1133 struct anon_vma
*anon_vma
;
1134 struct vm_area_struct
*near
;
1136 near
= vma
->vm_next
;
1140 anon_vma
= reusable_anon_vma(near
, vma
, near
);
1144 near
= vma
->vm_prev
;
1148 anon_vma
= reusable_anon_vma(near
, near
, vma
);
1153 * There's no absolute need to look only at touching neighbours:
1154 * we could search further afield for "compatible" anon_vmas.
1155 * But it would probably just be a waste of time searching,
1156 * or lead to too many vmas hanging off the same anon_vma.
1157 * We're trying to allow mprotect remerging later on,
1158 * not trying to minimize memory used for anon_vmas.
1163 #ifdef CONFIG_PROC_FS
1164 void vm_stat_account(struct mm_struct
*mm
, unsigned long flags
,
1165 struct file
*file
, long pages
)
1167 const unsigned long stack_flags
1168 = VM_STACK_FLAGS
& (VM_GROWSUP
|VM_GROWSDOWN
);
1170 mm
->total_vm
+= pages
;
1173 mm
->shared_vm
+= pages
;
1174 if ((flags
& (VM_EXEC
|VM_WRITE
)) == VM_EXEC
)
1175 mm
->exec_vm
+= pages
;
1176 } else if (flags
& stack_flags
)
1177 mm
->stack_vm
+= pages
;
1179 #endif /* CONFIG_PROC_FS */
1182 * If a hint addr is less than mmap_min_addr change hint to be as
1183 * low as possible but still greater than mmap_min_addr
1185 static inline unsigned long round_hint_to_min(unsigned long hint
)
1188 if (((void *)hint
!= NULL
) &&
1189 (hint
< mmap_min_addr
))
1190 return PAGE_ALIGN(mmap_min_addr
);
1194 static inline int mlock_future_check(struct mm_struct
*mm
,
1195 unsigned long flags
,
1198 unsigned long locked
, lock_limit
;
1200 /* mlock MCL_FUTURE? */
1201 if (flags
& VM_LOCKED
) {
1202 locked
= len
>> PAGE_SHIFT
;
1203 locked
+= mm
->locked_vm
;
1204 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1205 lock_limit
>>= PAGE_SHIFT
;
1206 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1213 * The caller must hold down_write(¤t->mm->mmap_sem).
1216 unsigned long do_mmap_pgoff(struct file
*file
, unsigned long addr
,
1217 unsigned long len
, unsigned long prot
,
1218 unsigned long flags
, unsigned long pgoff
,
1219 unsigned long *populate
)
1221 struct mm_struct
* mm
= current
->mm
;
1222 vm_flags_t vm_flags
;
1227 * Does the application expect PROT_READ to imply PROT_EXEC?
1229 * (the exception is when the underlying filesystem is noexec
1230 * mounted, in which case we dont add PROT_EXEC.)
1232 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1233 if (!(file
&& (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
)))
1239 if (!(flags
& MAP_FIXED
))
1240 addr
= round_hint_to_min(addr
);
1242 /* Careful about overflows.. */
1243 len
= PAGE_ALIGN(len
);
1247 /* offset overflow? */
1248 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1251 /* Too many mappings? */
1252 if (mm
->map_count
> sysctl_max_map_count
)
1255 /* Obtain the address to map to. we verify (or select) it and ensure
1256 * that it represents a valid section of the address space.
1258 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1259 if (addr
& ~PAGE_MASK
)
1262 /* Do simple checking here so the lower-level routines won't have
1263 * to. we assume access permissions have been handled by the open
1264 * of the memory object, so we don't do any here.
1266 vm_flags
= calc_vm_prot_bits(prot
) | calc_vm_flag_bits(flags
) |
1267 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1269 if (flags
& MAP_LOCKED
)
1270 if (!can_do_mlock())
1273 if (mlock_future_check(mm
, vm_flags
, len
))
1277 struct inode
*inode
= file_inode(file
);
1279 switch (flags
& MAP_TYPE
) {
1281 if ((prot
&PROT_WRITE
) && !(file
->f_mode
&FMODE_WRITE
))
1285 * Make sure we don't allow writing to an append-only
1288 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1292 * Make sure there are no mandatory locks on the file.
1294 if (locks_verify_locked(inode
))
1297 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1298 if (!(file
->f_mode
& FMODE_WRITE
))
1299 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1303 if (!(file
->f_mode
& FMODE_READ
))
1305 if (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
) {
1306 if (vm_flags
& VM_EXEC
)
1308 vm_flags
&= ~VM_MAYEXEC
;
1311 if (!file
->f_op
->mmap
)
1313 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1321 switch (flags
& MAP_TYPE
) {
1323 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1329 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1333 * Set pgoff according to addr for anon_vma.
1335 pgoff
= addr
>> PAGE_SHIFT
;
1343 * Set 'VM_NORESERVE' if we should not account for the
1344 * memory use of this mapping.
1346 if (flags
& MAP_NORESERVE
) {
1347 /* We honor MAP_NORESERVE if allowed to overcommit */
1348 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1349 vm_flags
|= VM_NORESERVE
;
1351 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1352 if (file
&& is_file_hugepages(file
))
1353 vm_flags
|= VM_NORESERVE
;
1356 addr
= mmap_region(file
, addr
, len
, vm_flags
, pgoff
);
1357 if (!IS_ERR_VALUE(addr
) &&
1358 ((vm_flags
& VM_LOCKED
) ||
1359 (flags
& (MAP_POPULATE
| MAP_NONBLOCK
)) == MAP_POPULATE
))
1364 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1365 unsigned long, prot
, unsigned long, flags
,
1366 unsigned long, fd
, unsigned long, pgoff
)
1368 struct file
*file
= NULL
;
1369 unsigned long retval
= -EBADF
;
1371 if (!(flags
& MAP_ANONYMOUS
)) {
1372 audit_mmap_fd(fd
, flags
);
1376 if (is_file_hugepages(file
))
1377 len
= ALIGN(len
, huge_page_size(hstate_file(file
)));
1379 if (unlikely(flags
& MAP_HUGETLB
&& !is_file_hugepages(file
)))
1381 } else if (flags
& MAP_HUGETLB
) {
1382 struct user_struct
*user
= NULL
;
1385 hs
= hstate_sizelog((flags
>> MAP_HUGE_SHIFT
) & SHM_HUGE_MASK
);
1389 len
= ALIGN(len
, huge_page_size(hs
));
1391 * VM_NORESERVE is used because the reservations will be
1392 * taken when vm_ops->mmap() is called
1393 * A dummy user value is used because we are not locking
1394 * memory so no accounting is necessary
1396 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, len
,
1398 &user
, HUGETLB_ANONHUGE_INODE
,
1399 (flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1401 return PTR_ERR(file
);
1404 flags
&= ~(MAP_EXECUTABLE
| MAP_DENYWRITE
);
1406 retval
= vm_mmap_pgoff(file
, addr
, len
, prot
, flags
, pgoff
);
1414 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1415 struct mmap_arg_struct
{
1419 unsigned long flags
;
1421 unsigned long offset
;
1424 SYSCALL_DEFINE1(old_mmap
, struct mmap_arg_struct __user
*, arg
)
1426 struct mmap_arg_struct a
;
1428 if (copy_from_user(&a
, arg
, sizeof(a
)))
1430 if (a
.offset
& ~PAGE_MASK
)
1433 return sys_mmap_pgoff(a
.addr
, a
.len
, a
.prot
, a
.flags
, a
.fd
,
1434 a
.offset
>> PAGE_SHIFT
);
1436 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1439 * Some shared mappigns will want the pages marked read-only
1440 * to track write events. If so, we'll downgrade vm_page_prot
1441 * to the private version (using protection_map[] without the
1444 int vma_wants_writenotify(struct vm_area_struct
*vma
)
1446 vm_flags_t vm_flags
= vma
->vm_flags
;
1448 /* If it was private or non-writable, the write bit is already clear */
1449 if ((vm_flags
& (VM_WRITE
|VM_SHARED
)) != ((VM_WRITE
|VM_SHARED
)))
1452 /* The backer wishes to know when pages are first written to? */
1453 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
)
1456 /* The open routine did something to the protections already? */
1457 if (pgprot_val(vma
->vm_page_prot
) !=
1458 pgprot_val(vm_get_page_prot(vm_flags
)))
1461 /* Specialty mapping? */
1462 if (vm_flags
& VM_PFNMAP
)
1465 /* Can the mapping track the dirty pages? */
1466 return vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
1467 mapping_cap_account_dirty(vma
->vm_file
->f_mapping
);
1471 * We account for memory if it's a private writeable mapping,
1472 * not hugepages and VM_NORESERVE wasn't set.
1474 static inline int accountable_mapping(struct file
*file
, vm_flags_t vm_flags
)
1477 * hugetlb has its own accounting separate from the core VM
1478 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1480 if (file
&& is_file_hugepages(file
))
1483 return (vm_flags
& (VM_NORESERVE
| VM_SHARED
| VM_WRITE
)) == VM_WRITE
;
1486 unsigned long mmap_region(struct file
*file
, unsigned long addr
,
1487 unsigned long len
, vm_flags_t vm_flags
, unsigned long pgoff
)
1489 struct mm_struct
*mm
= current
->mm
;
1490 struct vm_area_struct
*vma
, *prev
;
1492 struct rb_node
**rb_link
, *rb_parent
;
1493 unsigned long charged
= 0;
1495 /* Check against address space limit. */
1496 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
)) {
1497 unsigned long nr_pages
;
1500 * MAP_FIXED may remove pages of mappings that intersects with
1501 * requested mapping. Account for the pages it would unmap.
1503 if (!(vm_flags
& MAP_FIXED
))
1506 nr_pages
= count_vma_pages_range(mm
, addr
, addr
+ len
);
1508 if (!may_expand_vm(mm
, (len
>> PAGE_SHIFT
) - nr_pages
))
1512 /* Clear old maps */
1515 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
1516 if (do_munmap(mm
, addr
, len
))
1522 * Private writable mapping: check memory availability
1524 if (accountable_mapping(file
, vm_flags
)) {
1525 charged
= len
>> PAGE_SHIFT
;
1526 if (security_vm_enough_memory_mm(mm
, charged
))
1528 vm_flags
|= VM_ACCOUNT
;
1532 * Can we just expand an old mapping?
1534 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
, NULL
, file
, pgoff
, NULL
);
1539 * Determine the object being mapped and call the appropriate
1540 * specific mapper. the address has already been validated, but
1541 * not unmapped, but the maps are removed from the list.
1543 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
1550 vma
->vm_start
= addr
;
1551 vma
->vm_end
= addr
+ len
;
1552 vma
->vm_flags
= vm_flags
;
1553 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1554 vma
->vm_pgoff
= pgoff
;
1555 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
1558 if (vm_flags
& VM_DENYWRITE
) {
1559 error
= deny_write_access(file
);
1563 vma
->vm_file
= get_file(file
);
1564 error
= file
->f_op
->mmap(file
, vma
);
1566 goto unmap_and_free_vma
;
1568 /* Can addr have changed??
1570 * Answer: Yes, several device drivers can do it in their
1571 * f_op->mmap method. -DaveM
1572 * Bug: If addr is changed, prev, rb_link, rb_parent should
1573 * be updated for vma_link()
1575 WARN_ON_ONCE(addr
!= vma
->vm_start
);
1577 addr
= vma
->vm_start
;
1578 vm_flags
= vma
->vm_flags
;
1579 } else if (vm_flags
& VM_SHARED
) {
1580 error
= shmem_zero_setup(vma
);
1585 if (vma_wants_writenotify(vma
)) {
1586 pgprot_t pprot
= vma
->vm_page_prot
;
1588 /* Can vma->vm_page_prot have changed??
1590 * Answer: Yes, drivers may have changed it in their
1591 * f_op->mmap method.
1593 * Ensures that vmas marked as uncached stay that way.
1595 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
& ~VM_SHARED
);
1596 if (pgprot_val(pprot
) == pgprot_val(pgprot_noncached(pprot
)))
1597 vma
->vm_page_prot
= pgprot_noncached(vma
->vm_page_prot
);
1600 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1601 /* Once vma denies write, undo our temporary denial count */
1602 if (vm_flags
& VM_DENYWRITE
)
1603 allow_write_access(file
);
1604 file
= vma
->vm_file
;
1606 perf_event_mmap(vma
);
1608 vm_stat_account(mm
, vm_flags
, file
, len
>> PAGE_SHIFT
);
1609 if (vm_flags
& VM_LOCKED
) {
1610 if (!((vm_flags
& VM_SPECIAL
) || is_vm_hugetlb_page(vma
) ||
1611 vma
== get_gate_vma(current
->mm
)))
1612 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1614 vma
->vm_flags
&= ~VM_LOCKED
;
1621 * New (or expanded) vma always get soft dirty status.
1622 * Otherwise user-space soft-dirty page tracker won't
1623 * be able to distinguish situation when vma area unmapped,
1624 * then new mapped in-place (which must be aimed as
1625 * a completely new data area).
1627 vma
->vm_flags
|= VM_SOFTDIRTY
;
1632 if (vm_flags
& VM_DENYWRITE
)
1633 allow_write_access(file
);
1634 vma
->vm_file
= NULL
;
1637 /* Undo any partial mapping done by a device driver. */
1638 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1641 kmem_cache_free(vm_area_cachep
, vma
);
1644 vm_unacct_memory(charged
);
1648 unsigned long unmapped_area(struct vm_unmapped_area_info
*info
)
1651 * We implement the search by looking for an rbtree node that
1652 * immediately follows a suitable gap. That is,
1653 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1654 * - gap_end = vma->vm_start >= info->low_limit + length;
1655 * - gap_end - gap_start >= length
1658 struct mm_struct
*mm
= current
->mm
;
1659 struct vm_area_struct
*vma
;
1660 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1662 /* Adjust search length to account for worst case alignment overhead */
1663 length
= info
->length
+ info
->align_mask
;
1664 if (length
< info
->length
)
1667 /* Adjust search limits by the desired length */
1668 if (info
->high_limit
< length
)
1670 high_limit
= info
->high_limit
- length
;
1672 if (info
->low_limit
> high_limit
)
1674 low_limit
= info
->low_limit
+ length
;
1676 /* Check if rbtree root looks promising */
1677 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1679 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1680 if (vma
->rb_subtree_gap
< length
)
1684 /* Visit left subtree if it looks promising */
1685 gap_end
= vma
->vm_start
;
1686 if (gap_end
>= low_limit
&& vma
->vm_rb
.rb_left
) {
1687 struct vm_area_struct
*left
=
1688 rb_entry(vma
->vm_rb
.rb_left
,
1689 struct vm_area_struct
, vm_rb
);
1690 if (left
->rb_subtree_gap
>= length
) {
1696 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1698 /* Check if current node has a suitable gap */
1699 if (gap_start
> high_limit
)
1701 if (gap_end
>= low_limit
&& gap_end
- gap_start
>= length
)
1704 /* Visit right subtree if it looks promising */
1705 if (vma
->vm_rb
.rb_right
) {
1706 struct vm_area_struct
*right
=
1707 rb_entry(vma
->vm_rb
.rb_right
,
1708 struct vm_area_struct
, vm_rb
);
1709 if (right
->rb_subtree_gap
>= length
) {
1715 /* Go back up the rbtree to find next candidate node */
1717 struct rb_node
*prev
= &vma
->vm_rb
;
1718 if (!rb_parent(prev
))
1720 vma
= rb_entry(rb_parent(prev
),
1721 struct vm_area_struct
, vm_rb
);
1722 if (prev
== vma
->vm_rb
.rb_left
) {
1723 gap_start
= vma
->vm_prev
->vm_end
;
1724 gap_end
= vma
->vm_start
;
1731 /* Check highest gap, which does not precede any rbtree node */
1732 gap_start
= mm
->highest_vm_end
;
1733 gap_end
= ULONG_MAX
; /* Only for VM_BUG_ON below */
1734 if (gap_start
> high_limit
)
1738 /* We found a suitable gap. Clip it with the original low_limit. */
1739 if (gap_start
< info
->low_limit
)
1740 gap_start
= info
->low_limit
;
1742 /* Adjust gap address to the desired alignment */
1743 gap_start
+= (info
->align_offset
- gap_start
) & info
->align_mask
;
1745 VM_BUG_ON(gap_start
+ info
->length
> info
->high_limit
);
1746 VM_BUG_ON(gap_start
+ info
->length
> gap_end
);
1750 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info
*info
)
1752 struct mm_struct
*mm
= current
->mm
;
1753 struct vm_area_struct
*vma
;
1754 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1756 /* Adjust search length to account for worst case alignment overhead */
1757 length
= info
->length
+ info
->align_mask
;
1758 if (length
< info
->length
)
1762 * Adjust search limits by the desired length.
1763 * See implementation comment at top of unmapped_area().
1765 gap_end
= info
->high_limit
;
1766 if (gap_end
< length
)
1768 high_limit
= gap_end
- length
;
1770 if (info
->low_limit
> high_limit
)
1772 low_limit
= info
->low_limit
+ length
;
1774 /* Check highest gap, which does not precede any rbtree node */
1775 gap_start
= mm
->highest_vm_end
;
1776 if (gap_start
<= high_limit
)
1779 /* Check if rbtree root looks promising */
1780 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1782 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1783 if (vma
->rb_subtree_gap
< length
)
1787 /* Visit right subtree if it looks promising */
1788 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1789 if (gap_start
<= high_limit
&& vma
->vm_rb
.rb_right
) {
1790 struct vm_area_struct
*right
=
1791 rb_entry(vma
->vm_rb
.rb_right
,
1792 struct vm_area_struct
, vm_rb
);
1793 if (right
->rb_subtree_gap
>= length
) {
1800 /* Check if current node has a suitable gap */
1801 gap_end
= vma
->vm_start
;
1802 if (gap_end
< low_limit
)
1804 if (gap_start
<= high_limit
&& gap_end
- gap_start
>= length
)
1807 /* Visit left subtree if it looks promising */
1808 if (vma
->vm_rb
.rb_left
) {
1809 struct vm_area_struct
*left
=
1810 rb_entry(vma
->vm_rb
.rb_left
,
1811 struct vm_area_struct
, vm_rb
);
1812 if (left
->rb_subtree_gap
>= length
) {
1818 /* Go back up the rbtree to find next candidate node */
1820 struct rb_node
*prev
= &vma
->vm_rb
;
1821 if (!rb_parent(prev
))
1823 vma
= rb_entry(rb_parent(prev
),
1824 struct vm_area_struct
, vm_rb
);
1825 if (prev
== vma
->vm_rb
.rb_right
) {
1826 gap_start
= vma
->vm_prev
?
1827 vma
->vm_prev
->vm_end
: 0;
1834 /* We found a suitable gap. Clip it with the original high_limit. */
1835 if (gap_end
> info
->high_limit
)
1836 gap_end
= info
->high_limit
;
1839 /* Compute highest gap address at the desired alignment */
1840 gap_end
-= info
->length
;
1841 gap_end
-= (gap_end
- info
->align_offset
) & info
->align_mask
;
1843 VM_BUG_ON(gap_end
< info
->low_limit
);
1844 VM_BUG_ON(gap_end
< gap_start
);
1848 /* Get an address range which is currently unmapped.
1849 * For shmat() with addr=0.
1851 * Ugly calling convention alert:
1852 * Return value with the low bits set means error value,
1854 * if (ret & ~PAGE_MASK)
1857 * This function "knows" that -ENOMEM has the bits set.
1859 #ifndef HAVE_ARCH_UNMAPPED_AREA
1861 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
1862 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
1864 struct mm_struct
*mm
= current
->mm
;
1865 struct vm_area_struct
*vma
;
1866 struct vm_unmapped_area_info info
;
1868 if (len
> TASK_SIZE
- mmap_min_addr
)
1871 if (flags
& MAP_FIXED
)
1875 addr
= PAGE_ALIGN(addr
);
1876 vma
= find_vma(mm
, addr
);
1877 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1878 (!vma
|| addr
+ len
<= vma
->vm_start
))
1884 info
.low_limit
= mm
->mmap_base
;
1885 info
.high_limit
= TASK_SIZE
;
1886 info
.align_mask
= 0;
1887 return vm_unmapped_area(&info
);
1892 * This mmap-allocator allocates new areas top-down from below the
1893 * stack's low limit (the base):
1895 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1897 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
1898 const unsigned long len
, const unsigned long pgoff
,
1899 const unsigned long flags
)
1901 struct vm_area_struct
*vma
;
1902 struct mm_struct
*mm
= current
->mm
;
1903 unsigned long addr
= addr0
;
1904 struct vm_unmapped_area_info info
;
1906 /* requested length too big for entire address space */
1907 if (len
> TASK_SIZE
- mmap_min_addr
)
1910 if (flags
& MAP_FIXED
)
1913 /* requesting a specific address */
1915 addr
= PAGE_ALIGN(addr
);
1916 vma
= find_vma(mm
, addr
);
1917 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1918 (!vma
|| addr
+ len
<= vma
->vm_start
))
1922 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
1924 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
1925 info
.high_limit
= mm
->mmap_base
;
1926 info
.align_mask
= 0;
1927 addr
= vm_unmapped_area(&info
);
1930 * A failed mmap() very likely causes application failure,
1931 * so fall back to the bottom-up function here. This scenario
1932 * can happen with large stack limits and large mmap()
1935 if (addr
& ~PAGE_MASK
) {
1936 VM_BUG_ON(addr
!= -ENOMEM
);
1938 info
.low_limit
= TASK_UNMAPPED_BASE
;
1939 info
.high_limit
= TASK_SIZE
;
1940 addr
= vm_unmapped_area(&info
);
1948 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
1949 unsigned long pgoff
, unsigned long flags
)
1951 unsigned long (*get_area
)(struct file
*, unsigned long,
1952 unsigned long, unsigned long, unsigned long);
1954 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
1958 /* Careful about overflows.. */
1959 if (len
> TASK_SIZE
)
1962 get_area
= current
->mm
->get_unmapped_area
;
1963 if (file
&& file
->f_op
->get_unmapped_area
)
1964 get_area
= file
->f_op
->get_unmapped_area
;
1965 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
1966 if (IS_ERR_VALUE(addr
))
1969 if (addr
> TASK_SIZE
- len
)
1971 if (addr
& ~PAGE_MASK
)
1974 addr
= arch_rebalance_pgtables(addr
, len
);
1975 error
= security_mmap_addr(addr
);
1976 return error
? error
: addr
;
1979 EXPORT_SYMBOL(get_unmapped_area
);
1981 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1982 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
1984 struct vm_area_struct
*vma
= NULL
;
1986 /* Check the cache first. */
1987 /* (Cache hit rate is typically around 35%.) */
1988 vma
= ACCESS_ONCE(mm
->mmap_cache
);
1989 if (!(vma
&& vma
->vm_end
> addr
&& vma
->vm_start
<= addr
)) {
1990 struct rb_node
*rb_node
;
1992 rb_node
= mm
->mm_rb
.rb_node
;
1996 struct vm_area_struct
*vma_tmp
;
1998 vma_tmp
= rb_entry(rb_node
,
1999 struct vm_area_struct
, vm_rb
);
2001 if (vma_tmp
->vm_end
> addr
) {
2003 if (vma_tmp
->vm_start
<= addr
)
2005 rb_node
= rb_node
->rb_left
;
2007 rb_node
= rb_node
->rb_right
;
2010 mm
->mmap_cache
= vma
;
2015 EXPORT_SYMBOL(find_vma
);
2018 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2020 struct vm_area_struct
*
2021 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
2022 struct vm_area_struct
**pprev
)
2024 struct vm_area_struct
*vma
;
2026 vma
= find_vma(mm
, addr
);
2028 *pprev
= vma
->vm_prev
;
2030 struct rb_node
*rb_node
= mm
->mm_rb
.rb_node
;
2033 *pprev
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2034 rb_node
= rb_node
->rb_right
;
2041 * Verify that the stack growth is acceptable and
2042 * update accounting. This is shared with both the
2043 * grow-up and grow-down cases.
2045 static int acct_stack_growth(struct vm_area_struct
*vma
, unsigned long size
, unsigned long grow
)
2047 struct mm_struct
*mm
= vma
->vm_mm
;
2048 struct rlimit
*rlim
= current
->signal
->rlim
;
2049 unsigned long new_start
;
2051 /* address space limit tests */
2052 if (!may_expand_vm(mm
, grow
))
2055 /* Stack limit test */
2056 if (size
> ACCESS_ONCE(rlim
[RLIMIT_STACK
].rlim_cur
))
2059 /* mlock limit tests */
2060 if (vma
->vm_flags
& VM_LOCKED
) {
2061 unsigned long locked
;
2062 unsigned long limit
;
2063 locked
= mm
->locked_vm
+ grow
;
2064 limit
= ACCESS_ONCE(rlim
[RLIMIT_MEMLOCK
].rlim_cur
);
2065 limit
>>= PAGE_SHIFT
;
2066 if (locked
> limit
&& !capable(CAP_IPC_LOCK
))
2070 /* Check to ensure the stack will not grow into a hugetlb-only region */
2071 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
2073 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
2077 * Overcommit.. This must be the final test, as it will
2078 * update security statistics.
2080 if (security_vm_enough_memory_mm(mm
, grow
))
2083 /* Ok, everything looks good - let it rip */
2084 if (vma
->vm_flags
& VM_LOCKED
)
2085 mm
->locked_vm
+= grow
;
2086 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, grow
);
2090 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2092 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2093 * vma is the last one with address > vma->vm_end. Have to extend vma.
2095 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
)
2099 if (!(vma
->vm_flags
& VM_GROWSUP
))
2103 * We must make sure the anon_vma is allocated
2104 * so that the anon_vma locking is not a noop.
2106 if (unlikely(anon_vma_prepare(vma
)))
2108 vma_lock_anon_vma(vma
);
2111 * vma->vm_start/vm_end cannot change under us because the caller
2112 * is required to hold the mmap_sem in read mode. We need the
2113 * anon_vma lock to serialize against concurrent expand_stacks.
2114 * Also guard against wrapping around to address 0.
2116 if (address
< PAGE_ALIGN(address
+4))
2117 address
= PAGE_ALIGN(address
+4);
2119 vma_unlock_anon_vma(vma
);
2124 /* Somebody else might have raced and expanded it already */
2125 if (address
> vma
->vm_end
) {
2126 unsigned long size
, grow
;
2128 size
= address
- vma
->vm_start
;
2129 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
2132 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
2133 error
= acct_stack_growth(vma
, size
, grow
);
2136 * vma_gap_update() doesn't support concurrent
2137 * updates, but we only hold a shared mmap_sem
2138 * lock here, so we need to protect against
2139 * concurrent vma expansions.
2140 * vma_lock_anon_vma() doesn't help here, as
2141 * we don't guarantee that all growable vmas
2142 * in a mm share the same root anon vma.
2143 * So, we reuse mm->page_table_lock to guard
2144 * against concurrent vma expansions.
2146 spin_lock(&vma
->vm_mm
->page_table_lock
);
2147 anon_vma_interval_tree_pre_update_vma(vma
);
2148 vma
->vm_end
= address
;
2149 anon_vma_interval_tree_post_update_vma(vma
);
2151 vma_gap_update(vma
->vm_next
);
2153 vma
->vm_mm
->highest_vm_end
= address
;
2154 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2156 perf_event_mmap(vma
);
2160 vma_unlock_anon_vma(vma
);
2161 khugepaged_enter_vma_merge(vma
);
2162 validate_mm(vma
->vm_mm
);
2165 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2168 * vma is the first one with address < vma->vm_start. Have to extend vma.
2170 int expand_downwards(struct vm_area_struct
*vma
,
2171 unsigned long address
)
2176 * We must make sure the anon_vma is allocated
2177 * so that the anon_vma locking is not a noop.
2179 if (unlikely(anon_vma_prepare(vma
)))
2182 address
&= PAGE_MASK
;
2183 error
= security_mmap_addr(address
);
2187 vma_lock_anon_vma(vma
);
2190 * vma->vm_start/vm_end cannot change under us because the caller
2191 * is required to hold the mmap_sem in read mode. We need the
2192 * anon_vma lock to serialize against concurrent expand_stacks.
2195 /* Somebody else might have raced and expanded it already */
2196 if (address
< vma
->vm_start
) {
2197 unsigned long size
, grow
;
2199 size
= vma
->vm_end
- address
;
2200 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
2203 if (grow
<= vma
->vm_pgoff
) {
2204 error
= acct_stack_growth(vma
, size
, grow
);
2207 * vma_gap_update() doesn't support concurrent
2208 * updates, but we only hold a shared mmap_sem
2209 * lock here, so we need to protect against
2210 * concurrent vma expansions.
2211 * vma_lock_anon_vma() doesn't help here, as
2212 * we don't guarantee that all growable vmas
2213 * in a mm share the same root anon vma.
2214 * So, we reuse mm->page_table_lock to guard
2215 * against concurrent vma expansions.
2217 spin_lock(&vma
->vm_mm
->page_table_lock
);
2218 anon_vma_interval_tree_pre_update_vma(vma
);
2219 vma
->vm_start
= address
;
2220 vma
->vm_pgoff
-= grow
;
2221 anon_vma_interval_tree_post_update_vma(vma
);
2222 vma_gap_update(vma
);
2223 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2225 perf_event_mmap(vma
);
2229 vma_unlock_anon_vma(vma
);
2230 khugepaged_enter_vma_merge(vma
);
2231 validate_mm(vma
->vm_mm
);
2236 * Note how expand_stack() refuses to expand the stack all the way to
2237 * abut the next virtual mapping, *unless* that mapping itself is also
2238 * a stack mapping. We want to leave room for a guard page, after all
2239 * (the guard page itself is not added here, that is done by the
2240 * actual page faulting logic)
2242 * This matches the behavior of the guard page logic (see mm/memory.c:
2243 * check_stack_guard_page()), which only allows the guard page to be
2244 * removed under these circumstances.
2246 #ifdef CONFIG_STACK_GROWSUP
2247 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2249 struct vm_area_struct
*next
;
2251 address
&= PAGE_MASK
;
2252 next
= vma
->vm_next
;
2253 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
) {
2254 if (!(next
->vm_flags
& VM_GROWSUP
))
2257 return expand_upwards(vma
, address
);
2260 struct vm_area_struct
*
2261 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2263 struct vm_area_struct
*vma
, *prev
;
2266 vma
= find_vma_prev(mm
, addr
, &prev
);
2267 if (vma
&& (vma
->vm_start
<= addr
))
2269 if (!prev
|| expand_stack(prev
, addr
))
2271 if (prev
->vm_flags
& VM_LOCKED
)
2272 __mlock_vma_pages_range(prev
, addr
, prev
->vm_end
, NULL
);
2276 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2278 struct vm_area_struct
*prev
;
2280 address
&= PAGE_MASK
;
2281 prev
= vma
->vm_prev
;
2282 if (prev
&& prev
->vm_end
== address
) {
2283 if (!(prev
->vm_flags
& VM_GROWSDOWN
))
2286 return expand_downwards(vma
, address
);
2289 struct vm_area_struct
*
2290 find_extend_vma(struct mm_struct
* mm
, unsigned long addr
)
2292 struct vm_area_struct
* vma
;
2293 unsigned long start
;
2296 vma
= find_vma(mm
,addr
);
2299 if (vma
->vm_start
<= addr
)
2301 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2303 start
= vma
->vm_start
;
2304 if (expand_stack(vma
, addr
))
2306 if (vma
->vm_flags
& VM_LOCKED
)
2307 __mlock_vma_pages_range(vma
, addr
, start
, NULL
);
2313 * Ok - we have the memory areas we should free on the vma list,
2314 * so release them, and do the vma updates.
2316 * Called with the mm semaphore held.
2318 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2320 unsigned long nr_accounted
= 0;
2322 /* Update high watermark before we lower total_vm */
2323 update_hiwater_vm(mm
);
2325 long nrpages
= vma_pages(vma
);
2327 if (vma
->vm_flags
& VM_ACCOUNT
)
2328 nr_accounted
+= nrpages
;
2329 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, -nrpages
);
2330 vma
= remove_vma(vma
);
2332 vm_unacct_memory(nr_accounted
);
2337 * Get rid of page table information in the indicated region.
2339 * Called with the mm semaphore held.
2341 static void unmap_region(struct mm_struct
*mm
,
2342 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
2343 unsigned long start
, unsigned long end
)
2345 struct vm_area_struct
*next
= prev
? prev
->vm_next
: mm
->mmap
;
2346 struct mmu_gather tlb
;
2349 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2350 update_hiwater_rss(mm
);
2351 unmap_vmas(&tlb
, vma
, start
, end
);
2352 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
2353 next
? next
->vm_start
: USER_PGTABLES_CEILING
);
2354 tlb_finish_mmu(&tlb
, start
, end
);
2358 * Create a list of vma's touched by the unmap, removing them from the mm's
2359 * vma list as we go..
2362 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2363 struct vm_area_struct
*prev
, unsigned long end
)
2365 struct vm_area_struct
**insertion_point
;
2366 struct vm_area_struct
*tail_vma
= NULL
;
2368 insertion_point
= (prev
? &prev
->vm_next
: &mm
->mmap
);
2369 vma
->vm_prev
= NULL
;
2371 vma_rb_erase(vma
, &mm
->mm_rb
);
2375 } while (vma
&& vma
->vm_start
< end
);
2376 *insertion_point
= vma
;
2378 vma
->vm_prev
= prev
;
2379 vma_gap_update(vma
);
2381 mm
->highest_vm_end
= prev
? prev
->vm_end
: 0;
2382 tail_vma
->vm_next
= NULL
;
2383 mm
->mmap_cache
= NULL
; /* Kill the cache. */
2387 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2388 * munmap path where it doesn't make sense to fail.
2390 static int __split_vma(struct mm_struct
* mm
, struct vm_area_struct
* vma
,
2391 unsigned long addr
, int new_below
)
2393 struct vm_area_struct
*new;
2396 if (is_vm_hugetlb_page(vma
) && (addr
&
2397 ~(huge_page_mask(hstate_vma(vma
)))))
2400 new = kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2404 /* most fields are the same, copy all, and then fixup */
2407 INIT_LIST_HEAD(&new->anon_vma_chain
);
2412 new->vm_start
= addr
;
2413 new->vm_pgoff
+= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
);
2416 err
= vma_dup_policy(vma
, new);
2420 if (anon_vma_clone(new, vma
))
2424 get_file(new->vm_file
);
2426 if (new->vm_ops
&& new->vm_ops
->open
)
2427 new->vm_ops
->open(new);
2430 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2431 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2433 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2439 /* Clean everything up if vma_adjust failed. */
2440 if (new->vm_ops
&& new->vm_ops
->close
)
2441 new->vm_ops
->close(new);
2444 unlink_anon_vmas(new);
2446 mpol_put(vma_policy(new));
2448 kmem_cache_free(vm_area_cachep
, new);
2454 * Split a vma into two pieces at address 'addr', a new vma is allocated
2455 * either for the first part or the tail.
2457 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2458 unsigned long addr
, int new_below
)
2460 if (mm
->map_count
>= sysctl_max_map_count
)
2463 return __split_vma(mm
, vma
, addr
, new_below
);
2466 /* Munmap is split into 2 main parts -- this part which finds
2467 * what needs doing, and the areas themselves, which do the
2468 * work. This now handles partial unmappings.
2469 * Jeremy Fitzhardinge <jeremy@goop.org>
2471 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
)
2474 struct vm_area_struct
*vma
, *prev
, *last
;
2476 if ((start
& ~PAGE_MASK
) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2479 if ((len
= PAGE_ALIGN(len
)) == 0)
2482 /* Find the first overlapping VMA */
2483 vma
= find_vma(mm
, start
);
2486 prev
= vma
->vm_prev
;
2487 /* we have start < vma->vm_end */
2489 /* if it doesn't overlap, we have nothing.. */
2491 if (vma
->vm_start
>= end
)
2495 * If we need to split any vma, do it now to save pain later.
2497 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2498 * unmapped vm_area_struct will remain in use: so lower split_vma
2499 * places tmp vma above, and higher split_vma places tmp vma below.
2501 if (start
> vma
->vm_start
) {
2505 * Make sure that map_count on return from munmap() will
2506 * not exceed its limit; but let map_count go just above
2507 * its limit temporarily, to help free resources as expected.
2509 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2512 error
= __split_vma(mm
, vma
, start
, 0);
2518 /* Does it split the last one? */
2519 last
= find_vma(mm
, end
);
2520 if (last
&& end
> last
->vm_start
) {
2521 int error
= __split_vma(mm
, last
, end
, 1);
2525 vma
= prev
? prev
->vm_next
: mm
->mmap
;
2528 * unlock any mlock()ed ranges before detaching vmas
2530 if (mm
->locked_vm
) {
2531 struct vm_area_struct
*tmp
= vma
;
2532 while (tmp
&& tmp
->vm_start
< end
) {
2533 if (tmp
->vm_flags
& VM_LOCKED
) {
2534 mm
->locked_vm
-= vma_pages(tmp
);
2535 munlock_vma_pages_all(tmp
);
2542 * Remove the vma's, and unmap the actual pages
2544 detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
);
2545 unmap_region(mm
, vma
, prev
, start
, end
);
2547 /* Fix up all other VM information */
2548 remove_vma_list(mm
, vma
);
2553 int vm_munmap(unsigned long start
, size_t len
)
2556 struct mm_struct
*mm
= current
->mm
;
2558 down_write(&mm
->mmap_sem
);
2559 ret
= do_munmap(mm
, start
, len
);
2560 up_write(&mm
->mmap_sem
);
2563 EXPORT_SYMBOL(vm_munmap
);
2565 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2567 profile_munmap(addr
);
2568 return vm_munmap(addr
, len
);
2571 static inline void verify_mm_writelocked(struct mm_struct
*mm
)
2573 #ifdef CONFIG_DEBUG_VM
2574 if (unlikely(down_read_trylock(&mm
->mmap_sem
))) {
2576 up_read(&mm
->mmap_sem
);
2582 * this is really a simplified "do_mmap". it only handles
2583 * anonymous maps. eventually we may be able to do some
2584 * brk-specific accounting here.
2586 static unsigned long do_brk(unsigned long addr
, unsigned long len
)
2588 struct mm_struct
* mm
= current
->mm
;
2589 struct vm_area_struct
* vma
, * prev
;
2590 unsigned long flags
;
2591 struct rb_node
** rb_link
, * rb_parent
;
2592 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2595 len
= PAGE_ALIGN(len
);
2599 flags
= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2601 error
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2602 if (error
& ~PAGE_MASK
)
2605 error
= mlock_future_check(mm
, mm
->def_flags
, len
);
2610 * mm->mmap_sem is required to protect against another thread
2611 * changing the mappings in case we sleep.
2613 verify_mm_writelocked(mm
);
2616 * Clear old maps. this also does some error checking for us
2619 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
2620 if (do_munmap(mm
, addr
, len
))
2625 /* Check against address space limits *after* clearing old maps... */
2626 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
))
2629 if (mm
->map_count
> sysctl_max_map_count
)
2632 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
2635 /* Can we just expand an old private anonymous mapping? */
2636 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
2637 NULL
, NULL
, pgoff
, NULL
);
2642 * create a vma struct for an anonymous mapping
2644 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2646 vm_unacct_memory(len
>> PAGE_SHIFT
);
2650 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2652 vma
->vm_start
= addr
;
2653 vma
->vm_end
= addr
+ len
;
2654 vma
->vm_pgoff
= pgoff
;
2655 vma
->vm_flags
= flags
;
2656 vma
->vm_page_prot
= vm_get_page_prot(flags
);
2657 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2659 perf_event_mmap(vma
);
2660 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2661 if (flags
& VM_LOCKED
)
2662 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
2663 vma
->vm_flags
|= VM_SOFTDIRTY
;
2667 unsigned long vm_brk(unsigned long addr
, unsigned long len
)
2669 struct mm_struct
*mm
= current
->mm
;
2673 down_write(&mm
->mmap_sem
);
2674 ret
= do_brk(addr
, len
);
2675 populate
= ((mm
->def_flags
& VM_LOCKED
) != 0);
2676 up_write(&mm
->mmap_sem
);
2678 mm_populate(addr
, len
);
2681 EXPORT_SYMBOL(vm_brk
);
2683 /* Release all mmaps. */
2684 void exit_mmap(struct mm_struct
*mm
)
2686 struct mmu_gather tlb
;
2687 struct vm_area_struct
*vma
;
2688 unsigned long nr_accounted
= 0;
2690 /* mm's last user has gone, and its about to be pulled down */
2691 mmu_notifier_release(mm
);
2693 if (mm
->locked_vm
) {
2696 if (vma
->vm_flags
& VM_LOCKED
)
2697 munlock_vma_pages_all(vma
);
2705 if (!vma
) /* Can happen if dup_mmap() received an OOM */
2710 tlb_gather_mmu(&tlb
, mm
, 0, -1);
2711 /* update_hiwater_rss(mm) here? but nobody should be looking */
2712 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2713 unmap_vmas(&tlb
, vma
, 0, -1);
2715 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, USER_PGTABLES_CEILING
);
2716 tlb_finish_mmu(&tlb
, 0, -1);
2719 * Walk the list again, actually closing and freeing it,
2720 * with preemption enabled, without holding any MM locks.
2723 if (vma
->vm_flags
& VM_ACCOUNT
)
2724 nr_accounted
+= vma_pages(vma
);
2725 vma
= remove_vma(vma
);
2727 vm_unacct_memory(nr_accounted
);
2729 WARN_ON(atomic_long_read(&mm
->nr_ptes
) >
2730 (FIRST_USER_ADDRESS
+PMD_SIZE
-1)>>PMD_SHIFT
);
2733 /* Insert vm structure into process list sorted by address
2734 * and into the inode's i_mmap tree. If vm_file is non-NULL
2735 * then i_mmap_mutex is taken here.
2737 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2739 struct vm_area_struct
*prev
;
2740 struct rb_node
**rb_link
, *rb_parent
;
2743 * The vm_pgoff of a purely anonymous vma should be irrelevant
2744 * until its first write fault, when page's anon_vma and index
2745 * are set. But now set the vm_pgoff it will almost certainly
2746 * end up with (unless mremap moves it elsewhere before that
2747 * first wfault), so /proc/pid/maps tells a consistent story.
2749 * By setting it to reflect the virtual start address of the
2750 * vma, merges and splits can happen in a seamless way, just
2751 * using the existing file pgoff checks and manipulations.
2752 * Similarly in do_mmap_pgoff and in do_brk.
2754 if (!vma
->vm_file
) {
2755 BUG_ON(vma
->anon_vma
);
2756 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2758 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
2759 &prev
, &rb_link
, &rb_parent
))
2761 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
2762 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
2765 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2770 * Copy the vma structure to a new location in the same mm,
2771 * prior to moving page table entries, to effect an mremap move.
2773 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
2774 unsigned long addr
, unsigned long len
, pgoff_t pgoff
,
2775 bool *need_rmap_locks
)
2777 struct vm_area_struct
*vma
= *vmap
;
2778 unsigned long vma_start
= vma
->vm_start
;
2779 struct mm_struct
*mm
= vma
->vm_mm
;
2780 struct vm_area_struct
*new_vma
, *prev
;
2781 struct rb_node
**rb_link
, *rb_parent
;
2782 bool faulted_in_anon_vma
= true;
2785 * If anonymous vma has not yet been faulted, update new pgoff
2786 * to match new location, to increase its chance of merging.
2788 if (unlikely(!vma
->vm_file
&& !vma
->anon_vma
)) {
2789 pgoff
= addr
>> PAGE_SHIFT
;
2790 faulted_in_anon_vma
= false;
2793 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
2794 return NULL
; /* should never get here */
2795 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
2796 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
));
2799 * Source vma may have been merged into new_vma
2801 if (unlikely(vma_start
>= new_vma
->vm_start
&&
2802 vma_start
< new_vma
->vm_end
)) {
2804 * The only way we can get a vma_merge with
2805 * self during an mremap is if the vma hasn't
2806 * been faulted in yet and we were allowed to
2807 * reset the dst vma->vm_pgoff to the
2808 * destination address of the mremap to allow
2809 * the merge to happen. mremap must change the
2810 * vm_pgoff linearity between src and dst vmas
2811 * (in turn preventing a vma_merge) to be
2812 * safe. It is only safe to keep the vm_pgoff
2813 * linear if there are no pages mapped yet.
2815 VM_BUG_ON(faulted_in_anon_vma
);
2816 *vmap
= vma
= new_vma
;
2818 *need_rmap_locks
= (new_vma
->vm_pgoff
<= vma
->vm_pgoff
);
2820 new_vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2823 new_vma
->vm_start
= addr
;
2824 new_vma
->vm_end
= addr
+ len
;
2825 new_vma
->vm_pgoff
= pgoff
;
2826 if (vma_dup_policy(vma
, new_vma
))
2828 INIT_LIST_HEAD(&new_vma
->anon_vma_chain
);
2829 if (anon_vma_clone(new_vma
, vma
))
2830 goto out_free_mempol
;
2831 if (new_vma
->vm_file
)
2832 get_file(new_vma
->vm_file
);
2833 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
2834 new_vma
->vm_ops
->open(new_vma
);
2835 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
2836 *need_rmap_locks
= false;
2842 mpol_put(vma_policy(new_vma
));
2844 kmem_cache_free(vm_area_cachep
, new_vma
);
2849 * Return true if the calling process may expand its vm space by the passed
2852 int may_expand_vm(struct mm_struct
*mm
, unsigned long npages
)
2854 unsigned long cur
= mm
->total_vm
; /* pages */
2857 lim
= rlimit(RLIMIT_AS
) >> PAGE_SHIFT
;
2859 if (cur
+ npages
> lim
)
2865 static int special_mapping_fault(struct vm_area_struct
*vma
,
2866 struct vm_fault
*vmf
)
2869 struct page
**pages
;
2872 * special mappings have no vm_file, and in that case, the mm
2873 * uses vm_pgoff internally. So we have to subtract it from here.
2874 * We are allowed to do this because we are the mm; do not copy
2875 * this code into drivers!
2877 pgoff
= vmf
->pgoff
- vma
->vm_pgoff
;
2879 for (pages
= vma
->vm_private_data
; pgoff
&& *pages
; ++pages
)
2883 struct page
*page
= *pages
;
2889 return VM_FAULT_SIGBUS
;
2893 * Having a close hook prevents vma merging regardless of flags.
2895 static void special_mapping_close(struct vm_area_struct
*vma
)
2899 static const struct vm_operations_struct special_mapping_vmops
= {
2900 .close
= special_mapping_close
,
2901 .fault
= special_mapping_fault
,
2905 * Called with mm->mmap_sem held for writing.
2906 * Insert a new vma covering the given region, with the given flags.
2907 * Its pages are supplied by the given array of struct page *.
2908 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2909 * The region past the last page supplied will always produce SIGBUS.
2910 * The array pointer and the pages it points to are assumed to stay alive
2911 * for as long as this mapping might exist.
2913 int install_special_mapping(struct mm_struct
*mm
,
2914 unsigned long addr
, unsigned long len
,
2915 unsigned long vm_flags
, struct page
**pages
)
2918 struct vm_area_struct
*vma
;
2920 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2921 if (unlikely(vma
== NULL
))
2924 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2926 vma
->vm_start
= addr
;
2927 vma
->vm_end
= addr
+ len
;
2929 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
| VM_SOFTDIRTY
;
2930 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
2932 vma
->vm_ops
= &special_mapping_vmops
;
2933 vma
->vm_private_data
= pages
;
2935 ret
= insert_vm_struct(mm
, vma
);
2939 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2941 perf_event_mmap(vma
);
2946 kmem_cache_free(vm_area_cachep
, vma
);
2950 static DEFINE_MUTEX(mm_all_locks_mutex
);
2952 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
2954 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
2956 * The LSB of head.next can't change from under us
2957 * because we hold the mm_all_locks_mutex.
2959 down_write_nest_lock(&anon_vma
->root
->rwsem
, &mm
->mmap_sem
);
2961 * We can safely modify head.next after taking the
2962 * anon_vma->root->rwsem. If some other vma in this mm shares
2963 * the same anon_vma we won't take it again.
2965 * No need of atomic instructions here, head.next
2966 * can't change from under us thanks to the
2967 * anon_vma->root->rwsem.
2969 if (__test_and_set_bit(0, (unsigned long *)
2970 &anon_vma
->root
->rb_root
.rb_node
))
2975 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
2977 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
2979 * AS_MM_ALL_LOCKS can't change from under us because
2980 * we hold the mm_all_locks_mutex.
2982 * Operations on ->flags have to be atomic because
2983 * even if AS_MM_ALL_LOCKS is stable thanks to the
2984 * mm_all_locks_mutex, there may be other cpus
2985 * changing other bitflags in parallel to us.
2987 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
2989 mutex_lock_nest_lock(&mapping
->i_mmap_mutex
, &mm
->mmap_sem
);
2994 * This operation locks against the VM for all pte/vma/mm related
2995 * operations that could ever happen on a certain mm. This includes
2996 * vmtruncate, try_to_unmap, and all page faults.
2998 * The caller must take the mmap_sem in write mode before calling
2999 * mm_take_all_locks(). The caller isn't allowed to release the
3000 * mmap_sem until mm_drop_all_locks() returns.
3002 * mmap_sem in write mode is required in order to block all operations
3003 * that could modify pagetables and free pages without need of
3004 * altering the vma layout (for example populate_range() with
3005 * nonlinear vmas). It's also needed in write mode to avoid new
3006 * anon_vmas to be associated with existing vmas.
3008 * A single task can't take more than one mm_take_all_locks() in a row
3009 * or it would deadlock.
3011 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3012 * mapping->flags avoid to take the same lock twice, if more than one
3013 * vma in this mm is backed by the same anon_vma or address_space.
3015 * We can take all the locks in random order because the VM code
3016 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3017 * takes more than one of them in a row. Secondly we're protected
3018 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3020 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3021 * that may have to take thousand of locks.
3023 * mm_take_all_locks() can fail if it's interrupted by signals.
3025 int mm_take_all_locks(struct mm_struct
*mm
)
3027 struct vm_area_struct
*vma
;
3028 struct anon_vma_chain
*avc
;
3030 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3032 mutex_lock(&mm_all_locks_mutex
);
3034 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3035 if (signal_pending(current
))
3037 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3038 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3041 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3042 if (signal_pending(current
))
3045 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3046 vm_lock_anon_vma(mm
, avc
->anon_vma
);
3052 mm_drop_all_locks(mm
);
3056 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
3058 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3060 * The LSB of head.next can't change to 0 from under
3061 * us because we hold the mm_all_locks_mutex.
3063 * We must however clear the bitflag before unlocking
3064 * the vma so the users using the anon_vma->rb_root will
3065 * never see our bitflag.
3067 * No need of atomic instructions here, head.next
3068 * can't change from under us until we release the
3069 * anon_vma->root->rwsem.
3071 if (!__test_and_clear_bit(0, (unsigned long *)
3072 &anon_vma
->root
->rb_root
.rb_node
))
3074 anon_vma_unlock_write(anon_vma
);
3078 static void vm_unlock_mapping(struct address_space
*mapping
)
3080 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3082 * AS_MM_ALL_LOCKS can't change to 0 from under us
3083 * because we hold the mm_all_locks_mutex.
3085 mutex_unlock(&mapping
->i_mmap_mutex
);
3086 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
3093 * The mmap_sem cannot be released by the caller until
3094 * mm_drop_all_locks() returns.
3096 void mm_drop_all_locks(struct mm_struct
*mm
)
3098 struct vm_area_struct
*vma
;
3099 struct anon_vma_chain
*avc
;
3101 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3102 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
3104 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3106 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3107 vm_unlock_anon_vma(avc
->anon_vma
);
3108 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3109 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
3112 mutex_unlock(&mm_all_locks_mutex
);
3116 * initialise the VMA slab
3118 void __init
mmap_init(void)
3122 ret
= percpu_counter_init(&vm_committed_as
, 0);
3127 * Initialise sysctl_user_reserve_kbytes.
3129 * This is intended to prevent a user from starting a single memory hogging
3130 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3133 * The default value is min(3% of free memory, 128MB)
3134 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3136 static int init_user_reserve(void)
3138 unsigned long free_kbytes
;
3140 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3142 sysctl_user_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 17);
3145 subsys_initcall(init_user_reserve
);
3148 * Initialise sysctl_admin_reserve_kbytes.
3150 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3151 * to log in and kill a memory hogging process.
3153 * Systems with more than 256MB will reserve 8MB, enough to recover
3154 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3155 * only reserve 3% of free pages by default.
3157 static int init_admin_reserve(void)
3159 unsigned long free_kbytes
;
3161 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3163 sysctl_admin_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 13);
3166 subsys_initcall(init_admin_reserve
);
3169 * Reinititalise user and admin reserves if memory is added or removed.
3171 * The default user reserve max is 128MB, and the default max for the
3172 * admin reserve is 8MB. These are usually, but not always, enough to
3173 * enable recovery from a memory hogging process using login/sshd, a shell,
3174 * and tools like top. It may make sense to increase or even disable the
3175 * reserve depending on the existence of swap or variations in the recovery
3176 * tools. So, the admin may have changed them.
3178 * If memory is added and the reserves have been eliminated or increased above
3179 * the default max, then we'll trust the admin.
3181 * If memory is removed and there isn't enough free memory, then we
3182 * need to reset the reserves.
3184 * Otherwise keep the reserve set by the admin.
3186 static int reserve_mem_notifier(struct notifier_block
*nb
,
3187 unsigned long action
, void *data
)
3189 unsigned long tmp
, free_kbytes
;
3193 /* Default max is 128MB. Leave alone if modified by operator. */
3194 tmp
= sysctl_user_reserve_kbytes
;
3195 if (0 < tmp
&& tmp
< (1UL << 17))
3196 init_user_reserve();
3198 /* Default max is 8MB. Leave alone if modified by operator. */
3199 tmp
= sysctl_admin_reserve_kbytes
;
3200 if (0 < tmp
&& tmp
< (1UL << 13))
3201 init_admin_reserve();
3205 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3207 if (sysctl_user_reserve_kbytes
> free_kbytes
) {
3208 init_user_reserve();
3209 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3210 sysctl_user_reserve_kbytes
);
3213 if (sysctl_admin_reserve_kbytes
> free_kbytes
) {
3214 init_admin_reserve();
3215 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3216 sysctl_admin_reserve_kbytes
);
3225 static struct notifier_block reserve_mem_nb
= {
3226 .notifier_call
= reserve_mem_notifier
,
3229 static int __meminit
init_reserve_notifier(void)
3231 if (register_hotmemory_notifier(&reserve_mem_nb
))
3232 printk("Failed registering memory add/remove notifier for admin reserve");
3236 subsys_initcall(init_reserve_notifier
);