6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/profile.h>
29 #include <linux/export.h>
30 #include <linux/mount.h>
31 #include <linux/mempolicy.h>
32 #include <linux/rmap.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/mmdebug.h>
35 #include <linux/perf_event.h>
36 #include <linux/audit.h>
37 #include <linux/khugepaged.h>
38 #include <linux/uprobes.h>
39 #include <linux/rbtree_augmented.h>
40 #include <linux/sched/sysctl.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
45 #include <asm/uaccess.h>
46 #include <asm/cacheflush.h>
48 #include <asm/mmu_context.h>
52 #ifndef arch_mmap_check
53 #define arch_mmap_check(addr, len, flags) (0)
56 #ifndef arch_rebalance_pgtables
57 #define arch_rebalance_pgtables(addr, len) (addr)
60 static void unmap_region(struct mm_struct
*mm
,
61 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
62 unsigned long start
, unsigned long end
);
64 /* description of effects of mapping type and prot in current implementation.
65 * this is due to the limited x86 page protection hardware. The expected
66 * behavior is in parens:
69 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
70 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
71 * w: (no) no w: (no) no w: (yes) yes w: (no) no
72 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
74 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
75 * w: (no) no w: (no) no w: (copy) copy w: (no) no
76 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
79 pgprot_t protection_map
[16] = {
80 __P000
, __P001
, __P010
, __P011
, __P100
, __P101
, __P110
, __P111
,
81 __S000
, __S001
, __S010
, __S011
, __S100
, __S101
, __S110
, __S111
84 pgprot_t
vm_get_page_prot(unsigned long vm_flags
)
86 return __pgprot(pgprot_val(protection_map
[vm_flags
&
87 (VM_READ
|VM_WRITE
|VM_EXEC
|VM_SHARED
)]) |
88 pgprot_val(arch_vm_get_page_prot(vm_flags
)));
90 EXPORT_SYMBOL(vm_get_page_prot
);
92 static pgprot_t
vm_pgprot_modify(pgprot_t oldprot
, unsigned long vm_flags
)
94 return pgprot_modify(oldprot
, vm_get_page_prot(vm_flags
));
97 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
98 void vma_set_page_prot(struct vm_area_struct
*vma
)
100 unsigned long vm_flags
= vma
->vm_flags
;
102 vma
->vm_page_prot
= vm_pgprot_modify(vma
->vm_page_prot
, vm_flags
);
103 if (vma_wants_writenotify(vma
)) {
104 vm_flags
&= ~VM_SHARED
;
105 vma
->vm_page_prot
= vm_pgprot_modify(vma
->vm_page_prot
,
111 int sysctl_overcommit_memory __read_mostly
= OVERCOMMIT_GUESS
; /* heuristic overcommit */
112 int sysctl_overcommit_ratio __read_mostly
= 50; /* default is 50% */
113 unsigned long sysctl_overcommit_kbytes __read_mostly
;
114 int sysctl_max_map_count __read_mostly
= DEFAULT_MAX_MAP_COUNT
;
115 unsigned long sysctl_user_reserve_kbytes __read_mostly
= 1UL << 17; /* 128MB */
116 unsigned long sysctl_admin_reserve_kbytes __read_mostly
= 1UL << 13; /* 8MB */
118 * Make sure vm_committed_as in one cacheline and not cacheline shared with
119 * other variables. It can be updated by several CPUs frequently.
121 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp
;
124 * The global memory commitment made in the system can be a metric
125 * that can be used to drive ballooning decisions when Linux is hosted
126 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
127 * balancing memory across competing virtual machines that are hosted.
128 * Several metrics drive this policy engine including the guest reported
131 unsigned long vm_memory_committed(void)
133 return percpu_counter_read_positive(&vm_committed_as
);
135 EXPORT_SYMBOL_GPL(vm_memory_committed
);
138 * Check that a process has enough memory to allocate a new virtual
139 * mapping. 0 means there is enough memory for the allocation to
140 * succeed and -ENOMEM implies there is not.
142 * We currently support three overcommit policies, which are set via the
143 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
145 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
146 * Additional code 2002 Jul 20 by Robert Love.
148 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
150 * Note this is a helper function intended to be used by LSMs which
151 * wish to use this logic.
153 int __vm_enough_memory(struct mm_struct
*mm
, long pages
, int cap_sys_admin
)
155 unsigned long free
, allowed
, reserve
;
157 VM_WARN_ONCE(percpu_counter_read(&vm_committed_as
) <
158 -(s64
)vm_committed_as_batch
* num_online_cpus(),
159 "memory commitment underflow");
161 vm_acct_memory(pages
);
164 * Sometimes we want to use more memory than we have
166 if (sysctl_overcommit_memory
== OVERCOMMIT_ALWAYS
)
169 if (sysctl_overcommit_memory
== OVERCOMMIT_GUESS
) {
170 free
= global_page_state(NR_FREE_PAGES
);
171 free
+= global_page_state(NR_FILE_PAGES
);
174 * shmem pages shouldn't be counted as free in this
175 * case, they can't be purged, only swapped out, and
176 * that won't affect the overall amount of available
177 * memory in the system.
179 free
-= global_page_state(NR_SHMEM
);
181 free
+= get_nr_swap_pages();
184 * Any slabs which are created with the
185 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
186 * which are reclaimable, under pressure. The dentry
187 * cache and most inode caches should fall into this
189 free
+= global_page_state(NR_SLAB_RECLAIMABLE
);
192 * Leave reserved pages. The pages are not for anonymous pages.
194 if (free
<= totalreserve_pages
)
197 free
-= totalreserve_pages
;
200 * Reserve some for root
203 free
-= sysctl_admin_reserve_kbytes
>> (PAGE_SHIFT
- 10);
211 allowed
= vm_commit_limit();
213 * Reserve some for root
216 allowed
-= sysctl_admin_reserve_kbytes
>> (PAGE_SHIFT
- 10);
219 * Don't let a single process grow so big a user can't recover
222 reserve
= sysctl_user_reserve_kbytes
>> (PAGE_SHIFT
- 10);
223 allowed
-= min(mm
->total_vm
/ 32, reserve
);
226 if (percpu_counter_read_positive(&vm_committed_as
) < allowed
)
229 vm_unacct_memory(pages
);
235 * Requires inode->i_mapping->i_mmap_mutex
237 static void __remove_shared_vm_struct(struct vm_area_struct
*vma
,
238 struct file
*file
, struct address_space
*mapping
)
240 if (vma
->vm_flags
& VM_DENYWRITE
)
241 atomic_inc(&file_inode(file
)->i_writecount
);
242 if (vma
->vm_flags
& VM_SHARED
)
243 mapping_unmap_writable(mapping
);
245 flush_dcache_mmap_lock(mapping
);
246 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
247 list_del_init(&vma
->shared
.nonlinear
);
249 vma_interval_tree_remove(vma
, &mapping
->i_mmap
);
250 flush_dcache_mmap_unlock(mapping
);
254 * Unlink a file-based vm structure from its interval tree, to hide
255 * vma from rmap and vmtruncate before freeing its page tables.
257 void unlink_file_vma(struct vm_area_struct
*vma
)
259 struct file
*file
= vma
->vm_file
;
262 struct address_space
*mapping
= file
->f_mapping
;
263 mutex_lock(&mapping
->i_mmap_mutex
);
264 __remove_shared_vm_struct(vma
, file
, mapping
);
265 mutex_unlock(&mapping
->i_mmap_mutex
);
270 * Close a vm structure and free it, returning the next.
272 static struct vm_area_struct
*remove_vma(struct vm_area_struct
*vma
)
274 struct vm_area_struct
*next
= vma
->vm_next
;
277 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
278 vma
->vm_ops
->close(vma
);
281 mpol_put(vma_policy(vma
));
282 kmem_cache_free(vm_area_cachep
, vma
);
286 static unsigned long do_brk(unsigned long addr
, unsigned long len
);
288 SYSCALL_DEFINE1(brk
, unsigned long, brk
)
290 unsigned long retval
;
291 unsigned long newbrk
, oldbrk
;
292 struct mm_struct
*mm
= current
->mm
;
293 unsigned long min_brk
;
296 down_write(&mm
->mmap_sem
);
298 #ifdef CONFIG_COMPAT_BRK
300 * CONFIG_COMPAT_BRK can still be overridden by setting
301 * randomize_va_space to 2, which will still cause mm->start_brk
302 * to be arbitrarily shifted
304 if (current
->brk_randomized
)
305 min_brk
= mm
->start_brk
;
307 min_brk
= mm
->end_data
;
309 min_brk
= mm
->start_brk
;
315 * Check against rlimit here. If this check is done later after the test
316 * of oldbrk with newbrk then it can escape the test and let the data
317 * segment grow beyond its set limit the in case where the limit is
318 * not page aligned -Ram Gupta
320 if (check_data_rlimit(rlimit(RLIMIT_DATA
), brk
, mm
->start_brk
,
321 mm
->end_data
, mm
->start_data
))
324 newbrk
= PAGE_ALIGN(brk
);
325 oldbrk
= PAGE_ALIGN(mm
->brk
);
326 if (oldbrk
== newbrk
)
329 /* Always allow shrinking brk. */
330 if (brk
<= mm
->brk
) {
331 if (!do_munmap(mm
, newbrk
, oldbrk
-newbrk
))
336 /* Check against existing mmap mappings. */
337 if (find_vma_intersection(mm
, oldbrk
, newbrk
+PAGE_SIZE
))
340 /* Ok, looks good - let it rip. */
341 if (do_brk(oldbrk
, newbrk
-oldbrk
) != oldbrk
)
346 populate
= newbrk
> oldbrk
&& (mm
->def_flags
& VM_LOCKED
) != 0;
347 up_write(&mm
->mmap_sem
);
349 mm_populate(oldbrk
, newbrk
- oldbrk
);
354 up_write(&mm
->mmap_sem
);
358 static long vma_compute_subtree_gap(struct vm_area_struct
*vma
)
360 unsigned long max
, subtree_gap
;
363 max
-= vma
->vm_prev
->vm_end
;
364 if (vma
->vm_rb
.rb_left
) {
365 subtree_gap
= rb_entry(vma
->vm_rb
.rb_left
,
366 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
367 if (subtree_gap
> max
)
370 if (vma
->vm_rb
.rb_right
) {
371 subtree_gap
= rb_entry(vma
->vm_rb
.rb_right
,
372 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
373 if (subtree_gap
> max
)
379 #ifdef CONFIG_DEBUG_VM_RB
380 static int browse_rb(struct rb_root
*root
)
382 int i
= 0, j
, bug
= 0;
383 struct rb_node
*nd
, *pn
= NULL
;
384 unsigned long prev
= 0, pend
= 0;
386 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
387 struct vm_area_struct
*vma
;
388 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
389 if (vma
->vm_start
< prev
) {
390 pr_emerg("vm_start %lx < prev %lx\n",
391 vma
->vm_start
, prev
);
394 if (vma
->vm_start
< pend
) {
395 pr_emerg("vm_start %lx < pend %lx\n",
396 vma
->vm_start
, pend
);
399 if (vma
->vm_start
> vma
->vm_end
) {
400 pr_emerg("vm_start %lx > vm_end %lx\n",
401 vma
->vm_start
, vma
->vm_end
);
404 if (vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
)) {
405 pr_emerg("free gap %lx, correct %lx\n",
407 vma_compute_subtree_gap(vma
));
412 prev
= vma
->vm_start
;
416 for (nd
= pn
; nd
; nd
= rb_prev(nd
))
419 pr_emerg("backwards %d, forwards %d\n", j
, i
);
425 static void validate_mm_rb(struct rb_root
*root
, struct vm_area_struct
*ignore
)
429 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
430 struct vm_area_struct
*vma
;
431 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
432 VM_BUG_ON_VMA(vma
!= ignore
&&
433 vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
),
438 static void validate_mm(struct mm_struct
*mm
)
442 unsigned long highest_address
= 0;
443 struct vm_area_struct
*vma
= mm
->mmap
;
446 struct anon_vma_chain
*avc
;
448 vma_lock_anon_vma(vma
);
449 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
450 anon_vma_interval_tree_verify(avc
);
451 vma_unlock_anon_vma(vma
);
452 highest_address
= vma
->vm_end
;
456 if (i
!= mm
->map_count
) {
457 pr_emerg("map_count %d vm_next %d\n", mm
->map_count
, i
);
460 if (highest_address
!= mm
->highest_vm_end
) {
461 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
462 mm
->highest_vm_end
, highest_address
);
465 i
= browse_rb(&mm
->mm_rb
);
466 if (i
!= mm
->map_count
) {
468 pr_emerg("map_count %d rb %d\n", mm
->map_count
, i
);
471 VM_BUG_ON_MM(bug
, mm
);
474 #define validate_mm_rb(root, ignore) do { } while (0)
475 #define validate_mm(mm) do { } while (0)
478 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks
, struct vm_area_struct
, vm_rb
,
479 unsigned long, rb_subtree_gap
, vma_compute_subtree_gap
)
482 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
483 * vma->vm_prev->vm_end values changed, without modifying the vma's position
486 static void vma_gap_update(struct vm_area_struct
*vma
)
489 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
490 * function that does exacltly what we want.
492 vma_gap_callbacks_propagate(&vma
->vm_rb
, NULL
);
495 static inline void vma_rb_insert(struct vm_area_struct
*vma
,
496 struct rb_root
*root
)
498 /* All rb_subtree_gap values must be consistent prior to insertion */
499 validate_mm_rb(root
, NULL
);
501 rb_insert_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
504 static void vma_rb_erase(struct vm_area_struct
*vma
, struct rb_root
*root
)
507 * All rb_subtree_gap values must be consistent prior to erase,
508 * with the possible exception of the vma being erased.
510 validate_mm_rb(root
, vma
);
513 * Note rb_erase_augmented is a fairly large inline function,
514 * so make sure we instantiate it only once with our desired
515 * augmented rbtree callbacks.
517 rb_erase_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
521 * vma has some anon_vma assigned, and is already inserted on that
522 * anon_vma's interval trees.
524 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
525 * vma must be removed from the anon_vma's interval trees using
526 * anon_vma_interval_tree_pre_update_vma().
528 * After the update, the vma will be reinserted using
529 * anon_vma_interval_tree_post_update_vma().
531 * The entire update must be protected by exclusive mmap_sem and by
532 * the root anon_vma's mutex.
535 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct
*vma
)
537 struct anon_vma_chain
*avc
;
539 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
540 anon_vma_interval_tree_remove(avc
, &avc
->anon_vma
->rb_root
);
544 anon_vma_interval_tree_post_update_vma(struct vm_area_struct
*vma
)
546 struct anon_vma_chain
*avc
;
548 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
549 anon_vma_interval_tree_insert(avc
, &avc
->anon_vma
->rb_root
);
552 static int find_vma_links(struct mm_struct
*mm
, unsigned long addr
,
553 unsigned long end
, struct vm_area_struct
**pprev
,
554 struct rb_node
***rb_link
, struct rb_node
**rb_parent
)
556 struct rb_node
**__rb_link
, *__rb_parent
, *rb_prev
;
558 __rb_link
= &mm
->mm_rb
.rb_node
;
559 rb_prev
= __rb_parent
= NULL
;
562 struct vm_area_struct
*vma_tmp
;
564 __rb_parent
= *__rb_link
;
565 vma_tmp
= rb_entry(__rb_parent
, struct vm_area_struct
, vm_rb
);
567 if (vma_tmp
->vm_end
> addr
) {
568 /* Fail if an existing vma overlaps the area */
569 if (vma_tmp
->vm_start
< end
)
571 __rb_link
= &__rb_parent
->rb_left
;
573 rb_prev
= __rb_parent
;
574 __rb_link
= &__rb_parent
->rb_right
;
580 *pprev
= rb_entry(rb_prev
, struct vm_area_struct
, vm_rb
);
581 *rb_link
= __rb_link
;
582 *rb_parent
= __rb_parent
;
586 static unsigned long count_vma_pages_range(struct mm_struct
*mm
,
587 unsigned long addr
, unsigned long end
)
589 unsigned long nr_pages
= 0;
590 struct vm_area_struct
*vma
;
592 /* Find first overlaping mapping */
593 vma
= find_vma_intersection(mm
, addr
, end
);
597 nr_pages
= (min(end
, vma
->vm_end
) -
598 max(addr
, vma
->vm_start
)) >> PAGE_SHIFT
;
600 /* Iterate over the rest of the overlaps */
601 for (vma
= vma
->vm_next
; vma
; vma
= vma
->vm_next
) {
602 unsigned long overlap_len
;
604 if (vma
->vm_start
> end
)
607 overlap_len
= min(end
, vma
->vm_end
) - vma
->vm_start
;
608 nr_pages
+= overlap_len
>> PAGE_SHIFT
;
614 void __vma_link_rb(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
615 struct rb_node
**rb_link
, struct rb_node
*rb_parent
)
617 /* Update tracking information for the gap following the new vma. */
619 vma_gap_update(vma
->vm_next
);
621 mm
->highest_vm_end
= vma
->vm_end
;
624 * vma->vm_prev wasn't known when we followed the rbtree to find the
625 * correct insertion point for that vma. As a result, we could not
626 * update the vma vm_rb parents rb_subtree_gap values on the way down.
627 * So, we first insert the vma with a zero rb_subtree_gap value
628 * (to be consistent with what we did on the way down), and then
629 * immediately update the gap to the correct value. Finally we
630 * rebalance the rbtree after all augmented values have been set.
632 rb_link_node(&vma
->vm_rb
, rb_parent
, rb_link
);
633 vma
->rb_subtree_gap
= 0;
635 vma_rb_insert(vma
, &mm
->mm_rb
);
638 static void __vma_link_file(struct vm_area_struct
*vma
)
644 struct address_space
*mapping
= file
->f_mapping
;
646 if (vma
->vm_flags
& VM_DENYWRITE
)
647 atomic_dec(&file_inode(file
)->i_writecount
);
648 if (vma
->vm_flags
& VM_SHARED
)
649 atomic_inc(&mapping
->i_mmap_writable
);
651 flush_dcache_mmap_lock(mapping
);
652 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
653 vma_nonlinear_insert(vma
, &mapping
->i_mmap_nonlinear
);
655 vma_interval_tree_insert(vma
, &mapping
->i_mmap
);
656 flush_dcache_mmap_unlock(mapping
);
661 __vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
662 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
663 struct rb_node
*rb_parent
)
665 __vma_link_list(mm
, vma
, prev
, rb_parent
);
666 __vma_link_rb(mm
, vma
, rb_link
, rb_parent
);
669 static void vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
670 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
671 struct rb_node
*rb_parent
)
673 struct address_space
*mapping
= NULL
;
676 mapping
= vma
->vm_file
->f_mapping
;
677 mutex_lock(&mapping
->i_mmap_mutex
);
680 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
681 __vma_link_file(vma
);
684 mutex_unlock(&mapping
->i_mmap_mutex
);
691 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
692 * mm's list and rbtree. It has already been inserted into the interval tree.
694 static void __insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
696 struct vm_area_struct
*prev
;
697 struct rb_node
**rb_link
, *rb_parent
;
699 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
700 &prev
, &rb_link
, &rb_parent
))
702 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
707 __vma_unlink(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
708 struct vm_area_struct
*prev
)
710 struct vm_area_struct
*next
;
712 vma_rb_erase(vma
, &mm
->mm_rb
);
713 prev
->vm_next
= next
= vma
->vm_next
;
715 next
->vm_prev
= prev
;
718 vmacache_invalidate(mm
);
722 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
723 * is already present in an i_mmap tree without adjusting the tree.
724 * The following helper function should be used when such adjustments
725 * are necessary. The "insert" vma (if any) is to be inserted
726 * before we drop the necessary locks.
728 int vma_adjust(struct vm_area_struct
*vma
, unsigned long start
,
729 unsigned long end
, pgoff_t pgoff
, struct vm_area_struct
*insert
)
731 struct mm_struct
*mm
= vma
->vm_mm
;
732 struct vm_area_struct
*next
= vma
->vm_next
;
733 struct vm_area_struct
*importer
= NULL
;
734 struct address_space
*mapping
= NULL
;
735 struct rb_root
*root
= NULL
;
736 struct anon_vma
*anon_vma
= NULL
;
737 struct file
*file
= vma
->vm_file
;
738 bool start_changed
= false, end_changed
= false;
739 long adjust_next
= 0;
742 if (next
&& !insert
) {
743 struct vm_area_struct
*exporter
= NULL
;
745 if (end
>= next
->vm_end
) {
747 * vma expands, overlapping all the next, and
748 * perhaps the one after too (mprotect case 6).
750 again
: remove_next
= 1 + (end
> next
->vm_end
);
754 } else if (end
> next
->vm_start
) {
756 * vma expands, overlapping part of the next:
757 * mprotect case 5 shifting the boundary up.
759 adjust_next
= (end
- next
->vm_start
) >> PAGE_SHIFT
;
762 } else if (end
< vma
->vm_end
) {
764 * vma shrinks, and !insert tells it's not
765 * split_vma inserting another: so it must be
766 * mprotect case 4 shifting the boundary down.
768 adjust_next
= -((vma
->vm_end
- end
) >> PAGE_SHIFT
);
774 * Easily overlooked: when mprotect shifts the boundary,
775 * make sure the expanding vma has anon_vma set if the
776 * shrinking vma had, to cover any anon pages imported.
778 if (exporter
&& exporter
->anon_vma
&& !importer
->anon_vma
) {
779 if (anon_vma_clone(importer
, exporter
))
781 importer
->anon_vma
= exporter
->anon_vma
;
786 mapping
= file
->f_mapping
;
787 if (!(vma
->vm_flags
& VM_NONLINEAR
)) {
788 root
= &mapping
->i_mmap
;
789 uprobe_munmap(vma
, vma
->vm_start
, vma
->vm_end
);
792 uprobe_munmap(next
, next
->vm_start
,
796 mutex_lock(&mapping
->i_mmap_mutex
);
799 * Put into interval tree now, so instantiated pages
800 * are visible to arm/parisc __flush_dcache_page
801 * throughout; but we cannot insert into address
802 * space until vma start or end is updated.
804 __vma_link_file(insert
);
808 vma_adjust_trans_huge(vma
, start
, end
, adjust_next
);
810 anon_vma
= vma
->anon_vma
;
811 if (!anon_vma
&& adjust_next
)
812 anon_vma
= next
->anon_vma
;
814 VM_BUG_ON_VMA(adjust_next
&& next
->anon_vma
&&
815 anon_vma
!= next
->anon_vma
, next
);
816 anon_vma_lock_write(anon_vma
);
817 anon_vma_interval_tree_pre_update_vma(vma
);
819 anon_vma_interval_tree_pre_update_vma(next
);
823 flush_dcache_mmap_lock(mapping
);
824 vma_interval_tree_remove(vma
, root
);
826 vma_interval_tree_remove(next
, root
);
829 if (start
!= vma
->vm_start
) {
830 vma
->vm_start
= start
;
831 start_changed
= true;
833 if (end
!= vma
->vm_end
) {
837 vma
->vm_pgoff
= pgoff
;
839 next
->vm_start
+= adjust_next
<< PAGE_SHIFT
;
840 next
->vm_pgoff
+= adjust_next
;
845 vma_interval_tree_insert(next
, root
);
846 vma_interval_tree_insert(vma
, root
);
847 flush_dcache_mmap_unlock(mapping
);
852 * vma_merge has merged next into vma, and needs
853 * us to remove next before dropping the locks.
855 __vma_unlink(mm
, next
, vma
);
857 __remove_shared_vm_struct(next
, file
, mapping
);
860 * split_vma has split insert from vma, and needs
861 * us to insert it before dropping the locks
862 * (it may either follow vma or precede it).
864 __insert_vm_struct(mm
, insert
);
870 mm
->highest_vm_end
= end
;
871 else if (!adjust_next
)
872 vma_gap_update(next
);
877 anon_vma_interval_tree_post_update_vma(vma
);
879 anon_vma_interval_tree_post_update_vma(next
);
880 anon_vma_unlock_write(anon_vma
);
883 mutex_unlock(&mapping
->i_mmap_mutex
);
894 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
898 anon_vma_merge(vma
, next
);
900 mpol_put(vma_policy(next
));
901 kmem_cache_free(vm_area_cachep
, next
);
903 * In mprotect's case 6 (see comments on vma_merge),
904 * we must remove another next too. It would clutter
905 * up the code too much to do both in one go.
908 if (remove_next
== 2)
911 vma_gap_update(next
);
913 mm
->highest_vm_end
= end
;
924 * If the vma has a ->close operation then the driver probably needs to release
925 * per-vma resources, so we don't attempt to merge those.
927 static inline int is_mergeable_vma(struct vm_area_struct
*vma
,
928 struct file
*file
, unsigned long vm_flags
)
931 * VM_SOFTDIRTY should not prevent from VMA merging, if we
932 * match the flags but dirty bit -- the caller should mark
933 * merged VMA as dirty. If dirty bit won't be excluded from
934 * comparison, we increase pressue on the memory system forcing
935 * the kernel to generate new VMAs when old one could be
938 if ((vma
->vm_flags
^ vm_flags
) & ~VM_SOFTDIRTY
)
940 if (vma
->vm_file
!= file
)
942 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
947 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
948 struct anon_vma
*anon_vma2
,
949 struct vm_area_struct
*vma
)
952 * The list_is_singular() test is to avoid merging VMA cloned from
953 * parents. This can improve scalability caused by anon_vma lock.
955 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
956 list_is_singular(&vma
->anon_vma_chain
)))
958 return anon_vma1
== anon_vma2
;
962 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
963 * in front of (at a lower virtual address and file offset than) the vma.
965 * We cannot merge two vmas if they have differently assigned (non-NULL)
966 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
968 * We don't check here for the merged mmap wrapping around the end of pagecache
969 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
970 * wrap, nor mmaps which cover the final page at index -1UL.
973 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
974 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
976 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
977 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
978 if (vma
->vm_pgoff
== vm_pgoff
)
985 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
986 * beyond (at a higher virtual address and file offset than) the vma.
988 * We cannot merge two vmas if they have differently assigned (non-NULL)
989 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
992 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
993 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
995 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
996 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
998 vm_pglen
= vma_pages(vma
);
999 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
1006 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1007 * whether that can be merged with its predecessor or its successor.
1008 * Or both (it neatly fills a hole).
1010 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1011 * certain not to be mapped by the time vma_merge is called; but when
1012 * called for mprotect, it is certain to be already mapped (either at
1013 * an offset within prev, or at the start of next), and the flags of
1014 * this area are about to be changed to vm_flags - and the no-change
1015 * case has already been eliminated.
1017 * The following mprotect cases have to be considered, where AAAA is
1018 * the area passed down from mprotect_fixup, never extending beyond one
1019 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1021 * AAAA AAAA AAAA AAAA
1022 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1023 * cannot merge might become might become might become
1024 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1025 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1026 * mremap move: PPPPNNNNNNNN 8
1028 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1029 * might become case 1 below case 2 below case 3 below
1031 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1032 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1034 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
1035 struct vm_area_struct
*prev
, unsigned long addr
,
1036 unsigned long end
, unsigned long vm_flags
,
1037 struct anon_vma
*anon_vma
, struct file
*file
,
1038 pgoff_t pgoff
, struct mempolicy
*policy
)
1040 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
1041 struct vm_area_struct
*area
, *next
;
1045 * We later require that vma->vm_flags == vm_flags,
1046 * so this tests vma->vm_flags & VM_SPECIAL, too.
1048 if (vm_flags
& VM_SPECIAL
)
1052 next
= prev
->vm_next
;
1056 if (next
&& next
->vm_end
== end
) /* cases 6, 7, 8 */
1057 next
= next
->vm_next
;
1060 * Can it merge with the predecessor?
1062 if (prev
&& prev
->vm_end
== addr
&&
1063 mpol_equal(vma_policy(prev
), policy
) &&
1064 can_vma_merge_after(prev
, vm_flags
,
1065 anon_vma
, file
, pgoff
)) {
1067 * OK, it can. Can we now merge in the successor as well?
1069 if (next
&& end
== next
->vm_start
&&
1070 mpol_equal(policy
, vma_policy(next
)) &&
1071 can_vma_merge_before(next
, vm_flags
,
1072 anon_vma
, file
, pgoff
+pglen
) &&
1073 is_mergeable_anon_vma(prev
->anon_vma
,
1074 next
->anon_vma
, NULL
)) {
1076 err
= vma_adjust(prev
, prev
->vm_start
,
1077 next
->vm_end
, prev
->vm_pgoff
, NULL
);
1078 } else /* cases 2, 5, 7 */
1079 err
= vma_adjust(prev
, prev
->vm_start
,
1080 end
, prev
->vm_pgoff
, NULL
);
1083 khugepaged_enter_vma_merge(prev
);
1088 * Can this new request be merged in front of next?
1090 if (next
&& end
== next
->vm_start
&&
1091 mpol_equal(policy
, vma_policy(next
)) &&
1092 can_vma_merge_before(next
, vm_flags
,
1093 anon_vma
, file
, pgoff
+pglen
)) {
1094 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
1095 err
= vma_adjust(prev
, prev
->vm_start
,
1096 addr
, prev
->vm_pgoff
, NULL
);
1097 else /* cases 3, 8 */
1098 err
= vma_adjust(area
, addr
, next
->vm_end
,
1099 next
->vm_pgoff
- pglen
, NULL
);
1102 khugepaged_enter_vma_merge(area
);
1110 * Rough compatbility check to quickly see if it's even worth looking
1111 * at sharing an anon_vma.
1113 * They need to have the same vm_file, and the flags can only differ
1114 * in things that mprotect may change.
1116 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1117 * we can merge the two vma's. For example, we refuse to merge a vma if
1118 * there is a vm_ops->close() function, because that indicates that the
1119 * driver is doing some kind of reference counting. But that doesn't
1120 * really matter for the anon_vma sharing case.
1122 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1124 return a
->vm_end
== b
->vm_start
&&
1125 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
1126 a
->vm_file
== b
->vm_file
&&
1127 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_READ
|VM_WRITE
|VM_EXEC
|VM_SOFTDIRTY
)) &&
1128 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
1132 * Do some basic sanity checking to see if we can re-use the anon_vma
1133 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1134 * the same as 'old', the other will be the new one that is trying
1135 * to share the anon_vma.
1137 * NOTE! This runs with mm_sem held for reading, so it is possible that
1138 * the anon_vma of 'old' is concurrently in the process of being set up
1139 * by another page fault trying to merge _that_. But that's ok: if it
1140 * is being set up, that automatically means that it will be a singleton
1141 * acceptable for merging, so we can do all of this optimistically. But
1142 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1144 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1145 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1146 * is to return an anon_vma that is "complex" due to having gone through
1149 * We also make sure that the two vma's are compatible (adjacent,
1150 * and with the same memory policies). That's all stable, even with just
1151 * a read lock on the mm_sem.
1153 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1155 if (anon_vma_compatible(a
, b
)) {
1156 struct anon_vma
*anon_vma
= ACCESS_ONCE(old
->anon_vma
);
1158 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
1165 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1166 * neighbouring vmas for a suitable anon_vma, before it goes off
1167 * to allocate a new anon_vma. It checks because a repetitive
1168 * sequence of mprotects and faults may otherwise lead to distinct
1169 * anon_vmas being allocated, preventing vma merge in subsequent
1172 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
1174 struct anon_vma
*anon_vma
;
1175 struct vm_area_struct
*near
;
1177 near
= vma
->vm_next
;
1181 anon_vma
= reusable_anon_vma(near
, vma
, near
);
1185 near
= vma
->vm_prev
;
1189 anon_vma
= reusable_anon_vma(near
, near
, vma
);
1194 * There's no absolute need to look only at touching neighbours:
1195 * we could search further afield for "compatible" anon_vmas.
1196 * But it would probably just be a waste of time searching,
1197 * or lead to too many vmas hanging off the same anon_vma.
1198 * We're trying to allow mprotect remerging later on,
1199 * not trying to minimize memory used for anon_vmas.
1204 #ifdef CONFIG_PROC_FS
1205 void vm_stat_account(struct mm_struct
*mm
, unsigned long flags
,
1206 struct file
*file
, long pages
)
1208 const unsigned long stack_flags
1209 = VM_STACK_FLAGS
& (VM_GROWSUP
|VM_GROWSDOWN
);
1211 mm
->total_vm
+= pages
;
1214 mm
->shared_vm
+= pages
;
1215 if ((flags
& (VM_EXEC
|VM_WRITE
)) == VM_EXEC
)
1216 mm
->exec_vm
+= pages
;
1217 } else if (flags
& stack_flags
)
1218 mm
->stack_vm
+= pages
;
1220 #endif /* CONFIG_PROC_FS */
1223 * If a hint addr is less than mmap_min_addr change hint to be as
1224 * low as possible but still greater than mmap_min_addr
1226 static inline unsigned long round_hint_to_min(unsigned long hint
)
1229 if (((void *)hint
!= NULL
) &&
1230 (hint
< mmap_min_addr
))
1231 return PAGE_ALIGN(mmap_min_addr
);
1235 static inline int mlock_future_check(struct mm_struct
*mm
,
1236 unsigned long flags
,
1239 unsigned long locked
, lock_limit
;
1241 /* mlock MCL_FUTURE? */
1242 if (flags
& VM_LOCKED
) {
1243 locked
= len
>> PAGE_SHIFT
;
1244 locked
+= mm
->locked_vm
;
1245 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1246 lock_limit
>>= PAGE_SHIFT
;
1247 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1254 * The caller must hold down_write(¤t->mm->mmap_sem).
1257 unsigned long do_mmap_pgoff(struct file
*file
, unsigned long addr
,
1258 unsigned long len
, unsigned long prot
,
1259 unsigned long flags
, unsigned long pgoff
,
1260 unsigned long *populate
)
1262 struct mm_struct
*mm
= current
->mm
;
1263 vm_flags_t vm_flags
;
1268 * Does the application expect PROT_READ to imply PROT_EXEC?
1270 * (the exception is when the underlying filesystem is noexec
1271 * mounted, in which case we dont add PROT_EXEC.)
1273 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1274 if (!(file
&& (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
)))
1280 if (!(flags
& MAP_FIXED
))
1281 addr
= round_hint_to_min(addr
);
1283 /* Careful about overflows.. */
1284 len
= PAGE_ALIGN(len
);
1288 /* offset overflow? */
1289 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1292 /* Too many mappings? */
1293 if (mm
->map_count
> sysctl_max_map_count
)
1296 /* Obtain the address to map to. we verify (or select) it and ensure
1297 * that it represents a valid section of the address space.
1299 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1300 if (addr
& ~PAGE_MASK
)
1303 /* Do simple checking here so the lower-level routines won't have
1304 * to. we assume access permissions have been handled by the open
1305 * of the memory object, so we don't do any here.
1307 vm_flags
= calc_vm_prot_bits(prot
) | calc_vm_flag_bits(flags
) |
1308 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1310 if (flags
& MAP_LOCKED
)
1311 if (!can_do_mlock())
1314 if (mlock_future_check(mm
, vm_flags
, len
))
1318 struct inode
*inode
= file_inode(file
);
1320 switch (flags
& MAP_TYPE
) {
1322 if ((prot
&PROT_WRITE
) && !(file
->f_mode
&FMODE_WRITE
))
1326 * Make sure we don't allow writing to an append-only
1329 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1333 * Make sure there are no mandatory locks on the file.
1335 if (locks_verify_locked(file
))
1338 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1339 if (!(file
->f_mode
& FMODE_WRITE
))
1340 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1344 if (!(file
->f_mode
& FMODE_READ
))
1346 if (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
) {
1347 if (vm_flags
& VM_EXEC
)
1349 vm_flags
&= ~VM_MAYEXEC
;
1352 if (!file
->f_op
->mmap
)
1354 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1362 switch (flags
& MAP_TYPE
) {
1364 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1370 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1374 * Set pgoff according to addr for anon_vma.
1376 pgoff
= addr
>> PAGE_SHIFT
;
1384 * Set 'VM_NORESERVE' if we should not account for the
1385 * memory use of this mapping.
1387 if (flags
& MAP_NORESERVE
) {
1388 /* We honor MAP_NORESERVE if allowed to overcommit */
1389 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1390 vm_flags
|= VM_NORESERVE
;
1392 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1393 if (file
&& is_file_hugepages(file
))
1394 vm_flags
|= VM_NORESERVE
;
1397 addr
= mmap_region(file
, addr
, len
, vm_flags
, pgoff
);
1398 if (!IS_ERR_VALUE(addr
) &&
1399 ((vm_flags
& VM_LOCKED
) ||
1400 (flags
& (MAP_POPULATE
| MAP_NONBLOCK
)) == MAP_POPULATE
))
1405 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1406 unsigned long, prot
, unsigned long, flags
,
1407 unsigned long, fd
, unsigned long, pgoff
)
1409 struct file
*file
= NULL
;
1410 unsigned long retval
= -EBADF
;
1412 if (!(flags
& MAP_ANONYMOUS
)) {
1413 audit_mmap_fd(fd
, flags
);
1417 if (is_file_hugepages(file
))
1418 len
= ALIGN(len
, huge_page_size(hstate_file(file
)));
1420 if (unlikely(flags
& MAP_HUGETLB
&& !is_file_hugepages(file
)))
1422 } else if (flags
& MAP_HUGETLB
) {
1423 struct user_struct
*user
= NULL
;
1426 hs
= hstate_sizelog((flags
>> MAP_HUGE_SHIFT
) & SHM_HUGE_MASK
);
1430 len
= ALIGN(len
, huge_page_size(hs
));
1432 * VM_NORESERVE is used because the reservations will be
1433 * taken when vm_ops->mmap() is called
1434 * A dummy user value is used because we are not locking
1435 * memory so no accounting is necessary
1437 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, len
,
1439 &user
, HUGETLB_ANONHUGE_INODE
,
1440 (flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1442 return PTR_ERR(file
);
1445 flags
&= ~(MAP_EXECUTABLE
| MAP_DENYWRITE
);
1447 retval
= vm_mmap_pgoff(file
, addr
, len
, prot
, flags
, pgoff
);
1455 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1456 struct mmap_arg_struct
{
1460 unsigned long flags
;
1462 unsigned long offset
;
1465 SYSCALL_DEFINE1(old_mmap
, struct mmap_arg_struct __user
*, arg
)
1467 struct mmap_arg_struct a
;
1469 if (copy_from_user(&a
, arg
, sizeof(a
)))
1471 if (a
.offset
& ~PAGE_MASK
)
1474 return sys_mmap_pgoff(a
.addr
, a
.len
, a
.prot
, a
.flags
, a
.fd
,
1475 a
.offset
>> PAGE_SHIFT
);
1477 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1480 * Some shared mappigns will want the pages marked read-only
1481 * to track write events. If so, we'll downgrade vm_page_prot
1482 * to the private version (using protection_map[] without the
1485 int vma_wants_writenotify(struct vm_area_struct
*vma
)
1487 vm_flags_t vm_flags
= vma
->vm_flags
;
1489 /* If it was private or non-writable, the write bit is already clear */
1490 if ((vm_flags
& (VM_WRITE
|VM_SHARED
)) != ((VM_WRITE
|VM_SHARED
)))
1493 /* The backer wishes to know when pages are first written to? */
1494 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
)
1497 /* The open routine did something to the protections that pgprot_modify
1498 * won't preserve? */
1499 if (pgprot_val(vma
->vm_page_prot
) !=
1500 pgprot_val(vm_pgprot_modify(vma
->vm_page_prot
, vm_flags
)))
1503 /* Do we need to track softdirty? */
1504 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY
) && !(vm_flags
& VM_SOFTDIRTY
))
1507 /* Specialty mapping? */
1508 if (vm_flags
& VM_PFNMAP
)
1511 /* Can the mapping track the dirty pages? */
1512 return vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
1513 mapping_cap_account_dirty(vma
->vm_file
->f_mapping
);
1517 * We account for memory if it's a private writeable mapping,
1518 * not hugepages and VM_NORESERVE wasn't set.
1520 static inline int accountable_mapping(struct file
*file
, vm_flags_t vm_flags
)
1523 * hugetlb has its own accounting separate from the core VM
1524 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1526 if (file
&& is_file_hugepages(file
))
1529 return (vm_flags
& (VM_NORESERVE
| VM_SHARED
| VM_WRITE
)) == VM_WRITE
;
1532 unsigned long mmap_region(struct file
*file
, unsigned long addr
,
1533 unsigned long len
, vm_flags_t vm_flags
, unsigned long pgoff
)
1535 struct mm_struct
*mm
= current
->mm
;
1536 struct vm_area_struct
*vma
, *prev
;
1538 struct rb_node
**rb_link
, *rb_parent
;
1539 unsigned long charged
= 0;
1541 /* Check against address space limit. */
1542 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
)) {
1543 unsigned long nr_pages
;
1546 * MAP_FIXED may remove pages of mappings that intersects with
1547 * requested mapping. Account for the pages it would unmap.
1549 if (!(vm_flags
& MAP_FIXED
))
1552 nr_pages
= count_vma_pages_range(mm
, addr
, addr
+ len
);
1554 if (!may_expand_vm(mm
, (len
>> PAGE_SHIFT
) - nr_pages
))
1558 /* Clear old maps */
1561 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
1562 if (do_munmap(mm
, addr
, len
))
1568 * Private writable mapping: check memory availability
1570 if (accountable_mapping(file
, vm_flags
)) {
1571 charged
= len
>> PAGE_SHIFT
;
1572 if (security_vm_enough_memory_mm(mm
, charged
))
1574 vm_flags
|= VM_ACCOUNT
;
1578 * Can we just expand an old mapping?
1580 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
, NULL
, file
, pgoff
, NULL
);
1585 * Determine the object being mapped and call the appropriate
1586 * specific mapper. the address has already been validated, but
1587 * not unmapped, but the maps are removed from the list.
1589 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
1596 vma
->vm_start
= addr
;
1597 vma
->vm_end
= addr
+ len
;
1598 vma
->vm_flags
= vm_flags
;
1599 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1600 vma
->vm_pgoff
= pgoff
;
1601 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
1604 if (vm_flags
& VM_DENYWRITE
) {
1605 error
= deny_write_access(file
);
1609 if (vm_flags
& VM_SHARED
) {
1610 error
= mapping_map_writable(file
->f_mapping
);
1612 goto allow_write_and_free_vma
;
1615 /* ->mmap() can change vma->vm_file, but must guarantee that
1616 * vma_link() below can deny write-access if VM_DENYWRITE is set
1617 * and map writably if VM_SHARED is set. This usually means the
1618 * new file must not have been exposed to user-space, yet.
1620 vma
->vm_file
= get_file(file
);
1621 error
= file
->f_op
->mmap(file
, vma
);
1623 goto unmap_and_free_vma
;
1625 /* Can addr have changed??
1627 * Answer: Yes, several device drivers can do it in their
1628 * f_op->mmap method. -DaveM
1629 * Bug: If addr is changed, prev, rb_link, rb_parent should
1630 * be updated for vma_link()
1632 WARN_ON_ONCE(addr
!= vma
->vm_start
);
1634 addr
= vma
->vm_start
;
1635 vm_flags
= vma
->vm_flags
;
1636 } else if (vm_flags
& VM_SHARED
) {
1637 error
= shmem_zero_setup(vma
);
1642 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1643 /* Once vma denies write, undo our temporary denial count */
1645 if (vm_flags
& VM_SHARED
)
1646 mapping_unmap_writable(file
->f_mapping
);
1647 if (vm_flags
& VM_DENYWRITE
)
1648 allow_write_access(file
);
1650 file
= vma
->vm_file
;
1652 perf_event_mmap(vma
);
1654 vm_stat_account(mm
, vm_flags
, file
, len
>> PAGE_SHIFT
);
1655 if (vm_flags
& VM_LOCKED
) {
1656 if (!((vm_flags
& VM_SPECIAL
) || is_vm_hugetlb_page(vma
) ||
1657 vma
== get_gate_vma(current
->mm
)))
1658 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1660 vma
->vm_flags
&= ~VM_LOCKED
;
1667 * New (or expanded) vma always get soft dirty status.
1668 * Otherwise user-space soft-dirty page tracker won't
1669 * be able to distinguish situation when vma area unmapped,
1670 * then new mapped in-place (which must be aimed as
1671 * a completely new data area).
1673 vma
->vm_flags
|= VM_SOFTDIRTY
;
1675 vma_set_page_prot(vma
);
1680 vma
->vm_file
= NULL
;
1683 /* Undo any partial mapping done by a device driver. */
1684 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1686 if (vm_flags
& VM_SHARED
)
1687 mapping_unmap_writable(file
->f_mapping
);
1688 allow_write_and_free_vma
:
1689 if (vm_flags
& VM_DENYWRITE
)
1690 allow_write_access(file
);
1692 kmem_cache_free(vm_area_cachep
, vma
);
1695 vm_unacct_memory(charged
);
1699 unsigned long unmapped_area(struct vm_unmapped_area_info
*info
)
1702 * We implement the search by looking for an rbtree node that
1703 * immediately follows a suitable gap. That is,
1704 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1705 * - gap_end = vma->vm_start >= info->low_limit + length;
1706 * - gap_end - gap_start >= length
1709 struct mm_struct
*mm
= current
->mm
;
1710 struct vm_area_struct
*vma
;
1711 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1713 /* Adjust search length to account for worst case alignment overhead */
1714 length
= info
->length
+ info
->align_mask
;
1715 if (length
< info
->length
)
1718 /* Adjust search limits by the desired length */
1719 if (info
->high_limit
< length
)
1721 high_limit
= info
->high_limit
- length
;
1723 if (info
->low_limit
> high_limit
)
1725 low_limit
= info
->low_limit
+ length
;
1727 /* Check if rbtree root looks promising */
1728 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1730 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1731 if (vma
->rb_subtree_gap
< length
)
1735 /* Visit left subtree if it looks promising */
1736 gap_end
= vma
->vm_start
;
1737 if (gap_end
>= low_limit
&& vma
->vm_rb
.rb_left
) {
1738 struct vm_area_struct
*left
=
1739 rb_entry(vma
->vm_rb
.rb_left
,
1740 struct vm_area_struct
, vm_rb
);
1741 if (left
->rb_subtree_gap
>= length
) {
1747 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1749 /* Check if current node has a suitable gap */
1750 if (gap_start
> high_limit
)
1752 if (gap_end
>= low_limit
&& gap_end
- gap_start
>= length
)
1755 /* Visit right subtree if it looks promising */
1756 if (vma
->vm_rb
.rb_right
) {
1757 struct vm_area_struct
*right
=
1758 rb_entry(vma
->vm_rb
.rb_right
,
1759 struct vm_area_struct
, vm_rb
);
1760 if (right
->rb_subtree_gap
>= length
) {
1766 /* Go back up the rbtree to find next candidate node */
1768 struct rb_node
*prev
= &vma
->vm_rb
;
1769 if (!rb_parent(prev
))
1771 vma
= rb_entry(rb_parent(prev
),
1772 struct vm_area_struct
, vm_rb
);
1773 if (prev
== vma
->vm_rb
.rb_left
) {
1774 gap_start
= vma
->vm_prev
->vm_end
;
1775 gap_end
= vma
->vm_start
;
1782 /* Check highest gap, which does not precede any rbtree node */
1783 gap_start
= mm
->highest_vm_end
;
1784 gap_end
= ULONG_MAX
; /* Only for VM_BUG_ON below */
1785 if (gap_start
> high_limit
)
1789 /* We found a suitable gap. Clip it with the original low_limit. */
1790 if (gap_start
< info
->low_limit
)
1791 gap_start
= info
->low_limit
;
1793 /* Adjust gap address to the desired alignment */
1794 gap_start
+= (info
->align_offset
- gap_start
) & info
->align_mask
;
1796 VM_BUG_ON(gap_start
+ info
->length
> info
->high_limit
);
1797 VM_BUG_ON(gap_start
+ info
->length
> gap_end
);
1801 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info
*info
)
1803 struct mm_struct
*mm
= current
->mm
;
1804 struct vm_area_struct
*vma
;
1805 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1807 /* Adjust search length to account for worst case alignment overhead */
1808 length
= info
->length
+ info
->align_mask
;
1809 if (length
< info
->length
)
1813 * Adjust search limits by the desired length.
1814 * See implementation comment at top of unmapped_area().
1816 gap_end
= info
->high_limit
;
1817 if (gap_end
< length
)
1819 high_limit
= gap_end
- length
;
1821 if (info
->low_limit
> high_limit
)
1823 low_limit
= info
->low_limit
+ length
;
1825 /* Check highest gap, which does not precede any rbtree node */
1826 gap_start
= mm
->highest_vm_end
;
1827 if (gap_start
<= high_limit
)
1830 /* Check if rbtree root looks promising */
1831 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1833 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1834 if (vma
->rb_subtree_gap
< length
)
1838 /* Visit right subtree if it looks promising */
1839 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1840 if (gap_start
<= high_limit
&& vma
->vm_rb
.rb_right
) {
1841 struct vm_area_struct
*right
=
1842 rb_entry(vma
->vm_rb
.rb_right
,
1843 struct vm_area_struct
, vm_rb
);
1844 if (right
->rb_subtree_gap
>= length
) {
1851 /* Check if current node has a suitable gap */
1852 gap_end
= vma
->vm_start
;
1853 if (gap_end
< low_limit
)
1855 if (gap_start
<= high_limit
&& gap_end
- gap_start
>= length
)
1858 /* Visit left subtree if it looks promising */
1859 if (vma
->vm_rb
.rb_left
) {
1860 struct vm_area_struct
*left
=
1861 rb_entry(vma
->vm_rb
.rb_left
,
1862 struct vm_area_struct
, vm_rb
);
1863 if (left
->rb_subtree_gap
>= length
) {
1869 /* Go back up the rbtree to find next candidate node */
1871 struct rb_node
*prev
= &vma
->vm_rb
;
1872 if (!rb_parent(prev
))
1874 vma
= rb_entry(rb_parent(prev
),
1875 struct vm_area_struct
, vm_rb
);
1876 if (prev
== vma
->vm_rb
.rb_right
) {
1877 gap_start
= vma
->vm_prev
?
1878 vma
->vm_prev
->vm_end
: 0;
1885 /* We found a suitable gap. Clip it with the original high_limit. */
1886 if (gap_end
> info
->high_limit
)
1887 gap_end
= info
->high_limit
;
1890 /* Compute highest gap address at the desired alignment */
1891 gap_end
-= info
->length
;
1892 gap_end
-= (gap_end
- info
->align_offset
) & info
->align_mask
;
1894 VM_BUG_ON(gap_end
< info
->low_limit
);
1895 VM_BUG_ON(gap_end
< gap_start
);
1899 /* Get an address range which is currently unmapped.
1900 * For shmat() with addr=0.
1902 * Ugly calling convention alert:
1903 * Return value with the low bits set means error value,
1905 * if (ret & ~PAGE_MASK)
1908 * This function "knows" that -ENOMEM has the bits set.
1910 #ifndef HAVE_ARCH_UNMAPPED_AREA
1912 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
1913 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
1915 struct mm_struct
*mm
= current
->mm
;
1916 struct vm_area_struct
*vma
;
1917 struct vm_unmapped_area_info info
;
1919 if (len
> TASK_SIZE
- mmap_min_addr
)
1922 if (flags
& MAP_FIXED
)
1926 addr
= PAGE_ALIGN(addr
);
1927 vma
= find_vma(mm
, addr
);
1928 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1929 (!vma
|| addr
+ len
<= vma
->vm_start
))
1935 info
.low_limit
= mm
->mmap_base
;
1936 info
.high_limit
= TASK_SIZE
;
1937 info
.align_mask
= 0;
1938 return vm_unmapped_area(&info
);
1943 * This mmap-allocator allocates new areas top-down from below the
1944 * stack's low limit (the base):
1946 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1948 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
1949 const unsigned long len
, const unsigned long pgoff
,
1950 const unsigned long flags
)
1952 struct vm_area_struct
*vma
;
1953 struct mm_struct
*mm
= current
->mm
;
1954 unsigned long addr
= addr0
;
1955 struct vm_unmapped_area_info info
;
1957 /* requested length too big for entire address space */
1958 if (len
> TASK_SIZE
- mmap_min_addr
)
1961 if (flags
& MAP_FIXED
)
1964 /* requesting a specific address */
1966 addr
= PAGE_ALIGN(addr
);
1967 vma
= find_vma(mm
, addr
);
1968 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1969 (!vma
|| addr
+ len
<= vma
->vm_start
))
1973 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
1975 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
1976 info
.high_limit
= mm
->mmap_base
;
1977 info
.align_mask
= 0;
1978 addr
= vm_unmapped_area(&info
);
1981 * A failed mmap() very likely causes application failure,
1982 * so fall back to the bottom-up function here. This scenario
1983 * can happen with large stack limits and large mmap()
1986 if (addr
& ~PAGE_MASK
) {
1987 VM_BUG_ON(addr
!= -ENOMEM
);
1989 info
.low_limit
= TASK_UNMAPPED_BASE
;
1990 info
.high_limit
= TASK_SIZE
;
1991 addr
= vm_unmapped_area(&info
);
1999 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
2000 unsigned long pgoff
, unsigned long flags
)
2002 unsigned long (*get_area
)(struct file
*, unsigned long,
2003 unsigned long, unsigned long, unsigned long);
2005 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
2009 /* Careful about overflows.. */
2010 if (len
> TASK_SIZE
)
2013 get_area
= current
->mm
->get_unmapped_area
;
2014 if (file
&& file
->f_op
->get_unmapped_area
)
2015 get_area
= file
->f_op
->get_unmapped_area
;
2016 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
2017 if (IS_ERR_VALUE(addr
))
2020 if (addr
> TASK_SIZE
- len
)
2022 if (addr
& ~PAGE_MASK
)
2025 addr
= arch_rebalance_pgtables(addr
, len
);
2026 error
= security_mmap_addr(addr
);
2027 return error
? error
: addr
;
2030 EXPORT_SYMBOL(get_unmapped_area
);
2032 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2033 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
2035 struct rb_node
*rb_node
;
2036 struct vm_area_struct
*vma
;
2038 /* Check the cache first. */
2039 vma
= vmacache_find(mm
, addr
);
2043 rb_node
= mm
->mm_rb
.rb_node
;
2047 struct vm_area_struct
*tmp
;
2049 tmp
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2051 if (tmp
->vm_end
> addr
) {
2053 if (tmp
->vm_start
<= addr
)
2055 rb_node
= rb_node
->rb_left
;
2057 rb_node
= rb_node
->rb_right
;
2061 vmacache_update(addr
, vma
);
2065 EXPORT_SYMBOL(find_vma
);
2068 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2070 struct vm_area_struct
*
2071 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
2072 struct vm_area_struct
**pprev
)
2074 struct vm_area_struct
*vma
;
2076 vma
= find_vma(mm
, addr
);
2078 *pprev
= vma
->vm_prev
;
2080 struct rb_node
*rb_node
= mm
->mm_rb
.rb_node
;
2083 *pprev
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2084 rb_node
= rb_node
->rb_right
;
2091 * Verify that the stack growth is acceptable and
2092 * update accounting. This is shared with both the
2093 * grow-up and grow-down cases.
2095 static int acct_stack_growth(struct vm_area_struct
*vma
, unsigned long size
, unsigned long grow
)
2097 struct mm_struct
*mm
= vma
->vm_mm
;
2098 struct rlimit
*rlim
= current
->signal
->rlim
;
2099 unsigned long new_start
;
2101 /* address space limit tests */
2102 if (!may_expand_vm(mm
, grow
))
2105 /* Stack limit test */
2106 if (size
> ACCESS_ONCE(rlim
[RLIMIT_STACK
].rlim_cur
))
2109 /* mlock limit tests */
2110 if (vma
->vm_flags
& VM_LOCKED
) {
2111 unsigned long locked
;
2112 unsigned long limit
;
2113 locked
= mm
->locked_vm
+ grow
;
2114 limit
= ACCESS_ONCE(rlim
[RLIMIT_MEMLOCK
].rlim_cur
);
2115 limit
>>= PAGE_SHIFT
;
2116 if (locked
> limit
&& !capable(CAP_IPC_LOCK
))
2120 /* Check to ensure the stack will not grow into a hugetlb-only region */
2121 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
2123 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
2127 * Overcommit.. This must be the final test, as it will
2128 * update security statistics.
2130 if (security_vm_enough_memory_mm(mm
, grow
))
2133 /* Ok, everything looks good - let it rip */
2134 if (vma
->vm_flags
& VM_LOCKED
)
2135 mm
->locked_vm
+= grow
;
2136 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, grow
);
2140 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2142 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2143 * vma is the last one with address > vma->vm_end. Have to extend vma.
2145 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
)
2149 if (!(vma
->vm_flags
& VM_GROWSUP
))
2153 * We must make sure the anon_vma is allocated
2154 * so that the anon_vma locking is not a noop.
2156 if (unlikely(anon_vma_prepare(vma
)))
2158 vma_lock_anon_vma(vma
);
2161 * vma->vm_start/vm_end cannot change under us because the caller
2162 * is required to hold the mmap_sem in read mode. We need the
2163 * anon_vma lock to serialize against concurrent expand_stacks.
2164 * Also guard against wrapping around to address 0.
2166 if (address
< PAGE_ALIGN(address
+4))
2167 address
= PAGE_ALIGN(address
+4);
2169 vma_unlock_anon_vma(vma
);
2174 /* Somebody else might have raced and expanded it already */
2175 if (address
> vma
->vm_end
) {
2176 unsigned long size
, grow
;
2178 size
= address
- vma
->vm_start
;
2179 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
2182 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
2183 error
= acct_stack_growth(vma
, size
, grow
);
2186 * vma_gap_update() doesn't support concurrent
2187 * updates, but we only hold a shared mmap_sem
2188 * lock here, so we need to protect against
2189 * concurrent vma expansions.
2190 * vma_lock_anon_vma() doesn't help here, as
2191 * we don't guarantee that all growable vmas
2192 * in a mm share the same root anon vma.
2193 * So, we reuse mm->page_table_lock to guard
2194 * against concurrent vma expansions.
2196 spin_lock(&vma
->vm_mm
->page_table_lock
);
2197 anon_vma_interval_tree_pre_update_vma(vma
);
2198 vma
->vm_end
= address
;
2199 anon_vma_interval_tree_post_update_vma(vma
);
2201 vma_gap_update(vma
->vm_next
);
2203 vma
->vm_mm
->highest_vm_end
= address
;
2204 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2206 perf_event_mmap(vma
);
2210 vma_unlock_anon_vma(vma
);
2211 khugepaged_enter_vma_merge(vma
);
2212 validate_mm(vma
->vm_mm
);
2215 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2218 * vma is the first one with address < vma->vm_start. Have to extend vma.
2220 int expand_downwards(struct vm_area_struct
*vma
,
2221 unsigned long address
)
2226 * We must make sure the anon_vma is allocated
2227 * so that the anon_vma locking is not a noop.
2229 if (unlikely(anon_vma_prepare(vma
)))
2232 address
&= PAGE_MASK
;
2233 error
= security_mmap_addr(address
);
2237 vma_lock_anon_vma(vma
);
2240 * vma->vm_start/vm_end cannot change under us because the caller
2241 * is required to hold the mmap_sem in read mode. We need the
2242 * anon_vma lock to serialize against concurrent expand_stacks.
2245 /* Somebody else might have raced and expanded it already */
2246 if (address
< vma
->vm_start
) {
2247 unsigned long size
, grow
;
2249 size
= vma
->vm_end
- address
;
2250 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
2253 if (grow
<= vma
->vm_pgoff
) {
2254 error
= acct_stack_growth(vma
, size
, grow
);
2257 * vma_gap_update() doesn't support concurrent
2258 * updates, but we only hold a shared mmap_sem
2259 * lock here, so we need to protect against
2260 * concurrent vma expansions.
2261 * vma_lock_anon_vma() doesn't help here, as
2262 * we don't guarantee that all growable vmas
2263 * in a mm share the same root anon vma.
2264 * So, we reuse mm->page_table_lock to guard
2265 * against concurrent vma expansions.
2267 spin_lock(&vma
->vm_mm
->page_table_lock
);
2268 anon_vma_interval_tree_pre_update_vma(vma
);
2269 vma
->vm_start
= address
;
2270 vma
->vm_pgoff
-= grow
;
2271 anon_vma_interval_tree_post_update_vma(vma
);
2272 vma_gap_update(vma
);
2273 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2275 perf_event_mmap(vma
);
2279 vma_unlock_anon_vma(vma
);
2280 khugepaged_enter_vma_merge(vma
);
2281 validate_mm(vma
->vm_mm
);
2286 * Note how expand_stack() refuses to expand the stack all the way to
2287 * abut the next virtual mapping, *unless* that mapping itself is also
2288 * a stack mapping. We want to leave room for a guard page, after all
2289 * (the guard page itself is not added here, that is done by the
2290 * actual page faulting logic)
2292 * This matches the behavior of the guard page logic (see mm/memory.c:
2293 * check_stack_guard_page()), which only allows the guard page to be
2294 * removed under these circumstances.
2296 #ifdef CONFIG_STACK_GROWSUP
2297 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2299 struct vm_area_struct
*next
;
2301 address
&= PAGE_MASK
;
2302 next
= vma
->vm_next
;
2303 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
) {
2304 if (!(next
->vm_flags
& VM_GROWSUP
))
2307 return expand_upwards(vma
, address
);
2310 struct vm_area_struct
*
2311 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2313 struct vm_area_struct
*vma
, *prev
;
2316 vma
= find_vma_prev(mm
, addr
, &prev
);
2317 if (vma
&& (vma
->vm_start
<= addr
))
2319 if (!prev
|| expand_stack(prev
, addr
))
2321 if (prev
->vm_flags
& VM_LOCKED
)
2322 __mlock_vma_pages_range(prev
, addr
, prev
->vm_end
, NULL
);
2326 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2328 struct vm_area_struct
*prev
;
2330 address
&= PAGE_MASK
;
2331 prev
= vma
->vm_prev
;
2332 if (prev
&& prev
->vm_end
== address
) {
2333 if (!(prev
->vm_flags
& VM_GROWSDOWN
))
2336 return expand_downwards(vma
, address
);
2339 struct vm_area_struct
*
2340 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2342 struct vm_area_struct
*vma
;
2343 unsigned long start
;
2346 vma
= find_vma(mm
, addr
);
2349 if (vma
->vm_start
<= addr
)
2351 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2353 start
= vma
->vm_start
;
2354 if (expand_stack(vma
, addr
))
2356 if (vma
->vm_flags
& VM_LOCKED
)
2357 __mlock_vma_pages_range(vma
, addr
, start
, NULL
);
2363 * Ok - we have the memory areas we should free on the vma list,
2364 * so release them, and do the vma updates.
2366 * Called with the mm semaphore held.
2368 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2370 unsigned long nr_accounted
= 0;
2372 /* Update high watermark before we lower total_vm */
2373 update_hiwater_vm(mm
);
2375 long nrpages
= vma_pages(vma
);
2377 if (vma
->vm_flags
& VM_ACCOUNT
)
2378 nr_accounted
+= nrpages
;
2379 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, -nrpages
);
2380 vma
= remove_vma(vma
);
2382 vm_unacct_memory(nr_accounted
);
2387 * Get rid of page table information in the indicated region.
2389 * Called with the mm semaphore held.
2391 static void unmap_region(struct mm_struct
*mm
,
2392 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
2393 unsigned long start
, unsigned long end
)
2395 struct vm_area_struct
*next
= prev
? prev
->vm_next
: mm
->mmap
;
2396 struct mmu_gather tlb
;
2399 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2400 update_hiwater_rss(mm
);
2401 unmap_vmas(&tlb
, vma
, start
, end
);
2402 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
2403 next
? next
->vm_start
: USER_PGTABLES_CEILING
);
2404 tlb_finish_mmu(&tlb
, start
, end
);
2408 * Create a list of vma's touched by the unmap, removing them from the mm's
2409 * vma list as we go..
2412 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2413 struct vm_area_struct
*prev
, unsigned long end
)
2415 struct vm_area_struct
**insertion_point
;
2416 struct vm_area_struct
*tail_vma
= NULL
;
2418 insertion_point
= (prev
? &prev
->vm_next
: &mm
->mmap
);
2419 vma
->vm_prev
= NULL
;
2421 vma_rb_erase(vma
, &mm
->mm_rb
);
2425 } while (vma
&& vma
->vm_start
< end
);
2426 *insertion_point
= vma
;
2428 vma
->vm_prev
= prev
;
2429 vma_gap_update(vma
);
2431 mm
->highest_vm_end
= prev
? prev
->vm_end
: 0;
2432 tail_vma
->vm_next
= NULL
;
2434 /* Kill the cache */
2435 vmacache_invalidate(mm
);
2439 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2440 * munmap path where it doesn't make sense to fail.
2442 static int __split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2443 unsigned long addr
, int new_below
)
2445 struct vm_area_struct
*new;
2448 if (is_vm_hugetlb_page(vma
) && (addr
&
2449 ~(huge_page_mask(hstate_vma(vma
)))))
2452 new = kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2456 /* most fields are the same, copy all, and then fixup */
2459 INIT_LIST_HEAD(&new->anon_vma_chain
);
2464 new->vm_start
= addr
;
2465 new->vm_pgoff
+= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
);
2468 err
= vma_dup_policy(vma
, new);
2472 if (anon_vma_clone(new, vma
))
2476 get_file(new->vm_file
);
2478 if (new->vm_ops
&& new->vm_ops
->open
)
2479 new->vm_ops
->open(new);
2482 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2483 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2485 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2491 /* Clean everything up if vma_adjust failed. */
2492 if (new->vm_ops
&& new->vm_ops
->close
)
2493 new->vm_ops
->close(new);
2496 unlink_anon_vmas(new);
2498 mpol_put(vma_policy(new));
2500 kmem_cache_free(vm_area_cachep
, new);
2506 * Split a vma into two pieces at address 'addr', a new vma is allocated
2507 * either for the first part or the tail.
2509 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2510 unsigned long addr
, int new_below
)
2512 if (mm
->map_count
>= sysctl_max_map_count
)
2515 return __split_vma(mm
, vma
, addr
, new_below
);
2518 /* Munmap is split into 2 main parts -- this part which finds
2519 * what needs doing, and the areas themselves, which do the
2520 * work. This now handles partial unmappings.
2521 * Jeremy Fitzhardinge <jeremy@goop.org>
2523 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
)
2526 struct vm_area_struct
*vma
, *prev
, *last
;
2528 if ((start
& ~PAGE_MASK
) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2531 len
= PAGE_ALIGN(len
);
2535 /* Find the first overlapping VMA */
2536 vma
= find_vma(mm
, start
);
2539 prev
= vma
->vm_prev
;
2540 /* we have start < vma->vm_end */
2542 /* if it doesn't overlap, we have nothing.. */
2544 if (vma
->vm_start
>= end
)
2548 * If we need to split any vma, do it now to save pain later.
2550 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2551 * unmapped vm_area_struct will remain in use: so lower split_vma
2552 * places tmp vma above, and higher split_vma places tmp vma below.
2554 if (start
> vma
->vm_start
) {
2558 * Make sure that map_count on return from munmap() will
2559 * not exceed its limit; but let map_count go just above
2560 * its limit temporarily, to help free resources as expected.
2562 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2565 error
= __split_vma(mm
, vma
, start
, 0);
2571 /* Does it split the last one? */
2572 last
= find_vma(mm
, end
);
2573 if (last
&& end
> last
->vm_start
) {
2574 int error
= __split_vma(mm
, last
, end
, 1);
2578 vma
= prev
? prev
->vm_next
: mm
->mmap
;
2581 * unlock any mlock()ed ranges before detaching vmas
2583 if (mm
->locked_vm
) {
2584 struct vm_area_struct
*tmp
= vma
;
2585 while (tmp
&& tmp
->vm_start
< end
) {
2586 if (tmp
->vm_flags
& VM_LOCKED
) {
2587 mm
->locked_vm
-= vma_pages(tmp
);
2588 munlock_vma_pages_all(tmp
);
2595 * Remove the vma's, and unmap the actual pages
2597 detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
);
2598 unmap_region(mm
, vma
, prev
, start
, end
);
2600 /* Fix up all other VM information */
2601 remove_vma_list(mm
, vma
);
2606 int vm_munmap(unsigned long start
, size_t len
)
2609 struct mm_struct
*mm
= current
->mm
;
2611 down_write(&mm
->mmap_sem
);
2612 ret
= do_munmap(mm
, start
, len
);
2613 up_write(&mm
->mmap_sem
);
2616 EXPORT_SYMBOL(vm_munmap
);
2618 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2620 profile_munmap(addr
);
2621 return vm_munmap(addr
, len
);
2624 static inline void verify_mm_writelocked(struct mm_struct
*mm
)
2626 #ifdef CONFIG_DEBUG_VM
2627 if (unlikely(down_read_trylock(&mm
->mmap_sem
))) {
2629 up_read(&mm
->mmap_sem
);
2635 * this is really a simplified "do_mmap". it only handles
2636 * anonymous maps. eventually we may be able to do some
2637 * brk-specific accounting here.
2639 static unsigned long do_brk(unsigned long addr
, unsigned long len
)
2641 struct mm_struct
*mm
= current
->mm
;
2642 struct vm_area_struct
*vma
, *prev
;
2643 unsigned long flags
;
2644 struct rb_node
**rb_link
, *rb_parent
;
2645 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2648 len
= PAGE_ALIGN(len
);
2652 flags
= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2654 error
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2655 if (error
& ~PAGE_MASK
)
2658 error
= mlock_future_check(mm
, mm
->def_flags
, len
);
2663 * mm->mmap_sem is required to protect against another thread
2664 * changing the mappings in case we sleep.
2666 verify_mm_writelocked(mm
);
2669 * Clear old maps. this also does some error checking for us
2672 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
2673 if (do_munmap(mm
, addr
, len
))
2678 /* Check against address space limits *after* clearing old maps... */
2679 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
))
2682 if (mm
->map_count
> sysctl_max_map_count
)
2685 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
2688 /* Can we just expand an old private anonymous mapping? */
2689 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
2690 NULL
, NULL
, pgoff
, NULL
);
2695 * create a vma struct for an anonymous mapping
2697 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2699 vm_unacct_memory(len
>> PAGE_SHIFT
);
2703 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2705 vma
->vm_start
= addr
;
2706 vma
->vm_end
= addr
+ len
;
2707 vma
->vm_pgoff
= pgoff
;
2708 vma
->vm_flags
= flags
;
2709 vma
->vm_page_prot
= vm_get_page_prot(flags
);
2710 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2712 perf_event_mmap(vma
);
2713 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2714 if (flags
& VM_LOCKED
)
2715 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
2716 vma
->vm_flags
|= VM_SOFTDIRTY
;
2720 unsigned long vm_brk(unsigned long addr
, unsigned long len
)
2722 struct mm_struct
*mm
= current
->mm
;
2726 down_write(&mm
->mmap_sem
);
2727 ret
= do_brk(addr
, len
);
2728 populate
= ((mm
->def_flags
& VM_LOCKED
) != 0);
2729 up_write(&mm
->mmap_sem
);
2731 mm_populate(addr
, len
);
2734 EXPORT_SYMBOL(vm_brk
);
2736 /* Release all mmaps. */
2737 void exit_mmap(struct mm_struct
*mm
)
2739 struct mmu_gather tlb
;
2740 struct vm_area_struct
*vma
;
2741 unsigned long nr_accounted
= 0;
2743 /* mm's last user has gone, and its about to be pulled down */
2744 mmu_notifier_release(mm
);
2746 if (mm
->locked_vm
) {
2749 if (vma
->vm_flags
& VM_LOCKED
)
2750 munlock_vma_pages_all(vma
);
2758 if (!vma
) /* Can happen if dup_mmap() received an OOM */
2763 tlb_gather_mmu(&tlb
, mm
, 0, -1);
2764 /* update_hiwater_rss(mm) here? but nobody should be looking */
2765 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2766 unmap_vmas(&tlb
, vma
, 0, -1);
2768 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, USER_PGTABLES_CEILING
);
2769 tlb_finish_mmu(&tlb
, 0, -1);
2772 * Walk the list again, actually closing and freeing it,
2773 * with preemption enabled, without holding any MM locks.
2776 if (vma
->vm_flags
& VM_ACCOUNT
)
2777 nr_accounted
+= vma_pages(vma
);
2778 vma
= remove_vma(vma
);
2780 vm_unacct_memory(nr_accounted
);
2782 WARN_ON(atomic_long_read(&mm
->nr_ptes
) >
2783 (FIRST_USER_ADDRESS
+PMD_SIZE
-1)>>PMD_SHIFT
);
2786 /* Insert vm structure into process list sorted by address
2787 * and into the inode's i_mmap tree. If vm_file is non-NULL
2788 * then i_mmap_mutex is taken here.
2790 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2792 struct vm_area_struct
*prev
;
2793 struct rb_node
**rb_link
, *rb_parent
;
2796 * The vm_pgoff of a purely anonymous vma should be irrelevant
2797 * until its first write fault, when page's anon_vma and index
2798 * are set. But now set the vm_pgoff it will almost certainly
2799 * end up with (unless mremap moves it elsewhere before that
2800 * first wfault), so /proc/pid/maps tells a consistent story.
2802 * By setting it to reflect the virtual start address of the
2803 * vma, merges and splits can happen in a seamless way, just
2804 * using the existing file pgoff checks and manipulations.
2805 * Similarly in do_mmap_pgoff and in do_brk.
2807 if (!vma
->vm_file
) {
2808 BUG_ON(vma
->anon_vma
);
2809 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2811 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
2812 &prev
, &rb_link
, &rb_parent
))
2814 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
2815 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
2818 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2823 * Copy the vma structure to a new location in the same mm,
2824 * prior to moving page table entries, to effect an mremap move.
2826 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
2827 unsigned long addr
, unsigned long len
, pgoff_t pgoff
,
2828 bool *need_rmap_locks
)
2830 struct vm_area_struct
*vma
= *vmap
;
2831 unsigned long vma_start
= vma
->vm_start
;
2832 struct mm_struct
*mm
= vma
->vm_mm
;
2833 struct vm_area_struct
*new_vma
, *prev
;
2834 struct rb_node
**rb_link
, *rb_parent
;
2835 bool faulted_in_anon_vma
= true;
2838 * If anonymous vma has not yet been faulted, update new pgoff
2839 * to match new location, to increase its chance of merging.
2841 if (unlikely(!vma
->vm_file
&& !vma
->anon_vma
)) {
2842 pgoff
= addr
>> PAGE_SHIFT
;
2843 faulted_in_anon_vma
= false;
2846 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
2847 return NULL
; /* should never get here */
2848 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
2849 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
));
2852 * Source vma may have been merged into new_vma
2854 if (unlikely(vma_start
>= new_vma
->vm_start
&&
2855 vma_start
< new_vma
->vm_end
)) {
2857 * The only way we can get a vma_merge with
2858 * self during an mremap is if the vma hasn't
2859 * been faulted in yet and we were allowed to
2860 * reset the dst vma->vm_pgoff to the
2861 * destination address of the mremap to allow
2862 * the merge to happen. mremap must change the
2863 * vm_pgoff linearity between src and dst vmas
2864 * (in turn preventing a vma_merge) to be
2865 * safe. It is only safe to keep the vm_pgoff
2866 * linear if there are no pages mapped yet.
2868 VM_BUG_ON_VMA(faulted_in_anon_vma
, new_vma
);
2869 *vmap
= vma
= new_vma
;
2871 *need_rmap_locks
= (new_vma
->vm_pgoff
<= vma
->vm_pgoff
);
2873 new_vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2876 new_vma
->vm_start
= addr
;
2877 new_vma
->vm_end
= addr
+ len
;
2878 new_vma
->vm_pgoff
= pgoff
;
2879 if (vma_dup_policy(vma
, new_vma
))
2881 INIT_LIST_HEAD(&new_vma
->anon_vma_chain
);
2882 if (anon_vma_clone(new_vma
, vma
))
2883 goto out_free_mempol
;
2884 if (new_vma
->vm_file
)
2885 get_file(new_vma
->vm_file
);
2886 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
2887 new_vma
->vm_ops
->open(new_vma
);
2888 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
2889 *need_rmap_locks
= false;
2895 mpol_put(vma_policy(new_vma
));
2897 kmem_cache_free(vm_area_cachep
, new_vma
);
2902 * Return true if the calling process may expand its vm space by the passed
2905 int may_expand_vm(struct mm_struct
*mm
, unsigned long npages
)
2907 unsigned long cur
= mm
->total_vm
; /* pages */
2910 lim
= rlimit(RLIMIT_AS
) >> PAGE_SHIFT
;
2912 if (cur
+ npages
> lim
)
2917 static int special_mapping_fault(struct vm_area_struct
*vma
,
2918 struct vm_fault
*vmf
);
2921 * Having a close hook prevents vma merging regardless of flags.
2923 static void special_mapping_close(struct vm_area_struct
*vma
)
2927 static const char *special_mapping_name(struct vm_area_struct
*vma
)
2929 return ((struct vm_special_mapping
*)vma
->vm_private_data
)->name
;
2932 static const struct vm_operations_struct special_mapping_vmops
= {
2933 .close
= special_mapping_close
,
2934 .fault
= special_mapping_fault
,
2935 .name
= special_mapping_name
,
2938 static const struct vm_operations_struct legacy_special_mapping_vmops
= {
2939 .close
= special_mapping_close
,
2940 .fault
= special_mapping_fault
,
2943 static int special_mapping_fault(struct vm_area_struct
*vma
,
2944 struct vm_fault
*vmf
)
2947 struct page
**pages
;
2950 * special mappings have no vm_file, and in that case, the mm
2951 * uses vm_pgoff internally. So we have to subtract it from here.
2952 * We are allowed to do this because we are the mm; do not copy
2953 * this code into drivers!
2955 pgoff
= vmf
->pgoff
- vma
->vm_pgoff
;
2957 if (vma
->vm_ops
== &legacy_special_mapping_vmops
)
2958 pages
= vma
->vm_private_data
;
2960 pages
= ((struct vm_special_mapping
*)vma
->vm_private_data
)->
2963 for (; pgoff
&& *pages
; ++pages
)
2967 struct page
*page
= *pages
;
2973 return VM_FAULT_SIGBUS
;
2976 static struct vm_area_struct
*__install_special_mapping(
2977 struct mm_struct
*mm
,
2978 unsigned long addr
, unsigned long len
,
2979 unsigned long vm_flags
, const struct vm_operations_struct
*ops
,
2983 struct vm_area_struct
*vma
;
2985 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2986 if (unlikely(vma
== NULL
))
2987 return ERR_PTR(-ENOMEM
);
2989 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2991 vma
->vm_start
= addr
;
2992 vma
->vm_end
= addr
+ len
;
2994 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
| VM_SOFTDIRTY
;
2995 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
2998 vma
->vm_private_data
= priv
;
3000 ret
= insert_vm_struct(mm
, vma
);
3004 mm
->total_vm
+= len
>> PAGE_SHIFT
;
3006 perf_event_mmap(vma
);
3011 kmem_cache_free(vm_area_cachep
, vma
);
3012 return ERR_PTR(ret
);
3016 * Called with mm->mmap_sem held for writing.
3017 * Insert a new vma covering the given region, with the given flags.
3018 * Its pages are supplied by the given array of struct page *.
3019 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3020 * The region past the last page supplied will always produce SIGBUS.
3021 * The array pointer and the pages it points to are assumed to stay alive
3022 * for as long as this mapping might exist.
3024 struct vm_area_struct
*_install_special_mapping(
3025 struct mm_struct
*mm
,
3026 unsigned long addr
, unsigned long len
,
3027 unsigned long vm_flags
, const struct vm_special_mapping
*spec
)
3029 return __install_special_mapping(mm
, addr
, len
, vm_flags
,
3030 &special_mapping_vmops
, (void *)spec
);
3033 int install_special_mapping(struct mm_struct
*mm
,
3034 unsigned long addr
, unsigned long len
,
3035 unsigned long vm_flags
, struct page
**pages
)
3037 struct vm_area_struct
*vma
= __install_special_mapping(
3038 mm
, addr
, len
, vm_flags
, &legacy_special_mapping_vmops
,
3041 return PTR_ERR_OR_ZERO(vma
);
3044 static DEFINE_MUTEX(mm_all_locks_mutex
);
3046 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
3048 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3050 * The LSB of head.next can't change from under us
3051 * because we hold the mm_all_locks_mutex.
3053 down_write_nest_lock(&anon_vma
->root
->rwsem
, &mm
->mmap_sem
);
3055 * We can safely modify head.next after taking the
3056 * anon_vma->root->rwsem. If some other vma in this mm shares
3057 * the same anon_vma we won't take it again.
3059 * No need of atomic instructions here, head.next
3060 * can't change from under us thanks to the
3061 * anon_vma->root->rwsem.
3063 if (__test_and_set_bit(0, (unsigned long *)
3064 &anon_vma
->root
->rb_root
.rb_node
))
3069 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
3071 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3073 * AS_MM_ALL_LOCKS can't change from under us because
3074 * we hold the mm_all_locks_mutex.
3076 * Operations on ->flags have to be atomic because
3077 * even if AS_MM_ALL_LOCKS is stable thanks to the
3078 * mm_all_locks_mutex, there may be other cpus
3079 * changing other bitflags in parallel to us.
3081 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
3083 mutex_lock_nest_lock(&mapping
->i_mmap_mutex
, &mm
->mmap_sem
);
3088 * This operation locks against the VM for all pte/vma/mm related
3089 * operations that could ever happen on a certain mm. This includes
3090 * vmtruncate, try_to_unmap, and all page faults.
3092 * The caller must take the mmap_sem in write mode before calling
3093 * mm_take_all_locks(). The caller isn't allowed to release the
3094 * mmap_sem until mm_drop_all_locks() returns.
3096 * mmap_sem in write mode is required in order to block all operations
3097 * that could modify pagetables and free pages without need of
3098 * altering the vma layout (for example populate_range() with
3099 * nonlinear vmas). It's also needed in write mode to avoid new
3100 * anon_vmas to be associated with existing vmas.
3102 * A single task can't take more than one mm_take_all_locks() in a row
3103 * or it would deadlock.
3105 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3106 * mapping->flags avoid to take the same lock twice, if more than one
3107 * vma in this mm is backed by the same anon_vma or address_space.
3109 * We can take all the locks in random order because the VM code
3110 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3111 * takes more than one of them in a row. Secondly we're protected
3112 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3114 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3115 * that may have to take thousand of locks.
3117 * mm_take_all_locks() can fail if it's interrupted by signals.
3119 int mm_take_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
));
3126 mutex_lock(&mm_all_locks_mutex
);
3128 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3129 if (signal_pending(current
))
3131 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3132 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3135 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3136 if (signal_pending(current
))
3139 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3140 vm_lock_anon_vma(mm
, avc
->anon_vma
);
3146 mm_drop_all_locks(mm
);
3150 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
3152 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3154 * The LSB of head.next can't change to 0 from under
3155 * us because we hold the mm_all_locks_mutex.
3157 * We must however clear the bitflag before unlocking
3158 * the vma so the users using the anon_vma->rb_root will
3159 * never see our bitflag.
3161 * No need of atomic instructions here, head.next
3162 * can't change from under us until we release the
3163 * anon_vma->root->rwsem.
3165 if (!__test_and_clear_bit(0, (unsigned long *)
3166 &anon_vma
->root
->rb_root
.rb_node
))
3168 anon_vma_unlock_write(anon_vma
);
3172 static void vm_unlock_mapping(struct address_space
*mapping
)
3174 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3176 * AS_MM_ALL_LOCKS can't change to 0 from under us
3177 * because we hold the mm_all_locks_mutex.
3179 mutex_unlock(&mapping
->i_mmap_mutex
);
3180 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
3187 * The mmap_sem cannot be released by the caller until
3188 * mm_drop_all_locks() returns.
3190 void mm_drop_all_locks(struct mm_struct
*mm
)
3192 struct vm_area_struct
*vma
;
3193 struct anon_vma_chain
*avc
;
3195 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3196 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
3198 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3200 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3201 vm_unlock_anon_vma(avc
->anon_vma
);
3202 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3203 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
3206 mutex_unlock(&mm_all_locks_mutex
);
3210 * initialise the VMA slab
3212 void __init
mmap_init(void)
3216 ret
= percpu_counter_init(&vm_committed_as
, 0, GFP_KERNEL
);
3221 * Initialise sysctl_user_reserve_kbytes.
3223 * This is intended to prevent a user from starting a single memory hogging
3224 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3227 * The default value is min(3% of free memory, 128MB)
3228 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3230 static int init_user_reserve(void)
3232 unsigned long free_kbytes
;
3234 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3236 sysctl_user_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 17);
3239 subsys_initcall(init_user_reserve
);
3242 * Initialise sysctl_admin_reserve_kbytes.
3244 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3245 * to log in and kill a memory hogging process.
3247 * Systems with more than 256MB will reserve 8MB, enough to recover
3248 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3249 * only reserve 3% of free pages by default.
3251 static int init_admin_reserve(void)
3253 unsigned long free_kbytes
;
3255 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3257 sysctl_admin_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 13);
3260 subsys_initcall(init_admin_reserve
);
3263 * Reinititalise user and admin reserves if memory is added or removed.
3265 * The default user reserve max is 128MB, and the default max for the
3266 * admin reserve is 8MB. These are usually, but not always, enough to
3267 * enable recovery from a memory hogging process using login/sshd, a shell,
3268 * and tools like top. It may make sense to increase or even disable the
3269 * reserve depending on the existence of swap or variations in the recovery
3270 * tools. So, the admin may have changed them.
3272 * If memory is added and the reserves have been eliminated or increased above
3273 * the default max, then we'll trust the admin.
3275 * If memory is removed and there isn't enough free memory, then we
3276 * need to reset the reserves.
3278 * Otherwise keep the reserve set by the admin.
3280 static int reserve_mem_notifier(struct notifier_block
*nb
,
3281 unsigned long action
, void *data
)
3283 unsigned long tmp
, free_kbytes
;
3287 /* Default max is 128MB. Leave alone if modified by operator. */
3288 tmp
= sysctl_user_reserve_kbytes
;
3289 if (0 < tmp
&& tmp
< (1UL << 17))
3290 init_user_reserve();
3292 /* Default max is 8MB. Leave alone if modified by operator. */
3293 tmp
= sysctl_admin_reserve_kbytes
;
3294 if (0 < tmp
&& tmp
< (1UL << 13))
3295 init_admin_reserve();
3299 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3301 if (sysctl_user_reserve_kbytes
> free_kbytes
) {
3302 init_user_reserve();
3303 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3304 sysctl_user_reserve_kbytes
);
3307 if (sysctl_admin_reserve_kbytes
> free_kbytes
) {
3308 init_admin_reserve();
3309 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3310 sysctl_admin_reserve_kbytes
);
3319 static struct notifier_block reserve_mem_nb
= {
3320 .notifier_call
= reserve_mem_notifier
,
3323 static int __meminit
init_reserve_notifier(void)
3325 if (register_hotmemory_notifier(&reserve_mem_nb
))
3326 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3330 subsys_initcall(init_reserve_notifier
);