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 int sysctl_max_map_count __read_mostly
= DEFAULT_MAX_MAP_COUNT
;
90 unsigned long sysctl_user_reserve_kbytes __read_mostly
= 1UL << 17; /* 128MB */
91 unsigned long sysctl_admin_reserve_kbytes __read_mostly
= 1UL << 13; /* 8MB */
93 * Make sure vm_committed_as in one cacheline and not cacheline shared with
94 * other variables. It can be updated by several CPUs frequently.
96 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp
;
99 * The global memory commitment made in the system can be a metric
100 * that can be used to drive ballooning decisions when Linux is hosted
101 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
102 * balancing memory across competing virtual machines that are hosted.
103 * Several metrics drive this policy engine including the guest reported
106 unsigned long vm_memory_committed(void)
108 return percpu_counter_read_positive(&vm_committed_as
);
110 EXPORT_SYMBOL_GPL(vm_memory_committed
);
113 * Check that a process has enough memory to allocate a new virtual
114 * mapping. 0 means there is enough memory for the allocation to
115 * succeed and -ENOMEM implies there is not.
117 * We currently support three overcommit policies, which are set via the
118 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
120 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
121 * Additional code 2002 Jul 20 by Robert Love.
123 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
125 * Note this is a helper function intended to be used by LSMs which
126 * wish to use this logic.
128 int __vm_enough_memory(struct mm_struct
*mm
, long pages
, int cap_sys_admin
)
130 long free
, allowed
, reserve
;
132 vm_acct_memory(pages
);
135 * Sometimes we want to use more memory than we have
137 if (sysctl_overcommit_memory
== OVERCOMMIT_ALWAYS
)
140 if (sysctl_overcommit_memory
== OVERCOMMIT_GUESS
) {
141 free
= global_page_state(NR_FREE_PAGES
);
142 free
+= global_page_state(NR_FILE_PAGES
);
145 * shmem pages shouldn't be counted as free in this
146 * case, they can't be purged, only swapped out, and
147 * that won't affect the overall amount of available
148 * memory in the system.
150 free
-= global_page_state(NR_SHMEM
);
152 free
+= get_nr_swap_pages();
155 * Any slabs which are created with the
156 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
157 * which are reclaimable, under pressure. The dentry
158 * cache and most inode caches should fall into this
160 free
+= global_page_state(NR_SLAB_RECLAIMABLE
);
163 * Leave reserved pages. The pages are not for anonymous pages.
165 if (free
<= totalreserve_pages
)
168 free
-= totalreserve_pages
;
171 * Reserve some for root
174 free
-= sysctl_admin_reserve_kbytes
>> (PAGE_SHIFT
- 10);
182 allowed
= (totalram_pages
- hugetlb_total_pages())
183 * sysctl_overcommit_ratio
/ 100;
185 * Reserve some for root
188 allowed
-= sysctl_admin_reserve_kbytes
>> (PAGE_SHIFT
- 10);
189 allowed
+= total_swap_pages
;
192 * Don't let a single process grow so big a user can't recover
195 reserve
= sysctl_user_reserve_kbytes
>> (PAGE_SHIFT
- 10);
196 allowed
-= min_t(long, mm
->total_vm
/ 32, reserve
);
199 if (percpu_counter_read_positive(&vm_committed_as
) < allowed
)
202 vm_unacct_memory(pages
);
208 * Requires inode->i_mapping->i_mmap_mutex
210 static void __remove_shared_vm_struct(struct vm_area_struct
*vma
,
211 struct file
*file
, struct address_space
*mapping
)
213 if (vma
->vm_flags
& VM_DENYWRITE
)
214 atomic_inc(&file_inode(file
)->i_writecount
);
215 if (vma
->vm_flags
& VM_SHARED
)
216 mapping
->i_mmap_writable
--;
218 flush_dcache_mmap_lock(mapping
);
219 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
220 list_del_init(&vma
->shared
.nonlinear
);
222 vma_interval_tree_remove(vma
, &mapping
->i_mmap
);
223 flush_dcache_mmap_unlock(mapping
);
227 * Unlink a file-based vm structure from its interval tree, to hide
228 * vma from rmap and vmtruncate before freeing its page tables.
230 void unlink_file_vma(struct vm_area_struct
*vma
)
232 struct file
*file
= vma
->vm_file
;
235 struct address_space
*mapping
= file
->f_mapping
;
236 mutex_lock(&mapping
->i_mmap_mutex
);
237 __remove_shared_vm_struct(vma
, file
, mapping
);
238 mutex_unlock(&mapping
->i_mmap_mutex
);
243 * Close a vm structure and free it, returning the next.
245 static struct vm_area_struct
*remove_vma(struct vm_area_struct
*vma
)
247 struct vm_area_struct
*next
= vma
->vm_next
;
250 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
251 vma
->vm_ops
->close(vma
);
254 mpol_put(vma_policy(vma
));
255 kmem_cache_free(vm_area_cachep
, vma
);
259 static unsigned long do_brk(unsigned long addr
, unsigned long len
);
261 SYSCALL_DEFINE1(brk
, unsigned long, brk
)
263 unsigned long rlim
, retval
;
264 unsigned long newbrk
, oldbrk
;
265 struct mm_struct
*mm
= current
->mm
;
266 struct vm_area_struct
*next
;
267 unsigned long min_brk
;
270 down_write(&mm
->mmap_sem
);
272 #ifdef CONFIG_COMPAT_BRK
274 * CONFIG_COMPAT_BRK can still be overridden by setting
275 * randomize_va_space to 2, which will still cause mm->start_brk
276 * to be arbitrarily shifted
278 if (current
->brk_randomized
)
279 min_brk
= mm
->start_brk
;
281 min_brk
= mm
->end_data
;
283 min_brk
= mm
->start_brk
;
289 * Check against rlimit here. If this check is done later after the test
290 * of oldbrk with newbrk then it can escape the test and let the data
291 * segment grow beyond its set limit the in case where the limit is
292 * not page aligned -Ram Gupta
294 rlim
= rlimit(RLIMIT_DATA
);
295 if (rlim
< RLIM_INFINITY
&& (brk
- mm
->start_brk
) +
296 (mm
->end_data
- mm
->start_data
) > rlim
)
299 newbrk
= PAGE_ALIGN(brk
);
300 oldbrk
= PAGE_ALIGN(mm
->brk
);
301 if (oldbrk
== newbrk
)
304 /* Always allow shrinking brk. */
305 if (brk
<= mm
->brk
) {
306 if (!do_munmap(mm
, newbrk
, oldbrk
-newbrk
))
311 /* Check against existing mmap mappings. */
312 next
= find_vma(mm
, oldbrk
);
313 if (next
&& newbrk
+ PAGE_SIZE
> vm_start_gap(next
))
316 /* Ok, looks good - let it rip. */
317 if (do_brk(oldbrk
, newbrk
-oldbrk
) != oldbrk
)
322 populate
= newbrk
> oldbrk
&& (mm
->def_flags
& VM_LOCKED
) != 0;
323 up_write(&mm
->mmap_sem
);
325 mm_populate(oldbrk
, newbrk
- oldbrk
);
330 up_write(&mm
->mmap_sem
);
334 static long vma_compute_subtree_gap(struct vm_area_struct
*vma
)
336 unsigned long max
, prev_end
, subtree_gap
;
339 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
340 * allow two stack_guard_gaps between them here, and when choosing
341 * an unmapped area; whereas when expanding we only require one.
342 * That's a little inconsistent, but keeps the code here simpler.
344 max
= vm_start_gap(vma
);
346 prev_end
= vm_end_gap(vma
->vm_prev
);
352 if (vma
->vm_rb
.rb_left
) {
353 subtree_gap
= rb_entry(vma
->vm_rb
.rb_left
,
354 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
355 if (subtree_gap
> max
)
358 if (vma
->vm_rb
.rb_right
) {
359 subtree_gap
= rb_entry(vma
->vm_rb
.rb_right
,
360 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
361 if (subtree_gap
> max
)
367 #ifdef CONFIG_DEBUG_VM_RB
368 static int browse_rb(struct rb_root
*root
)
370 int i
= 0, j
, bug
= 0;
371 struct rb_node
*nd
, *pn
= NULL
;
372 unsigned long prev
= 0, pend
= 0;
374 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
375 struct vm_area_struct
*vma
;
376 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
377 if (vma
->vm_start
< prev
) {
378 printk("vm_start %lx prev %lx\n", vma
->vm_start
, prev
);
381 if (vma
->vm_start
< pend
) {
382 printk("vm_start %lx pend %lx\n", vma
->vm_start
, pend
);
385 if (vma
->vm_start
> vma
->vm_end
) {
386 printk("vm_end %lx < vm_start %lx\n",
387 vma
->vm_end
, vma
->vm_start
);
390 if (vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
)) {
391 printk("free gap %lx, correct %lx\n",
393 vma_compute_subtree_gap(vma
));
398 prev
= vma
->vm_start
;
402 for (nd
= pn
; nd
; nd
= rb_prev(nd
))
405 printk("backwards %d, forwards %d\n", j
, i
);
411 static void validate_mm_rb(struct rb_root
*root
, struct vm_area_struct
*ignore
)
415 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
416 struct vm_area_struct
*vma
;
417 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
418 BUG_ON(vma
!= ignore
&&
419 vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
));
423 void validate_mm(struct mm_struct
*mm
)
427 unsigned long highest_address
= 0;
428 struct vm_area_struct
*vma
= mm
->mmap
;
430 struct anon_vma_chain
*avc
;
431 vma_lock_anon_vma(vma
);
432 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
433 anon_vma_interval_tree_verify(avc
);
434 vma_unlock_anon_vma(vma
);
435 highest_address
= vm_end_gap(vma
);
439 if (i
!= mm
->map_count
) {
440 printk("map_count %d vm_next %d\n", mm
->map_count
, i
);
443 if (highest_address
!= mm
->highest_vm_end
) {
444 printk("mm->highest_vm_end %lx, found %lx\n",
445 mm
->highest_vm_end
, highest_address
);
448 i
= browse_rb(&mm
->mm_rb
);
449 if (i
!= mm
->map_count
) {
450 printk("map_count %d rb %d\n", mm
->map_count
, i
);
456 #define validate_mm_rb(root, ignore) do { } while (0)
457 #define validate_mm(mm) do { } while (0)
460 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks
, struct vm_area_struct
, vm_rb
,
461 unsigned long, rb_subtree_gap
, vma_compute_subtree_gap
)
464 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
465 * vma->vm_prev->vm_end values changed, without modifying the vma's position
468 static void vma_gap_update(struct vm_area_struct
*vma
)
471 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
472 * function that does exacltly what we want.
474 vma_gap_callbacks_propagate(&vma
->vm_rb
, NULL
);
477 static inline void vma_rb_insert(struct vm_area_struct
*vma
,
478 struct rb_root
*root
)
480 /* All rb_subtree_gap values must be consistent prior to insertion */
481 validate_mm_rb(root
, NULL
);
483 rb_insert_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
486 static void vma_rb_erase(struct vm_area_struct
*vma
, struct rb_root
*root
)
489 * All rb_subtree_gap values must be consistent prior to erase,
490 * with the possible exception of the vma being erased.
492 validate_mm_rb(root
, vma
);
495 * Note rb_erase_augmented is a fairly large inline function,
496 * so make sure we instantiate it only once with our desired
497 * augmented rbtree callbacks.
499 rb_erase_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
503 * vma has some anon_vma assigned, and is already inserted on that
504 * anon_vma's interval trees.
506 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
507 * vma must be removed from the anon_vma's interval trees using
508 * anon_vma_interval_tree_pre_update_vma().
510 * After the update, the vma will be reinserted using
511 * anon_vma_interval_tree_post_update_vma().
513 * The entire update must be protected by exclusive mmap_sem and by
514 * the root anon_vma's mutex.
517 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct
*vma
)
519 struct anon_vma_chain
*avc
;
521 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
522 anon_vma_interval_tree_remove(avc
, &avc
->anon_vma
->rb_root
);
526 anon_vma_interval_tree_post_update_vma(struct vm_area_struct
*vma
)
528 struct anon_vma_chain
*avc
;
530 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
531 anon_vma_interval_tree_insert(avc
, &avc
->anon_vma
->rb_root
);
534 static int find_vma_links(struct mm_struct
*mm
, unsigned long addr
,
535 unsigned long end
, struct vm_area_struct
**pprev
,
536 struct rb_node
***rb_link
, struct rb_node
**rb_parent
)
538 struct rb_node
**__rb_link
, *__rb_parent
, *rb_prev
;
540 __rb_link
= &mm
->mm_rb
.rb_node
;
541 rb_prev
= __rb_parent
= NULL
;
544 struct vm_area_struct
*vma_tmp
;
546 __rb_parent
= *__rb_link
;
547 vma_tmp
= rb_entry(__rb_parent
, struct vm_area_struct
, vm_rb
);
549 if (vma_tmp
->vm_end
> addr
) {
550 /* Fail if an existing vma overlaps the area */
551 if (vma_tmp
->vm_start
< end
)
553 __rb_link
= &__rb_parent
->rb_left
;
555 rb_prev
= __rb_parent
;
556 __rb_link
= &__rb_parent
->rb_right
;
562 *pprev
= rb_entry(rb_prev
, struct vm_area_struct
, vm_rb
);
563 *rb_link
= __rb_link
;
564 *rb_parent
= __rb_parent
;
568 static unsigned long count_vma_pages_range(struct mm_struct
*mm
,
569 unsigned long addr
, unsigned long end
)
571 unsigned long nr_pages
= 0;
572 struct vm_area_struct
*vma
;
574 /* Find first overlaping mapping */
575 vma
= find_vma_intersection(mm
, addr
, end
);
579 nr_pages
= (min(end
, vma
->vm_end
) -
580 max(addr
, vma
->vm_start
)) >> PAGE_SHIFT
;
582 /* Iterate over the rest of the overlaps */
583 for (vma
= vma
->vm_next
; vma
; vma
= vma
->vm_next
) {
584 unsigned long overlap_len
;
586 if (vma
->vm_start
> end
)
589 overlap_len
= min(end
, vma
->vm_end
) - vma
->vm_start
;
590 nr_pages
+= overlap_len
>> PAGE_SHIFT
;
596 void __vma_link_rb(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
597 struct rb_node
**rb_link
, struct rb_node
*rb_parent
)
599 /* Update tracking information for the gap following the new vma. */
601 vma_gap_update(vma
->vm_next
);
603 mm
->highest_vm_end
= vm_end_gap(vma
);
606 * vma->vm_prev wasn't known when we followed the rbtree to find the
607 * correct insertion point for that vma. As a result, we could not
608 * update the vma vm_rb parents rb_subtree_gap values on the way down.
609 * So, we first insert the vma with a zero rb_subtree_gap value
610 * (to be consistent with what we did on the way down), and then
611 * immediately update the gap to the correct value. Finally we
612 * rebalance the rbtree after all augmented values have been set.
614 rb_link_node(&vma
->vm_rb
, rb_parent
, rb_link
);
615 vma
->rb_subtree_gap
= 0;
617 vma_rb_insert(vma
, &mm
->mm_rb
);
620 static void __vma_link_file(struct vm_area_struct
*vma
)
626 struct address_space
*mapping
= file
->f_mapping
;
628 if (vma
->vm_flags
& VM_DENYWRITE
)
629 atomic_dec(&file_inode(file
)->i_writecount
);
630 if (vma
->vm_flags
& VM_SHARED
)
631 mapping
->i_mmap_writable
++;
633 flush_dcache_mmap_lock(mapping
);
634 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
635 vma_nonlinear_insert(vma
, &mapping
->i_mmap_nonlinear
);
637 vma_interval_tree_insert(vma
, &mapping
->i_mmap
);
638 flush_dcache_mmap_unlock(mapping
);
643 __vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
644 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
645 struct rb_node
*rb_parent
)
647 __vma_link_list(mm
, vma
, prev
, rb_parent
);
648 __vma_link_rb(mm
, vma
, rb_link
, rb_parent
);
651 static void vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
652 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
653 struct rb_node
*rb_parent
)
655 struct address_space
*mapping
= NULL
;
658 mapping
= vma
->vm_file
->f_mapping
;
661 mutex_lock(&mapping
->i_mmap_mutex
);
663 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
664 __vma_link_file(vma
);
667 mutex_unlock(&mapping
->i_mmap_mutex
);
674 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
675 * mm's list and rbtree. It has already been inserted into the interval tree.
677 static void __insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
679 struct vm_area_struct
*prev
;
680 struct rb_node
**rb_link
, *rb_parent
;
682 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
683 &prev
, &rb_link
, &rb_parent
))
685 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
690 __vma_unlink(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
691 struct vm_area_struct
*prev
)
693 struct vm_area_struct
*next
;
695 vma_rb_erase(vma
, &mm
->mm_rb
);
696 prev
->vm_next
= next
= vma
->vm_next
;
698 next
->vm_prev
= prev
;
699 if (mm
->mmap_cache
== vma
)
700 mm
->mmap_cache
= prev
;
704 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
705 * is already present in an i_mmap tree without adjusting the tree.
706 * The following helper function should be used when such adjustments
707 * are necessary. The "insert" vma (if any) is to be inserted
708 * before we drop the necessary locks.
710 int vma_adjust(struct vm_area_struct
*vma
, unsigned long start
,
711 unsigned long end
, pgoff_t pgoff
, struct vm_area_struct
*insert
)
713 struct mm_struct
*mm
= vma
->vm_mm
;
714 struct vm_area_struct
*next
= vma
->vm_next
;
715 struct vm_area_struct
*importer
= NULL
;
716 struct address_space
*mapping
= NULL
;
717 struct rb_root
*root
= NULL
;
718 struct anon_vma
*anon_vma
= NULL
;
719 struct file
*file
= vma
->vm_file
;
720 bool start_changed
= false, end_changed
= false;
721 long adjust_next
= 0;
724 if (next
&& !insert
) {
725 struct vm_area_struct
*exporter
= NULL
;
727 if (end
>= next
->vm_end
) {
729 * vma expands, overlapping all the next, and
730 * perhaps the one after too (mprotect case 6).
732 again
: remove_next
= 1 + (end
> next
->vm_end
);
736 } else if (end
> next
->vm_start
) {
738 * vma expands, overlapping part of the next:
739 * mprotect case 5 shifting the boundary up.
741 adjust_next
= (end
- next
->vm_start
) >> PAGE_SHIFT
;
744 } else if (end
< vma
->vm_end
) {
746 * vma shrinks, and !insert tells it's not
747 * split_vma inserting another: so it must be
748 * mprotect case 4 shifting the boundary down.
750 adjust_next
= - ((vma
->vm_end
- end
) >> PAGE_SHIFT
);
756 * Easily overlooked: when mprotect shifts the boundary,
757 * make sure the expanding vma has anon_vma set if the
758 * shrinking vma had, to cover any anon pages imported.
760 if (exporter
&& exporter
->anon_vma
&& !importer
->anon_vma
) {
761 if (anon_vma_clone(importer
, exporter
))
763 importer
->anon_vma
= exporter
->anon_vma
;
768 mapping
= file
->f_mapping
;
769 if (!(vma
->vm_flags
& VM_NONLINEAR
)) {
770 root
= &mapping
->i_mmap
;
771 uprobe_munmap(vma
, vma
->vm_start
, vma
->vm_end
);
774 uprobe_munmap(next
, next
->vm_start
,
778 mutex_lock(&mapping
->i_mmap_mutex
);
781 * Put into interval tree now, so instantiated pages
782 * are visible to arm/parisc __flush_dcache_page
783 * throughout; but we cannot insert into address
784 * space until vma start or end is updated.
786 __vma_link_file(insert
);
790 vma_adjust_trans_huge(vma
, start
, end
, adjust_next
);
792 anon_vma
= vma
->anon_vma
;
793 if (!anon_vma
&& adjust_next
)
794 anon_vma
= next
->anon_vma
;
796 VM_BUG_ON(adjust_next
&& next
->anon_vma
&&
797 anon_vma
!= next
->anon_vma
);
798 anon_vma_lock_write(anon_vma
);
799 anon_vma_interval_tree_pre_update_vma(vma
);
801 anon_vma_interval_tree_pre_update_vma(next
);
805 flush_dcache_mmap_lock(mapping
);
806 vma_interval_tree_remove(vma
, root
);
808 vma_interval_tree_remove(next
, root
);
811 if (start
!= vma
->vm_start
) {
812 vma
->vm_start
= start
;
813 start_changed
= true;
815 if (end
!= vma
->vm_end
) {
819 vma
->vm_pgoff
= pgoff
;
821 next
->vm_start
+= adjust_next
<< PAGE_SHIFT
;
822 next
->vm_pgoff
+= adjust_next
;
827 vma_interval_tree_insert(next
, root
);
828 vma_interval_tree_insert(vma
, root
);
829 flush_dcache_mmap_unlock(mapping
);
834 * vma_merge has merged next into vma, and needs
835 * us to remove next before dropping the locks.
837 __vma_unlink(mm
, next
, vma
);
839 __remove_shared_vm_struct(next
, file
, mapping
);
842 * split_vma has split insert from vma, and needs
843 * us to insert it before dropping the locks
844 * (it may either follow vma or precede it).
846 __insert_vm_struct(mm
, insert
);
852 mm
->highest_vm_end
= vm_end_gap(vma
);
853 else if (!adjust_next
)
854 vma_gap_update(next
);
859 anon_vma_interval_tree_post_update_vma(vma
);
861 anon_vma_interval_tree_post_update_vma(next
);
862 anon_vma_unlock_write(anon_vma
);
865 mutex_unlock(&mapping
->i_mmap_mutex
);
876 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
880 anon_vma_merge(vma
, next
);
882 mpol_put(vma_policy(next
));
883 kmem_cache_free(vm_area_cachep
, next
);
885 * In mprotect's case 6 (see comments on vma_merge),
886 * we must remove another next too. It would clutter
887 * up the code too much to do both in one go.
890 if (remove_next
== 2)
893 vma_gap_update(next
);
895 WARN_ON(mm
->highest_vm_end
!= vm_end_gap(vma
));
906 * If the vma has a ->close operation then the driver probably needs to release
907 * per-vma resources, so we don't attempt to merge those.
909 static inline int is_mergeable_vma(struct vm_area_struct
*vma
,
910 struct file
*file
, unsigned long vm_flags
)
912 if (vma
->vm_flags
^ vm_flags
)
914 if (vma
->vm_file
!= file
)
916 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
921 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
922 struct anon_vma
*anon_vma2
,
923 struct vm_area_struct
*vma
)
926 * The list_is_singular() test is to avoid merging VMA cloned from
927 * parents. This can improve scalability caused by anon_vma lock.
929 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
930 list_is_singular(&vma
->anon_vma_chain
)))
932 return anon_vma1
== anon_vma2
;
936 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
937 * in front of (at a lower virtual address and file offset than) the vma.
939 * We cannot merge two vmas if they have differently assigned (non-NULL)
940 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
942 * We don't check here for the merged mmap wrapping around the end of pagecache
943 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
944 * wrap, nor mmaps which cover the final page at index -1UL.
947 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
948 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
950 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
951 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
952 if (vma
->vm_pgoff
== vm_pgoff
)
959 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
960 * beyond (at a higher virtual address and file offset than) the vma.
962 * We cannot merge two vmas if they have differently assigned (non-NULL)
963 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
966 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
967 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
969 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
970 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
972 vm_pglen
= (vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
;
973 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
980 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
981 * whether that can be merged with its predecessor or its successor.
982 * Or both (it neatly fills a hole).
984 * In most cases - when called for mmap, brk or mremap - [addr,end) is
985 * certain not to be mapped by the time vma_merge is called; but when
986 * called for mprotect, it is certain to be already mapped (either at
987 * an offset within prev, or at the start of next), and the flags of
988 * this area are about to be changed to vm_flags - and the no-change
989 * case has already been eliminated.
991 * The following mprotect cases have to be considered, where AAAA is
992 * the area passed down from mprotect_fixup, never extending beyond one
993 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
995 * AAAA AAAA AAAA AAAA
996 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
997 * cannot merge might become might become might become
998 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
999 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1000 * mremap move: PPPPNNNNNNNN 8
1002 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1003 * might become case 1 below case 2 below case 3 below
1005 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1006 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1008 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
1009 struct vm_area_struct
*prev
, unsigned long addr
,
1010 unsigned long end
, unsigned long vm_flags
,
1011 struct anon_vma
*anon_vma
, struct file
*file
,
1012 pgoff_t pgoff
, struct mempolicy
*policy
)
1014 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
1015 struct vm_area_struct
*area
, *next
;
1019 * We later require that vma->vm_flags == vm_flags,
1020 * so this tests vma->vm_flags & VM_SPECIAL, too.
1022 if (vm_flags
& VM_SPECIAL
)
1026 next
= prev
->vm_next
;
1030 if (next
&& next
->vm_end
== end
) /* cases 6, 7, 8 */
1031 next
= next
->vm_next
;
1034 * Can it merge with the predecessor?
1036 if (prev
&& prev
->vm_end
== addr
&&
1037 mpol_equal(vma_policy(prev
), policy
) &&
1038 can_vma_merge_after(prev
, vm_flags
,
1039 anon_vma
, file
, pgoff
)) {
1041 * OK, it can. Can we now merge in the successor as well?
1043 if (next
&& end
== next
->vm_start
&&
1044 mpol_equal(policy
, vma_policy(next
)) &&
1045 can_vma_merge_before(next
, vm_flags
,
1046 anon_vma
, file
, pgoff
+pglen
) &&
1047 is_mergeable_anon_vma(prev
->anon_vma
,
1048 next
->anon_vma
, NULL
)) {
1050 err
= vma_adjust(prev
, prev
->vm_start
,
1051 next
->vm_end
, prev
->vm_pgoff
, NULL
);
1052 } else /* cases 2, 5, 7 */
1053 err
= vma_adjust(prev
, prev
->vm_start
,
1054 end
, prev
->vm_pgoff
, NULL
);
1057 khugepaged_enter_vma_merge(prev
);
1062 * Can this new request be merged in front of next?
1064 if (next
&& end
== next
->vm_start
&&
1065 mpol_equal(policy
, vma_policy(next
)) &&
1066 can_vma_merge_before(next
, vm_flags
,
1067 anon_vma
, file
, pgoff
+pglen
)) {
1068 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
1069 err
= vma_adjust(prev
, prev
->vm_start
,
1070 addr
, prev
->vm_pgoff
, NULL
);
1071 else /* cases 3, 8 */
1072 err
= vma_adjust(area
, addr
, next
->vm_end
,
1073 next
->vm_pgoff
- pglen
, NULL
);
1076 khugepaged_enter_vma_merge(area
);
1084 * Rough compatbility check to quickly see if it's even worth looking
1085 * at sharing an anon_vma.
1087 * They need to have the same vm_file, and the flags can only differ
1088 * in things that mprotect may change.
1090 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1091 * we can merge the two vma's. For example, we refuse to merge a vma if
1092 * there is a vm_ops->close() function, because that indicates that the
1093 * driver is doing some kind of reference counting. But that doesn't
1094 * really matter for the anon_vma sharing case.
1096 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1098 return a
->vm_end
== b
->vm_start
&&
1099 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
1100 a
->vm_file
== b
->vm_file
&&
1101 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_READ
|VM_WRITE
|VM_EXEC
)) &&
1102 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
1106 * Do some basic sanity checking to see if we can re-use the anon_vma
1107 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1108 * the same as 'old', the other will be the new one that is trying
1109 * to share the anon_vma.
1111 * NOTE! This runs with mm_sem held for reading, so it is possible that
1112 * the anon_vma of 'old' is concurrently in the process of being set up
1113 * by another page fault trying to merge _that_. But that's ok: if it
1114 * is being set up, that automatically means that it will be a singleton
1115 * acceptable for merging, so we can do all of this optimistically. But
1116 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1118 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1119 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1120 * is to return an anon_vma that is "complex" due to having gone through
1123 * We also make sure that the two vma's are compatible (adjacent,
1124 * and with the same memory policies). That's all stable, even with just
1125 * a read lock on the mm_sem.
1127 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1129 if (anon_vma_compatible(a
, b
)) {
1130 struct anon_vma
*anon_vma
= ACCESS_ONCE(old
->anon_vma
);
1132 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
1139 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1140 * neighbouring vmas for a suitable anon_vma, before it goes off
1141 * to allocate a new anon_vma. It checks because a repetitive
1142 * sequence of mprotects and faults may otherwise lead to distinct
1143 * anon_vmas being allocated, preventing vma merge in subsequent
1146 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
1148 struct anon_vma
*anon_vma
;
1149 struct vm_area_struct
*near
;
1151 near
= vma
->vm_next
;
1155 anon_vma
= reusable_anon_vma(near
, vma
, near
);
1159 near
= vma
->vm_prev
;
1163 anon_vma
= reusable_anon_vma(near
, near
, vma
);
1168 * There's no absolute need to look only at touching neighbours:
1169 * we could search further afield for "compatible" anon_vmas.
1170 * But it would probably just be a waste of time searching,
1171 * or lead to too many vmas hanging off the same anon_vma.
1172 * We're trying to allow mprotect remerging later on,
1173 * not trying to minimize memory used for anon_vmas.
1178 #ifdef CONFIG_PROC_FS
1179 void vm_stat_account(struct mm_struct
*mm
, unsigned long flags
,
1180 struct file
*file
, long pages
)
1182 const unsigned long stack_flags
1183 = VM_STACK_FLAGS
& (VM_GROWSUP
|VM_GROWSDOWN
);
1185 mm
->total_vm
+= pages
;
1188 mm
->shared_vm
+= pages
;
1189 if ((flags
& (VM_EXEC
|VM_WRITE
)) == VM_EXEC
)
1190 mm
->exec_vm
+= pages
;
1191 } else if (flags
& stack_flags
)
1192 mm
->stack_vm
+= pages
;
1194 #endif /* CONFIG_PROC_FS */
1197 * If a hint addr is less than mmap_min_addr change hint to be as
1198 * low as possible but still greater than mmap_min_addr
1200 static inline unsigned long round_hint_to_min(unsigned long hint
)
1203 if (((void *)hint
!= NULL
) &&
1204 (hint
< mmap_min_addr
))
1205 return PAGE_ALIGN(mmap_min_addr
);
1210 * The caller must hold down_write(¤t->mm->mmap_sem).
1213 unsigned long do_mmap_pgoff(struct file
*file
, unsigned long addr
,
1214 unsigned long len
, unsigned long prot
,
1215 unsigned long flags
, unsigned long pgoff
,
1216 unsigned long *populate
)
1218 struct mm_struct
* mm
= current
->mm
;
1219 struct inode
*inode
;
1220 vm_flags_t vm_flags
;
1225 * Does the application expect PROT_READ to imply PROT_EXEC?
1227 * (the exception is when the underlying filesystem is noexec
1228 * mounted, in which case we dont add PROT_EXEC.)
1230 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1231 if (!(file
&& (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
)))
1237 if (!(flags
& MAP_FIXED
))
1238 addr
= round_hint_to_min(addr
);
1240 /* Careful about overflows.. */
1241 len
= PAGE_ALIGN(len
);
1245 /* offset overflow? */
1246 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1249 /* Too many mappings? */
1250 if (mm
->map_count
> sysctl_max_map_count
)
1253 /* Obtain the address to map to. we verify (or select) it and ensure
1254 * that it represents a valid section of the address space.
1256 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1257 if (addr
& ~PAGE_MASK
)
1260 /* Do simple checking here so the lower-level routines won't have
1261 * to. we assume access permissions have been handled by the open
1262 * of the memory object, so we don't do any here.
1264 vm_flags
= calc_vm_prot_bits(prot
) | calc_vm_flag_bits(flags
) |
1265 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1267 if (flags
& MAP_LOCKED
)
1268 if (!can_do_mlock())
1271 /* mlock MCL_FUTURE? */
1272 if (vm_flags
& VM_LOCKED
) {
1273 unsigned long locked
, lock_limit
;
1274 locked
= len
>> PAGE_SHIFT
;
1275 locked
+= mm
->locked_vm
;
1276 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1277 lock_limit
>>= PAGE_SHIFT
;
1278 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1282 inode
= file
? file_inode(file
) : NULL
;
1285 switch (flags
& MAP_TYPE
) {
1287 if ((prot
&PROT_WRITE
) && !(file
->f_mode
&FMODE_WRITE
))
1291 * Make sure we don't allow writing to an append-only
1294 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1298 * Make sure there are no mandatory locks on the file.
1300 if (locks_verify_locked(inode
))
1303 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1304 if (!(file
->f_mode
& FMODE_WRITE
))
1305 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1309 if (!(file
->f_mode
& FMODE_READ
))
1311 if (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
) {
1312 if (vm_flags
& VM_EXEC
)
1314 vm_flags
&= ~VM_MAYEXEC
;
1317 if (!file
->f_op
|| !file
->f_op
->mmap
)
1325 switch (flags
& MAP_TYPE
) {
1331 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1335 * Set pgoff according to addr for anon_vma.
1337 pgoff
= addr
>> PAGE_SHIFT
;
1345 * Set 'VM_NORESERVE' if we should not account for the
1346 * memory use of this mapping.
1348 if (flags
& MAP_NORESERVE
) {
1349 /* We honor MAP_NORESERVE if allowed to overcommit */
1350 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1351 vm_flags
|= VM_NORESERVE
;
1353 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1354 if (file
&& is_file_hugepages(file
))
1355 vm_flags
|= VM_NORESERVE
;
1358 addr
= mmap_region(file
, addr
, len
, vm_flags
, pgoff
);
1359 if (!IS_ERR_VALUE(addr
) &&
1360 ((vm_flags
& VM_LOCKED
) ||
1361 (flags
& (MAP_POPULATE
| MAP_NONBLOCK
)) == MAP_POPULATE
))
1366 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1367 unsigned long, prot
, unsigned long, flags
,
1368 unsigned long, fd
, unsigned long, pgoff
)
1370 struct file
*file
= NULL
;
1371 unsigned long retval
= -EBADF
;
1373 if (!(flags
& MAP_ANONYMOUS
)) {
1374 audit_mmap_fd(fd
, flags
);
1375 if (unlikely(flags
& MAP_HUGETLB
))
1380 if (is_file_hugepages(file
))
1381 len
= ALIGN(len
, huge_page_size(hstate_file(file
)));
1382 } else if (flags
& MAP_HUGETLB
) {
1383 struct user_struct
*user
= NULL
;
1384 struct hstate
*hs
= hstate_sizelog((flags
>> MAP_HUGE_SHIFT
) &
1390 len
= ALIGN(len
, huge_page_size(hs
));
1392 * VM_NORESERVE is used because the reservations will be
1393 * taken when vm_ops->mmap() is called
1394 * A dummy user value is used because we are not locking
1395 * memory so no accounting is necessary
1397 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, len
,
1399 &user
, HUGETLB_ANONHUGE_INODE
,
1400 (flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1402 return PTR_ERR(file
);
1405 flags
&= ~(MAP_EXECUTABLE
| MAP_DENYWRITE
);
1407 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
;
1491 int correct_wcount
= 0;
1493 struct rb_node
**rb_link
, *rb_parent
;
1494 unsigned long charged
= 0;
1495 struct inode
*inode
= file
? file_inode(file
) : NULL
;
1497 /* Check against address space limit. */
1498 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
)) {
1499 unsigned long nr_pages
;
1502 * MAP_FIXED may remove pages of mappings that intersects with
1503 * requested mapping. Account for the pages it would unmap.
1505 if (!(vm_flags
& MAP_FIXED
))
1508 nr_pages
= count_vma_pages_range(mm
, addr
, addr
+ len
);
1510 if (!may_expand_vm(mm
, (len
>> PAGE_SHIFT
) - nr_pages
))
1514 /* Clear old maps */
1517 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
1518 if (do_munmap(mm
, addr
, len
))
1524 * Private writable mapping: check memory availability
1526 if (accountable_mapping(file
, vm_flags
)) {
1527 charged
= len
>> PAGE_SHIFT
;
1528 if (security_vm_enough_memory_mm(mm
, charged
))
1530 vm_flags
|= VM_ACCOUNT
;
1534 * Can we just expand an old mapping?
1536 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
, NULL
, file
, pgoff
, NULL
);
1541 * Determine the object being mapped and call the appropriate
1542 * specific mapper. the address has already been validated, but
1543 * not unmapped, but the maps are removed from the list.
1545 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
1552 vma
->vm_start
= addr
;
1553 vma
->vm_end
= addr
+ len
;
1554 vma
->vm_flags
= vm_flags
;
1555 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1556 vma
->vm_pgoff
= pgoff
;
1557 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
1559 error
= -EINVAL
; /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1562 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1564 if (vm_flags
& VM_DENYWRITE
) {
1565 error
= deny_write_access(file
);
1570 vma
->vm_file
= get_file(file
);
1571 error
= file
->f_op
->mmap(file
, vma
);
1573 goto unmap_and_free_vma
;
1575 /* Can addr have changed??
1577 * Answer: Yes, several device drivers can do it in their
1578 * f_op->mmap method. -DaveM
1579 * Bug: If addr is changed, prev, rb_link, rb_parent should
1580 * be updated for vma_link()
1582 WARN_ON_ONCE(addr
!= vma
->vm_start
);
1584 addr
= vma
->vm_start
;
1585 pgoff
= vma
->vm_pgoff
;
1586 vm_flags
= vma
->vm_flags
;
1587 } else if (vm_flags
& VM_SHARED
) {
1588 if (unlikely(vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
)))
1590 error
= shmem_zero_setup(vma
);
1595 if (vma_wants_writenotify(vma
)) {
1596 pgprot_t pprot
= vma
->vm_page_prot
;
1598 /* Can vma->vm_page_prot have changed??
1600 * Answer: Yes, drivers may have changed it in their
1601 * f_op->mmap method.
1603 * Ensures that vmas marked as uncached stay that way.
1605 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
& ~VM_SHARED
);
1606 if (pgprot_val(pprot
) == pgprot_val(pgprot_noncached(pprot
)))
1607 vma
->vm_page_prot
= pgprot_noncached(vma
->vm_page_prot
);
1610 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1611 file
= vma
->vm_file
;
1613 /* Once vma denies write, undo our temporary denial count */
1615 atomic_inc(&inode
->i_writecount
);
1617 perf_event_mmap(vma
);
1619 vm_stat_account(mm
, vm_flags
, file
, len
>> PAGE_SHIFT
);
1620 if (vm_flags
& VM_LOCKED
) {
1621 if (!((vm_flags
& VM_SPECIAL
) || is_vm_hugetlb_page(vma
) ||
1622 vma
== get_gate_vma(current
->mm
)))
1623 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1625 vma
->vm_flags
&= ~VM_LOCKED
;
1635 atomic_inc(&inode
->i_writecount
);
1636 vma
->vm_file
= NULL
;
1639 /* Undo any partial mapping done by a device driver. */
1640 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1643 kmem_cache_free(vm_area_cachep
, vma
);
1646 vm_unacct_memory(charged
);
1650 unsigned long unmapped_area(struct vm_unmapped_area_info
*info
)
1653 * We implement the search by looking for an rbtree node that
1654 * immediately follows a suitable gap. That is,
1655 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1656 * - gap_end = vma->vm_start >= info->low_limit + length;
1657 * - gap_end - gap_start >= length
1660 struct mm_struct
*mm
= current
->mm
;
1661 struct vm_area_struct
*vma
;
1662 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1664 /* Adjust search length to account for worst case alignment overhead */
1665 length
= info
->length
+ info
->align_mask
;
1666 if (length
< info
->length
)
1669 /* Adjust search limits by the desired length */
1670 if (info
->high_limit
< length
)
1672 high_limit
= info
->high_limit
- length
;
1674 if (info
->low_limit
> high_limit
)
1676 low_limit
= info
->low_limit
+ length
;
1678 /* Check if rbtree root looks promising */
1679 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1681 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1682 if (vma
->rb_subtree_gap
< length
)
1686 /* Visit left subtree if it looks promising */
1687 gap_end
= vm_start_gap(vma
);
1688 if (gap_end
>= low_limit
&& vma
->vm_rb
.rb_left
) {
1689 struct vm_area_struct
*left
=
1690 rb_entry(vma
->vm_rb
.rb_left
,
1691 struct vm_area_struct
, vm_rb
);
1692 if (left
->rb_subtree_gap
>= length
) {
1698 gap_start
= vma
->vm_prev
? vm_end_gap(vma
->vm_prev
) : 0;
1700 /* Check if current node has a suitable gap */
1701 if (gap_start
> high_limit
)
1703 if (gap_end
>= low_limit
&&
1704 gap_end
> gap_start
&& gap_end
- gap_start
>= length
)
1707 /* Visit right subtree if it looks promising */
1708 if (vma
->vm_rb
.rb_right
) {
1709 struct vm_area_struct
*right
=
1710 rb_entry(vma
->vm_rb
.rb_right
,
1711 struct vm_area_struct
, vm_rb
);
1712 if (right
->rb_subtree_gap
>= length
) {
1718 /* Go back up the rbtree to find next candidate node */
1720 struct rb_node
*prev
= &vma
->vm_rb
;
1721 if (!rb_parent(prev
))
1723 vma
= rb_entry(rb_parent(prev
),
1724 struct vm_area_struct
, vm_rb
);
1725 if (prev
== vma
->vm_rb
.rb_left
) {
1726 gap_start
= vm_end_gap(vma
->vm_prev
);
1727 gap_end
= vm_start_gap(vma
);
1734 /* Check highest gap, which does not precede any rbtree node */
1735 gap_start
= mm
->highest_vm_end
;
1736 gap_end
= ULONG_MAX
; /* Only for VM_BUG_ON below */
1737 if (gap_start
> high_limit
)
1741 /* We found a suitable gap. Clip it with the original low_limit. */
1742 if (gap_start
< info
->low_limit
)
1743 gap_start
= info
->low_limit
;
1745 /* Adjust gap address to the desired alignment */
1746 gap_start
+= (info
->align_offset
- gap_start
) & info
->align_mask
;
1748 VM_BUG_ON(gap_start
+ info
->length
> info
->high_limit
);
1749 VM_BUG_ON(gap_start
+ info
->length
> gap_end
);
1753 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info
*info
)
1755 struct mm_struct
*mm
= current
->mm
;
1756 struct vm_area_struct
*vma
;
1757 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1759 /* Adjust search length to account for worst case alignment overhead */
1760 length
= info
->length
+ info
->align_mask
;
1761 if (length
< info
->length
)
1765 * Adjust search limits by the desired length.
1766 * See implementation comment at top of unmapped_area().
1768 gap_end
= info
->high_limit
;
1769 if (gap_end
< length
)
1771 high_limit
= gap_end
- length
;
1773 if (info
->low_limit
> high_limit
)
1775 low_limit
= info
->low_limit
+ length
;
1777 /* Check highest gap, which does not precede any rbtree node */
1778 gap_start
= mm
->highest_vm_end
;
1779 if (gap_start
<= high_limit
)
1782 /* Check if rbtree root looks promising */
1783 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1785 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1786 if (vma
->rb_subtree_gap
< length
)
1790 /* Visit right subtree if it looks promising */
1791 gap_start
= vma
->vm_prev
? vm_end_gap(vma
->vm_prev
) : 0;
1792 if (gap_start
<= high_limit
&& vma
->vm_rb
.rb_right
) {
1793 struct vm_area_struct
*right
=
1794 rb_entry(vma
->vm_rb
.rb_right
,
1795 struct vm_area_struct
, vm_rb
);
1796 if (right
->rb_subtree_gap
>= length
) {
1803 /* Check if current node has a suitable gap */
1804 gap_end
= vm_start_gap(vma
);
1805 if (gap_end
< low_limit
)
1807 if (gap_start
<= high_limit
&&
1808 gap_end
> gap_start
&& gap_end
- gap_start
>= length
)
1811 /* Visit left subtree if it looks promising */
1812 if (vma
->vm_rb
.rb_left
) {
1813 struct vm_area_struct
*left
=
1814 rb_entry(vma
->vm_rb
.rb_left
,
1815 struct vm_area_struct
, vm_rb
);
1816 if (left
->rb_subtree_gap
>= length
) {
1822 /* Go back up the rbtree to find next candidate node */
1824 struct rb_node
*prev
= &vma
->vm_rb
;
1825 if (!rb_parent(prev
))
1827 vma
= rb_entry(rb_parent(prev
),
1828 struct vm_area_struct
, vm_rb
);
1829 if (prev
== vma
->vm_rb
.rb_right
) {
1830 gap_start
= vma
->vm_prev
?
1831 vm_end_gap(vma
->vm_prev
) : 0;
1838 /* We found a suitable gap. Clip it with the original high_limit. */
1839 if (gap_end
> info
->high_limit
)
1840 gap_end
= info
->high_limit
;
1843 /* Compute highest gap address at the desired alignment */
1844 gap_end
-= info
->length
;
1845 gap_end
-= (gap_end
- info
->align_offset
) & info
->align_mask
;
1847 VM_BUG_ON(gap_end
< info
->low_limit
);
1848 VM_BUG_ON(gap_end
< gap_start
);
1852 /* Get an address range which is currently unmapped.
1853 * For shmat() with addr=0.
1855 * Ugly calling convention alert:
1856 * Return value with the low bits set means error value,
1858 * if (ret & ~PAGE_MASK)
1861 * This function "knows" that -ENOMEM has the bits set.
1863 #ifndef HAVE_ARCH_UNMAPPED_AREA
1865 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
1866 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
1868 struct mm_struct
*mm
= current
->mm
;
1869 struct vm_area_struct
*vma
, *prev
;
1870 struct vm_unmapped_area_info info
;
1872 if (len
> TASK_SIZE
- mmap_min_addr
)
1875 if (flags
& MAP_FIXED
)
1879 addr
= PAGE_ALIGN(addr
);
1880 vma
= find_vma_prev(mm
, addr
, &prev
);
1881 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1882 (!vma
|| addr
+ len
<= vm_start_gap(vma
)) &&
1883 (!prev
|| addr
>= vm_end_gap(prev
)))
1889 info
.low_limit
= TASK_UNMAPPED_BASE
;
1890 info
.high_limit
= TASK_SIZE
;
1891 info
.align_mask
= 0;
1892 return vm_unmapped_area(&info
);
1896 void arch_unmap_area(struct mm_struct
*mm
, unsigned long addr
)
1899 * Is this a new hole at the lowest possible address?
1901 if (addr
>= TASK_UNMAPPED_BASE
&& addr
< mm
->free_area_cache
)
1902 mm
->free_area_cache
= addr
;
1906 * This mmap-allocator allocates new areas top-down from below the
1907 * stack's low limit (the base):
1909 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1911 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
1912 const unsigned long len
, const unsigned long pgoff
,
1913 const unsigned long flags
)
1915 struct vm_area_struct
*vma
, *prev
;
1916 struct mm_struct
*mm
= current
->mm
;
1917 unsigned long addr
= addr0
;
1918 struct vm_unmapped_area_info info
;
1920 /* requested length too big for entire address space */
1921 if (len
> TASK_SIZE
- mmap_min_addr
)
1924 if (flags
& MAP_FIXED
)
1927 /* requesting a specific address */
1929 addr
= PAGE_ALIGN(addr
);
1930 vma
= find_vma_prev(mm
, addr
, &prev
);
1931 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1932 (!vma
|| addr
+ len
<= vm_start_gap(vma
)) &&
1933 (!prev
|| addr
>= vm_end_gap(prev
)))
1937 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
1939 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
1940 info
.high_limit
= mm
->mmap_base
;
1941 info
.align_mask
= 0;
1942 addr
= vm_unmapped_area(&info
);
1945 * A failed mmap() very likely causes application failure,
1946 * so fall back to the bottom-up function here. This scenario
1947 * can happen with large stack limits and large mmap()
1950 if (addr
& ~PAGE_MASK
) {
1951 VM_BUG_ON(addr
!= -ENOMEM
);
1953 info
.low_limit
= TASK_UNMAPPED_BASE
;
1954 info
.high_limit
= TASK_SIZE
;
1955 addr
= vm_unmapped_area(&info
);
1962 void arch_unmap_area_topdown(struct mm_struct
*mm
, unsigned long addr
)
1965 * Is this a new hole at the highest possible address?
1967 if (addr
> mm
->free_area_cache
)
1968 mm
->free_area_cache
= addr
;
1970 /* dont allow allocations above current base */
1971 if (mm
->free_area_cache
> mm
->mmap_base
)
1972 mm
->free_area_cache
= mm
->mmap_base
;
1976 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
1977 unsigned long pgoff
, unsigned long flags
)
1979 unsigned long (*get_area
)(struct file
*, unsigned long,
1980 unsigned long, unsigned long, unsigned long);
1982 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
1986 /* Careful about overflows.. */
1987 if (len
> TASK_SIZE
)
1990 get_area
= current
->mm
->get_unmapped_area
;
1991 if (file
&& file
->f_op
&& file
->f_op
->get_unmapped_area
)
1992 get_area
= file
->f_op
->get_unmapped_area
;
1993 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
1994 if (IS_ERR_VALUE(addr
))
1997 if (addr
> TASK_SIZE
- len
)
1999 if (addr
& ~PAGE_MASK
)
2002 addr
= arch_rebalance_pgtables(addr
, len
);
2003 error
= security_mmap_addr(addr
);
2004 return error
? error
: addr
;
2007 EXPORT_SYMBOL(get_unmapped_area
);
2009 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2010 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
2012 struct vm_area_struct
*vma
= NULL
;
2014 /* Check the cache first. */
2015 /* (Cache hit rate is typically around 35%.) */
2016 vma
= ACCESS_ONCE(mm
->mmap_cache
);
2017 if (!(vma
&& vma
->vm_end
> addr
&& vma
->vm_start
<= addr
)) {
2018 struct rb_node
*rb_node
;
2020 rb_node
= mm
->mm_rb
.rb_node
;
2024 struct vm_area_struct
*vma_tmp
;
2026 vma_tmp
= rb_entry(rb_node
,
2027 struct vm_area_struct
, vm_rb
);
2029 if (vma_tmp
->vm_end
> addr
) {
2031 if (vma_tmp
->vm_start
<= addr
)
2033 rb_node
= rb_node
->rb_left
;
2035 rb_node
= rb_node
->rb_right
;
2038 mm
->mmap_cache
= vma
;
2043 EXPORT_SYMBOL(find_vma
);
2046 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2048 struct vm_area_struct
*
2049 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
2050 struct vm_area_struct
**pprev
)
2052 struct vm_area_struct
*vma
;
2054 vma
= find_vma(mm
, addr
);
2056 *pprev
= vma
->vm_prev
;
2058 struct rb_node
*rb_node
= mm
->mm_rb
.rb_node
;
2061 *pprev
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2062 rb_node
= rb_node
->rb_right
;
2069 * Verify that the stack growth is acceptable and
2070 * update accounting. This is shared with both the
2071 * grow-up and grow-down cases.
2073 static int acct_stack_growth(struct vm_area_struct
*vma
,
2074 unsigned long size
, unsigned long grow
)
2076 struct mm_struct
*mm
= vma
->vm_mm
;
2077 struct rlimit
*rlim
= current
->signal
->rlim
;
2078 unsigned long new_start
;
2080 /* address space limit tests */
2081 if (!may_expand_vm(mm
, grow
))
2084 /* Stack limit test */
2085 if (size
> ACCESS_ONCE(rlim
[RLIMIT_STACK
].rlim_cur
))
2088 /* mlock limit tests */
2089 if (vma
->vm_flags
& VM_LOCKED
) {
2090 unsigned long locked
;
2091 unsigned long limit
;
2092 locked
= mm
->locked_vm
+ grow
;
2093 limit
= ACCESS_ONCE(rlim
[RLIMIT_MEMLOCK
].rlim_cur
);
2094 limit
>>= PAGE_SHIFT
;
2095 if (locked
> limit
&& !capable(CAP_IPC_LOCK
))
2099 /* Check to ensure the stack will not grow into a hugetlb-only region */
2100 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
2102 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
2106 * Overcommit.. This must be the final test, as it will
2107 * update security statistics.
2109 if (security_vm_enough_memory_mm(mm
, grow
))
2112 /* Ok, everything looks good - let it rip */
2113 if (vma
->vm_flags
& VM_LOCKED
)
2114 mm
->locked_vm
+= grow
;
2115 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, grow
);
2119 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2121 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2122 * vma is the last one with address > vma->vm_end. Have to extend vma.
2124 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
)
2126 struct vm_area_struct
*next
;
2127 unsigned long gap_addr
;
2130 if (!(vma
->vm_flags
& VM_GROWSUP
))
2133 /* Guard against exceeding limits of the address space. */
2134 address
&= PAGE_MASK
;
2135 if (address
>= TASK_SIZE
)
2137 address
+= PAGE_SIZE
;
2139 /* Enforce stack_guard_gap */
2140 gap_addr
= address
+ stack_guard_gap
;
2142 /* Guard against overflow */
2143 if (gap_addr
< address
|| gap_addr
> TASK_SIZE
)
2144 gap_addr
= TASK_SIZE
;
2146 next
= vma
->vm_next
;
2147 if (next
&& next
->vm_start
< gap_addr
) {
2148 if (!(next
->vm_flags
& VM_GROWSUP
))
2150 /* Check that both stack segments have the same anon_vma? */
2153 /* We must make sure the anon_vma is allocated. */
2154 if (unlikely(anon_vma_prepare(vma
)))
2158 * vma->vm_start/vm_end cannot change under us because the caller
2159 * is required to hold the mmap_sem in read mode. We need the
2160 * anon_vma lock to serialize against concurrent expand_stacks.
2162 vma_lock_anon_vma(vma
);
2164 /* Somebody else might have raced and expanded it already */
2165 if (address
> vma
->vm_end
) {
2166 unsigned long size
, grow
;
2168 size
= address
- vma
->vm_start
;
2169 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
2172 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
2173 error
= acct_stack_growth(vma
, size
, grow
);
2176 * vma_gap_update() doesn't support concurrent
2177 * updates, but we only hold a shared mmap_sem
2178 * lock here, so we need to protect against
2179 * concurrent vma expansions.
2180 * vma_lock_anon_vma() doesn't help here, as
2181 * we don't guarantee that all growable vmas
2182 * in a mm share the same root anon vma.
2183 * So, we reuse mm->page_table_lock to guard
2184 * against concurrent vma expansions.
2186 spin_lock(&vma
->vm_mm
->page_table_lock
);
2187 anon_vma_interval_tree_pre_update_vma(vma
);
2188 vma
->vm_end
= address
;
2189 anon_vma_interval_tree_post_update_vma(vma
);
2191 vma_gap_update(vma
->vm_next
);
2193 vma
->vm_mm
->highest_vm_end
= vm_end_gap(vma
);
2194 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2196 perf_event_mmap(vma
);
2200 vma_unlock_anon_vma(vma
);
2201 khugepaged_enter_vma_merge(vma
);
2202 validate_mm(vma
->vm_mm
);
2205 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2208 * vma is the first one with address < vma->vm_start. Have to extend vma.
2210 int expand_downwards(struct vm_area_struct
*vma
,
2211 unsigned long address
)
2213 struct vm_area_struct
*prev
;
2214 unsigned long gap_addr
;
2217 address
&= PAGE_MASK
;
2218 error
= security_mmap_addr(address
);
2222 /* Enforce stack_guard_gap */
2223 gap_addr
= address
- stack_guard_gap
;
2224 if (gap_addr
> address
)
2226 prev
= vma
->vm_prev
;
2227 if (prev
&& prev
->vm_end
> gap_addr
) {
2228 if (!(prev
->vm_flags
& VM_GROWSDOWN
))
2230 /* Check that both stack segments have the same anon_vma? */
2233 /* We must make sure the anon_vma is allocated. */
2234 if (unlikely(anon_vma_prepare(vma
)))
2238 * vma->vm_start/vm_end cannot change under us because the caller
2239 * is required to hold the mmap_sem in read mode. We need the
2240 * anon_vma lock to serialize against concurrent expand_stacks.
2242 vma_lock_anon_vma(vma
);
2244 /* Somebody else might have raced and expanded it already */
2245 if (address
< vma
->vm_start
) {
2246 unsigned long size
, grow
;
2248 size
= vma
->vm_end
- address
;
2249 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
2252 if (grow
<= vma
->vm_pgoff
) {
2253 error
= acct_stack_growth(vma
, size
, grow
);
2256 * vma_gap_update() doesn't support concurrent
2257 * updates, but we only hold a shared mmap_sem
2258 * lock here, so we need to protect against
2259 * concurrent vma expansions.
2260 * vma_lock_anon_vma() doesn't help here, as
2261 * we don't guarantee that all growable vmas
2262 * in a mm share the same root anon vma.
2263 * So, we reuse mm->page_table_lock to guard
2264 * against concurrent vma expansions.
2266 spin_lock(&vma
->vm_mm
->page_table_lock
);
2267 anon_vma_interval_tree_pre_update_vma(vma
);
2268 vma
->vm_start
= address
;
2269 vma
->vm_pgoff
-= grow
;
2270 anon_vma_interval_tree_post_update_vma(vma
);
2271 vma_gap_update(vma
);
2272 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2274 perf_event_mmap(vma
);
2278 vma_unlock_anon_vma(vma
);
2279 khugepaged_enter_vma_merge(vma
);
2280 validate_mm(vma
->vm_mm
);
2284 /* enforced gap between the expanding stack and other mappings. */
2285 unsigned long stack_guard_gap
= 256UL<<PAGE_SHIFT
;
2287 static int __init
cmdline_parse_stack_guard_gap(char *p
)
2292 val
= simple_strtoul(p
, &endptr
, 10);
2294 stack_guard_gap
= val
<< PAGE_SHIFT
;
2298 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap
);
2300 #ifdef CONFIG_STACK_GROWSUP
2301 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2303 return expand_upwards(vma
, address
);
2306 struct vm_area_struct
*
2307 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2309 struct vm_area_struct
*vma
, *prev
;
2312 vma
= find_vma_prev(mm
, addr
, &prev
);
2313 if (vma
&& (vma
->vm_start
<= addr
))
2315 if (!prev
|| expand_stack(prev
, addr
))
2317 if (prev
->vm_flags
& VM_LOCKED
)
2318 __mlock_vma_pages_range(prev
, addr
, prev
->vm_end
, NULL
);
2322 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2324 return expand_downwards(vma
, address
);
2327 struct vm_area_struct
*
2328 find_extend_vma(struct mm_struct
* mm
, unsigned long addr
)
2330 struct vm_area_struct
* vma
;
2331 unsigned long start
;
2334 vma
= find_vma(mm
,addr
);
2337 if (vma
->vm_start
<= addr
)
2339 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2341 start
= vma
->vm_start
;
2342 if (expand_stack(vma
, addr
))
2344 if (vma
->vm_flags
& VM_LOCKED
)
2345 __mlock_vma_pages_range(vma
, addr
, start
, NULL
);
2351 * Ok - we have the memory areas we should free on the vma list,
2352 * so release them, and do the vma updates.
2354 * Called with the mm semaphore held.
2356 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2358 unsigned long nr_accounted
= 0;
2360 /* Update high watermark before we lower total_vm */
2361 update_hiwater_vm(mm
);
2363 long nrpages
= vma_pages(vma
);
2365 if (vma
->vm_flags
& VM_ACCOUNT
)
2366 nr_accounted
+= nrpages
;
2367 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, -nrpages
);
2368 vma
= remove_vma(vma
);
2370 vm_unacct_memory(nr_accounted
);
2375 * Get rid of page table information in the indicated region.
2377 * Called with the mm semaphore held.
2379 static void unmap_region(struct mm_struct
*mm
,
2380 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
2381 unsigned long start
, unsigned long end
)
2383 struct vm_area_struct
*next
= prev
? prev
->vm_next
: mm
->mmap
;
2384 struct mmu_gather tlb
;
2387 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2388 update_hiwater_rss(mm
);
2389 unmap_vmas(&tlb
, vma
, start
, end
);
2390 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
2391 next
? next
->vm_start
: USER_PGTABLES_CEILING
);
2392 tlb_finish_mmu(&tlb
, start
, end
);
2396 * Create a list of vma's touched by the unmap, removing them from the mm's
2397 * vma list as we go..
2400 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2401 struct vm_area_struct
*prev
, unsigned long end
)
2403 struct vm_area_struct
**insertion_point
;
2404 struct vm_area_struct
*tail_vma
= NULL
;
2407 insertion_point
= (prev
? &prev
->vm_next
: &mm
->mmap
);
2408 vma
->vm_prev
= NULL
;
2410 vma_rb_erase(vma
, &mm
->mm_rb
);
2414 } while (vma
&& vma
->vm_start
< end
);
2415 *insertion_point
= vma
;
2417 vma
->vm_prev
= prev
;
2418 vma_gap_update(vma
);
2420 mm
->highest_vm_end
= prev
? vm_end_gap(prev
) : 0;
2421 tail_vma
->vm_next
= NULL
;
2422 if (mm
->unmap_area
== arch_unmap_area
)
2423 addr
= prev
? prev
->vm_end
: mm
->mmap_base
;
2425 addr
= vma
? vma
->vm_start
: mm
->mmap_base
;
2426 mm
->unmap_area(mm
, addr
);
2427 mm
->mmap_cache
= NULL
; /* Kill the cache. */
2431 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2432 * munmap path where it doesn't make sense to fail.
2434 static int __split_vma(struct mm_struct
* mm
, struct vm_area_struct
* vma
,
2435 unsigned long addr
, int new_below
)
2437 struct mempolicy
*pol
;
2438 struct vm_area_struct
*new;
2441 if (is_vm_hugetlb_page(vma
) && (addr
&
2442 ~(huge_page_mask(hstate_vma(vma
)))))
2445 new = kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2449 /* most fields are the same, copy all, and then fixup */
2452 INIT_LIST_HEAD(&new->anon_vma_chain
);
2457 new->vm_start
= addr
;
2458 new->vm_pgoff
+= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
);
2461 pol
= mpol_dup(vma_policy(vma
));
2466 vma_set_policy(new, pol
);
2468 if (anon_vma_clone(new, vma
))
2472 get_file(new->vm_file
);
2474 if (new->vm_ops
&& new->vm_ops
->open
)
2475 new->vm_ops
->open(new);
2478 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2479 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2481 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2487 /* Clean everything up if vma_adjust failed. */
2488 if (new->vm_ops
&& new->vm_ops
->close
)
2489 new->vm_ops
->close(new);
2492 unlink_anon_vmas(new);
2496 kmem_cache_free(vm_area_cachep
, new);
2502 * Split a vma into two pieces at address 'addr', a new vma is allocated
2503 * either for the first part or the tail.
2505 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2506 unsigned long addr
, int new_below
)
2508 if (mm
->map_count
>= sysctl_max_map_count
)
2511 return __split_vma(mm
, vma
, addr
, new_below
);
2514 /* Munmap is split into 2 main parts -- this part which finds
2515 * what needs doing, and the areas themselves, which do the
2516 * work. This now handles partial unmappings.
2517 * Jeremy Fitzhardinge <jeremy@goop.org>
2519 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
)
2522 struct vm_area_struct
*vma
, *prev
, *last
;
2524 if ((start
& ~PAGE_MASK
) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2527 if ((len
= PAGE_ALIGN(len
)) == 0)
2530 /* Find the first overlapping VMA */
2531 vma
= find_vma(mm
, start
);
2534 prev
= vma
->vm_prev
;
2535 /* we have start < vma->vm_end */
2537 /* if it doesn't overlap, we have nothing.. */
2539 if (vma
->vm_start
>= end
)
2543 * If we need to split any vma, do it now to save pain later.
2545 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2546 * unmapped vm_area_struct will remain in use: so lower split_vma
2547 * places tmp vma above, and higher split_vma places tmp vma below.
2549 if (start
> vma
->vm_start
) {
2553 * Make sure that map_count on return from munmap() will
2554 * not exceed its limit; but let map_count go just above
2555 * its limit temporarily, to help free resources as expected.
2557 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2560 error
= __split_vma(mm
, vma
, start
, 0);
2566 /* Does it split the last one? */
2567 last
= find_vma(mm
, end
);
2568 if (last
&& end
> last
->vm_start
) {
2569 int error
= __split_vma(mm
, last
, end
, 1);
2573 vma
= prev
? prev
->vm_next
: mm
->mmap
;
2576 * unlock any mlock()ed ranges before detaching vmas
2578 if (mm
->locked_vm
) {
2579 struct vm_area_struct
*tmp
= vma
;
2580 while (tmp
&& tmp
->vm_start
< end
) {
2581 if (tmp
->vm_flags
& VM_LOCKED
) {
2582 mm
->locked_vm
-= vma_pages(tmp
);
2583 munlock_vma_pages_all(tmp
);
2590 * Remove the vma's, and unmap the actual pages
2592 detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
);
2593 unmap_region(mm
, vma
, prev
, start
, end
);
2595 /* Fix up all other VM information */
2596 remove_vma_list(mm
, vma
);
2601 int vm_munmap(unsigned long start
, size_t len
)
2604 struct mm_struct
*mm
= current
->mm
;
2606 down_write(&mm
->mmap_sem
);
2607 ret
= do_munmap(mm
, start
, len
);
2608 up_write(&mm
->mmap_sem
);
2611 EXPORT_SYMBOL(vm_munmap
);
2613 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2615 profile_munmap(addr
);
2616 return vm_munmap(addr
, len
);
2619 static inline void verify_mm_writelocked(struct mm_struct
*mm
)
2621 #ifdef CONFIG_DEBUG_VM
2622 if (unlikely(down_read_trylock(&mm
->mmap_sem
))) {
2624 up_read(&mm
->mmap_sem
);
2630 * this is really a simplified "do_mmap". it only handles
2631 * anonymous maps. eventually we may be able to do some
2632 * brk-specific accounting here.
2634 static unsigned long do_brk(unsigned long addr
, unsigned long len
)
2636 struct mm_struct
* mm
= current
->mm
;
2637 struct vm_area_struct
* vma
, * prev
;
2638 unsigned long flags
;
2639 struct rb_node
** rb_link
, * rb_parent
;
2640 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2643 len
= PAGE_ALIGN(len
);
2647 flags
= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2649 error
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2650 if (error
& ~PAGE_MASK
)
2656 if (mm
->def_flags
& VM_LOCKED
) {
2657 unsigned long locked
, lock_limit
;
2658 locked
= len
>> PAGE_SHIFT
;
2659 locked
+= mm
->locked_vm
;
2660 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
2661 lock_limit
>>= PAGE_SHIFT
;
2662 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
2667 * mm->mmap_sem is required to protect against another thread
2668 * changing the mappings in case we sleep.
2670 verify_mm_writelocked(mm
);
2673 * Clear old maps. this also does some error checking for us
2676 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
2677 if (do_munmap(mm
, addr
, len
))
2682 /* Check against address space limits *after* clearing old maps... */
2683 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
))
2686 if (mm
->map_count
> sysctl_max_map_count
)
2689 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
2692 /* Can we just expand an old private anonymous mapping? */
2693 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
2694 NULL
, NULL
, pgoff
, NULL
);
2699 * create a vma struct for an anonymous mapping
2701 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2703 vm_unacct_memory(len
>> PAGE_SHIFT
);
2707 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2709 vma
->vm_start
= addr
;
2710 vma
->vm_end
= addr
+ len
;
2711 vma
->vm_pgoff
= pgoff
;
2712 vma
->vm_flags
= flags
;
2713 vma
->vm_page_prot
= vm_get_page_prot(flags
);
2714 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2716 perf_event_mmap(vma
);
2717 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2718 if (flags
& VM_LOCKED
)
2719 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
2723 unsigned long vm_brk(unsigned long addr
, unsigned long len
)
2725 struct mm_struct
*mm
= current
->mm
;
2729 down_write(&mm
->mmap_sem
);
2730 ret
= do_brk(addr
, len
);
2731 populate
= ((mm
->def_flags
& VM_LOCKED
) != 0);
2732 up_write(&mm
->mmap_sem
);
2734 mm_populate(addr
, len
);
2737 EXPORT_SYMBOL(vm_brk
);
2739 /* Release all mmaps. */
2740 void exit_mmap(struct mm_struct
*mm
)
2742 struct mmu_gather tlb
;
2743 struct vm_area_struct
*vma
;
2744 unsigned long nr_accounted
= 0;
2746 /* mm's last user has gone, and its about to be pulled down */
2747 mmu_notifier_release(mm
);
2749 if (mm
->locked_vm
) {
2752 if (vma
->vm_flags
& VM_LOCKED
)
2753 munlock_vma_pages_all(vma
);
2761 if (!vma
) /* Can happen if dup_mmap() received an OOM */
2766 tlb_gather_mmu(&tlb
, mm
, 0, -1);
2767 /* update_hiwater_rss(mm) here? but nobody should be looking */
2768 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2769 unmap_vmas(&tlb
, vma
, 0, -1);
2771 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, USER_PGTABLES_CEILING
);
2772 tlb_finish_mmu(&tlb
, 0, -1);
2775 * Walk the list again, actually closing and freeing it,
2776 * with preemption enabled, without holding any MM locks.
2779 if (vma
->vm_flags
& VM_ACCOUNT
)
2780 nr_accounted
+= vma_pages(vma
);
2781 vma
= remove_vma(vma
);
2783 vm_unacct_memory(nr_accounted
);
2785 WARN_ON(mm
->nr_ptes
> (FIRST_USER_ADDRESS
+PMD_SIZE
-1)>>PMD_SHIFT
);
2788 /* Insert vm structure into process list sorted by address
2789 * and into the inode's i_mmap tree. If vm_file is non-NULL
2790 * then i_mmap_mutex is taken here.
2792 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2794 struct vm_area_struct
*prev
;
2795 struct rb_node
**rb_link
, *rb_parent
;
2798 * The vm_pgoff of a purely anonymous vma should be irrelevant
2799 * until its first write fault, when page's anon_vma and index
2800 * are set. But now set the vm_pgoff it will almost certainly
2801 * end up with (unless mremap moves it elsewhere before that
2802 * first wfault), so /proc/pid/maps tells a consistent story.
2804 * By setting it to reflect the virtual start address of the
2805 * vma, merges and splits can happen in a seamless way, just
2806 * using the existing file pgoff checks and manipulations.
2807 * Similarly in do_mmap_pgoff and in do_brk.
2809 if (!vma
->vm_file
) {
2810 BUG_ON(vma
->anon_vma
);
2811 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2813 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
2814 &prev
, &rb_link
, &rb_parent
))
2816 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
2817 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
2820 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2825 * Copy the vma structure to a new location in the same mm,
2826 * prior to moving page table entries, to effect an mremap move.
2828 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
2829 unsigned long addr
, unsigned long len
, pgoff_t pgoff
,
2830 bool *need_rmap_locks
)
2832 struct vm_area_struct
*vma
= *vmap
;
2833 unsigned long vma_start
= vma
->vm_start
;
2834 struct mm_struct
*mm
= vma
->vm_mm
;
2835 struct vm_area_struct
*new_vma
, *prev
;
2836 struct rb_node
**rb_link
, *rb_parent
;
2837 struct mempolicy
*pol
;
2838 bool faulted_in_anon_vma
= true;
2841 * If anonymous vma has not yet been faulted, update new pgoff
2842 * to match new location, to increase its chance of merging.
2844 if (unlikely(!vma
->vm_file
&& !vma
->anon_vma
)) {
2845 pgoff
= addr
>> PAGE_SHIFT
;
2846 faulted_in_anon_vma
= false;
2849 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
2850 return NULL
; /* should never get here */
2851 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
2852 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
));
2855 * Source vma may have been merged into new_vma
2857 if (unlikely(vma_start
>= new_vma
->vm_start
&&
2858 vma_start
< new_vma
->vm_end
)) {
2860 * The only way we can get a vma_merge with
2861 * self during an mremap is if the vma hasn't
2862 * been faulted in yet and we were allowed to
2863 * reset the dst vma->vm_pgoff to the
2864 * destination address of the mremap to allow
2865 * the merge to happen. mremap must change the
2866 * vm_pgoff linearity between src and dst vmas
2867 * (in turn preventing a vma_merge) to be
2868 * safe. It is only safe to keep the vm_pgoff
2869 * linear if there are no pages mapped yet.
2871 VM_BUG_ON(faulted_in_anon_vma
);
2872 *vmap
= vma
= new_vma
;
2874 *need_rmap_locks
= (new_vma
->vm_pgoff
<= vma
->vm_pgoff
);
2876 new_vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2879 new_vma
->vm_start
= addr
;
2880 new_vma
->vm_end
= addr
+ len
;
2881 new_vma
->vm_pgoff
= pgoff
;
2882 pol
= mpol_dup(vma_policy(vma
));
2885 vma_set_policy(new_vma
, pol
);
2886 INIT_LIST_HEAD(&new_vma
->anon_vma_chain
);
2887 if (anon_vma_clone(new_vma
, vma
))
2888 goto out_free_mempol
;
2889 if (new_vma
->vm_file
)
2890 get_file(new_vma
->vm_file
);
2891 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
2892 new_vma
->vm_ops
->open(new_vma
);
2893 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
2894 *need_rmap_locks
= false;
2902 kmem_cache_free(vm_area_cachep
, new_vma
);
2907 * Return true if the calling process may expand its vm space by the passed
2910 int may_expand_vm(struct mm_struct
*mm
, unsigned long npages
)
2912 unsigned long cur
= mm
->total_vm
; /* pages */
2915 lim
= rlimit(RLIMIT_AS
) >> PAGE_SHIFT
;
2917 if (cur
+ npages
> lim
)
2923 static int special_mapping_fault(struct vm_area_struct
*vma
,
2924 struct vm_fault
*vmf
)
2927 struct page
**pages
;
2930 * special mappings have no vm_file, and in that case, the mm
2931 * uses vm_pgoff internally. So we have to subtract it from here.
2932 * We are allowed to do this because we are the mm; do not copy
2933 * this code into drivers!
2935 pgoff
= vmf
->pgoff
- vma
->vm_pgoff
;
2937 for (pages
= vma
->vm_private_data
; pgoff
&& *pages
; ++pages
)
2941 struct page
*page
= *pages
;
2947 return VM_FAULT_SIGBUS
;
2951 * Having a close hook prevents vma merging regardless of flags.
2953 static void special_mapping_close(struct vm_area_struct
*vma
)
2957 static const struct vm_operations_struct special_mapping_vmops
= {
2958 .close
= special_mapping_close
,
2959 .fault
= special_mapping_fault
,
2963 * Called with mm->mmap_sem held for writing.
2964 * Insert a new vma covering the given region, with the given flags.
2965 * Its pages are supplied by the given array of struct page *.
2966 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2967 * The region past the last page supplied will always produce SIGBUS.
2968 * The array pointer and the pages it points to are assumed to stay alive
2969 * for as long as this mapping might exist.
2971 int install_special_mapping(struct mm_struct
*mm
,
2972 unsigned long addr
, unsigned long len
,
2973 unsigned long vm_flags
, struct page
**pages
)
2976 struct vm_area_struct
*vma
;
2978 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2979 if (unlikely(vma
== NULL
))
2982 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2984 vma
->vm_start
= addr
;
2985 vma
->vm_end
= addr
+ len
;
2987 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
;
2988 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
2990 vma
->vm_ops
= &special_mapping_vmops
;
2991 vma
->vm_private_data
= pages
;
2993 ret
= insert_vm_struct(mm
, vma
);
2997 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2999 perf_event_mmap(vma
);
3004 kmem_cache_free(vm_area_cachep
, vma
);
3008 static DEFINE_MUTEX(mm_all_locks_mutex
);
3010 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
3012 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3014 * The LSB of head.next can't change from under us
3015 * because we hold the mm_all_locks_mutex.
3017 down_write_nest_lock(&anon_vma
->root
->rwsem
, &mm
->mmap_sem
);
3019 * We can safely modify head.next after taking the
3020 * anon_vma->root->rwsem. If some other vma in this mm shares
3021 * the same anon_vma we won't take it again.
3023 * No need of atomic instructions here, head.next
3024 * can't change from under us thanks to the
3025 * anon_vma->root->rwsem.
3027 if (__test_and_set_bit(0, (unsigned long *)
3028 &anon_vma
->root
->rb_root
.rb_node
))
3033 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
3035 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3037 * AS_MM_ALL_LOCKS can't change from under us because
3038 * we hold the mm_all_locks_mutex.
3040 * Operations on ->flags have to be atomic because
3041 * even if AS_MM_ALL_LOCKS is stable thanks to the
3042 * mm_all_locks_mutex, there may be other cpus
3043 * changing other bitflags in parallel to us.
3045 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
3047 mutex_lock_nest_lock(&mapping
->i_mmap_mutex
, &mm
->mmap_sem
);
3052 * This operation locks against the VM for all pte/vma/mm related
3053 * operations that could ever happen on a certain mm. This includes
3054 * vmtruncate, try_to_unmap, and all page faults.
3056 * The caller must take the mmap_sem in write mode before calling
3057 * mm_take_all_locks(). The caller isn't allowed to release the
3058 * mmap_sem until mm_drop_all_locks() returns.
3060 * mmap_sem in write mode is required in order to block all operations
3061 * that could modify pagetables and free pages without need of
3062 * altering the vma layout (for example populate_range() with
3063 * nonlinear vmas). It's also needed in write mode to avoid new
3064 * anon_vmas to be associated with existing vmas.
3066 * A single task can't take more than one mm_take_all_locks() in a row
3067 * or it would deadlock.
3069 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3070 * mapping->flags avoid to take the same lock twice, if more than one
3071 * vma in this mm is backed by the same anon_vma or address_space.
3073 * We can take all the locks in random order because the VM code
3074 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3075 * takes more than one of them in a row. Secondly we're protected
3076 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3078 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3079 * that may have to take thousand of locks.
3081 * mm_take_all_locks() can fail if it's interrupted by signals.
3083 int mm_take_all_locks(struct mm_struct
*mm
)
3085 struct vm_area_struct
*vma
;
3086 struct anon_vma_chain
*avc
;
3088 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3090 mutex_lock(&mm_all_locks_mutex
);
3092 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3093 if (signal_pending(current
))
3095 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3096 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3099 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3100 if (signal_pending(current
))
3103 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3104 vm_lock_anon_vma(mm
, avc
->anon_vma
);
3110 mm_drop_all_locks(mm
);
3114 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
3116 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3118 * The LSB of head.next can't change to 0 from under
3119 * us because we hold the mm_all_locks_mutex.
3121 * We must however clear the bitflag before unlocking
3122 * the vma so the users using the anon_vma->rb_root will
3123 * never see our bitflag.
3125 * No need of atomic instructions here, head.next
3126 * can't change from under us until we release the
3127 * anon_vma->root->rwsem.
3129 if (!__test_and_clear_bit(0, (unsigned long *)
3130 &anon_vma
->root
->rb_root
.rb_node
))
3132 anon_vma_unlock_write(anon_vma
);
3136 static void vm_unlock_mapping(struct address_space
*mapping
)
3138 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3140 * AS_MM_ALL_LOCKS can't change to 0 from under us
3141 * because we hold the mm_all_locks_mutex.
3143 mutex_unlock(&mapping
->i_mmap_mutex
);
3144 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
3151 * The mmap_sem cannot be released by the caller until
3152 * mm_drop_all_locks() returns.
3154 void mm_drop_all_locks(struct mm_struct
*mm
)
3156 struct vm_area_struct
*vma
;
3157 struct anon_vma_chain
*avc
;
3159 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3160 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
3162 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3164 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3165 vm_unlock_anon_vma(avc
->anon_vma
);
3166 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3167 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
3170 mutex_unlock(&mm_all_locks_mutex
);
3174 * initialise the VMA slab
3176 void __init
mmap_init(void)
3180 ret
= percpu_counter_init(&vm_committed_as
, 0);
3185 * Initialise sysctl_user_reserve_kbytes.
3187 * This is intended to prevent a user from starting a single memory hogging
3188 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3191 * The default value is min(3% of free memory, 128MB)
3192 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3194 static int init_user_reserve(void)
3196 unsigned long free_kbytes
;
3198 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3200 sysctl_user_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 17);
3203 module_init(init_user_reserve
)
3206 * Initialise sysctl_admin_reserve_kbytes.
3208 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3209 * to log in and kill a memory hogging process.
3211 * Systems with more than 256MB will reserve 8MB, enough to recover
3212 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3213 * only reserve 3% of free pages by default.
3215 static int init_admin_reserve(void)
3217 unsigned long free_kbytes
;
3219 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3221 sysctl_admin_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 13);
3224 module_init(init_admin_reserve
)
3227 * Reinititalise user and admin reserves if memory is added or removed.
3229 * The default user reserve max is 128MB, and the default max for the
3230 * admin reserve is 8MB. These are usually, but not always, enough to
3231 * enable recovery from a memory hogging process using login/sshd, a shell,
3232 * and tools like top. It may make sense to increase or even disable the
3233 * reserve depending on the existence of swap or variations in the recovery
3234 * tools. So, the admin may have changed them.
3236 * If memory is added and the reserves have been eliminated or increased above
3237 * the default max, then we'll trust the admin.
3239 * If memory is removed and there isn't enough free memory, then we
3240 * need to reset the reserves.
3242 * Otherwise keep the reserve set by the admin.
3244 static int reserve_mem_notifier(struct notifier_block
*nb
,
3245 unsigned long action
, void *data
)
3247 unsigned long tmp
, free_kbytes
;
3251 /* Default max is 128MB. Leave alone if modified by operator. */
3252 tmp
= sysctl_user_reserve_kbytes
;
3253 if (0 < tmp
&& tmp
< (1UL << 17))
3254 init_user_reserve();
3256 /* Default max is 8MB. Leave alone if modified by operator. */
3257 tmp
= sysctl_admin_reserve_kbytes
;
3258 if (0 < tmp
&& tmp
< (1UL << 13))
3259 init_admin_reserve();
3263 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3265 if (sysctl_user_reserve_kbytes
> free_kbytes
) {
3266 init_user_reserve();
3267 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3268 sysctl_user_reserve_kbytes
);
3271 if (sysctl_admin_reserve_kbytes
> free_kbytes
) {
3272 init_admin_reserve();
3273 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3274 sysctl_admin_reserve_kbytes
);
3283 static struct notifier_block reserve_mem_nb
= {
3284 .notifier_call
= reserve_mem_notifier
,
3287 static int __meminit
init_reserve_notifier(void)
3289 if (register_hotmemory_notifier(&reserve_mem_nb
))
3290 printk("Failed registering memory add/remove notifier for admin reserve");
3294 module_init(init_reserve_notifier
)