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
,
911 const char __user
*anon_name
)
913 if (vma
->vm_flags
^ vm_flags
)
915 if (vma
->vm_file
!= file
)
917 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
919 if (vma_get_anon_name(vma
) != anon_name
)
924 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
925 struct anon_vma
*anon_vma2
,
926 struct vm_area_struct
*vma
)
929 * The list_is_singular() test is to avoid merging VMA cloned from
930 * parents. This can improve scalability caused by anon_vma lock.
932 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
933 list_is_singular(&vma
->anon_vma_chain
)))
935 return anon_vma1
== anon_vma2
;
939 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
940 * in front of (at a lower virtual address and file offset than) the vma.
942 * We cannot merge two vmas if they have differently assigned (non-NULL)
943 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
945 * We don't check here for the merged mmap wrapping around the end of pagecache
946 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
947 * wrap, nor mmaps which cover the final page at index -1UL.
950 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
951 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
,
952 const char __user
*anon_name
)
954 if (is_mergeable_vma(vma
, file
, vm_flags
, anon_name
) &&
955 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
956 if (vma
->vm_pgoff
== vm_pgoff
)
963 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
964 * beyond (at a higher virtual address and file offset than) the vma.
966 * We cannot merge two vmas if they have differently assigned (non-NULL)
967 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
970 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
971 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
,
972 const char __user
*anon_name
)
974 if (is_mergeable_vma(vma
, file
, vm_flags
, anon_name
) &&
975 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
977 vm_pglen
= (vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
;
978 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
985 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
986 * figure out whether that can be merged with its predecessor or its
987 * successor. Or both (it neatly fills a hole).
989 * In most cases - when called for mmap, brk or mremap - [addr,end) is
990 * certain not to be mapped by the time vma_merge is called; but when
991 * called for mprotect, it is certain to be already mapped (either at
992 * an offset within prev, or at the start of next), and the flags of
993 * this area are about to be changed to vm_flags - and the no-change
994 * case has already been eliminated.
996 * The following mprotect cases have to be considered, where AAAA is
997 * the area passed down from mprotect_fixup, never extending beyond one
998 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1000 * AAAA AAAA AAAA AAAA
1001 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1002 * cannot merge might become might become might become
1003 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1004 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1005 * mremap move: PPPPNNNNNNNN 8
1007 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1008 * might become case 1 below case 2 below case 3 below
1010 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1011 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1013 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
1014 struct vm_area_struct
*prev
, unsigned long addr
,
1015 unsigned long end
, unsigned long vm_flags
,
1016 struct anon_vma
*anon_vma
, struct file
*file
,
1017 pgoff_t pgoff
, struct mempolicy
*policy
,
1018 const char __user
*anon_name
)
1020 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
1021 struct vm_area_struct
*area
, *next
;
1025 * We later require that vma->vm_flags == vm_flags,
1026 * so this tests vma->vm_flags & VM_SPECIAL, too.
1028 if (vm_flags
& VM_SPECIAL
)
1032 next
= prev
->vm_next
;
1036 if (next
&& next
->vm_end
== end
) /* cases 6, 7, 8 */
1037 next
= next
->vm_next
;
1040 * Can it merge with the predecessor?
1042 if (prev
&& prev
->vm_end
== addr
&&
1043 mpol_equal(vma_policy(prev
), policy
) &&
1044 can_vma_merge_after(prev
, vm_flags
, anon_vma
,
1045 file
, pgoff
, anon_name
)) {
1047 * OK, it can. Can we now merge in the successor as well?
1049 if (next
&& end
== next
->vm_start
&&
1050 mpol_equal(policy
, vma_policy(next
)) &&
1051 can_vma_merge_before(next
, vm_flags
, anon_vma
,
1052 file
, pgoff
+pglen
, anon_name
) &&
1053 is_mergeable_anon_vma(prev
->anon_vma
,
1054 next
->anon_vma
, NULL
)) {
1056 err
= vma_adjust(prev
, prev
->vm_start
,
1057 next
->vm_end
, prev
->vm_pgoff
, NULL
);
1058 } else /* cases 2, 5, 7 */
1059 err
= vma_adjust(prev
, prev
->vm_start
,
1060 end
, prev
->vm_pgoff
, NULL
);
1063 khugepaged_enter_vma_merge(prev
);
1068 * Can this new request be merged in front of next?
1070 if (next
&& end
== next
->vm_start
&&
1071 mpol_equal(policy
, vma_policy(next
)) &&
1072 can_vma_merge_before(next
, vm_flags
, anon_vma
,
1073 file
, pgoff
+pglen
, anon_name
)) {
1074 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
1075 err
= vma_adjust(prev
, prev
->vm_start
,
1076 addr
, prev
->vm_pgoff
, NULL
);
1077 else /* cases 3, 8 */
1078 err
= vma_adjust(area
, addr
, next
->vm_end
,
1079 next
->vm_pgoff
- pglen
, NULL
);
1082 khugepaged_enter_vma_merge(area
);
1090 * Rough compatbility check to quickly see if it's even worth looking
1091 * at sharing an anon_vma.
1093 * They need to have the same vm_file, and the flags can only differ
1094 * in things that mprotect may change.
1096 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1097 * we can merge the two vma's. For example, we refuse to merge a vma if
1098 * there is a vm_ops->close() function, because that indicates that the
1099 * driver is doing some kind of reference counting. But that doesn't
1100 * really matter for the anon_vma sharing case.
1102 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1104 return a
->vm_end
== b
->vm_start
&&
1105 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
1106 a
->vm_file
== b
->vm_file
&&
1107 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_READ
|VM_WRITE
|VM_EXEC
)) &&
1108 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
1112 * Do some basic sanity checking to see if we can re-use the anon_vma
1113 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1114 * the same as 'old', the other will be the new one that is trying
1115 * to share the anon_vma.
1117 * NOTE! This runs with mm_sem held for reading, so it is possible that
1118 * the anon_vma of 'old' is concurrently in the process of being set up
1119 * by another page fault trying to merge _that_. But that's ok: if it
1120 * is being set up, that automatically means that it will be a singleton
1121 * acceptable for merging, so we can do all of this optimistically. But
1122 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1124 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1125 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1126 * is to return an anon_vma that is "complex" due to having gone through
1129 * We also make sure that the two vma's are compatible (adjacent,
1130 * and with the same memory policies). That's all stable, even with just
1131 * a read lock on the mm_sem.
1133 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1135 if (anon_vma_compatible(a
, b
)) {
1136 struct anon_vma
*anon_vma
= ACCESS_ONCE(old
->anon_vma
);
1138 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
1145 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1146 * neighbouring vmas for a suitable anon_vma, before it goes off
1147 * to allocate a new anon_vma. It checks because a repetitive
1148 * sequence of mprotects and faults may otherwise lead to distinct
1149 * anon_vmas being allocated, preventing vma merge in subsequent
1152 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
1154 struct anon_vma
*anon_vma
;
1155 struct vm_area_struct
*near
;
1157 near
= vma
->vm_next
;
1161 anon_vma
= reusable_anon_vma(near
, vma
, near
);
1165 near
= vma
->vm_prev
;
1169 anon_vma
= reusable_anon_vma(near
, near
, vma
);
1174 * There's no absolute need to look only at touching neighbours:
1175 * we could search further afield for "compatible" anon_vmas.
1176 * But it would probably just be a waste of time searching,
1177 * or lead to too many vmas hanging off the same anon_vma.
1178 * We're trying to allow mprotect remerging later on,
1179 * not trying to minimize memory used for anon_vmas.
1184 #ifdef CONFIG_PROC_FS
1185 void vm_stat_account(struct mm_struct
*mm
, unsigned long flags
,
1186 struct file
*file
, long pages
)
1188 const unsigned long stack_flags
1189 = VM_STACK_FLAGS
& (VM_GROWSUP
|VM_GROWSDOWN
);
1191 mm
->total_vm
+= pages
;
1194 mm
->shared_vm
+= pages
;
1195 if ((flags
& (VM_EXEC
|VM_WRITE
)) == VM_EXEC
)
1196 mm
->exec_vm
+= pages
;
1197 } else if (flags
& stack_flags
)
1198 mm
->stack_vm
+= pages
;
1200 #endif /* CONFIG_PROC_FS */
1203 * If a hint addr is less than mmap_min_addr change hint to be as
1204 * low as possible but still greater than mmap_min_addr
1206 static inline unsigned long round_hint_to_min(unsigned long hint
)
1209 if (((void *)hint
!= NULL
) &&
1210 (hint
< mmap_min_addr
))
1211 return PAGE_ALIGN(mmap_min_addr
);
1216 * The caller must hold down_write(¤t->mm->mmap_sem).
1219 unsigned long do_mmap_pgoff(struct file
*file
, unsigned long addr
,
1220 unsigned long len
, unsigned long prot
,
1221 unsigned long flags
, unsigned long pgoff
,
1222 unsigned long *populate
)
1224 struct mm_struct
* mm
= current
->mm
;
1225 struct inode
*inode
;
1226 vm_flags_t vm_flags
;
1231 * Does the application expect PROT_READ to imply PROT_EXEC?
1233 * (the exception is when the underlying filesystem is noexec
1234 * mounted, in which case we dont add PROT_EXEC.)
1236 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1237 if (!(file
&& (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
)))
1243 if (!(flags
& MAP_FIXED
))
1244 addr
= round_hint_to_min(addr
);
1246 /* Careful about overflows.. */
1247 len
= PAGE_ALIGN(len
);
1251 /* offset overflow? */
1252 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1255 /* Too many mappings? */
1256 if (mm
->map_count
> sysctl_max_map_count
)
1259 /* Obtain the address to map to. we verify (or select) it and ensure
1260 * that it represents a valid section of the address space.
1262 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1263 if (addr
& ~PAGE_MASK
)
1266 /* Do simple checking here so the lower-level routines won't have
1267 * to. we assume access permissions have been handled by the open
1268 * of the memory object, so we don't do any here.
1270 vm_flags
= calc_vm_prot_bits(prot
) | calc_vm_flag_bits(flags
) |
1271 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1273 if (flags
& MAP_LOCKED
)
1274 if (!can_do_mlock())
1277 /* mlock MCL_FUTURE? */
1278 if (vm_flags
& VM_LOCKED
) {
1279 unsigned long locked
, lock_limit
;
1280 locked
= len
>> PAGE_SHIFT
;
1281 locked
+= mm
->locked_vm
;
1282 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1283 lock_limit
>>= PAGE_SHIFT
;
1284 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1288 inode
= file
? file_inode(file
) : NULL
;
1291 switch (flags
& MAP_TYPE
) {
1293 if ((prot
&PROT_WRITE
) && !(file
->f_mode
&FMODE_WRITE
))
1297 * Make sure we don't allow writing to an append-only
1300 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1304 * Make sure there are no mandatory locks on the file.
1306 if (locks_verify_locked(inode
))
1309 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1310 if (!(file
->f_mode
& FMODE_WRITE
))
1311 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1315 if (!(file
->f_mode
& FMODE_READ
))
1317 if (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
) {
1318 if (vm_flags
& VM_EXEC
)
1320 vm_flags
&= ~VM_MAYEXEC
;
1323 if (!file
->f_op
|| !file
->f_op
->mmap
)
1331 switch (flags
& MAP_TYPE
) {
1337 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1341 * Set pgoff according to addr for anon_vma.
1343 pgoff
= addr
>> PAGE_SHIFT
;
1351 * Set 'VM_NORESERVE' if we should not account for the
1352 * memory use of this mapping.
1354 if (flags
& MAP_NORESERVE
) {
1355 /* We honor MAP_NORESERVE if allowed to overcommit */
1356 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1357 vm_flags
|= VM_NORESERVE
;
1359 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1360 if (file
&& is_file_hugepages(file
))
1361 vm_flags
|= VM_NORESERVE
;
1364 addr
= mmap_region(file
, addr
, len
, vm_flags
, pgoff
);
1365 if (!IS_ERR_VALUE(addr
) &&
1366 ((vm_flags
& VM_LOCKED
) ||
1367 (flags
& (MAP_POPULATE
| MAP_NONBLOCK
)) == MAP_POPULATE
))
1372 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1373 unsigned long, prot
, unsigned long, flags
,
1374 unsigned long, fd
, unsigned long, pgoff
)
1376 struct file
*file
= NULL
;
1377 unsigned long retval
= -EBADF
;
1379 if (!(flags
& MAP_ANONYMOUS
)) {
1380 audit_mmap_fd(fd
, flags
);
1381 if (unlikely(flags
& MAP_HUGETLB
))
1386 if (is_file_hugepages(file
))
1387 len
= ALIGN(len
, huge_page_size(hstate_file(file
)));
1388 } else if (flags
& MAP_HUGETLB
) {
1389 struct user_struct
*user
= NULL
;
1390 struct hstate
*hs
= hstate_sizelog((flags
>> MAP_HUGE_SHIFT
) &
1396 len
= ALIGN(len
, huge_page_size(hs
));
1398 * VM_NORESERVE is used because the reservations will be
1399 * taken when vm_ops->mmap() is called
1400 * A dummy user value is used because we are not locking
1401 * memory so no accounting is necessary
1403 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, len
,
1405 &user
, HUGETLB_ANONHUGE_INODE
,
1406 (flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1408 return PTR_ERR(file
);
1411 flags
&= ~(MAP_EXECUTABLE
| MAP_DENYWRITE
);
1413 retval
= vm_mmap_pgoff(file
, addr
, len
, prot
, flags
, pgoff
);
1420 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1421 struct mmap_arg_struct
{
1425 unsigned long flags
;
1427 unsigned long offset
;
1430 SYSCALL_DEFINE1(old_mmap
, struct mmap_arg_struct __user
*, arg
)
1432 struct mmap_arg_struct a
;
1434 if (copy_from_user(&a
, arg
, sizeof(a
)))
1436 if (a
.offset
& ~PAGE_MASK
)
1439 return sys_mmap_pgoff(a
.addr
, a
.len
, a
.prot
, a
.flags
, a
.fd
,
1440 a
.offset
>> PAGE_SHIFT
);
1442 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1445 * Some shared mappigns will want the pages marked read-only
1446 * to track write events. If so, we'll downgrade vm_page_prot
1447 * to the private version (using protection_map[] without the
1450 int vma_wants_writenotify(struct vm_area_struct
*vma
)
1452 vm_flags_t vm_flags
= vma
->vm_flags
;
1454 /* If it was private or non-writable, the write bit is already clear */
1455 if ((vm_flags
& (VM_WRITE
|VM_SHARED
)) != ((VM_WRITE
|VM_SHARED
)))
1458 /* The backer wishes to know when pages are first written to? */
1459 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
)
1462 /* The open routine did something to the protections already? */
1463 if (pgprot_val(vma
->vm_page_prot
) !=
1464 pgprot_val(vm_get_page_prot(vm_flags
)))
1467 /* Specialty mapping? */
1468 if (vm_flags
& VM_PFNMAP
)
1471 /* Can the mapping track the dirty pages? */
1472 return vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
1473 mapping_cap_account_dirty(vma
->vm_file
->f_mapping
);
1477 * We account for memory if it's a private writeable mapping,
1478 * not hugepages and VM_NORESERVE wasn't set.
1480 static inline int accountable_mapping(struct file
*file
, vm_flags_t vm_flags
)
1483 * hugetlb has its own accounting separate from the core VM
1484 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1486 if (file
&& is_file_hugepages(file
))
1489 return (vm_flags
& (VM_NORESERVE
| VM_SHARED
| VM_WRITE
)) == VM_WRITE
;
1492 unsigned long mmap_region(struct file
*file
, unsigned long addr
,
1493 unsigned long len
, vm_flags_t vm_flags
, unsigned long pgoff
)
1495 struct mm_struct
*mm
= current
->mm
;
1496 struct vm_area_struct
*vma
, *prev
;
1497 int correct_wcount
= 0;
1499 struct rb_node
**rb_link
, *rb_parent
;
1500 unsigned long charged
= 0;
1501 struct inode
*inode
= file
? file_inode(file
) : NULL
;
1503 /* Check against address space limit. */
1504 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
)) {
1505 unsigned long nr_pages
;
1508 * MAP_FIXED may remove pages of mappings that intersects with
1509 * requested mapping. Account for the pages it would unmap.
1511 if (!(vm_flags
& MAP_FIXED
))
1514 nr_pages
= count_vma_pages_range(mm
, addr
, addr
+ len
);
1516 if (!may_expand_vm(mm
, (len
>> PAGE_SHIFT
) - nr_pages
))
1520 /* Clear old maps */
1523 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
1524 if (do_munmap(mm
, addr
, len
))
1530 * Private writable mapping: check memory availability
1532 if (accountable_mapping(file
, vm_flags
)) {
1533 charged
= len
>> PAGE_SHIFT
;
1534 if (security_vm_enough_memory_mm(mm
, charged
))
1536 vm_flags
|= VM_ACCOUNT
;
1540 * Can we just expand an old mapping?
1542 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
, NULL
, file
, pgoff
,
1548 * Determine the object being mapped and call the appropriate
1549 * specific mapper. the address has already been validated, but
1550 * not unmapped, but the maps are removed from the list.
1552 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
1559 vma
->vm_start
= addr
;
1560 vma
->vm_end
= addr
+ len
;
1561 vma
->vm_flags
= vm_flags
;
1562 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1563 vma
->vm_pgoff
= pgoff
;
1564 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
1566 error
= -EINVAL
; /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1569 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1571 if (vm_flags
& VM_DENYWRITE
) {
1572 error
= deny_write_access(file
);
1577 vma
->vm_file
= get_file(file
);
1578 error
= file
->f_op
->mmap(file
, vma
);
1580 goto unmap_and_free_vma
;
1582 /* Can addr have changed??
1584 * Answer: Yes, several device drivers can do it in their
1585 * f_op->mmap method. -DaveM
1586 * Bug: If addr is changed, prev, rb_link, rb_parent should
1587 * be updated for vma_link()
1589 WARN_ON_ONCE(addr
!= vma
->vm_start
);
1591 addr
= vma
->vm_start
;
1592 pgoff
= vma
->vm_pgoff
;
1593 vm_flags
= vma
->vm_flags
;
1594 } else if (vm_flags
& VM_SHARED
) {
1595 if (unlikely(vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
)))
1597 error
= shmem_zero_setup(vma
);
1602 if (vma_wants_writenotify(vma
)) {
1603 pgprot_t pprot
= vma
->vm_page_prot
;
1605 /* Can vma->vm_page_prot have changed??
1607 * Answer: Yes, drivers may have changed it in their
1608 * f_op->mmap method.
1610 * Ensures that vmas marked as uncached stay that way.
1612 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
& ~VM_SHARED
);
1613 if (pgprot_val(pprot
) == pgprot_val(pgprot_noncached(pprot
)))
1614 vma
->vm_page_prot
= pgprot_noncached(vma
->vm_page_prot
);
1617 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1618 file
= vma
->vm_file
;
1620 /* Once vma denies write, undo our temporary denial count */
1622 atomic_inc(&inode
->i_writecount
);
1624 perf_event_mmap(vma
);
1626 vm_stat_account(mm
, vm_flags
, file
, len
>> PAGE_SHIFT
);
1627 if (vm_flags
& VM_LOCKED
) {
1628 if (!((vm_flags
& VM_SPECIAL
) || is_vm_hugetlb_page(vma
) ||
1629 vma
== get_gate_vma(current
->mm
)))
1630 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1632 vma
->vm_flags
&= ~VM_LOCKED
;
1642 atomic_inc(&inode
->i_writecount
);
1643 vma
->vm_file
= NULL
;
1646 /* Undo any partial mapping done by a device driver. */
1647 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1650 kmem_cache_free(vm_area_cachep
, vma
);
1653 vm_unacct_memory(charged
);
1657 unsigned long unmapped_area(struct vm_unmapped_area_info
*info
)
1660 * We implement the search by looking for an rbtree node that
1661 * immediately follows a suitable gap. That is,
1662 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1663 * - gap_end = vma->vm_start >= info->low_limit + length;
1664 * - gap_end - gap_start >= length
1667 struct mm_struct
*mm
= current
->mm
;
1668 struct vm_area_struct
*vma
;
1669 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1671 /* Adjust search length to account for worst case alignment overhead */
1672 length
= info
->length
+ info
->align_mask
;
1673 if (length
< info
->length
)
1676 /* Adjust search limits by the desired length */
1677 if (info
->high_limit
< length
)
1679 high_limit
= info
->high_limit
- length
;
1681 if (info
->low_limit
> high_limit
)
1683 low_limit
= info
->low_limit
+ length
;
1685 /* Check if rbtree root looks promising */
1686 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1688 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1689 if (vma
->rb_subtree_gap
< length
)
1693 /* Visit left subtree if it looks promising */
1694 gap_end
= vm_start_gap(vma
);
1695 if (gap_end
>= low_limit
&& vma
->vm_rb
.rb_left
) {
1696 struct vm_area_struct
*left
=
1697 rb_entry(vma
->vm_rb
.rb_left
,
1698 struct vm_area_struct
, vm_rb
);
1699 if (left
->rb_subtree_gap
>= length
) {
1705 gap_start
= vma
->vm_prev
? vm_end_gap(vma
->vm_prev
) : 0;
1707 /* Check if current node has a suitable gap */
1708 if (gap_start
> high_limit
)
1710 if (gap_end
>= low_limit
&&
1711 gap_end
> gap_start
&& gap_end
- gap_start
>= length
)
1714 /* Visit right subtree if it looks promising */
1715 if (vma
->vm_rb
.rb_right
) {
1716 struct vm_area_struct
*right
=
1717 rb_entry(vma
->vm_rb
.rb_right
,
1718 struct vm_area_struct
, vm_rb
);
1719 if (right
->rb_subtree_gap
>= length
) {
1725 /* Go back up the rbtree to find next candidate node */
1727 struct rb_node
*prev
= &vma
->vm_rb
;
1728 if (!rb_parent(prev
))
1730 vma
= rb_entry(rb_parent(prev
),
1731 struct vm_area_struct
, vm_rb
);
1732 if (prev
== vma
->vm_rb
.rb_left
) {
1733 gap_start
= vm_end_gap(vma
->vm_prev
);
1734 gap_end
= vm_start_gap(vma
);
1741 /* Check highest gap, which does not precede any rbtree node */
1742 gap_start
= mm
->highest_vm_end
;
1743 gap_end
= ULONG_MAX
; /* Only for VM_BUG_ON below */
1744 if (gap_start
> high_limit
)
1748 /* We found a suitable gap. Clip it with the original low_limit. */
1749 if (gap_start
< info
->low_limit
)
1750 gap_start
= info
->low_limit
;
1752 /* Adjust gap address to the desired alignment */
1753 gap_start
+= (info
->align_offset
- gap_start
) & info
->align_mask
;
1755 VM_BUG_ON(gap_start
+ info
->length
> info
->high_limit
);
1756 VM_BUG_ON(gap_start
+ info
->length
> gap_end
);
1760 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info
*info
)
1762 struct mm_struct
*mm
= current
->mm
;
1763 struct vm_area_struct
*vma
;
1764 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1766 /* Adjust search length to account for worst case alignment overhead */
1767 length
= info
->length
+ info
->align_mask
;
1768 if (length
< info
->length
)
1772 * Adjust search limits by the desired length.
1773 * See implementation comment at top of unmapped_area().
1775 gap_end
= info
->high_limit
;
1776 if (gap_end
< length
)
1778 high_limit
= gap_end
- length
;
1780 if (info
->low_limit
> high_limit
)
1782 low_limit
= info
->low_limit
+ length
;
1784 /* Check highest gap, which does not precede any rbtree node */
1785 gap_start
= mm
->highest_vm_end
;
1786 if (gap_start
<= high_limit
)
1789 /* Check if rbtree root looks promising */
1790 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1792 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1793 if (vma
->rb_subtree_gap
< length
)
1797 /* Visit right subtree if it looks promising */
1798 gap_start
= vma
->vm_prev
? vm_end_gap(vma
->vm_prev
) : 0;
1799 if (gap_start
<= high_limit
&& vma
->vm_rb
.rb_right
) {
1800 struct vm_area_struct
*right
=
1801 rb_entry(vma
->vm_rb
.rb_right
,
1802 struct vm_area_struct
, vm_rb
);
1803 if (right
->rb_subtree_gap
>= length
) {
1810 /* Check if current node has a suitable gap */
1811 gap_end
= vm_start_gap(vma
);
1812 if (gap_end
< low_limit
)
1814 if (gap_start
<= high_limit
&&
1815 gap_end
> gap_start
&& gap_end
- gap_start
>= length
)
1818 /* Visit left subtree if it looks promising */
1819 if (vma
->vm_rb
.rb_left
) {
1820 struct vm_area_struct
*left
=
1821 rb_entry(vma
->vm_rb
.rb_left
,
1822 struct vm_area_struct
, vm_rb
);
1823 if (left
->rb_subtree_gap
>= length
) {
1829 /* Go back up the rbtree to find next candidate node */
1831 struct rb_node
*prev
= &vma
->vm_rb
;
1832 if (!rb_parent(prev
))
1834 vma
= rb_entry(rb_parent(prev
),
1835 struct vm_area_struct
, vm_rb
);
1836 if (prev
== vma
->vm_rb
.rb_right
) {
1837 gap_start
= vma
->vm_prev
?
1838 vm_end_gap(vma
->vm_prev
) : 0;
1845 /* We found a suitable gap. Clip it with the original high_limit. */
1846 if (gap_end
> info
->high_limit
)
1847 gap_end
= info
->high_limit
;
1850 /* Compute highest gap address at the desired alignment */
1851 gap_end
-= info
->length
;
1852 gap_end
-= (gap_end
- info
->align_offset
) & info
->align_mask
;
1854 VM_BUG_ON(gap_end
< info
->low_limit
);
1855 VM_BUG_ON(gap_end
< gap_start
);
1859 /* Get an address range which is currently unmapped.
1860 * For shmat() with addr=0.
1862 * Ugly calling convention alert:
1863 * Return value with the low bits set means error value,
1865 * if (ret & ~PAGE_MASK)
1868 * This function "knows" that -ENOMEM has the bits set.
1870 #ifndef HAVE_ARCH_UNMAPPED_AREA
1872 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
1873 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
1875 struct mm_struct
*mm
= current
->mm
;
1876 struct vm_area_struct
*vma
, *prev
;
1877 struct vm_unmapped_area_info info
;
1879 if (len
> TASK_SIZE
- mmap_min_addr
)
1882 if (flags
& MAP_FIXED
)
1886 addr
= PAGE_ALIGN(addr
);
1887 vma
= find_vma_prev(mm
, addr
, &prev
);
1888 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1889 (!vma
|| addr
+ len
<= vm_start_gap(vma
)) &&
1890 (!prev
|| addr
>= vm_end_gap(prev
)))
1896 info
.low_limit
= TASK_UNMAPPED_BASE
;
1897 info
.high_limit
= TASK_SIZE
;
1898 info
.align_mask
= 0;
1899 return vm_unmapped_area(&info
);
1903 void arch_unmap_area(struct mm_struct
*mm
, unsigned long addr
)
1906 * Is this a new hole at the lowest possible address?
1908 if (addr
>= TASK_UNMAPPED_BASE
&& addr
< mm
->free_area_cache
)
1909 mm
->free_area_cache
= addr
;
1913 * This mmap-allocator allocates new areas top-down from below the
1914 * stack's low limit (the base):
1916 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1918 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
1919 const unsigned long len
, const unsigned long pgoff
,
1920 const unsigned long flags
)
1922 struct vm_area_struct
*vma
, *prev
;
1923 struct mm_struct
*mm
= current
->mm
;
1924 unsigned long addr
= addr0
;
1925 struct vm_unmapped_area_info info
;
1927 /* requested length too big for entire address space */
1928 if (len
> TASK_SIZE
- mmap_min_addr
)
1931 if (flags
& MAP_FIXED
)
1934 /* requesting a specific address */
1936 addr
= PAGE_ALIGN(addr
);
1937 vma
= find_vma_prev(mm
, addr
, &prev
);
1938 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1939 (!vma
|| addr
+ len
<= vm_start_gap(vma
)) &&
1940 (!prev
|| addr
>= vm_end_gap(prev
)))
1944 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
1946 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
1947 info
.high_limit
= mm
->mmap_base
;
1948 info
.align_mask
= 0;
1949 addr
= vm_unmapped_area(&info
);
1952 * A failed mmap() very likely causes application failure,
1953 * so fall back to the bottom-up function here. This scenario
1954 * can happen with large stack limits and large mmap()
1957 if (addr
& ~PAGE_MASK
) {
1958 VM_BUG_ON(addr
!= -ENOMEM
);
1960 info
.low_limit
= TASK_UNMAPPED_BASE
;
1961 info
.high_limit
= TASK_SIZE
;
1962 addr
= vm_unmapped_area(&info
);
1969 void arch_unmap_area_topdown(struct mm_struct
*mm
, unsigned long addr
)
1972 * Is this a new hole at the highest possible address?
1974 if (addr
> mm
->free_area_cache
)
1975 mm
->free_area_cache
= addr
;
1977 /* dont allow allocations above current base */
1978 if (mm
->free_area_cache
> mm
->mmap_base
)
1979 mm
->free_area_cache
= mm
->mmap_base
;
1983 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
1984 unsigned long pgoff
, unsigned long flags
)
1986 unsigned long (*get_area
)(struct file
*, unsigned long,
1987 unsigned long, unsigned long, unsigned long);
1989 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
1993 /* Careful about overflows.. */
1994 if (len
> TASK_SIZE
)
1997 get_area
= current
->mm
->get_unmapped_area
;
1998 if (file
&& file
->f_op
&& file
->f_op
->get_unmapped_area
)
1999 get_area
= file
->f_op
->get_unmapped_area
;
2000 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
2001 if (IS_ERR_VALUE(addr
))
2004 if (addr
> TASK_SIZE
- len
)
2006 if (addr
& ~PAGE_MASK
)
2009 addr
= arch_rebalance_pgtables(addr
, len
);
2010 error
= security_mmap_addr(addr
);
2011 return error
? error
: addr
;
2014 EXPORT_SYMBOL(get_unmapped_area
);
2016 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2017 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
2019 struct vm_area_struct
*vma
= NULL
;
2021 /* Check the cache first. */
2022 /* (Cache hit rate is typically around 35%.) */
2023 vma
= ACCESS_ONCE(mm
->mmap_cache
);
2024 if (!(vma
&& vma
->vm_end
> addr
&& vma
->vm_start
<= addr
)) {
2025 struct rb_node
*rb_node
;
2027 rb_node
= mm
->mm_rb
.rb_node
;
2031 struct vm_area_struct
*vma_tmp
;
2033 vma_tmp
= rb_entry(rb_node
,
2034 struct vm_area_struct
, vm_rb
);
2036 if (vma_tmp
->vm_end
> addr
) {
2038 if (vma_tmp
->vm_start
<= addr
)
2040 rb_node
= rb_node
->rb_left
;
2042 rb_node
= rb_node
->rb_right
;
2045 mm
->mmap_cache
= vma
;
2050 EXPORT_SYMBOL(find_vma
);
2053 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2055 struct vm_area_struct
*
2056 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
2057 struct vm_area_struct
**pprev
)
2059 struct vm_area_struct
*vma
;
2061 vma
= find_vma(mm
, addr
);
2063 *pprev
= vma
->vm_prev
;
2065 struct rb_node
*rb_node
= mm
->mm_rb
.rb_node
;
2068 *pprev
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2069 rb_node
= rb_node
->rb_right
;
2076 * Verify that the stack growth is acceptable and
2077 * update accounting. This is shared with both the
2078 * grow-up and grow-down cases.
2080 static int acct_stack_growth(struct vm_area_struct
*vma
,
2081 unsigned long size
, unsigned long grow
)
2083 struct mm_struct
*mm
= vma
->vm_mm
;
2084 struct rlimit
*rlim
= current
->signal
->rlim
;
2085 unsigned long new_start
;
2087 /* address space limit tests */
2088 if (!may_expand_vm(mm
, grow
))
2091 /* Stack limit test */
2092 if (size
> ACCESS_ONCE(rlim
[RLIMIT_STACK
].rlim_cur
))
2095 /* mlock limit tests */
2096 if (vma
->vm_flags
& VM_LOCKED
) {
2097 unsigned long locked
;
2098 unsigned long limit
;
2099 locked
= mm
->locked_vm
+ grow
;
2100 limit
= ACCESS_ONCE(rlim
[RLIMIT_MEMLOCK
].rlim_cur
);
2101 limit
>>= PAGE_SHIFT
;
2102 if (locked
> limit
&& !capable(CAP_IPC_LOCK
))
2106 /* Check to ensure the stack will not grow into a hugetlb-only region */
2107 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
2109 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
2113 * Overcommit.. This must be the final test, as it will
2114 * update security statistics.
2116 if (security_vm_enough_memory_mm(mm
, grow
))
2119 /* Ok, everything looks good - let it rip */
2120 if (vma
->vm_flags
& VM_LOCKED
)
2121 mm
->locked_vm
+= grow
;
2122 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, grow
);
2126 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2128 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2129 * vma is the last one with address > vma->vm_end. Have to extend vma.
2131 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
)
2133 struct vm_area_struct
*next
;
2134 unsigned long gap_addr
;
2137 if (!(vma
->vm_flags
& VM_GROWSUP
))
2140 /* Guard against exceeding limits of the address space. */
2141 address
&= PAGE_MASK
;
2142 if (address
>= TASK_SIZE
)
2144 address
+= PAGE_SIZE
;
2146 /* Enforce stack_guard_gap */
2147 gap_addr
= address
+ stack_guard_gap
;
2149 /* Guard against overflow */
2150 if (gap_addr
< address
|| gap_addr
> TASK_SIZE
)
2151 gap_addr
= TASK_SIZE
;
2153 next
= vma
->vm_next
;
2154 if (next
&& next
->vm_start
< gap_addr
) {
2155 if (!(next
->vm_flags
& VM_GROWSUP
))
2157 /* Check that both stack segments have the same anon_vma? */
2160 /* We must make sure the anon_vma is allocated. */
2161 if (unlikely(anon_vma_prepare(vma
)))
2165 * vma->vm_start/vm_end cannot change under us because the caller
2166 * is required to hold the mmap_sem in read mode. We need the
2167 * anon_vma lock to serialize against concurrent expand_stacks.
2169 vma_lock_anon_vma(vma
);
2171 /* Somebody else might have raced and expanded it already */
2172 if (address
> vma
->vm_end
) {
2173 unsigned long size
, grow
;
2175 size
= address
- vma
->vm_start
;
2176 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
2179 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
2180 error
= acct_stack_growth(vma
, size
, grow
);
2183 * vma_gap_update() doesn't support concurrent
2184 * updates, but we only hold a shared mmap_sem
2185 * lock here, so we need to protect against
2186 * concurrent vma expansions.
2187 * vma_lock_anon_vma() doesn't help here, as
2188 * we don't guarantee that all growable vmas
2189 * in a mm share the same root anon vma.
2190 * So, we reuse mm->page_table_lock to guard
2191 * against concurrent vma expansions.
2193 spin_lock(&vma
->vm_mm
->page_table_lock
);
2194 anon_vma_interval_tree_pre_update_vma(vma
);
2195 vma
->vm_end
= address
;
2196 anon_vma_interval_tree_post_update_vma(vma
);
2198 vma_gap_update(vma
->vm_next
);
2200 vma
->vm_mm
->highest_vm_end
= vm_end_gap(vma
);
2201 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2203 perf_event_mmap(vma
);
2207 vma_unlock_anon_vma(vma
);
2208 khugepaged_enter_vma_merge(vma
);
2209 validate_mm(vma
->vm_mm
);
2212 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2215 * vma is the first one with address < vma->vm_start. Have to extend vma.
2217 int expand_downwards(struct vm_area_struct
*vma
,
2218 unsigned long address
)
2220 struct vm_area_struct
*prev
;
2221 unsigned long gap_addr
;
2224 address
&= PAGE_MASK
;
2225 error
= security_mmap_addr(address
);
2229 /* Enforce stack_guard_gap */
2230 gap_addr
= address
- stack_guard_gap
;
2231 if (gap_addr
> address
)
2233 prev
= vma
->vm_prev
;
2234 if (prev
&& prev
->vm_end
> gap_addr
) {
2235 if (!(prev
->vm_flags
& VM_GROWSDOWN
))
2237 /* Check that both stack segments have the same anon_vma? */
2240 /* We must make sure the anon_vma is allocated. */
2241 if (unlikely(anon_vma_prepare(vma
)))
2245 * vma->vm_start/vm_end cannot change under us because the caller
2246 * is required to hold the mmap_sem in read mode. We need the
2247 * anon_vma lock to serialize against concurrent expand_stacks.
2249 vma_lock_anon_vma(vma
);
2251 /* Somebody else might have raced and expanded it already */
2252 if (address
< vma
->vm_start
) {
2253 unsigned long size
, grow
;
2255 size
= vma
->vm_end
- address
;
2256 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
2259 if (grow
<= vma
->vm_pgoff
) {
2260 error
= acct_stack_growth(vma
, size
, grow
);
2263 * vma_gap_update() doesn't support concurrent
2264 * updates, but we only hold a shared mmap_sem
2265 * lock here, so we need to protect against
2266 * concurrent vma expansions.
2267 * vma_lock_anon_vma() doesn't help here, as
2268 * we don't guarantee that all growable vmas
2269 * in a mm share the same root anon vma.
2270 * So, we reuse mm->page_table_lock to guard
2271 * against concurrent vma expansions.
2273 spin_lock(&vma
->vm_mm
->page_table_lock
);
2274 anon_vma_interval_tree_pre_update_vma(vma
);
2275 vma
->vm_start
= address
;
2276 vma
->vm_pgoff
-= grow
;
2277 anon_vma_interval_tree_post_update_vma(vma
);
2278 vma_gap_update(vma
);
2279 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2281 perf_event_mmap(vma
);
2285 vma_unlock_anon_vma(vma
);
2286 khugepaged_enter_vma_merge(vma
);
2287 validate_mm(vma
->vm_mm
);
2291 /* enforced gap between the expanding stack and other mappings. */
2292 unsigned long stack_guard_gap
= 256UL<<PAGE_SHIFT
;
2294 static int __init
cmdline_parse_stack_guard_gap(char *p
)
2299 val
= simple_strtoul(p
, &endptr
, 10);
2301 stack_guard_gap
= val
<< PAGE_SHIFT
;
2305 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap
);
2307 #ifdef CONFIG_STACK_GROWSUP
2308 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2310 return expand_upwards(vma
, address
);
2313 struct vm_area_struct
*
2314 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2316 struct vm_area_struct
*vma
, *prev
;
2319 vma
= find_vma_prev(mm
, addr
, &prev
);
2320 if (vma
&& (vma
->vm_start
<= addr
))
2322 if (!prev
|| expand_stack(prev
, addr
))
2324 if (prev
->vm_flags
& VM_LOCKED
)
2325 __mlock_vma_pages_range(prev
, addr
, prev
->vm_end
, NULL
);
2329 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2331 return expand_downwards(vma
, address
);
2334 struct vm_area_struct
*
2335 find_extend_vma(struct mm_struct
* mm
, unsigned long addr
)
2337 struct vm_area_struct
* vma
;
2338 unsigned long start
;
2341 vma
= find_vma(mm
,addr
);
2344 if (vma
->vm_start
<= addr
)
2346 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2348 start
= vma
->vm_start
;
2349 if (expand_stack(vma
, addr
))
2351 if (vma
->vm_flags
& VM_LOCKED
)
2352 __mlock_vma_pages_range(vma
, addr
, start
, NULL
);
2358 * Ok - we have the memory areas we should free on the vma list,
2359 * so release them, and do the vma updates.
2361 * Called with the mm semaphore held.
2363 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2365 unsigned long nr_accounted
= 0;
2367 /* Update high watermark before we lower total_vm */
2368 update_hiwater_vm(mm
);
2370 long nrpages
= vma_pages(vma
);
2372 if (vma
->vm_flags
& VM_ACCOUNT
)
2373 nr_accounted
+= nrpages
;
2374 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, -nrpages
);
2375 vma
= remove_vma(vma
);
2377 vm_unacct_memory(nr_accounted
);
2382 * Get rid of page table information in the indicated region.
2384 * Called with the mm semaphore held.
2386 static void unmap_region(struct mm_struct
*mm
,
2387 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
2388 unsigned long start
, unsigned long end
)
2390 struct vm_area_struct
*next
= prev
? prev
->vm_next
: mm
->mmap
;
2391 struct mmu_gather tlb
;
2394 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2395 update_hiwater_rss(mm
);
2396 unmap_vmas(&tlb
, vma
, start
, end
);
2397 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
2398 next
? next
->vm_start
: USER_PGTABLES_CEILING
);
2399 tlb_finish_mmu(&tlb
, start
, end
);
2403 * Create a list of vma's touched by the unmap, removing them from the mm's
2404 * vma list as we go..
2407 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2408 struct vm_area_struct
*prev
, unsigned long end
)
2410 struct vm_area_struct
**insertion_point
;
2411 struct vm_area_struct
*tail_vma
= NULL
;
2414 insertion_point
= (prev
? &prev
->vm_next
: &mm
->mmap
);
2415 vma
->vm_prev
= NULL
;
2417 vma_rb_erase(vma
, &mm
->mm_rb
);
2421 } while (vma
&& vma
->vm_start
< end
);
2422 *insertion_point
= vma
;
2424 vma
->vm_prev
= prev
;
2425 vma_gap_update(vma
);
2427 mm
->highest_vm_end
= prev
? vm_end_gap(prev
) : 0;
2428 tail_vma
->vm_next
= NULL
;
2429 if (mm
->unmap_area
== arch_unmap_area
)
2430 addr
= prev
? prev
->vm_end
: mm
->mmap_base
;
2432 addr
= vma
? vma
->vm_start
: mm
->mmap_base
;
2433 mm
->unmap_area(mm
, addr
);
2434 mm
->mmap_cache
= NULL
; /* Kill the cache. */
2438 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2439 * munmap path where it doesn't make sense to fail.
2441 static int __split_vma(struct mm_struct
* mm
, struct vm_area_struct
* vma
,
2442 unsigned long addr
, int new_below
)
2444 struct mempolicy
*pol
;
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 pol
= mpol_dup(vma_policy(vma
));
2473 vma_set_policy(new, pol
);
2475 if (anon_vma_clone(new, vma
))
2479 get_file(new->vm_file
);
2481 if (new->vm_ops
&& new->vm_ops
->open
)
2482 new->vm_ops
->open(new);
2485 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2486 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2488 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2494 /* Clean everything up if vma_adjust failed. */
2495 if (new->vm_ops
&& new->vm_ops
->close
)
2496 new->vm_ops
->close(new);
2499 unlink_anon_vmas(new);
2503 kmem_cache_free(vm_area_cachep
, new);
2509 * Split a vma into two pieces at address 'addr', a new vma is allocated
2510 * either for the first part or the tail.
2512 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2513 unsigned long addr
, int new_below
)
2515 if (mm
->map_count
>= sysctl_max_map_count
)
2518 return __split_vma(mm
, vma
, addr
, new_below
);
2521 /* Munmap is split into 2 main parts -- this part which finds
2522 * what needs doing, and the areas themselves, which do the
2523 * work. This now handles partial unmappings.
2524 * Jeremy Fitzhardinge <jeremy@goop.org>
2526 #ifdef CONFIG_MTK_EXTMEM
2527 extern bool extmem_in_mspace(struct vm_area_struct
*vma
);
2528 extern void * get_virt_from_mspace(void * pa
);
2529 extern size_t extmem_get_mem_size(unsigned long pgoff
);
2530 extern void extmem_free(void* mem
);
2533 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
)
2537 struct vm_area_struct
*vma
, *prev
, *last
;
2539 if ((start
& ~PAGE_MASK
) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2542 if ((len
= PAGE_ALIGN(len
)) == 0)
2545 /* Find the first overlapping VMA */
2546 vma
= find_vma(mm
, start
);
2552 const char *name
=file
->f_path
.dentry
->d_iname
;
2553 if(name
&& (strstr(name
,"app_process") || strstr(name
,"app_process64") || strstr(name
,"main") || strstr(name
,"Binder_")))
2554 printk("name:%s unmap vm_start %lx end: %lx\n", name
, vma
->vm_start
, vma
->vm_end
);
2558 const char *name
= arch_vma_name(vma
);
2559 if(name
&& (strstr(name
,"app_process") || strstr(name
,"app_process64") || strstr(name
,"main") || strstr(name
,"Binder_")))
2560 printk("name:%s unmap vm_start %lx end: %lx\n", name
, vma
->vm_start
, vma
->vm_end
);
2562 prev
= vma
->vm_prev
;
2563 /* we have start < vma->vm_end */
2565 #ifdef CONFIG_MTK_EXTMEM
2566 /* get correct mmap size if in mspace. */
2567 if (extmem_in_mspace(vma
))
2568 len
= extmem_get_mem_size(vma
->vm_pgoff
);
2571 /* if it doesn't overlap, we have nothing.. */
2573 if (vma
->vm_start
>= end
)
2577 * If we need to split any vma, do it now to save pain later.
2579 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2580 * unmapped vm_area_struct will remain in use: so lower split_vma
2581 * places tmp vma above, and higher split_vma places tmp vma below.
2583 if (start
> vma
->vm_start
) {
2587 * Make sure that map_count on return from munmap() will
2588 * not exceed its limit; but let map_count go just above
2589 * its limit temporarily, to help free resources as expected.
2591 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2594 error
= __split_vma(mm
, vma
, start
, 0);
2600 /* Does it split the last one? */
2601 last
= find_vma(mm
, end
);
2602 if (last
&& end
> last
->vm_start
) {
2603 int error
= __split_vma(mm
, last
, end
, 1);
2607 vma
= prev
? prev
->vm_next
: mm
->mmap
;
2610 * unlock any mlock()ed ranges before detaching vmas
2612 if (mm
->locked_vm
) {
2613 struct vm_area_struct
*tmp
= vma
;
2614 while (tmp
&& tmp
->vm_start
< end
) {
2615 if (tmp
->vm_flags
& VM_LOCKED
) {
2616 mm
->locked_vm
-= vma_pages(tmp
);
2617 munlock_vma_pages_all(tmp
);
2624 * Remove the vma's, and unmap the actual pages
2626 detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
);
2627 unmap_region(mm
, vma
, prev
, start
, end
);
2629 /* Fix up all other VM information */
2630 remove_vma_list(mm
, vma
);
2635 int vm_munmap(unsigned long start
, size_t len
)
2638 struct mm_struct
*mm
= current
->mm
;
2640 down_write(&mm
->mmap_sem
);
2641 ret
= do_munmap(mm
, start
, len
);
2642 up_write(&mm
->mmap_sem
);
2645 EXPORT_SYMBOL(vm_munmap
);
2647 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2649 profile_munmap(addr
);
2650 return vm_munmap(addr
, len
);
2653 static inline void verify_mm_writelocked(struct mm_struct
*mm
)
2655 #ifdef CONFIG_DEBUG_VM
2656 if (unlikely(down_read_trylock(&mm
->mmap_sem
))) {
2658 up_read(&mm
->mmap_sem
);
2664 * this is really a simplified "do_mmap". it only handles
2665 * anonymous maps. eventually we may be able to do some
2666 * brk-specific accounting here.
2668 static unsigned long do_brk(unsigned long addr
, unsigned long len
)
2670 struct mm_struct
* mm
= current
->mm
;
2671 struct vm_area_struct
* vma
, * prev
;
2672 unsigned long flags
;
2673 struct rb_node
** rb_link
, * rb_parent
;
2674 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2677 len
= PAGE_ALIGN(len
);
2681 flags
= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2683 error
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2684 if (error
& ~PAGE_MASK
)
2690 if (mm
->def_flags
& VM_LOCKED
) {
2691 unsigned long locked
, lock_limit
;
2692 locked
= len
>> PAGE_SHIFT
;
2693 locked
+= mm
->locked_vm
;
2694 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
2695 lock_limit
>>= PAGE_SHIFT
;
2696 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
2701 * mm->mmap_sem is required to protect against another thread
2702 * changing the mappings in case we sleep.
2704 verify_mm_writelocked(mm
);
2707 * Clear old maps. this also does some error checking for us
2710 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
2711 if (do_munmap(mm
, addr
, len
))
2716 /* Check against address space limits *after* clearing old maps... */
2717 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
))
2720 if (mm
->map_count
> sysctl_max_map_count
)
2723 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
2726 /* Can we just expand an old private anonymous mapping? */
2727 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
2728 NULL
, NULL
, pgoff
, NULL
, NULL
);
2733 * create a vma struct for an anonymous mapping
2735 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2737 vm_unacct_memory(len
>> PAGE_SHIFT
);
2741 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2743 vma
->vm_start
= addr
;
2744 vma
->vm_end
= addr
+ len
;
2745 vma
->vm_pgoff
= pgoff
;
2746 vma
->vm_flags
= flags
;
2747 vma
->vm_page_prot
= vm_get_page_prot(flags
);
2748 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2750 perf_event_mmap(vma
);
2751 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2752 if (flags
& VM_LOCKED
)
2753 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
2757 unsigned long vm_brk(unsigned long addr
, unsigned long len
)
2759 struct mm_struct
*mm
= current
->mm
;
2763 down_write(&mm
->mmap_sem
);
2764 ret
= do_brk(addr
, len
);
2765 populate
= ((mm
->def_flags
& VM_LOCKED
) != 0);
2766 up_write(&mm
->mmap_sem
);
2768 mm_populate(addr
, len
);
2771 EXPORT_SYMBOL(vm_brk
);
2773 /* Release all mmaps. */
2774 void exit_mmap(struct mm_struct
*mm
)
2776 struct mmu_gather tlb
;
2777 struct vm_area_struct
*vma
;
2778 unsigned long nr_accounted
= 0;
2780 /* mm's last user has gone, and its about to be pulled down */
2781 mmu_notifier_release(mm
);
2783 if (mm
->locked_vm
) {
2786 if (vma
->vm_flags
& VM_LOCKED
)
2787 munlock_vma_pages_all(vma
);
2795 if (!vma
) /* Can happen if dup_mmap() received an OOM */
2800 tlb_gather_mmu(&tlb
, mm
, 0, -1);
2801 /* update_hiwater_rss(mm) here? but nobody should be looking */
2802 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2803 unmap_vmas(&tlb
, vma
, 0, -1);
2805 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, USER_PGTABLES_CEILING
);
2806 tlb_finish_mmu(&tlb
, 0, -1);
2809 * Walk the list again, actually closing and freeing it,
2810 * with preemption enabled, without holding any MM locks.
2813 if (vma
->vm_flags
& VM_ACCOUNT
)
2814 nr_accounted
+= vma_pages(vma
);
2815 vma
= remove_vma(vma
);
2817 vm_unacct_memory(nr_accounted
);
2819 WARN_ON(mm
->nr_ptes
> (FIRST_USER_ADDRESS
+PMD_SIZE
-1)>>PMD_SHIFT
);
2822 /* Insert vm structure into process list sorted by address
2823 * and into the inode's i_mmap tree. If vm_file is non-NULL
2824 * then i_mmap_mutex is taken here.
2826 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2828 struct vm_area_struct
*prev
;
2829 struct rb_node
**rb_link
, *rb_parent
;
2832 * The vm_pgoff of a purely anonymous vma should be irrelevant
2833 * until its first write fault, when page's anon_vma and index
2834 * are set. But now set the vm_pgoff it will almost certainly
2835 * end up with (unless mremap moves it elsewhere before that
2836 * first wfault), so /proc/pid/maps tells a consistent story.
2838 * By setting it to reflect the virtual start address of the
2839 * vma, merges and splits can happen in a seamless way, just
2840 * using the existing file pgoff checks and manipulations.
2841 * Similarly in do_mmap_pgoff and in do_brk.
2843 if (!vma
->vm_file
) {
2844 BUG_ON(vma
->anon_vma
);
2845 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2847 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
2848 &prev
, &rb_link
, &rb_parent
))
2850 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
2851 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
2854 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2859 * Copy the vma structure to a new location in the same mm,
2860 * prior to moving page table entries, to effect an mremap move.
2862 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
2863 unsigned long addr
, unsigned long len
, pgoff_t pgoff
,
2864 bool *need_rmap_locks
)
2866 struct vm_area_struct
*vma
= *vmap
;
2867 unsigned long vma_start
= vma
->vm_start
;
2868 struct mm_struct
*mm
= vma
->vm_mm
;
2869 struct vm_area_struct
*new_vma
, *prev
;
2870 struct rb_node
**rb_link
, *rb_parent
;
2871 struct mempolicy
*pol
;
2872 bool faulted_in_anon_vma
= true;
2875 * If anonymous vma has not yet been faulted, update new pgoff
2876 * to match new location, to increase its chance of merging.
2878 if (unlikely(!vma
->vm_file
&& !vma
->anon_vma
)) {
2879 pgoff
= addr
>> PAGE_SHIFT
;
2880 faulted_in_anon_vma
= false;
2883 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
2884 return NULL
; /* should never get here */
2885 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
2886 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
),
2887 vma_get_anon_name(vma
));
2890 * Source vma may have been merged into new_vma
2892 if (unlikely(vma_start
>= new_vma
->vm_start
&&
2893 vma_start
< new_vma
->vm_end
)) {
2895 * The only way we can get a vma_merge with
2896 * self during an mremap is if the vma hasn't
2897 * been faulted in yet and we were allowed to
2898 * reset the dst vma->vm_pgoff to the
2899 * destination address of the mremap to allow
2900 * the merge to happen. mremap must change the
2901 * vm_pgoff linearity between src and dst vmas
2902 * (in turn preventing a vma_merge) to be
2903 * safe. It is only safe to keep the vm_pgoff
2904 * linear if there are no pages mapped yet.
2906 VM_BUG_ON(faulted_in_anon_vma
);
2907 *vmap
= vma
= new_vma
;
2909 *need_rmap_locks
= (new_vma
->vm_pgoff
<= vma
->vm_pgoff
);
2911 new_vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2914 new_vma
->vm_start
= addr
;
2915 new_vma
->vm_end
= addr
+ len
;
2916 new_vma
->vm_pgoff
= pgoff
;
2917 pol
= mpol_dup(vma_policy(vma
));
2920 vma_set_policy(new_vma
, pol
);
2921 INIT_LIST_HEAD(&new_vma
->anon_vma_chain
);
2922 if (anon_vma_clone(new_vma
, vma
))
2923 goto out_free_mempol
;
2924 if (new_vma
->vm_file
)
2925 get_file(new_vma
->vm_file
);
2926 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
2927 new_vma
->vm_ops
->open(new_vma
);
2928 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
2929 *need_rmap_locks
= false;
2937 kmem_cache_free(vm_area_cachep
, new_vma
);
2942 * Return true if the calling process may expand its vm space by the passed
2945 int may_expand_vm(struct mm_struct
*mm
, unsigned long npages
)
2947 unsigned long cur
= mm
->total_vm
; /* pages */
2950 lim
= rlimit(RLIMIT_AS
) >> PAGE_SHIFT
;
2952 if (cur
+ npages
> lim
)
2958 static int special_mapping_fault(struct vm_area_struct
*vma
,
2959 struct vm_fault
*vmf
)
2962 struct page
**pages
;
2965 * special mappings have no vm_file, and in that case, the mm
2966 * uses vm_pgoff internally. So we have to subtract it from here.
2967 * We are allowed to do this because we are the mm; do not copy
2968 * this code into drivers!
2970 pgoff
= vmf
->pgoff
- vma
->vm_pgoff
;
2972 for (pages
= vma
->vm_private_data
; pgoff
&& *pages
; ++pages
)
2976 struct page
*page
= *pages
;
2982 return VM_FAULT_SIGBUS
;
2986 * Having a close hook prevents vma merging regardless of flags.
2988 static void special_mapping_close(struct vm_area_struct
*vma
)
2992 static const struct vm_operations_struct special_mapping_vmops
= {
2993 .close
= special_mapping_close
,
2994 .fault
= special_mapping_fault
,
2998 * Called with mm->mmap_sem held for writing.
2999 * Insert a new vma covering the given region, with the given flags.
3000 * Its pages are supplied by the given array of struct page *.
3001 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3002 * The region past the last page supplied will always produce SIGBUS.
3003 * The array pointer and the pages it points to are assumed to stay alive
3004 * for as long as this mapping might exist.
3006 int install_special_mapping(struct mm_struct
*mm
,
3007 unsigned long addr
, unsigned long len
,
3008 unsigned long vm_flags
, struct page
**pages
)
3011 struct vm_area_struct
*vma
;
3013 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
3014 if (unlikely(vma
== NULL
))
3017 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
3019 vma
->vm_start
= addr
;
3020 vma
->vm_end
= addr
+ len
;
3022 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
;
3023 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
3025 vma
->vm_ops
= &special_mapping_vmops
;
3026 vma
->vm_private_data
= pages
;
3028 ret
= insert_vm_struct(mm
, vma
);
3032 mm
->total_vm
+= len
>> PAGE_SHIFT
;
3034 perf_event_mmap(vma
);
3039 kmem_cache_free(vm_area_cachep
, vma
);
3043 static DEFINE_MUTEX(mm_all_locks_mutex
);
3045 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
3047 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3049 * The LSB of head.next can't change from under us
3050 * because we hold the mm_all_locks_mutex.
3052 down_write_nest_lock(&anon_vma
->root
->rwsem
, &mm
->mmap_sem
);
3054 * We can safely modify head.next after taking the
3055 * anon_vma->root->rwsem. If some other vma in this mm shares
3056 * the same anon_vma we won't take it again.
3058 * No need of atomic instructions here, head.next
3059 * can't change from under us thanks to the
3060 * anon_vma->root->rwsem.
3062 if (__test_and_set_bit(0, (unsigned long *)
3063 &anon_vma
->root
->rb_root
.rb_node
))
3068 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
3070 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3072 * AS_MM_ALL_LOCKS can't change from under us because
3073 * we hold the mm_all_locks_mutex.
3075 * Operations on ->flags have to be atomic because
3076 * even if AS_MM_ALL_LOCKS is stable thanks to the
3077 * mm_all_locks_mutex, there may be other cpus
3078 * changing other bitflags in parallel to us.
3080 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
3082 mutex_lock_nest_lock(&mapping
->i_mmap_mutex
, &mm
->mmap_sem
);
3087 * This operation locks against the VM for all pte/vma/mm related
3088 * operations that could ever happen on a certain mm. This includes
3089 * vmtruncate, try_to_unmap, and all page faults.
3091 * The caller must take the mmap_sem in write mode before calling
3092 * mm_take_all_locks(). The caller isn't allowed to release the
3093 * mmap_sem until mm_drop_all_locks() returns.
3095 * mmap_sem in write mode is required in order to block all operations
3096 * that could modify pagetables and free pages without need of
3097 * altering the vma layout (for example populate_range() with
3098 * nonlinear vmas). It's also needed in write mode to avoid new
3099 * anon_vmas to be associated with existing vmas.
3101 * A single task can't take more than one mm_take_all_locks() in a row
3102 * or it would deadlock.
3104 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3105 * mapping->flags avoid to take the same lock twice, if more than one
3106 * vma in this mm is backed by the same anon_vma or address_space.
3108 * We can take all the locks in random order because the VM code
3109 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3110 * takes more than one of them in a row. Secondly we're protected
3111 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3113 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3114 * that may have to take thousand of locks.
3116 * mm_take_all_locks() can fail if it's interrupted by signals.
3118 int mm_take_all_locks(struct mm_struct
*mm
)
3120 struct vm_area_struct
*vma
;
3121 struct anon_vma_chain
*avc
;
3123 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3125 mutex_lock(&mm_all_locks_mutex
);
3127 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3128 if (signal_pending(current
))
3130 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3131 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3134 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3135 if (signal_pending(current
))
3138 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3139 vm_lock_anon_vma(mm
, avc
->anon_vma
);
3145 mm_drop_all_locks(mm
);
3149 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
3151 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3153 * The LSB of head.next can't change to 0 from under
3154 * us because we hold the mm_all_locks_mutex.
3156 * We must however clear the bitflag before unlocking
3157 * the vma so the users using the anon_vma->rb_root will
3158 * never see our bitflag.
3160 * No need of atomic instructions here, head.next
3161 * can't change from under us until we release the
3162 * anon_vma->root->rwsem.
3164 if (!__test_and_clear_bit(0, (unsigned long *)
3165 &anon_vma
->root
->rb_root
.rb_node
))
3167 anon_vma_unlock_write(anon_vma
);
3171 static void vm_unlock_mapping(struct address_space
*mapping
)
3173 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3175 * AS_MM_ALL_LOCKS can't change to 0 from under us
3176 * because we hold the mm_all_locks_mutex.
3178 mutex_unlock(&mapping
->i_mmap_mutex
);
3179 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
3186 * The mmap_sem cannot be released by the caller until
3187 * mm_drop_all_locks() returns.
3189 void mm_drop_all_locks(struct mm_struct
*mm
)
3191 struct vm_area_struct
*vma
;
3192 struct anon_vma_chain
*avc
;
3194 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3195 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
3197 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3199 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3200 vm_unlock_anon_vma(avc
->anon_vma
);
3201 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3202 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
3205 mutex_unlock(&mm_all_locks_mutex
);
3209 * initialise the VMA slab
3211 void __init
mmap_init(void)
3215 ret
= percpu_counter_init(&vm_committed_as
, 0);
3220 * Initialise sysctl_user_reserve_kbytes.
3222 * This is intended to prevent a user from starting a single memory hogging
3223 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3226 * The default value is min(3% of free memory, 128MB)
3227 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3229 static int init_user_reserve(void)
3231 unsigned long free_kbytes
;
3233 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3235 sysctl_user_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 17);
3238 module_init(init_user_reserve
)
3241 * Initialise sysctl_admin_reserve_kbytes.
3243 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3244 * to log in and kill a memory hogging process.
3246 * Systems with more than 256MB will reserve 8MB, enough to recover
3247 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3248 * only reserve 3% of free pages by default.
3250 static int init_admin_reserve(void)
3252 unsigned long free_kbytes
;
3254 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3256 sysctl_admin_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 13);
3259 module_init(init_admin_reserve
)
3262 * Reinititalise user and admin reserves if memory is added or removed.
3264 * The default user reserve max is 128MB, and the default max for the
3265 * admin reserve is 8MB. These are usually, but not always, enough to
3266 * enable recovery from a memory hogging process using login/sshd, a shell,
3267 * and tools like top. It may make sense to increase or even disable the
3268 * reserve depending on the existence of swap or variations in the recovery
3269 * tools. So, the admin may have changed them.
3271 * If memory is added and the reserves have been eliminated or increased above
3272 * the default max, then we'll trust the admin.
3274 * If memory is removed and there isn't enough free memory, then we
3275 * need to reset the reserves.
3277 * Otherwise keep the reserve set by the admin.
3279 static int reserve_mem_notifier(struct notifier_block
*nb
,
3280 unsigned long action
, void *data
)
3282 unsigned long tmp
, free_kbytes
;
3286 /* Default max is 128MB. Leave alone if modified by operator. */
3287 tmp
= sysctl_user_reserve_kbytes
;
3288 if (0 < tmp
&& tmp
< (1UL << 17))
3289 init_user_reserve();
3291 /* Default max is 8MB. Leave alone if modified by operator. */
3292 tmp
= sysctl_admin_reserve_kbytes
;
3293 if (0 < tmp
&& tmp
< (1UL << 13))
3294 init_admin_reserve();
3298 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3300 if (sysctl_user_reserve_kbytes
> free_kbytes
) {
3301 init_user_reserve();
3302 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3303 sysctl_user_reserve_kbytes
);
3306 if (sysctl_admin_reserve_kbytes
> free_kbytes
) {
3307 init_admin_reserve();
3308 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3309 sysctl_admin_reserve_kbytes
);
3318 static struct notifier_block reserve_mem_nb
= {
3319 .notifier_call
= reserve_mem_notifier
,
3322 static int __meminit
init_reserve_notifier(void)
3324 if (register_hotmemory_notifier(&reserve_mem_nb
))
3325 printk("Failed registering memory add/remove notifier for admin reserve");
3329 module_init(init_reserve_notifier
)