6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/profile.h>
29 #include <linux/export.h>
30 #include <linux/mount.h>
31 #include <linux/mempolicy.h>
32 #include <linux/rmap.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/mmdebug.h>
35 #include <linux/perf_event.h>
36 #include <linux/audit.h>
37 #include <linux/khugepaged.h>
38 #include <linux/uprobes.h>
39 #include <linux/rbtree_augmented.h>
40 #include <linux/sched/sysctl.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
45 #include <asm/uaccess.h>
46 #include <asm/cacheflush.h>
48 #include <asm/mmu_context.h>
52 #ifndef arch_mmap_check
53 #define arch_mmap_check(addr, len, flags) (0)
56 #ifndef arch_rebalance_pgtables
57 #define arch_rebalance_pgtables(addr, len) (addr)
60 static void unmap_region(struct mm_struct
*mm
,
61 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
62 unsigned long start
, unsigned long end
);
64 /* description of effects of mapping type and prot in current implementation.
65 * this is due to the limited x86 page protection hardware. The expected
66 * behavior is in parens:
69 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
70 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
71 * w: (no) no w: (no) no w: (yes) yes w: (no) no
72 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
74 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
75 * w: (no) no w: (no) no w: (copy) copy w: (no) no
76 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
79 pgprot_t protection_map
[16] = {
80 __P000
, __P001
, __P010
, __P011
, __P100
, __P101
, __P110
, __P111
,
81 __S000
, __S001
, __S010
, __S011
, __S100
, __S101
, __S110
, __S111
84 pgprot_t
vm_get_page_prot(unsigned long vm_flags
)
86 return __pgprot(pgprot_val(protection_map
[vm_flags
&
87 (VM_READ
|VM_WRITE
|VM_EXEC
|VM_SHARED
)]) |
88 pgprot_val(arch_vm_get_page_prot(vm_flags
)));
90 EXPORT_SYMBOL(vm_get_page_prot
);
92 int sysctl_overcommit_memory __read_mostly
= OVERCOMMIT_GUESS
; /* heuristic overcommit */
93 int sysctl_overcommit_ratio __read_mostly
= 50; /* default is 50% */
94 unsigned long sysctl_overcommit_kbytes __read_mostly
;
95 int sysctl_max_map_count __read_mostly
= DEFAULT_MAX_MAP_COUNT
;
96 unsigned long sysctl_user_reserve_kbytes __read_mostly
= 1UL << 17; /* 128MB */
97 unsigned long sysctl_admin_reserve_kbytes __read_mostly
= 1UL << 13; /* 8MB */
99 * Make sure vm_committed_as in one cacheline and not cacheline shared with
100 * other variables. It can be updated by several CPUs frequently.
102 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp
;
105 * The global memory commitment made in the system can be a metric
106 * that can be used to drive ballooning decisions when Linux is hosted
107 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
108 * balancing memory across competing virtual machines that are hosted.
109 * Several metrics drive this policy engine including the guest reported
112 unsigned long vm_memory_committed(void)
114 return percpu_counter_read_positive(&vm_committed_as
);
116 EXPORT_SYMBOL_GPL(vm_memory_committed
);
119 * Check that a process has enough memory to allocate a new virtual
120 * mapping. 0 means there is enough memory for the allocation to
121 * succeed and -ENOMEM implies there is not.
123 * We currently support three overcommit policies, which are set via the
124 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
126 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
127 * Additional code 2002 Jul 20 by Robert Love.
129 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
131 * Note this is a helper function intended to be used by LSMs which
132 * wish to use this logic.
134 int __vm_enough_memory(struct mm_struct
*mm
, long pages
, int cap_sys_admin
)
136 unsigned long free
, allowed
, reserve
;
138 VM_WARN_ONCE(percpu_counter_read(&vm_committed_as
) <
139 -(s64
)vm_committed_as_batch
* num_online_cpus(),
140 "memory commitment underflow");
142 vm_acct_memory(pages
);
145 * Sometimes we want to use more memory than we have
147 if (sysctl_overcommit_memory
== OVERCOMMIT_ALWAYS
)
150 if (sysctl_overcommit_memory
== OVERCOMMIT_GUESS
) {
151 free
= global_page_state(NR_FREE_PAGES
);
152 free
+= global_page_state(NR_FILE_PAGES
);
155 * shmem pages shouldn't be counted as free in this
156 * case, they can't be purged, only swapped out, and
157 * that won't affect the overall amount of available
158 * memory in the system.
160 free
-= global_page_state(NR_SHMEM
);
162 free
+= get_nr_swap_pages();
165 * Any slabs which are created with the
166 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
167 * which are reclaimable, under pressure. The dentry
168 * cache and most inode caches should fall into this
170 free
+= global_page_state(NR_SLAB_RECLAIMABLE
);
173 * Leave reserved pages. The pages are not for anonymous pages.
175 if (free
<= totalreserve_pages
)
178 free
-= totalreserve_pages
;
181 * Reserve some for root
184 free
-= sysctl_admin_reserve_kbytes
>> (PAGE_SHIFT
- 10);
192 allowed
= vm_commit_limit();
194 * Reserve some for root
197 allowed
-= sysctl_admin_reserve_kbytes
>> (PAGE_SHIFT
- 10);
200 * Don't let a single process grow so big a user can't recover
203 reserve
= sysctl_user_reserve_kbytes
>> (PAGE_SHIFT
- 10);
204 allowed
-= min(mm
->total_vm
/ 32, reserve
);
207 if (percpu_counter_read_positive(&vm_committed_as
) < allowed
)
210 vm_unacct_memory(pages
);
216 * Requires inode->i_mapping->i_mmap_mutex
218 static void __remove_shared_vm_struct(struct vm_area_struct
*vma
,
219 struct file
*file
, struct address_space
*mapping
)
221 if (vma
->vm_flags
& VM_DENYWRITE
)
222 atomic_inc(&file_inode(file
)->i_writecount
);
223 if (vma
->vm_flags
& VM_SHARED
)
224 mapping_unmap_writable(mapping
);
226 flush_dcache_mmap_lock(mapping
);
227 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
228 list_del_init(&vma
->shared
.nonlinear
);
230 vma_interval_tree_remove(vma
, &mapping
->i_mmap
);
231 flush_dcache_mmap_unlock(mapping
);
235 * Unlink a file-based vm structure from its interval tree, to hide
236 * vma from rmap and vmtruncate before freeing its page tables.
238 void unlink_file_vma(struct vm_area_struct
*vma
)
240 struct file
*file
= vma
->vm_file
;
243 struct address_space
*mapping
= file
->f_mapping
;
244 mutex_lock(&mapping
->i_mmap_mutex
);
245 __remove_shared_vm_struct(vma
, file
, mapping
);
246 mutex_unlock(&mapping
->i_mmap_mutex
);
251 * Close a vm structure and free it, returning the next.
253 static struct vm_area_struct
*remove_vma(struct vm_area_struct
*vma
)
255 struct vm_area_struct
*next
= vma
->vm_next
;
258 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
259 vma
->vm_ops
->close(vma
);
262 mpol_put(vma_policy(vma
));
263 kmem_cache_free(vm_area_cachep
, vma
);
267 static unsigned long do_brk(unsigned long addr
, unsigned long len
);
269 SYSCALL_DEFINE1(brk
, unsigned long, brk
)
271 unsigned long retval
;
272 unsigned long newbrk
, oldbrk
;
273 struct mm_struct
*mm
= current
->mm
;
274 unsigned long min_brk
;
277 down_write(&mm
->mmap_sem
);
279 #ifdef CONFIG_COMPAT_BRK
281 * CONFIG_COMPAT_BRK can still be overridden by setting
282 * randomize_va_space to 2, which will still cause mm->start_brk
283 * to be arbitrarily shifted
285 if (current
->brk_randomized
)
286 min_brk
= mm
->start_brk
;
288 min_brk
= mm
->end_data
;
290 min_brk
= mm
->start_brk
;
296 * Check against rlimit here. If this check is done later after the test
297 * of oldbrk with newbrk then it can escape the test and let the data
298 * segment grow beyond its set limit the in case where the limit is
299 * not page aligned -Ram Gupta
301 if (check_data_rlimit(rlimit(RLIMIT_DATA
), brk
, mm
->start_brk
,
302 mm
->end_data
, mm
->start_data
))
305 newbrk
= PAGE_ALIGN(brk
);
306 oldbrk
= PAGE_ALIGN(mm
->brk
);
307 if (oldbrk
== newbrk
)
310 /* Always allow shrinking brk. */
311 if (brk
<= mm
->brk
) {
312 if (!do_munmap(mm
, newbrk
, oldbrk
-newbrk
))
317 /* Check against existing mmap mappings. */
318 if (find_vma_intersection(mm
, oldbrk
, newbrk
+PAGE_SIZE
))
321 /* Ok, looks good - let it rip. */
322 if (do_brk(oldbrk
, newbrk
-oldbrk
) != oldbrk
)
327 populate
= newbrk
> oldbrk
&& (mm
->def_flags
& VM_LOCKED
) != 0;
328 up_write(&mm
->mmap_sem
);
330 mm_populate(oldbrk
, newbrk
- oldbrk
);
335 up_write(&mm
->mmap_sem
);
339 static long vma_compute_subtree_gap(struct vm_area_struct
*vma
)
341 unsigned long max
, subtree_gap
;
344 max
-= vma
->vm_prev
->vm_end
;
345 if (vma
->vm_rb
.rb_left
) {
346 subtree_gap
= rb_entry(vma
->vm_rb
.rb_left
,
347 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
348 if (subtree_gap
> max
)
351 if (vma
->vm_rb
.rb_right
) {
352 subtree_gap
= rb_entry(vma
->vm_rb
.rb_right
,
353 struct vm_area_struct
, vm_rb
)->rb_subtree_gap
;
354 if (subtree_gap
> max
)
360 #ifdef CONFIG_DEBUG_VM_RB
361 static int browse_rb(struct rb_root
*root
)
363 int i
= 0, j
, bug
= 0;
364 struct rb_node
*nd
, *pn
= NULL
;
365 unsigned long prev
= 0, pend
= 0;
367 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
368 struct vm_area_struct
*vma
;
369 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
370 if (vma
->vm_start
< prev
) {
371 pr_emerg("vm_start %lx < prev %lx\n",
372 vma
->vm_start
, prev
);
375 if (vma
->vm_start
< pend
) {
376 pr_emerg("vm_start %lx < pend %lx\n",
377 vma
->vm_start
, pend
);
380 if (vma
->vm_start
> vma
->vm_end
) {
381 pr_emerg("vm_start %lx > vm_end %lx\n",
382 vma
->vm_start
, vma
->vm_end
);
385 if (vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
)) {
386 pr_emerg("free gap %lx, correct %lx\n",
388 vma_compute_subtree_gap(vma
));
393 prev
= vma
->vm_start
;
397 for (nd
= pn
; nd
; nd
= rb_prev(nd
))
400 pr_emerg("backwards %d, forwards %d\n", j
, i
);
406 static void validate_mm_rb(struct rb_root
*root
, struct vm_area_struct
*ignore
)
410 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
411 struct vm_area_struct
*vma
;
412 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
413 VM_BUG_ON_VMA(vma
!= ignore
&&
414 vma
->rb_subtree_gap
!= vma_compute_subtree_gap(vma
),
419 static void validate_mm(struct mm_struct
*mm
)
423 unsigned long highest_address
= 0;
424 struct vm_area_struct
*vma
= mm
->mmap
;
427 struct anon_vma_chain
*avc
;
429 vma_lock_anon_vma(vma
);
430 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
431 anon_vma_interval_tree_verify(avc
);
432 vma_unlock_anon_vma(vma
);
433 highest_address
= vma
->vm_end
;
437 if (i
!= mm
->map_count
) {
438 pr_emerg("map_count %d vm_next %d\n", mm
->map_count
, i
);
441 if (highest_address
!= mm
->highest_vm_end
) {
442 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
443 mm
->highest_vm_end
, highest_address
);
446 i
= browse_rb(&mm
->mm_rb
);
447 if (i
!= mm
->map_count
) {
449 pr_emerg("map_count %d rb %d\n", mm
->map_count
, i
);
452 VM_BUG_ON_MM(bug
, mm
);
455 #define validate_mm_rb(root, ignore) do { } while (0)
456 #define validate_mm(mm) do { } while (0)
459 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks
, struct vm_area_struct
, vm_rb
,
460 unsigned long, rb_subtree_gap
, vma_compute_subtree_gap
)
463 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
464 * vma->vm_prev->vm_end values changed, without modifying the vma's position
467 static void vma_gap_update(struct vm_area_struct
*vma
)
470 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
471 * function that does exacltly what we want.
473 vma_gap_callbacks_propagate(&vma
->vm_rb
, NULL
);
476 static inline void vma_rb_insert(struct vm_area_struct
*vma
,
477 struct rb_root
*root
)
479 /* All rb_subtree_gap values must be consistent prior to insertion */
480 validate_mm_rb(root
, NULL
);
482 rb_insert_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
485 static void vma_rb_erase(struct vm_area_struct
*vma
, struct rb_root
*root
)
488 * All rb_subtree_gap values must be consistent prior to erase,
489 * with the possible exception of the vma being erased.
491 validate_mm_rb(root
, vma
);
494 * Note rb_erase_augmented is a fairly large inline function,
495 * so make sure we instantiate it only once with our desired
496 * augmented rbtree callbacks.
498 rb_erase_augmented(&vma
->vm_rb
, root
, &vma_gap_callbacks
);
502 * vma has some anon_vma assigned, and is already inserted on that
503 * anon_vma's interval trees.
505 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
506 * vma must be removed from the anon_vma's interval trees using
507 * anon_vma_interval_tree_pre_update_vma().
509 * After the update, the vma will be reinserted using
510 * anon_vma_interval_tree_post_update_vma().
512 * The entire update must be protected by exclusive mmap_sem and by
513 * the root anon_vma's mutex.
516 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct
*vma
)
518 struct anon_vma_chain
*avc
;
520 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
521 anon_vma_interval_tree_remove(avc
, &avc
->anon_vma
->rb_root
);
525 anon_vma_interval_tree_post_update_vma(struct vm_area_struct
*vma
)
527 struct anon_vma_chain
*avc
;
529 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
530 anon_vma_interval_tree_insert(avc
, &avc
->anon_vma
->rb_root
);
533 static int find_vma_links(struct mm_struct
*mm
, unsigned long addr
,
534 unsigned long end
, struct vm_area_struct
**pprev
,
535 struct rb_node
***rb_link
, struct rb_node
**rb_parent
)
537 struct rb_node
**__rb_link
, *__rb_parent
, *rb_prev
;
539 __rb_link
= &mm
->mm_rb
.rb_node
;
540 rb_prev
= __rb_parent
= NULL
;
543 struct vm_area_struct
*vma_tmp
;
545 __rb_parent
= *__rb_link
;
546 vma_tmp
= rb_entry(__rb_parent
, struct vm_area_struct
, vm_rb
);
548 if (vma_tmp
->vm_end
> addr
) {
549 /* Fail if an existing vma overlaps the area */
550 if (vma_tmp
->vm_start
< end
)
552 __rb_link
= &__rb_parent
->rb_left
;
554 rb_prev
= __rb_parent
;
555 __rb_link
= &__rb_parent
->rb_right
;
561 *pprev
= rb_entry(rb_prev
, struct vm_area_struct
, vm_rb
);
562 *rb_link
= __rb_link
;
563 *rb_parent
= __rb_parent
;
567 static unsigned long count_vma_pages_range(struct mm_struct
*mm
,
568 unsigned long addr
, unsigned long end
)
570 unsigned long nr_pages
= 0;
571 struct vm_area_struct
*vma
;
573 /* Find first overlaping mapping */
574 vma
= find_vma_intersection(mm
, addr
, end
);
578 nr_pages
= (min(end
, vma
->vm_end
) -
579 max(addr
, vma
->vm_start
)) >> PAGE_SHIFT
;
581 /* Iterate over the rest of the overlaps */
582 for (vma
= vma
->vm_next
; vma
; vma
= vma
->vm_next
) {
583 unsigned long overlap_len
;
585 if (vma
->vm_start
> end
)
588 overlap_len
= min(end
, vma
->vm_end
) - vma
->vm_start
;
589 nr_pages
+= overlap_len
>> PAGE_SHIFT
;
595 void __vma_link_rb(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
596 struct rb_node
**rb_link
, struct rb_node
*rb_parent
)
598 /* Update tracking information for the gap following the new vma. */
600 vma_gap_update(vma
->vm_next
);
602 mm
->highest_vm_end
= vma
->vm_end
;
605 * vma->vm_prev wasn't known when we followed the rbtree to find the
606 * correct insertion point for that vma. As a result, we could not
607 * update the vma vm_rb parents rb_subtree_gap values on the way down.
608 * So, we first insert the vma with a zero rb_subtree_gap value
609 * (to be consistent with what we did on the way down), and then
610 * immediately update the gap to the correct value. Finally we
611 * rebalance the rbtree after all augmented values have been set.
613 rb_link_node(&vma
->vm_rb
, rb_parent
, rb_link
);
614 vma
->rb_subtree_gap
= 0;
616 vma_rb_insert(vma
, &mm
->mm_rb
);
619 static void __vma_link_file(struct vm_area_struct
*vma
)
625 struct address_space
*mapping
= file
->f_mapping
;
627 if (vma
->vm_flags
& VM_DENYWRITE
)
628 atomic_dec(&file_inode(file
)->i_writecount
);
629 if (vma
->vm_flags
& VM_SHARED
)
630 atomic_inc(&mapping
->i_mmap_writable
);
632 flush_dcache_mmap_lock(mapping
);
633 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
634 vma_nonlinear_insert(vma
, &mapping
->i_mmap_nonlinear
);
636 vma_interval_tree_insert(vma
, &mapping
->i_mmap
);
637 flush_dcache_mmap_unlock(mapping
);
642 __vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
643 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
644 struct rb_node
*rb_parent
)
646 __vma_link_list(mm
, vma
, prev
, rb_parent
);
647 __vma_link_rb(mm
, vma
, rb_link
, rb_parent
);
650 static void vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
651 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
652 struct rb_node
*rb_parent
)
654 struct address_space
*mapping
= NULL
;
657 mapping
= vma
->vm_file
->f_mapping
;
658 mutex_lock(&mapping
->i_mmap_mutex
);
661 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
662 __vma_link_file(vma
);
665 mutex_unlock(&mapping
->i_mmap_mutex
);
672 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
673 * mm's list and rbtree. It has already been inserted into the interval tree.
675 static void __insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
677 struct vm_area_struct
*prev
;
678 struct rb_node
**rb_link
, *rb_parent
;
680 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
681 &prev
, &rb_link
, &rb_parent
))
683 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
688 __vma_unlink(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
689 struct vm_area_struct
*prev
)
691 struct vm_area_struct
*next
;
693 vma_rb_erase(vma
, &mm
->mm_rb
);
694 prev
->vm_next
= next
= vma
->vm_next
;
696 next
->vm_prev
= prev
;
699 vmacache_invalidate(mm
);
703 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
704 * is already present in an i_mmap tree without adjusting the tree.
705 * The following helper function should be used when such adjustments
706 * are necessary. The "insert" vma (if any) is to be inserted
707 * before we drop the necessary locks.
709 int vma_adjust(struct vm_area_struct
*vma
, unsigned long start
,
710 unsigned long end
, pgoff_t pgoff
, struct vm_area_struct
*insert
)
712 struct mm_struct
*mm
= vma
->vm_mm
;
713 struct vm_area_struct
*next
= vma
->vm_next
;
714 struct vm_area_struct
*importer
= NULL
;
715 struct address_space
*mapping
= NULL
;
716 struct rb_root
*root
= NULL
;
717 struct anon_vma
*anon_vma
= NULL
;
718 struct file
*file
= vma
->vm_file
;
719 bool start_changed
= false, end_changed
= false;
720 long adjust_next
= 0;
723 if (next
&& !insert
) {
724 struct vm_area_struct
*exporter
= NULL
;
726 if (end
>= next
->vm_end
) {
728 * vma expands, overlapping all the next, and
729 * perhaps the one after too (mprotect case 6).
731 again
: remove_next
= 1 + (end
> next
->vm_end
);
735 } else if (end
> next
->vm_start
) {
737 * vma expands, overlapping part of the next:
738 * mprotect case 5 shifting the boundary up.
740 adjust_next
= (end
- next
->vm_start
) >> PAGE_SHIFT
;
743 } else if (end
< vma
->vm_end
) {
745 * vma shrinks, and !insert tells it's not
746 * split_vma inserting another: so it must be
747 * mprotect case 4 shifting the boundary down.
749 adjust_next
= -((vma
->vm_end
- end
) >> PAGE_SHIFT
);
755 * Easily overlooked: when mprotect shifts the boundary,
756 * make sure the expanding vma has anon_vma set if the
757 * shrinking vma had, to cover any anon pages imported.
759 if (exporter
&& exporter
->anon_vma
&& !importer
->anon_vma
) {
760 if (anon_vma_clone(importer
, exporter
))
762 importer
->anon_vma
= exporter
->anon_vma
;
767 mapping
= file
->f_mapping
;
768 if (!(vma
->vm_flags
& VM_NONLINEAR
)) {
769 root
= &mapping
->i_mmap
;
770 uprobe_munmap(vma
, vma
->vm_start
, vma
->vm_end
);
773 uprobe_munmap(next
, next
->vm_start
,
777 mutex_lock(&mapping
->i_mmap_mutex
);
780 * Put into interval tree now, so instantiated pages
781 * are visible to arm/parisc __flush_dcache_page
782 * throughout; but we cannot insert into address
783 * space until vma start or end is updated.
785 __vma_link_file(insert
);
789 vma_adjust_trans_huge(vma
, start
, end
, adjust_next
);
791 anon_vma
= vma
->anon_vma
;
792 if (!anon_vma
&& adjust_next
)
793 anon_vma
= next
->anon_vma
;
795 VM_BUG_ON_VMA(adjust_next
&& next
->anon_vma
&&
796 anon_vma
!= next
->anon_vma
, next
);
797 anon_vma_lock_write(anon_vma
);
798 anon_vma_interval_tree_pre_update_vma(vma
);
800 anon_vma_interval_tree_pre_update_vma(next
);
804 flush_dcache_mmap_lock(mapping
);
805 vma_interval_tree_remove(vma
, root
);
807 vma_interval_tree_remove(next
, root
);
810 if (start
!= vma
->vm_start
) {
811 vma
->vm_start
= start
;
812 start_changed
= true;
814 if (end
!= vma
->vm_end
) {
818 vma
->vm_pgoff
= pgoff
;
820 next
->vm_start
+= adjust_next
<< PAGE_SHIFT
;
821 next
->vm_pgoff
+= adjust_next
;
826 vma_interval_tree_insert(next
, root
);
827 vma_interval_tree_insert(vma
, root
);
828 flush_dcache_mmap_unlock(mapping
);
833 * vma_merge has merged next into vma, and needs
834 * us to remove next before dropping the locks.
836 __vma_unlink(mm
, next
, vma
);
838 __remove_shared_vm_struct(next
, file
, mapping
);
841 * split_vma has split insert from vma, and needs
842 * us to insert it before dropping the locks
843 * (it may either follow vma or precede it).
845 __insert_vm_struct(mm
, insert
);
851 mm
->highest_vm_end
= end
;
852 else if (!adjust_next
)
853 vma_gap_update(next
);
858 anon_vma_interval_tree_post_update_vma(vma
);
860 anon_vma_interval_tree_post_update_vma(next
);
861 anon_vma_unlock_write(anon_vma
);
864 mutex_unlock(&mapping
->i_mmap_mutex
);
875 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
879 anon_vma_merge(vma
, next
);
881 mpol_put(vma_policy(next
));
882 kmem_cache_free(vm_area_cachep
, next
);
884 * In mprotect's case 6 (see comments on vma_merge),
885 * we must remove another next too. It would clutter
886 * up the code too much to do both in one go.
889 if (remove_next
== 2)
892 vma_gap_update(next
);
894 mm
->highest_vm_end
= end
;
905 * If the vma has a ->close operation then the driver probably needs to release
906 * per-vma resources, so we don't attempt to merge those.
908 static inline int is_mergeable_vma(struct vm_area_struct
*vma
,
909 struct file
*file
, unsigned long vm_flags
)
912 * VM_SOFTDIRTY should not prevent from VMA merging, if we
913 * match the flags but dirty bit -- the caller should mark
914 * merged VMA as dirty. If dirty bit won't be excluded from
915 * comparison, we increase pressue on the memory system forcing
916 * the kernel to generate new VMAs when old one could be
919 if ((vma
->vm_flags
^ vm_flags
) & ~VM_SOFTDIRTY
)
921 if (vma
->vm_file
!= file
)
923 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
928 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
929 struct anon_vma
*anon_vma2
,
930 struct vm_area_struct
*vma
)
933 * The list_is_singular() test is to avoid merging VMA cloned from
934 * parents. This can improve scalability caused by anon_vma lock.
936 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
937 list_is_singular(&vma
->anon_vma_chain
)))
939 return anon_vma1
== anon_vma2
;
943 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
944 * in front of (at a lower virtual address and file offset than) the vma.
946 * We cannot merge two vmas if they have differently assigned (non-NULL)
947 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
949 * We don't check here for the merged mmap wrapping around the end of pagecache
950 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
951 * wrap, nor mmaps which cover the final page at index -1UL.
954 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
955 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
957 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
958 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
959 if (vma
->vm_pgoff
== vm_pgoff
)
966 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
967 * beyond (at a higher virtual address and file offset than) the vma.
969 * We cannot merge two vmas if they have differently assigned (non-NULL)
970 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
973 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
974 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
976 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
977 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
979 vm_pglen
= vma_pages(vma
);
980 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
987 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
988 * whether that can be merged with its predecessor or its successor.
989 * Or both (it neatly fills a hole).
991 * In most cases - when called for mmap, brk or mremap - [addr,end) is
992 * certain not to be mapped by the time vma_merge is called; but when
993 * called for mprotect, it is certain to be already mapped (either at
994 * an offset within prev, or at the start of next), and the flags of
995 * this area are about to be changed to vm_flags - and the no-change
996 * case has already been eliminated.
998 * The following mprotect cases have to be considered, where AAAA is
999 * the area passed down from mprotect_fixup, never extending beyond one
1000 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1002 * AAAA AAAA AAAA AAAA
1003 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1004 * cannot merge might become might become might become
1005 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1006 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1007 * mremap move: PPPPNNNNNNNN 8
1009 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1010 * might become case 1 below case 2 below case 3 below
1012 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1013 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1015 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
1016 struct vm_area_struct
*prev
, unsigned long addr
,
1017 unsigned long end
, unsigned long vm_flags
,
1018 struct anon_vma
*anon_vma
, struct file
*file
,
1019 pgoff_t pgoff
, struct mempolicy
*policy
)
1021 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
1022 struct vm_area_struct
*area
, *next
;
1026 * We later require that vma->vm_flags == vm_flags,
1027 * so this tests vma->vm_flags & VM_SPECIAL, too.
1029 if (vm_flags
& VM_SPECIAL
)
1033 next
= prev
->vm_next
;
1037 if (next
&& next
->vm_end
== end
) /* cases 6, 7, 8 */
1038 next
= next
->vm_next
;
1041 * Can it merge with the predecessor?
1043 if (prev
&& prev
->vm_end
== addr
&&
1044 mpol_equal(vma_policy(prev
), policy
) &&
1045 can_vma_merge_after(prev
, vm_flags
,
1046 anon_vma
, file
, pgoff
)) {
1048 * OK, it can. Can we now merge in the successor as well?
1050 if (next
&& end
== next
->vm_start
&&
1051 mpol_equal(policy
, vma_policy(next
)) &&
1052 can_vma_merge_before(next
, vm_flags
,
1053 anon_vma
, file
, pgoff
+pglen
) &&
1054 is_mergeable_anon_vma(prev
->anon_vma
,
1055 next
->anon_vma
, NULL
)) {
1057 err
= vma_adjust(prev
, prev
->vm_start
,
1058 next
->vm_end
, prev
->vm_pgoff
, NULL
);
1059 } else /* cases 2, 5, 7 */
1060 err
= vma_adjust(prev
, prev
->vm_start
,
1061 end
, prev
->vm_pgoff
, NULL
);
1064 khugepaged_enter_vma_merge(prev
);
1069 * Can this new request be merged in front of next?
1071 if (next
&& end
== next
->vm_start
&&
1072 mpol_equal(policy
, vma_policy(next
)) &&
1073 can_vma_merge_before(next
, vm_flags
,
1074 anon_vma
, file
, pgoff
+pglen
)) {
1075 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
1076 err
= vma_adjust(prev
, prev
->vm_start
,
1077 addr
, prev
->vm_pgoff
, NULL
);
1078 else /* cases 3, 8 */
1079 err
= vma_adjust(area
, addr
, next
->vm_end
,
1080 next
->vm_pgoff
- pglen
, NULL
);
1083 khugepaged_enter_vma_merge(area
);
1091 * Rough compatbility check to quickly see if it's even worth looking
1092 * at sharing an anon_vma.
1094 * They need to have the same vm_file, and the flags can only differ
1095 * in things that mprotect may change.
1097 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1098 * we can merge the two vma's. For example, we refuse to merge a vma if
1099 * there is a vm_ops->close() function, because that indicates that the
1100 * driver is doing some kind of reference counting. But that doesn't
1101 * really matter for the anon_vma sharing case.
1103 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1105 return a
->vm_end
== b
->vm_start
&&
1106 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
1107 a
->vm_file
== b
->vm_file
&&
1108 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_READ
|VM_WRITE
|VM_EXEC
|VM_SOFTDIRTY
)) &&
1109 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
1113 * Do some basic sanity checking to see if we can re-use the anon_vma
1114 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1115 * the same as 'old', the other will be the new one that is trying
1116 * to share the anon_vma.
1118 * NOTE! This runs with mm_sem held for reading, so it is possible that
1119 * the anon_vma of 'old' is concurrently in the process of being set up
1120 * by another page fault trying to merge _that_. But that's ok: if it
1121 * is being set up, that automatically means that it will be a singleton
1122 * acceptable for merging, so we can do all of this optimistically. But
1123 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1125 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1126 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1127 * is to return an anon_vma that is "complex" due to having gone through
1130 * We also make sure that the two vma's are compatible (adjacent,
1131 * and with the same memory policies). That's all stable, even with just
1132 * a read lock on the mm_sem.
1134 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
1136 if (anon_vma_compatible(a
, b
)) {
1137 struct anon_vma
*anon_vma
= ACCESS_ONCE(old
->anon_vma
);
1139 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
1146 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1147 * neighbouring vmas for a suitable anon_vma, before it goes off
1148 * to allocate a new anon_vma. It checks because a repetitive
1149 * sequence of mprotects and faults may otherwise lead to distinct
1150 * anon_vmas being allocated, preventing vma merge in subsequent
1153 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
1155 struct anon_vma
*anon_vma
;
1156 struct vm_area_struct
*near
;
1158 near
= vma
->vm_next
;
1162 anon_vma
= reusable_anon_vma(near
, vma
, near
);
1166 near
= vma
->vm_prev
;
1170 anon_vma
= reusable_anon_vma(near
, near
, vma
);
1175 * There's no absolute need to look only at touching neighbours:
1176 * we could search further afield for "compatible" anon_vmas.
1177 * But it would probably just be a waste of time searching,
1178 * or lead to too many vmas hanging off the same anon_vma.
1179 * We're trying to allow mprotect remerging later on,
1180 * not trying to minimize memory used for anon_vmas.
1185 #ifdef CONFIG_PROC_FS
1186 void vm_stat_account(struct mm_struct
*mm
, unsigned long flags
,
1187 struct file
*file
, long pages
)
1189 const unsigned long stack_flags
1190 = VM_STACK_FLAGS
& (VM_GROWSUP
|VM_GROWSDOWN
);
1192 mm
->total_vm
+= pages
;
1195 mm
->shared_vm
+= pages
;
1196 if ((flags
& (VM_EXEC
|VM_WRITE
)) == VM_EXEC
)
1197 mm
->exec_vm
+= pages
;
1198 } else if (flags
& stack_flags
)
1199 mm
->stack_vm
+= pages
;
1201 #endif /* CONFIG_PROC_FS */
1204 * If a hint addr is less than mmap_min_addr change hint to be as
1205 * low as possible but still greater than mmap_min_addr
1207 static inline unsigned long round_hint_to_min(unsigned long hint
)
1210 if (((void *)hint
!= NULL
) &&
1211 (hint
< mmap_min_addr
))
1212 return PAGE_ALIGN(mmap_min_addr
);
1216 static inline int mlock_future_check(struct mm_struct
*mm
,
1217 unsigned long flags
,
1220 unsigned long locked
, lock_limit
;
1222 /* mlock MCL_FUTURE? */
1223 if (flags
& VM_LOCKED
) {
1224 locked
= len
>> PAGE_SHIFT
;
1225 locked
+= mm
->locked_vm
;
1226 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1227 lock_limit
>>= PAGE_SHIFT
;
1228 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1235 * The caller must hold down_write(¤t->mm->mmap_sem).
1238 unsigned long do_mmap_pgoff(struct file
*file
, unsigned long addr
,
1239 unsigned long len
, unsigned long prot
,
1240 unsigned long flags
, unsigned long pgoff
,
1241 unsigned long *populate
)
1243 struct mm_struct
*mm
= current
->mm
;
1244 vm_flags_t vm_flags
;
1249 * Does the application expect PROT_READ to imply PROT_EXEC?
1251 * (the exception is when the underlying filesystem is noexec
1252 * mounted, in which case we dont add PROT_EXEC.)
1254 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1255 if (!(file
&& (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
)))
1261 if (!(flags
& MAP_FIXED
))
1262 addr
= round_hint_to_min(addr
);
1264 /* Careful about overflows.. */
1265 len
= PAGE_ALIGN(len
);
1269 /* offset overflow? */
1270 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1273 /* Too many mappings? */
1274 if (mm
->map_count
> sysctl_max_map_count
)
1277 /* Obtain the address to map to. we verify (or select) it and ensure
1278 * that it represents a valid section of the address space.
1280 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1281 if (addr
& ~PAGE_MASK
)
1284 /* Do simple checking here so the lower-level routines won't have
1285 * to. we assume access permissions have been handled by the open
1286 * of the memory object, so we don't do any here.
1288 vm_flags
= calc_vm_prot_bits(prot
) | calc_vm_flag_bits(flags
) |
1289 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1291 if (flags
& MAP_LOCKED
)
1292 if (!can_do_mlock())
1295 if (mlock_future_check(mm
, vm_flags
, len
))
1299 struct inode
*inode
= file_inode(file
);
1301 switch (flags
& MAP_TYPE
) {
1303 if ((prot
&PROT_WRITE
) && !(file
->f_mode
&FMODE_WRITE
))
1307 * Make sure we don't allow writing to an append-only
1310 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1314 * Make sure there are no mandatory locks on the file.
1316 if (locks_verify_locked(file
))
1319 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1320 if (!(file
->f_mode
& FMODE_WRITE
))
1321 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1325 if (!(file
->f_mode
& FMODE_READ
))
1327 if (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
) {
1328 if (vm_flags
& VM_EXEC
)
1330 vm_flags
&= ~VM_MAYEXEC
;
1333 if (!file
->f_op
->mmap
)
1335 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1343 switch (flags
& MAP_TYPE
) {
1345 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1351 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1355 * Set pgoff according to addr for anon_vma.
1357 pgoff
= addr
>> PAGE_SHIFT
;
1365 * Set 'VM_NORESERVE' if we should not account for the
1366 * memory use of this mapping.
1368 if (flags
& MAP_NORESERVE
) {
1369 /* We honor MAP_NORESERVE if allowed to overcommit */
1370 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1371 vm_flags
|= VM_NORESERVE
;
1373 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1374 if (file
&& is_file_hugepages(file
))
1375 vm_flags
|= VM_NORESERVE
;
1378 addr
= mmap_region(file
, addr
, len
, vm_flags
, pgoff
);
1379 if (!IS_ERR_VALUE(addr
) &&
1380 ((vm_flags
& VM_LOCKED
) ||
1381 (flags
& (MAP_POPULATE
| MAP_NONBLOCK
)) == MAP_POPULATE
))
1386 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1387 unsigned long, prot
, unsigned long, flags
,
1388 unsigned long, fd
, unsigned long, pgoff
)
1390 struct file
*file
= NULL
;
1391 unsigned long retval
= -EBADF
;
1393 if (!(flags
& MAP_ANONYMOUS
)) {
1394 audit_mmap_fd(fd
, flags
);
1398 if (is_file_hugepages(file
))
1399 len
= ALIGN(len
, huge_page_size(hstate_file(file
)));
1401 if (unlikely(flags
& MAP_HUGETLB
&& !is_file_hugepages(file
)))
1403 } else if (flags
& MAP_HUGETLB
) {
1404 struct user_struct
*user
= NULL
;
1407 hs
= hstate_sizelog((flags
>> MAP_HUGE_SHIFT
) & SHM_HUGE_MASK
);
1411 len
= ALIGN(len
, huge_page_size(hs
));
1413 * VM_NORESERVE is used because the reservations will be
1414 * taken when vm_ops->mmap() is called
1415 * A dummy user value is used because we are not locking
1416 * memory so no accounting is necessary
1418 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, len
,
1420 &user
, HUGETLB_ANONHUGE_INODE
,
1421 (flags
>> MAP_HUGE_SHIFT
) & MAP_HUGE_MASK
);
1423 return PTR_ERR(file
);
1426 flags
&= ~(MAP_EXECUTABLE
| MAP_DENYWRITE
);
1428 retval
= vm_mmap_pgoff(file
, addr
, len
, prot
, flags
, pgoff
);
1436 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1437 struct mmap_arg_struct
{
1441 unsigned long flags
;
1443 unsigned long offset
;
1446 SYSCALL_DEFINE1(old_mmap
, struct mmap_arg_struct __user
*, arg
)
1448 struct mmap_arg_struct a
;
1450 if (copy_from_user(&a
, arg
, sizeof(a
)))
1452 if (a
.offset
& ~PAGE_MASK
)
1455 return sys_mmap_pgoff(a
.addr
, a
.len
, a
.prot
, a
.flags
, a
.fd
,
1456 a
.offset
>> PAGE_SHIFT
);
1458 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1461 * Some shared mappigns will want the pages marked read-only
1462 * to track write events. If so, we'll downgrade vm_page_prot
1463 * to the private version (using protection_map[] without the
1466 int vma_wants_writenotify(struct vm_area_struct
*vma
)
1468 vm_flags_t vm_flags
= vma
->vm_flags
;
1470 /* If it was private or non-writable, the write bit is already clear */
1471 if ((vm_flags
& (VM_WRITE
|VM_SHARED
)) != ((VM_WRITE
|VM_SHARED
)))
1474 /* The backer wishes to know when pages are first written to? */
1475 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
)
1478 /* The open routine did something to the protections already? */
1479 if (pgprot_val(vma
->vm_page_prot
) !=
1480 pgprot_val(vm_get_page_prot(vm_flags
)))
1483 /* Specialty mapping? */
1484 if (vm_flags
& VM_PFNMAP
)
1487 /* Can the mapping track the dirty pages? */
1488 return vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
1489 mapping_cap_account_dirty(vma
->vm_file
->f_mapping
);
1493 * We account for memory if it's a private writeable mapping,
1494 * not hugepages and VM_NORESERVE wasn't set.
1496 static inline int accountable_mapping(struct file
*file
, vm_flags_t vm_flags
)
1499 * hugetlb has its own accounting separate from the core VM
1500 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1502 if (file
&& is_file_hugepages(file
))
1505 return (vm_flags
& (VM_NORESERVE
| VM_SHARED
| VM_WRITE
)) == VM_WRITE
;
1508 unsigned long mmap_region(struct file
*file
, unsigned long addr
,
1509 unsigned long len
, vm_flags_t vm_flags
, unsigned long pgoff
)
1511 struct mm_struct
*mm
= current
->mm
;
1512 struct vm_area_struct
*vma
, *prev
;
1514 struct rb_node
**rb_link
, *rb_parent
;
1515 unsigned long charged
= 0;
1517 /* Check against address space limit. */
1518 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
)) {
1519 unsigned long nr_pages
;
1522 * MAP_FIXED may remove pages of mappings that intersects with
1523 * requested mapping. Account for the pages it would unmap.
1525 if (!(vm_flags
& MAP_FIXED
))
1528 nr_pages
= count_vma_pages_range(mm
, addr
, addr
+ len
);
1530 if (!may_expand_vm(mm
, (len
>> PAGE_SHIFT
) - nr_pages
))
1534 /* Clear old maps */
1537 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
1538 if (do_munmap(mm
, addr
, len
))
1544 * Private writable mapping: check memory availability
1546 if (accountable_mapping(file
, vm_flags
)) {
1547 charged
= len
>> PAGE_SHIFT
;
1548 if (security_vm_enough_memory_mm(mm
, charged
))
1550 vm_flags
|= VM_ACCOUNT
;
1554 * Can we just expand an old mapping?
1556 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
, NULL
, file
, pgoff
, NULL
);
1561 * Determine the object being mapped and call the appropriate
1562 * specific mapper. the address has already been validated, but
1563 * not unmapped, but the maps are removed from the list.
1565 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
1572 vma
->vm_start
= addr
;
1573 vma
->vm_end
= addr
+ len
;
1574 vma
->vm_flags
= vm_flags
;
1575 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1576 vma
->vm_pgoff
= pgoff
;
1577 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
1580 if (vm_flags
& VM_DENYWRITE
) {
1581 error
= deny_write_access(file
);
1585 if (vm_flags
& VM_SHARED
) {
1586 error
= mapping_map_writable(file
->f_mapping
);
1588 goto allow_write_and_free_vma
;
1591 /* ->mmap() can change vma->vm_file, but must guarantee that
1592 * vma_link() below can deny write-access if VM_DENYWRITE is set
1593 * and map writably if VM_SHARED is set. This usually means the
1594 * new file must not have been exposed to user-space, yet.
1596 vma
->vm_file
= get_file(file
);
1597 error
= file
->f_op
->mmap(file
, vma
);
1599 goto unmap_and_free_vma
;
1601 /* Can addr have changed??
1603 * Answer: Yes, several device drivers can do it in their
1604 * f_op->mmap method. -DaveM
1605 * Bug: If addr is changed, prev, rb_link, rb_parent should
1606 * be updated for vma_link()
1608 WARN_ON_ONCE(addr
!= vma
->vm_start
);
1610 addr
= vma
->vm_start
;
1611 vm_flags
= vma
->vm_flags
;
1612 } else if (vm_flags
& VM_SHARED
) {
1613 error
= shmem_zero_setup(vma
);
1618 if (vma_wants_writenotify(vma
)) {
1619 pgprot_t pprot
= vma
->vm_page_prot
;
1621 /* Can vma->vm_page_prot have changed??
1623 * Answer: Yes, drivers may have changed it in their
1624 * f_op->mmap method.
1626 * Ensures that vmas marked as uncached stay that way.
1628 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
& ~VM_SHARED
);
1629 if (pgprot_val(pprot
) == pgprot_val(pgprot_noncached(pprot
)))
1630 vma
->vm_page_prot
= pgprot_noncached(vma
->vm_page_prot
);
1633 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1634 /* Once vma denies write, undo our temporary denial count */
1636 if (vm_flags
& VM_SHARED
)
1637 mapping_unmap_writable(file
->f_mapping
);
1638 if (vm_flags
& VM_DENYWRITE
)
1639 allow_write_access(file
);
1641 file
= vma
->vm_file
;
1643 perf_event_mmap(vma
);
1645 vm_stat_account(mm
, vm_flags
, file
, len
>> PAGE_SHIFT
);
1646 if (vm_flags
& VM_LOCKED
) {
1647 if (!((vm_flags
& VM_SPECIAL
) || is_vm_hugetlb_page(vma
) ||
1648 vma
== get_gate_vma(current
->mm
)))
1649 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1651 vma
->vm_flags
&= ~VM_LOCKED
;
1658 * New (or expanded) vma always get soft dirty status.
1659 * Otherwise user-space soft-dirty page tracker won't
1660 * be able to distinguish situation when vma area unmapped,
1661 * then new mapped in-place (which must be aimed as
1662 * a completely new data area).
1664 vma
->vm_flags
|= VM_SOFTDIRTY
;
1669 vma
->vm_file
= NULL
;
1672 /* Undo any partial mapping done by a device driver. */
1673 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1675 if (vm_flags
& VM_SHARED
)
1676 mapping_unmap_writable(file
->f_mapping
);
1677 allow_write_and_free_vma
:
1678 if (vm_flags
& VM_DENYWRITE
)
1679 allow_write_access(file
);
1681 kmem_cache_free(vm_area_cachep
, vma
);
1684 vm_unacct_memory(charged
);
1688 unsigned long unmapped_area(struct vm_unmapped_area_info
*info
)
1691 * We implement the search by looking for an rbtree node that
1692 * immediately follows a suitable gap. That is,
1693 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1694 * - gap_end = vma->vm_start >= info->low_limit + length;
1695 * - gap_end - gap_start >= length
1698 struct mm_struct
*mm
= current
->mm
;
1699 struct vm_area_struct
*vma
;
1700 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1702 /* Adjust search length to account for worst case alignment overhead */
1703 length
= info
->length
+ info
->align_mask
;
1704 if (length
< info
->length
)
1707 /* Adjust search limits by the desired length */
1708 if (info
->high_limit
< length
)
1710 high_limit
= info
->high_limit
- length
;
1712 if (info
->low_limit
> high_limit
)
1714 low_limit
= info
->low_limit
+ length
;
1716 /* Check if rbtree root looks promising */
1717 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1719 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1720 if (vma
->rb_subtree_gap
< length
)
1724 /* Visit left subtree if it looks promising */
1725 gap_end
= vma
->vm_start
;
1726 if (gap_end
>= low_limit
&& vma
->vm_rb
.rb_left
) {
1727 struct vm_area_struct
*left
=
1728 rb_entry(vma
->vm_rb
.rb_left
,
1729 struct vm_area_struct
, vm_rb
);
1730 if (left
->rb_subtree_gap
>= length
) {
1736 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1738 /* Check if current node has a suitable gap */
1739 if (gap_start
> high_limit
)
1741 if (gap_end
>= low_limit
&& gap_end
- gap_start
>= length
)
1744 /* Visit right subtree if it looks promising */
1745 if (vma
->vm_rb
.rb_right
) {
1746 struct vm_area_struct
*right
=
1747 rb_entry(vma
->vm_rb
.rb_right
,
1748 struct vm_area_struct
, vm_rb
);
1749 if (right
->rb_subtree_gap
>= length
) {
1755 /* Go back up the rbtree to find next candidate node */
1757 struct rb_node
*prev
= &vma
->vm_rb
;
1758 if (!rb_parent(prev
))
1760 vma
= rb_entry(rb_parent(prev
),
1761 struct vm_area_struct
, vm_rb
);
1762 if (prev
== vma
->vm_rb
.rb_left
) {
1763 gap_start
= vma
->vm_prev
->vm_end
;
1764 gap_end
= vma
->vm_start
;
1771 /* Check highest gap, which does not precede any rbtree node */
1772 gap_start
= mm
->highest_vm_end
;
1773 gap_end
= ULONG_MAX
; /* Only for VM_BUG_ON below */
1774 if (gap_start
> high_limit
)
1778 /* We found a suitable gap. Clip it with the original low_limit. */
1779 if (gap_start
< info
->low_limit
)
1780 gap_start
= info
->low_limit
;
1782 /* Adjust gap address to the desired alignment */
1783 gap_start
+= (info
->align_offset
- gap_start
) & info
->align_mask
;
1785 VM_BUG_ON(gap_start
+ info
->length
> info
->high_limit
);
1786 VM_BUG_ON(gap_start
+ info
->length
> gap_end
);
1790 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info
*info
)
1792 struct mm_struct
*mm
= current
->mm
;
1793 struct vm_area_struct
*vma
;
1794 unsigned long length
, low_limit
, high_limit
, gap_start
, gap_end
;
1796 /* Adjust search length to account for worst case alignment overhead */
1797 length
= info
->length
+ info
->align_mask
;
1798 if (length
< info
->length
)
1802 * Adjust search limits by the desired length.
1803 * See implementation comment at top of unmapped_area().
1805 gap_end
= info
->high_limit
;
1806 if (gap_end
< length
)
1808 high_limit
= gap_end
- length
;
1810 if (info
->low_limit
> high_limit
)
1812 low_limit
= info
->low_limit
+ length
;
1814 /* Check highest gap, which does not precede any rbtree node */
1815 gap_start
= mm
->highest_vm_end
;
1816 if (gap_start
<= high_limit
)
1819 /* Check if rbtree root looks promising */
1820 if (RB_EMPTY_ROOT(&mm
->mm_rb
))
1822 vma
= rb_entry(mm
->mm_rb
.rb_node
, struct vm_area_struct
, vm_rb
);
1823 if (vma
->rb_subtree_gap
< length
)
1827 /* Visit right subtree if it looks promising */
1828 gap_start
= vma
->vm_prev
? vma
->vm_prev
->vm_end
: 0;
1829 if (gap_start
<= high_limit
&& vma
->vm_rb
.rb_right
) {
1830 struct vm_area_struct
*right
=
1831 rb_entry(vma
->vm_rb
.rb_right
,
1832 struct vm_area_struct
, vm_rb
);
1833 if (right
->rb_subtree_gap
>= length
) {
1840 /* Check if current node has a suitable gap */
1841 gap_end
= vma
->vm_start
;
1842 if (gap_end
< low_limit
)
1844 if (gap_start
<= high_limit
&& gap_end
- gap_start
>= length
)
1847 /* Visit left subtree if it looks promising */
1848 if (vma
->vm_rb
.rb_left
) {
1849 struct vm_area_struct
*left
=
1850 rb_entry(vma
->vm_rb
.rb_left
,
1851 struct vm_area_struct
, vm_rb
);
1852 if (left
->rb_subtree_gap
>= length
) {
1858 /* Go back up the rbtree to find next candidate node */
1860 struct rb_node
*prev
= &vma
->vm_rb
;
1861 if (!rb_parent(prev
))
1863 vma
= rb_entry(rb_parent(prev
),
1864 struct vm_area_struct
, vm_rb
);
1865 if (prev
== vma
->vm_rb
.rb_right
) {
1866 gap_start
= vma
->vm_prev
?
1867 vma
->vm_prev
->vm_end
: 0;
1874 /* We found a suitable gap. Clip it with the original high_limit. */
1875 if (gap_end
> info
->high_limit
)
1876 gap_end
= info
->high_limit
;
1879 /* Compute highest gap address at the desired alignment */
1880 gap_end
-= info
->length
;
1881 gap_end
-= (gap_end
- info
->align_offset
) & info
->align_mask
;
1883 VM_BUG_ON(gap_end
< info
->low_limit
);
1884 VM_BUG_ON(gap_end
< gap_start
);
1888 /* Get an address range which is currently unmapped.
1889 * For shmat() with addr=0.
1891 * Ugly calling convention alert:
1892 * Return value with the low bits set means error value,
1894 * if (ret & ~PAGE_MASK)
1897 * This function "knows" that -ENOMEM has the bits set.
1899 #ifndef HAVE_ARCH_UNMAPPED_AREA
1901 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
1902 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
1904 struct mm_struct
*mm
= current
->mm
;
1905 struct vm_area_struct
*vma
;
1906 struct vm_unmapped_area_info info
;
1908 if (len
> TASK_SIZE
- mmap_min_addr
)
1911 if (flags
& MAP_FIXED
)
1915 addr
= PAGE_ALIGN(addr
);
1916 vma
= find_vma(mm
, addr
);
1917 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1918 (!vma
|| addr
+ len
<= vma
->vm_start
))
1924 info
.low_limit
= mm
->mmap_base
;
1925 info
.high_limit
= TASK_SIZE
;
1926 info
.align_mask
= 0;
1927 return vm_unmapped_area(&info
);
1932 * This mmap-allocator allocates new areas top-down from below the
1933 * stack's low limit (the base):
1935 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1937 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
1938 const unsigned long len
, const unsigned long pgoff
,
1939 const unsigned long flags
)
1941 struct vm_area_struct
*vma
;
1942 struct mm_struct
*mm
= current
->mm
;
1943 unsigned long addr
= addr0
;
1944 struct vm_unmapped_area_info info
;
1946 /* requested length too big for entire address space */
1947 if (len
> TASK_SIZE
- mmap_min_addr
)
1950 if (flags
& MAP_FIXED
)
1953 /* requesting a specific address */
1955 addr
= PAGE_ALIGN(addr
);
1956 vma
= find_vma(mm
, addr
);
1957 if (TASK_SIZE
- len
>= addr
&& addr
>= mmap_min_addr
&&
1958 (!vma
|| addr
+ len
<= vma
->vm_start
))
1962 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
1964 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
1965 info
.high_limit
= mm
->mmap_base
;
1966 info
.align_mask
= 0;
1967 addr
= vm_unmapped_area(&info
);
1970 * A failed mmap() very likely causes application failure,
1971 * so fall back to the bottom-up function here. This scenario
1972 * can happen with large stack limits and large mmap()
1975 if (addr
& ~PAGE_MASK
) {
1976 VM_BUG_ON(addr
!= -ENOMEM
);
1978 info
.low_limit
= TASK_UNMAPPED_BASE
;
1979 info
.high_limit
= TASK_SIZE
;
1980 addr
= vm_unmapped_area(&info
);
1988 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
1989 unsigned long pgoff
, unsigned long flags
)
1991 unsigned long (*get_area
)(struct file
*, unsigned long,
1992 unsigned long, unsigned long, unsigned long);
1994 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
1998 /* Careful about overflows.. */
1999 if (len
> TASK_SIZE
)
2002 get_area
= current
->mm
->get_unmapped_area
;
2003 if (file
&& file
->f_op
->get_unmapped_area
)
2004 get_area
= file
->f_op
->get_unmapped_area
;
2005 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
2006 if (IS_ERR_VALUE(addr
))
2009 if (addr
> TASK_SIZE
- len
)
2011 if (addr
& ~PAGE_MASK
)
2014 addr
= arch_rebalance_pgtables(addr
, len
);
2015 error
= security_mmap_addr(addr
);
2016 return error
? error
: addr
;
2019 EXPORT_SYMBOL(get_unmapped_area
);
2021 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2022 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
2024 struct rb_node
*rb_node
;
2025 struct vm_area_struct
*vma
;
2027 /* Check the cache first. */
2028 vma
= vmacache_find(mm
, addr
);
2032 rb_node
= mm
->mm_rb
.rb_node
;
2036 struct vm_area_struct
*tmp
;
2038 tmp
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2040 if (tmp
->vm_end
> addr
) {
2042 if (tmp
->vm_start
<= addr
)
2044 rb_node
= rb_node
->rb_left
;
2046 rb_node
= rb_node
->rb_right
;
2050 vmacache_update(addr
, vma
);
2054 EXPORT_SYMBOL(find_vma
);
2057 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2059 struct vm_area_struct
*
2060 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
2061 struct vm_area_struct
**pprev
)
2063 struct vm_area_struct
*vma
;
2065 vma
= find_vma(mm
, addr
);
2067 *pprev
= vma
->vm_prev
;
2069 struct rb_node
*rb_node
= mm
->mm_rb
.rb_node
;
2072 *pprev
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
2073 rb_node
= rb_node
->rb_right
;
2080 * Verify that the stack growth is acceptable and
2081 * update accounting. This is shared with both the
2082 * grow-up and grow-down cases.
2084 static int acct_stack_growth(struct vm_area_struct
*vma
, unsigned long size
, unsigned long grow
)
2086 struct mm_struct
*mm
= vma
->vm_mm
;
2087 struct rlimit
*rlim
= current
->signal
->rlim
;
2088 unsigned long new_start
;
2090 /* address space limit tests */
2091 if (!may_expand_vm(mm
, grow
))
2094 /* Stack limit test */
2095 if (size
> ACCESS_ONCE(rlim
[RLIMIT_STACK
].rlim_cur
))
2098 /* mlock limit tests */
2099 if (vma
->vm_flags
& VM_LOCKED
) {
2100 unsigned long locked
;
2101 unsigned long limit
;
2102 locked
= mm
->locked_vm
+ grow
;
2103 limit
= ACCESS_ONCE(rlim
[RLIMIT_MEMLOCK
].rlim_cur
);
2104 limit
>>= PAGE_SHIFT
;
2105 if (locked
> limit
&& !capable(CAP_IPC_LOCK
))
2109 /* Check to ensure the stack will not grow into a hugetlb-only region */
2110 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
2112 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
2116 * Overcommit.. This must be the final test, as it will
2117 * update security statistics.
2119 if (security_vm_enough_memory_mm(mm
, grow
))
2122 /* Ok, everything looks good - let it rip */
2123 if (vma
->vm_flags
& VM_LOCKED
)
2124 mm
->locked_vm
+= grow
;
2125 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, grow
);
2129 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2131 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2132 * vma is the last one with address > vma->vm_end. Have to extend vma.
2134 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
)
2138 if (!(vma
->vm_flags
& VM_GROWSUP
))
2142 * We must make sure the anon_vma is allocated
2143 * so that the anon_vma locking is not a noop.
2145 if (unlikely(anon_vma_prepare(vma
)))
2147 vma_lock_anon_vma(vma
);
2150 * vma->vm_start/vm_end cannot change under us because the caller
2151 * is required to hold the mmap_sem in read mode. We need the
2152 * anon_vma lock to serialize against concurrent expand_stacks.
2153 * Also guard against wrapping around to address 0.
2155 if (address
< PAGE_ALIGN(address
+4))
2156 address
= PAGE_ALIGN(address
+4);
2158 vma_unlock_anon_vma(vma
);
2163 /* Somebody else might have raced and expanded it already */
2164 if (address
> vma
->vm_end
) {
2165 unsigned long size
, grow
;
2167 size
= address
- vma
->vm_start
;
2168 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
2171 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
2172 error
= acct_stack_growth(vma
, size
, grow
);
2175 * vma_gap_update() doesn't support concurrent
2176 * updates, but we only hold a shared mmap_sem
2177 * lock here, so we need to protect against
2178 * concurrent vma expansions.
2179 * vma_lock_anon_vma() doesn't help here, as
2180 * we don't guarantee that all growable vmas
2181 * in a mm share the same root anon vma.
2182 * So, we reuse mm->page_table_lock to guard
2183 * against concurrent vma expansions.
2185 spin_lock(&vma
->vm_mm
->page_table_lock
);
2186 anon_vma_interval_tree_pre_update_vma(vma
);
2187 vma
->vm_end
= address
;
2188 anon_vma_interval_tree_post_update_vma(vma
);
2190 vma_gap_update(vma
->vm_next
);
2192 vma
->vm_mm
->highest_vm_end
= address
;
2193 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2195 perf_event_mmap(vma
);
2199 vma_unlock_anon_vma(vma
);
2200 khugepaged_enter_vma_merge(vma
);
2201 validate_mm(vma
->vm_mm
);
2204 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2207 * vma is the first one with address < vma->vm_start. Have to extend vma.
2209 int expand_downwards(struct vm_area_struct
*vma
,
2210 unsigned long address
)
2215 * We must make sure the anon_vma is allocated
2216 * so that the anon_vma locking is not a noop.
2218 if (unlikely(anon_vma_prepare(vma
)))
2221 address
&= PAGE_MASK
;
2222 error
= security_mmap_addr(address
);
2226 vma_lock_anon_vma(vma
);
2229 * vma->vm_start/vm_end cannot change under us because the caller
2230 * is required to hold the mmap_sem in read mode. We need the
2231 * anon_vma lock to serialize against concurrent expand_stacks.
2234 /* Somebody else might have raced and expanded it already */
2235 if (address
< vma
->vm_start
) {
2236 unsigned long size
, grow
;
2238 size
= vma
->vm_end
- address
;
2239 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
2242 if (grow
<= vma
->vm_pgoff
) {
2243 error
= acct_stack_growth(vma
, size
, grow
);
2246 * vma_gap_update() doesn't support concurrent
2247 * updates, but we only hold a shared mmap_sem
2248 * lock here, so we need to protect against
2249 * concurrent vma expansions.
2250 * vma_lock_anon_vma() doesn't help here, as
2251 * we don't guarantee that all growable vmas
2252 * in a mm share the same root anon vma.
2253 * So, we reuse mm->page_table_lock to guard
2254 * against concurrent vma expansions.
2256 spin_lock(&vma
->vm_mm
->page_table_lock
);
2257 anon_vma_interval_tree_pre_update_vma(vma
);
2258 vma
->vm_start
= address
;
2259 vma
->vm_pgoff
-= grow
;
2260 anon_vma_interval_tree_post_update_vma(vma
);
2261 vma_gap_update(vma
);
2262 spin_unlock(&vma
->vm_mm
->page_table_lock
);
2264 perf_event_mmap(vma
);
2268 vma_unlock_anon_vma(vma
);
2269 khugepaged_enter_vma_merge(vma
);
2270 validate_mm(vma
->vm_mm
);
2275 * Note how expand_stack() refuses to expand the stack all the way to
2276 * abut the next virtual mapping, *unless* that mapping itself is also
2277 * a stack mapping. We want to leave room for a guard page, after all
2278 * (the guard page itself is not added here, that is done by the
2279 * actual page faulting logic)
2281 * This matches the behavior of the guard page logic (see mm/memory.c:
2282 * check_stack_guard_page()), which only allows the guard page to be
2283 * removed under these circumstances.
2285 #ifdef CONFIG_STACK_GROWSUP
2286 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2288 struct vm_area_struct
*next
;
2290 address
&= PAGE_MASK
;
2291 next
= vma
->vm_next
;
2292 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
) {
2293 if (!(next
->vm_flags
& VM_GROWSUP
))
2296 return expand_upwards(vma
, address
);
2299 struct vm_area_struct
*
2300 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2302 struct vm_area_struct
*vma
, *prev
;
2305 vma
= find_vma_prev(mm
, addr
, &prev
);
2306 if (vma
&& (vma
->vm_start
<= addr
))
2308 if (!prev
|| expand_stack(prev
, addr
))
2310 if (prev
->vm_flags
& VM_LOCKED
)
2311 __mlock_vma_pages_range(prev
, addr
, prev
->vm_end
, NULL
);
2315 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
2317 struct vm_area_struct
*prev
;
2319 address
&= PAGE_MASK
;
2320 prev
= vma
->vm_prev
;
2321 if (prev
&& prev
->vm_end
== address
) {
2322 if (!(prev
->vm_flags
& VM_GROWSDOWN
))
2325 return expand_downwards(vma
, address
);
2328 struct vm_area_struct
*
2329 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
2331 struct vm_area_struct
*vma
;
2332 unsigned long start
;
2335 vma
= find_vma(mm
, addr
);
2338 if (vma
->vm_start
<= addr
)
2340 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
2342 start
= vma
->vm_start
;
2343 if (expand_stack(vma
, addr
))
2345 if (vma
->vm_flags
& VM_LOCKED
)
2346 __mlock_vma_pages_range(vma
, addr
, start
, NULL
);
2352 * Ok - we have the memory areas we should free on the vma list,
2353 * so release them, and do the vma updates.
2355 * Called with the mm semaphore held.
2357 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2359 unsigned long nr_accounted
= 0;
2361 /* Update high watermark before we lower total_vm */
2362 update_hiwater_vm(mm
);
2364 long nrpages
= vma_pages(vma
);
2366 if (vma
->vm_flags
& VM_ACCOUNT
)
2367 nr_accounted
+= nrpages
;
2368 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, -nrpages
);
2369 vma
= remove_vma(vma
);
2371 vm_unacct_memory(nr_accounted
);
2376 * Get rid of page table information in the indicated region.
2378 * Called with the mm semaphore held.
2380 static void unmap_region(struct mm_struct
*mm
,
2381 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
2382 unsigned long start
, unsigned long end
)
2384 struct vm_area_struct
*next
= prev
? prev
->vm_next
: mm
->mmap
;
2385 struct mmu_gather tlb
;
2388 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2389 update_hiwater_rss(mm
);
2390 unmap_vmas(&tlb
, vma
, start
, end
);
2391 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
2392 next
? next
->vm_start
: USER_PGTABLES_CEILING
);
2393 tlb_finish_mmu(&tlb
, start
, end
);
2397 * Create a list of vma's touched by the unmap, removing them from the mm's
2398 * vma list as we go..
2401 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2402 struct vm_area_struct
*prev
, unsigned long end
)
2404 struct vm_area_struct
**insertion_point
;
2405 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
? prev
->vm_end
: 0;
2421 tail_vma
->vm_next
= NULL
;
2423 /* Kill the cache */
2424 vmacache_invalidate(mm
);
2428 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2429 * munmap path where it doesn't make sense to fail.
2431 static int __split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2432 unsigned long addr
, int new_below
)
2434 struct vm_area_struct
*new;
2437 if (is_vm_hugetlb_page(vma
) && (addr
&
2438 ~(huge_page_mask(hstate_vma(vma
)))))
2441 new = kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2445 /* most fields are the same, copy all, and then fixup */
2448 INIT_LIST_HEAD(&new->anon_vma_chain
);
2453 new->vm_start
= addr
;
2454 new->vm_pgoff
+= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
);
2457 err
= vma_dup_policy(vma
, new);
2461 if (anon_vma_clone(new, vma
))
2465 get_file(new->vm_file
);
2467 if (new->vm_ops
&& new->vm_ops
->open
)
2468 new->vm_ops
->open(new);
2471 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2472 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2474 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2480 /* Clean everything up if vma_adjust failed. */
2481 if (new->vm_ops
&& new->vm_ops
->close
)
2482 new->vm_ops
->close(new);
2485 unlink_anon_vmas(new);
2487 mpol_put(vma_policy(new));
2489 kmem_cache_free(vm_area_cachep
, new);
2495 * Split a vma into two pieces at address 'addr', a new vma is allocated
2496 * either for the first part or the tail.
2498 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2499 unsigned long addr
, int new_below
)
2501 if (mm
->map_count
>= sysctl_max_map_count
)
2504 return __split_vma(mm
, vma
, addr
, new_below
);
2507 /* Munmap is split into 2 main parts -- this part which finds
2508 * what needs doing, and the areas themselves, which do the
2509 * work. This now handles partial unmappings.
2510 * Jeremy Fitzhardinge <jeremy@goop.org>
2512 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
)
2515 struct vm_area_struct
*vma
, *prev
, *last
;
2517 if ((start
& ~PAGE_MASK
) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2520 len
= PAGE_ALIGN(len
);
2524 /* Find the first overlapping VMA */
2525 vma
= find_vma(mm
, start
);
2528 prev
= vma
->vm_prev
;
2529 /* we have start < vma->vm_end */
2531 /* if it doesn't overlap, we have nothing.. */
2533 if (vma
->vm_start
>= end
)
2537 * If we need to split any vma, do it now to save pain later.
2539 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2540 * unmapped vm_area_struct will remain in use: so lower split_vma
2541 * places tmp vma above, and higher split_vma places tmp vma below.
2543 if (start
> vma
->vm_start
) {
2547 * Make sure that map_count on return from munmap() will
2548 * not exceed its limit; but let map_count go just above
2549 * its limit temporarily, to help free resources as expected.
2551 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2554 error
= __split_vma(mm
, vma
, start
, 0);
2560 /* Does it split the last one? */
2561 last
= find_vma(mm
, end
);
2562 if (last
&& end
> last
->vm_start
) {
2563 int error
= __split_vma(mm
, last
, end
, 1);
2567 vma
= prev
? prev
->vm_next
: mm
->mmap
;
2570 * unlock any mlock()ed ranges before detaching vmas
2572 if (mm
->locked_vm
) {
2573 struct vm_area_struct
*tmp
= vma
;
2574 while (tmp
&& tmp
->vm_start
< end
) {
2575 if (tmp
->vm_flags
& VM_LOCKED
) {
2576 mm
->locked_vm
-= vma_pages(tmp
);
2577 munlock_vma_pages_all(tmp
);
2584 * Remove the vma's, and unmap the actual pages
2586 detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
);
2587 unmap_region(mm
, vma
, prev
, start
, end
);
2589 /* Fix up all other VM information */
2590 remove_vma_list(mm
, vma
);
2595 int vm_munmap(unsigned long start
, size_t len
)
2598 struct mm_struct
*mm
= current
->mm
;
2600 down_write(&mm
->mmap_sem
);
2601 ret
= do_munmap(mm
, start
, len
);
2602 up_write(&mm
->mmap_sem
);
2605 EXPORT_SYMBOL(vm_munmap
);
2607 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2609 profile_munmap(addr
);
2610 return vm_munmap(addr
, len
);
2613 static inline void verify_mm_writelocked(struct mm_struct
*mm
)
2615 #ifdef CONFIG_DEBUG_VM
2616 if (unlikely(down_read_trylock(&mm
->mmap_sem
))) {
2618 up_read(&mm
->mmap_sem
);
2624 * this is really a simplified "do_mmap". it only handles
2625 * anonymous maps. eventually we may be able to do some
2626 * brk-specific accounting here.
2628 static unsigned long do_brk(unsigned long addr
, unsigned long len
)
2630 struct mm_struct
*mm
= current
->mm
;
2631 struct vm_area_struct
*vma
, *prev
;
2632 unsigned long flags
;
2633 struct rb_node
**rb_link
, *rb_parent
;
2634 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2637 len
= PAGE_ALIGN(len
);
2641 flags
= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2643 error
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2644 if (error
& ~PAGE_MASK
)
2647 error
= mlock_future_check(mm
, mm
->def_flags
, len
);
2652 * mm->mmap_sem is required to protect against another thread
2653 * changing the mappings in case we sleep.
2655 verify_mm_writelocked(mm
);
2658 * Clear old maps. this also does some error checking for us
2661 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
2662 if (do_munmap(mm
, addr
, len
))
2667 /* Check against address space limits *after* clearing old maps... */
2668 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
))
2671 if (mm
->map_count
> sysctl_max_map_count
)
2674 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
2677 /* Can we just expand an old private anonymous mapping? */
2678 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
2679 NULL
, NULL
, pgoff
, NULL
);
2684 * create a vma struct for an anonymous mapping
2686 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2688 vm_unacct_memory(len
>> PAGE_SHIFT
);
2692 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2694 vma
->vm_start
= addr
;
2695 vma
->vm_end
= addr
+ len
;
2696 vma
->vm_pgoff
= pgoff
;
2697 vma
->vm_flags
= flags
;
2698 vma
->vm_page_prot
= vm_get_page_prot(flags
);
2699 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2701 perf_event_mmap(vma
);
2702 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2703 if (flags
& VM_LOCKED
)
2704 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
2705 vma
->vm_flags
|= VM_SOFTDIRTY
;
2709 unsigned long vm_brk(unsigned long addr
, unsigned long len
)
2711 struct mm_struct
*mm
= current
->mm
;
2715 down_write(&mm
->mmap_sem
);
2716 ret
= do_brk(addr
, len
);
2717 populate
= ((mm
->def_flags
& VM_LOCKED
) != 0);
2718 up_write(&mm
->mmap_sem
);
2720 mm_populate(addr
, len
);
2723 EXPORT_SYMBOL(vm_brk
);
2725 /* Release all mmaps. */
2726 void exit_mmap(struct mm_struct
*mm
)
2728 struct mmu_gather tlb
;
2729 struct vm_area_struct
*vma
;
2730 unsigned long nr_accounted
= 0;
2732 /* mm's last user has gone, and its about to be pulled down */
2733 mmu_notifier_release(mm
);
2735 if (mm
->locked_vm
) {
2738 if (vma
->vm_flags
& VM_LOCKED
)
2739 munlock_vma_pages_all(vma
);
2747 if (!vma
) /* Can happen if dup_mmap() received an OOM */
2752 tlb_gather_mmu(&tlb
, mm
, 0, -1);
2753 /* update_hiwater_rss(mm) here? but nobody should be looking */
2754 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2755 unmap_vmas(&tlb
, vma
, 0, -1);
2757 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, USER_PGTABLES_CEILING
);
2758 tlb_finish_mmu(&tlb
, 0, -1);
2761 * Walk the list again, actually closing and freeing it,
2762 * with preemption enabled, without holding any MM locks.
2765 if (vma
->vm_flags
& VM_ACCOUNT
)
2766 nr_accounted
+= vma_pages(vma
);
2767 vma
= remove_vma(vma
);
2769 vm_unacct_memory(nr_accounted
);
2771 WARN_ON(atomic_long_read(&mm
->nr_ptes
) >
2772 (FIRST_USER_ADDRESS
+PMD_SIZE
-1)>>PMD_SHIFT
);
2775 /* Insert vm structure into process list sorted by address
2776 * and into the inode's i_mmap tree. If vm_file is non-NULL
2777 * then i_mmap_mutex is taken here.
2779 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2781 struct vm_area_struct
*prev
;
2782 struct rb_node
**rb_link
, *rb_parent
;
2785 * The vm_pgoff of a purely anonymous vma should be irrelevant
2786 * until its first write fault, when page's anon_vma and index
2787 * are set. But now set the vm_pgoff it will almost certainly
2788 * end up with (unless mremap moves it elsewhere before that
2789 * first wfault), so /proc/pid/maps tells a consistent story.
2791 * By setting it to reflect the virtual start address of the
2792 * vma, merges and splits can happen in a seamless way, just
2793 * using the existing file pgoff checks and manipulations.
2794 * Similarly in do_mmap_pgoff and in do_brk.
2796 if (!vma
->vm_file
) {
2797 BUG_ON(vma
->anon_vma
);
2798 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2800 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
2801 &prev
, &rb_link
, &rb_parent
))
2803 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
2804 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
2807 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2812 * Copy the vma structure to a new location in the same mm,
2813 * prior to moving page table entries, to effect an mremap move.
2815 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
2816 unsigned long addr
, unsigned long len
, pgoff_t pgoff
,
2817 bool *need_rmap_locks
)
2819 struct vm_area_struct
*vma
= *vmap
;
2820 unsigned long vma_start
= vma
->vm_start
;
2821 struct mm_struct
*mm
= vma
->vm_mm
;
2822 struct vm_area_struct
*new_vma
, *prev
;
2823 struct rb_node
**rb_link
, *rb_parent
;
2824 bool faulted_in_anon_vma
= true;
2827 * If anonymous vma has not yet been faulted, update new pgoff
2828 * to match new location, to increase its chance of merging.
2830 if (unlikely(!vma
->vm_file
&& !vma
->anon_vma
)) {
2831 pgoff
= addr
>> PAGE_SHIFT
;
2832 faulted_in_anon_vma
= false;
2835 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
2836 return NULL
; /* should never get here */
2837 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
2838 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
));
2841 * Source vma may have been merged into new_vma
2843 if (unlikely(vma_start
>= new_vma
->vm_start
&&
2844 vma_start
< new_vma
->vm_end
)) {
2846 * The only way we can get a vma_merge with
2847 * self during an mremap is if the vma hasn't
2848 * been faulted in yet and we were allowed to
2849 * reset the dst vma->vm_pgoff to the
2850 * destination address of the mremap to allow
2851 * the merge to happen. mremap must change the
2852 * vm_pgoff linearity between src and dst vmas
2853 * (in turn preventing a vma_merge) to be
2854 * safe. It is only safe to keep the vm_pgoff
2855 * linear if there are no pages mapped yet.
2857 VM_BUG_ON_VMA(faulted_in_anon_vma
, new_vma
);
2858 *vmap
= vma
= new_vma
;
2860 *need_rmap_locks
= (new_vma
->vm_pgoff
<= vma
->vm_pgoff
);
2862 new_vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2865 new_vma
->vm_start
= addr
;
2866 new_vma
->vm_end
= addr
+ len
;
2867 new_vma
->vm_pgoff
= pgoff
;
2868 if (vma_dup_policy(vma
, new_vma
))
2870 INIT_LIST_HEAD(&new_vma
->anon_vma_chain
);
2871 if (anon_vma_clone(new_vma
, vma
))
2872 goto out_free_mempol
;
2873 if (new_vma
->vm_file
)
2874 get_file(new_vma
->vm_file
);
2875 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
2876 new_vma
->vm_ops
->open(new_vma
);
2877 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
2878 *need_rmap_locks
= false;
2884 mpol_put(vma_policy(new_vma
));
2886 kmem_cache_free(vm_area_cachep
, new_vma
);
2891 * Return true if the calling process may expand its vm space by the passed
2894 int may_expand_vm(struct mm_struct
*mm
, unsigned long npages
)
2896 unsigned long cur
= mm
->total_vm
; /* pages */
2899 lim
= rlimit(RLIMIT_AS
) >> PAGE_SHIFT
;
2901 if (cur
+ npages
> lim
)
2906 static int special_mapping_fault(struct vm_area_struct
*vma
,
2907 struct vm_fault
*vmf
);
2910 * Having a close hook prevents vma merging regardless of flags.
2912 static void special_mapping_close(struct vm_area_struct
*vma
)
2916 static const char *special_mapping_name(struct vm_area_struct
*vma
)
2918 return ((struct vm_special_mapping
*)vma
->vm_private_data
)->name
;
2921 static const struct vm_operations_struct special_mapping_vmops
= {
2922 .close
= special_mapping_close
,
2923 .fault
= special_mapping_fault
,
2924 .name
= special_mapping_name
,
2927 static const struct vm_operations_struct legacy_special_mapping_vmops
= {
2928 .close
= special_mapping_close
,
2929 .fault
= special_mapping_fault
,
2932 static int special_mapping_fault(struct vm_area_struct
*vma
,
2933 struct vm_fault
*vmf
)
2936 struct page
**pages
;
2939 * special mappings have no vm_file, and in that case, the mm
2940 * uses vm_pgoff internally. So we have to subtract it from here.
2941 * We are allowed to do this because we are the mm; do not copy
2942 * this code into drivers!
2944 pgoff
= vmf
->pgoff
- vma
->vm_pgoff
;
2946 if (vma
->vm_ops
== &legacy_special_mapping_vmops
)
2947 pages
= vma
->vm_private_data
;
2949 pages
= ((struct vm_special_mapping
*)vma
->vm_private_data
)->
2952 for (; pgoff
&& *pages
; ++pages
)
2956 struct page
*page
= *pages
;
2962 return VM_FAULT_SIGBUS
;
2965 static struct vm_area_struct
*__install_special_mapping(
2966 struct mm_struct
*mm
,
2967 unsigned long addr
, unsigned long len
,
2968 unsigned long vm_flags
, const struct vm_operations_struct
*ops
,
2972 struct vm_area_struct
*vma
;
2974 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2975 if (unlikely(vma
== NULL
))
2976 return ERR_PTR(-ENOMEM
);
2978 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2980 vma
->vm_start
= addr
;
2981 vma
->vm_end
= addr
+ len
;
2983 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
| VM_SOFTDIRTY
;
2984 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
2987 vma
->vm_private_data
= priv
;
2989 ret
= insert_vm_struct(mm
, vma
);
2993 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2995 perf_event_mmap(vma
);
3000 kmem_cache_free(vm_area_cachep
, vma
);
3001 return ERR_PTR(ret
);
3005 * Called with mm->mmap_sem held for writing.
3006 * Insert a new vma covering the given region, with the given flags.
3007 * Its pages are supplied by the given array of struct page *.
3008 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3009 * The region past the last page supplied will always produce SIGBUS.
3010 * The array pointer and the pages it points to are assumed to stay alive
3011 * for as long as this mapping might exist.
3013 struct vm_area_struct
*_install_special_mapping(
3014 struct mm_struct
*mm
,
3015 unsigned long addr
, unsigned long len
,
3016 unsigned long vm_flags
, const struct vm_special_mapping
*spec
)
3018 return __install_special_mapping(mm
, addr
, len
, vm_flags
,
3019 &special_mapping_vmops
, (void *)spec
);
3022 int install_special_mapping(struct mm_struct
*mm
,
3023 unsigned long addr
, unsigned long len
,
3024 unsigned long vm_flags
, struct page
**pages
)
3026 struct vm_area_struct
*vma
= __install_special_mapping(
3027 mm
, addr
, len
, vm_flags
, &legacy_special_mapping_vmops
,
3030 return PTR_ERR_OR_ZERO(vma
);
3033 static DEFINE_MUTEX(mm_all_locks_mutex
);
3035 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
3037 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3039 * The LSB of head.next can't change from under us
3040 * because we hold the mm_all_locks_mutex.
3042 down_write_nest_lock(&anon_vma
->root
->rwsem
, &mm
->mmap_sem
);
3044 * We can safely modify head.next after taking the
3045 * anon_vma->root->rwsem. If some other vma in this mm shares
3046 * the same anon_vma we won't take it again.
3048 * No need of atomic instructions here, head.next
3049 * can't change from under us thanks to the
3050 * anon_vma->root->rwsem.
3052 if (__test_and_set_bit(0, (unsigned long *)
3053 &anon_vma
->root
->rb_root
.rb_node
))
3058 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
3060 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3062 * AS_MM_ALL_LOCKS can't change from under us because
3063 * we hold the mm_all_locks_mutex.
3065 * Operations on ->flags have to be atomic because
3066 * even if AS_MM_ALL_LOCKS is stable thanks to the
3067 * mm_all_locks_mutex, there may be other cpus
3068 * changing other bitflags in parallel to us.
3070 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
3072 mutex_lock_nest_lock(&mapping
->i_mmap_mutex
, &mm
->mmap_sem
);
3077 * This operation locks against the VM for all pte/vma/mm related
3078 * operations that could ever happen on a certain mm. This includes
3079 * vmtruncate, try_to_unmap, and all page faults.
3081 * The caller must take the mmap_sem in write mode before calling
3082 * mm_take_all_locks(). The caller isn't allowed to release the
3083 * mmap_sem until mm_drop_all_locks() returns.
3085 * mmap_sem in write mode is required in order to block all operations
3086 * that could modify pagetables and free pages without need of
3087 * altering the vma layout (for example populate_range() with
3088 * nonlinear vmas). It's also needed in write mode to avoid new
3089 * anon_vmas to be associated with existing vmas.
3091 * A single task can't take more than one mm_take_all_locks() in a row
3092 * or it would deadlock.
3094 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3095 * mapping->flags avoid to take the same lock twice, if more than one
3096 * vma in this mm is backed by the same anon_vma or address_space.
3098 * We can take all the locks in random order because the VM code
3099 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3100 * takes more than one of them in a row. Secondly we're protected
3101 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3103 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3104 * that may have to take thousand of locks.
3106 * mm_take_all_locks() can fail if it's interrupted by signals.
3108 int mm_take_all_locks(struct mm_struct
*mm
)
3110 struct vm_area_struct
*vma
;
3111 struct anon_vma_chain
*avc
;
3113 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3115 mutex_lock(&mm_all_locks_mutex
);
3117 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3118 if (signal_pending(current
))
3120 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3121 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
3124 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3125 if (signal_pending(current
))
3128 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3129 vm_lock_anon_vma(mm
, avc
->anon_vma
);
3135 mm_drop_all_locks(mm
);
3139 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
3141 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
3143 * The LSB of head.next can't change to 0 from under
3144 * us because we hold the mm_all_locks_mutex.
3146 * We must however clear the bitflag before unlocking
3147 * the vma so the users using the anon_vma->rb_root will
3148 * never see our bitflag.
3150 * No need of atomic instructions here, head.next
3151 * can't change from under us until we release the
3152 * anon_vma->root->rwsem.
3154 if (!__test_and_clear_bit(0, (unsigned long *)
3155 &anon_vma
->root
->rb_root
.rb_node
))
3157 anon_vma_unlock_write(anon_vma
);
3161 static void vm_unlock_mapping(struct address_space
*mapping
)
3163 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
3165 * AS_MM_ALL_LOCKS can't change to 0 from under us
3166 * because we hold the mm_all_locks_mutex.
3168 mutex_unlock(&mapping
->i_mmap_mutex
);
3169 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
3176 * The mmap_sem cannot be released by the caller until
3177 * mm_drop_all_locks() returns.
3179 void mm_drop_all_locks(struct mm_struct
*mm
)
3181 struct vm_area_struct
*vma
;
3182 struct anon_vma_chain
*avc
;
3184 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
3185 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
3187 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
3189 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
3190 vm_unlock_anon_vma(avc
->anon_vma
);
3191 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
3192 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
3195 mutex_unlock(&mm_all_locks_mutex
);
3199 * initialise the VMA slab
3201 void __init
mmap_init(void)
3205 ret
= percpu_counter_init(&vm_committed_as
, 0, GFP_KERNEL
);
3210 * Initialise sysctl_user_reserve_kbytes.
3212 * This is intended to prevent a user from starting a single memory hogging
3213 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3216 * The default value is min(3% of free memory, 128MB)
3217 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3219 static int init_user_reserve(void)
3221 unsigned long free_kbytes
;
3223 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3225 sysctl_user_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 17);
3228 subsys_initcall(init_user_reserve
);
3231 * Initialise sysctl_admin_reserve_kbytes.
3233 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3234 * to log in and kill a memory hogging process.
3236 * Systems with more than 256MB will reserve 8MB, enough to recover
3237 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3238 * only reserve 3% of free pages by default.
3240 static int init_admin_reserve(void)
3242 unsigned long free_kbytes
;
3244 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3246 sysctl_admin_reserve_kbytes
= min(free_kbytes
/ 32, 1UL << 13);
3249 subsys_initcall(init_admin_reserve
);
3252 * Reinititalise user and admin reserves if memory is added or removed.
3254 * The default user reserve max is 128MB, and the default max for the
3255 * admin reserve is 8MB. These are usually, but not always, enough to
3256 * enable recovery from a memory hogging process using login/sshd, a shell,
3257 * and tools like top. It may make sense to increase or even disable the
3258 * reserve depending on the existence of swap or variations in the recovery
3259 * tools. So, the admin may have changed them.
3261 * If memory is added and the reserves have been eliminated or increased above
3262 * the default max, then we'll trust the admin.
3264 * If memory is removed and there isn't enough free memory, then we
3265 * need to reset the reserves.
3267 * Otherwise keep the reserve set by the admin.
3269 static int reserve_mem_notifier(struct notifier_block
*nb
,
3270 unsigned long action
, void *data
)
3272 unsigned long tmp
, free_kbytes
;
3276 /* Default max is 128MB. Leave alone if modified by operator. */
3277 tmp
= sysctl_user_reserve_kbytes
;
3278 if (0 < tmp
&& tmp
< (1UL << 17))
3279 init_user_reserve();
3281 /* Default max is 8MB. Leave alone if modified by operator. */
3282 tmp
= sysctl_admin_reserve_kbytes
;
3283 if (0 < tmp
&& tmp
< (1UL << 13))
3284 init_admin_reserve();
3288 free_kbytes
= global_page_state(NR_FREE_PAGES
) << (PAGE_SHIFT
- 10);
3290 if (sysctl_user_reserve_kbytes
> free_kbytes
) {
3291 init_user_reserve();
3292 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3293 sysctl_user_reserve_kbytes
);
3296 if (sysctl_admin_reserve_kbytes
> free_kbytes
) {
3297 init_admin_reserve();
3298 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3299 sysctl_admin_reserve_kbytes
);
3308 static struct notifier_block reserve_mem_nb
= {
3309 .notifier_call
= reserve_mem_notifier
,
3312 static int __meminit
init_reserve_notifier(void)
3314 if (register_hotmemory_notifier(&reserve_mem_nb
))
3315 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3319 subsys_initcall(init_reserve_notifier
);