6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/export.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.h>
31 #include <linux/audit.h>
32 #include <linux/khugepaged.h>
33 #include <linux/uprobes.h>
35 #include <asm/uaccess.h>
36 #include <asm/cacheflush.h>
38 #include <asm/mmu_context.h>
42 #ifndef arch_mmap_check
43 #define arch_mmap_check(addr, len, flags) (0)
46 #ifndef arch_rebalance_pgtables
47 #define arch_rebalance_pgtables(addr, len) (addr)
50 static void unmap_region(struct mm_struct
*mm
,
51 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
52 unsigned long start
, unsigned long end
);
54 /* description of effects of mapping type and prot in current implementation.
55 * this is due to the limited x86 page protection hardware. The expected
56 * behavior is in parens:
59 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
60 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
61 * w: (no) no w: (no) no w: (yes) yes w: (no) no
62 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
64 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
65 * w: (no) no w: (no) no w: (copy) copy w: (no) no
66 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
69 pgprot_t protection_map
[16] = {
70 __P000
, __P001
, __P010
, __P011
, __P100
, __P101
, __P110
, __P111
,
71 __S000
, __S001
, __S010
, __S011
, __S100
, __S101
, __S110
, __S111
74 pgprot_t
vm_get_page_prot(unsigned long vm_flags
)
76 return __pgprot(pgprot_val(protection_map
[vm_flags
&
77 (VM_READ
|VM_WRITE
|VM_EXEC
|VM_SHARED
)]) |
78 pgprot_val(arch_vm_get_page_prot(vm_flags
)));
80 EXPORT_SYMBOL(vm_get_page_prot
);
82 int sysctl_overcommit_memory __read_mostly
= OVERCOMMIT_GUESS
; /* heuristic overcommit */
83 int sysctl_overcommit_ratio __read_mostly
= 50; /* default is 50% */
84 int sysctl_max_map_count __read_mostly
= DEFAULT_MAX_MAP_COUNT
;
86 * Make sure vm_committed_as in one cacheline and not cacheline shared with
87 * other variables. It can be updated by several CPUs frequently.
89 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp
;
92 * Check that a process has enough memory to allocate a new virtual
93 * mapping. 0 means there is enough memory for the allocation to
94 * succeed and -ENOMEM implies there is not.
96 * We currently support three overcommit policies, which are set via the
97 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
99 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
100 * Additional code 2002 Jul 20 by Robert Love.
102 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
104 * Note this is a helper function intended to be used by LSMs which
105 * wish to use this logic.
107 int __vm_enough_memory(struct mm_struct
*mm
, long pages
, int cap_sys_admin
)
109 unsigned long free
, allowed
;
111 vm_acct_memory(pages
);
114 * Sometimes we want to use more memory than we have
116 if (sysctl_overcommit_memory
== OVERCOMMIT_ALWAYS
)
119 if (sysctl_overcommit_memory
== OVERCOMMIT_GUESS
) {
120 free
= global_page_state(NR_FREE_PAGES
);
121 free
+= global_page_state(NR_FILE_PAGES
);
124 * shmem pages shouldn't be counted as free in this
125 * case, they can't be purged, only swapped out, and
126 * that won't affect the overall amount of available
127 * memory in the system.
129 free
-= global_page_state(NR_SHMEM
);
131 free
+= nr_swap_pages
;
134 * Any slabs which are created with the
135 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
136 * which are reclaimable, under pressure. The dentry
137 * cache and most inode caches should fall into this
139 free
+= global_page_state(NR_SLAB_RECLAIMABLE
);
142 * Leave reserved pages. The pages are not for anonymous pages.
144 if (free
<= totalreserve_pages
)
147 free
-= totalreserve_pages
;
150 * Leave the last 3% for root
161 allowed
= (totalram_pages
- hugetlb_total_pages())
162 * sysctl_overcommit_ratio
/ 100;
164 * Leave the last 3% for root
167 allowed
-= allowed
/ 32;
168 allowed
+= total_swap_pages
;
170 /* Don't let a single process grow too big:
171 leave 3% of the size of this process for other processes */
173 allowed
-= mm
->total_vm
/ 32;
175 if (percpu_counter_read_positive(&vm_committed_as
) < allowed
)
178 vm_unacct_memory(pages
);
184 * Requires inode->i_mapping->i_mmap_mutex
186 static void __remove_shared_vm_struct(struct vm_area_struct
*vma
,
187 struct file
*file
, struct address_space
*mapping
)
189 if (vma
->vm_flags
& VM_DENYWRITE
)
190 atomic_inc(&file
->f_path
.dentry
->d_inode
->i_writecount
);
191 if (vma
->vm_flags
& VM_SHARED
)
192 mapping
->i_mmap_writable
--;
194 flush_dcache_mmap_lock(mapping
);
195 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
196 list_del_init(&vma
->shared
.nonlinear
);
198 vma_interval_tree_remove(vma
, &mapping
->i_mmap
);
199 flush_dcache_mmap_unlock(mapping
);
203 * Unlink a file-based vm structure from its interval tree, to hide
204 * vma from rmap and vmtruncate before freeing its page tables.
206 void unlink_file_vma(struct vm_area_struct
*vma
)
208 struct file
*file
= vma
->vm_file
;
211 struct address_space
*mapping
= file
->f_mapping
;
212 mutex_lock(&mapping
->i_mmap_mutex
);
213 __remove_shared_vm_struct(vma
, file
, mapping
);
214 mutex_unlock(&mapping
->i_mmap_mutex
);
219 * Close a vm structure and free it, returning the next.
221 static struct vm_area_struct
*remove_vma(struct vm_area_struct
*vma
)
223 struct vm_area_struct
*next
= vma
->vm_next
;
226 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
227 vma
->vm_ops
->close(vma
);
230 mpol_put(vma_policy(vma
));
231 kmem_cache_free(vm_area_cachep
, vma
);
235 static unsigned long do_brk(unsigned long addr
, unsigned long len
);
237 SYSCALL_DEFINE1(brk
, unsigned long, brk
)
239 unsigned long rlim
, retval
;
240 unsigned long newbrk
, oldbrk
;
241 struct mm_struct
*mm
= current
->mm
;
242 unsigned long min_brk
;
244 down_write(&mm
->mmap_sem
);
246 #ifdef CONFIG_COMPAT_BRK
248 * CONFIG_COMPAT_BRK can still be overridden by setting
249 * randomize_va_space to 2, which will still cause mm->start_brk
250 * to be arbitrarily shifted
252 if (current
->brk_randomized
)
253 min_brk
= mm
->start_brk
;
255 min_brk
= mm
->end_data
;
257 min_brk
= mm
->start_brk
;
263 * Check against rlimit here. If this check is done later after the test
264 * of oldbrk with newbrk then it can escape the test and let the data
265 * segment grow beyond its set limit the in case where the limit is
266 * not page aligned -Ram Gupta
268 rlim
= rlimit(RLIMIT_DATA
);
269 if (rlim
< RLIM_INFINITY
&& (brk
- mm
->start_brk
) +
270 (mm
->end_data
- mm
->start_data
) > rlim
)
273 newbrk
= PAGE_ALIGN(brk
);
274 oldbrk
= PAGE_ALIGN(mm
->brk
);
275 if (oldbrk
== newbrk
)
278 /* Always allow shrinking brk. */
279 if (brk
<= mm
->brk
) {
280 if (!do_munmap(mm
, newbrk
, oldbrk
-newbrk
))
285 /* Check against existing mmap mappings. */
286 if (find_vma_intersection(mm
, oldbrk
, newbrk
+PAGE_SIZE
))
289 /* Ok, looks good - let it rip. */
290 if (do_brk(oldbrk
, newbrk
-oldbrk
) != oldbrk
)
296 up_write(&mm
->mmap_sem
);
300 #ifdef CONFIG_DEBUG_VM_RB
301 static int browse_rb(struct rb_root
*root
)
304 struct rb_node
*nd
, *pn
= NULL
;
305 unsigned long prev
= 0, pend
= 0;
307 for (nd
= rb_first(root
); nd
; nd
= rb_next(nd
)) {
308 struct vm_area_struct
*vma
;
309 vma
= rb_entry(nd
, struct vm_area_struct
, vm_rb
);
310 if (vma
->vm_start
< prev
)
311 printk("vm_start %lx prev %lx\n", vma
->vm_start
, prev
), i
= -1;
312 if (vma
->vm_start
< pend
)
313 printk("vm_start %lx pend %lx\n", vma
->vm_start
, pend
);
314 if (vma
->vm_start
> vma
->vm_end
)
315 printk("vm_end %lx < vm_start %lx\n", vma
->vm_end
, vma
->vm_start
);
318 prev
= vma
->vm_start
;
322 for (nd
= pn
; nd
; nd
= rb_prev(nd
)) {
326 printk("backwards %d, forwards %d\n", j
, i
), i
= 0;
330 void validate_mm(struct mm_struct
*mm
)
334 struct vm_area_struct
*vma
= mm
->mmap
;
336 struct anon_vma_chain
*avc
;
337 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
338 anon_vma_interval_tree_verify(avc
);
342 if (i
!= mm
->map_count
)
343 printk("map_count %d vm_next %d\n", mm
->map_count
, i
), bug
= 1;
344 i
= browse_rb(&mm
->mm_rb
);
345 if (i
!= mm
->map_count
)
346 printk("map_count %d rb %d\n", mm
->map_count
, i
), bug
= 1;
350 #define validate_mm(mm) do { } while (0)
354 * vma has some anon_vma assigned, and is already inserted on that
355 * anon_vma's interval trees.
357 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
358 * vma must be removed from the anon_vma's interval trees using
359 * anon_vma_interval_tree_pre_update_vma().
361 * After the update, the vma will be reinserted using
362 * anon_vma_interval_tree_post_update_vma().
364 * The entire update must be protected by exclusive mmap_sem and by
365 * the root anon_vma's mutex.
368 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct
*vma
)
370 struct anon_vma_chain
*avc
;
372 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
373 anon_vma_interval_tree_remove(avc
, &avc
->anon_vma
->rb_root
);
377 anon_vma_interval_tree_post_update_vma(struct vm_area_struct
*vma
)
379 struct anon_vma_chain
*avc
;
381 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
382 anon_vma_interval_tree_insert(avc
, &avc
->anon_vma
->rb_root
);
385 static int find_vma_links(struct mm_struct
*mm
, unsigned long addr
,
386 unsigned long end
, struct vm_area_struct
**pprev
,
387 struct rb_node
***rb_link
, struct rb_node
**rb_parent
)
389 struct rb_node
**__rb_link
, *__rb_parent
, *rb_prev
;
391 __rb_link
= &mm
->mm_rb
.rb_node
;
392 rb_prev
= __rb_parent
= NULL
;
395 struct vm_area_struct
*vma_tmp
;
397 __rb_parent
= *__rb_link
;
398 vma_tmp
= rb_entry(__rb_parent
, struct vm_area_struct
, vm_rb
);
400 if (vma_tmp
->vm_end
> addr
) {
401 /* Fail if an existing vma overlaps the area */
402 if (vma_tmp
->vm_start
< end
)
404 __rb_link
= &__rb_parent
->rb_left
;
406 rb_prev
= __rb_parent
;
407 __rb_link
= &__rb_parent
->rb_right
;
413 *pprev
= rb_entry(rb_prev
, struct vm_area_struct
, vm_rb
);
414 *rb_link
= __rb_link
;
415 *rb_parent
= __rb_parent
;
419 void __vma_link_rb(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
420 struct rb_node
**rb_link
, struct rb_node
*rb_parent
)
422 rb_link_node(&vma
->vm_rb
, rb_parent
, rb_link
);
423 rb_insert_color(&vma
->vm_rb
, &mm
->mm_rb
);
426 static void __vma_link_file(struct vm_area_struct
*vma
)
432 struct address_space
*mapping
= file
->f_mapping
;
434 if (vma
->vm_flags
& VM_DENYWRITE
)
435 atomic_dec(&file
->f_path
.dentry
->d_inode
->i_writecount
);
436 if (vma
->vm_flags
& VM_SHARED
)
437 mapping
->i_mmap_writable
++;
439 flush_dcache_mmap_lock(mapping
);
440 if (unlikely(vma
->vm_flags
& VM_NONLINEAR
))
441 vma_nonlinear_insert(vma
, &mapping
->i_mmap_nonlinear
);
443 vma_interval_tree_insert(vma
, &mapping
->i_mmap
);
444 flush_dcache_mmap_unlock(mapping
);
449 __vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
450 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
451 struct rb_node
*rb_parent
)
453 __vma_link_list(mm
, vma
, prev
, rb_parent
);
454 __vma_link_rb(mm
, vma
, rb_link
, rb_parent
);
457 static void vma_link(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
458 struct vm_area_struct
*prev
, struct rb_node
**rb_link
,
459 struct rb_node
*rb_parent
)
461 struct address_space
*mapping
= NULL
;
464 mapping
= vma
->vm_file
->f_mapping
;
467 mutex_lock(&mapping
->i_mmap_mutex
);
469 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
470 __vma_link_file(vma
);
473 mutex_unlock(&mapping
->i_mmap_mutex
);
480 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
481 * mm's list and rbtree. It has already been inserted into the interval tree.
483 static void __insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
485 struct vm_area_struct
*prev
;
486 struct rb_node
**rb_link
, *rb_parent
;
488 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
489 &prev
, &rb_link
, &rb_parent
))
491 __vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
496 __vma_unlink(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
497 struct vm_area_struct
*prev
)
499 struct vm_area_struct
*next
= vma
->vm_next
;
501 prev
->vm_next
= next
;
503 next
->vm_prev
= prev
;
504 rb_erase(&vma
->vm_rb
, &mm
->mm_rb
);
505 if (mm
->mmap_cache
== vma
)
506 mm
->mmap_cache
= prev
;
510 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
511 * is already present in an i_mmap tree without adjusting the tree.
512 * The following helper function should be used when such adjustments
513 * are necessary. The "insert" vma (if any) is to be inserted
514 * before we drop the necessary locks.
516 int vma_adjust(struct vm_area_struct
*vma
, unsigned long start
,
517 unsigned long end
, pgoff_t pgoff
, struct vm_area_struct
*insert
)
519 struct mm_struct
*mm
= vma
->vm_mm
;
520 struct vm_area_struct
*next
= vma
->vm_next
;
521 struct vm_area_struct
*importer
= NULL
;
522 struct address_space
*mapping
= NULL
;
523 struct rb_root
*root
= NULL
;
524 struct anon_vma
*anon_vma
= NULL
;
525 struct file
*file
= vma
->vm_file
;
526 long adjust_next
= 0;
529 if (next
&& !insert
) {
530 struct vm_area_struct
*exporter
= NULL
;
532 if (end
>= next
->vm_end
) {
534 * vma expands, overlapping all the next, and
535 * perhaps the one after too (mprotect case 6).
537 again
: remove_next
= 1 + (end
> next
->vm_end
);
541 } else if (end
> next
->vm_start
) {
543 * vma expands, overlapping part of the next:
544 * mprotect case 5 shifting the boundary up.
546 adjust_next
= (end
- next
->vm_start
) >> PAGE_SHIFT
;
549 } else if (end
< vma
->vm_end
) {
551 * vma shrinks, and !insert tells it's not
552 * split_vma inserting another: so it must be
553 * mprotect case 4 shifting the boundary down.
555 adjust_next
= - ((vma
->vm_end
- end
) >> PAGE_SHIFT
);
561 * Easily overlooked: when mprotect shifts the boundary,
562 * make sure the expanding vma has anon_vma set if the
563 * shrinking vma had, to cover any anon pages imported.
565 if (exporter
&& exporter
->anon_vma
&& !importer
->anon_vma
) {
566 if (anon_vma_clone(importer
, exporter
))
568 importer
->anon_vma
= exporter
->anon_vma
;
573 mapping
= file
->f_mapping
;
574 if (!(vma
->vm_flags
& VM_NONLINEAR
)) {
575 root
= &mapping
->i_mmap
;
576 uprobe_munmap(vma
, vma
->vm_start
, vma
->vm_end
);
579 uprobe_munmap(next
, next
->vm_start
,
583 mutex_lock(&mapping
->i_mmap_mutex
);
586 * Put into interval tree now, so instantiated pages
587 * are visible to arm/parisc __flush_dcache_page
588 * throughout; but we cannot insert into address
589 * space until vma start or end is updated.
591 __vma_link_file(insert
);
595 vma_adjust_trans_huge(vma
, start
, end
, adjust_next
);
597 anon_vma
= vma
->anon_vma
;
598 if (!anon_vma
&& adjust_next
)
599 anon_vma
= next
->anon_vma
;
601 VM_BUG_ON(adjust_next
&& next
->anon_vma
&&
602 anon_vma
!= next
->anon_vma
);
603 anon_vma_lock(anon_vma
);
604 anon_vma_interval_tree_pre_update_vma(vma
);
606 anon_vma_interval_tree_pre_update_vma(next
);
610 flush_dcache_mmap_lock(mapping
);
611 vma_interval_tree_remove(vma
, root
);
613 vma_interval_tree_remove(next
, root
);
616 vma
->vm_start
= start
;
618 vma
->vm_pgoff
= pgoff
;
620 next
->vm_start
+= adjust_next
<< PAGE_SHIFT
;
621 next
->vm_pgoff
+= adjust_next
;
626 vma_interval_tree_insert(next
, root
);
627 vma_interval_tree_insert(vma
, root
);
628 flush_dcache_mmap_unlock(mapping
);
633 * vma_merge has merged next into vma, and needs
634 * us to remove next before dropping the locks.
636 __vma_unlink(mm
, next
, vma
);
638 __remove_shared_vm_struct(next
, file
, mapping
);
641 * split_vma has split insert from vma, and needs
642 * us to insert it before dropping the locks
643 * (it may either follow vma or precede it).
645 __insert_vm_struct(mm
, insert
);
649 anon_vma_interval_tree_post_update_vma(vma
);
651 anon_vma_interval_tree_post_update_vma(next
);
652 anon_vma_unlock(anon_vma
);
655 mutex_unlock(&mapping
->i_mmap_mutex
);
666 uprobe_munmap(next
, next
->vm_start
, next
->vm_end
);
670 anon_vma_merge(vma
, next
);
672 mpol_put(vma_policy(next
));
673 kmem_cache_free(vm_area_cachep
, next
);
675 * In mprotect's case 6 (see comments on vma_merge),
676 * we must remove another next too. It would clutter
677 * up the code too much to do both in one go.
679 if (remove_next
== 2) {
693 * If the vma has a ->close operation then the driver probably needs to release
694 * per-vma resources, so we don't attempt to merge those.
696 static inline int is_mergeable_vma(struct vm_area_struct
*vma
,
697 struct file
*file
, unsigned long vm_flags
)
699 if (vma
->vm_flags
^ vm_flags
)
701 if (vma
->vm_file
!= file
)
703 if (vma
->vm_ops
&& vma
->vm_ops
->close
)
708 static inline int is_mergeable_anon_vma(struct anon_vma
*anon_vma1
,
709 struct anon_vma
*anon_vma2
,
710 struct vm_area_struct
*vma
)
713 * The list_is_singular() test is to avoid merging VMA cloned from
714 * parents. This can improve scalability caused by anon_vma lock.
716 if ((!anon_vma1
|| !anon_vma2
) && (!vma
||
717 list_is_singular(&vma
->anon_vma_chain
)))
719 return anon_vma1
== anon_vma2
;
723 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
724 * in front of (at a lower virtual address and file offset than) the vma.
726 * We cannot merge two vmas if they have differently assigned (non-NULL)
727 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
729 * We don't check here for the merged mmap wrapping around the end of pagecache
730 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
731 * wrap, nor mmaps which cover the final page at index -1UL.
734 can_vma_merge_before(struct vm_area_struct
*vma
, unsigned long vm_flags
,
735 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
737 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
738 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
739 if (vma
->vm_pgoff
== vm_pgoff
)
746 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
747 * beyond (at a higher virtual address and file offset than) the vma.
749 * We cannot merge two vmas if they have differently assigned (non-NULL)
750 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
753 can_vma_merge_after(struct vm_area_struct
*vma
, unsigned long vm_flags
,
754 struct anon_vma
*anon_vma
, struct file
*file
, pgoff_t vm_pgoff
)
756 if (is_mergeable_vma(vma
, file
, vm_flags
) &&
757 is_mergeable_anon_vma(anon_vma
, vma
->anon_vma
, vma
)) {
759 vm_pglen
= (vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
;
760 if (vma
->vm_pgoff
+ vm_pglen
== vm_pgoff
)
767 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
768 * whether that can be merged with its predecessor or its successor.
769 * Or both (it neatly fills a hole).
771 * In most cases - when called for mmap, brk or mremap - [addr,end) is
772 * certain not to be mapped by the time vma_merge is called; but when
773 * called for mprotect, it is certain to be already mapped (either at
774 * an offset within prev, or at the start of next), and the flags of
775 * this area are about to be changed to vm_flags - and the no-change
776 * case has already been eliminated.
778 * The following mprotect cases have to be considered, where AAAA is
779 * the area passed down from mprotect_fixup, never extending beyond one
780 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
782 * AAAA AAAA AAAA AAAA
783 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
784 * cannot merge might become might become might become
785 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
786 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
787 * mremap move: PPPPNNNNNNNN 8
789 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
790 * might become case 1 below case 2 below case 3 below
792 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
793 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
795 struct vm_area_struct
*vma_merge(struct mm_struct
*mm
,
796 struct vm_area_struct
*prev
, unsigned long addr
,
797 unsigned long end
, unsigned long vm_flags
,
798 struct anon_vma
*anon_vma
, struct file
*file
,
799 pgoff_t pgoff
, struct mempolicy
*policy
)
801 pgoff_t pglen
= (end
- addr
) >> PAGE_SHIFT
;
802 struct vm_area_struct
*area
, *next
;
806 * We later require that vma->vm_flags == vm_flags,
807 * so this tests vma->vm_flags & VM_SPECIAL, too.
809 if (vm_flags
& VM_SPECIAL
)
813 next
= prev
->vm_next
;
817 if (next
&& next
->vm_end
== end
) /* cases 6, 7, 8 */
818 next
= next
->vm_next
;
821 * Can it merge with the predecessor?
823 if (prev
&& prev
->vm_end
== addr
&&
824 mpol_equal(vma_policy(prev
), policy
) &&
825 can_vma_merge_after(prev
, vm_flags
,
826 anon_vma
, file
, pgoff
)) {
828 * OK, it can. Can we now merge in the successor as well?
830 if (next
&& end
== next
->vm_start
&&
831 mpol_equal(policy
, vma_policy(next
)) &&
832 can_vma_merge_before(next
, vm_flags
,
833 anon_vma
, file
, pgoff
+pglen
) &&
834 is_mergeable_anon_vma(prev
->anon_vma
,
835 next
->anon_vma
, NULL
)) {
837 err
= vma_adjust(prev
, prev
->vm_start
,
838 next
->vm_end
, prev
->vm_pgoff
, NULL
);
839 } else /* cases 2, 5, 7 */
840 err
= vma_adjust(prev
, prev
->vm_start
,
841 end
, prev
->vm_pgoff
, NULL
);
844 khugepaged_enter_vma_merge(prev
);
849 * Can this new request be merged in front of next?
851 if (next
&& end
== next
->vm_start
&&
852 mpol_equal(policy
, vma_policy(next
)) &&
853 can_vma_merge_before(next
, vm_flags
,
854 anon_vma
, file
, pgoff
+pglen
)) {
855 if (prev
&& addr
< prev
->vm_end
) /* case 4 */
856 err
= vma_adjust(prev
, prev
->vm_start
,
857 addr
, prev
->vm_pgoff
, NULL
);
858 else /* cases 3, 8 */
859 err
= vma_adjust(area
, addr
, next
->vm_end
,
860 next
->vm_pgoff
- pglen
, NULL
);
863 khugepaged_enter_vma_merge(area
);
871 * Rough compatbility check to quickly see if it's even worth looking
872 * at sharing an anon_vma.
874 * They need to have the same vm_file, and the flags can only differ
875 * in things that mprotect may change.
877 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
878 * we can merge the two vma's. For example, we refuse to merge a vma if
879 * there is a vm_ops->close() function, because that indicates that the
880 * driver is doing some kind of reference counting. But that doesn't
881 * really matter for the anon_vma sharing case.
883 static int anon_vma_compatible(struct vm_area_struct
*a
, struct vm_area_struct
*b
)
885 return a
->vm_end
== b
->vm_start
&&
886 mpol_equal(vma_policy(a
), vma_policy(b
)) &&
887 a
->vm_file
== b
->vm_file
&&
888 !((a
->vm_flags
^ b
->vm_flags
) & ~(VM_READ
|VM_WRITE
|VM_EXEC
)) &&
889 b
->vm_pgoff
== a
->vm_pgoff
+ ((b
->vm_start
- a
->vm_start
) >> PAGE_SHIFT
);
893 * Do some basic sanity checking to see if we can re-use the anon_vma
894 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
895 * the same as 'old', the other will be the new one that is trying
896 * to share the anon_vma.
898 * NOTE! This runs with mm_sem held for reading, so it is possible that
899 * the anon_vma of 'old' is concurrently in the process of being set up
900 * by another page fault trying to merge _that_. But that's ok: if it
901 * is being set up, that automatically means that it will be a singleton
902 * acceptable for merging, so we can do all of this optimistically. But
903 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
905 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
906 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
907 * is to return an anon_vma that is "complex" due to having gone through
910 * We also make sure that the two vma's are compatible (adjacent,
911 * and with the same memory policies). That's all stable, even with just
912 * a read lock on the mm_sem.
914 static struct anon_vma
*reusable_anon_vma(struct vm_area_struct
*old
, struct vm_area_struct
*a
, struct vm_area_struct
*b
)
916 if (anon_vma_compatible(a
, b
)) {
917 struct anon_vma
*anon_vma
= ACCESS_ONCE(old
->anon_vma
);
919 if (anon_vma
&& list_is_singular(&old
->anon_vma_chain
))
926 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
927 * neighbouring vmas for a suitable anon_vma, before it goes off
928 * to allocate a new anon_vma. It checks because a repetitive
929 * sequence of mprotects and faults may otherwise lead to distinct
930 * anon_vmas being allocated, preventing vma merge in subsequent
933 struct anon_vma
*find_mergeable_anon_vma(struct vm_area_struct
*vma
)
935 struct anon_vma
*anon_vma
;
936 struct vm_area_struct
*near
;
942 anon_vma
= reusable_anon_vma(near
, vma
, near
);
950 anon_vma
= reusable_anon_vma(near
, near
, vma
);
955 * There's no absolute need to look only at touching neighbours:
956 * we could search further afield for "compatible" anon_vmas.
957 * But it would probably just be a waste of time searching,
958 * or lead to too many vmas hanging off the same anon_vma.
959 * We're trying to allow mprotect remerging later on,
960 * not trying to minimize memory used for anon_vmas.
965 #ifdef CONFIG_PROC_FS
966 void vm_stat_account(struct mm_struct
*mm
, unsigned long flags
,
967 struct file
*file
, long pages
)
969 const unsigned long stack_flags
970 = VM_STACK_FLAGS
& (VM_GROWSUP
|VM_GROWSDOWN
);
972 mm
->total_vm
+= pages
;
975 mm
->shared_vm
+= pages
;
976 if ((flags
& (VM_EXEC
|VM_WRITE
)) == VM_EXEC
)
977 mm
->exec_vm
+= pages
;
978 } else if (flags
& stack_flags
)
979 mm
->stack_vm
+= pages
;
981 #endif /* CONFIG_PROC_FS */
984 * If a hint addr is less than mmap_min_addr change hint to be as
985 * low as possible but still greater than mmap_min_addr
987 static inline unsigned long round_hint_to_min(unsigned long hint
)
990 if (((void *)hint
!= NULL
) &&
991 (hint
< mmap_min_addr
))
992 return PAGE_ALIGN(mmap_min_addr
);
997 * The caller must hold down_write(¤t->mm->mmap_sem).
1000 unsigned long do_mmap_pgoff(struct file
*file
, unsigned long addr
,
1001 unsigned long len
, unsigned long prot
,
1002 unsigned long flags
, unsigned long pgoff
)
1004 struct mm_struct
* mm
= current
->mm
;
1005 struct inode
*inode
;
1006 vm_flags_t vm_flags
;
1009 * Does the application expect PROT_READ to imply PROT_EXEC?
1011 * (the exception is when the underlying filesystem is noexec
1012 * mounted, in which case we dont add PROT_EXEC.)
1014 if ((prot
& PROT_READ
) && (current
->personality
& READ_IMPLIES_EXEC
))
1015 if (!(file
&& (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
)))
1021 if (!(flags
& MAP_FIXED
))
1022 addr
= round_hint_to_min(addr
);
1024 /* Careful about overflows.. */
1025 len
= PAGE_ALIGN(len
);
1029 /* offset overflow? */
1030 if ((pgoff
+ (len
>> PAGE_SHIFT
)) < pgoff
)
1033 /* Too many mappings? */
1034 if (mm
->map_count
> sysctl_max_map_count
)
1037 /* Obtain the address to map to. we verify (or select) it and ensure
1038 * that it represents a valid section of the address space.
1040 addr
= get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
1041 if (addr
& ~PAGE_MASK
)
1044 /* Do simple checking here so the lower-level routines won't have
1045 * to. we assume access permissions have been handled by the open
1046 * of the memory object, so we don't do any here.
1048 vm_flags
= calc_vm_prot_bits(prot
) | calc_vm_flag_bits(flags
) |
1049 mm
->def_flags
| VM_MAYREAD
| VM_MAYWRITE
| VM_MAYEXEC
;
1051 if (flags
& MAP_LOCKED
)
1052 if (!can_do_mlock())
1055 /* mlock MCL_FUTURE? */
1056 if (vm_flags
& VM_LOCKED
) {
1057 unsigned long locked
, lock_limit
;
1058 locked
= len
>> PAGE_SHIFT
;
1059 locked
+= mm
->locked_vm
;
1060 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
1061 lock_limit
>>= PAGE_SHIFT
;
1062 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
1066 inode
= file
? file
->f_path
.dentry
->d_inode
: NULL
;
1069 switch (flags
& MAP_TYPE
) {
1071 if ((prot
&PROT_WRITE
) && !(file
->f_mode
&FMODE_WRITE
))
1075 * Make sure we don't allow writing to an append-only
1078 if (IS_APPEND(inode
) && (file
->f_mode
& FMODE_WRITE
))
1082 * Make sure there are no mandatory locks on the file.
1084 if (locks_verify_locked(inode
))
1087 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1088 if (!(file
->f_mode
& FMODE_WRITE
))
1089 vm_flags
&= ~(VM_MAYWRITE
| VM_SHARED
);
1093 if (!(file
->f_mode
& FMODE_READ
))
1095 if (file
->f_path
.mnt
->mnt_flags
& MNT_NOEXEC
) {
1096 if (vm_flags
& VM_EXEC
)
1098 vm_flags
&= ~VM_MAYEXEC
;
1101 if (!file
->f_op
|| !file
->f_op
->mmap
)
1109 switch (flags
& MAP_TYPE
) {
1115 vm_flags
|= VM_SHARED
| VM_MAYSHARE
;
1119 * Set pgoff according to addr for anon_vma.
1121 pgoff
= addr
>> PAGE_SHIFT
;
1128 return mmap_region(file
, addr
, len
, flags
, vm_flags
, pgoff
);
1131 SYSCALL_DEFINE6(mmap_pgoff
, unsigned long, addr
, unsigned long, len
,
1132 unsigned long, prot
, unsigned long, flags
,
1133 unsigned long, fd
, unsigned long, pgoff
)
1135 struct file
*file
= NULL
;
1136 unsigned long retval
= -EBADF
;
1138 if (!(flags
& MAP_ANONYMOUS
)) {
1139 audit_mmap_fd(fd
, flags
);
1140 if (unlikely(flags
& MAP_HUGETLB
))
1145 } else if (flags
& MAP_HUGETLB
) {
1146 struct user_struct
*user
= NULL
;
1148 * VM_NORESERVE is used because the reservations will be
1149 * taken when vm_ops->mmap() is called
1150 * A dummy user value is used because we are not locking
1151 * memory so no accounting is necessary
1153 file
= hugetlb_file_setup(HUGETLB_ANON_FILE
, addr
, len
,
1154 VM_NORESERVE
, &user
,
1155 HUGETLB_ANONHUGE_INODE
);
1157 return PTR_ERR(file
);
1160 flags
&= ~(MAP_EXECUTABLE
| MAP_DENYWRITE
);
1162 retval
= vm_mmap_pgoff(file
, addr
, len
, prot
, flags
, pgoff
);
1169 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1170 struct mmap_arg_struct
{
1174 unsigned long flags
;
1176 unsigned long offset
;
1179 SYSCALL_DEFINE1(old_mmap
, struct mmap_arg_struct __user
*, arg
)
1181 struct mmap_arg_struct a
;
1183 if (copy_from_user(&a
, arg
, sizeof(a
)))
1185 if (a
.offset
& ~PAGE_MASK
)
1188 return sys_mmap_pgoff(a
.addr
, a
.len
, a
.prot
, a
.flags
, a
.fd
,
1189 a
.offset
>> PAGE_SHIFT
);
1191 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1194 * Some shared mappigns will want the pages marked read-only
1195 * to track write events. If so, we'll downgrade vm_page_prot
1196 * to the private version (using protection_map[] without the
1199 int vma_wants_writenotify(struct vm_area_struct
*vma
)
1201 vm_flags_t vm_flags
= vma
->vm_flags
;
1203 /* If it was private or non-writable, the write bit is already clear */
1204 if ((vm_flags
& (VM_WRITE
|VM_SHARED
)) != ((VM_WRITE
|VM_SHARED
)))
1207 /* The backer wishes to know when pages are first written to? */
1208 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
)
1211 /* The open routine did something to the protections already? */
1212 if (pgprot_val(vma
->vm_page_prot
) !=
1213 pgprot_val(vm_get_page_prot(vm_flags
)))
1216 /* Specialty mapping? */
1217 if (vm_flags
& VM_PFNMAP
)
1220 /* Can the mapping track the dirty pages? */
1221 return vma
->vm_file
&& vma
->vm_file
->f_mapping
&&
1222 mapping_cap_account_dirty(vma
->vm_file
->f_mapping
);
1226 * We account for memory if it's a private writeable mapping,
1227 * not hugepages and VM_NORESERVE wasn't set.
1229 static inline int accountable_mapping(struct file
*file
, vm_flags_t vm_flags
)
1232 * hugetlb has its own accounting separate from the core VM
1233 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1235 if (file
&& is_file_hugepages(file
))
1238 return (vm_flags
& (VM_NORESERVE
| VM_SHARED
| VM_WRITE
)) == VM_WRITE
;
1241 unsigned long mmap_region(struct file
*file
, unsigned long addr
,
1242 unsigned long len
, unsigned long flags
,
1243 vm_flags_t vm_flags
, unsigned long pgoff
)
1245 struct mm_struct
*mm
= current
->mm
;
1246 struct vm_area_struct
*vma
, *prev
;
1247 int correct_wcount
= 0;
1249 struct rb_node
**rb_link
, *rb_parent
;
1250 unsigned long charged
= 0;
1251 struct inode
*inode
= file
? file
->f_path
.dentry
->d_inode
: NULL
;
1253 /* Clear old maps */
1256 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
1257 if (do_munmap(mm
, addr
, len
))
1262 /* Check against address space limit. */
1263 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
))
1267 * Set 'VM_NORESERVE' if we should not account for the
1268 * memory use of this mapping.
1270 if ((flags
& MAP_NORESERVE
)) {
1271 /* We honor MAP_NORESERVE if allowed to overcommit */
1272 if (sysctl_overcommit_memory
!= OVERCOMMIT_NEVER
)
1273 vm_flags
|= VM_NORESERVE
;
1275 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1276 if (file
&& is_file_hugepages(file
))
1277 vm_flags
|= VM_NORESERVE
;
1281 * Private writable mapping: check memory availability
1283 if (accountable_mapping(file
, vm_flags
)) {
1284 charged
= len
>> PAGE_SHIFT
;
1285 if (security_vm_enough_memory_mm(mm
, charged
))
1287 vm_flags
|= VM_ACCOUNT
;
1291 * Can we just expand an old mapping?
1293 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vm_flags
, NULL
, file
, pgoff
, NULL
);
1298 * Determine the object being mapped and call the appropriate
1299 * specific mapper. the address has already been validated, but
1300 * not unmapped, but the maps are removed from the list.
1302 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
1309 vma
->vm_start
= addr
;
1310 vma
->vm_end
= addr
+ len
;
1311 vma
->vm_flags
= vm_flags
;
1312 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
);
1313 vma
->vm_pgoff
= pgoff
;
1314 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
1316 error
= -EINVAL
; /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1319 if (vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
))
1321 if (vm_flags
& VM_DENYWRITE
) {
1322 error
= deny_write_access(file
);
1327 vma
->vm_file
= get_file(file
);
1328 error
= file
->f_op
->mmap(file
, vma
);
1330 goto unmap_and_free_vma
;
1332 /* Can addr have changed??
1334 * Answer: Yes, several device drivers can do it in their
1335 * f_op->mmap method. -DaveM
1337 addr
= vma
->vm_start
;
1338 pgoff
= vma
->vm_pgoff
;
1339 vm_flags
= vma
->vm_flags
;
1340 } else if (vm_flags
& VM_SHARED
) {
1341 if (unlikely(vm_flags
& (VM_GROWSDOWN
|VM_GROWSUP
)))
1343 error
= shmem_zero_setup(vma
);
1348 if (vma_wants_writenotify(vma
)) {
1349 pgprot_t pprot
= vma
->vm_page_prot
;
1351 /* Can vma->vm_page_prot have changed??
1353 * Answer: Yes, drivers may have changed it in their
1354 * f_op->mmap method.
1356 * Ensures that vmas marked as uncached stay that way.
1358 vma
->vm_page_prot
= vm_get_page_prot(vm_flags
& ~VM_SHARED
);
1359 if (pgprot_val(pprot
) == pgprot_val(pgprot_noncached(pprot
)))
1360 vma
->vm_page_prot
= pgprot_noncached(vma
->vm_page_prot
);
1363 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
1364 file
= vma
->vm_file
;
1366 /* Once vma denies write, undo our temporary denial count */
1368 atomic_inc(&inode
->i_writecount
);
1370 perf_event_mmap(vma
);
1372 vm_stat_account(mm
, vm_flags
, file
, len
>> PAGE_SHIFT
);
1373 if (vm_flags
& VM_LOCKED
) {
1374 if (!mlock_vma_pages_range(vma
, addr
, addr
+ len
))
1375 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
1376 } else if ((flags
& MAP_POPULATE
) && !(flags
& MAP_NONBLOCK
))
1377 make_pages_present(addr
, addr
+ len
);
1386 atomic_inc(&inode
->i_writecount
);
1387 vma
->vm_file
= NULL
;
1390 /* Undo any partial mapping done by a device driver. */
1391 unmap_region(mm
, vma
, prev
, vma
->vm_start
, vma
->vm_end
);
1394 kmem_cache_free(vm_area_cachep
, vma
);
1397 vm_unacct_memory(charged
);
1401 /* Get an address range which is currently unmapped.
1402 * For shmat() with addr=0.
1404 * Ugly calling convention alert:
1405 * Return value with the low bits set means error value,
1407 * if (ret & ~PAGE_MASK)
1410 * This function "knows" that -ENOMEM has the bits set.
1412 #ifndef HAVE_ARCH_UNMAPPED_AREA
1414 arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
1415 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
1417 struct mm_struct
*mm
= current
->mm
;
1418 struct vm_area_struct
*vma
;
1419 unsigned long start_addr
;
1421 if (len
> TASK_SIZE
)
1424 if (flags
& MAP_FIXED
)
1428 addr
= PAGE_ALIGN(addr
);
1429 vma
= find_vma(mm
, addr
);
1430 if (TASK_SIZE
- len
>= addr
&&
1431 (!vma
|| addr
+ len
<= vma
->vm_start
))
1434 if (len
> mm
->cached_hole_size
) {
1435 start_addr
= addr
= mm
->free_area_cache
;
1437 start_addr
= addr
= TASK_UNMAPPED_BASE
;
1438 mm
->cached_hole_size
= 0;
1442 for (vma
= find_vma(mm
, addr
); ; vma
= vma
->vm_next
) {
1443 /* At this point: (!vma || addr < vma->vm_end). */
1444 if (TASK_SIZE
- len
< addr
) {
1446 * Start a new search - just in case we missed
1449 if (start_addr
!= TASK_UNMAPPED_BASE
) {
1450 addr
= TASK_UNMAPPED_BASE
;
1452 mm
->cached_hole_size
= 0;
1457 if (!vma
|| addr
+ len
<= vma
->vm_start
) {
1459 * Remember the place where we stopped the search:
1461 mm
->free_area_cache
= addr
+ len
;
1464 if (addr
+ mm
->cached_hole_size
< vma
->vm_start
)
1465 mm
->cached_hole_size
= vma
->vm_start
- addr
;
1471 void arch_unmap_area(struct mm_struct
*mm
, unsigned long addr
)
1474 * Is this a new hole at the lowest possible address?
1476 if (addr
>= TASK_UNMAPPED_BASE
&& addr
< mm
->free_area_cache
)
1477 mm
->free_area_cache
= addr
;
1481 * This mmap-allocator allocates new areas top-down from below the
1482 * stack's low limit (the base):
1484 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1486 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
1487 const unsigned long len
, const unsigned long pgoff
,
1488 const unsigned long flags
)
1490 struct vm_area_struct
*vma
;
1491 struct mm_struct
*mm
= current
->mm
;
1492 unsigned long addr
= addr0
, start_addr
;
1494 /* requested length too big for entire address space */
1495 if (len
> TASK_SIZE
)
1498 if (flags
& MAP_FIXED
)
1501 /* requesting a specific address */
1503 addr
= PAGE_ALIGN(addr
);
1504 vma
= find_vma(mm
, addr
);
1505 if (TASK_SIZE
- len
>= addr
&&
1506 (!vma
|| addr
+ len
<= vma
->vm_start
))
1510 /* check if free_area_cache is useful for us */
1511 if (len
<= mm
->cached_hole_size
) {
1512 mm
->cached_hole_size
= 0;
1513 mm
->free_area_cache
= mm
->mmap_base
;
1517 /* either no address requested or can't fit in requested address hole */
1518 start_addr
= addr
= mm
->free_area_cache
;
1526 * Lookup failure means no vma is above this address,
1527 * else if new region fits below vma->vm_start,
1528 * return with success:
1530 vma
= find_vma(mm
, addr
);
1531 if (!vma
|| addr
+len
<= vma
->vm_start
)
1532 /* remember the address as a hint for next time */
1533 return (mm
->free_area_cache
= addr
);
1535 /* remember the largest hole we saw so far */
1536 if (addr
+ mm
->cached_hole_size
< vma
->vm_start
)
1537 mm
->cached_hole_size
= vma
->vm_start
- addr
;
1539 /* try just below the current vma->vm_start */
1540 addr
= vma
->vm_start
-len
;
1541 } while (len
< vma
->vm_start
);
1545 * if hint left us with no space for the requested
1546 * mapping then try again:
1548 * Note: this is different with the case of bottomup
1549 * which does the fully line-search, but we use find_vma
1550 * here that causes some holes skipped.
1552 if (start_addr
!= mm
->mmap_base
) {
1553 mm
->free_area_cache
= mm
->mmap_base
;
1554 mm
->cached_hole_size
= 0;
1559 * A failed mmap() very likely causes application failure,
1560 * so fall back to the bottom-up function here. This scenario
1561 * can happen with large stack limits and large mmap()
1564 mm
->cached_hole_size
= ~0UL;
1565 mm
->free_area_cache
= TASK_UNMAPPED_BASE
;
1566 addr
= arch_get_unmapped_area(filp
, addr0
, len
, pgoff
, flags
);
1568 * Restore the topdown base:
1570 mm
->free_area_cache
= mm
->mmap_base
;
1571 mm
->cached_hole_size
= ~0UL;
1577 void arch_unmap_area_topdown(struct mm_struct
*mm
, unsigned long addr
)
1580 * Is this a new hole at the highest possible address?
1582 if (addr
> mm
->free_area_cache
)
1583 mm
->free_area_cache
= addr
;
1585 /* dont allow allocations above current base */
1586 if (mm
->free_area_cache
> mm
->mmap_base
)
1587 mm
->free_area_cache
= mm
->mmap_base
;
1591 get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
,
1592 unsigned long pgoff
, unsigned long flags
)
1594 unsigned long (*get_area
)(struct file
*, unsigned long,
1595 unsigned long, unsigned long, unsigned long);
1597 unsigned long error
= arch_mmap_check(addr
, len
, flags
);
1601 /* Careful about overflows.. */
1602 if (len
> TASK_SIZE
)
1605 get_area
= current
->mm
->get_unmapped_area
;
1606 if (file
&& file
->f_op
&& file
->f_op
->get_unmapped_area
)
1607 get_area
= file
->f_op
->get_unmapped_area
;
1608 addr
= get_area(file
, addr
, len
, pgoff
, flags
);
1609 if (IS_ERR_VALUE(addr
))
1612 if (addr
> TASK_SIZE
- len
)
1614 if (addr
& ~PAGE_MASK
)
1617 addr
= arch_rebalance_pgtables(addr
, len
);
1618 error
= security_mmap_addr(addr
);
1619 return error
? error
: addr
;
1622 EXPORT_SYMBOL(get_unmapped_area
);
1624 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1625 struct vm_area_struct
*find_vma(struct mm_struct
*mm
, unsigned long addr
)
1627 struct vm_area_struct
*vma
= NULL
;
1629 if (WARN_ON_ONCE(!mm
)) /* Remove this in linux-3.6 */
1632 /* Check the cache first. */
1633 /* (Cache hit rate is typically around 35%.) */
1634 vma
= mm
->mmap_cache
;
1635 if (!(vma
&& vma
->vm_end
> addr
&& vma
->vm_start
<= addr
)) {
1636 struct rb_node
*rb_node
;
1638 rb_node
= mm
->mm_rb
.rb_node
;
1642 struct vm_area_struct
*vma_tmp
;
1644 vma_tmp
= rb_entry(rb_node
,
1645 struct vm_area_struct
, vm_rb
);
1647 if (vma_tmp
->vm_end
> addr
) {
1649 if (vma_tmp
->vm_start
<= addr
)
1651 rb_node
= rb_node
->rb_left
;
1653 rb_node
= rb_node
->rb_right
;
1656 mm
->mmap_cache
= vma
;
1661 EXPORT_SYMBOL(find_vma
);
1664 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1666 struct vm_area_struct
*
1667 find_vma_prev(struct mm_struct
*mm
, unsigned long addr
,
1668 struct vm_area_struct
**pprev
)
1670 struct vm_area_struct
*vma
;
1672 vma
= find_vma(mm
, addr
);
1674 *pprev
= vma
->vm_prev
;
1676 struct rb_node
*rb_node
= mm
->mm_rb
.rb_node
;
1679 *pprev
= rb_entry(rb_node
, struct vm_area_struct
, vm_rb
);
1680 rb_node
= rb_node
->rb_right
;
1687 * Verify that the stack growth is acceptable and
1688 * update accounting. This is shared with both the
1689 * grow-up and grow-down cases.
1691 static int acct_stack_growth(struct vm_area_struct
*vma
, unsigned long size
, unsigned long grow
)
1693 struct mm_struct
*mm
= vma
->vm_mm
;
1694 struct rlimit
*rlim
= current
->signal
->rlim
;
1695 unsigned long new_start
;
1697 /* address space limit tests */
1698 if (!may_expand_vm(mm
, grow
))
1701 /* Stack limit test */
1702 if (size
> ACCESS_ONCE(rlim
[RLIMIT_STACK
].rlim_cur
))
1705 /* mlock limit tests */
1706 if (vma
->vm_flags
& VM_LOCKED
) {
1707 unsigned long locked
;
1708 unsigned long limit
;
1709 locked
= mm
->locked_vm
+ grow
;
1710 limit
= ACCESS_ONCE(rlim
[RLIMIT_MEMLOCK
].rlim_cur
);
1711 limit
>>= PAGE_SHIFT
;
1712 if (locked
> limit
&& !capable(CAP_IPC_LOCK
))
1716 /* Check to ensure the stack will not grow into a hugetlb-only region */
1717 new_start
= (vma
->vm_flags
& VM_GROWSUP
) ? vma
->vm_start
:
1719 if (is_hugepage_only_range(vma
->vm_mm
, new_start
, size
))
1723 * Overcommit.. This must be the final test, as it will
1724 * update security statistics.
1726 if (security_vm_enough_memory_mm(mm
, grow
))
1729 /* Ok, everything looks good - let it rip */
1730 if (vma
->vm_flags
& VM_LOCKED
)
1731 mm
->locked_vm
+= grow
;
1732 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, grow
);
1736 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1738 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1739 * vma is the last one with address > vma->vm_end. Have to extend vma.
1741 int expand_upwards(struct vm_area_struct
*vma
, unsigned long address
)
1745 if (!(vma
->vm_flags
& VM_GROWSUP
))
1749 * We must make sure the anon_vma is allocated
1750 * so that the anon_vma locking is not a noop.
1752 if (unlikely(anon_vma_prepare(vma
)))
1754 vma_lock_anon_vma(vma
);
1757 * vma->vm_start/vm_end cannot change under us because the caller
1758 * is required to hold the mmap_sem in read mode. We need the
1759 * anon_vma lock to serialize against concurrent expand_stacks.
1760 * Also guard against wrapping around to address 0.
1762 if (address
< PAGE_ALIGN(address
+4))
1763 address
= PAGE_ALIGN(address
+4);
1765 vma_unlock_anon_vma(vma
);
1770 /* Somebody else might have raced and expanded it already */
1771 if (address
> vma
->vm_end
) {
1772 unsigned long size
, grow
;
1774 size
= address
- vma
->vm_start
;
1775 grow
= (address
- vma
->vm_end
) >> PAGE_SHIFT
;
1778 if (vma
->vm_pgoff
+ (size
>> PAGE_SHIFT
) >= vma
->vm_pgoff
) {
1779 error
= acct_stack_growth(vma
, size
, grow
);
1781 anon_vma_interval_tree_pre_update_vma(vma
);
1782 vma
->vm_end
= address
;
1783 anon_vma_interval_tree_post_update_vma(vma
);
1784 perf_event_mmap(vma
);
1788 vma_unlock_anon_vma(vma
);
1789 khugepaged_enter_vma_merge(vma
);
1790 validate_mm(vma
->vm_mm
);
1793 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1796 * vma is the first one with address < vma->vm_start. Have to extend vma.
1798 int expand_downwards(struct vm_area_struct
*vma
,
1799 unsigned long address
)
1804 * We must make sure the anon_vma is allocated
1805 * so that the anon_vma locking is not a noop.
1807 if (unlikely(anon_vma_prepare(vma
)))
1810 address
&= PAGE_MASK
;
1811 error
= security_mmap_addr(address
);
1815 vma_lock_anon_vma(vma
);
1818 * vma->vm_start/vm_end cannot change under us because the caller
1819 * is required to hold the mmap_sem in read mode. We need the
1820 * anon_vma lock to serialize against concurrent expand_stacks.
1823 /* Somebody else might have raced and expanded it already */
1824 if (address
< vma
->vm_start
) {
1825 unsigned long size
, grow
;
1827 size
= vma
->vm_end
- address
;
1828 grow
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
1831 if (grow
<= vma
->vm_pgoff
) {
1832 error
= acct_stack_growth(vma
, size
, grow
);
1834 anon_vma_interval_tree_pre_update_vma(vma
);
1835 vma
->vm_start
= address
;
1836 vma
->vm_pgoff
-= grow
;
1837 anon_vma_interval_tree_post_update_vma(vma
);
1838 perf_event_mmap(vma
);
1842 vma_unlock_anon_vma(vma
);
1843 khugepaged_enter_vma_merge(vma
);
1844 validate_mm(vma
->vm_mm
);
1848 #ifdef CONFIG_STACK_GROWSUP
1849 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
1851 return expand_upwards(vma
, address
);
1854 struct vm_area_struct
*
1855 find_extend_vma(struct mm_struct
*mm
, unsigned long addr
)
1857 struct vm_area_struct
*vma
, *prev
;
1860 vma
= find_vma_prev(mm
, addr
, &prev
);
1861 if (vma
&& (vma
->vm_start
<= addr
))
1863 if (!prev
|| expand_stack(prev
, addr
))
1865 if (prev
->vm_flags
& VM_LOCKED
) {
1866 mlock_vma_pages_range(prev
, addr
, prev
->vm_end
);
1871 int expand_stack(struct vm_area_struct
*vma
, unsigned long address
)
1873 return expand_downwards(vma
, address
);
1876 struct vm_area_struct
*
1877 find_extend_vma(struct mm_struct
* mm
, unsigned long addr
)
1879 struct vm_area_struct
* vma
;
1880 unsigned long start
;
1883 vma
= find_vma(mm
,addr
);
1886 if (vma
->vm_start
<= addr
)
1888 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
1890 start
= vma
->vm_start
;
1891 if (expand_stack(vma
, addr
))
1893 if (vma
->vm_flags
& VM_LOCKED
) {
1894 mlock_vma_pages_range(vma
, addr
, start
);
1901 * Ok - we have the memory areas we should free on the vma list,
1902 * so release them, and do the vma updates.
1904 * Called with the mm semaphore held.
1906 static void remove_vma_list(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
1908 unsigned long nr_accounted
= 0;
1910 /* Update high watermark before we lower total_vm */
1911 update_hiwater_vm(mm
);
1913 long nrpages
= vma_pages(vma
);
1915 if (vma
->vm_flags
& VM_ACCOUNT
)
1916 nr_accounted
+= nrpages
;
1917 vm_stat_account(mm
, vma
->vm_flags
, vma
->vm_file
, -nrpages
);
1918 vma
= remove_vma(vma
);
1920 vm_unacct_memory(nr_accounted
);
1925 * Get rid of page table information in the indicated region.
1927 * Called with the mm semaphore held.
1929 static void unmap_region(struct mm_struct
*mm
,
1930 struct vm_area_struct
*vma
, struct vm_area_struct
*prev
,
1931 unsigned long start
, unsigned long end
)
1933 struct vm_area_struct
*next
= prev
? prev
->vm_next
: mm
->mmap
;
1934 struct mmu_gather tlb
;
1937 tlb_gather_mmu(&tlb
, mm
, 0);
1938 update_hiwater_rss(mm
);
1939 unmap_vmas(&tlb
, vma
, start
, end
);
1940 free_pgtables(&tlb
, vma
, prev
? prev
->vm_end
: FIRST_USER_ADDRESS
,
1941 next
? next
->vm_start
: 0);
1942 tlb_finish_mmu(&tlb
, start
, end
);
1946 * Create a list of vma's touched by the unmap, removing them from the mm's
1947 * vma list as we go..
1950 detach_vmas_to_be_unmapped(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1951 struct vm_area_struct
*prev
, unsigned long end
)
1953 struct vm_area_struct
**insertion_point
;
1954 struct vm_area_struct
*tail_vma
= NULL
;
1957 insertion_point
= (prev
? &prev
->vm_next
: &mm
->mmap
);
1958 vma
->vm_prev
= NULL
;
1960 rb_erase(&vma
->vm_rb
, &mm
->mm_rb
);
1964 } while (vma
&& vma
->vm_start
< end
);
1965 *insertion_point
= vma
;
1967 vma
->vm_prev
= prev
;
1968 tail_vma
->vm_next
= NULL
;
1969 if (mm
->unmap_area
== arch_unmap_area
)
1970 addr
= prev
? prev
->vm_end
: mm
->mmap_base
;
1972 addr
= vma
? vma
->vm_start
: mm
->mmap_base
;
1973 mm
->unmap_area(mm
, addr
);
1974 mm
->mmap_cache
= NULL
; /* Kill the cache. */
1978 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1979 * munmap path where it doesn't make sense to fail.
1981 static int __split_vma(struct mm_struct
* mm
, struct vm_area_struct
* vma
,
1982 unsigned long addr
, int new_below
)
1984 struct mempolicy
*pol
;
1985 struct vm_area_struct
*new;
1988 if (is_vm_hugetlb_page(vma
) && (addr
&
1989 ~(huge_page_mask(hstate_vma(vma
)))))
1992 new = kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
1996 /* most fields are the same, copy all, and then fixup */
1999 INIT_LIST_HEAD(&new->anon_vma_chain
);
2004 new->vm_start
= addr
;
2005 new->vm_pgoff
+= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
);
2008 pol
= mpol_dup(vma_policy(vma
));
2013 vma_set_policy(new, pol
);
2015 if (anon_vma_clone(new, vma
))
2019 get_file(new->vm_file
);
2021 if (new->vm_ops
&& new->vm_ops
->open
)
2022 new->vm_ops
->open(new);
2025 err
= vma_adjust(vma
, addr
, vma
->vm_end
, vma
->vm_pgoff
+
2026 ((addr
- new->vm_start
) >> PAGE_SHIFT
), new);
2028 err
= vma_adjust(vma
, vma
->vm_start
, addr
, vma
->vm_pgoff
, new);
2034 /* Clean everything up if vma_adjust failed. */
2035 if (new->vm_ops
&& new->vm_ops
->close
)
2036 new->vm_ops
->close(new);
2039 unlink_anon_vmas(new);
2043 kmem_cache_free(vm_area_cachep
, new);
2049 * Split a vma into two pieces at address 'addr', a new vma is allocated
2050 * either for the first part or the tail.
2052 int split_vma(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2053 unsigned long addr
, int new_below
)
2055 if (mm
->map_count
>= sysctl_max_map_count
)
2058 return __split_vma(mm
, vma
, addr
, new_below
);
2061 /* Munmap is split into 2 main parts -- this part which finds
2062 * what needs doing, and the areas themselves, which do the
2063 * work. This now handles partial unmappings.
2064 * Jeremy Fitzhardinge <jeremy@goop.org>
2066 int do_munmap(struct mm_struct
*mm
, unsigned long start
, size_t len
)
2069 struct vm_area_struct
*vma
, *prev
, *last
;
2071 if ((start
& ~PAGE_MASK
) || start
> TASK_SIZE
|| len
> TASK_SIZE
-start
)
2074 if ((len
= PAGE_ALIGN(len
)) == 0)
2077 /* Find the first overlapping VMA */
2078 vma
= find_vma(mm
, start
);
2081 prev
= vma
->vm_prev
;
2082 /* we have start < vma->vm_end */
2084 /* if it doesn't overlap, we have nothing.. */
2086 if (vma
->vm_start
>= end
)
2090 * If we need to split any vma, do it now to save pain later.
2092 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2093 * unmapped vm_area_struct will remain in use: so lower split_vma
2094 * places tmp vma above, and higher split_vma places tmp vma below.
2096 if (start
> vma
->vm_start
) {
2100 * Make sure that map_count on return from munmap() will
2101 * not exceed its limit; but let map_count go just above
2102 * its limit temporarily, to help free resources as expected.
2104 if (end
< vma
->vm_end
&& mm
->map_count
>= sysctl_max_map_count
)
2107 error
= __split_vma(mm
, vma
, start
, 0);
2113 /* Does it split the last one? */
2114 last
= find_vma(mm
, end
);
2115 if (last
&& end
> last
->vm_start
) {
2116 int error
= __split_vma(mm
, last
, end
, 1);
2120 vma
= prev
? prev
->vm_next
: mm
->mmap
;
2123 * unlock any mlock()ed ranges before detaching vmas
2125 if (mm
->locked_vm
) {
2126 struct vm_area_struct
*tmp
= vma
;
2127 while (tmp
&& tmp
->vm_start
< end
) {
2128 if (tmp
->vm_flags
& VM_LOCKED
) {
2129 mm
->locked_vm
-= vma_pages(tmp
);
2130 munlock_vma_pages_all(tmp
);
2137 * Remove the vma's, and unmap the actual pages
2139 detach_vmas_to_be_unmapped(mm
, vma
, prev
, end
);
2140 unmap_region(mm
, vma
, prev
, start
, end
);
2142 /* Fix up all other VM information */
2143 remove_vma_list(mm
, vma
);
2148 int vm_munmap(unsigned long start
, size_t len
)
2151 struct mm_struct
*mm
= current
->mm
;
2153 down_write(&mm
->mmap_sem
);
2154 ret
= do_munmap(mm
, start
, len
);
2155 up_write(&mm
->mmap_sem
);
2158 EXPORT_SYMBOL(vm_munmap
);
2160 SYSCALL_DEFINE2(munmap
, unsigned long, addr
, size_t, len
)
2162 profile_munmap(addr
);
2163 return vm_munmap(addr
, len
);
2166 static inline void verify_mm_writelocked(struct mm_struct
*mm
)
2168 #ifdef CONFIG_DEBUG_VM
2169 if (unlikely(down_read_trylock(&mm
->mmap_sem
))) {
2171 up_read(&mm
->mmap_sem
);
2177 * this is really a simplified "do_mmap". it only handles
2178 * anonymous maps. eventually we may be able to do some
2179 * brk-specific accounting here.
2181 static unsigned long do_brk(unsigned long addr
, unsigned long len
)
2183 struct mm_struct
* mm
= current
->mm
;
2184 struct vm_area_struct
* vma
, * prev
;
2185 unsigned long flags
;
2186 struct rb_node
** rb_link
, * rb_parent
;
2187 pgoff_t pgoff
= addr
>> PAGE_SHIFT
;
2190 len
= PAGE_ALIGN(len
);
2194 flags
= VM_DATA_DEFAULT_FLAGS
| VM_ACCOUNT
| mm
->def_flags
;
2196 error
= get_unmapped_area(NULL
, addr
, len
, 0, MAP_FIXED
);
2197 if (error
& ~PAGE_MASK
)
2203 if (mm
->def_flags
& VM_LOCKED
) {
2204 unsigned long locked
, lock_limit
;
2205 locked
= len
>> PAGE_SHIFT
;
2206 locked
+= mm
->locked_vm
;
2207 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
2208 lock_limit
>>= PAGE_SHIFT
;
2209 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
2214 * mm->mmap_sem is required to protect against another thread
2215 * changing the mappings in case we sleep.
2217 verify_mm_writelocked(mm
);
2220 * Clear old maps. this also does some error checking for us
2223 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
)) {
2224 if (do_munmap(mm
, addr
, len
))
2229 /* Check against address space limits *after* clearing old maps... */
2230 if (!may_expand_vm(mm
, len
>> PAGE_SHIFT
))
2233 if (mm
->map_count
> sysctl_max_map_count
)
2236 if (security_vm_enough_memory_mm(mm
, len
>> PAGE_SHIFT
))
2239 /* Can we just expand an old private anonymous mapping? */
2240 vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, flags
,
2241 NULL
, NULL
, pgoff
, NULL
);
2246 * create a vma struct for an anonymous mapping
2248 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2250 vm_unacct_memory(len
>> PAGE_SHIFT
);
2254 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2256 vma
->vm_start
= addr
;
2257 vma
->vm_end
= addr
+ len
;
2258 vma
->vm_pgoff
= pgoff
;
2259 vma
->vm_flags
= flags
;
2260 vma
->vm_page_prot
= vm_get_page_prot(flags
);
2261 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2263 perf_event_mmap(vma
);
2264 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2265 if (flags
& VM_LOCKED
) {
2266 if (!mlock_vma_pages_range(vma
, addr
, addr
+ len
))
2267 mm
->locked_vm
+= (len
>> PAGE_SHIFT
);
2272 unsigned long vm_brk(unsigned long addr
, unsigned long len
)
2274 struct mm_struct
*mm
= current
->mm
;
2277 down_write(&mm
->mmap_sem
);
2278 ret
= do_brk(addr
, len
);
2279 up_write(&mm
->mmap_sem
);
2282 EXPORT_SYMBOL(vm_brk
);
2284 /* Release all mmaps. */
2285 void exit_mmap(struct mm_struct
*mm
)
2287 struct mmu_gather tlb
;
2288 struct vm_area_struct
*vma
;
2289 unsigned long nr_accounted
= 0;
2291 /* mm's last user has gone, and its about to be pulled down */
2292 mmu_notifier_release(mm
);
2294 if (mm
->locked_vm
) {
2297 if (vma
->vm_flags
& VM_LOCKED
)
2298 munlock_vma_pages_all(vma
);
2306 if (!vma
) /* Can happen if dup_mmap() received an OOM */
2311 tlb_gather_mmu(&tlb
, mm
, 1);
2312 /* update_hiwater_rss(mm) here? but nobody should be looking */
2313 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2314 unmap_vmas(&tlb
, vma
, 0, -1);
2316 free_pgtables(&tlb
, vma
, FIRST_USER_ADDRESS
, 0);
2317 tlb_finish_mmu(&tlb
, 0, -1);
2320 * Walk the list again, actually closing and freeing it,
2321 * with preemption enabled, without holding any MM locks.
2324 if (vma
->vm_flags
& VM_ACCOUNT
)
2325 nr_accounted
+= vma_pages(vma
);
2326 vma
= remove_vma(vma
);
2328 vm_unacct_memory(nr_accounted
);
2330 WARN_ON(mm
->nr_ptes
> (FIRST_USER_ADDRESS
+PMD_SIZE
-1)>>PMD_SHIFT
);
2333 /* Insert vm structure into process list sorted by address
2334 * and into the inode's i_mmap tree. If vm_file is non-NULL
2335 * then i_mmap_mutex is taken here.
2337 int insert_vm_struct(struct mm_struct
*mm
, struct vm_area_struct
*vma
)
2339 struct vm_area_struct
*prev
;
2340 struct rb_node
**rb_link
, *rb_parent
;
2343 * The vm_pgoff of a purely anonymous vma should be irrelevant
2344 * until its first write fault, when page's anon_vma and index
2345 * are set. But now set the vm_pgoff it will almost certainly
2346 * end up with (unless mremap moves it elsewhere before that
2347 * first wfault), so /proc/pid/maps tells a consistent story.
2349 * By setting it to reflect the virtual start address of the
2350 * vma, merges and splits can happen in a seamless way, just
2351 * using the existing file pgoff checks and manipulations.
2352 * Similarly in do_mmap_pgoff and in do_brk.
2354 if (!vma
->vm_file
) {
2355 BUG_ON(vma
->anon_vma
);
2356 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2358 if (find_vma_links(mm
, vma
->vm_start
, vma
->vm_end
,
2359 &prev
, &rb_link
, &rb_parent
))
2361 if ((vma
->vm_flags
& VM_ACCOUNT
) &&
2362 security_vm_enough_memory_mm(mm
, vma_pages(vma
)))
2365 vma_link(mm
, vma
, prev
, rb_link
, rb_parent
);
2370 * Copy the vma structure to a new location in the same mm,
2371 * prior to moving page table entries, to effect an mremap move.
2373 struct vm_area_struct
*copy_vma(struct vm_area_struct
**vmap
,
2374 unsigned long addr
, unsigned long len
, pgoff_t pgoff
)
2376 struct vm_area_struct
*vma
= *vmap
;
2377 unsigned long vma_start
= vma
->vm_start
;
2378 struct mm_struct
*mm
= vma
->vm_mm
;
2379 struct vm_area_struct
*new_vma
, *prev
;
2380 struct rb_node
**rb_link
, *rb_parent
;
2381 struct mempolicy
*pol
;
2382 bool faulted_in_anon_vma
= true;
2385 * If anonymous vma has not yet been faulted, update new pgoff
2386 * to match new location, to increase its chance of merging.
2388 if (unlikely(!vma
->vm_file
&& !vma
->anon_vma
)) {
2389 pgoff
= addr
>> PAGE_SHIFT
;
2390 faulted_in_anon_vma
= false;
2393 if (find_vma_links(mm
, addr
, addr
+ len
, &prev
, &rb_link
, &rb_parent
))
2394 return NULL
; /* should never get here */
2395 new_vma
= vma_merge(mm
, prev
, addr
, addr
+ len
, vma
->vm_flags
,
2396 vma
->anon_vma
, vma
->vm_file
, pgoff
, vma_policy(vma
));
2399 * Source vma may have been merged into new_vma
2401 if (unlikely(vma_start
>= new_vma
->vm_start
&&
2402 vma_start
< new_vma
->vm_end
)) {
2404 * The only way we can get a vma_merge with
2405 * self during an mremap is if the vma hasn't
2406 * been faulted in yet and we were allowed to
2407 * reset the dst vma->vm_pgoff to the
2408 * destination address of the mremap to allow
2409 * the merge to happen. mremap must change the
2410 * vm_pgoff linearity between src and dst vmas
2411 * (in turn preventing a vma_merge) to be
2412 * safe. It is only safe to keep the vm_pgoff
2413 * linear if there are no pages mapped yet.
2415 VM_BUG_ON(faulted_in_anon_vma
);
2419 new_vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
2422 new_vma
->vm_start
= addr
;
2423 new_vma
->vm_end
= addr
+ len
;
2424 new_vma
->vm_pgoff
= pgoff
;
2425 pol
= mpol_dup(vma_policy(vma
));
2428 vma_set_policy(new_vma
, pol
);
2429 INIT_LIST_HEAD(&new_vma
->anon_vma_chain
);
2430 if (anon_vma_clone(new_vma
, vma
))
2431 goto out_free_mempol
;
2432 if (new_vma
->vm_file
)
2433 get_file(new_vma
->vm_file
);
2434 if (new_vma
->vm_ops
&& new_vma
->vm_ops
->open
)
2435 new_vma
->vm_ops
->open(new_vma
);
2436 vma_link(mm
, new_vma
, prev
, rb_link
, rb_parent
);
2444 kmem_cache_free(vm_area_cachep
, new_vma
);
2449 * Return true if the calling process may expand its vm space by the passed
2452 int may_expand_vm(struct mm_struct
*mm
, unsigned long npages
)
2454 unsigned long cur
= mm
->total_vm
; /* pages */
2457 lim
= rlimit(RLIMIT_AS
) >> PAGE_SHIFT
;
2459 if (cur
+ npages
> lim
)
2465 static int special_mapping_fault(struct vm_area_struct
*vma
,
2466 struct vm_fault
*vmf
)
2469 struct page
**pages
;
2472 * special mappings have no vm_file, and in that case, the mm
2473 * uses vm_pgoff internally. So we have to subtract it from here.
2474 * We are allowed to do this because we are the mm; do not copy
2475 * this code into drivers!
2477 pgoff
= vmf
->pgoff
- vma
->vm_pgoff
;
2479 for (pages
= vma
->vm_private_data
; pgoff
&& *pages
; ++pages
)
2483 struct page
*page
= *pages
;
2489 return VM_FAULT_SIGBUS
;
2493 * Having a close hook prevents vma merging regardless of flags.
2495 static void special_mapping_close(struct vm_area_struct
*vma
)
2499 static const struct vm_operations_struct special_mapping_vmops
= {
2500 .close
= special_mapping_close
,
2501 .fault
= special_mapping_fault
,
2505 * Called with mm->mmap_sem held for writing.
2506 * Insert a new vma covering the given region, with the given flags.
2507 * Its pages are supplied by the given array of struct page *.
2508 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2509 * The region past the last page supplied will always produce SIGBUS.
2510 * The array pointer and the pages it points to are assumed to stay alive
2511 * for as long as this mapping might exist.
2513 int install_special_mapping(struct mm_struct
*mm
,
2514 unsigned long addr
, unsigned long len
,
2515 unsigned long vm_flags
, struct page
**pages
)
2518 struct vm_area_struct
*vma
;
2520 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2521 if (unlikely(vma
== NULL
))
2524 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2526 vma
->vm_start
= addr
;
2527 vma
->vm_end
= addr
+ len
;
2529 vma
->vm_flags
= vm_flags
| mm
->def_flags
| VM_DONTEXPAND
;
2530 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
2532 vma
->vm_ops
= &special_mapping_vmops
;
2533 vma
->vm_private_data
= pages
;
2535 ret
= insert_vm_struct(mm
, vma
);
2539 mm
->total_vm
+= len
>> PAGE_SHIFT
;
2541 perf_event_mmap(vma
);
2546 kmem_cache_free(vm_area_cachep
, vma
);
2550 static DEFINE_MUTEX(mm_all_locks_mutex
);
2552 static void vm_lock_anon_vma(struct mm_struct
*mm
, struct anon_vma
*anon_vma
)
2554 if (!test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
2556 * The LSB of head.next can't change from under us
2557 * because we hold the mm_all_locks_mutex.
2559 mutex_lock_nest_lock(&anon_vma
->root
->mutex
, &mm
->mmap_sem
);
2561 * We can safely modify head.next after taking the
2562 * anon_vma->root->mutex. If some other vma in this mm shares
2563 * the same anon_vma we won't take it again.
2565 * No need of atomic instructions here, head.next
2566 * can't change from under us thanks to the
2567 * anon_vma->root->mutex.
2569 if (__test_and_set_bit(0, (unsigned long *)
2570 &anon_vma
->root
->rb_root
.rb_node
))
2575 static void vm_lock_mapping(struct mm_struct
*mm
, struct address_space
*mapping
)
2577 if (!test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
2579 * AS_MM_ALL_LOCKS can't change from under us because
2580 * we hold the mm_all_locks_mutex.
2582 * Operations on ->flags have to be atomic because
2583 * even if AS_MM_ALL_LOCKS is stable thanks to the
2584 * mm_all_locks_mutex, there may be other cpus
2585 * changing other bitflags in parallel to us.
2587 if (test_and_set_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
))
2589 mutex_lock_nest_lock(&mapping
->i_mmap_mutex
, &mm
->mmap_sem
);
2594 * This operation locks against the VM for all pte/vma/mm related
2595 * operations that could ever happen on a certain mm. This includes
2596 * vmtruncate, try_to_unmap, and all page faults.
2598 * The caller must take the mmap_sem in write mode before calling
2599 * mm_take_all_locks(). The caller isn't allowed to release the
2600 * mmap_sem until mm_drop_all_locks() returns.
2602 * mmap_sem in write mode is required in order to block all operations
2603 * that could modify pagetables and free pages without need of
2604 * altering the vma layout (for example populate_range() with
2605 * nonlinear vmas). It's also needed in write mode to avoid new
2606 * anon_vmas to be associated with existing vmas.
2608 * A single task can't take more than one mm_take_all_locks() in a row
2609 * or it would deadlock.
2611 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
2612 * mapping->flags avoid to take the same lock twice, if more than one
2613 * vma in this mm is backed by the same anon_vma or address_space.
2615 * We can take all the locks in random order because the VM code
2616 * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2617 * takes more than one of them in a row. Secondly we're protected
2618 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2620 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2621 * that may have to take thousand of locks.
2623 * mm_take_all_locks() can fail if it's interrupted by signals.
2625 int mm_take_all_locks(struct mm_struct
*mm
)
2627 struct vm_area_struct
*vma
;
2628 struct anon_vma_chain
*avc
;
2630 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
2632 mutex_lock(&mm_all_locks_mutex
);
2634 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
2635 if (signal_pending(current
))
2637 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
2638 vm_lock_mapping(mm
, vma
->vm_file
->f_mapping
);
2641 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
2642 if (signal_pending(current
))
2645 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
2646 vm_lock_anon_vma(mm
, avc
->anon_vma
);
2652 mm_drop_all_locks(mm
);
2656 static void vm_unlock_anon_vma(struct anon_vma
*anon_vma
)
2658 if (test_bit(0, (unsigned long *) &anon_vma
->root
->rb_root
.rb_node
)) {
2660 * The LSB of head.next can't change to 0 from under
2661 * us because we hold the mm_all_locks_mutex.
2663 * We must however clear the bitflag before unlocking
2664 * the vma so the users using the anon_vma->rb_root will
2665 * never see our bitflag.
2667 * No need of atomic instructions here, head.next
2668 * can't change from under us until we release the
2669 * anon_vma->root->mutex.
2671 if (!__test_and_clear_bit(0, (unsigned long *)
2672 &anon_vma
->root
->rb_root
.rb_node
))
2674 anon_vma_unlock(anon_vma
);
2678 static void vm_unlock_mapping(struct address_space
*mapping
)
2680 if (test_bit(AS_MM_ALL_LOCKS
, &mapping
->flags
)) {
2682 * AS_MM_ALL_LOCKS can't change to 0 from under us
2683 * because we hold the mm_all_locks_mutex.
2685 mutex_unlock(&mapping
->i_mmap_mutex
);
2686 if (!test_and_clear_bit(AS_MM_ALL_LOCKS
,
2693 * The mmap_sem cannot be released by the caller until
2694 * mm_drop_all_locks() returns.
2696 void mm_drop_all_locks(struct mm_struct
*mm
)
2698 struct vm_area_struct
*vma
;
2699 struct anon_vma_chain
*avc
;
2701 BUG_ON(down_read_trylock(&mm
->mmap_sem
));
2702 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex
));
2704 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
2706 list_for_each_entry(avc
, &vma
->anon_vma_chain
, same_vma
)
2707 vm_unlock_anon_vma(avc
->anon_vma
);
2708 if (vma
->vm_file
&& vma
->vm_file
->f_mapping
)
2709 vm_unlock_mapping(vma
->vm_file
->f_mapping
);
2712 mutex_unlock(&mm_all_locks_mutex
);
2716 * initialise the VMA slab
2718 void __init
mmap_init(void)
2722 ret
= percpu_counter_init(&vm_committed_as
, 0);