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>
34 #include <linux/rbtree_augmented.h>
35 #include <linux/sched/sysctl.h>
37 #include <asm/uaccess.h>
38 #include <asm/cacheflush.h>
40 #include <asm/mmu_context.h>
44 #ifndef arch_mmap_check
45 #define arch_mmap_check(addr, len, flags) (0)
48 #ifndef arch_rebalance_pgtables
49 #define arch_rebalance_pgtables(addr, len) (addr)
52 static void unmap_region(struct mm_struct *mm,
53 struct vm_area_struct *vma, struct vm_area_struct *prev,
54 unsigned long start, unsigned long end);
56 /* description of effects of mapping type and prot in current implementation.
57 * this is due to the limited x86 page protection hardware. The expected
58 * behavior is in parens:
61 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
62 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
63 * w: (no) no w: (no) no w: (yes) yes w: (no) no
64 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
66 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
67 * w: (no) no w: (no) no w: (copy) copy w: (no) no
68 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
71 pgprot_t protection_map[16] = {
72 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
73 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
76 pgprot_t vm_get_page_prot(unsigned long vm_flags)
78 return __pgprot(pgprot_val(protection_map[vm_flags &
79 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
80 pgprot_val(arch_vm_get_page_prot(vm_flags)));
82 EXPORT_SYMBOL(vm_get_page_prot);
84 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */
85 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
86 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
88 * Make sure vm_committed_as in one cacheline and not cacheline shared with
89 * other variables. It can be updated by several CPUs frequently.
91 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
94 * The global memory commitment made in the system can be a metric
95 * that can be used to drive ballooning decisions when Linux is hosted
96 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
97 * balancing memory across competing virtual machines that are hosted.
98 * Several metrics drive this policy engine including the guest reported
101 unsigned long vm_memory_committed(void)
103 return percpu_counter_read_positive(&vm_committed_as);
105 EXPORT_SYMBOL_GPL(vm_memory_committed);
108 * Check that a process has enough memory to allocate a new virtual
109 * mapping. 0 means there is enough memory for the allocation to
110 * succeed and -ENOMEM implies there is not.
112 * We currently support three overcommit policies, which are set via the
113 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
115 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
116 * Additional code 2002 Jul 20 by Robert Love.
118 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
120 * Note this is a helper function intended to be used by LSMs which
121 * wish to use this logic.
123 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
125 unsigned long free, allowed;
127 vm_acct_memory(pages);
130 * Sometimes we want to use more memory than we have
132 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
135 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
136 free = global_page_state(NR_FREE_PAGES);
137 free += global_page_state(NR_FILE_PAGES);
140 * shmem pages shouldn't be counted as free in this
141 * case, they can't be purged, only swapped out, and
142 * that won't affect the overall amount of available
143 * memory in the system.
145 free -= global_page_state(NR_SHMEM);
147 free += nr_swap_pages;
150 * Any slabs which are created with the
151 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
152 * which are reclaimable, under pressure. The dentry
153 * cache and most inode caches should fall into this
155 free += global_page_state(NR_SLAB_RECLAIMABLE);
158 * Leave reserved pages. The pages are not for anonymous pages.
160 if (free <= totalreserve_pages)
163 free -= totalreserve_pages;
166 * Leave the last 3% for root
177 allowed = (totalram_pages - hugetlb_total_pages())
178 * sysctl_overcommit_ratio / 100;
180 * Leave the last 3% for root
183 allowed -= allowed / 32;
184 allowed += total_swap_pages;
186 /* Don't let a single process grow too big:
187 leave 3% of the size of this process for other processes */
189 allowed -= mm->total_vm / 32;
191 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
194 vm_unacct_memory(pages);
200 * Requires inode->i_mapping->i_mmap_mutex
202 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
203 struct file *file, struct address_space *mapping)
205 if (vma->vm_flags & VM_DENYWRITE)
206 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
207 if (vma->vm_flags & VM_SHARED)
208 mapping->i_mmap_writable--;
210 flush_dcache_mmap_lock(mapping);
211 if (unlikely(vma->vm_flags & VM_NONLINEAR))
212 list_del_init(&vma->shared.nonlinear);
214 vma_interval_tree_remove(vma, &mapping->i_mmap);
215 flush_dcache_mmap_unlock(mapping);
219 * Unlink a file-based vm structure from its interval tree, to hide
220 * vma from rmap and vmtruncate before freeing its page tables.
222 void unlink_file_vma(struct vm_area_struct *vma)
224 struct file *file = vma->vm_file;
227 struct address_space *mapping = file->f_mapping;
228 mutex_lock(&mapping->i_mmap_mutex);
229 __remove_shared_vm_struct(vma, file, mapping);
230 mutex_unlock(&mapping->i_mmap_mutex);
235 * Close a vm structure and free it, returning the next.
237 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
239 struct vm_area_struct *next = vma->vm_next;
242 if (vma->vm_ops && vma->vm_ops->close)
243 vma->vm_ops->close(vma);
246 mpol_put(vma_policy(vma));
247 kmem_cache_free(vm_area_cachep, vma);
251 static unsigned long do_brk(unsigned long addr, unsigned long len);
253 SYSCALL_DEFINE1(brk, unsigned long, brk)
255 unsigned long rlim, retval;
256 unsigned long newbrk, oldbrk;
257 struct mm_struct *mm = current->mm;
258 unsigned long min_brk;
260 down_write(&mm->mmap_sem);
262 #ifdef CONFIG_COMPAT_BRK
264 * CONFIG_COMPAT_BRK can still be overridden by setting
265 * randomize_va_space to 2, which will still cause mm->start_brk
266 * to be arbitrarily shifted
268 if (current->brk_randomized)
269 min_brk = mm->start_brk;
271 min_brk = mm->end_data;
273 min_brk = mm->start_brk;
279 * Check against rlimit here. If this check is done later after the test
280 * of oldbrk with newbrk then it can escape the test and let the data
281 * segment grow beyond its set limit the in case where the limit is
282 * not page aligned -Ram Gupta
284 rlim = rlimit(RLIMIT_DATA);
285 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
286 (mm->end_data - mm->start_data) > rlim)
289 newbrk = PAGE_ALIGN(brk);
290 oldbrk = PAGE_ALIGN(mm->brk);
291 if (oldbrk == newbrk)
294 /* Always allow shrinking brk. */
295 if (brk <= mm->brk) {
296 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
301 /* Check against existing mmap mappings. */
302 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
305 /* Ok, looks good - let it rip. */
306 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
312 up_write(&mm->mmap_sem);
316 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
318 unsigned long max, subtree_gap;
321 max -= vma->vm_prev->vm_end;
322 if (vma->vm_rb.rb_left) {
323 subtree_gap = rb_entry(vma->vm_rb.rb_left,
324 struct vm_area_struct, vm_rb)->rb_subtree_gap;
325 if (subtree_gap > max)
328 if (vma->vm_rb.rb_right) {
329 subtree_gap = rb_entry(vma->vm_rb.rb_right,
330 struct vm_area_struct, vm_rb)->rb_subtree_gap;
331 if (subtree_gap > max)
337 #ifdef CONFIG_DEBUG_VM_RB
338 static int browse_rb(struct rb_root *root)
340 int i = 0, j, bug = 0;
341 struct rb_node *nd, *pn = NULL;
342 unsigned long prev = 0, pend = 0;
344 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
345 struct vm_area_struct *vma;
346 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
347 if (vma->vm_start < prev) {
348 printk("vm_start %lx prev %lx\n", vma->vm_start, prev);
351 if (vma->vm_start < pend) {
352 printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
355 if (vma->vm_start > vma->vm_end) {
356 printk("vm_end %lx < vm_start %lx\n",
357 vma->vm_end, vma->vm_start);
360 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
361 printk("free gap %lx, correct %lx\n",
363 vma_compute_subtree_gap(vma));
368 prev = vma->vm_start;
372 for (nd = pn; nd; nd = rb_prev(nd))
375 printk("backwards %d, forwards %d\n", j, i);
381 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
385 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
386 struct vm_area_struct *vma;
387 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
388 BUG_ON(vma != ignore &&
389 vma->rb_subtree_gap != vma_compute_subtree_gap(vma));
393 void validate_mm(struct mm_struct *mm)
397 unsigned long highest_address = 0;
398 struct vm_area_struct *vma = mm->mmap;
400 struct anon_vma_chain *avc;
401 vma_lock_anon_vma(vma);
402 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
403 anon_vma_interval_tree_verify(avc);
404 vma_unlock_anon_vma(vma);
405 highest_address = vma->vm_end;
409 if (i != mm->map_count) {
410 printk("map_count %d vm_next %d\n", mm->map_count, i);
413 if (highest_address != mm->highest_vm_end) {
414 printk("mm->highest_vm_end %lx, found %lx\n",
415 mm->highest_vm_end, highest_address);
418 i = browse_rb(&mm->mm_rb);
419 if (i != mm->map_count) {
420 printk("map_count %d rb %d\n", mm->map_count, i);
426 #define validate_mm_rb(root, ignore) do { } while (0)
427 #define validate_mm(mm) do { } while (0)
430 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
431 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
434 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
435 * vma->vm_prev->vm_end values changed, without modifying the vma's position
438 static void vma_gap_update(struct vm_area_struct *vma)
441 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
442 * function that does exacltly what we want.
444 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
447 static inline void vma_rb_insert(struct vm_area_struct *vma,
448 struct rb_root *root)
450 /* All rb_subtree_gap values must be consistent prior to insertion */
451 validate_mm_rb(root, NULL);
453 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
456 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
459 * All rb_subtree_gap values must be consistent prior to erase,
460 * with the possible exception of the vma being erased.
462 validate_mm_rb(root, vma);
465 * Note rb_erase_augmented is a fairly large inline function,
466 * so make sure we instantiate it only once with our desired
467 * augmented rbtree callbacks.
469 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
473 * vma has some anon_vma assigned, and is already inserted on that
474 * anon_vma's interval trees.
476 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
477 * vma must be removed from the anon_vma's interval trees using
478 * anon_vma_interval_tree_pre_update_vma().
480 * After the update, the vma will be reinserted using
481 * anon_vma_interval_tree_post_update_vma().
483 * The entire update must be protected by exclusive mmap_sem and by
484 * the root anon_vma's mutex.
487 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
489 struct anon_vma_chain *avc;
491 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
492 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
496 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
498 struct anon_vma_chain *avc;
500 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
501 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
504 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
505 unsigned long end, struct vm_area_struct **pprev,
506 struct rb_node ***rb_link, struct rb_node **rb_parent)
508 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
510 __rb_link = &mm->mm_rb.rb_node;
511 rb_prev = __rb_parent = NULL;
514 struct vm_area_struct *vma_tmp;
516 __rb_parent = *__rb_link;
517 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
519 if (vma_tmp->vm_end > addr) {
520 /* Fail if an existing vma overlaps the area */
521 if (vma_tmp->vm_start < end)
523 __rb_link = &__rb_parent->rb_left;
525 rb_prev = __rb_parent;
526 __rb_link = &__rb_parent->rb_right;
532 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
533 *rb_link = __rb_link;
534 *rb_parent = __rb_parent;
538 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
539 struct rb_node **rb_link, struct rb_node *rb_parent)
541 /* Update tracking information for the gap following the new vma. */
543 vma_gap_update(vma->vm_next);
545 mm->highest_vm_end = vma->vm_end;
548 * vma->vm_prev wasn't known when we followed the rbtree to find the
549 * correct insertion point for that vma. As a result, we could not
550 * update the vma vm_rb parents rb_subtree_gap values on the way down.
551 * So, we first insert the vma with a zero rb_subtree_gap value
552 * (to be consistent with what we did on the way down), and then
553 * immediately update the gap to the correct value. Finally we
554 * rebalance the rbtree after all augmented values have been set.
556 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
557 vma->rb_subtree_gap = 0;
559 vma_rb_insert(vma, &mm->mm_rb);
562 static void __vma_link_file(struct vm_area_struct *vma)
568 struct address_space *mapping = file->f_mapping;
570 if (vma->vm_flags & VM_DENYWRITE)
571 atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
572 if (vma->vm_flags & VM_SHARED)
573 mapping->i_mmap_writable++;
575 flush_dcache_mmap_lock(mapping);
576 if (unlikely(vma->vm_flags & VM_NONLINEAR))
577 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
579 vma_interval_tree_insert(vma, &mapping->i_mmap);
580 flush_dcache_mmap_unlock(mapping);
585 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
586 struct vm_area_struct *prev, struct rb_node **rb_link,
587 struct rb_node *rb_parent)
589 __vma_link_list(mm, vma, prev, rb_parent);
590 __vma_link_rb(mm, vma, rb_link, rb_parent);
593 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
594 struct vm_area_struct *prev, struct rb_node **rb_link,
595 struct rb_node *rb_parent)
597 struct address_space *mapping = NULL;
600 mapping = vma->vm_file->f_mapping;
603 mutex_lock(&mapping->i_mmap_mutex);
605 __vma_link(mm, vma, prev, rb_link, rb_parent);
606 __vma_link_file(vma);
609 mutex_unlock(&mapping->i_mmap_mutex);
616 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
617 * mm's list and rbtree. It has already been inserted into the interval tree.
619 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
621 struct vm_area_struct *prev;
622 struct rb_node **rb_link, *rb_parent;
624 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
625 &prev, &rb_link, &rb_parent))
627 __vma_link(mm, vma, prev, rb_link, rb_parent);
632 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
633 struct vm_area_struct *prev)
635 struct vm_area_struct *next;
637 vma_rb_erase(vma, &mm->mm_rb);
638 prev->vm_next = next = vma->vm_next;
640 next->vm_prev = prev;
641 if (mm->mmap_cache == vma)
642 mm->mmap_cache = prev;
646 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
647 * is already present in an i_mmap tree without adjusting the tree.
648 * The following helper function should be used when such adjustments
649 * are necessary. The "insert" vma (if any) is to be inserted
650 * before we drop the necessary locks.
652 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
653 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
655 struct mm_struct *mm = vma->vm_mm;
656 struct vm_area_struct *next = vma->vm_next;
657 struct vm_area_struct *importer = NULL;
658 struct address_space *mapping = NULL;
659 struct rb_root *root = NULL;
660 struct anon_vma *anon_vma = NULL;
661 struct file *file = vma->vm_file;
662 bool start_changed = false, end_changed = false;
663 long adjust_next = 0;
666 if (next && !insert) {
667 struct vm_area_struct *exporter = NULL;
669 if (end >= next->vm_end) {
671 * vma expands, overlapping all the next, and
672 * perhaps the one after too (mprotect case 6).
674 again: remove_next = 1 + (end > next->vm_end);
678 } else if (end > next->vm_start) {
680 * vma expands, overlapping part of the next:
681 * mprotect case 5 shifting the boundary up.
683 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
686 } else if (end < vma->vm_end) {
688 * vma shrinks, and !insert tells it's not
689 * split_vma inserting another: so it must be
690 * mprotect case 4 shifting the boundary down.
692 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
698 * Easily overlooked: when mprotect shifts the boundary,
699 * make sure the expanding vma has anon_vma set if the
700 * shrinking vma had, to cover any anon pages imported.
702 if (exporter && exporter->anon_vma && !importer->anon_vma) {
703 if (anon_vma_clone(importer, exporter))
705 importer->anon_vma = exporter->anon_vma;
710 mapping = file->f_mapping;
711 if (!(vma->vm_flags & VM_NONLINEAR)) {
712 root = &mapping->i_mmap;
713 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
716 uprobe_munmap(next, next->vm_start,
720 mutex_lock(&mapping->i_mmap_mutex);
723 * Put into interval tree now, so instantiated pages
724 * are visible to arm/parisc __flush_dcache_page
725 * throughout; but we cannot insert into address
726 * space until vma start or end is updated.
728 __vma_link_file(insert);
732 vma_adjust_trans_huge(vma, start, end, adjust_next);
734 anon_vma = vma->anon_vma;
735 if (!anon_vma && adjust_next)
736 anon_vma = next->anon_vma;
738 VM_BUG_ON(adjust_next && next->anon_vma &&
739 anon_vma != next->anon_vma);
740 anon_vma_lock_write(anon_vma);
741 anon_vma_interval_tree_pre_update_vma(vma);
743 anon_vma_interval_tree_pre_update_vma(next);
747 flush_dcache_mmap_lock(mapping);
748 vma_interval_tree_remove(vma, root);
750 vma_interval_tree_remove(next, root);
753 if (start != vma->vm_start) {
754 vma->vm_start = start;
755 start_changed = true;
757 if (end != vma->vm_end) {
761 vma->vm_pgoff = pgoff;
763 next->vm_start += adjust_next << PAGE_SHIFT;
764 next->vm_pgoff += adjust_next;
769 vma_interval_tree_insert(next, root);
770 vma_interval_tree_insert(vma, root);
771 flush_dcache_mmap_unlock(mapping);
776 * vma_merge has merged next into vma, and needs
777 * us to remove next before dropping the locks.
779 __vma_unlink(mm, next, vma);
781 __remove_shared_vm_struct(next, file, mapping);
784 * split_vma has split insert from vma, and needs
785 * us to insert it before dropping the locks
786 * (it may either follow vma or precede it).
788 __insert_vm_struct(mm, insert);
794 mm->highest_vm_end = end;
795 else if (!adjust_next)
796 vma_gap_update(next);
801 anon_vma_interval_tree_post_update_vma(vma);
803 anon_vma_interval_tree_post_update_vma(next);
804 anon_vma_unlock(anon_vma);
807 mutex_unlock(&mapping->i_mmap_mutex);
818 uprobe_munmap(next, next->vm_start, next->vm_end);
822 anon_vma_merge(vma, next);
824 mpol_put(vma_policy(next));
825 kmem_cache_free(vm_area_cachep, next);
827 * In mprotect's case 6 (see comments on vma_merge),
828 * we must remove another next too. It would clutter
829 * up the code too much to do both in one go.
832 if (remove_next == 2)
835 vma_gap_update(next);
837 mm->highest_vm_end = end;
848 * If the vma has a ->close operation then the driver probably needs to release
849 * per-vma resources, so we don't attempt to merge those.
851 static inline int is_mergeable_vma(struct vm_area_struct *vma,
852 struct file *file, unsigned long vm_flags)
854 if (vma->vm_flags ^ vm_flags)
856 if (vma->vm_file != file)
858 if (vma->vm_ops && vma->vm_ops->close)
863 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
864 struct anon_vma *anon_vma2,
865 struct vm_area_struct *vma)
868 * The list_is_singular() test is to avoid merging VMA cloned from
869 * parents. This can improve scalability caused by anon_vma lock.
871 if ((!anon_vma1 || !anon_vma2) && (!vma ||
872 list_is_singular(&vma->anon_vma_chain)))
874 return anon_vma1 == anon_vma2;
878 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
879 * in front of (at a lower virtual address and file offset than) the vma.
881 * We cannot merge two vmas if they have differently assigned (non-NULL)
882 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
884 * We don't check here for the merged mmap wrapping around the end of pagecache
885 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
886 * wrap, nor mmaps which cover the final page at index -1UL.
889 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
890 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
892 if (is_mergeable_vma(vma, file, vm_flags) &&
893 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
894 if (vma->vm_pgoff == vm_pgoff)
901 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
902 * beyond (at a higher virtual address and file offset than) the vma.
904 * We cannot merge two vmas if they have differently assigned (non-NULL)
905 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
908 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
909 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
911 if (is_mergeable_vma(vma, file, vm_flags) &&
912 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
914 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
915 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
922 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
923 * whether that can be merged with its predecessor or its successor.
924 * Or both (it neatly fills a hole).
926 * In most cases - when called for mmap, brk or mremap - [addr,end) is
927 * certain not to be mapped by the time vma_merge is called; but when
928 * called for mprotect, it is certain to be already mapped (either at
929 * an offset within prev, or at the start of next), and the flags of
930 * this area are about to be changed to vm_flags - and the no-change
931 * case has already been eliminated.
933 * The following mprotect cases have to be considered, where AAAA is
934 * the area passed down from mprotect_fixup, never extending beyond one
935 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
937 * AAAA AAAA AAAA AAAA
938 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
939 * cannot merge might become might become might become
940 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
941 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
942 * mremap move: PPPPNNNNNNNN 8
944 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
945 * might become case 1 below case 2 below case 3 below
947 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
948 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
950 struct vm_area_struct *vma_merge(struct mm_struct *mm,
951 struct vm_area_struct *prev, unsigned long addr,
952 unsigned long end, unsigned long vm_flags,
953 struct anon_vma *anon_vma, struct file *file,
954 pgoff_t pgoff, struct mempolicy *policy)
956 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
957 struct vm_area_struct *area, *next;
961 * We later require that vma->vm_flags == vm_flags,
962 * so this tests vma->vm_flags & VM_SPECIAL, too.
964 if (vm_flags & VM_SPECIAL)
968 next = prev->vm_next;
972 if (next && next->vm_end == end) /* cases 6, 7, 8 */
973 next = next->vm_next;
976 * Can it merge with the predecessor?
978 if (prev && prev->vm_end == addr &&
979 mpol_equal(vma_policy(prev), policy) &&
980 can_vma_merge_after(prev, vm_flags,
981 anon_vma, file, pgoff)) {
983 * OK, it can. Can we now merge in the successor as well?
985 if (next && end == next->vm_start &&
986 mpol_equal(policy, vma_policy(next)) &&
987 can_vma_merge_before(next, vm_flags,
988 anon_vma, file, pgoff+pglen) &&
989 is_mergeable_anon_vma(prev->anon_vma,
990 next->anon_vma, NULL)) {
992 err = vma_adjust(prev, prev->vm_start,
993 next->vm_end, prev->vm_pgoff, NULL);
994 } else /* cases 2, 5, 7 */
995 err = vma_adjust(prev, prev->vm_start,
996 end, prev->vm_pgoff, NULL);
999 khugepaged_enter_vma_merge(prev);
1004 * Can this new request be merged in front of next?
1006 if (next && end == next->vm_start &&
1007 mpol_equal(policy, vma_policy(next)) &&
1008 can_vma_merge_before(next, vm_flags,
1009 anon_vma, file, pgoff+pglen)) {
1010 if (prev && addr < prev->vm_end) /* case 4 */
1011 err = vma_adjust(prev, prev->vm_start,
1012 addr, prev->vm_pgoff, NULL);
1013 else /* cases 3, 8 */
1014 err = vma_adjust(area, addr, next->vm_end,
1015 next->vm_pgoff - pglen, NULL);
1018 khugepaged_enter_vma_merge(area);
1026 * Rough compatbility check to quickly see if it's even worth looking
1027 * at sharing an anon_vma.
1029 * They need to have the same vm_file, and the flags can only differ
1030 * in things that mprotect may change.
1032 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1033 * we can merge the two vma's. For example, we refuse to merge a vma if
1034 * there is a vm_ops->close() function, because that indicates that the
1035 * driver is doing some kind of reference counting. But that doesn't
1036 * really matter for the anon_vma sharing case.
1038 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1040 return a->vm_end == b->vm_start &&
1041 mpol_equal(vma_policy(a), vma_policy(b)) &&
1042 a->vm_file == b->vm_file &&
1043 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
1044 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1048 * Do some basic sanity checking to see if we can re-use the anon_vma
1049 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1050 * the same as 'old', the other will be the new one that is trying
1051 * to share the anon_vma.
1053 * NOTE! This runs with mm_sem held for reading, so it is possible that
1054 * the anon_vma of 'old' is concurrently in the process of being set up
1055 * by another page fault trying to merge _that_. But that's ok: if it
1056 * is being set up, that automatically means that it will be a singleton
1057 * acceptable for merging, so we can do all of this optimistically. But
1058 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1060 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1061 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1062 * is to return an anon_vma that is "complex" due to having gone through
1065 * We also make sure that the two vma's are compatible (adjacent,
1066 * and with the same memory policies). That's all stable, even with just
1067 * a read lock on the mm_sem.
1069 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1071 if (anon_vma_compatible(a, b)) {
1072 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
1074 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1081 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1082 * neighbouring vmas for a suitable anon_vma, before it goes off
1083 * to allocate a new anon_vma. It checks because a repetitive
1084 * sequence of mprotects and faults may otherwise lead to distinct
1085 * anon_vmas being allocated, preventing vma merge in subsequent
1088 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1090 struct anon_vma *anon_vma;
1091 struct vm_area_struct *near;
1093 near = vma->vm_next;
1097 anon_vma = reusable_anon_vma(near, vma, near);
1101 near = vma->vm_prev;
1105 anon_vma = reusable_anon_vma(near, near, vma);
1110 * There's no absolute need to look only at touching neighbours:
1111 * we could search further afield for "compatible" anon_vmas.
1112 * But it would probably just be a waste of time searching,
1113 * or lead to too many vmas hanging off the same anon_vma.
1114 * We're trying to allow mprotect remerging later on,
1115 * not trying to minimize memory used for anon_vmas.
1120 #ifdef CONFIG_PROC_FS
1121 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1122 struct file *file, long pages)
1124 const unsigned long stack_flags
1125 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1127 mm->total_vm += pages;
1130 mm->shared_vm += pages;
1131 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1132 mm->exec_vm += pages;
1133 } else if (flags & stack_flags)
1134 mm->stack_vm += pages;
1136 #endif /* CONFIG_PROC_FS */
1139 * If a hint addr is less than mmap_min_addr change hint to be as
1140 * low as possible but still greater than mmap_min_addr
1142 static inline unsigned long round_hint_to_min(unsigned long hint)
1145 if (((void *)hint != NULL) &&
1146 (hint < mmap_min_addr))
1147 return PAGE_ALIGN(mmap_min_addr);
1152 * The caller must hold down_write(¤t->mm->mmap_sem).
1155 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
1156 unsigned long len, unsigned long prot,
1157 unsigned long flags, unsigned long pgoff,
1160 struct mm_struct * mm = current->mm;
1161 struct inode *inode;
1162 vm_flags_t vm_flags;
1167 * Does the application expect PROT_READ to imply PROT_EXEC?
1169 * (the exception is when the underlying filesystem is noexec
1170 * mounted, in which case we dont add PROT_EXEC.)
1172 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1173 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1179 if (!(flags & MAP_FIXED))
1180 addr = round_hint_to_min(addr);
1182 /* Careful about overflows.. */
1183 len = PAGE_ALIGN(len);
1187 /* offset overflow? */
1188 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1191 /* Too many mappings? */
1192 if (mm->map_count > sysctl_max_map_count)
1195 /* Obtain the address to map to. we verify (or select) it and ensure
1196 * that it represents a valid section of the address space.
1198 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1199 if (addr & ~PAGE_MASK)
1202 /* Do simple checking here so the lower-level routines won't have
1203 * to. we assume access permissions have been handled by the open
1204 * of the memory object, so we don't do any here.
1206 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1207 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1209 if (flags & MAP_LOCKED)
1210 if (!can_do_mlock())
1213 /* mlock MCL_FUTURE? */
1214 if (vm_flags & VM_LOCKED) {
1215 unsigned long locked, lock_limit;
1216 locked = len >> PAGE_SHIFT;
1217 locked += mm->locked_vm;
1218 lock_limit = rlimit(RLIMIT_MEMLOCK);
1219 lock_limit >>= PAGE_SHIFT;
1220 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1224 inode = file ? file->f_path.dentry->d_inode : NULL;
1227 switch (flags & MAP_TYPE) {
1229 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1233 * Make sure we don't allow writing to an append-only
1236 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1240 * Make sure there are no mandatory locks on the file.
1242 if (locks_verify_locked(inode))
1245 vm_flags |= VM_SHARED | VM_MAYSHARE;
1246 if (!(file->f_mode & FMODE_WRITE))
1247 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1251 if (!(file->f_mode & FMODE_READ))
1253 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1254 if (vm_flags & VM_EXEC)
1256 vm_flags &= ~VM_MAYEXEC;
1259 if (!file->f_op || !file->f_op->mmap)
1267 switch (flags & MAP_TYPE) {
1273 vm_flags |= VM_SHARED | VM_MAYSHARE;
1277 * Set pgoff according to addr for anon_vma.
1279 pgoff = addr >> PAGE_SHIFT;
1286 addr = mmap_region(file, addr, len, flags, vm_flags, pgoff);
1287 if (!IS_ERR_VALUE(addr) &&
1288 ((vm_flags & VM_LOCKED) ||
1289 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1294 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1295 unsigned long, prot, unsigned long, flags,
1296 unsigned long, fd, unsigned long, pgoff)
1298 struct file *file = NULL;
1299 unsigned long retval = -EBADF;
1301 if (!(flags & MAP_ANONYMOUS)) {
1302 audit_mmap_fd(fd, flags);
1303 if (unlikely(flags & MAP_HUGETLB))
1308 } else if (flags & MAP_HUGETLB) {
1309 struct user_struct *user = NULL;
1311 * VM_NORESERVE is used because the reservations will be
1312 * taken when vm_ops->mmap() is called
1313 * A dummy user value is used because we are not locking
1314 * memory so no accounting is necessary
1316 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1318 &user, HUGETLB_ANONHUGE_INODE,
1319 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1321 return PTR_ERR(file);
1324 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1326 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1333 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1334 struct mmap_arg_struct {
1338 unsigned long flags;
1340 unsigned long offset;
1343 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1345 struct mmap_arg_struct a;
1347 if (copy_from_user(&a, arg, sizeof(a)))
1349 if (a.offset & ~PAGE_MASK)
1352 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1353 a.offset >> PAGE_SHIFT);
1355 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1358 * Some shared mappigns will want the pages marked read-only
1359 * to track write events. If so, we'll downgrade vm_page_prot
1360 * to the private version (using protection_map[] without the
1363 int vma_wants_writenotify(struct vm_area_struct *vma)
1365 vm_flags_t vm_flags = vma->vm_flags;
1367 /* If it was private or non-writable, the write bit is already clear */
1368 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1371 /* The backer wishes to know when pages are first written to? */
1372 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1375 /* The open routine did something to the protections already? */
1376 if (pgprot_val(vma->vm_page_prot) !=
1377 pgprot_val(vm_get_page_prot(vm_flags)))
1380 /* Specialty mapping? */
1381 if (vm_flags & VM_PFNMAP)
1384 /* Can the mapping track the dirty pages? */
1385 return vma->vm_file && vma->vm_file->f_mapping &&
1386 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1390 * We account for memory if it's a private writeable mapping,
1391 * not hugepages and VM_NORESERVE wasn't set.
1393 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1396 * hugetlb has its own accounting separate from the core VM
1397 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1399 if (file && is_file_hugepages(file))
1402 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1405 unsigned long mmap_region(struct file *file, unsigned long addr,
1406 unsigned long len, unsigned long flags,
1407 vm_flags_t vm_flags, unsigned long pgoff)
1409 struct mm_struct *mm = current->mm;
1410 struct vm_area_struct *vma, *prev;
1411 int correct_wcount = 0;
1413 struct rb_node **rb_link, *rb_parent;
1414 unsigned long charged = 0;
1415 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1417 /* Clear old maps */
1420 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
1421 if (do_munmap(mm, addr, len))
1426 /* Check against address space limit. */
1427 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1431 * Set 'VM_NORESERVE' if we should not account for the
1432 * memory use of this mapping.
1434 if ((flags & MAP_NORESERVE)) {
1435 /* We honor MAP_NORESERVE if allowed to overcommit */
1436 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1437 vm_flags |= VM_NORESERVE;
1439 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1440 if (file && is_file_hugepages(file))
1441 vm_flags |= VM_NORESERVE;
1445 * Private writable mapping: check memory availability
1447 if (accountable_mapping(file, vm_flags)) {
1448 charged = len >> PAGE_SHIFT;
1449 if (security_vm_enough_memory_mm(mm, charged))
1451 vm_flags |= VM_ACCOUNT;
1455 * Can we just expand an old mapping?
1457 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1462 * Determine the object being mapped and call the appropriate
1463 * specific mapper. the address has already been validated, but
1464 * not unmapped, but the maps are removed from the list.
1466 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1473 vma->vm_start = addr;
1474 vma->vm_end = addr + len;
1475 vma->vm_flags = vm_flags;
1476 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1477 vma->vm_pgoff = pgoff;
1478 INIT_LIST_HEAD(&vma->anon_vma_chain);
1480 error = -EINVAL; /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1483 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1485 if (vm_flags & VM_DENYWRITE) {
1486 error = deny_write_access(file);
1491 vma->vm_file = get_file(file);
1492 error = file->f_op->mmap(file, vma);
1494 goto unmap_and_free_vma;
1496 /* Can addr have changed??
1498 * Answer: Yes, several device drivers can do it in their
1499 * f_op->mmap method. -DaveM
1500 * Bug: If addr is changed, prev, rb_link, rb_parent should
1501 * be updated for vma_link()
1503 WARN_ON_ONCE(addr != vma->vm_start);
1505 addr = vma->vm_start;
1506 pgoff = vma->vm_pgoff;
1507 vm_flags = vma->vm_flags;
1508 } else if (vm_flags & VM_SHARED) {
1509 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1511 error = shmem_zero_setup(vma);
1516 if (vma_wants_writenotify(vma)) {
1517 pgprot_t pprot = vma->vm_page_prot;
1519 /* Can vma->vm_page_prot have changed??
1521 * Answer: Yes, drivers may have changed it in their
1522 * f_op->mmap method.
1524 * Ensures that vmas marked as uncached stay that way.
1526 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1527 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1528 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1531 vma_link(mm, vma, prev, rb_link, rb_parent);
1532 file = vma->vm_file;
1534 /* Once vma denies write, undo our temporary denial count */
1536 atomic_inc(&inode->i_writecount);
1538 perf_event_mmap(vma);
1540 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1541 if (vm_flags & VM_LOCKED) {
1542 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1543 vma == get_gate_vma(current->mm)))
1544 mm->locked_vm += (len >> PAGE_SHIFT);
1546 vma->vm_flags &= ~VM_LOCKED;
1556 atomic_inc(&inode->i_writecount);
1557 vma->vm_file = NULL;
1560 /* Undo any partial mapping done by a device driver. */
1561 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1564 kmem_cache_free(vm_area_cachep, vma);
1567 vm_unacct_memory(charged);
1571 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1574 * We implement the search by looking for an rbtree node that
1575 * immediately follows a suitable gap. That is,
1576 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1577 * - gap_end = vma->vm_start >= info->low_limit + length;
1578 * - gap_end - gap_start >= length
1581 struct mm_struct *mm = current->mm;
1582 struct vm_area_struct *vma;
1583 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1585 /* Adjust search length to account for worst case alignment overhead */
1586 length = info->length + info->align_mask;
1587 if (length < info->length)
1590 /* Adjust search limits by the desired length */
1591 if (info->high_limit < length)
1593 high_limit = info->high_limit - length;
1595 if (info->low_limit > high_limit)
1597 low_limit = info->low_limit + length;
1599 /* Check if rbtree root looks promising */
1600 if (RB_EMPTY_ROOT(&mm->mm_rb))
1602 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1603 if (vma->rb_subtree_gap < length)
1607 /* Visit left subtree if it looks promising */
1608 gap_end = vma->vm_start;
1609 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1610 struct vm_area_struct *left =
1611 rb_entry(vma->vm_rb.rb_left,
1612 struct vm_area_struct, vm_rb);
1613 if (left->rb_subtree_gap >= length) {
1619 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1621 /* Check if current node has a suitable gap */
1622 if (gap_start > high_limit)
1624 if (gap_end >= low_limit && gap_end - gap_start >= length)
1627 /* Visit right subtree if it looks promising */
1628 if (vma->vm_rb.rb_right) {
1629 struct vm_area_struct *right =
1630 rb_entry(vma->vm_rb.rb_right,
1631 struct vm_area_struct, vm_rb);
1632 if (right->rb_subtree_gap >= length) {
1638 /* Go back up the rbtree to find next candidate node */
1640 struct rb_node *prev = &vma->vm_rb;
1641 if (!rb_parent(prev))
1643 vma = rb_entry(rb_parent(prev),
1644 struct vm_area_struct, vm_rb);
1645 if (prev == vma->vm_rb.rb_left) {
1646 gap_start = vma->vm_prev->vm_end;
1647 gap_end = vma->vm_start;
1654 /* Check highest gap, which does not precede any rbtree node */
1655 gap_start = mm->highest_vm_end;
1656 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1657 if (gap_start > high_limit)
1661 /* We found a suitable gap. Clip it with the original low_limit. */
1662 if (gap_start < info->low_limit)
1663 gap_start = info->low_limit;
1665 /* Adjust gap address to the desired alignment */
1666 gap_start += (info->align_offset - gap_start) & info->align_mask;
1668 VM_BUG_ON(gap_start + info->length > info->high_limit);
1669 VM_BUG_ON(gap_start + info->length > gap_end);
1673 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1675 struct mm_struct *mm = current->mm;
1676 struct vm_area_struct *vma;
1677 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1679 /* Adjust search length to account for worst case alignment overhead */
1680 length = info->length + info->align_mask;
1681 if (length < info->length)
1685 * Adjust search limits by the desired length.
1686 * See implementation comment at top of unmapped_area().
1688 gap_end = info->high_limit;
1689 if (gap_end < length)
1691 high_limit = gap_end - length;
1693 if (info->low_limit > high_limit)
1695 low_limit = info->low_limit + length;
1697 /* Check highest gap, which does not precede any rbtree node */
1698 gap_start = mm->highest_vm_end;
1699 if (gap_start <= high_limit)
1702 /* Check if rbtree root looks promising */
1703 if (RB_EMPTY_ROOT(&mm->mm_rb))
1705 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1706 if (vma->rb_subtree_gap < length)
1710 /* Visit right subtree if it looks promising */
1711 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1712 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1713 struct vm_area_struct *right =
1714 rb_entry(vma->vm_rb.rb_right,
1715 struct vm_area_struct, vm_rb);
1716 if (right->rb_subtree_gap >= length) {
1723 /* Check if current node has a suitable gap */
1724 gap_end = vma->vm_start;
1725 if (gap_end < low_limit)
1727 if (gap_start <= high_limit && gap_end - gap_start >= length)
1730 /* Visit left subtree if it looks promising */
1731 if (vma->vm_rb.rb_left) {
1732 struct vm_area_struct *left =
1733 rb_entry(vma->vm_rb.rb_left,
1734 struct vm_area_struct, vm_rb);
1735 if (left->rb_subtree_gap >= length) {
1741 /* Go back up the rbtree to find next candidate node */
1743 struct rb_node *prev = &vma->vm_rb;
1744 if (!rb_parent(prev))
1746 vma = rb_entry(rb_parent(prev),
1747 struct vm_area_struct, vm_rb);
1748 if (prev == vma->vm_rb.rb_right) {
1749 gap_start = vma->vm_prev ?
1750 vma->vm_prev->vm_end : 0;
1757 /* We found a suitable gap. Clip it with the original high_limit. */
1758 if (gap_end > info->high_limit)
1759 gap_end = info->high_limit;
1762 /* Compute highest gap address at the desired alignment */
1763 gap_end -= info->length;
1764 gap_end -= (gap_end - info->align_offset) & info->align_mask;
1766 VM_BUG_ON(gap_end < info->low_limit);
1767 VM_BUG_ON(gap_end < gap_start);
1771 /* Get an address range which is currently unmapped.
1772 * For shmat() with addr=0.
1774 * Ugly calling convention alert:
1775 * Return value with the low bits set means error value,
1777 * if (ret & ~PAGE_MASK)
1780 * This function "knows" that -ENOMEM has the bits set.
1782 #ifndef HAVE_ARCH_UNMAPPED_AREA
1784 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1785 unsigned long len, unsigned long pgoff, unsigned long flags)
1787 struct mm_struct *mm = current->mm;
1788 struct vm_area_struct *vma;
1789 struct vm_unmapped_area_info info;
1791 if (len > TASK_SIZE)
1794 if (flags & MAP_FIXED)
1798 addr = PAGE_ALIGN(addr);
1799 vma = find_vma(mm, addr);
1800 if (TASK_SIZE - len >= addr &&
1801 (!vma || addr + len <= vma->vm_start))
1807 info.low_limit = TASK_UNMAPPED_BASE;
1808 info.high_limit = TASK_SIZE;
1809 info.align_mask = 0;
1810 return vm_unmapped_area(&info);
1814 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1817 * Is this a new hole at the lowest possible address?
1819 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1820 mm->free_area_cache = addr;
1824 * This mmap-allocator allocates new areas top-down from below the
1825 * stack's low limit (the base):
1827 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1829 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1830 const unsigned long len, const unsigned long pgoff,
1831 const unsigned long flags)
1833 struct vm_area_struct *vma;
1834 struct mm_struct *mm = current->mm;
1835 unsigned long addr = addr0;
1836 struct vm_unmapped_area_info info;
1838 /* requested length too big for entire address space */
1839 if (len > TASK_SIZE)
1842 if (flags & MAP_FIXED)
1845 /* requesting a specific address */
1847 addr = PAGE_ALIGN(addr);
1848 vma = find_vma(mm, addr);
1849 if (TASK_SIZE - len >= addr &&
1850 (!vma || addr + len <= vma->vm_start))
1854 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1856 info.low_limit = PAGE_SIZE;
1857 info.high_limit = mm->mmap_base;
1858 info.align_mask = 0;
1859 addr = vm_unmapped_area(&info);
1862 * A failed mmap() very likely causes application failure,
1863 * so fall back to the bottom-up function here. This scenario
1864 * can happen with large stack limits and large mmap()
1867 if (addr & ~PAGE_MASK) {
1868 VM_BUG_ON(addr != -ENOMEM);
1870 info.low_limit = TASK_UNMAPPED_BASE;
1871 info.high_limit = TASK_SIZE;
1872 addr = vm_unmapped_area(&info);
1879 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1882 * Is this a new hole at the highest possible address?
1884 if (addr > mm->free_area_cache)
1885 mm->free_area_cache = addr;
1887 /* dont allow allocations above current base */
1888 if (mm->free_area_cache > mm->mmap_base)
1889 mm->free_area_cache = mm->mmap_base;
1893 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1894 unsigned long pgoff, unsigned long flags)
1896 unsigned long (*get_area)(struct file *, unsigned long,
1897 unsigned long, unsigned long, unsigned long);
1899 unsigned long error = arch_mmap_check(addr, len, flags);
1903 /* Careful about overflows.. */
1904 if (len > TASK_SIZE)
1907 get_area = current->mm->get_unmapped_area;
1908 if (file && file->f_op && file->f_op->get_unmapped_area)
1909 get_area = file->f_op->get_unmapped_area;
1910 addr = get_area(file, addr, len, pgoff, flags);
1911 if (IS_ERR_VALUE(addr))
1914 if (addr > TASK_SIZE - len)
1916 if (addr & ~PAGE_MASK)
1919 addr = arch_rebalance_pgtables(addr, len);
1920 error = security_mmap_addr(addr);
1921 return error ? error : addr;
1924 EXPORT_SYMBOL(get_unmapped_area);
1926 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1927 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1929 struct vm_area_struct *vma = NULL;
1931 if (WARN_ON_ONCE(!mm)) /* Remove this in linux-3.6 */
1934 /* Check the cache first. */
1935 /* (Cache hit rate is typically around 35%.) */
1936 vma = mm->mmap_cache;
1937 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1938 struct rb_node *rb_node;
1940 rb_node = mm->mm_rb.rb_node;
1944 struct vm_area_struct *vma_tmp;
1946 vma_tmp = rb_entry(rb_node,
1947 struct vm_area_struct, vm_rb);
1949 if (vma_tmp->vm_end > addr) {
1951 if (vma_tmp->vm_start <= addr)
1953 rb_node = rb_node->rb_left;
1955 rb_node = rb_node->rb_right;
1958 mm->mmap_cache = vma;
1963 EXPORT_SYMBOL(find_vma);
1966 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1968 struct vm_area_struct *
1969 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1970 struct vm_area_struct **pprev)
1972 struct vm_area_struct *vma;
1974 vma = find_vma(mm, addr);
1976 *pprev = vma->vm_prev;
1978 struct rb_node *rb_node = mm->mm_rb.rb_node;
1981 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1982 rb_node = rb_node->rb_right;
1989 * Verify that the stack growth is acceptable and
1990 * update accounting. This is shared with both the
1991 * grow-up and grow-down cases.
1993 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1995 struct mm_struct *mm = vma->vm_mm;
1996 struct rlimit *rlim = current->signal->rlim;
1997 unsigned long new_start;
1999 /* address space limit tests */
2000 if (!may_expand_vm(mm, grow))
2003 /* Stack limit test */
2004 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2007 /* mlock limit tests */
2008 if (vma->vm_flags & VM_LOCKED) {
2009 unsigned long locked;
2010 unsigned long limit;
2011 locked = mm->locked_vm + grow;
2012 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2013 limit >>= PAGE_SHIFT;
2014 if (locked > limit && !capable(CAP_IPC_LOCK))
2018 /* Check to ensure the stack will not grow into a hugetlb-only region */
2019 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2021 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2025 * Overcommit.. This must be the final test, as it will
2026 * update security statistics.
2028 if (security_vm_enough_memory_mm(mm, grow))
2031 /* Ok, everything looks good - let it rip */
2032 if (vma->vm_flags & VM_LOCKED)
2033 mm->locked_vm += grow;
2034 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
2038 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2040 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2041 * vma is the last one with address > vma->vm_end. Have to extend vma.
2043 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2047 if (!(vma->vm_flags & VM_GROWSUP))
2051 * We must make sure the anon_vma is allocated
2052 * so that the anon_vma locking is not a noop.
2054 if (unlikely(anon_vma_prepare(vma)))
2056 vma_lock_anon_vma(vma);
2059 * vma->vm_start/vm_end cannot change under us because the caller
2060 * is required to hold the mmap_sem in read mode. We need the
2061 * anon_vma lock to serialize against concurrent expand_stacks.
2062 * Also guard against wrapping around to address 0.
2064 if (address < PAGE_ALIGN(address+4))
2065 address = PAGE_ALIGN(address+4);
2067 vma_unlock_anon_vma(vma);
2072 /* Somebody else might have raced and expanded it already */
2073 if (address > vma->vm_end) {
2074 unsigned long size, grow;
2076 size = address - vma->vm_start;
2077 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2080 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2081 error = acct_stack_growth(vma, size, grow);
2084 * vma_gap_update() doesn't support concurrent
2085 * updates, but we only hold a shared mmap_sem
2086 * lock here, so we need to protect against
2087 * concurrent vma expansions.
2088 * vma_lock_anon_vma() doesn't help here, as
2089 * we don't guarantee that all growable vmas
2090 * in a mm share the same root anon vma.
2091 * So, we reuse mm->page_table_lock to guard
2092 * against concurrent vma expansions.
2094 spin_lock(&vma->vm_mm->page_table_lock);
2095 anon_vma_interval_tree_pre_update_vma(vma);
2096 vma->vm_end = address;
2097 anon_vma_interval_tree_post_update_vma(vma);
2099 vma_gap_update(vma->vm_next);
2101 vma->vm_mm->highest_vm_end = address;
2102 spin_unlock(&vma->vm_mm->page_table_lock);
2104 perf_event_mmap(vma);
2108 vma_unlock_anon_vma(vma);
2109 khugepaged_enter_vma_merge(vma);
2110 validate_mm(vma->vm_mm);
2113 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2116 * vma is the first one with address < vma->vm_start. Have to extend vma.
2118 int expand_downwards(struct vm_area_struct *vma,
2119 unsigned long address)
2124 * We must make sure the anon_vma is allocated
2125 * so that the anon_vma locking is not a noop.
2127 if (unlikely(anon_vma_prepare(vma)))
2130 address &= PAGE_MASK;
2131 error = security_mmap_addr(address);
2135 vma_lock_anon_vma(vma);
2138 * vma->vm_start/vm_end cannot change under us because the caller
2139 * is required to hold the mmap_sem in read mode. We need the
2140 * anon_vma lock to serialize against concurrent expand_stacks.
2143 /* Somebody else might have raced and expanded it already */
2144 if (address < vma->vm_start) {
2145 unsigned long size, grow;
2147 size = vma->vm_end - address;
2148 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2151 if (grow <= vma->vm_pgoff) {
2152 error = acct_stack_growth(vma, size, grow);
2155 * vma_gap_update() doesn't support concurrent
2156 * updates, but we only hold a shared mmap_sem
2157 * lock here, so we need to protect against
2158 * concurrent vma expansions.
2159 * vma_lock_anon_vma() doesn't help here, as
2160 * we don't guarantee that all growable vmas
2161 * in a mm share the same root anon vma.
2162 * So, we reuse mm->page_table_lock to guard
2163 * against concurrent vma expansions.
2165 spin_lock(&vma->vm_mm->page_table_lock);
2166 anon_vma_interval_tree_pre_update_vma(vma);
2167 vma->vm_start = address;
2168 vma->vm_pgoff -= grow;
2169 anon_vma_interval_tree_post_update_vma(vma);
2170 vma_gap_update(vma);
2171 spin_unlock(&vma->vm_mm->page_table_lock);
2173 perf_event_mmap(vma);
2177 vma_unlock_anon_vma(vma);
2178 khugepaged_enter_vma_merge(vma);
2179 validate_mm(vma->vm_mm);
2183 #ifdef CONFIG_STACK_GROWSUP
2184 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2186 return expand_upwards(vma, address);
2189 struct vm_area_struct *
2190 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2192 struct vm_area_struct *vma, *prev;
2195 vma = find_vma_prev(mm, addr, &prev);
2196 if (vma && (vma->vm_start <= addr))
2198 if (!prev || expand_stack(prev, addr))
2200 if (prev->vm_flags & VM_LOCKED) {
2201 mlock_vma_pages_range(prev, addr, prev->vm_end);
2206 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2208 return expand_downwards(vma, address);
2211 struct vm_area_struct *
2212 find_extend_vma(struct mm_struct * mm, unsigned long addr)
2214 struct vm_area_struct * vma;
2215 unsigned long start;
2218 vma = find_vma(mm,addr);
2221 if (vma->vm_start <= addr)
2223 if (!(vma->vm_flags & VM_GROWSDOWN))
2225 start = vma->vm_start;
2226 if (expand_stack(vma, addr))
2228 if (vma->vm_flags & VM_LOCKED) {
2229 mlock_vma_pages_range(vma, addr, start);
2236 * Ok - we have the memory areas we should free on the vma list,
2237 * so release them, and do the vma updates.
2239 * Called with the mm semaphore held.
2241 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2243 unsigned long nr_accounted = 0;
2245 /* Update high watermark before we lower total_vm */
2246 update_hiwater_vm(mm);
2248 long nrpages = vma_pages(vma);
2250 if (vma->vm_flags & VM_ACCOUNT)
2251 nr_accounted += nrpages;
2252 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2253 vma = remove_vma(vma);
2255 vm_unacct_memory(nr_accounted);
2260 * Get rid of page table information in the indicated region.
2262 * Called with the mm semaphore held.
2264 static void unmap_region(struct mm_struct *mm,
2265 struct vm_area_struct *vma, struct vm_area_struct *prev,
2266 unsigned long start, unsigned long end)
2268 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
2269 struct mmu_gather tlb;
2272 tlb_gather_mmu(&tlb, mm, 0);
2273 update_hiwater_rss(mm);
2274 unmap_vmas(&tlb, vma, start, end);
2275 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2276 next ? next->vm_start : 0);
2277 tlb_finish_mmu(&tlb, start, end);
2281 * Create a list of vma's touched by the unmap, removing them from the mm's
2282 * vma list as we go..
2285 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2286 struct vm_area_struct *prev, unsigned long end)
2288 struct vm_area_struct **insertion_point;
2289 struct vm_area_struct *tail_vma = NULL;
2292 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2293 vma->vm_prev = NULL;
2295 vma_rb_erase(vma, &mm->mm_rb);
2299 } while (vma && vma->vm_start < end);
2300 *insertion_point = vma;
2302 vma->vm_prev = prev;
2303 vma_gap_update(vma);
2305 mm->highest_vm_end = prev ? prev->vm_end : 0;
2306 tail_vma->vm_next = NULL;
2307 if (mm->unmap_area == arch_unmap_area)
2308 addr = prev ? prev->vm_end : mm->mmap_base;
2310 addr = vma ? vma->vm_start : mm->mmap_base;
2311 mm->unmap_area(mm, addr);
2312 mm->mmap_cache = NULL; /* Kill the cache. */
2316 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2317 * munmap path where it doesn't make sense to fail.
2319 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
2320 unsigned long addr, int new_below)
2322 struct mempolicy *pol;
2323 struct vm_area_struct *new;
2326 if (is_vm_hugetlb_page(vma) && (addr &
2327 ~(huge_page_mask(hstate_vma(vma)))))
2330 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2334 /* most fields are the same, copy all, and then fixup */
2337 INIT_LIST_HEAD(&new->anon_vma_chain);
2342 new->vm_start = addr;
2343 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2346 pol = mpol_dup(vma_policy(vma));
2351 vma_set_policy(new, pol);
2353 if (anon_vma_clone(new, vma))
2357 get_file(new->vm_file);
2359 if (new->vm_ops && new->vm_ops->open)
2360 new->vm_ops->open(new);
2363 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2364 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2366 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2372 /* Clean everything up if vma_adjust failed. */
2373 if (new->vm_ops && new->vm_ops->close)
2374 new->vm_ops->close(new);
2377 unlink_anon_vmas(new);
2381 kmem_cache_free(vm_area_cachep, new);
2387 * Split a vma into two pieces at address 'addr', a new vma is allocated
2388 * either for the first part or the tail.
2390 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2391 unsigned long addr, int new_below)
2393 if (mm->map_count >= sysctl_max_map_count)
2396 return __split_vma(mm, vma, addr, new_below);
2399 /* Munmap is split into 2 main parts -- this part which finds
2400 * what needs doing, and the areas themselves, which do the
2401 * work. This now handles partial unmappings.
2402 * Jeremy Fitzhardinge <jeremy@goop.org>
2404 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2407 struct vm_area_struct *vma, *prev, *last;
2409 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2412 if ((len = PAGE_ALIGN(len)) == 0)
2415 /* Find the first overlapping VMA */
2416 vma = find_vma(mm, start);
2419 prev = vma->vm_prev;
2420 /* we have start < vma->vm_end */
2422 /* if it doesn't overlap, we have nothing.. */
2424 if (vma->vm_start >= end)
2428 * If we need to split any vma, do it now to save pain later.
2430 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2431 * unmapped vm_area_struct will remain in use: so lower split_vma
2432 * places tmp vma above, and higher split_vma places tmp vma below.
2434 if (start > vma->vm_start) {
2438 * Make sure that map_count on return from munmap() will
2439 * not exceed its limit; but let map_count go just above
2440 * its limit temporarily, to help free resources as expected.
2442 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2445 error = __split_vma(mm, vma, start, 0);
2451 /* Does it split the last one? */
2452 last = find_vma(mm, end);
2453 if (last && end > last->vm_start) {
2454 int error = __split_vma(mm, last, end, 1);
2458 vma = prev? prev->vm_next: mm->mmap;
2461 * unlock any mlock()ed ranges before detaching vmas
2463 if (mm->locked_vm) {
2464 struct vm_area_struct *tmp = vma;
2465 while (tmp && tmp->vm_start < end) {
2466 if (tmp->vm_flags & VM_LOCKED) {
2467 mm->locked_vm -= vma_pages(tmp);
2468 munlock_vma_pages_all(tmp);
2475 * Remove the vma's, and unmap the actual pages
2477 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2478 unmap_region(mm, vma, prev, start, end);
2480 /* Fix up all other VM information */
2481 remove_vma_list(mm, vma);
2486 int vm_munmap(unsigned long start, size_t len)
2489 struct mm_struct *mm = current->mm;
2491 down_write(&mm->mmap_sem);
2492 ret = do_munmap(mm, start, len);
2493 up_write(&mm->mmap_sem);
2496 EXPORT_SYMBOL(vm_munmap);
2498 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2500 profile_munmap(addr);
2501 return vm_munmap(addr, len);
2504 static inline void verify_mm_writelocked(struct mm_struct *mm)
2506 #ifdef CONFIG_DEBUG_VM
2507 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2509 up_read(&mm->mmap_sem);
2515 * this is really a simplified "do_mmap". it only handles
2516 * anonymous maps. eventually we may be able to do some
2517 * brk-specific accounting here.
2519 static unsigned long do_brk(unsigned long addr, unsigned long len)
2521 struct mm_struct * mm = current->mm;
2522 struct vm_area_struct * vma, * prev;
2523 unsigned long flags;
2524 struct rb_node ** rb_link, * rb_parent;
2525 pgoff_t pgoff = addr >> PAGE_SHIFT;
2528 len = PAGE_ALIGN(len);
2532 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2534 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2535 if (error & ~PAGE_MASK)
2541 if (mm->def_flags & VM_LOCKED) {
2542 unsigned long locked, lock_limit;
2543 locked = len >> PAGE_SHIFT;
2544 locked += mm->locked_vm;
2545 lock_limit = rlimit(RLIMIT_MEMLOCK);
2546 lock_limit >>= PAGE_SHIFT;
2547 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2552 * mm->mmap_sem is required to protect against another thread
2553 * changing the mappings in case we sleep.
2555 verify_mm_writelocked(mm);
2558 * Clear old maps. this also does some error checking for us
2561 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
2562 if (do_munmap(mm, addr, len))
2567 /* Check against address space limits *after* clearing old maps... */
2568 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2571 if (mm->map_count > sysctl_max_map_count)
2574 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2577 /* Can we just expand an old private anonymous mapping? */
2578 vma = vma_merge(mm, prev, addr, addr + len, flags,
2579 NULL, NULL, pgoff, NULL);
2584 * create a vma struct for an anonymous mapping
2586 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2588 vm_unacct_memory(len >> PAGE_SHIFT);
2592 INIT_LIST_HEAD(&vma->anon_vma_chain);
2594 vma->vm_start = addr;
2595 vma->vm_end = addr + len;
2596 vma->vm_pgoff = pgoff;
2597 vma->vm_flags = flags;
2598 vma->vm_page_prot = vm_get_page_prot(flags);
2599 vma_link(mm, vma, prev, rb_link, rb_parent);
2601 perf_event_mmap(vma);
2602 mm->total_vm += len >> PAGE_SHIFT;
2603 if (flags & VM_LOCKED) {
2604 if (!mlock_vma_pages_range(vma, addr, addr + len))
2605 mm->locked_vm += (len >> PAGE_SHIFT);
2610 unsigned long vm_brk(unsigned long addr, unsigned long len)
2612 struct mm_struct *mm = current->mm;
2615 down_write(&mm->mmap_sem);
2616 ret = do_brk(addr, len);
2617 up_write(&mm->mmap_sem);
2620 EXPORT_SYMBOL(vm_brk);
2622 /* Release all mmaps. */
2623 void exit_mmap(struct mm_struct *mm)
2625 struct mmu_gather tlb;
2626 struct vm_area_struct *vma;
2627 unsigned long nr_accounted = 0;
2629 /* mm's last user has gone, and its about to be pulled down */
2630 mmu_notifier_release(mm);
2632 if (mm->locked_vm) {
2635 if (vma->vm_flags & VM_LOCKED)
2636 munlock_vma_pages_all(vma);
2644 if (!vma) /* Can happen if dup_mmap() received an OOM */
2649 tlb_gather_mmu(&tlb, mm, 1);
2650 /* update_hiwater_rss(mm) here? but nobody should be looking */
2651 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2652 unmap_vmas(&tlb, vma, 0, -1);
2654 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2655 tlb_finish_mmu(&tlb, 0, -1);
2658 * Walk the list again, actually closing and freeing it,
2659 * with preemption enabled, without holding any MM locks.
2662 if (vma->vm_flags & VM_ACCOUNT)
2663 nr_accounted += vma_pages(vma);
2664 vma = remove_vma(vma);
2666 vm_unacct_memory(nr_accounted);
2668 WARN_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2671 /* Insert vm structure into process list sorted by address
2672 * and into the inode's i_mmap tree. If vm_file is non-NULL
2673 * then i_mmap_mutex is taken here.
2675 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2677 struct vm_area_struct *prev;
2678 struct rb_node **rb_link, *rb_parent;
2681 * The vm_pgoff of a purely anonymous vma should be irrelevant
2682 * until its first write fault, when page's anon_vma and index
2683 * are set. But now set the vm_pgoff it will almost certainly
2684 * end up with (unless mremap moves it elsewhere before that
2685 * first wfault), so /proc/pid/maps tells a consistent story.
2687 * By setting it to reflect the virtual start address of the
2688 * vma, merges and splits can happen in a seamless way, just
2689 * using the existing file pgoff checks and manipulations.
2690 * Similarly in do_mmap_pgoff and in do_brk.
2692 if (!vma->vm_file) {
2693 BUG_ON(vma->anon_vma);
2694 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2696 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2697 &prev, &rb_link, &rb_parent))
2699 if ((vma->vm_flags & VM_ACCOUNT) &&
2700 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2703 vma_link(mm, vma, prev, rb_link, rb_parent);
2708 * Copy the vma structure to a new location in the same mm,
2709 * prior to moving page table entries, to effect an mremap move.
2711 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2712 unsigned long addr, unsigned long len, pgoff_t pgoff,
2713 bool *need_rmap_locks)
2715 struct vm_area_struct *vma = *vmap;
2716 unsigned long vma_start = vma->vm_start;
2717 struct mm_struct *mm = vma->vm_mm;
2718 struct vm_area_struct *new_vma, *prev;
2719 struct rb_node **rb_link, *rb_parent;
2720 struct mempolicy *pol;
2721 bool faulted_in_anon_vma = true;
2724 * If anonymous vma has not yet been faulted, update new pgoff
2725 * to match new location, to increase its chance of merging.
2727 if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2728 pgoff = addr >> PAGE_SHIFT;
2729 faulted_in_anon_vma = false;
2732 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2733 return NULL; /* should never get here */
2734 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2735 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2738 * Source vma may have been merged into new_vma
2740 if (unlikely(vma_start >= new_vma->vm_start &&
2741 vma_start < new_vma->vm_end)) {
2743 * The only way we can get a vma_merge with
2744 * self during an mremap is if the vma hasn't
2745 * been faulted in yet and we were allowed to
2746 * reset the dst vma->vm_pgoff to the
2747 * destination address of the mremap to allow
2748 * the merge to happen. mremap must change the
2749 * vm_pgoff linearity between src and dst vmas
2750 * (in turn preventing a vma_merge) to be
2751 * safe. It is only safe to keep the vm_pgoff
2752 * linear if there are no pages mapped yet.
2754 VM_BUG_ON(faulted_in_anon_vma);
2755 *vmap = vma = new_vma;
2757 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2759 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2762 new_vma->vm_start = addr;
2763 new_vma->vm_end = addr + len;
2764 new_vma->vm_pgoff = pgoff;
2765 pol = mpol_dup(vma_policy(vma));
2768 vma_set_policy(new_vma, pol);
2769 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2770 if (anon_vma_clone(new_vma, vma))
2771 goto out_free_mempol;
2772 if (new_vma->vm_file)
2773 get_file(new_vma->vm_file);
2774 if (new_vma->vm_ops && new_vma->vm_ops->open)
2775 new_vma->vm_ops->open(new_vma);
2776 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2777 *need_rmap_locks = false;
2785 kmem_cache_free(vm_area_cachep, new_vma);
2790 * Return true if the calling process may expand its vm space by the passed
2793 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2795 unsigned long cur = mm->total_vm; /* pages */
2798 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2800 if (cur + npages > lim)
2806 static int special_mapping_fault(struct vm_area_struct *vma,
2807 struct vm_fault *vmf)
2810 struct page **pages;
2813 * special mappings have no vm_file, and in that case, the mm
2814 * uses vm_pgoff internally. So we have to subtract it from here.
2815 * We are allowed to do this because we are the mm; do not copy
2816 * this code into drivers!
2818 pgoff = vmf->pgoff - vma->vm_pgoff;
2820 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2824 struct page *page = *pages;
2830 return VM_FAULT_SIGBUS;
2834 * Having a close hook prevents vma merging regardless of flags.
2836 static void special_mapping_close(struct vm_area_struct *vma)
2840 static const struct vm_operations_struct special_mapping_vmops = {
2841 .close = special_mapping_close,
2842 .fault = special_mapping_fault,
2846 * Called with mm->mmap_sem held for writing.
2847 * Insert a new vma covering the given region, with the given flags.
2848 * Its pages are supplied by the given array of struct page *.
2849 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2850 * The region past the last page supplied will always produce SIGBUS.
2851 * The array pointer and the pages it points to are assumed to stay alive
2852 * for as long as this mapping might exist.
2854 int install_special_mapping(struct mm_struct *mm,
2855 unsigned long addr, unsigned long len,
2856 unsigned long vm_flags, struct page **pages)
2859 struct vm_area_struct *vma;
2861 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2862 if (unlikely(vma == NULL))
2865 INIT_LIST_HEAD(&vma->anon_vma_chain);
2867 vma->vm_start = addr;
2868 vma->vm_end = addr + len;
2870 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2871 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2873 vma->vm_ops = &special_mapping_vmops;
2874 vma->vm_private_data = pages;
2876 ret = insert_vm_struct(mm, vma);
2880 mm->total_vm += len >> PAGE_SHIFT;
2882 perf_event_mmap(vma);
2887 kmem_cache_free(vm_area_cachep, vma);
2891 static DEFINE_MUTEX(mm_all_locks_mutex);
2893 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2895 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2897 * The LSB of head.next can't change from under us
2898 * because we hold the mm_all_locks_mutex.
2900 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
2902 * We can safely modify head.next after taking the
2903 * anon_vma->root->rwsem. If some other vma in this mm shares
2904 * the same anon_vma we won't take it again.
2906 * No need of atomic instructions here, head.next
2907 * can't change from under us thanks to the
2908 * anon_vma->root->rwsem.
2910 if (__test_and_set_bit(0, (unsigned long *)
2911 &anon_vma->root->rb_root.rb_node))
2916 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2918 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2920 * AS_MM_ALL_LOCKS can't change from under us because
2921 * we hold the mm_all_locks_mutex.
2923 * Operations on ->flags have to be atomic because
2924 * even if AS_MM_ALL_LOCKS is stable thanks to the
2925 * mm_all_locks_mutex, there may be other cpus
2926 * changing other bitflags in parallel to us.
2928 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2930 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2935 * This operation locks against the VM for all pte/vma/mm related
2936 * operations that could ever happen on a certain mm. This includes
2937 * vmtruncate, try_to_unmap, and all page faults.
2939 * The caller must take the mmap_sem in write mode before calling
2940 * mm_take_all_locks(). The caller isn't allowed to release the
2941 * mmap_sem until mm_drop_all_locks() returns.
2943 * mmap_sem in write mode is required in order to block all operations
2944 * that could modify pagetables and free pages without need of
2945 * altering the vma layout (for example populate_range() with
2946 * nonlinear vmas). It's also needed in write mode to avoid new
2947 * anon_vmas to be associated with existing vmas.
2949 * A single task can't take more than one mm_take_all_locks() in a row
2950 * or it would deadlock.
2952 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
2953 * mapping->flags avoid to take the same lock twice, if more than one
2954 * vma in this mm is backed by the same anon_vma or address_space.
2956 * We can take all the locks in random order because the VM code
2957 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
2958 * takes more than one of them in a row. Secondly we're protected
2959 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2961 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2962 * that may have to take thousand of locks.
2964 * mm_take_all_locks() can fail if it's interrupted by signals.
2966 int mm_take_all_locks(struct mm_struct *mm)
2968 struct vm_area_struct *vma;
2969 struct anon_vma_chain *avc;
2971 BUG_ON(down_read_trylock(&mm->mmap_sem));
2973 mutex_lock(&mm_all_locks_mutex);
2975 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2976 if (signal_pending(current))
2978 if (vma->vm_file && vma->vm_file->f_mapping)
2979 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2982 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2983 if (signal_pending(current))
2986 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2987 vm_lock_anon_vma(mm, avc->anon_vma);
2993 mm_drop_all_locks(mm);
2997 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2999 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3001 * The LSB of head.next can't change to 0 from under
3002 * us because we hold the mm_all_locks_mutex.
3004 * We must however clear the bitflag before unlocking
3005 * the vma so the users using the anon_vma->rb_root will
3006 * never see our bitflag.
3008 * No need of atomic instructions here, head.next
3009 * can't change from under us until we release the
3010 * anon_vma->root->rwsem.
3012 if (!__test_and_clear_bit(0, (unsigned long *)
3013 &anon_vma->root->rb_root.rb_node))
3015 anon_vma_unlock(anon_vma);
3019 static void vm_unlock_mapping(struct address_space *mapping)
3021 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3023 * AS_MM_ALL_LOCKS can't change to 0 from under us
3024 * because we hold the mm_all_locks_mutex.
3026 mutex_unlock(&mapping->i_mmap_mutex);
3027 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3034 * The mmap_sem cannot be released by the caller until
3035 * mm_drop_all_locks() returns.
3037 void mm_drop_all_locks(struct mm_struct *mm)
3039 struct vm_area_struct *vma;
3040 struct anon_vma_chain *avc;
3042 BUG_ON(down_read_trylock(&mm->mmap_sem));
3043 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3045 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3047 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3048 vm_unlock_anon_vma(avc->anon_vma);
3049 if (vma->vm_file && vma->vm_file->f_mapping)
3050 vm_unlock_mapping(vma->vm_file->f_mapping);
3053 mutex_unlock(&mm_all_locks_mutex);
3057 * initialise the VMA slab
3059 void __init mmap_init(void)
3063 ret = percpu_counter_init(&vm_committed_as, 0);