6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/shmem_fs.h>
29 #include <linux/profile.h>
30 #include <linux/export.h>
31 #include <linux/mount.h>
32 #include <linux/mempolicy.h>
33 #include <linux/rmap.h>
34 #include <linux/mmu_notifier.h>
35 #include <linux/mmdebug.h>
36 #include <linux/perf_event.h>
37 #include <linux/audit.h>
38 #include <linux/khugepaged.h>
39 #include <linux/uprobes.h>
40 #include <linux/rbtree_augmented.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/moduleparam.h>
46 #include <linux/pkeys.h>
48 #include <asm/uaccess.h>
49 #include <asm/cacheflush.h>
51 #include <asm/mmu_context.h>
55 #ifndef arch_mmap_check
56 #define arch_mmap_check(addr, len, flags) (0)
59 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
60 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
61 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
62 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
64 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
65 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
66 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
67 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
70 static bool ignore_rlimit_data;
71 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
73 static void unmap_region(struct mm_struct *mm,
74 struct vm_area_struct *vma, struct vm_area_struct *prev,
75 unsigned long start, unsigned long end);
77 /* description of effects of mapping type and prot in current implementation.
78 * this is due to the limited x86 page protection hardware. The expected
79 * behavior is in parens:
82 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
83 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
84 * w: (no) no w: (no) no w: (yes) yes w: (no) no
85 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
87 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
88 * w: (no) no w: (no) no w: (copy) copy w: (no) no
89 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
91 pgprot_t protection_map[16] = {
92 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
93 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
96 pgprot_t vm_get_page_prot(unsigned long vm_flags)
98 return __pgprot(pgprot_val(protection_map[vm_flags &
99 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
100 pgprot_val(arch_vm_get_page_prot(vm_flags)));
102 EXPORT_SYMBOL(vm_get_page_prot);
104 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
106 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
109 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
110 void vma_set_page_prot(struct vm_area_struct *vma)
112 unsigned long vm_flags = vma->vm_flags;
113 pgprot_t vm_page_prot;
115 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
116 if (vma_wants_writenotify(vma, vm_page_prot)) {
117 vm_flags &= ~VM_SHARED;
118 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
120 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
121 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
125 * Requires inode->i_mapping->i_mmap_rwsem
127 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
128 struct file *file, struct address_space *mapping)
130 if (vma->vm_flags & VM_DENYWRITE)
131 atomic_inc(&file_inode(file)->i_writecount);
132 if (vma->vm_flags & VM_SHARED)
133 mapping_unmap_writable(mapping);
135 flush_dcache_mmap_lock(mapping);
136 vma_interval_tree_remove(vma, &mapping->i_mmap);
137 flush_dcache_mmap_unlock(mapping);
141 * Unlink a file-based vm structure from its interval tree, to hide
142 * vma from rmap and vmtruncate before freeing its page tables.
144 void unlink_file_vma(struct vm_area_struct *vma)
146 struct file *file = vma->vm_file;
149 struct address_space *mapping = file->f_mapping;
150 i_mmap_lock_write(mapping);
151 __remove_shared_vm_struct(vma, file, mapping);
152 i_mmap_unlock_write(mapping);
157 * Close a vm structure and free it, returning the next.
159 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
161 struct vm_area_struct *next = vma->vm_next;
164 if (vma->vm_ops && vma->vm_ops->close)
165 vma->vm_ops->close(vma);
168 mpol_put(vma_policy(vma));
169 kmem_cache_free(vm_area_cachep, vma);
173 static int do_brk(unsigned long addr, unsigned long len);
175 SYSCALL_DEFINE1(brk, unsigned long, brk)
177 unsigned long retval;
178 unsigned long newbrk, oldbrk;
179 struct mm_struct *mm = current->mm;
180 struct vm_area_struct *next;
181 unsigned long min_brk;
184 if (down_write_killable(&mm->mmap_sem))
187 #ifdef CONFIG_COMPAT_BRK
189 * CONFIG_COMPAT_BRK can still be overridden by setting
190 * randomize_va_space to 2, which will still cause mm->start_brk
191 * to be arbitrarily shifted
193 if (current->brk_randomized)
194 min_brk = mm->start_brk;
196 min_brk = mm->end_data;
198 min_brk = mm->start_brk;
204 * Check against rlimit here. If this check is done later after the test
205 * of oldbrk with newbrk then it can escape the test and let the data
206 * segment grow beyond its set limit the in case where the limit is
207 * not page aligned -Ram Gupta
209 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
210 mm->end_data, mm->start_data))
213 newbrk = PAGE_ALIGN(brk);
214 oldbrk = PAGE_ALIGN(mm->brk);
215 if (oldbrk == newbrk)
218 /* Always allow shrinking brk. */
219 if (brk <= mm->brk) {
220 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
225 /* Check against existing mmap mappings. */
226 next = find_vma(mm, oldbrk);
227 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
230 /* Ok, looks good - let it rip. */
231 if (do_brk(oldbrk, newbrk-oldbrk) < 0)
236 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
237 up_write(&mm->mmap_sem);
239 mm_populate(oldbrk, newbrk - oldbrk);
244 up_write(&mm->mmap_sem);
248 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
250 unsigned long max, prev_end, subtree_gap;
253 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
254 * allow two stack_guard_gaps between them here, and when choosing
255 * an unmapped area; whereas when expanding we only require one.
256 * That's a little inconsistent, but keeps the code here simpler.
258 max = vm_start_gap(vma);
260 prev_end = vm_end_gap(vma->vm_prev);
266 if (vma->vm_rb.rb_left) {
267 subtree_gap = rb_entry(vma->vm_rb.rb_left,
268 struct vm_area_struct, vm_rb)->rb_subtree_gap;
269 if (subtree_gap > max)
272 if (vma->vm_rb.rb_right) {
273 subtree_gap = rb_entry(vma->vm_rb.rb_right,
274 struct vm_area_struct, vm_rb)->rb_subtree_gap;
275 if (subtree_gap > max)
281 #ifdef CONFIG_DEBUG_VM_RB
282 static int browse_rb(struct mm_struct *mm)
284 struct rb_root *root = &mm->mm_rb;
285 int i = 0, j, bug = 0;
286 struct rb_node *nd, *pn = NULL;
287 unsigned long prev = 0, pend = 0;
289 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
290 struct vm_area_struct *vma;
291 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
292 if (vma->vm_start < prev) {
293 pr_emerg("vm_start %lx < prev %lx\n",
294 vma->vm_start, prev);
297 if (vma->vm_start < pend) {
298 pr_emerg("vm_start %lx < pend %lx\n",
299 vma->vm_start, pend);
302 if (vma->vm_start > vma->vm_end) {
303 pr_emerg("vm_start %lx > vm_end %lx\n",
304 vma->vm_start, vma->vm_end);
307 spin_lock(&mm->page_table_lock);
308 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
309 pr_emerg("free gap %lx, correct %lx\n",
311 vma_compute_subtree_gap(vma));
314 spin_unlock(&mm->page_table_lock);
317 prev = vma->vm_start;
321 for (nd = pn; nd; nd = rb_prev(nd))
324 pr_emerg("backwards %d, forwards %d\n", j, i);
330 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
334 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
335 struct vm_area_struct *vma;
336 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
337 VM_BUG_ON_VMA(vma != ignore &&
338 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
343 static void validate_mm(struct mm_struct *mm)
347 unsigned long highest_address = 0;
348 struct vm_area_struct *vma = mm->mmap;
351 struct anon_vma *anon_vma = vma->anon_vma;
352 struct anon_vma_chain *avc;
355 anon_vma_lock_read(anon_vma);
356 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
357 anon_vma_interval_tree_verify(avc);
358 anon_vma_unlock_read(anon_vma);
361 highest_address = vm_end_gap(vma);
365 if (i != mm->map_count) {
366 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
369 if (highest_address != mm->highest_vm_end) {
370 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
371 mm->highest_vm_end, highest_address);
375 if (i != mm->map_count) {
377 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
380 VM_BUG_ON_MM(bug, mm);
383 #define validate_mm_rb(root, ignore) do { } while (0)
384 #define validate_mm(mm) do { } while (0)
387 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
388 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
391 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
392 * vma->vm_prev->vm_end values changed, without modifying the vma's position
395 static void vma_gap_update(struct vm_area_struct *vma)
398 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
399 * function that does exacltly what we want.
401 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
404 static inline void vma_rb_insert(struct vm_area_struct *vma,
405 struct rb_root *root)
407 /* All rb_subtree_gap values must be consistent prior to insertion */
408 validate_mm_rb(root, NULL);
410 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
413 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
416 * Note rb_erase_augmented is a fairly large inline function,
417 * so make sure we instantiate it only once with our desired
418 * augmented rbtree callbacks.
420 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
423 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
424 struct rb_root *root,
425 struct vm_area_struct *ignore)
428 * All rb_subtree_gap values must be consistent prior to erase,
429 * with the possible exception of the "next" vma being erased if
430 * next->vm_start was reduced.
432 validate_mm_rb(root, ignore);
434 __vma_rb_erase(vma, root);
437 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
438 struct rb_root *root)
441 * All rb_subtree_gap values must be consistent prior to erase,
442 * with the possible exception of the vma being erased.
444 validate_mm_rb(root, vma);
446 __vma_rb_erase(vma, root);
450 * vma has some anon_vma assigned, and is already inserted on that
451 * anon_vma's interval trees.
453 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
454 * vma must be removed from the anon_vma's interval trees using
455 * anon_vma_interval_tree_pre_update_vma().
457 * After the update, the vma will be reinserted using
458 * anon_vma_interval_tree_post_update_vma().
460 * The entire update must be protected by exclusive mmap_sem and by
461 * the root anon_vma's mutex.
464 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
466 struct anon_vma_chain *avc;
468 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
469 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
473 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
475 struct anon_vma_chain *avc;
477 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
478 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
481 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
482 unsigned long end, struct vm_area_struct **pprev,
483 struct rb_node ***rb_link, struct rb_node **rb_parent)
485 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
487 __rb_link = &mm->mm_rb.rb_node;
488 rb_prev = __rb_parent = NULL;
491 struct vm_area_struct *vma_tmp;
493 __rb_parent = *__rb_link;
494 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
496 if (vma_tmp->vm_end > addr) {
497 /* Fail if an existing vma overlaps the area */
498 if (vma_tmp->vm_start < end)
500 __rb_link = &__rb_parent->rb_left;
502 rb_prev = __rb_parent;
503 __rb_link = &__rb_parent->rb_right;
509 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
510 *rb_link = __rb_link;
511 *rb_parent = __rb_parent;
515 static unsigned long count_vma_pages_range(struct mm_struct *mm,
516 unsigned long addr, unsigned long end)
518 unsigned long nr_pages = 0;
519 struct vm_area_struct *vma;
521 /* Find first overlaping mapping */
522 vma = find_vma_intersection(mm, addr, end);
526 nr_pages = (min(end, vma->vm_end) -
527 max(addr, vma->vm_start)) >> PAGE_SHIFT;
529 /* Iterate over the rest of the overlaps */
530 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
531 unsigned long overlap_len;
533 if (vma->vm_start > end)
536 overlap_len = min(end, vma->vm_end) - vma->vm_start;
537 nr_pages += overlap_len >> PAGE_SHIFT;
543 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
544 struct rb_node **rb_link, struct rb_node *rb_parent)
546 /* Update tracking information for the gap following the new vma. */
548 vma_gap_update(vma->vm_next);
550 mm->highest_vm_end = vm_end_gap(vma);
553 * vma->vm_prev wasn't known when we followed the rbtree to find the
554 * correct insertion point for that vma. As a result, we could not
555 * update the vma vm_rb parents rb_subtree_gap values on the way down.
556 * So, we first insert the vma with a zero rb_subtree_gap value
557 * (to be consistent with what we did on the way down), and then
558 * immediately update the gap to the correct value. Finally we
559 * rebalance the rbtree after all augmented values have been set.
561 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
562 vma->rb_subtree_gap = 0;
564 vma_rb_insert(vma, &mm->mm_rb);
567 static void __vma_link_file(struct vm_area_struct *vma)
573 struct address_space *mapping = file->f_mapping;
575 if (vma->vm_flags & VM_DENYWRITE)
576 atomic_dec(&file_inode(file)->i_writecount);
577 if (vma->vm_flags & VM_SHARED)
578 atomic_inc(&mapping->i_mmap_writable);
580 flush_dcache_mmap_lock(mapping);
581 vma_interval_tree_insert(vma, &mapping->i_mmap);
582 flush_dcache_mmap_unlock(mapping);
587 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
588 struct vm_area_struct *prev, struct rb_node **rb_link,
589 struct rb_node *rb_parent)
591 __vma_link_list(mm, vma, prev, rb_parent);
592 __vma_link_rb(mm, vma, rb_link, rb_parent);
595 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
596 struct vm_area_struct *prev, struct rb_node **rb_link,
597 struct rb_node *rb_parent)
599 struct address_space *mapping = NULL;
602 mapping = vma->vm_file->f_mapping;
603 i_mmap_lock_write(mapping);
606 __vma_link(mm, vma, prev, rb_link, rb_parent);
607 __vma_link_file(vma);
610 i_mmap_unlock_write(mapping);
617 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
618 * mm's list and rbtree. It has already been inserted into the interval tree.
620 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
622 struct vm_area_struct *prev;
623 struct rb_node **rb_link, *rb_parent;
625 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
626 &prev, &rb_link, &rb_parent))
628 __vma_link(mm, vma, prev, rb_link, rb_parent);
632 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
633 struct vm_area_struct *vma,
634 struct vm_area_struct *prev,
636 struct vm_area_struct *ignore)
638 struct vm_area_struct *next;
640 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
643 prev->vm_next = next;
647 prev->vm_next = next;
652 next->vm_prev = prev;
655 vmacache_invalidate(mm);
658 static inline void __vma_unlink_prev(struct mm_struct *mm,
659 struct vm_area_struct *vma,
660 struct vm_area_struct *prev)
662 __vma_unlink_common(mm, vma, prev, true, vma);
666 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
667 * is already present in an i_mmap tree without adjusting the tree.
668 * The following helper function should be used when such adjustments
669 * are necessary. The "insert" vma (if any) is to be inserted
670 * before we drop the necessary locks.
672 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
673 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
674 struct vm_area_struct *expand)
676 struct mm_struct *mm = vma->vm_mm;
677 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
678 struct address_space *mapping = NULL;
679 struct rb_root *root = NULL;
680 struct anon_vma *anon_vma = NULL;
681 struct file *file = vma->vm_file;
682 bool start_changed = false, end_changed = false;
683 long adjust_next = 0;
686 if (next && !insert) {
687 struct vm_area_struct *exporter = NULL, *importer = NULL;
689 if (end >= next->vm_end) {
691 * vma expands, overlapping all the next, and
692 * perhaps the one after too (mprotect case 6).
693 * The only other cases that gets here are
694 * case 1, case 7 and case 8.
696 if (next == expand) {
698 * The only case where we don't expand "vma"
699 * and we expand "next" instead is case 8.
701 VM_WARN_ON(end != next->vm_end);
703 * remove_next == 3 means we're
704 * removing "vma" and that to do so we
705 * swapped "vma" and "next".
708 VM_WARN_ON(file != next->vm_file);
711 VM_WARN_ON(expand != vma);
713 * case 1, 6, 7, remove_next == 2 is case 6,
714 * remove_next == 1 is case 1 or 7.
716 remove_next = 1 + (end > next->vm_end);
717 VM_WARN_ON(remove_next == 2 &&
718 end != next->vm_next->vm_end);
719 VM_WARN_ON(remove_next == 1 &&
720 end != next->vm_end);
721 /* trim end to next, for case 6 first pass */
729 * If next doesn't have anon_vma, import from vma after
730 * next, if the vma overlaps with it.
732 if (remove_next == 2 && !next->anon_vma)
733 exporter = next->vm_next;
735 } else if (end > next->vm_start) {
737 * vma expands, overlapping part of the next:
738 * mprotect case 5 shifting the boundary up.
740 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
743 VM_WARN_ON(expand != importer);
744 } else if (end < vma->vm_end) {
746 * vma shrinks, and !insert tells it's not
747 * split_vma inserting another: so it must be
748 * mprotect case 4 shifting the boundary down.
750 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
753 VM_WARN_ON(expand != importer);
757 * Easily overlooked: when mprotect shifts the boundary,
758 * make sure the expanding vma has anon_vma set if the
759 * shrinking vma had, to cover any anon pages imported.
761 if (exporter && exporter->anon_vma && !importer->anon_vma) {
764 importer->anon_vma = exporter->anon_vma;
765 error = anon_vma_clone(importer, exporter);
771 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
774 mapping = file->f_mapping;
775 root = &mapping->i_mmap;
776 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
779 uprobe_munmap(next, next->vm_start, next->vm_end);
781 i_mmap_lock_write(mapping);
784 * Put into interval tree now, so instantiated pages
785 * are visible to arm/parisc __flush_dcache_page
786 * throughout; but we cannot insert into address
787 * space until vma start or end is updated.
789 __vma_link_file(insert);
793 anon_vma = vma->anon_vma;
794 if (!anon_vma && adjust_next)
795 anon_vma = next->anon_vma;
797 VM_WARN_ON(adjust_next && next->anon_vma &&
798 anon_vma != next->anon_vma);
799 anon_vma_lock_write(anon_vma);
800 anon_vma_interval_tree_pre_update_vma(vma);
802 anon_vma_interval_tree_pre_update_vma(next);
806 flush_dcache_mmap_lock(mapping);
807 vma_interval_tree_remove(vma, root);
809 vma_interval_tree_remove(next, root);
812 if (start != vma->vm_start) {
813 vma->vm_start = start;
814 start_changed = true;
816 if (end != vma->vm_end) {
820 vma->vm_pgoff = pgoff;
822 next->vm_start += adjust_next << PAGE_SHIFT;
823 next->vm_pgoff += adjust_next;
828 vma_interval_tree_insert(next, root);
829 vma_interval_tree_insert(vma, root);
830 flush_dcache_mmap_unlock(mapping);
835 * vma_merge has merged next into vma, and needs
836 * us to remove next before dropping the locks.
838 if (remove_next != 3)
839 __vma_unlink_prev(mm, next, vma);
842 * vma is not before next if they've been
845 * pre-swap() next->vm_start was reduced so
846 * tell validate_mm_rb to ignore pre-swap()
847 * "next" (which is stored in post-swap()
850 __vma_unlink_common(mm, next, NULL, false, vma);
852 __remove_shared_vm_struct(next, file, mapping);
855 * split_vma has split insert from vma, and needs
856 * us to insert it before dropping the locks
857 * (it may either follow vma or precede it).
859 __insert_vm_struct(mm, insert);
865 mm->highest_vm_end = vm_end_gap(vma);
866 else if (!adjust_next)
867 vma_gap_update(next);
872 anon_vma_interval_tree_post_update_vma(vma);
874 anon_vma_interval_tree_post_update_vma(next);
875 anon_vma_unlock_write(anon_vma);
878 i_mmap_unlock_write(mapping);
889 uprobe_munmap(next, next->vm_start, next->vm_end);
893 anon_vma_merge(vma, next);
895 mpol_put(vma_policy(next));
896 kmem_cache_free(vm_area_cachep, next);
898 * In mprotect's case 6 (see comments on vma_merge),
899 * we must remove another next too. It would clutter
900 * up the code too much to do both in one go.
902 if (remove_next != 3) {
904 * If "next" was removed and vma->vm_end was
905 * expanded (up) over it, in turn
906 * "next->vm_prev->vm_end" changed and the
907 * "vma->vm_next" gap must be updated.
912 * For the scope of the comment "next" and
913 * "vma" considered pre-swap(): if "vma" was
914 * removed, next->vm_start was expanded (down)
915 * over it and the "next" gap must be updated.
916 * Because of the swap() the post-swap() "vma"
917 * actually points to pre-swap() "next"
918 * (post-swap() "next" as opposed is now a
923 if (remove_next == 2) {
929 vma_gap_update(next);
932 * If remove_next == 2 we obviously can't
935 * If remove_next == 3 we can't reach this
936 * path because pre-swap() next is always not
937 * NULL. pre-swap() "next" is not being
938 * removed and its next->vm_end is not altered
939 * (and furthermore "end" already matches
940 * next->vm_end in remove_next == 3).
942 * We reach this only in the remove_next == 1
943 * case if the "next" vma that was removed was
944 * the highest vma of the mm. However in such
945 * case next->vm_end == "end" and the extended
946 * "vma" has vma->vm_end == next->vm_end so
947 * mm->highest_vm_end doesn't need any update
948 * in remove_next == 1 case.
950 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
962 * If the vma has a ->close operation then the driver probably needs to release
963 * per-vma resources, so we don't attempt to merge those.
965 static inline int is_mergeable_vma(struct vm_area_struct *vma,
966 struct file *file, unsigned long vm_flags,
967 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
968 const char __user *anon_name)
971 * VM_SOFTDIRTY should not prevent from VMA merging, if we
972 * match the flags but dirty bit -- the caller should mark
973 * merged VMA as dirty. If dirty bit won't be excluded from
974 * comparison, we increase pressue on the memory system forcing
975 * the kernel to generate new VMAs when old one could be
978 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
980 if (vma->vm_file != file)
982 if (vma->vm_ops && vma->vm_ops->close)
984 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
986 if (vma_get_anon_name(vma) != anon_name)
991 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
992 struct anon_vma *anon_vma2,
993 struct vm_area_struct *vma)
996 * The list_is_singular() test is to avoid merging VMA cloned from
997 * parents. This can improve scalability caused by anon_vma lock.
999 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1000 list_is_singular(&vma->anon_vma_chain)))
1002 return anon_vma1 == anon_vma2;
1006 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1007 * in front of (at a lower virtual address and file offset than) the vma.
1009 * We cannot merge two vmas if they have differently assigned (non-NULL)
1010 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1012 * We don't check here for the merged mmap wrapping around the end of pagecache
1013 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1014 * wrap, nor mmaps which cover the final page at index -1UL.
1017 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1018 struct anon_vma *anon_vma, struct file *file,
1020 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1021 const char __user *anon_name)
1023 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
1024 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1025 if (vma->vm_pgoff == vm_pgoff)
1032 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1033 * beyond (at a higher virtual address and file offset than) the vma.
1035 * We cannot merge two vmas if they have differently assigned (non-NULL)
1036 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1039 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1040 struct anon_vma *anon_vma, struct file *file,
1042 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1043 const char __user *anon_name)
1045 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
1046 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1048 vm_pglen = vma_pages(vma);
1049 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1056 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
1057 * figure out whether that can be merged with its predecessor or its
1058 * successor. Or both (it neatly fills a hole).
1060 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1061 * certain not to be mapped by the time vma_merge is called; but when
1062 * called for mprotect, it is certain to be already mapped (either at
1063 * an offset within prev, or at the start of next), and the flags of
1064 * this area are about to be changed to vm_flags - and the no-change
1065 * case has already been eliminated.
1067 * The following mprotect cases have to be considered, where AAAA is
1068 * the area passed down from mprotect_fixup, never extending beyond one
1069 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1071 * AAAA AAAA AAAA AAAA
1072 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1073 * cannot merge might become might become might become
1074 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1075 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1076 * mremap move: PPPPXXXXXXXX 8
1078 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1079 * might become case 1 below case 2 below case 3 below
1081 * It is important for case 8 that the the vma NNNN overlapping the
1082 * region AAAA is never going to extended over XXXX. Instead XXXX must
1083 * be extended in region AAAA and NNNN must be removed. This way in
1084 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1085 * rmap_locks, the properties of the merged vma will be already
1086 * correct for the whole merged range. Some of those properties like
1087 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1088 * be correct for the whole merged range immediately after the
1089 * rmap_locks are released. Otherwise if XXXX would be removed and
1090 * NNNN would be extended over the XXXX range, remove_migration_ptes
1091 * or other rmap walkers (if working on addresses beyond the "end"
1092 * parameter) may establish ptes with the wrong permissions of NNNN
1093 * instead of the right permissions of XXXX.
1095 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1096 struct vm_area_struct *prev, unsigned long addr,
1097 unsigned long end, unsigned long vm_flags,
1098 struct anon_vma *anon_vma, struct file *file,
1099 pgoff_t pgoff, struct mempolicy *policy,
1100 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1101 const char __user *anon_name)
1103 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1104 struct vm_area_struct *area, *next;
1108 * We later require that vma->vm_flags == vm_flags,
1109 * so this tests vma->vm_flags & VM_SPECIAL, too.
1111 if (vm_flags & VM_SPECIAL)
1115 next = prev->vm_next;
1119 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1120 next = next->vm_next;
1122 /* verify some invariant that must be enforced by the caller */
1123 VM_WARN_ON(prev && addr <= prev->vm_start);
1124 VM_WARN_ON(area && end > area->vm_end);
1125 VM_WARN_ON(addr >= end);
1128 * Can it merge with the predecessor?
1130 if (prev && prev->vm_end == addr &&
1131 mpol_equal(vma_policy(prev), policy) &&
1132 can_vma_merge_after(prev, vm_flags,
1133 anon_vma, file, pgoff,
1137 * OK, it can. Can we now merge in the successor as well?
1139 if (next && end == next->vm_start &&
1140 mpol_equal(policy, vma_policy(next)) &&
1141 can_vma_merge_before(next, vm_flags,
1146 is_mergeable_anon_vma(prev->anon_vma,
1147 next->anon_vma, NULL)) {
1149 err = __vma_adjust(prev, prev->vm_start,
1150 next->vm_end, prev->vm_pgoff, NULL,
1152 } else /* cases 2, 5, 7 */
1153 err = __vma_adjust(prev, prev->vm_start,
1154 end, prev->vm_pgoff, NULL, prev);
1157 khugepaged_enter_vma_merge(prev, vm_flags);
1162 * Can this new request be merged in front of next?
1164 if (next && end == next->vm_start &&
1165 mpol_equal(policy, vma_policy(next)) &&
1166 can_vma_merge_before(next, vm_flags,
1167 anon_vma, file, pgoff+pglen,
1170 if (prev && addr < prev->vm_end) /* case 4 */
1171 err = __vma_adjust(prev, prev->vm_start,
1172 addr, prev->vm_pgoff, NULL, next);
1173 else { /* cases 3, 8 */
1174 err = __vma_adjust(area, addr, next->vm_end,
1175 next->vm_pgoff - pglen, NULL, next);
1177 * In case 3 area is already equal to next and
1178 * this is a noop, but in case 8 "area" has
1179 * been removed and next was expanded over it.
1185 khugepaged_enter_vma_merge(area, vm_flags);
1193 * Rough compatbility check to quickly see if it's even worth looking
1194 * at sharing an anon_vma.
1196 * They need to have the same vm_file, and the flags can only differ
1197 * in things that mprotect may change.
1199 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1200 * we can merge the two vma's. For example, we refuse to merge a vma if
1201 * there is a vm_ops->close() function, because that indicates that the
1202 * driver is doing some kind of reference counting. But that doesn't
1203 * really matter for the anon_vma sharing case.
1205 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1207 return a->vm_end == b->vm_start &&
1208 mpol_equal(vma_policy(a), vma_policy(b)) &&
1209 a->vm_file == b->vm_file &&
1210 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1211 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1215 * Do some basic sanity checking to see if we can re-use the anon_vma
1216 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1217 * the same as 'old', the other will be the new one that is trying
1218 * to share the anon_vma.
1220 * NOTE! This runs with mm_sem held for reading, so it is possible that
1221 * the anon_vma of 'old' is concurrently in the process of being set up
1222 * by another page fault trying to merge _that_. But that's ok: if it
1223 * is being set up, that automatically means that it will be a singleton
1224 * acceptable for merging, so we can do all of this optimistically. But
1225 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1227 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1228 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1229 * is to return an anon_vma that is "complex" due to having gone through
1232 * We also make sure that the two vma's are compatible (adjacent,
1233 * and with the same memory policies). That's all stable, even with just
1234 * a read lock on the mm_sem.
1236 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1238 if (anon_vma_compatible(a, b)) {
1239 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1241 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1248 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1249 * neighbouring vmas for a suitable anon_vma, before it goes off
1250 * to allocate a new anon_vma. It checks because a repetitive
1251 * sequence of mprotects and faults may otherwise lead to distinct
1252 * anon_vmas being allocated, preventing vma merge in subsequent
1255 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1257 struct anon_vma *anon_vma;
1258 struct vm_area_struct *near;
1260 near = vma->vm_next;
1264 anon_vma = reusable_anon_vma(near, vma, near);
1268 near = vma->vm_prev;
1272 anon_vma = reusable_anon_vma(near, near, vma);
1277 * There's no absolute need to look only at touching neighbours:
1278 * we could search further afield for "compatible" anon_vmas.
1279 * But it would probably just be a waste of time searching,
1280 * or lead to too many vmas hanging off the same anon_vma.
1281 * We're trying to allow mprotect remerging later on,
1282 * not trying to minimize memory used for anon_vmas.
1288 * If a hint addr is less than mmap_min_addr change hint to be as
1289 * low as possible but still greater than mmap_min_addr
1291 static inline unsigned long round_hint_to_min(unsigned long hint)
1294 if (((void *)hint != NULL) &&
1295 (hint < mmap_min_addr))
1296 return PAGE_ALIGN(mmap_min_addr);
1300 static inline int mlock_future_check(struct mm_struct *mm,
1301 unsigned long flags,
1304 unsigned long locked, lock_limit;
1306 /* mlock MCL_FUTURE? */
1307 if (flags & VM_LOCKED) {
1308 locked = len >> PAGE_SHIFT;
1309 locked += mm->locked_vm;
1310 lock_limit = rlimit(RLIMIT_MEMLOCK);
1311 lock_limit >>= PAGE_SHIFT;
1312 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1318 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1320 if (S_ISREG(inode->i_mode))
1321 return MAX_LFS_FILESIZE;
1323 if (S_ISBLK(inode->i_mode))
1324 return MAX_LFS_FILESIZE;
1326 /* Special "we do even unsigned file positions" case */
1327 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1330 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1334 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1335 unsigned long pgoff, unsigned long len)
1337 u64 maxsize = file_mmap_size_max(file, inode);
1339 if (maxsize && len > maxsize)
1342 if (pgoff > maxsize >> PAGE_SHIFT)
1348 * The caller must hold down_write(¤t->mm->mmap_sem).
1350 unsigned long do_mmap(struct file *file, unsigned long addr,
1351 unsigned long len, unsigned long prot,
1352 unsigned long flags, vm_flags_t vm_flags,
1353 unsigned long pgoff, unsigned long *populate)
1355 struct mm_struct *mm = current->mm;
1364 * Does the application expect PROT_READ to imply PROT_EXEC?
1366 * (the exception is when the underlying filesystem is noexec
1367 * mounted, in which case we dont add PROT_EXEC.)
1369 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1370 if (!(file && path_noexec(&file->f_path)))
1373 if (!(flags & MAP_FIXED))
1374 addr = round_hint_to_min(addr);
1376 /* Careful about overflows.. */
1377 len = PAGE_ALIGN(len);
1381 /* offset overflow? */
1382 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1385 /* Too many mappings? */
1386 if (mm->map_count > sysctl_max_map_count)
1389 /* Obtain the address to map to. we verify (or select) it and ensure
1390 * that it represents a valid section of the address space.
1392 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1393 if (offset_in_page(addr))
1396 if (prot == PROT_EXEC) {
1397 pkey = execute_only_pkey(mm);
1402 /* Do simple checking here so the lower-level routines won't have
1403 * to. we assume access permissions have been handled by the open
1404 * of the memory object, so we don't do any here.
1406 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1407 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1409 if (flags & MAP_LOCKED)
1410 if (!can_do_mlock())
1413 if (mlock_future_check(mm, vm_flags, len))
1417 struct inode *inode = file_inode(file);
1419 if (!file_mmap_ok(file, inode, pgoff, len))
1422 switch (flags & MAP_TYPE) {
1424 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1428 * Make sure we don't allow writing to an append-only
1431 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1435 * Make sure there are no mandatory locks on the file.
1437 if (locks_verify_locked(file))
1440 vm_flags |= VM_SHARED | VM_MAYSHARE;
1441 if (!(file->f_mode & FMODE_WRITE))
1442 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1446 if (!(file->f_mode & FMODE_READ))
1448 if (path_noexec(&file->f_path)) {
1449 if (vm_flags & VM_EXEC)
1451 vm_flags &= ~VM_MAYEXEC;
1454 if (!file->f_op->mmap)
1456 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1464 switch (flags & MAP_TYPE) {
1466 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1472 vm_flags |= VM_SHARED | VM_MAYSHARE;
1476 * Set pgoff according to addr for anon_vma.
1478 pgoff = addr >> PAGE_SHIFT;
1486 * Set 'VM_NORESERVE' if we should not account for the
1487 * memory use of this mapping.
1489 if (flags & MAP_NORESERVE) {
1490 /* We honor MAP_NORESERVE if allowed to overcommit */
1491 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1492 vm_flags |= VM_NORESERVE;
1494 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1495 if (file && is_file_hugepages(file))
1496 vm_flags |= VM_NORESERVE;
1499 addr = mmap_region(file, addr, len, vm_flags, pgoff);
1500 if (!IS_ERR_VALUE(addr) &&
1501 ((vm_flags & VM_LOCKED) ||
1502 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1507 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1508 unsigned long, prot, unsigned long, flags,
1509 unsigned long, fd, unsigned long, pgoff)
1511 struct file *file = NULL;
1512 unsigned long retval;
1514 if (!(flags & MAP_ANONYMOUS)) {
1515 audit_mmap_fd(fd, flags);
1519 if (is_file_hugepages(file))
1520 len = ALIGN(len, huge_page_size(hstate_file(file)));
1522 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1524 } else if (flags & MAP_HUGETLB) {
1525 struct user_struct *user = NULL;
1528 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1532 len = ALIGN(len, huge_page_size(hs));
1534 * VM_NORESERVE is used because the reservations will be
1535 * taken when vm_ops->mmap() is called
1536 * A dummy user value is used because we are not locking
1537 * memory so no accounting is necessary
1539 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1541 &user, HUGETLB_ANONHUGE_INODE,
1542 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1544 return PTR_ERR(file);
1547 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1549 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1556 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1557 struct mmap_arg_struct {
1561 unsigned long flags;
1563 unsigned long offset;
1566 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1568 struct mmap_arg_struct a;
1570 if (copy_from_user(&a, arg, sizeof(a)))
1572 if (offset_in_page(a.offset))
1575 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1576 a.offset >> PAGE_SHIFT);
1578 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1581 * Some shared mappigns will want the pages marked read-only
1582 * to track write events. If so, we'll downgrade vm_page_prot
1583 * to the private version (using protection_map[] without the
1586 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1588 vm_flags_t vm_flags = vma->vm_flags;
1589 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1591 /* If it was private or non-writable, the write bit is already clear */
1592 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1595 /* The backer wishes to know when pages are first written to? */
1596 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1599 /* The open routine did something to the protections that pgprot_modify
1600 * won't preserve? */
1601 if (pgprot_val(vm_page_prot) !=
1602 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1605 /* Do we need to track softdirty? */
1606 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1609 /* Specialty mapping? */
1610 if (vm_flags & VM_PFNMAP)
1613 /* Can the mapping track the dirty pages? */
1614 return vma->vm_file && vma->vm_file->f_mapping &&
1615 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1619 * We account for memory if it's a private writeable mapping,
1620 * not hugepages and VM_NORESERVE wasn't set.
1622 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1625 * hugetlb has its own accounting separate from the core VM
1626 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1628 if (file && is_file_hugepages(file))
1631 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1634 unsigned long mmap_region(struct file *file, unsigned long addr,
1635 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1637 struct mm_struct *mm = current->mm;
1638 struct vm_area_struct *vma, *prev;
1640 struct rb_node **rb_link, *rb_parent;
1641 unsigned long charged = 0;
1643 /* Check against address space limit. */
1644 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1645 unsigned long nr_pages;
1648 * MAP_FIXED may remove pages of mappings that intersects with
1649 * requested mapping. Account for the pages it would unmap.
1651 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1653 if (!may_expand_vm(mm, vm_flags,
1654 (len >> PAGE_SHIFT) - nr_pages))
1658 /* Clear old maps */
1659 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1661 if (do_munmap(mm, addr, len))
1666 * Private writable mapping: check memory availability
1668 if (accountable_mapping(file, vm_flags)) {
1669 charged = len >> PAGE_SHIFT;
1670 if (security_vm_enough_memory_mm(mm, charged))
1672 vm_flags |= VM_ACCOUNT;
1676 * Can we just expand an old mapping?
1678 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1679 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
1684 * Determine the object being mapped and call the appropriate
1685 * specific mapper. the address has already been validated, but
1686 * not unmapped, but the maps are removed from the list.
1688 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1695 vma->vm_start = addr;
1696 vma->vm_end = addr + len;
1697 vma->vm_flags = vm_flags;
1698 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1699 vma->vm_pgoff = pgoff;
1700 INIT_LIST_HEAD(&vma->anon_vma_chain);
1703 if (vm_flags & VM_DENYWRITE) {
1704 error = deny_write_access(file);
1708 if (vm_flags & VM_SHARED) {
1709 error = mapping_map_writable(file->f_mapping);
1711 goto allow_write_and_free_vma;
1714 /* ->mmap() can change vma->vm_file, but must guarantee that
1715 * vma_link() below can deny write-access if VM_DENYWRITE is set
1716 * and map writably if VM_SHARED is set. This usually means the
1717 * new file must not have been exposed to user-space, yet.
1719 vma->vm_file = get_file(file);
1720 error = file->f_op->mmap(file, vma);
1722 goto unmap_and_free_vma;
1724 /* Can addr have changed??
1726 * Answer: Yes, several device drivers can do it in their
1727 * f_op->mmap method. -DaveM
1728 * Bug: If addr is changed, prev, rb_link, rb_parent should
1729 * be updated for vma_link()
1731 WARN_ON_ONCE(addr != vma->vm_start);
1733 addr = vma->vm_start;
1734 vm_flags = vma->vm_flags;
1735 } else if (vm_flags & VM_SHARED) {
1736 error = shmem_zero_setup(vma);
1741 vma_link(mm, vma, prev, rb_link, rb_parent);
1742 /* Once vma denies write, undo our temporary denial count */
1744 if (vm_flags & VM_SHARED)
1745 mapping_unmap_writable(file->f_mapping);
1746 if (vm_flags & VM_DENYWRITE)
1747 allow_write_access(file);
1749 file = vma->vm_file;
1751 perf_event_mmap(vma);
1753 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1754 if (vm_flags & VM_LOCKED) {
1755 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1756 vma == get_gate_vma(current->mm)))
1757 mm->locked_vm += (len >> PAGE_SHIFT);
1759 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1766 * New (or expanded) vma always get soft dirty status.
1767 * Otherwise user-space soft-dirty page tracker won't
1768 * be able to distinguish situation when vma area unmapped,
1769 * then new mapped in-place (which must be aimed as
1770 * a completely new data area).
1772 vma->vm_flags |= VM_SOFTDIRTY;
1774 vma_set_page_prot(vma);
1779 vma->vm_file = NULL;
1782 /* Undo any partial mapping done by a device driver. */
1783 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1785 if (vm_flags & VM_SHARED)
1786 mapping_unmap_writable(file->f_mapping);
1787 allow_write_and_free_vma:
1788 if (vm_flags & VM_DENYWRITE)
1789 allow_write_access(file);
1791 kmem_cache_free(vm_area_cachep, vma);
1794 vm_unacct_memory(charged);
1798 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1801 * We implement the search by looking for an rbtree node that
1802 * immediately follows a suitable gap. That is,
1803 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1804 * - gap_end = vma->vm_start >= info->low_limit + length;
1805 * - gap_end - gap_start >= length
1808 struct mm_struct *mm = current->mm;
1809 struct vm_area_struct *vma;
1810 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1812 /* Adjust search length to account for worst case alignment overhead */
1813 length = info->length + info->align_mask;
1814 if (length < info->length)
1817 /* Adjust search limits by the desired length */
1818 if (info->high_limit < length)
1820 high_limit = info->high_limit - length;
1822 if (info->low_limit > high_limit)
1824 low_limit = info->low_limit + length;
1826 /* Check if rbtree root looks promising */
1827 if (RB_EMPTY_ROOT(&mm->mm_rb))
1829 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1830 if (vma->rb_subtree_gap < length)
1834 /* Visit left subtree if it looks promising */
1835 gap_end = vm_start_gap(vma);
1836 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1837 struct vm_area_struct *left =
1838 rb_entry(vma->vm_rb.rb_left,
1839 struct vm_area_struct, vm_rb);
1840 if (left->rb_subtree_gap >= length) {
1846 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1848 /* Check if current node has a suitable gap */
1849 if (gap_start > high_limit)
1851 if (gap_end >= low_limit &&
1852 gap_end > gap_start && gap_end - gap_start >= length)
1855 /* Visit right subtree if it looks promising */
1856 if (vma->vm_rb.rb_right) {
1857 struct vm_area_struct *right =
1858 rb_entry(vma->vm_rb.rb_right,
1859 struct vm_area_struct, vm_rb);
1860 if (right->rb_subtree_gap >= length) {
1866 /* Go back up the rbtree to find next candidate node */
1868 struct rb_node *prev = &vma->vm_rb;
1869 if (!rb_parent(prev))
1871 vma = rb_entry(rb_parent(prev),
1872 struct vm_area_struct, vm_rb);
1873 if (prev == vma->vm_rb.rb_left) {
1874 gap_start = vm_end_gap(vma->vm_prev);
1875 gap_end = vm_start_gap(vma);
1882 /* Check highest gap, which does not precede any rbtree node */
1883 gap_start = mm->highest_vm_end;
1884 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1885 if (gap_start > high_limit)
1889 /* We found a suitable gap. Clip it with the original low_limit. */
1890 if (gap_start < info->low_limit)
1891 gap_start = info->low_limit;
1893 /* Adjust gap address to the desired alignment */
1894 gap_start += (info->align_offset - gap_start) & info->align_mask;
1896 VM_BUG_ON(gap_start + info->length > info->high_limit);
1897 VM_BUG_ON(gap_start + info->length > gap_end);
1901 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1903 struct mm_struct *mm = current->mm;
1904 struct vm_area_struct *vma;
1905 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1907 /* Adjust search length to account for worst case alignment overhead */
1908 length = info->length + info->align_mask;
1909 if (length < info->length)
1913 * Adjust search limits by the desired length.
1914 * See implementation comment at top of unmapped_area().
1916 gap_end = info->high_limit;
1917 if (gap_end < length)
1919 high_limit = gap_end - length;
1921 if (info->low_limit > high_limit)
1923 low_limit = info->low_limit + length;
1925 /* Check highest gap, which does not precede any rbtree node */
1926 gap_start = mm->highest_vm_end;
1927 if (gap_start <= high_limit)
1930 /* Check if rbtree root looks promising */
1931 if (RB_EMPTY_ROOT(&mm->mm_rb))
1933 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1934 if (vma->rb_subtree_gap < length)
1938 /* Visit right subtree if it looks promising */
1939 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1940 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1941 struct vm_area_struct *right =
1942 rb_entry(vma->vm_rb.rb_right,
1943 struct vm_area_struct, vm_rb);
1944 if (right->rb_subtree_gap >= length) {
1951 /* Check if current node has a suitable gap */
1952 gap_end = vm_start_gap(vma);
1953 if (gap_end < low_limit)
1955 if (gap_start <= high_limit &&
1956 gap_end > gap_start && gap_end - gap_start >= length)
1959 /* Visit left subtree if it looks promising */
1960 if (vma->vm_rb.rb_left) {
1961 struct vm_area_struct *left =
1962 rb_entry(vma->vm_rb.rb_left,
1963 struct vm_area_struct, vm_rb);
1964 if (left->rb_subtree_gap >= length) {
1970 /* Go back up the rbtree to find next candidate node */
1972 struct rb_node *prev = &vma->vm_rb;
1973 if (!rb_parent(prev))
1975 vma = rb_entry(rb_parent(prev),
1976 struct vm_area_struct, vm_rb);
1977 if (prev == vma->vm_rb.rb_right) {
1978 gap_start = vma->vm_prev ?
1979 vm_end_gap(vma->vm_prev) : 0;
1986 /* We found a suitable gap. Clip it with the original high_limit. */
1987 if (gap_end > info->high_limit)
1988 gap_end = info->high_limit;
1991 /* Compute highest gap address at the desired alignment */
1992 gap_end -= info->length;
1993 gap_end -= (gap_end - info->align_offset) & info->align_mask;
1995 VM_BUG_ON(gap_end < info->low_limit);
1996 VM_BUG_ON(gap_end < gap_start);
2000 /* Get an address range which is currently unmapped.
2001 * For shmat() with addr=0.
2003 * Ugly calling convention alert:
2004 * Return value with the low bits set means error value,
2006 * if (ret & ~PAGE_MASK)
2009 * This function "knows" that -ENOMEM has the bits set.
2011 #ifndef HAVE_ARCH_UNMAPPED_AREA
2013 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2014 unsigned long len, unsigned long pgoff, unsigned long flags)
2016 struct mm_struct *mm = current->mm;
2017 struct vm_area_struct *vma, *prev;
2018 struct vm_unmapped_area_info info;
2020 if (len > TASK_SIZE - mmap_min_addr)
2023 if (flags & MAP_FIXED)
2027 addr = PAGE_ALIGN(addr);
2028 vma = find_vma_prev(mm, addr, &prev);
2029 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2030 (!vma || addr + len <= vm_start_gap(vma)) &&
2031 (!prev || addr >= vm_end_gap(prev)))
2037 info.low_limit = mm->mmap_base;
2038 info.high_limit = TASK_SIZE;
2039 info.align_mask = 0;
2040 return vm_unmapped_area(&info);
2045 * This mmap-allocator allocates new areas top-down from below the
2046 * stack's low limit (the base):
2048 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2050 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2051 const unsigned long len, const unsigned long pgoff,
2052 const unsigned long flags)
2054 struct vm_area_struct *vma, *prev;
2055 struct mm_struct *mm = current->mm;
2056 unsigned long addr = addr0;
2057 struct vm_unmapped_area_info info;
2059 /* requested length too big for entire address space */
2060 if (len > TASK_SIZE - mmap_min_addr)
2063 if (flags & MAP_FIXED)
2066 /* requesting a specific address */
2068 addr = PAGE_ALIGN(addr);
2069 vma = find_vma_prev(mm, addr, &prev);
2070 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2071 (!vma || addr + len <= vm_start_gap(vma)) &&
2072 (!prev || addr >= vm_end_gap(prev)))
2076 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2078 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2079 info.high_limit = mm->mmap_base;
2080 info.align_mask = 0;
2081 addr = vm_unmapped_area(&info);
2084 * A failed mmap() very likely causes application failure,
2085 * so fall back to the bottom-up function here. This scenario
2086 * can happen with large stack limits and large mmap()
2089 if (offset_in_page(addr)) {
2090 VM_BUG_ON(addr != -ENOMEM);
2092 info.low_limit = TASK_UNMAPPED_BASE;
2093 info.high_limit = TASK_SIZE;
2094 addr = vm_unmapped_area(&info);
2102 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2103 unsigned long pgoff, unsigned long flags)
2105 unsigned long (*get_area)(struct file *, unsigned long,
2106 unsigned long, unsigned long, unsigned long);
2108 unsigned long error = arch_mmap_check(addr, len, flags);
2112 /* Careful about overflows.. */
2113 if (len > TASK_SIZE)
2116 get_area = current->mm->get_unmapped_area;
2118 if (file->f_op->get_unmapped_area)
2119 get_area = file->f_op->get_unmapped_area;
2120 } else if (flags & MAP_SHARED) {
2122 * mmap_region() will call shmem_zero_setup() to create a file,
2123 * so use shmem's get_unmapped_area in case it can be huge.
2124 * do_mmap_pgoff() will clear pgoff, so match alignment.
2127 get_area = shmem_get_unmapped_area;
2130 addr = get_area(file, addr, len, pgoff, flags);
2131 if (IS_ERR_VALUE(addr))
2134 if (addr > TASK_SIZE - len)
2136 if (offset_in_page(addr))
2139 error = security_mmap_addr(addr);
2140 return error ? error : addr;
2143 EXPORT_SYMBOL(get_unmapped_area);
2145 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2146 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2148 struct rb_node *rb_node;
2149 struct vm_area_struct *vma;
2151 /* Check the cache first. */
2152 vma = vmacache_find(mm, addr);
2156 rb_node = mm->mm_rb.rb_node;
2159 struct vm_area_struct *tmp;
2161 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2163 if (tmp->vm_end > addr) {
2165 if (tmp->vm_start <= addr)
2167 rb_node = rb_node->rb_left;
2169 rb_node = rb_node->rb_right;
2173 vmacache_update(addr, vma);
2177 EXPORT_SYMBOL(find_vma);
2180 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2182 struct vm_area_struct *
2183 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2184 struct vm_area_struct **pprev)
2186 struct vm_area_struct *vma;
2188 vma = find_vma(mm, addr);
2190 *pprev = vma->vm_prev;
2192 struct rb_node *rb_node = mm->mm_rb.rb_node;
2195 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2196 rb_node = rb_node->rb_right;
2203 * Verify that the stack growth is acceptable and
2204 * update accounting. This is shared with both the
2205 * grow-up and grow-down cases.
2207 static int acct_stack_growth(struct vm_area_struct *vma,
2208 unsigned long size, unsigned long grow)
2210 struct mm_struct *mm = vma->vm_mm;
2211 struct rlimit *rlim = current->signal->rlim;
2212 unsigned long new_start;
2214 /* address space limit tests */
2215 if (!may_expand_vm(mm, vma->vm_flags, grow))
2218 /* Stack limit test */
2219 if (size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2222 /* mlock limit tests */
2223 if (vma->vm_flags & VM_LOCKED) {
2224 unsigned long locked;
2225 unsigned long limit;
2226 locked = mm->locked_vm + grow;
2227 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2228 limit >>= PAGE_SHIFT;
2229 if (locked > limit && !capable(CAP_IPC_LOCK))
2233 /* Check to ensure the stack will not grow into a hugetlb-only region */
2234 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2236 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2240 * Overcommit.. This must be the final test, as it will
2241 * update security statistics.
2243 if (security_vm_enough_memory_mm(mm, grow))
2249 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2251 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2252 * vma is the last one with address > vma->vm_end. Have to extend vma.
2254 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2256 struct mm_struct *mm = vma->vm_mm;
2257 struct vm_area_struct *next;
2258 unsigned long gap_addr;
2261 if (!(vma->vm_flags & VM_GROWSUP))
2264 /* Guard against exceeding limits of the address space. */
2265 address &= PAGE_MASK;
2266 if (address >= (TASK_SIZE & PAGE_MASK))
2268 address += PAGE_SIZE;
2270 /* Enforce stack_guard_gap */
2271 gap_addr = address + stack_guard_gap;
2273 /* Guard against overflow */
2274 if (gap_addr < address || gap_addr > TASK_SIZE)
2275 gap_addr = TASK_SIZE;
2277 next = vma->vm_next;
2278 if (next && next->vm_start < gap_addr &&
2279 (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2280 if (!(next->vm_flags & VM_GROWSUP))
2282 /* Check that both stack segments have the same anon_vma? */
2285 /* We must make sure the anon_vma is allocated. */
2286 if (unlikely(anon_vma_prepare(vma)))
2290 * vma->vm_start/vm_end cannot change under us because the caller
2291 * is required to hold the mmap_sem in read mode. We need the
2292 * anon_vma lock to serialize against concurrent expand_stacks.
2294 anon_vma_lock_write(vma->anon_vma);
2296 /* Somebody else might have raced and expanded it already */
2297 if (address > vma->vm_end) {
2298 unsigned long size, grow;
2300 size = address - vma->vm_start;
2301 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2304 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2305 error = acct_stack_growth(vma, size, grow);
2308 * vma_gap_update() doesn't support concurrent
2309 * updates, but we only hold a shared mmap_sem
2310 * lock here, so we need to protect against
2311 * concurrent vma expansions.
2312 * anon_vma_lock_write() doesn't help here, as
2313 * we don't guarantee that all growable vmas
2314 * in a mm share the same root anon vma.
2315 * So, we reuse mm->page_table_lock to guard
2316 * against concurrent vma expansions.
2318 spin_lock(&mm->page_table_lock);
2319 if (vma->vm_flags & VM_LOCKED)
2320 mm->locked_vm += grow;
2321 vm_stat_account(mm, vma->vm_flags, grow);
2322 anon_vma_interval_tree_pre_update_vma(vma);
2323 vma->vm_end = address;
2324 anon_vma_interval_tree_post_update_vma(vma);
2326 vma_gap_update(vma->vm_next);
2328 mm->highest_vm_end = vm_end_gap(vma);
2329 spin_unlock(&mm->page_table_lock);
2331 perf_event_mmap(vma);
2335 anon_vma_unlock_write(vma->anon_vma);
2336 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2340 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2343 * vma is the first one with address < vma->vm_start. Have to extend vma.
2345 int expand_downwards(struct vm_area_struct *vma,
2346 unsigned long address)
2348 struct mm_struct *mm = vma->vm_mm;
2349 struct vm_area_struct *prev;
2350 unsigned long gap_addr;
2353 address &= PAGE_MASK;
2354 if (address < mmap_min_addr)
2357 /* Enforce stack_guard_gap */
2358 gap_addr = address - stack_guard_gap;
2359 if (gap_addr > address)
2361 prev = vma->vm_prev;
2362 if (prev && prev->vm_end > gap_addr &&
2363 (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2364 if (!(prev->vm_flags & VM_GROWSDOWN))
2366 /* Check that both stack segments have the same anon_vma? */
2369 /* We must make sure the anon_vma is allocated. */
2370 if (unlikely(anon_vma_prepare(vma)))
2374 * vma->vm_start/vm_end cannot change under us because the caller
2375 * is required to hold the mmap_sem in read mode. We need the
2376 * anon_vma lock to serialize against concurrent expand_stacks.
2378 anon_vma_lock_write(vma->anon_vma);
2380 /* Somebody else might have raced and expanded it already */
2381 if (address < vma->vm_start) {
2382 unsigned long size, grow;
2384 size = vma->vm_end - address;
2385 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2388 if (grow <= vma->vm_pgoff) {
2389 error = acct_stack_growth(vma, size, grow);
2392 * vma_gap_update() doesn't support concurrent
2393 * updates, but we only hold a shared mmap_sem
2394 * lock here, so we need to protect against
2395 * concurrent vma expansions.
2396 * anon_vma_lock_write() doesn't help here, as
2397 * we don't guarantee that all growable vmas
2398 * in a mm share the same root anon vma.
2399 * So, we reuse mm->page_table_lock to guard
2400 * against concurrent vma expansions.
2402 spin_lock(&mm->page_table_lock);
2403 if (vma->vm_flags & VM_LOCKED)
2404 mm->locked_vm += grow;
2405 vm_stat_account(mm, vma->vm_flags, grow);
2406 anon_vma_interval_tree_pre_update_vma(vma);
2407 vma->vm_start = address;
2408 vma->vm_pgoff -= grow;
2409 anon_vma_interval_tree_post_update_vma(vma);
2410 vma_gap_update(vma);
2411 spin_unlock(&mm->page_table_lock);
2413 perf_event_mmap(vma);
2417 anon_vma_unlock_write(vma->anon_vma);
2418 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2423 /* enforced gap between the expanding stack and other mappings. */
2424 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2426 static int __init cmdline_parse_stack_guard_gap(char *p)
2431 val = simple_strtoul(p, &endptr, 10);
2433 stack_guard_gap = val << PAGE_SHIFT;
2437 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2439 #ifdef CONFIG_STACK_GROWSUP
2440 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2442 return expand_upwards(vma, address);
2445 struct vm_area_struct *
2446 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2448 struct vm_area_struct *vma, *prev;
2451 vma = find_vma_prev(mm, addr, &prev);
2452 if (vma && (vma->vm_start <= addr))
2454 /* don't alter vm_end if the coredump is running */
2455 if (!prev || !mmget_still_valid(mm) || expand_stack(prev, addr))
2457 if (prev->vm_flags & VM_LOCKED)
2458 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2462 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2464 return expand_downwards(vma, address);
2467 struct vm_area_struct *
2468 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2470 struct vm_area_struct *vma;
2471 unsigned long start;
2474 vma = find_vma(mm, addr);
2477 if (vma->vm_start <= addr)
2479 if (!(vma->vm_flags & VM_GROWSDOWN))
2481 /* don't alter vm_start if the coredump is running */
2482 if (!mmget_still_valid(mm))
2484 start = vma->vm_start;
2485 if (expand_stack(vma, addr))
2487 if (vma->vm_flags & VM_LOCKED)
2488 populate_vma_page_range(vma, addr, start, NULL);
2493 EXPORT_SYMBOL_GPL(find_extend_vma);
2496 * Ok - we have the memory areas we should free on the vma list,
2497 * so release them, and do the vma updates.
2499 * Called with the mm semaphore held.
2501 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2503 unsigned long nr_accounted = 0;
2505 /* Update high watermark before we lower total_vm */
2506 update_hiwater_vm(mm);
2508 long nrpages = vma_pages(vma);
2510 if (vma->vm_flags & VM_ACCOUNT)
2511 nr_accounted += nrpages;
2512 vm_stat_account(mm, vma->vm_flags, -nrpages);
2513 vma = remove_vma(vma);
2515 vm_unacct_memory(nr_accounted);
2520 * Get rid of page table information in the indicated region.
2522 * Called with the mm semaphore held.
2524 static void unmap_region(struct mm_struct *mm,
2525 struct vm_area_struct *vma, struct vm_area_struct *prev,
2526 unsigned long start, unsigned long end)
2528 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2529 struct mmu_gather tlb;
2532 tlb_gather_mmu(&tlb, mm, start, end);
2533 update_hiwater_rss(mm);
2534 unmap_vmas(&tlb, vma, start, end);
2535 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2536 next ? next->vm_start : USER_PGTABLES_CEILING);
2537 tlb_finish_mmu(&tlb, start, end);
2541 * Create a list of vma's touched by the unmap, removing them from the mm's
2542 * vma list as we go..
2545 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2546 struct vm_area_struct *prev, unsigned long end)
2548 struct vm_area_struct **insertion_point;
2549 struct vm_area_struct *tail_vma = NULL;
2551 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2552 vma->vm_prev = NULL;
2554 vma_rb_erase(vma, &mm->mm_rb);
2558 } while (vma && vma->vm_start < end);
2559 *insertion_point = vma;
2561 vma->vm_prev = prev;
2562 vma_gap_update(vma);
2564 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2565 tail_vma->vm_next = NULL;
2567 /* Kill the cache */
2568 vmacache_invalidate(mm);
2572 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2573 * munmap path where it doesn't make sense to fail.
2575 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2576 unsigned long addr, int new_below)
2578 struct vm_area_struct *new;
2581 if (vma->vm_ops && vma->vm_ops->split) {
2582 err = vma->vm_ops->split(vma, addr);
2587 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2591 /* most fields are the same, copy all, and then fixup */
2594 INIT_LIST_HEAD(&new->anon_vma_chain);
2599 new->vm_start = addr;
2600 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2603 err = vma_dup_policy(vma, new);
2607 err = anon_vma_clone(new, vma);
2612 get_file(new->vm_file);
2614 if (new->vm_ops && new->vm_ops->open)
2615 new->vm_ops->open(new);
2618 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2619 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2621 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2627 /* Clean everything up if vma_adjust failed. */
2628 if (new->vm_ops && new->vm_ops->close)
2629 new->vm_ops->close(new);
2632 unlink_anon_vmas(new);
2634 mpol_put(vma_policy(new));
2636 kmem_cache_free(vm_area_cachep, new);
2641 * Split a vma into two pieces at address 'addr', a new vma is allocated
2642 * either for the first part or the tail.
2644 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2645 unsigned long addr, int new_below)
2647 if (mm->map_count >= sysctl_max_map_count)
2650 return __split_vma(mm, vma, addr, new_below);
2653 /* Munmap is split into 2 main parts -- this part which finds
2654 * what needs doing, and the areas themselves, which do the
2655 * work. This now handles partial unmappings.
2656 * Jeremy Fitzhardinge <jeremy@goop.org>
2658 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2661 struct vm_area_struct *vma, *prev, *last;
2663 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2666 len = PAGE_ALIGN(len);
2670 /* Find the first overlapping VMA */
2671 vma = find_vma(mm, start);
2674 prev = vma->vm_prev;
2675 /* we have start < vma->vm_end */
2677 /* if it doesn't overlap, we have nothing.. */
2679 if (vma->vm_start >= end)
2683 * If we need to split any vma, do it now to save pain later.
2685 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2686 * unmapped vm_area_struct will remain in use: so lower split_vma
2687 * places tmp vma above, and higher split_vma places tmp vma below.
2689 if (start > vma->vm_start) {
2693 * Make sure that map_count on return from munmap() will
2694 * not exceed its limit; but let map_count go just above
2695 * its limit temporarily, to help free resources as expected.
2697 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2700 error = __split_vma(mm, vma, start, 0);
2706 /* Does it split the last one? */
2707 last = find_vma(mm, end);
2708 if (last && end > last->vm_start) {
2709 int error = __split_vma(mm, last, end, 1);
2713 vma = prev ? prev->vm_next : mm->mmap;
2716 * unlock any mlock()ed ranges before detaching vmas
2718 if (mm->locked_vm) {
2719 struct vm_area_struct *tmp = vma;
2720 while (tmp && tmp->vm_start < end) {
2721 if (tmp->vm_flags & VM_LOCKED) {
2722 mm->locked_vm -= vma_pages(tmp);
2723 munlock_vma_pages_all(tmp);
2730 * Remove the vma's, and unmap the actual pages
2732 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2733 unmap_region(mm, vma, prev, start, end);
2735 arch_unmap(mm, vma, start, end);
2737 /* Fix up all other VM information */
2738 remove_vma_list(mm, vma);
2742 EXPORT_SYMBOL(do_munmap);
2744 int vm_munmap(unsigned long start, size_t len)
2747 struct mm_struct *mm = current->mm;
2749 if (down_write_killable(&mm->mmap_sem))
2752 ret = do_munmap(mm, start, len);
2753 up_write(&mm->mmap_sem);
2756 EXPORT_SYMBOL(vm_munmap);
2758 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2761 struct mm_struct *mm = current->mm;
2763 profile_munmap(addr);
2764 if (down_write_killable(&mm->mmap_sem))
2766 ret = do_munmap(mm, addr, len);
2767 up_write(&mm->mmap_sem);
2773 * Emulation of deprecated remap_file_pages() syscall.
2775 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2776 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2779 struct mm_struct *mm = current->mm;
2780 struct vm_area_struct *vma;
2781 unsigned long populate = 0;
2782 unsigned long ret = -EINVAL;
2785 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2786 current->comm, current->pid);
2790 start = start & PAGE_MASK;
2791 size = size & PAGE_MASK;
2793 if (start + size <= start)
2796 /* Does pgoff wrap? */
2797 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2800 if (down_write_killable(&mm->mmap_sem))
2803 vma = find_vma(mm, start);
2805 if (!vma || !(vma->vm_flags & VM_SHARED))
2808 if (start < vma->vm_start)
2811 if (start + size > vma->vm_end) {
2812 struct vm_area_struct *next;
2814 for (next = vma->vm_next; next; next = next->vm_next) {
2815 /* hole between vmas ? */
2816 if (next->vm_start != next->vm_prev->vm_end)
2819 if (next->vm_file != vma->vm_file)
2822 if (next->vm_flags != vma->vm_flags)
2825 if (start + size <= next->vm_end)
2833 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2834 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2835 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2837 flags &= MAP_NONBLOCK;
2838 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2839 if (vma->vm_flags & VM_LOCKED) {
2840 struct vm_area_struct *tmp;
2841 flags |= MAP_LOCKED;
2843 /* drop PG_Mlocked flag for over-mapped range */
2844 for (tmp = vma; tmp->vm_start >= start + size;
2845 tmp = tmp->vm_next) {
2847 * Split pmd and munlock page on the border
2850 vma_adjust_trans_huge(tmp, start, start + size, 0);
2852 munlock_vma_pages_range(tmp,
2853 max(tmp->vm_start, start),
2854 min(tmp->vm_end, start + size));
2858 file = get_file(vma->vm_file);
2859 ret = do_mmap_pgoff(vma->vm_file, start, size,
2860 prot, flags, pgoff, &populate);
2863 up_write(&mm->mmap_sem);
2865 mm_populate(ret, populate);
2866 if (!IS_ERR_VALUE(ret))
2871 static inline void verify_mm_writelocked(struct mm_struct *mm)
2873 #ifdef CONFIG_DEBUG_VM
2874 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2876 up_read(&mm->mmap_sem);
2882 * this is really a simplified "do_mmap". it only handles
2883 * anonymous maps. eventually we may be able to do some
2884 * brk-specific accounting here.
2886 static int do_brk(unsigned long addr, unsigned long len)
2888 struct mm_struct *mm = current->mm;
2889 struct vm_area_struct *vma, *prev;
2890 unsigned long flags;
2891 struct rb_node **rb_link, *rb_parent;
2892 pgoff_t pgoff = addr >> PAGE_SHIFT;
2895 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2897 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2898 if (offset_in_page(error))
2901 error = mlock_future_check(mm, mm->def_flags, len);
2906 * mm->mmap_sem is required to protect against another thread
2907 * changing the mappings in case we sleep.
2909 verify_mm_writelocked(mm);
2912 * Clear old maps. this also does some error checking for us
2914 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2916 if (do_munmap(mm, addr, len))
2920 /* Check against address space limits *after* clearing old maps... */
2921 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2924 if (mm->map_count > sysctl_max_map_count)
2927 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2930 /* Can we just expand an old private anonymous mapping? */
2931 vma = vma_merge(mm, prev, addr, addr + len, flags,
2932 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
2937 * create a vma struct for an anonymous mapping
2939 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2941 vm_unacct_memory(len >> PAGE_SHIFT);
2945 INIT_LIST_HEAD(&vma->anon_vma_chain);
2947 vma->vm_start = addr;
2948 vma->vm_end = addr + len;
2949 vma->vm_pgoff = pgoff;
2950 vma->vm_flags = flags;
2951 vma->vm_page_prot = vm_get_page_prot(flags);
2952 vma_link(mm, vma, prev, rb_link, rb_parent);
2954 perf_event_mmap(vma);
2955 mm->total_vm += len >> PAGE_SHIFT;
2956 mm->data_vm += len >> PAGE_SHIFT;
2957 if (flags & VM_LOCKED)
2958 mm->locked_vm += (len >> PAGE_SHIFT);
2959 vma->vm_flags |= VM_SOFTDIRTY;
2963 int vm_brk(unsigned long addr, unsigned long request)
2965 struct mm_struct *mm = current->mm;
2970 len = PAGE_ALIGN(request);
2976 if (down_write_killable(&mm->mmap_sem))
2979 ret = do_brk(addr, len);
2980 populate = ((mm->def_flags & VM_LOCKED) != 0);
2981 up_write(&mm->mmap_sem);
2982 if (populate && !ret)
2983 mm_populate(addr, len);
2986 EXPORT_SYMBOL(vm_brk);
2988 /* Release all mmaps. */
2989 void exit_mmap(struct mm_struct *mm)
2991 struct mmu_gather tlb;
2992 struct vm_area_struct *vma;
2993 unsigned long nr_accounted = 0;
2995 /* mm's last user has gone, and its about to be pulled down */
2996 mmu_notifier_release(mm);
2998 if (mm->locked_vm) {
3001 if (vma->vm_flags & VM_LOCKED)
3002 munlock_vma_pages_all(vma);
3010 if (!vma) /* Can happen if dup_mmap() received an OOM */
3015 tlb_gather_mmu(&tlb, mm, 0, -1);
3016 /* update_hiwater_rss(mm) here? but nobody should be looking */
3017 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3018 unmap_vmas(&tlb, vma, 0, -1);
3020 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3021 tlb_finish_mmu(&tlb, 0, -1);
3024 * Walk the list again, actually closing and freeing it,
3025 * with preemption enabled, without holding any MM locks.
3028 if (vma->vm_flags & VM_ACCOUNT)
3029 nr_accounted += vma_pages(vma);
3030 vma = remove_vma(vma);
3032 vm_unacct_memory(nr_accounted);
3035 /* Insert vm structure into process list sorted by address
3036 * and into the inode's i_mmap tree. If vm_file is non-NULL
3037 * then i_mmap_rwsem is taken here.
3039 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3041 struct vm_area_struct *prev;
3042 struct rb_node **rb_link, *rb_parent;
3044 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3045 &prev, &rb_link, &rb_parent))
3047 if ((vma->vm_flags & VM_ACCOUNT) &&
3048 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3052 * The vm_pgoff of a purely anonymous vma should be irrelevant
3053 * until its first write fault, when page's anon_vma and index
3054 * are set. But now set the vm_pgoff it will almost certainly
3055 * end up with (unless mremap moves it elsewhere before that
3056 * first wfault), so /proc/pid/maps tells a consistent story.
3058 * By setting it to reflect the virtual start address of the
3059 * vma, merges and splits can happen in a seamless way, just
3060 * using the existing file pgoff checks and manipulations.
3061 * Similarly in do_mmap_pgoff and in do_brk.
3063 if (vma_is_anonymous(vma)) {
3064 BUG_ON(vma->anon_vma);
3065 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3068 vma_link(mm, vma, prev, rb_link, rb_parent);
3073 * Copy the vma structure to a new location in the same mm,
3074 * prior to moving page table entries, to effect an mremap move.
3076 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3077 unsigned long addr, unsigned long len, pgoff_t pgoff,
3078 bool *need_rmap_locks)
3080 struct vm_area_struct *vma = *vmap;
3081 unsigned long vma_start = vma->vm_start;
3082 struct mm_struct *mm = vma->vm_mm;
3083 struct vm_area_struct *new_vma, *prev;
3084 struct rb_node **rb_link, *rb_parent;
3085 bool faulted_in_anon_vma = true;
3088 * If anonymous vma has not yet been faulted, update new pgoff
3089 * to match new location, to increase its chance of merging.
3091 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3092 pgoff = addr >> PAGE_SHIFT;
3093 faulted_in_anon_vma = false;
3096 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3097 return NULL; /* should never get here */
3098 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3099 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3100 vma->vm_userfaultfd_ctx, vma_get_anon_name(vma));
3103 * Source vma may have been merged into new_vma
3105 if (unlikely(vma_start >= new_vma->vm_start &&
3106 vma_start < new_vma->vm_end)) {
3108 * The only way we can get a vma_merge with
3109 * self during an mremap is if the vma hasn't
3110 * been faulted in yet and we were allowed to
3111 * reset the dst vma->vm_pgoff to the
3112 * destination address of the mremap to allow
3113 * the merge to happen. mremap must change the
3114 * vm_pgoff linearity between src and dst vmas
3115 * (in turn preventing a vma_merge) to be
3116 * safe. It is only safe to keep the vm_pgoff
3117 * linear if there are no pages mapped yet.
3119 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3120 *vmap = vma = new_vma;
3122 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3124 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3128 new_vma->vm_start = addr;
3129 new_vma->vm_end = addr + len;
3130 new_vma->vm_pgoff = pgoff;
3131 if (vma_dup_policy(vma, new_vma))
3133 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3134 if (anon_vma_clone(new_vma, vma))
3135 goto out_free_mempol;
3136 if (new_vma->vm_file)
3137 get_file(new_vma->vm_file);
3138 if (new_vma->vm_ops && new_vma->vm_ops->open)
3139 new_vma->vm_ops->open(new_vma);
3140 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3141 *need_rmap_locks = false;
3146 mpol_put(vma_policy(new_vma));
3148 kmem_cache_free(vm_area_cachep, new_vma);
3154 * Return true if the calling process may expand its vm space by the passed
3157 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3159 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3162 if (is_data_mapping(flags) &&
3163 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3164 /* Workaround for Valgrind */
3165 if (rlimit(RLIMIT_DATA) == 0 &&
3166 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3168 if (!ignore_rlimit_data) {
3169 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits or use boot option ignore_rlimit_data.\n",
3170 current->comm, current->pid,
3171 (mm->data_vm + npages) << PAGE_SHIFT,
3172 rlimit(RLIMIT_DATA));
3180 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3182 mm->total_vm += npages;
3184 if (is_exec_mapping(flags))
3185 mm->exec_vm += npages;
3186 else if (is_stack_mapping(flags))
3187 mm->stack_vm += npages;
3188 else if (is_data_mapping(flags))
3189 mm->data_vm += npages;
3192 static int special_mapping_fault(struct vm_area_struct *vma,
3193 struct vm_fault *vmf);
3196 * Having a close hook prevents vma merging regardless of flags.
3198 static void special_mapping_close(struct vm_area_struct *vma)
3202 static const char *special_mapping_name(struct vm_area_struct *vma)
3204 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3207 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3209 struct vm_special_mapping *sm = new_vma->vm_private_data;
3212 return sm->mremap(sm, new_vma);
3216 static const struct vm_operations_struct special_mapping_vmops = {
3217 .close = special_mapping_close,
3218 .fault = special_mapping_fault,
3219 .mremap = special_mapping_mremap,
3220 .name = special_mapping_name,
3223 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3224 .close = special_mapping_close,
3225 .fault = special_mapping_fault,
3228 static int special_mapping_fault(struct vm_area_struct *vma,
3229 struct vm_fault *vmf)
3232 struct page **pages;
3234 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3235 pages = vma->vm_private_data;
3237 struct vm_special_mapping *sm = vma->vm_private_data;
3240 return sm->fault(sm, vma, vmf);
3245 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3249 struct page *page = *pages;
3255 return VM_FAULT_SIGBUS;
3258 static struct vm_area_struct *__install_special_mapping(
3259 struct mm_struct *mm,
3260 unsigned long addr, unsigned long len,
3261 unsigned long vm_flags, void *priv,
3262 const struct vm_operations_struct *ops)
3265 struct vm_area_struct *vma;
3267 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3268 if (unlikely(vma == NULL))
3269 return ERR_PTR(-ENOMEM);
3271 INIT_LIST_HEAD(&vma->anon_vma_chain);
3273 vma->vm_start = addr;
3274 vma->vm_end = addr + len;
3276 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3277 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3280 vma->vm_private_data = priv;
3282 ret = insert_vm_struct(mm, vma);
3286 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3288 perf_event_mmap(vma);
3293 kmem_cache_free(vm_area_cachep, vma);
3294 return ERR_PTR(ret);
3297 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3298 const struct vm_special_mapping *sm)
3300 return vma->vm_private_data == sm &&
3301 (vma->vm_ops == &special_mapping_vmops ||
3302 vma->vm_ops == &legacy_special_mapping_vmops);
3306 * Called with mm->mmap_sem held for writing.
3307 * Insert a new vma covering the given region, with the given flags.
3308 * Its pages are supplied by the given array of struct page *.
3309 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3310 * The region past the last page supplied will always produce SIGBUS.
3311 * The array pointer and the pages it points to are assumed to stay alive
3312 * for as long as this mapping might exist.
3314 struct vm_area_struct *_install_special_mapping(
3315 struct mm_struct *mm,
3316 unsigned long addr, unsigned long len,
3317 unsigned long vm_flags, const struct vm_special_mapping *spec)
3319 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3320 &special_mapping_vmops);
3323 int install_special_mapping(struct mm_struct *mm,
3324 unsigned long addr, unsigned long len,
3325 unsigned long vm_flags, struct page **pages)
3327 struct vm_area_struct *vma = __install_special_mapping(
3328 mm, addr, len, vm_flags, (void *)pages,
3329 &legacy_special_mapping_vmops);
3331 return PTR_ERR_OR_ZERO(vma);
3334 static DEFINE_MUTEX(mm_all_locks_mutex);
3336 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3338 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3340 * The LSB of head.next can't change from under us
3341 * because we hold the mm_all_locks_mutex.
3343 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3345 * We can safely modify head.next after taking the
3346 * anon_vma->root->rwsem. If some other vma in this mm shares
3347 * the same anon_vma we won't take it again.
3349 * No need of atomic instructions here, head.next
3350 * can't change from under us thanks to the
3351 * anon_vma->root->rwsem.
3353 if (__test_and_set_bit(0, (unsigned long *)
3354 &anon_vma->root->rb_root.rb_node))
3359 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3361 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3363 * AS_MM_ALL_LOCKS can't change from under us because
3364 * we hold the mm_all_locks_mutex.
3366 * Operations on ->flags have to be atomic because
3367 * even if AS_MM_ALL_LOCKS is stable thanks to the
3368 * mm_all_locks_mutex, there may be other cpus
3369 * changing other bitflags in parallel to us.
3371 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3373 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3378 * This operation locks against the VM for all pte/vma/mm related
3379 * operations that could ever happen on a certain mm. This includes
3380 * vmtruncate, try_to_unmap, and all page faults.
3382 * The caller must take the mmap_sem in write mode before calling
3383 * mm_take_all_locks(). The caller isn't allowed to release the
3384 * mmap_sem until mm_drop_all_locks() returns.
3386 * mmap_sem in write mode is required in order to block all operations
3387 * that could modify pagetables and free pages without need of
3388 * altering the vma layout. It's also needed in write mode to avoid new
3389 * anon_vmas to be associated with existing vmas.
3391 * A single task can't take more than one mm_take_all_locks() in a row
3392 * or it would deadlock.
3394 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3395 * mapping->flags avoid to take the same lock twice, if more than one
3396 * vma in this mm is backed by the same anon_vma or address_space.
3398 * We take locks in following order, accordingly to comment at beginning
3400 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3402 * - all i_mmap_rwsem locks;
3403 * - all anon_vma->rwseml
3405 * We can take all locks within these types randomly because the VM code
3406 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3407 * mm_all_locks_mutex.
3409 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3410 * that may have to take thousand of locks.
3412 * mm_take_all_locks() can fail if it's interrupted by signals.
3414 int mm_take_all_locks(struct mm_struct *mm)
3416 struct vm_area_struct *vma;
3417 struct anon_vma_chain *avc;
3419 BUG_ON(down_read_trylock(&mm->mmap_sem));
3421 mutex_lock(&mm_all_locks_mutex);
3423 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3424 if (signal_pending(current))
3426 if (vma->vm_file && vma->vm_file->f_mapping &&
3427 is_vm_hugetlb_page(vma))
3428 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3431 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3432 if (signal_pending(current))
3434 if (vma->vm_file && vma->vm_file->f_mapping &&
3435 !is_vm_hugetlb_page(vma))
3436 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3439 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3440 if (signal_pending(current))
3443 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3444 vm_lock_anon_vma(mm, avc->anon_vma);
3450 mm_drop_all_locks(mm);
3454 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3456 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3458 * The LSB of head.next can't change to 0 from under
3459 * us because we hold the mm_all_locks_mutex.
3461 * We must however clear the bitflag before unlocking
3462 * the vma so the users using the anon_vma->rb_root will
3463 * never see our bitflag.
3465 * No need of atomic instructions here, head.next
3466 * can't change from under us until we release the
3467 * anon_vma->root->rwsem.
3469 if (!__test_and_clear_bit(0, (unsigned long *)
3470 &anon_vma->root->rb_root.rb_node))
3472 anon_vma_unlock_write(anon_vma);
3476 static void vm_unlock_mapping(struct address_space *mapping)
3478 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3480 * AS_MM_ALL_LOCKS can't change to 0 from under us
3481 * because we hold the mm_all_locks_mutex.
3483 i_mmap_unlock_write(mapping);
3484 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3491 * The mmap_sem cannot be released by the caller until
3492 * mm_drop_all_locks() returns.
3494 void mm_drop_all_locks(struct mm_struct *mm)
3496 struct vm_area_struct *vma;
3497 struct anon_vma_chain *avc;
3499 BUG_ON(down_read_trylock(&mm->mmap_sem));
3500 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3502 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3504 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3505 vm_unlock_anon_vma(avc->anon_vma);
3506 if (vma->vm_file && vma->vm_file->f_mapping)
3507 vm_unlock_mapping(vma->vm_file->f_mapping);
3510 mutex_unlock(&mm_all_locks_mutex);
3514 * initialise the VMA slab
3516 void __init mmap_init(void)
3520 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3525 * Initialise sysctl_user_reserve_kbytes.
3527 * This is intended to prevent a user from starting a single memory hogging
3528 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3531 * The default value is min(3% of free memory, 128MB)
3532 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3534 static int init_user_reserve(void)
3536 unsigned long free_kbytes;
3538 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3540 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3543 subsys_initcall(init_user_reserve);
3546 * Initialise sysctl_admin_reserve_kbytes.
3548 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3549 * to log in and kill a memory hogging process.
3551 * Systems with more than 256MB will reserve 8MB, enough to recover
3552 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3553 * only reserve 3% of free pages by default.
3555 static int init_admin_reserve(void)
3557 unsigned long free_kbytes;
3559 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3561 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3564 subsys_initcall(init_admin_reserve);
3567 * Reinititalise user and admin reserves if memory is added or removed.
3569 * The default user reserve max is 128MB, and the default max for the
3570 * admin reserve is 8MB. These are usually, but not always, enough to
3571 * enable recovery from a memory hogging process using login/sshd, a shell,
3572 * and tools like top. It may make sense to increase or even disable the
3573 * reserve depending on the existence of swap or variations in the recovery
3574 * tools. So, the admin may have changed them.
3576 * If memory is added and the reserves have been eliminated or increased above
3577 * the default max, then we'll trust the admin.
3579 * If memory is removed and there isn't enough free memory, then we
3580 * need to reset the reserves.
3582 * Otherwise keep the reserve set by the admin.
3584 static int reserve_mem_notifier(struct notifier_block *nb,
3585 unsigned long action, void *data)
3587 unsigned long tmp, free_kbytes;
3591 /* Default max is 128MB. Leave alone if modified by operator. */
3592 tmp = sysctl_user_reserve_kbytes;
3593 if (0 < tmp && tmp < (1UL << 17))
3594 init_user_reserve();
3596 /* Default max is 8MB. Leave alone if modified by operator. */
3597 tmp = sysctl_admin_reserve_kbytes;
3598 if (0 < tmp && tmp < (1UL << 13))
3599 init_admin_reserve();
3603 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3605 if (sysctl_user_reserve_kbytes > free_kbytes) {
3606 init_user_reserve();
3607 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3608 sysctl_user_reserve_kbytes);
3611 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3612 init_admin_reserve();
3613 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3614 sysctl_admin_reserve_kbytes);
3623 static struct notifier_block reserve_mem_nb = {
3624 .notifier_call = reserve_mem_notifier,
3627 static int __meminit init_reserve_notifier(void)
3629 if (register_hotmemory_notifier(&reserve_mem_nb))
3630 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3634 subsys_initcall(init_reserve_notifier);