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 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
97 pgprot_t protection_map[16] = {
98 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
99 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
102 pgprot_t vm_get_page_prot(unsigned long vm_flags)
104 return __pgprot(pgprot_val(protection_map[vm_flags &
105 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
106 pgprot_val(arch_vm_get_page_prot(vm_flags)));
108 EXPORT_SYMBOL(vm_get_page_prot);
110 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
112 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
115 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
116 void vma_set_page_prot(struct vm_area_struct *vma)
118 unsigned long vm_flags = vma->vm_flags;
119 pgprot_t vm_page_prot;
121 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
122 if (vma_wants_writenotify(vma, vm_page_prot)) {
123 vm_flags &= ~VM_SHARED;
124 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
126 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
127 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
131 * Requires inode->i_mapping->i_mmap_rwsem
133 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
134 struct file *file, struct address_space *mapping)
136 if (vma->vm_flags & VM_DENYWRITE)
137 atomic_inc(&file_inode(file)->i_writecount);
138 if (vma->vm_flags & VM_SHARED)
139 mapping_unmap_writable(mapping);
141 flush_dcache_mmap_lock(mapping);
142 vma_interval_tree_remove(vma, &mapping->i_mmap);
143 flush_dcache_mmap_unlock(mapping);
147 * Unlink a file-based vm structure from its interval tree, to hide
148 * vma from rmap and vmtruncate before freeing its page tables.
150 void unlink_file_vma(struct vm_area_struct *vma)
152 struct file *file = vma->vm_file;
155 struct address_space *mapping = file->f_mapping;
156 i_mmap_lock_write(mapping);
157 __remove_shared_vm_struct(vma, file, mapping);
158 i_mmap_unlock_write(mapping);
163 * Close a vm structure and free it, returning the next.
165 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
167 struct vm_area_struct *next = vma->vm_next;
170 if (vma->vm_ops && vma->vm_ops->close)
171 vma->vm_ops->close(vma);
174 mpol_put(vma_policy(vma));
175 kmem_cache_free(vm_area_cachep, vma);
179 static int do_brk(unsigned long addr, unsigned long len);
181 SYSCALL_DEFINE1(brk, unsigned long, brk)
183 unsigned long retval;
184 unsigned long newbrk, oldbrk;
185 struct mm_struct *mm = current->mm;
186 struct vm_area_struct *next;
187 unsigned long min_brk;
190 if (down_write_killable(&mm->mmap_sem))
193 #ifdef CONFIG_COMPAT_BRK
195 * CONFIG_COMPAT_BRK can still be overridden by setting
196 * randomize_va_space to 2, which will still cause mm->start_brk
197 * to be arbitrarily shifted
199 if (current->brk_randomized)
200 min_brk = mm->start_brk;
202 min_brk = mm->end_data;
204 min_brk = mm->start_brk;
210 * Check against rlimit here. If this check is done later after the test
211 * of oldbrk with newbrk then it can escape the test and let the data
212 * segment grow beyond its set limit the in case where the limit is
213 * not page aligned -Ram Gupta
215 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
216 mm->end_data, mm->start_data))
219 newbrk = PAGE_ALIGN(brk);
220 oldbrk = PAGE_ALIGN(mm->brk);
221 if (oldbrk == newbrk)
224 /* Always allow shrinking brk. */
225 if (brk <= mm->brk) {
226 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
231 /* Check against existing mmap mappings. */
232 next = find_vma(mm, oldbrk);
233 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
236 /* Ok, looks good - let it rip. */
237 if (do_brk(oldbrk, newbrk-oldbrk) < 0)
242 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
243 up_write(&mm->mmap_sem);
245 mm_populate(oldbrk, newbrk - oldbrk);
250 up_write(&mm->mmap_sem);
254 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
256 unsigned long max, prev_end, subtree_gap;
259 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
260 * allow two stack_guard_gaps between them here, and when choosing
261 * an unmapped area; whereas when expanding we only require one.
262 * That's a little inconsistent, but keeps the code here simpler.
264 max = vm_start_gap(vma);
266 prev_end = vm_end_gap(vma->vm_prev);
272 if (vma->vm_rb.rb_left) {
273 subtree_gap = rb_entry(vma->vm_rb.rb_left,
274 struct vm_area_struct, vm_rb)->rb_subtree_gap;
275 if (subtree_gap > max)
278 if (vma->vm_rb.rb_right) {
279 subtree_gap = rb_entry(vma->vm_rb.rb_right,
280 struct vm_area_struct, vm_rb)->rb_subtree_gap;
281 if (subtree_gap > max)
287 #ifdef CONFIG_DEBUG_VM_RB
288 static int browse_rb(struct mm_struct *mm)
290 struct rb_root *root = &mm->mm_rb;
291 int i = 0, j, bug = 0;
292 struct rb_node *nd, *pn = NULL;
293 unsigned long prev = 0, pend = 0;
295 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
296 struct vm_area_struct *vma;
297 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
298 if (vma->vm_start < prev) {
299 pr_emerg("vm_start %lx < prev %lx\n",
300 vma->vm_start, prev);
303 if (vma->vm_start < pend) {
304 pr_emerg("vm_start %lx < pend %lx\n",
305 vma->vm_start, pend);
308 if (vma->vm_start > vma->vm_end) {
309 pr_emerg("vm_start %lx > vm_end %lx\n",
310 vma->vm_start, vma->vm_end);
313 spin_lock(&mm->page_table_lock);
314 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
315 pr_emerg("free gap %lx, correct %lx\n",
317 vma_compute_subtree_gap(vma));
320 spin_unlock(&mm->page_table_lock);
323 prev = vma->vm_start;
327 for (nd = pn; nd; nd = rb_prev(nd))
330 pr_emerg("backwards %d, forwards %d\n", j, i);
336 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
340 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
341 struct vm_area_struct *vma;
342 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
343 VM_BUG_ON_VMA(vma != ignore &&
344 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
349 static void validate_mm(struct mm_struct *mm)
353 unsigned long highest_address = 0;
354 struct vm_area_struct *vma = mm->mmap;
357 struct anon_vma *anon_vma = vma->anon_vma;
358 struct anon_vma_chain *avc;
361 anon_vma_lock_read(anon_vma);
362 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
363 anon_vma_interval_tree_verify(avc);
364 anon_vma_unlock_read(anon_vma);
367 highest_address = vm_end_gap(vma);
371 if (i != mm->map_count) {
372 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
375 if (highest_address != mm->highest_vm_end) {
376 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
377 mm->highest_vm_end, highest_address);
381 if (i != mm->map_count) {
383 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
386 VM_BUG_ON_MM(bug, mm);
389 #define validate_mm_rb(root, ignore) do { } while (0)
390 #define validate_mm(mm) do { } while (0)
393 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
394 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
397 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
398 * vma->vm_prev->vm_end values changed, without modifying the vma's position
401 static void vma_gap_update(struct vm_area_struct *vma)
404 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
405 * function that does exacltly what we want.
407 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
410 static inline void vma_rb_insert(struct vm_area_struct *vma,
411 struct rb_root *root)
413 /* All rb_subtree_gap values must be consistent prior to insertion */
414 validate_mm_rb(root, NULL);
416 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
419 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
422 * Note rb_erase_augmented is a fairly large inline function,
423 * so make sure we instantiate it only once with our desired
424 * augmented rbtree callbacks.
426 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
429 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
430 struct rb_root *root,
431 struct vm_area_struct *ignore)
434 * All rb_subtree_gap values must be consistent prior to erase,
435 * with the possible exception of the "next" vma being erased if
436 * next->vm_start was reduced.
438 validate_mm_rb(root, ignore);
440 __vma_rb_erase(vma, root);
443 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
444 struct rb_root *root)
447 * All rb_subtree_gap values must be consistent prior to erase,
448 * with the possible exception of the vma being erased.
450 validate_mm_rb(root, vma);
452 __vma_rb_erase(vma, root);
456 * vma has some anon_vma assigned, and is already inserted on that
457 * anon_vma's interval trees.
459 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
460 * vma must be removed from the anon_vma's interval trees using
461 * anon_vma_interval_tree_pre_update_vma().
463 * After the update, the vma will be reinserted using
464 * anon_vma_interval_tree_post_update_vma().
466 * The entire update must be protected by exclusive mmap_sem and by
467 * the root anon_vma's mutex.
470 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
472 struct anon_vma_chain *avc;
474 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
475 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
479 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
481 struct anon_vma_chain *avc;
483 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
484 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
487 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
488 unsigned long end, struct vm_area_struct **pprev,
489 struct rb_node ***rb_link, struct rb_node **rb_parent)
491 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
493 __rb_link = &mm->mm_rb.rb_node;
494 rb_prev = __rb_parent = NULL;
497 struct vm_area_struct *vma_tmp;
499 __rb_parent = *__rb_link;
500 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
502 if (vma_tmp->vm_end > addr) {
503 /* Fail if an existing vma overlaps the area */
504 if (vma_tmp->vm_start < end)
506 __rb_link = &__rb_parent->rb_left;
508 rb_prev = __rb_parent;
509 __rb_link = &__rb_parent->rb_right;
515 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
516 *rb_link = __rb_link;
517 *rb_parent = __rb_parent;
521 static unsigned long count_vma_pages_range(struct mm_struct *mm,
522 unsigned long addr, unsigned long end)
524 unsigned long nr_pages = 0;
525 struct vm_area_struct *vma;
527 /* Find first overlaping mapping */
528 vma = find_vma_intersection(mm, addr, end);
532 nr_pages = (min(end, vma->vm_end) -
533 max(addr, vma->vm_start)) >> PAGE_SHIFT;
535 /* Iterate over the rest of the overlaps */
536 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
537 unsigned long overlap_len;
539 if (vma->vm_start > end)
542 overlap_len = min(end, vma->vm_end) - vma->vm_start;
543 nr_pages += overlap_len >> PAGE_SHIFT;
549 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
550 struct rb_node **rb_link, struct rb_node *rb_parent)
552 /* Update tracking information for the gap following the new vma. */
554 vma_gap_update(vma->vm_next);
556 mm->highest_vm_end = vm_end_gap(vma);
559 * vma->vm_prev wasn't known when we followed the rbtree to find the
560 * correct insertion point for that vma. As a result, we could not
561 * update the vma vm_rb parents rb_subtree_gap values on the way down.
562 * So, we first insert the vma with a zero rb_subtree_gap value
563 * (to be consistent with what we did on the way down), and then
564 * immediately update the gap to the correct value. Finally we
565 * rebalance the rbtree after all augmented values have been set.
567 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
568 vma->rb_subtree_gap = 0;
570 vma_rb_insert(vma, &mm->mm_rb);
573 static void __vma_link_file(struct vm_area_struct *vma)
579 struct address_space *mapping = file->f_mapping;
581 if (vma->vm_flags & VM_DENYWRITE)
582 atomic_dec(&file_inode(file)->i_writecount);
583 if (vma->vm_flags & VM_SHARED)
584 atomic_inc(&mapping->i_mmap_writable);
586 flush_dcache_mmap_lock(mapping);
587 vma_interval_tree_insert(vma, &mapping->i_mmap);
588 flush_dcache_mmap_unlock(mapping);
593 __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 __vma_link_list(mm, vma, prev, rb_parent);
598 __vma_link_rb(mm, vma, rb_link, rb_parent);
601 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
602 struct vm_area_struct *prev, struct rb_node **rb_link,
603 struct rb_node *rb_parent)
605 struct address_space *mapping = NULL;
608 mapping = vma->vm_file->f_mapping;
609 i_mmap_lock_write(mapping);
612 __vma_link(mm, vma, prev, rb_link, rb_parent);
613 __vma_link_file(vma);
616 i_mmap_unlock_write(mapping);
623 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
624 * mm's list and rbtree. It has already been inserted into the interval tree.
626 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
628 struct vm_area_struct *prev;
629 struct rb_node **rb_link, *rb_parent;
631 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
632 &prev, &rb_link, &rb_parent))
634 __vma_link(mm, vma, prev, rb_link, rb_parent);
638 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
639 struct vm_area_struct *vma,
640 struct vm_area_struct *prev,
642 struct vm_area_struct *ignore)
644 struct vm_area_struct *next;
646 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
649 prev->vm_next = next;
653 prev->vm_next = next;
658 next->vm_prev = prev;
661 vmacache_invalidate(mm);
664 static inline void __vma_unlink_prev(struct mm_struct *mm,
665 struct vm_area_struct *vma,
666 struct vm_area_struct *prev)
668 __vma_unlink_common(mm, vma, prev, true, vma);
672 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
673 * is already present in an i_mmap tree without adjusting the tree.
674 * The following helper function should be used when such adjustments
675 * are necessary. The "insert" vma (if any) is to be inserted
676 * before we drop the necessary locks.
678 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
679 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
680 struct vm_area_struct *expand)
682 struct mm_struct *mm = vma->vm_mm;
683 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
684 struct address_space *mapping = NULL;
685 struct rb_root *root = NULL;
686 struct anon_vma *anon_vma = NULL;
687 struct file *file = vma->vm_file;
688 bool start_changed = false, end_changed = false;
689 long adjust_next = 0;
692 if (next && !insert) {
693 struct vm_area_struct *exporter = NULL, *importer = NULL;
695 if (end >= next->vm_end) {
697 * vma expands, overlapping all the next, and
698 * perhaps the one after too (mprotect case 6).
699 * The only other cases that gets here are
700 * case 1, case 7 and case 8.
702 if (next == expand) {
704 * The only case where we don't expand "vma"
705 * and we expand "next" instead is case 8.
707 VM_WARN_ON(end != next->vm_end);
709 * remove_next == 3 means we're
710 * removing "vma" and that to do so we
711 * swapped "vma" and "next".
714 VM_WARN_ON(file != next->vm_file);
717 VM_WARN_ON(expand != vma);
719 * case 1, 6, 7, remove_next == 2 is case 6,
720 * remove_next == 1 is case 1 or 7.
722 remove_next = 1 + (end > next->vm_end);
723 VM_WARN_ON(remove_next == 2 &&
724 end != next->vm_next->vm_end);
725 VM_WARN_ON(remove_next == 1 &&
726 end != next->vm_end);
727 /* trim end to next, for case 6 first pass */
735 * If next doesn't have anon_vma, import from vma after
736 * next, if the vma overlaps with it.
738 if (remove_next == 2 && !next->anon_vma)
739 exporter = next->vm_next;
741 } else if (end > next->vm_start) {
743 * vma expands, overlapping part of the next:
744 * mprotect case 5 shifting the boundary up.
746 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
749 VM_WARN_ON(expand != importer);
750 } else if (end < vma->vm_end) {
752 * vma shrinks, and !insert tells it's not
753 * split_vma inserting another: so it must be
754 * mprotect case 4 shifting the boundary down.
756 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
759 VM_WARN_ON(expand != importer);
763 * Easily overlooked: when mprotect shifts the boundary,
764 * make sure the expanding vma has anon_vma set if the
765 * shrinking vma had, to cover any anon pages imported.
767 if (exporter && exporter->anon_vma && !importer->anon_vma) {
770 importer->anon_vma = exporter->anon_vma;
771 error = anon_vma_clone(importer, exporter);
777 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
780 mapping = file->f_mapping;
781 root = &mapping->i_mmap;
782 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
785 uprobe_munmap(next, next->vm_start, next->vm_end);
787 i_mmap_lock_write(mapping);
790 * Put into interval tree now, so instantiated pages
791 * are visible to arm/parisc __flush_dcache_page
792 * throughout; but we cannot insert into address
793 * space until vma start or end is updated.
795 __vma_link_file(insert);
799 anon_vma = vma->anon_vma;
800 if (!anon_vma && adjust_next)
801 anon_vma = next->anon_vma;
803 VM_WARN_ON(adjust_next && next->anon_vma &&
804 anon_vma != next->anon_vma);
805 anon_vma_lock_write(anon_vma);
806 anon_vma_interval_tree_pre_update_vma(vma);
808 anon_vma_interval_tree_pre_update_vma(next);
812 flush_dcache_mmap_lock(mapping);
813 vma_interval_tree_remove(vma, root);
815 vma_interval_tree_remove(next, root);
818 if (start != vma->vm_start) {
819 vma->vm_start = start;
820 start_changed = true;
822 if (end != vma->vm_end) {
826 vma->vm_pgoff = pgoff;
828 next->vm_start += adjust_next << PAGE_SHIFT;
829 next->vm_pgoff += adjust_next;
834 vma_interval_tree_insert(next, root);
835 vma_interval_tree_insert(vma, root);
836 flush_dcache_mmap_unlock(mapping);
841 * vma_merge has merged next into vma, and needs
842 * us to remove next before dropping the locks.
844 if (remove_next != 3)
845 __vma_unlink_prev(mm, next, vma);
848 * vma is not before next if they've been
851 * pre-swap() next->vm_start was reduced so
852 * tell validate_mm_rb to ignore pre-swap()
853 * "next" (which is stored in post-swap()
856 __vma_unlink_common(mm, next, NULL, false, vma);
858 __remove_shared_vm_struct(next, file, mapping);
861 * split_vma has split insert from vma, and needs
862 * us to insert it before dropping the locks
863 * (it may either follow vma or precede it).
865 __insert_vm_struct(mm, insert);
871 mm->highest_vm_end = vm_end_gap(vma);
872 else if (!adjust_next)
873 vma_gap_update(next);
878 anon_vma_interval_tree_post_update_vma(vma);
880 anon_vma_interval_tree_post_update_vma(next);
881 anon_vma_unlock_write(anon_vma);
884 i_mmap_unlock_write(mapping);
895 uprobe_munmap(next, next->vm_start, next->vm_end);
899 anon_vma_merge(vma, next);
901 mpol_put(vma_policy(next));
902 kmem_cache_free(vm_area_cachep, next);
904 * In mprotect's case 6 (see comments on vma_merge),
905 * we must remove another next too. It would clutter
906 * up the code too much to do both in one go.
908 if (remove_next != 3) {
910 * If "next" was removed and vma->vm_end was
911 * expanded (up) over it, in turn
912 * "next->vm_prev->vm_end" changed and the
913 * "vma->vm_next" gap must be updated.
918 * For the scope of the comment "next" and
919 * "vma" considered pre-swap(): if "vma" was
920 * removed, next->vm_start was expanded (down)
921 * over it and the "next" gap must be updated.
922 * Because of the swap() the post-swap() "vma"
923 * actually points to pre-swap() "next"
924 * (post-swap() "next" as opposed is now a
929 if (remove_next == 2) {
935 vma_gap_update(next);
938 * If remove_next == 2 we obviously can't
941 * If remove_next == 3 we can't reach this
942 * path because pre-swap() next is always not
943 * NULL. pre-swap() "next" is not being
944 * removed and its next->vm_end is not altered
945 * (and furthermore "end" already matches
946 * next->vm_end in remove_next == 3).
948 * We reach this only in the remove_next == 1
949 * case if the "next" vma that was removed was
950 * the highest vma of the mm. However in such
951 * case next->vm_end == "end" and the extended
952 * "vma" has vma->vm_end == next->vm_end so
953 * mm->highest_vm_end doesn't need any update
954 * in remove_next == 1 case.
956 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
968 * If the vma has a ->close operation then the driver probably needs to release
969 * per-vma resources, so we don't attempt to merge those.
971 static inline int is_mergeable_vma(struct vm_area_struct *vma,
972 struct file *file, unsigned long vm_flags,
973 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
976 * VM_SOFTDIRTY should not prevent from VMA merging, if we
977 * match the flags but dirty bit -- the caller should mark
978 * merged VMA as dirty. If dirty bit won't be excluded from
979 * comparison, we increase pressue on the memory system forcing
980 * the kernel to generate new VMAs when old one could be
983 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
985 if (vma->vm_file != file)
987 if (vma->vm_ops && vma->vm_ops->close)
989 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
994 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
995 struct anon_vma *anon_vma2,
996 struct vm_area_struct *vma)
999 * The list_is_singular() test is to avoid merging VMA cloned from
1000 * parents. This can improve scalability caused by anon_vma lock.
1002 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1003 list_is_singular(&vma->anon_vma_chain)))
1005 return anon_vma1 == anon_vma2;
1009 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1010 * in front of (at a lower virtual address and file offset than) the vma.
1012 * We cannot merge two vmas if they have differently assigned (non-NULL)
1013 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1015 * We don't check here for the merged mmap wrapping around the end of pagecache
1016 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1017 * wrap, nor mmaps which cover the final page at index -1UL.
1020 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1021 struct anon_vma *anon_vma, struct file *file,
1023 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1025 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1026 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1027 if (vma->vm_pgoff == vm_pgoff)
1034 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1035 * beyond (at a higher virtual address and file offset than) the vma.
1037 * We cannot merge two vmas if they have differently assigned (non-NULL)
1038 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1041 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1042 struct anon_vma *anon_vma, struct file *file,
1044 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1046 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1047 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1049 vm_pglen = vma_pages(vma);
1050 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1057 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1058 * whether that can be merged with its predecessor or its successor.
1059 * Or both (it neatly fills a hole).
1061 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1062 * certain not to be mapped by the time vma_merge is called; but when
1063 * called for mprotect, it is certain to be already mapped (either at
1064 * an offset within prev, or at the start of next), and the flags of
1065 * this area are about to be changed to vm_flags - and the no-change
1066 * case has already been eliminated.
1068 * The following mprotect cases have to be considered, where AAAA is
1069 * the area passed down from mprotect_fixup, never extending beyond one
1070 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1072 * AAAA AAAA AAAA AAAA
1073 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1074 * cannot merge might become might become might become
1075 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1076 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1077 * mremap move: PPPPXXXXXXXX 8
1079 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1080 * might become case 1 below case 2 below case 3 below
1082 * It is important for case 8 that the the vma NNNN overlapping the
1083 * region AAAA is never going to extended over XXXX. Instead XXXX must
1084 * be extended in region AAAA and NNNN must be removed. This way in
1085 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1086 * rmap_locks, the properties of the merged vma will be already
1087 * correct for the whole merged range. Some of those properties like
1088 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1089 * be correct for the whole merged range immediately after the
1090 * rmap_locks are released. Otherwise if XXXX would be removed and
1091 * NNNN would be extended over the XXXX range, remove_migration_ptes
1092 * or other rmap walkers (if working on addresses beyond the "end"
1093 * parameter) may establish ptes with the wrong permissions of NNNN
1094 * instead of the right permissions of XXXX.
1096 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1097 struct vm_area_struct *prev, unsigned long addr,
1098 unsigned long end, unsigned long vm_flags,
1099 struct anon_vma *anon_vma, struct file *file,
1100 pgoff_t pgoff, struct mempolicy *policy,
1101 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
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,
1134 vm_userfaultfd_ctx)) {
1136 * OK, it can. Can we now merge in the successor as well?
1138 if (next && end == next->vm_start &&
1139 mpol_equal(policy, vma_policy(next)) &&
1140 can_vma_merge_before(next, vm_flags,
1143 vm_userfaultfd_ctx) &&
1144 is_mergeable_anon_vma(prev->anon_vma,
1145 next->anon_vma, NULL)) {
1147 err = __vma_adjust(prev, prev->vm_start,
1148 next->vm_end, prev->vm_pgoff, NULL,
1150 } else /* cases 2, 5, 7 */
1151 err = __vma_adjust(prev, prev->vm_start,
1152 end, prev->vm_pgoff, NULL, prev);
1155 khugepaged_enter_vma_merge(prev, vm_flags);
1160 * Can this new request be merged in front of next?
1162 if (next && end == next->vm_start &&
1163 mpol_equal(policy, vma_policy(next)) &&
1164 can_vma_merge_before(next, vm_flags,
1165 anon_vma, file, pgoff+pglen,
1166 vm_userfaultfd_ctx)) {
1167 if (prev && addr < prev->vm_end) /* case 4 */
1168 err = __vma_adjust(prev, prev->vm_start,
1169 addr, prev->vm_pgoff, NULL, next);
1170 else { /* cases 3, 8 */
1171 err = __vma_adjust(area, addr, next->vm_end,
1172 next->vm_pgoff - pglen, NULL, next);
1174 * In case 3 area is already equal to next and
1175 * this is a noop, but in case 8 "area" has
1176 * been removed and next was expanded over it.
1182 khugepaged_enter_vma_merge(area, vm_flags);
1190 * Rough compatbility check to quickly see if it's even worth looking
1191 * at sharing an anon_vma.
1193 * They need to have the same vm_file, and the flags can only differ
1194 * in things that mprotect may change.
1196 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1197 * we can merge the two vma's. For example, we refuse to merge a vma if
1198 * there is a vm_ops->close() function, because that indicates that the
1199 * driver is doing some kind of reference counting. But that doesn't
1200 * really matter for the anon_vma sharing case.
1202 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1204 return a->vm_end == b->vm_start &&
1205 mpol_equal(vma_policy(a), vma_policy(b)) &&
1206 a->vm_file == b->vm_file &&
1207 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1208 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1212 * Do some basic sanity checking to see if we can re-use the anon_vma
1213 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1214 * the same as 'old', the other will be the new one that is trying
1215 * to share the anon_vma.
1217 * NOTE! This runs with mm_sem held for reading, so it is possible that
1218 * the anon_vma of 'old' is concurrently in the process of being set up
1219 * by another page fault trying to merge _that_. But that's ok: if it
1220 * is being set up, that automatically means that it will be a singleton
1221 * acceptable for merging, so we can do all of this optimistically. But
1222 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1224 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1225 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1226 * is to return an anon_vma that is "complex" due to having gone through
1229 * We also make sure that the two vma's are compatible (adjacent,
1230 * and with the same memory policies). That's all stable, even with just
1231 * a read lock on the mm_sem.
1233 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1235 if (anon_vma_compatible(a, b)) {
1236 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1238 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1245 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1246 * neighbouring vmas for a suitable anon_vma, before it goes off
1247 * to allocate a new anon_vma. It checks because a repetitive
1248 * sequence of mprotects and faults may otherwise lead to distinct
1249 * anon_vmas being allocated, preventing vma merge in subsequent
1252 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1254 struct anon_vma *anon_vma;
1255 struct vm_area_struct *near;
1257 near = vma->vm_next;
1261 anon_vma = reusable_anon_vma(near, vma, near);
1265 near = vma->vm_prev;
1269 anon_vma = reusable_anon_vma(near, near, vma);
1274 * There's no absolute need to look only at touching neighbours:
1275 * we could search further afield for "compatible" anon_vmas.
1276 * But it would probably just be a waste of time searching,
1277 * or lead to too many vmas hanging off the same anon_vma.
1278 * We're trying to allow mprotect remerging later on,
1279 * not trying to minimize memory used for anon_vmas.
1285 * If a hint addr is less than mmap_min_addr change hint to be as
1286 * low as possible but still greater than mmap_min_addr
1288 static inline unsigned long round_hint_to_min(unsigned long hint)
1291 if (((void *)hint != NULL) &&
1292 (hint < mmap_min_addr))
1293 return PAGE_ALIGN(mmap_min_addr);
1297 static inline int mlock_future_check(struct mm_struct *mm,
1298 unsigned long flags,
1301 unsigned long locked, lock_limit;
1303 /* mlock MCL_FUTURE? */
1304 if (flags & VM_LOCKED) {
1305 locked = len >> PAGE_SHIFT;
1306 locked += mm->locked_vm;
1307 lock_limit = rlimit(RLIMIT_MEMLOCK);
1308 lock_limit >>= PAGE_SHIFT;
1309 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1315 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1317 if (S_ISREG(inode->i_mode))
1318 return MAX_LFS_FILESIZE;
1320 if (S_ISBLK(inode->i_mode))
1321 return MAX_LFS_FILESIZE;
1323 /* Special "we do even unsigned file positions" case */
1324 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1327 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1331 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1332 unsigned long pgoff, unsigned long len)
1334 u64 maxsize = file_mmap_size_max(file, inode);
1336 if (maxsize && len > maxsize)
1339 if (pgoff > maxsize >> PAGE_SHIFT)
1345 * The caller must hold down_write(¤t->mm->mmap_sem).
1347 unsigned long do_mmap(struct file *file, unsigned long addr,
1348 unsigned long len, unsigned long prot,
1349 unsigned long flags, vm_flags_t vm_flags,
1350 unsigned long pgoff, unsigned long *populate)
1352 struct mm_struct *mm = current->mm;
1361 * Does the application expect PROT_READ to imply PROT_EXEC?
1363 * (the exception is when the underlying filesystem is noexec
1364 * mounted, in which case we dont add PROT_EXEC.)
1366 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1367 if (!(file && path_noexec(&file->f_path)))
1370 if (!(flags & MAP_FIXED))
1371 addr = round_hint_to_min(addr);
1373 /* Careful about overflows.. */
1374 len = PAGE_ALIGN(len);
1378 /* offset overflow? */
1379 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1382 /* Too many mappings? */
1383 if (mm->map_count > sysctl_max_map_count)
1386 /* Obtain the address to map to. we verify (or select) it and ensure
1387 * that it represents a valid section of the address space.
1389 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1390 if (offset_in_page(addr))
1393 if (prot == PROT_EXEC) {
1394 pkey = execute_only_pkey(mm);
1399 /* Do simple checking here so the lower-level routines won't have
1400 * to. we assume access permissions have been handled by the open
1401 * of the memory object, so we don't do any here.
1403 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1404 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1406 if (flags & MAP_LOCKED)
1407 if (!can_do_mlock())
1410 if (mlock_future_check(mm, vm_flags, len))
1414 struct inode *inode = file_inode(file);
1416 if (!file_mmap_ok(file, inode, pgoff, len))
1419 switch (flags & MAP_TYPE) {
1421 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1425 * Make sure we don't allow writing to an append-only
1428 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1432 * Make sure there are no mandatory locks on the file.
1434 if (locks_verify_locked(file))
1437 vm_flags |= VM_SHARED | VM_MAYSHARE;
1438 if (!(file->f_mode & FMODE_WRITE))
1439 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1443 if (!(file->f_mode & FMODE_READ))
1445 if (path_noexec(&file->f_path)) {
1446 if (vm_flags & VM_EXEC)
1448 vm_flags &= ~VM_MAYEXEC;
1451 if (!file->f_op->mmap)
1453 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1461 switch (flags & MAP_TYPE) {
1463 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1469 vm_flags |= VM_SHARED | VM_MAYSHARE;
1473 * Set pgoff according to addr for anon_vma.
1475 pgoff = addr >> PAGE_SHIFT;
1483 * Set 'VM_NORESERVE' if we should not account for the
1484 * memory use of this mapping.
1486 if (flags & MAP_NORESERVE) {
1487 /* We honor MAP_NORESERVE if allowed to overcommit */
1488 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1489 vm_flags |= VM_NORESERVE;
1491 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1492 if (file && is_file_hugepages(file))
1493 vm_flags |= VM_NORESERVE;
1496 addr = mmap_region(file, addr, len, vm_flags, pgoff);
1497 if (!IS_ERR_VALUE(addr) &&
1498 ((vm_flags & VM_LOCKED) ||
1499 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1504 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1505 unsigned long, prot, unsigned long, flags,
1506 unsigned long, fd, unsigned long, pgoff)
1508 struct file *file = NULL;
1509 unsigned long retval;
1511 if (!(flags & MAP_ANONYMOUS)) {
1512 audit_mmap_fd(fd, flags);
1516 if (is_file_hugepages(file))
1517 len = ALIGN(len, huge_page_size(hstate_file(file)));
1519 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1521 } else if (flags & MAP_HUGETLB) {
1522 struct user_struct *user = NULL;
1525 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1529 len = ALIGN(len, huge_page_size(hs));
1531 * VM_NORESERVE is used because the reservations will be
1532 * taken when vm_ops->mmap() is called
1533 * A dummy user value is used because we are not locking
1534 * memory so no accounting is necessary
1536 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1538 &user, HUGETLB_ANONHUGE_INODE,
1539 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1541 return PTR_ERR(file);
1544 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1546 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1553 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1554 struct mmap_arg_struct {
1558 unsigned long flags;
1560 unsigned long offset;
1563 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1565 struct mmap_arg_struct a;
1567 if (copy_from_user(&a, arg, sizeof(a)))
1569 if (offset_in_page(a.offset))
1572 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1573 a.offset >> PAGE_SHIFT);
1575 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1578 * Some shared mappigns will want the pages marked read-only
1579 * to track write events. If so, we'll downgrade vm_page_prot
1580 * to the private version (using protection_map[] without the
1583 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1585 vm_flags_t vm_flags = vma->vm_flags;
1586 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1588 /* If it was private or non-writable, the write bit is already clear */
1589 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1592 /* The backer wishes to know when pages are first written to? */
1593 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1596 /* The open routine did something to the protections that pgprot_modify
1597 * won't preserve? */
1598 if (pgprot_val(vm_page_prot) !=
1599 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1602 /* Do we need to track softdirty? */
1603 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1606 /* Specialty mapping? */
1607 if (vm_flags & VM_PFNMAP)
1610 /* Can the mapping track the dirty pages? */
1611 return vma->vm_file && vma->vm_file->f_mapping &&
1612 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1616 * We account for memory if it's a private writeable mapping,
1617 * not hugepages and VM_NORESERVE wasn't set.
1619 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1622 * hugetlb has its own accounting separate from the core VM
1623 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1625 if (file && is_file_hugepages(file))
1628 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1631 unsigned long mmap_region(struct file *file, unsigned long addr,
1632 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1634 struct mm_struct *mm = current->mm;
1635 struct vm_area_struct *vma, *prev;
1637 struct rb_node **rb_link, *rb_parent;
1638 unsigned long charged = 0;
1640 /* Check against address space limit. */
1641 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1642 unsigned long nr_pages;
1645 * MAP_FIXED may remove pages of mappings that intersects with
1646 * requested mapping. Account for the pages it would unmap.
1648 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1650 if (!may_expand_vm(mm, vm_flags,
1651 (len >> PAGE_SHIFT) - nr_pages))
1655 /* Clear old maps */
1656 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1658 if (do_munmap(mm, addr, len))
1663 * Private writable mapping: check memory availability
1665 if (accountable_mapping(file, vm_flags)) {
1666 charged = len >> PAGE_SHIFT;
1667 if (security_vm_enough_memory_mm(mm, charged))
1669 vm_flags |= VM_ACCOUNT;
1673 * Can we just expand an old mapping?
1675 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1676 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1681 * Determine the object being mapped and call the appropriate
1682 * specific mapper. the address has already been validated, but
1683 * not unmapped, but the maps are removed from the list.
1685 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1692 vma->vm_start = addr;
1693 vma->vm_end = addr + len;
1694 vma->vm_flags = vm_flags;
1695 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1696 vma->vm_pgoff = pgoff;
1697 INIT_LIST_HEAD(&vma->anon_vma_chain);
1700 if (vm_flags & VM_DENYWRITE) {
1701 error = deny_write_access(file);
1705 if (vm_flags & VM_SHARED) {
1706 error = mapping_map_writable(file->f_mapping);
1708 goto allow_write_and_free_vma;
1711 /* ->mmap() can change vma->vm_file, but must guarantee that
1712 * vma_link() below can deny write-access if VM_DENYWRITE is set
1713 * and map writably if VM_SHARED is set. This usually means the
1714 * new file must not have been exposed to user-space, yet.
1716 vma->vm_file = get_file(file);
1717 error = file->f_op->mmap(file, vma);
1719 goto unmap_and_free_vma;
1721 /* Can addr have changed??
1723 * Answer: Yes, several device drivers can do it in their
1724 * f_op->mmap method. -DaveM
1725 * Bug: If addr is changed, prev, rb_link, rb_parent should
1726 * be updated for vma_link()
1728 WARN_ON_ONCE(addr != vma->vm_start);
1730 addr = vma->vm_start;
1731 vm_flags = vma->vm_flags;
1732 } else if (vm_flags & VM_SHARED) {
1733 error = shmem_zero_setup(vma);
1738 vma_link(mm, vma, prev, rb_link, rb_parent);
1739 /* Once vma denies write, undo our temporary denial count */
1741 if (vm_flags & VM_SHARED)
1742 mapping_unmap_writable(file->f_mapping);
1743 if (vm_flags & VM_DENYWRITE)
1744 allow_write_access(file);
1746 file = vma->vm_file;
1748 perf_event_mmap(vma);
1750 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1751 if (vm_flags & VM_LOCKED) {
1752 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1753 vma == get_gate_vma(current->mm)))
1754 mm->locked_vm += (len >> PAGE_SHIFT);
1756 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1763 * New (or expanded) vma always get soft dirty status.
1764 * Otherwise user-space soft-dirty page tracker won't
1765 * be able to distinguish situation when vma area unmapped,
1766 * then new mapped in-place (which must be aimed as
1767 * a completely new data area).
1769 vma->vm_flags |= VM_SOFTDIRTY;
1771 vma_set_page_prot(vma);
1776 vma->vm_file = NULL;
1779 /* Undo any partial mapping done by a device driver. */
1780 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1782 if (vm_flags & VM_SHARED)
1783 mapping_unmap_writable(file->f_mapping);
1784 allow_write_and_free_vma:
1785 if (vm_flags & VM_DENYWRITE)
1786 allow_write_access(file);
1788 kmem_cache_free(vm_area_cachep, vma);
1791 vm_unacct_memory(charged);
1795 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1798 * We implement the search by looking for an rbtree node that
1799 * immediately follows a suitable gap. That is,
1800 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1801 * - gap_end = vma->vm_start >= info->low_limit + length;
1802 * - gap_end - gap_start >= length
1805 struct mm_struct *mm = current->mm;
1806 struct vm_area_struct *vma;
1807 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1809 /* Adjust search length to account for worst case alignment overhead */
1810 length = info->length + info->align_mask;
1811 if (length < info->length)
1814 /* Adjust search limits by the desired length */
1815 if (info->high_limit < length)
1817 high_limit = info->high_limit - length;
1819 if (info->low_limit > high_limit)
1821 low_limit = info->low_limit + length;
1823 /* Check if rbtree root looks promising */
1824 if (RB_EMPTY_ROOT(&mm->mm_rb))
1826 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1827 if (vma->rb_subtree_gap < length)
1831 /* Visit left subtree if it looks promising */
1832 gap_end = vm_start_gap(vma);
1833 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1834 struct vm_area_struct *left =
1835 rb_entry(vma->vm_rb.rb_left,
1836 struct vm_area_struct, vm_rb);
1837 if (left->rb_subtree_gap >= length) {
1843 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1845 /* Check if current node has a suitable gap */
1846 if (gap_start > high_limit)
1848 if (gap_end >= low_limit &&
1849 gap_end > gap_start && gap_end - gap_start >= length)
1852 /* Visit right subtree if it looks promising */
1853 if (vma->vm_rb.rb_right) {
1854 struct vm_area_struct *right =
1855 rb_entry(vma->vm_rb.rb_right,
1856 struct vm_area_struct, vm_rb);
1857 if (right->rb_subtree_gap >= length) {
1863 /* Go back up the rbtree to find next candidate node */
1865 struct rb_node *prev = &vma->vm_rb;
1866 if (!rb_parent(prev))
1868 vma = rb_entry(rb_parent(prev),
1869 struct vm_area_struct, vm_rb);
1870 if (prev == vma->vm_rb.rb_left) {
1871 gap_start = vm_end_gap(vma->vm_prev);
1872 gap_end = vm_start_gap(vma);
1879 /* Check highest gap, which does not precede any rbtree node */
1880 gap_start = mm->highest_vm_end;
1881 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1882 if (gap_start > high_limit)
1886 /* We found a suitable gap. Clip it with the original low_limit. */
1887 if (gap_start < info->low_limit)
1888 gap_start = info->low_limit;
1890 /* Adjust gap address to the desired alignment */
1891 gap_start += (info->align_offset - gap_start) & info->align_mask;
1893 VM_BUG_ON(gap_start + info->length > info->high_limit);
1894 VM_BUG_ON(gap_start + info->length > gap_end);
1898 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1900 struct mm_struct *mm = current->mm;
1901 struct vm_area_struct *vma;
1902 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1904 /* Adjust search length to account for worst case alignment overhead */
1905 length = info->length + info->align_mask;
1906 if (length < info->length)
1910 * Adjust search limits by the desired length.
1911 * See implementation comment at top of unmapped_area().
1913 gap_end = info->high_limit;
1914 if (gap_end < length)
1916 high_limit = gap_end - length;
1918 if (info->low_limit > high_limit)
1920 low_limit = info->low_limit + length;
1922 /* Check highest gap, which does not precede any rbtree node */
1923 gap_start = mm->highest_vm_end;
1924 if (gap_start <= high_limit)
1927 /* Check if rbtree root looks promising */
1928 if (RB_EMPTY_ROOT(&mm->mm_rb))
1930 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1931 if (vma->rb_subtree_gap < length)
1935 /* Visit right subtree if it looks promising */
1936 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1937 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1938 struct vm_area_struct *right =
1939 rb_entry(vma->vm_rb.rb_right,
1940 struct vm_area_struct, vm_rb);
1941 if (right->rb_subtree_gap >= length) {
1948 /* Check if current node has a suitable gap */
1949 gap_end = vm_start_gap(vma);
1950 if (gap_end < low_limit)
1952 if (gap_start <= high_limit &&
1953 gap_end > gap_start && gap_end - gap_start >= length)
1956 /* Visit left subtree if it looks promising */
1957 if (vma->vm_rb.rb_left) {
1958 struct vm_area_struct *left =
1959 rb_entry(vma->vm_rb.rb_left,
1960 struct vm_area_struct, vm_rb);
1961 if (left->rb_subtree_gap >= length) {
1967 /* Go back up the rbtree to find next candidate node */
1969 struct rb_node *prev = &vma->vm_rb;
1970 if (!rb_parent(prev))
1972 vma = rb_entry(rb_parent(prev),
1973 struct vm_area_struct, vm_rb);
1974 if (prev == vma->vm_rb.rb_right) {
1975 gap_start = vma->vm_prev ?
1976 vm_end_gap(vma->vm_prev) : 0;
1983 /* We found a suitable gap. Clip it with the original high_limit. */
1984 if (gap_end > info->high_limit)
1985 gap_end = info->high_limit;
1988 /* Compute highest gap address at the desired alignment */
1989 gap_end -= info->length;
1990 gap_end -= (gap_end - info->align_offset) & info->align_mask;
1992 VM_BUG_ON(gap_end < info->low_limit);
1993 VM_BUG_ON(gap_end < gap_start);
1997 /* Get an address range which is currently unmapped.
1998 * For shmat() with addr=0.
2000 * Ugly calling convention alert:
2001 * Return value with the low bits set means error value,
2003 * if (ret & ~PAGE_MASK)
2006 * This function "knows" that -ENOMEM has the bits set.
2008 #ifndef HAVE_ARCH_UNMAPPED_AREA
2010 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2011 unsigned long len, unsigned long pgoff, unsigned long flags)
2013 struct mm_struct *mm = current->mm;
2014 struct vm_area_struct *vma, *prev;
2015 struct vm_unmapped_area_info info;
2017 if (len > TASK_SIZE - mmap_min_addr)
2020 if (flags & MAP_FIXED)
2024 addr = PAGE_ALIGN(addr);
2025 vma = find_vma_prev(mm, addr, &prev);
2026 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2027 (!vma || addr + len <= vm_start_gap(vma)) &&
2028 (!prev || addr >= vm_end_gap(prev)))
2034 info.low_limit = mm->mmap_base;
2035 info.high_limit = TASK_SIZE;
2036 info.align_mask = 0;
2037 return vm_unmapped_area(&info);
2042 * This mmap-allocator allocates new areas top-down from below the
2043 * stack's low limit (the base):
2045 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2047 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2048 const unsigned long len, const unsigned long pgoff,
2049 const unsigned long flags)
2051 struct vm_area_struct *vma, *prev;
2052 struct mm_struct *mm = current->mm;
2053 unsigned long addr = addr0;
2054 struct vm_unmapped_area_info info;
2056 /* requested length too big for entire address space */
2057 if (len > TASK_SIZE - mmap_min_addr)
2060 if (flags & MAP_FIXED)
2063 /* requesting a specific address */
2065 addr = PAGE_ALIGN(addr);
2066 vma = find_vma_prev(mm, addr, &prev);
2067 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2068 (!vma || addr + len <= vm_start_gap(vma)) &&
2069 (!prev || addr >= vm_end_gap(prev)))
2073 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2075 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2076 info.high_limit = mm->mmap_base;
2077 info.align_mask = 0;
2078 addr = vm_unmapped_area(&info);
2081 * A failed mmap() very likely causes application failure,
2082 * so fall back to the bottom-up function here. This scenario
2083 * can happen with large stack limits and large mmap()
2086 if (offset_in_page(addr)) {
2087 VM_BUG_ON(addr != -ENOMEM);
2089 info.low_limit = TASK_UNMAPPED_BASE;
2090 info.high_limit = TASK_SIZE;
2091 addr = vm_unmapped_area(&info);
2099 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2100 unsigned long pgoff, unsigned long flags)
2102 unsigned long (*get_area)(struct file *, unsigned long,
2103 unsigned long, unsigned long, unsigned long);
2105 unsigned long error = arch_mmap_check(addr, len, flags);
2109 /* Careful about overflows.. */
2110 if (len > TASK_SIZE)
2113 get_area = current->mm->get_unmapped_area;
2115 if (file->f_op->get_unmapped_area)
2116 get_area = file->f_op->get_unmapped_area;
2117 } else if (flags & MAP_SHARED) {
2119 * mmap_region() will call shmem_zero_setup() to create a file,
2120 * so use shmem's get_unmapped_area in case it can be huge.
2121 * do_mmap_pgoff() will clear pgoff, so match alignment.
2124 get_area = shmem_get_unmapped_area;
2127 addr = get_area(file, addr, len, pgoff, flags);
2128 if (IS_ERR_VALUE(addr))
2131 if (addr > TASK_SIZE - len)
2133 if (offset_in_page(addr))
2136 error = security_mmap_addr(addr);
2137 return error ? error : addr;
2140 EXPORT_SYMBOL(get_unmapped_area);
2142 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2143 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2145 struct rb_node *rb_node;
2146 struct vm_area_struct *vma;
2148 /* Check the cache first. */
2149 vma = vmacache_find(mm, addr);
2153 rb_node = mm->mm_rb.rb_node;
2156 struct vm_area_struct *tmp;
2158 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2160 if (tmp->vm_end > addr) {
2162 if (tmp->vm_start <= addr)
2164 rb_node = rb_node->rb_left;
2166 rb_node = rb_node->rb_right;
2170 vmacache_update(addr, vma);
2174 EXPORT_SYMBOL(find_vma);
2177 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2179 struct vm_area_struct *
2180 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2181 struct vm_area_struct **pprev)
2183 struct vm_area_struct *vma;
2185 vma = find_vma(mm, addr);
2187 *pprev = vma->vm_prev;
2189 struct rb_node *rb_node = mm->mm_rb.rb_node;
2192 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2193 rb_node = rb_node->rb_right;
2200 * Verify that the stack growth is acceptable and
2201 * update accounting. This is shared with both the
2202 * grow-up and grow-down cases.
2204 static int acct_stack_growth(struct vm_area_struct *vma,
2205 unsigned long size, unsigned long grow)
2207 struct mm_struct *mm = vma->vm_mm;
2208 struct rlimit *rlim = current->signal->rlim;
2209 unsigned long new_start;
2211 /* address space limit tests */
2212 if (!may_expand_vm(mm, vma->vm_flags, grow))
2215 /* Stack limit test */
2216 if (size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2219 /* mlock limit tests */
2220 if (vma->vm_flags & VM_LOCKED) {
2221 unsigned long locked;
2222 unsigned long limit;
2223 locked = mm->locked_vm + grow;
2224 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2225 limit >>= PAGE_SHIFT;
2226 if (locked > limit && !capable(CAP_IPC_LOCK))
2230 /* Check to ensure the stack will not grow into a hugetlb-only region */
2231 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2233 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2237 * Overcommit.. This must be the final test, as it will
2238 * update security statistics.
2240 if (security_vm_enough_memory_mm(mm, grow))
2246 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2248 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2249 * vma is the last one with address > vma->vm_end. Have to extend vma.
2251 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2253 struct mm_struct *mm = vma->vm_mm;
2254 struct vm_area_struct *next;
2255 unsigned long gap_addr;
2258 if (!(vma->vm_flags & VM_GROWSUP))
2261 /* Guard against exceeding limits of the address space. */
2262 address &= PAGE_MASK;
2263 if (address >= (TASK_SIZE & PAGE_MASK))
2265 address += PAGE_SIZE;
2267 /* Enforce stack_guard_gap */
2268 gap_addr = address + stack_guard_gap;
2270 /* Guard against overflow */
2271 if (gap_addr < address || gap_addr > TASK_SIZE)
2272 gap_addr = TASK_SIZE;
2274 next = vma->vm_next;
2275 if (next && next->vm_start < gap_addr &&
2276 (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2277 if (!(next->vm_flags & VM_GROWSUP))
2279 /* Check that both stack segments have the same anon_vma? */
2282 /* We must make sure the anon_vma is allocated. */
2283 if (unlikely(anon_vma_prepare(vma)))
2287 * vma->vm_start/vm_end cannot change under us because the caller
2288 * is required to hold the mmap_sem in read mode. We need the
2289 * anon_vma lock to serialize against concurrent expand_stacks.
2291 anon_vma_lock_write(vma->anon_vma);
2293 /* Somebody else might have raced and expanded it already */
2294 if (address > vma->vm_end) {
2295 unsigned long size, grow;
2297 size = address - vma->vm_start;
2298 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2301 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2302 error = acct_stack_growth(vma, size, grow);
2305 * vma_gap_update() doesn't support concurrent
2306 * updates, but we only hold a shared mmap_sem
2307 * lock here, so we need to protect against
2308 * concurrent vma expansions.
2309 * anon_vma_lock_write() doesn't help here, as
2310 * we don't guarantee that all growable vmas
2311 * in a mm share the same root anon vma.
2312 * So, we reuse mm->page_table_lock to guard
2313 * against concurrent vma expansions.
2315 spin_lock(&mm->page_table_lock);
2316 if (vma->vm_flags & VM_LOCKED)
2317 mm->locked_vm += grow;
2318 vm_stat_account(mm, vma->vm_flags, grow);
2319 anon_vma_interval_tree_pre_update_vma(vma);
2320 vma->vm_end = address;
2321 anon_vma_interval_tree_post_update_vma(vma);
2323 vma_gap_update(vma->vm_next);
2325 mm->highest_vm_end = vm_end_gap(vma);
2326 spin_unlock(&mm->page_table_lock);
2328 perf_event_mmap(vma);
2332 anon_vma_unlock_write(vma->anon_vma);
2333 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2337 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2340 * vma is the first one with address < vma->vm_start. Have to extend vma.
2342 int expand_downwards(struct vm_area_struct *vma,
2343 unsigned long address)
2345 struct mm_struct *mm = vma->vm_mm;
2346 struct vm_area_struct *prev;
2347 unsigned long gap_addr;
2350 address &= PAGE_MASK;
2351 if (address < mmap_min_addr)
2354 /* Enforce stack_guard_gap */
2355 gap_addr = address - stack_guard_gap;
2356 if (gap_addr > address)
2358 prev = vma->vm_prev;
2359 if (prev && prev->vm_end > gap_addr &&
2360 (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2361 if (!(prev->vm_flags & VM_GROWSDOWN))
2363 /* Check that both stack segments have the same anon_vma? */
2366 /* We must make sure the anon_vma is allocated. */
2367 if (unlikely(anon_vma_prepare(vma)))
2371 * vma->vm_start/vm_end cannot change under us because the caller
2372 * is required to hold the mmap_sem in read mode. We need the
2373 * anon_vma lock to serialize against concurrent expand_stacks.
2375 anon_vma_lock_write(vma->anon_vma);
2377 /* Somebody else might have raced and expanded it already */
2378 if (address < vma->vm_start) {
2379 unsigned long size, grow;
2381 size = vma->vm_end - address;
2382 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2385 if (grow <= vma->vm_pgoff) {
2386 error = acct_stack_growth(vma, size, grow);
2389 * vma_gap_update() doesn't support concurrent
2390 * updates, but we only hold a shared mmap_sem
2391 * lock here, so we need to protect against
2392 * concurrent vma expansions.
2393 * anon_vma_lock_write() doesn't help here, as
2394 * we don't guarantee that all growable vmas
2395 * in a mm share the same root anon vma.
2396 * So, we reuse mm->page_table_lock to guard
2397 * against concurrent vma expansions.
2399 spin_lock(&mm->page_table_lock);
2400 if (vma->vm_flags & VM_LOCKED)
2401 mm->locked_vm += grow;
2402 vm_stat_account(mm, vma->vm_flags, grow);
2403 anon_vma_interval_tree_pre_update_vma(vma);
2404 vma->vm_start = address;
2405 vma->vm_pgoff -= grow;
2406 anon_vma_interval_tree_post_update_vma(vma);
2407 vma_gap_update(vma);
2408 spin_unlock(&mm->page_table_lock);
2410 perf_event_mmap(vma);
2414 anon_vma_unlock_write(vma->anon_vma);
2415 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2420 /* enforced gap between the expanding stack and other mappings. */
2421 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2423 static int __init cmdline_parse_stack_guard_gap(char *p)
2428 val = simple_strtoul(p, &endptr, 10);
2430 stack_guard_gap = val << PAGE_SHIFT;
2434 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2436 #ifdef CONFIG_STACK_GROWSUP
2437 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2439 return expand_upwards(vma, address);
2442 struct vm_area_struct *
2443 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2445 struct vm_area_struct *vma, *prev;
2448 vma = find_vma_prev(mm, addr, &prev);
2449 if (vma && (vma->vm_start <= addr))
2451 /* don't alter vm_end if the coredump is running */
2452 if (!prev || !mmget_still_valid(mm) || expand_stack(prev, addr))
2454 if (prev->vm_flags & VM_LOCKED)
2455 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2459 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2461 return expand_downwards(vma, address);
2464 struct vm_area_struct *
2465 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2467 struct vm_area_struct *vma;
2468 unsigned long start;
2471 vma = find_vma(mm, addr);
2474 if (vma->vm_start <= addr)
2476 if (!(vma->vm_flags & VM_GROWSDOWN))
2478 /* don't alter vm_start if the coredump is running */
2479 if (!mmget_still_valid(mm))
2481 start = vma->vm_start;
2482 if (expand_stack(vma, addr))
2484 if (vma->vm_flags & VM_LOCKED)
2485 populate_vma_page_range(vma, addr, start, NULL);
2490 EXPORT_SYMBOL_GPL(find_extend_vma);
2493 * Ok - we have the memory areas we should free on the vma list,
2494 * so release them, and do the vma updates.
2496 * Called with the mm semaphore held.
2498 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2500 unsigned long nr_accounted = 0;
2502 /* Update high watermark before we lower total_vm */
2503 update_hiwater_vm(mm);
2505 long nrpages = vma_pages(vma);
2507 if (vma->vm_flags & VM_ACCOUNT)
2508 nr_accounted += nrpages;
2509 vm_stat_account(mm, vma->vm_flags, -nrpages);
2510 vma = remove_vma(vma);
2512 vm_unacct_memory(nr_accounted);
2517 * Get rid of page table information in the indicated region.
2519 * Called with the mm semaphore held.
2521 static void unmap_region(struct mm_struct *mm,
2522 struct vm_area_struct *vma, struct vm_area_struct *prev,
2523 unsigned long start, unsigned long end)
2525 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2526 struct mmu_gather tlb;
2529 tlb_gather_mmu(&tlb, mm, start, end);
2530 update_hiwater_rss(mm);
2531 unmap_vmas(&tlb, vma, start, end);
2532 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2533 next ? next->vm_start : USER_PGTABLES_CEILING);
2534 tlb_finish_mmu(&tlb, start, end);
2538 * Create a list of vma's touched by the unmap, removing them from the mm's
2539 * vma list as we go..
2542 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2543 struct vm_area_struct *prev, unsigned long end)
2545 struct vm_area_struct **insertion_point;
2546 struct vm_area_struct *tail_vma = NULL;
2548 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2549 vma->vm_prev = NULL;
2551 vma_rb_erase(vma, &mm->mm_rb);
2555 } while (vma && vma->vm_start < end);
2556 *insertion_point = vma;
2558 vma->vm_prev = prev;
2559 vma_gap_update(vma);
2561 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2562 tail_vma->vm_next = NULL;
2564 /* Kill the cache */
2565 vmacache_invalidate(mm);
2569 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2570 * munmap path where it doesn't make sense to fail.
2572 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2573 unsigned long addr, int new_below)
2575 struct vm_area_struct *new;
2578 if (vma->vm_ops && vma->vm_ops->split) {
2579 err = vma->vm_ops->split(vma, addr);
2584 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2588 /* most fields are the same, copy all, and then fixup */
2591 INIT_LIST_HEAD(&new->anon_vma_chain);
2596 new->vm_start = addr;
2597 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2600 err = vma_dup_policy(vma, new);
2604 err = anon_vma_clone(new, vma);
2609 get_file(new->vm_file);
2611 if (new->vm_ops && new->vm_ops->open)
2612 new->vm_ops->open(new);
2615 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2616 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2618 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2624 /* Clean everything up if vma_adjust failed. */
2625 if (new->vm_ops && new->vm_ops->close)
2626 new->vm_ops->close(new);
2629 unlink_anon_vmas(new);
2631 mpol_put(vma_policy(new));
2633 kmem_cache_free(vm_area_cachep, new);
2638 * Split a vma into two pieces at address 'addr', a new vma is allocated
2639 * either for the first part or the tail.
2641 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2642 unsigned long addr, int new_below)
2644 if (mm->map_count >= sysctl_max_map_count)
2647 return __split_vma(mm, vma, addr, new_below);
2650 /* Munmap is split into 2 main parts -- this part which finds
2651 * what needs doing, and the areas themselves, which do the
2652 * work. This now handles partial unmappings.
2653 * Jeremy Fitzhardinge <jeremy@goop.org>
2655 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2658 struct vm_area_struct *vma, *prev, *last;
2660 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2663 len = PAGE_ALIGN(len);
2667 /* Find the first overlapping VMA */
2668 vma = find_vma(mm, start);
2671 prev = vma->vm_prev;
2672 /* we have start < vma->vm_end */
2674 /* if it doesn't overlap, we have nothing.. */
2676 if (vma->vm_start >= end)
2680 * If we need to split any vma, do it now to save pain later.
2682 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2683 * unmapped vm_area_struct will remain in use: so lower split_vma
2684 * places tmp vma above, and higher split_vma places tmp vma below.
2686 if (start > vma->vm_start) {
2690 * Make sure that map_count on return from munmap() will
2691 * not exceed its limit; but let map_count go just above
2692 * its limit temporarily, to help free resources as expected.
2694 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2697 error = __split_vma(mm, vma, start, 0);
2703 /* Does it split the last one? */
2704 last = find_vma(mm, end);
2705 if (last && end > last->vm_start) {
2706 int error = __split_vma(mm, last, end, 1);
2710 vma = prev ? prev->vm_next : mm->mmap;
2713 * unlock any mlock()ed ranges before detaching vmas
2715 if (mm->locked_vm) {
2716 struct vm_area_struct *tmp = vma;
2717 while (tmp && tmp->vm_start < end) {
2718 if (tmp->vm_flags & VM_LOCKED) {
2719 mm->locked_vm -= vma_pages(tmp);
2720 munlock_vma_pages_all(tmp);
2727 * Remove the vma's, and unmap the actual pages
2729 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2730 unmap_region(mm, vma, prev, start, end);
2732 arch_unmap(mm, vma, start, end);
2734 /* Fix up all other VM information */
2735 remove_vma_list(mm, vma);
2740 int vm_munmap(unsigned long start, size_t len)
2743 struct mm_struct *mm = current->mm;
2745 if (down_write_killable(&mm->mmap_sem))
2748 ret = do_munmap(mm, start, len);
2749 up_write(&mm->mmap_sem);
2752 EXPORT_SYMBOL(vm_munmap);
2754 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2757 struct mm_struct *mm = current->mm;
2759 profile_munmap(addr);
2760 if (down_write_killable(&mm->mmap_sem))
2762 ret = do_munmap(mm, addr, len);
2763 up_write(&mm->mmap_sem);
2769 * Emulation of deprecated remap_file_pages() syscall.
2771 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2772 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2775 struct mm_struct *mm = current->mm;
2776 struct vm_area_struct *vma;
2777 unsigned long populate = 0;
2778 unsigned long ret = -EINVAL;
2781 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2782 current->comm, current->pid);
2786 start = start & PAGE_MASK;
2787 size = size & PAGE_MASK;
2789 if (start + size <= start)
2792 /* Does pgoff wrap? */
2793 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2796 if (down_write_killable(&mm->mmap_sem))
2799 vma = find_vma(mm, start);
2801 if (!vma || !(vma->vm_flags & VM_SHARED))
2804 if (start < vma->vm_start)
2807 if (start + size > vma->vm_end) {
2808 struct vm_area_struct *next;
2810 for (next = vma->vm_next; next; next = next->vm_next) {
2811 /* hole between vmas ? */
2812 if (next->vm_start != next->vm_prev->vm_end)
2815 if (next->vm_file != vma->vm_file)
2818 if (next->vm_flags != vma->vm_flags)
2821 if (start + size <= next->vm_end)
2829 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2830 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2831 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2833 flags &= MAP_NONBLOCK;
2834 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2835 if (vma->vm_flags & VM_LOCKED) {
2836 struct vm_area_struct *tmp;
2837 flags |= MAP_LOCKED;
2839 /* drop PG_Mlocked flag for over-mapped range */
2840 for (tmp = vma; tmp->vm_start >= start + size;
2841 tmp = tmp->vm_next) {
2843 * Split pmd and munlock page on the border
2846 vma_adjust_trans_huge(tmp, start, start + size, 0);
2848 munlock_vma_pages_range(tmp,
2849 max(tmp->vm_start, start),
2850 min(tmp->vm_end, start + size));
2854 file = get_file(vma->vm_file);
2855 ret = do_mmap_pgoff(vma->vm_file, start, size,
2856 prot, flags, pgoff, &populate);
2859 up_write(&mm->mmap_sem);
2861 mm_populate(ret, populate);
2862 if (!IS_ERR_VALUE(ret))
2867 static inline void verify_mm_writelocked(struct mm_struct *mm)
2869 #ifdef CONFIG_DEBUG_VM
2870 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2872 up_read(&mm->mmap_sem);
2878 * this is really a simplified "do_mmap". it only handles
2879 * anonymous maps. eventually we may be able to do some
2880 * brk-specific accounting here.
2882 static int do_brk(unsigned long addr, unsigned long len)
2884 struct mm_struct *mm = current->mm;
2885 struct vm_area_struct *vma, *prev;
2886 unsigned long flags;
2887 struct rb_node **rb_link, *rb_parent;
2888 pgoff_t pgoff = addr >> PAGE_SHIFT;
2891 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2893 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2894 if (offset_in_page(error))
2897 error = mlock_future_check(mm, mm->def_flags, len);
2902 * mm->mmap_sem is required to protect against another thread
2903 * changing the mappings in case we sleep.
2905 verify_mm_writelocked(mm);
2908 * Clear old maps. this also does some error checking for us
2910 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2912 if (do_munmap(mm, addr, len))
2916 /* Check against address space limits *after* clearing old maps... */
2917 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2920 if (mm->map_count > sysctl_max_map_count)
2923 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2926 /* Can we just expand an old private anonymous mapping? */
2927 vma = vma_merge(mm, prev, addr, addr + len, flags,
2928 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2933 * create a vma struct for an anonymous mapping
2935 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2937 vm_unacct_memory(len >> PAGE_SHIFT);
2941 INIT_LIST_HEAD(&vma->anon_vma_chain);
2943 vma->vm_start = addr;
2944 vma->vm_end = addr + len;
2945 vma->vm_pgoff = pgoff;
2946 vma->vm_flags = flags;
2947 vma->vm_page_prot = vm_get_page_prot(flags);
2948 vma_link(mm, vma, prev, rb_link, rb_parent);
2950 perf_event_mmap(vma);
2951 mm->total_vm += len >> PAGE_SHIFT;
2952 mm->data_vm += len >> PAGE_SHIFT;
2953 if (flags & VM_LOCKED)
2954 mm->locked_vm += (len >> PAGE_SHIFT);
2955 vma->vm_flags |= VM_SOFTDIRTY;
2959 int vm_brk(unsigned long addr, unsigned long request)
2961 struct mm_struct *mm = current->mm;
2966 len = PAGE_ALIGN(request);
2972 if (down_write_killable(&mm->mmap_sem))
2975 ret = do_brk(addr, len);
2976 populate = ((mm->def_flags & VM_LOCKED) != 0);
2977 up_write(&mm->mmap_sem);
2978 if (populate && !ret)
2979 mm_populate(addr, len);
2982 EXPORT_SYMBOL(vm_brk);
2984 /* Release all mmaps. */
2985 void exit_mmap(struct mm_struct *mm)
2987 struct mmu_gather tlb;
2988 struct vm_area_struct *vma;
2989 unsigned long nr_accounted = 0;
2991 /* mm's last user has gone, and its about to be pulled down */
2992 mmu_notifier_release(mm);
2994 if (mm->locked_vm) {
2997 if (vma->vm_flags & VM_LOCKED)
2998 munlock_vma_pages_all(vma);
3006 if (!vma) /* Can happen if dup_mmap() received an OOM */
3011 tlb_gather_mmu(&tlb, mm, 0, -1);
3012 /* update_hiwater_rss(mm) here? but nobody should be looking */
3013 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3014 unmap_vmas(&tlb, vma, 0, -1);
3016 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3017 tlb_finish_mmu(&tlb, 0, -1);
3020 * Walk the list again, actually closing and freeing it,
3021 * with preemption enabled, without holding any MM locks.
3024 if (vma->vm_flags & VM_ACCOUNT)
3025 nr_accounted += vma_pages(vma);
3026 vma = remove_vma(vma);
3028 vm_unacct_memory(nr_accounted);
3031 /* Insert vm structure into process list sorted by address
3032 * and into the inode's i_mmap tree. If vm_file is non-NULL
3033 * then i_mmap_rwsem is taken here.
3035 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3037 struct vm_area_struct *prev;
3038 struct rb_node **rb_link, *rb_parent;
3040 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3041 &prev, &rb_link, &rb_parent))
3043 if ((vma->vm_flags & VM_ACCOUNT) &&
3044 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3048 * The vm_pgoff of a purely anonymous vma should be irrelevant
3049 * until its first write fault, when page's anon_vma and index
3050 * are set. But now set the vm_pgoff it will almost certainly
3051 * end up with (unless mremap moves it elsewhere before that
3052 * first wfault), so /proc/pid/maps tells a consistent story.
3054 * By setting it to reflect the virtual start address of the
3055 * vma, merges and splits can happen in a seamless way, just
3056 * using the existing file pgoff checks and manipulations.
3057 * Similarly in do_mmap_pgoff and in do_brk.
3059 if (vma_is_anonymous(vma)) {
3060 BUG_ON(vma->anon_vma);
3061 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3064 vma_link(mm, vma, prev, rb_link, rb_parent);
3069 * Copy the vma structure to a new location in the same mm,
3070 * prior to moving page table entries, to effect an mremap move.
3072 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3073 unsigned long addr, unsigned long len, pgoff_t pgoff,
3074 bool *need_rmap_locks)
3076 struct vm_area_struct *vma = *vmap;
3077 unsigned long vma_start = vma->vm_start;
3078 struct mm_struct *mm = vma->vm_mm;
3079 struct vm_area_struct *new_vma, *prev;
3080 struct rb_node **rb_link, *rb_parent;
3081 bool faulted_in_anon_vma = true;
3084 * If anonymous vma has not yet been faulted, update new pgoff
3085 * to match new location, to increase its chance of merging.
3087 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3088 pgoff = addr >> PAGE_SHIFT;
3089 faulted_in_anon_vma = false;
3092 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3093 return NULL; /* should never get here */
3094 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3095 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3096 vma->vm_userfaultfd_ctx);
3099 * Source vma may have been merged into new_vma
3101 if (unlikely(vma_start >= new_vma->vm_start &&
3102 vma_start < new_vma->vm_end)) {
3104 * The only way we can get a vma_merge with
3105 * self during an mremap is if the vma hasn't
3106 * been faulted in yet and we were allowed to
3107 * reset the dst vma->vm_pgoff to the
3108 * destination address of the mremap to allow
3109 * the merge to happen. mremap must change the
3110 * vm_pgoff linearity between src and dst vmas
3111 * (in turn preventing a vma_merge) to be
3112 * safe. It is only safe to keep the vm_pgoff
3113 * linear if there are no pages mapped yet.
3115 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3116 *vmap = vma = new_vma;
3118 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3120 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3124 new_vma->vm_start = addr;
3125 new_vma->vm_end = addr + len;
3126 new_vma->vm_pgoff = pgoff;
3127 if (vma_dup_policy(vma, new_vma))
3129 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3130 if (anon_vma_clone(new_vma, vma))
3131 goto out_free_mempol;
3132 if (new_vma->vm_file)
3133 get_file(new_vma->vm_file);
3134 if (new_vma->vm_ops && new_vma->vm_ops->open)
3135 new_vma->vm_ops->open(new_vma);
3136 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3137 *need_rmap_locks = false;
3142 mpol_put(vma_policy(new_vma));
3144 kmem_cache_free(vm_area_cachep, new_vma);
3150 * Return true if the calling process may expand its vm space by the passed
3153 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3155 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3158 if (is_data_mapping(flags) &&
3159 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3160 /* Workaround for Valgrind */
3161 if (rlimit(RLIMIT_DATA) == 0 &&
3162 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3164 if (!ignore_rlimit_data) {
3165 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits or use boot option ignore_rlimit_data.\n",
3166 current->comm, current->pid,
3167 (mm->data_vm + npages) << PAGE_SHIFT,
3168 rlimit(RLIMIT_DATA));
3176 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3178 mm->total_vm += npages;
3180 if (is_exec_mapping(flags))
3181 mm->exec_vm += npages;
3182 else if (is_stack_mapping(flags))
3183 mm->stack_vm += npages;
3184 else if (is_data_mapping(flags))
3185 mm->data_vm += npages;
3188 static int special_mapping_fault(struct vm_area_struct *vma,
3189 struct vm_fault *vmf);
3192 * Having a close hook prevents vma merging regardless of flags.
3194 static void special_mapping_close(struct vm_area_struct *vma)
3198 static const char *special_mapping_name(struct vm_area_struct *vma)
3200 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3203 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3205 struct vm_special_mapping *sm = new_vma->vm_private_data;
3208 return sm->mremap(sm, new_vma);
3212 static const struct vm_operations_struct special_mapping_vmops = {
3213 .close = special_mapping_close,
3214 .fault = special_mapping_fault,
3215 .mremap = special_mapping_mremap,
3216 .name = special_mapping_name,
3219 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3220 .close = special_mapping_close,
3221 .fault = special_mapping_fault,
3224 static int special_mapping_fault(struct vm_area_struct *vma,
3225 struct vm_fault *vmf)
3228 struct page **pages;
3230 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3231 pages = vma->vm_private_data;
3233 struct vm_special_mapping *sm = vma->vm_private_data;
3236 return sm->fault(sm, vma, vmf);
3241 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3245 struct page *page = *pages;
3251 return VM_FAULT_SIGBUS;
3254 static struct vm_area_struct *__install_special_mapping(
3255 struct mm_struct *mm,
3256 unsigned long addr, unsigned long len,
3257 unsigned long vm_flags, void *priv,
3258 const struct vm_operations_struct *ops)
3261 struct vm_area_struct *vma;
3263 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3264 if (unlikely(vma == NULL))
3265 return ERR_PTR(-ENOMEM);
3267 INIT_LIST_HEAD(&vma->anon_vma_chain);
3269 vma->vm_start = addr;
3270 vma->vm_end = addr + len;
3272 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3273 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3276 vma->vm_private_data = priv;
3278 ret = insert_vm_struct(mm, vma);
3282 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3284 perf_event_mmap(vma);
3289 kmem_cache_free(vm_area_cachep, vma);
3290 return ERR_PTR(ret);
3293 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3294 const struct vm_special_mapping *sm)
3296 return vma->vm_private_data == sm &&
3297 (vma->vm_ops == &special_mapping_vmops ||
3298 vma->vm_ops == &legacy_special_mapping_vmops);
3302 * Called with mm->mmap_sem held for writing.
3303 * Insert a new vma covering the given region, with the given flags.
3304 * Its pages are supplied by the given array of struct page *.
3305 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3306 * The region past the last page supplied will always produce SIGBUS.
3307 * The array pointer and the pages it points to are assumed to stay alive
3308 * for as long as this mapping might exist.
3310 struct vm_area_struct *_install_special_mapping(
3311 struct mm_struct *mm,
3312 unsigned long addr, unsigned long len,
3313 unsigned long vm_flags, const struct vm_special_mapping *spec)
3315 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3316 &special_mapping_vmops);
3319 int install_special_mapping(struct mm_struct *mm,
3320 unsigned long addr, unsigned long len,
3321 unsigned long vm_flags, struct page **pages)
3323 struct vm_area_struct *vma = __install_special_mapping(
3324 mm, addr, len, vm_flags, (void *)pages,
3325 &legacy_special_mapping_vmops);
3327 return PTR_ERR_OR_ZERO(vma);
3330 static DEFINE_MUTEX(mm_all_locks_mutex);
3332 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3334 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3336 * The LSB of head.next can't change from under us
3337 * because we hold the mm_all_locks_mutex.
3339 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3341 * We can safely modify head.next after taking the
3342 * anon_vma->root->rwsem. If some other vma in this mm shares
3343 * the same anon_vma we won't take it again.
3345 * No need of atomic instructions here, head.next
3346 * can't change from under us thanks to the
3347 * anon_vma->root->rwsem.
3349 if (__test_and_set_bit(0, (unsigned long *)
3350 &anon_vma->root->rb_root.rb_node))
3355 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3357 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3359 * AS_MM_ALL_LOCKS can't change from under us because
3360 * we hold the mm_all_locks_mutex.
3362 * Operations on ->flags have to be atomic because
3363 * even if AS_MM_ALL_LOCKS is stable thanks to the
3364 * mm_all_locks_mutex, there may be other cpus
3365 * changing other bitflags in parallel to us.
3367 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3369 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3374 * This operation locks against the VM for all pte/vma/mm related
3375 * operations that could ever happen on a certain mm. This includes
3376 * vmtruncate, try_to_unmap, and all page faults.
3378 * The caller must take the mmap_sem in write mode before calling
3379 * mm_take_all_locks(). The caller isn't allowed to release the
3380 * mmap_sem until mm_drop_all_locks() returns.
3382 * mmap_sem in write mode is required in order to block all operations
3383 * that could modify pagetables and free pages without need of
3384 * altering the vma layout. It's also needed in write mode to avoid new
3385 * anon_vmas to be associated with existing vmas.
3387 * A single task can't take more than one mm_take_all_locks() in a row
3388 * or it would deadlock.
3390 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3391 * mapping->flags avoid to take the same lock twice, if more than one
3392 * vma in this mm is backed by the same anon_vma or address_space.
3394 * We take locks in following order, accordingly to comment at beginning
3396 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3398 * - all i_mmap_rwsem locks;
3399 * - all anon_vma->rwseml
3401 * We can take all locks within these types randomly because the VM code
3402 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3403 * mm_all_locks_mutex.
3405 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3406 * that may have to take thousand of locks.
3408 * mm_take_all_locks() can fail if it's interrupted by signals.
3410 int mm_take_all_locks(struct mm_struct *mm)
3412 struct vm_area_struct *vma;
3413 struct anon_vma_chain *avc;
3415 BUG_ON(down_read_trylock(&mm->mmap_sem));
3417 mutex_lock(&mm_all_locks_mutex);
3419 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3420 if (signal_pending(current))
3422 if (vma->vm_file && vma->vm_file->f_mapping &&
3423 is_vm_hugetlb_page(vma))
3424 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3427 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3428 if (signal_pending(current))
3430 if (vma->vm_file && vma->vm_file->f_mapping &&
3431 !is_vm_hugetlb_page(vma))
3432 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3435 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3436 if (signal_pending(current))
3439 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3440 vm_lock_anon_vma(mm, avc->anon_vma);
3446 mm_drop_all_locks(mm);
3450 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3452 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3454 * The LSB of head.next can't change to 0 from under
3455 * us because we hold the mm_all_locks_mutex.
3457 * We must however clear the bitflag before unlocking
3458 * the vma so the users using the anon_vma->rb_root will
3459 * never see our bitflag.
3461 * No need of atomic instructions here, head.next
3462 * can't change from under us until we release the
3463 * anon_vma->root->rwsem.
3465 if (!__test_and_clear_bit(0, (unsigned long *)
3466 &anon_vma->root->rb_root.rb_node))
3468 anon_vma_unlock_write(anon_vma);
3472 static void vm_unlock_mapping(struct address_space *mapping)
3474 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3476 * AS_MM_ALL_LOCKS can't change to 0 from under us
3477 * because we hold the mm_all_locks_mutex.
3479 i_mmap_unlock_write(mapping);
3480 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3487 * The mmap_sem cannot be released by the caller until
3488 * mm_drop_all_locks() returns.
3490 void mm_drop_all_locks(struct mm_struct *mm)
3492 struct vm_area_struct *vma;
3493 struct anon_vma_chain *avc;
3495 BUG_ON(down_read_trylock(&mm->mmap_sem));
3496 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3498 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3500 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3501 vm_unlock_anon_vma(avc->anon_vma);
3502 if (vma->vm_file && vma->vm_file->f_mapping)
3503 vm_unlock_mapping(vma->vm_file->f_mapping);
3506 mutex_unlock(&mm_all_locks_mutex);
3510 * initialise the VMA slab
3512 void __init mmap_init(void)
3516 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3521 * Initialise sysctl_user_reserve_kbytes.
3523 * This is intended to prevent a user from starting a single memory hogging
3524 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3527 * The default value is min(3% of free memory, 128MB)
3528 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3530 static int init_user_reserve(void)
3532 unsigned long free_kbytes;
3534 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3536 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3539 subsys_initcall(init_user_reserve);
3542 * Initialise sysctl_admin_reserve_kbytes.
3544 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3545 * to log in and kill a memory hogging process.
3547 * Systems with more than 256MB will reserve 8MB, enough to recover
3548 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3549 * only reserve 3% of free pages by default.
3551 static int init_admin_reserve(void)
3553 unsigned long free_kbytes;
3555 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3557 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3560 subsys_initcall(init_admin_reserve);
3563 * Reinititalise user and admin reserves if memory is added or removed.
3565 * The default user reserve max is 128MB, and the default max for the
3566 * admin reserve is 8MB. These are usually, but not always, enough to
3567 * enable recovery from a memory hogging process using login/sshd, a shell,
3568 * and tools like top. It may make sense to increase or even disable the
3569 * reserve depending on the existence of swap or variations in the recovery
3570 * tools. So, the admin may have changed them.
3572 * If memory is added and the reserves have been eliminated or increased above
3573 * the default max, then we'll trust the admin.
3575 * If memory is removed and there isn't enough free memory, then we
3576 * need to reset the reserves.
3578 * Otherwise keep the reserve set by the admin.
3580 static int reserve_mem_notifier(struct notifier_block *nb,
3581 unsigned long action, void *data)
3583 unsigned long tmp, free_kbytes;
3587 /* Default max is 128MB. Leave alone if modified by operator. */
3588 tmp = sysctl_user_reserve_kbytes;
3589 if (0 < tmp && tmp < (1UL << 17))
3590 init_user_reserve();
3592 /* Default max is 8MB. Leave alone if modified by operator. */
3593 tmp = sysctl_admin_reserve_kbytes;
3594 if (0 < tmp && tmp < (1UL << 13))
3595 init_admin_reserve();
3599 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3601 if (sysctl_user_reserve_kbytes > free_kbytes) {
3602 init_user_reserve();
3603 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3604 sysctl_user_reserve_kbytes);
3607 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3608 init_admin_reserve();
3609 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3610 sysctl_admin_reserve_kbytes);
3619 static struct notifier_block reserve_mem_nb = {
3620 .notifier_call = reserve_mem_notifier,
3623 static int __meminit init_reserve_notifier(void)
3625 if (register_hotmemory_notifier(&reserve_mem_nb))
3626 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3630 subsys_initcall(init_reserve_notifier);