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>
47 #include <linux/oom.h>
48 #include <linux/sched/mm.h>
50 #include <linux/uaccess.h>
51 #include <asm/cacheflush.h>
53 #include <asm/mmu_context.h>
57 #ifndef arch_mmap_check
58 #define arch_mmap_check(addr, len, flags) (0)
61 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
62 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
63 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
64 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
66 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
67 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
68 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
69 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
72 static bool ignore_rlimit_data;
73 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
75 static void unmap_region(struct mm_struct *mm,
76 struct vm_area_struct *vma, struct vm_area_struct *prev,
77 unsigned long start, unsigned long end);
79 /* description of effects of mapping type and prot in current implementation.
80 * this is due to the limited x86 page protection hardware. The expected
81 * behavior is in parens:
84 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
85 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
86 * w: (no) no w: (no) no w: (yes) yes w: (no) no
87 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
89 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
90 * w: (no) no w: (no) no w: (copy) copy w: (no) no
91 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
93 pgprot_t protection_map[16] __ro_after_init = {
94 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
95 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
98 pgprot_t vm_get_page_prot(unsigned long vm_flags)
100 return __pgprot(pgprot_val(protection_map[vm_flags &
101 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
102 pgprot_val(arch_vm_get_page_prot(vm_flags)));
104 EXPORT_SYMBOL(vm_get_page_prot);
106 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
108 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
111 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
112 void vma_set_page_prot(struct vm_area_struct *vma)
114 unsigned long vm_flags = vma->vm_flags;
115 pgprot_t vm_page_prot;
117 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
118 if (vma_wants_writenotify(vma, vm_page_prot)) {
119 vm_flags &= ~VM_SHARED;
120 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
122 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
123 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
127 * Requires inode->i_mapping->i_mmap_rwsem
129 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
130 struct file *file, struct address_space *mapping)
132 if (vma->vm_flags & VM_DENYWRITE)
133 atomic_inc(&file_inode(file)->i_writecount);
134 if (vma->vm_flags & VM_SHARED)
135 mapping_unmap_writable(mapping);
137 flush_dcache_mmap_lock(mapping);
138 vma_interval_tree_remove(vma, &mapping->i_mmap);
139 flush_dcache_mmap_unlock(mapping);
143 * Unlink a file-based vm structure from its interval tree, to hide
144 * vma from rmap and vmtruncate before freeing its page tables.
146 void unlink_file_vma(struct vm_area_struct *vma)
148 struct file *file = vma->vm_file;
151 struct address_space *mapping = file->f_mapping;
152 i_mmap_lock_write(mapping);
153 __remove_shared_vm_struct(vma, file, mapping);
154 i_mmap_unlock_write(mapping);
159 * Close a vm structure and free it, returning the next.
161 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
163 struct vm_area_struct *next = vma->vm_next;
166 if (vma->vm_ops && vma->vm_ops->close)
167 vma->vm_ops->close(vma);
170 mpol_put(vma_policy(vma));
171 kmem_cache_free(vm_area_cachep, vma);
175 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
176 struct list_head *uf);
177 SYSCALL_DEFINE1(brk, unsigned long, brk)
179 unsigned long retval;
180 unsigned long newbrk, oldbrk;
181 struct mm_struct *mm = current->mm;
182 struct vm_area_struct *next;
183 unsigned long min_brk;
187 if (down_write_killable(&mm->mmap_sem))
190 #ifdef CONFIG_COMPAT_BRK
192 * CONFIG_COMPAT_BRK can still be overridden by setting
193 * randomize_va_space to 2, which will still cause mm->start_brk
194 * to be arbitrarily shifted
196 if (current->brk_randomized)
197 min_brk = mm->start_brk;
199 min_brk = mm->end_data;
201 min_brk = mm->start_brk;
207 * Check against rlimit here. If this check is done later after the test
208 * of oldbrk with newbrk then it can escape the test and let the data
209 * segment grow beyond its set limit the in case where the limit is
210 * not page aligned -Ram Gupta
212 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
213 mm->end_data, mm->start_data))
216 newbrk = PAGE_ALIGN(brk);
217 oldbrk = PAGE_ALIGN(mm->brk);
218 if (oldbrk == newbrk)
221 /* Always allow shrinking brk. */
222 if (brk <= mm->brk) {
223 if (!do_munmap(mm, newbrk, oldbrk-newbrk, &uf))
228 /* Check against existing mmap mappings. */
229 next = find_vma(mm, oldbrk);
230 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
233 /* Ok, looks good - let it rip. */
234 if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
239 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
240 up_write(&mm->mmap_sem);
241 userfaultfd_unmap_complete(mm, &uf);
243 mm_populate(oldbrk, newbrk - oldbrk);
248 up_write(&mm->mmap_sem);
252 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
254 unsigned long max, prev_end, subtree_gap;
257 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
258 * allow two stack_guard_gaps between them here, and when choosing
259 * an unmapped area; whereas when expanding we only require one.
260 * That's a little inconsistent, but keeps the code here simpler.
262 max = vm_start_gap(vma);
264 prev_end = vm_end_gap(vma->vm_prev);
270 if (vma->vm_rb.rb_left) {
271 subtree_gap = rb_entry(vma->vm_rb.rb_left,
272 struct vm_area_struct, vm_rb)->rb_subtree_gap;
273 if (subtree_gap > max)
276 if (vma->vm_rb.rb_right) {
277 subtree_gap = rb_entry(vma->vm_rb.rb_right,
278 struct vm_area_struct, vm_rb)->rb_subtree_gap;
279 if (subtree_gap > max)
285 #ifdef CONFIG_DEBUG_VM_RB
286 static int browse_rb(struct mm_struct *mm)
288 struct rb_root *root = &mm->mm_rb;
289 int i = 0, j, bug = 0;
290 struct rb_node *nd, *pn = NULL;
291 unsigned long prev = 0, pend = 0;
293 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
294 struct vm_area_struct *vma;
295 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
296 if (vma->vm_start < prev) {
297 pr_emerg("vm_start %lx < prev %lx\n",
298 vma->vm_start, prev);
301 if (vma->vm_start < pend) {
302 pr_emerg("vm_start %lx < pend %lx\n",
303 vma->vm_start, pend);
306 if (vma->vm_start > vma->vm_end) {
307 pr_emerg("vm_start %lx > vm_end %lx\n",
308 vma->vm_start, vma->vm_end);
311 spin_lock(&mm->page_table_lock);
312 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
313 pr_emerg("free gap %lx, correct %lx\n",
315 vma_compute_subtree_gap(vma));
318 spin_unlock(&mm->page_table_lock);
321 prev = vma->vm_start;
325 for (nd = pn; nd; nd = rb_prev(nd))
328 pr_emerg("backwards %d, forwards %d\n", j, i);
334 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
338 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
339 struct vm_area_struct *vma;
340 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
341 VM_BUG_ON_VMA(vma != ignore &&
342 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
347 static void validate_mm(struct mm_struct *mm)
351 unsigned long highest_address = 0;
352 struct vm_area_struct *vma = mm->mmap;
355 struct anon_vma *anon_vma = vma->anon_vma;
356 struct anon_vma_chain *avc;
359 anon_vma_lock_read(anon_vma);
360 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
361 anon_vma_interval_tree_verify(avc);
362 anon_vma_unlock_read(anon_vma);
365 highest_address = vm_end_gap(vma);
369 if (i != mm->map_count) {
370 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
373 if (highest_address != mm->highest_vm_end) {
374 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
375 mm->highest_vm_end, highest_address);
379 if (i != mm->map_count) {
381 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
384 VM_BUG_ON_MM(bug, mm);
387 #define validate_mm_rb(root, ignore) do { } while (0)
388 #define validate_mm(mm) do { } while (0)
391 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
392 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
395 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
396 * vma->vm_prev->vm_end values changed, without modifying the vma's position
399 static void vma_gap_update(struct vm_area_struct *vma)
402 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
403 * function that does exacltly what we want.
405 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
408 static inline void vma_rb_insert(struct vm_area_struct *vma,
409 struct rb_root *root)
411 /* All rb_subtree_gap values must be consistent prior to insertion */
412 validate_mm_rb(root, NULL);
414 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
417 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
420 * Note rb_erase_augmented is a fairly large inline function,
421 * so make sure we instantiate it only once with our desired
422 * augmented rbtree callbacks.
424 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
427 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
428 struct rb_root *root,
429 struct vm_area_struct *ignore)
432 * All rb_subtree_gap values must be consistent prior to erase,
433 * with the possible exception of the "next" vma being erased if
434 * next->vm_start was reduced.
436 validate_mm_rb(root, ignore);
438 __vma_rb_erase(vma, root);
441 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
442 struct rb_root *root)
445 * All rb_subtree_gap values must be consistent prior to erase,
446 * with the possible exception of the vma being erased.
448 validate_mm_rb(root, vma);
450 __vma_rb_erase(vma, root);
454 * vma has some anon_vma assigned, and is already inserted on that
455 * anon_vma's interval trees.
457 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
458 * vma must be removed from the anon_vma's interval trees using
459 * anon_vma_interval_tree_pre_update_vma().
461 * After the update, the vma will be reinserted using
462 * anon_vma_interval_tree_post_update_vma().
464 * The entire update must be protected by exclusive mmap_sem and by
465 * the root anon_vma's mutex.
468 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
470 struct anon_vma_chain *avc;
472 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
473 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
477 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
479 struct anon_vma_chain *avc;
481 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
482 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
485 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
486 unsigned long end, struct vm_area_struct **pprev,
487 struct rb_node ***rb_link, struct rb_node **rb_parent)
489 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
491 __rb_link = &mm->mm_rb.rb_node;
492 rb_prev = __rb_parent = NULL;
495 struct vm_area_struct *vma_tmp;
497 __rb_parent = *__rb_link;
498 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
500 if (vma_tmp->vm_end > addr) {
501 /* Fail if an existing vma overlaps the area */
502 if (vma_tmp->vm_start < end)
504 __rb_link = &__rb_parent->rb_left;
506 rb_prev = __rb_parent;
507 __rb_link = &__rb_parent->rb_right;
513 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
514 *rb_link = __rb_link;
515 *rb_parent = __rb_parent;
519 static unsigned long count_vma_pages_range(struct mm_struct *mm,
520 unsigned long addr, unsigned long end)
522 unsigned long nr_pages = 0;
523 struct vm_area_struct *vma;
525 /* Find first overlaping mapping */
526 vma = find_vma_intersection(mm, addr, end);
530 nr_pages = (min(end, vma->vm_end) -
531 max(addr, vma->vm_start)) >> PAGE_SHIFT;
533 /* Iterate over the rest of the overlaps */
534 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
535 unsigned long overlap_len;
537 if (vma->vm_start > end)
540 overlap_len = min(end, vma->vm_end) - vma->vm_start;
541 nr_pages += overlap_len >> PAGE_SHIFT;
547 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
548 struct rb_node **rb_link, struct rb_node *rb_parent)
550 /* Update tracking information for the gap following the new vma. */
552 vma_gap_update(vma->vm_next);
554 mm->highest_vm_end = vm_end_gap(vma);
557 * vma->vm_prev wasn't known when we followed the rbtree to find the
558 * correct insertion point for that vma. As a result, we could not
559 * update the vma vm_rb parents rb_subtree_gap values on the way down.
560 * So, we first insert the vma with a zero rb_subtree_gap value
561 * (to be consistent with what we did on the way down), and then
562 * immediately update the gap to the correct value. Finally we
563 * rebalance the rbtree after all augmented values have been set.
565 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
566 vma->rb_subtree_gap = 0;
568 vma_rb_insert(vma, &mm->mm_rb);
571 static void __vma_link_file(struct vm_area_struct *vma)
577 struct address_space *mapping = file->f_mapping;
579 if (vma->vm_flags & VM_DENYWRITE)
580 atomic_dec(&file_inode(file)->i_writecount);
581 if (vma->vm_flags & VM_SHARED)
582 atomic_inc(&mapping->i_mmap_writable);
584 flush_dcache_mmap_lock(mapping);
585 vma_interval_tree_insert(vma, &mapping->i_mmap);
586 flush_dcache_mmap_unlock(mapping);
591 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
592 struct vm_area_struct *prev, struct rb_node **rb_link,
593 struct rb_node *rb_parent)
595 __vma_link_list(mm, vma, prev, rb_parent);
596 __vma_link_rb(mm, vma, rb_link, rb_parent);
599 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
600 struct vm_area_struct *prev, struct rb_node **rb_link,
601 struct rb_node *rb_parent)
603 struct address_space *mapping = NULL;
606 mapping = vma->vm_file->f_mapping;
607 i_mmap_lock_write(mapping);
610 __vma_link(mm, vma, prev, rb_link, rb_parent);
611 __vma_link_file(vma);
614 i_mmap_unlock_write(mapping);
621 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
622 * mm's list and rbtree. It has already been inserted into the interval tree.
624 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
626 struct vm_area_struct *prev;
627 struct rb_node **rb_link, *rb_parent;
629 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
630 &prev, &rb_link, &rb_parent))
632 __vma_link(mm, vma, prev, rb_link, rb_parent);
636 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
637 struct vm_area_struct *vma,
638 struct vm_area_struct *prev,
640 struct vm_area_struct *ignore)
642 struct vm_area_struct *next;
644 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
647 prev->vm_next = next;
651 prev->vm_next = next;
656 next->vm_prev = prev;
659 vmacache_invalidate(mm);
662 static inline void __vma_unlink_prev(struct mm_struct *mm,
663 struct vm_area_struct *vma,
664 struct vm_area_struct *prev)
666 __vma_unlink_common(mm, vma, prev, true, vma);
670 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
671 * is already present in an i_mmap tree without adjusting the tree.
672 * The following helper function should be used when such adjustments
673 * are necessary. The "insert" vma (if any) is to be inserted
674 * before we drop the necessary locks.
676 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
677 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
678 struct vm_area_struct *expand)
680 struct mm_struct *mm = vma->vm_mm;
681 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
682 struct address_space *mapping = NULL;
683 struct rb_root_cached *root = NULL;
684 struct anon_vma *anon_vma = NULL;
685 struct file *file = vma->vm_file;
686 bool start_changed = false, end_changed = false;
687 long adjust_next = 0;
690 if (next && !insert) {
691 struct vm_area_struct *exporter = NULL, *importer = NULL;
693 if (end >= next->vm_end) {
695 * vma expands, overlapping all the next, and
696 * perhaps the one after too (mprotect case 6).
697 * The only other cases that gets here are
698 * case 1, case 7 and case 8.
700 if (next == expand) {
702 * The only case where we don't expand "vma"
703 * and we expand "next" instead is case 8.
705 VM_WARN_ON(end != next->vm_end);
707 * remove_next == 3 means we're
708 * removing "vma" and that to do so we
709 * swapped "vma" and "next".
712 VM_WARN_ON(file != next->vm_file);
715 VM_WARN_ON(expand != vma);
717 * case 1, 6, 7, remove_next == 2 is case 6,
718 * remove_next == 1 is case 1 or 7.
720 remove_next = 1 + (end > next->vm_end);
721 VM_WARN_ON(remove_next == 2 &&
722 end != next->vm_next->vm_end);
723 VM_WARN_ON(remove_next == 1 &&
724 end != next->vm_end);
725 /* trim end to next, for case 6 first pass */
733 * If next doesn't have anon_vma, import from vma after
734 * next, if the vma overlaps with it.
736 if (remove_next == 2 && !next->anon_vma)
737 exporter = next->vm_next;
739 } else if (end > next->vm_start) {
741 * vma expands, overlapping part of the next:
742 * mprotect case 5 shifting the boundary up.
744 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
747 VM_WARN_ON(expand != importer);
748 } else if (end < vma->vm_end) {
750 * vma shrinks, and !insert tells it's not
751 * split_vma inserting another: so it must be
752 * mprotect case 4 shifting the boundary down.
754 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
757 VM_WARN_ON(expand != importer);
761 * Easily overlooked: when mprotect shifts the boundary,
762 * make sure the expanding vma has anon_vma set if the
763 * shrinking vma had, to cover any anon pages imported.
765 if (exporter && exporter->anon_vma && !importer->anon_vma) {
768 importer->anon_vma = exporter->anon_vma;
769 error = anon_vma_clone(importer, exporter);
775 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
778 mapping = file->f_mapping;
779 root = &mapping->i_mmap;
780 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
783 uprobe_munmap(next, next->vm_start, next->vm_end);
785 i_mmap_lock_write(mapping);
788 * Put into interval tree now, so instantiated pages
789 * are visible to arm/parisc __flush_dcache_page
790 * throughout; but we cannot insert into address
791 * space until vma start or end is updated.
793 __vma_link_file(insert);
797 anon_vma = vma->anon_vma;
798 if (!anon_vma && adjust_next)
799 anon_vma = next->anon_vma;
801 VM_WARN_ON(adjust_next && next->anon_vma &&
802 anon_vma != next->anon_vma);
803 anon_vma_lock_write(anon_vma);
804 anon_vma_interval_tree_pre_update_vma(vma);
806 anon_vma_interval_tree_pre_update_vma(next);
810 flush_dcache_mmap_lock(mapping);
811 vma_interval_tree_remove(vma, root);
813 vma_interval_tree_remove(next, root);
816 if (start != vma->vm_start) {
817 vma->vm_start = start;
818 start_changed = true;
820 if (end != vma->vm_end) {
824 vma->vm_pgoff = pgoff;
826 next->vm_start += adjust_next << PAGE_SHIFT;
827 next->vm_pgoff += adjust_next;
832 vma_interval_tree_insert(next, root);
833 vma_interval_tree_insert(vma, root);
834 flush_dcache_mmap_unlock(mapping);
839 * vma_merge has merged next into vma, and needs
840 * us to remove next before dropping the locks.
842 if (remove_next != 3)
843 __vma_unlink_prev(mm, next, vma);
846 * vma is not before next if they've been
849 * pre-swap() next->vm_start was reduced so
850 * tell validate_mm_rb to ignore pre-swap()
851 * "next" (which is stored in post-swap()
854 __vma_unlink_common(mm, next, NULL, false, vma);
856 __remove_shared_vm_struct(next, file, mapping);
859 * split_vma has split insert from vma, and needs
860 * us to insert it before dropping the locks
861 * (it may either follow vma or precede it).
863 __insert_vm_struct(mm, insert);
869 mm->highest_vm_end = vm_end_gap(vma);
870 else if (!adjust_next)
871 vma_gap_update(next);
876 anon_vma_interval_tree_post_update_vma(vma);
878 anon_vma_interval_tree_post_update_vma(next);
879 anon_vma_unlock_write(anon_vma);
882 i_mmap_unlock_write(mapping);
893 uprobe_munmap(next, next->vm_start, next->vm_end);
897 anon_vma_merge(vma, next);
899 mpol_put(vma_policy(next));
900 kmem_cache_free(vm_area_cachep, next);
902 * In mprotect's case 6 (see comments on vma_merge),
903 * we must remove another next too. It would clutter
904 * up the code too much to do both in one go.
906 if (remove_next != 3) {
908 * If "next" was removed and vma->vm_end was
909 * expanded (up) over it, in turn
910 * "next->vm_prev->vm_end" changed and the
911 * "vma->vm_next" gap must be updated.
916 * For the scope of the comment "next" and
917 * "vma" considered pre-swap(): if "vma" was
918 * removed, next->vm_start was expanded (down)
919 * over it and the "next" gap must be updated.
920 * Because of the swap() the post-swap() "vma"
921 * actually points to pre-swap() "next"
922 * (post-swap() "next" as opposed is now a
927 if (remove_next == 2) {
933 vma_gap_update(next);
936 * If remove_next == 2 we obviously can't
939 * If remove_next == 3 we can't reach this
940 * path because pre-swap() next is always not
941 * NULL. pre-swap() "next" is not being
942 * removed and its next->vm_end is not altered
943 * (and furthermore "end" already matches
944 * next->vm_end in remove_next == 3).
946 * We reach this only in the remove_next == 1
947 * case if the "next" vma that was removed was
948 * the highest vma of the mm. However in such
949 * case next->vm_end == "end" and the extended
950 * "vma" has vma->vm_end == next->vm_end so
951 * mm->highest_vm_end doesn't need any update
952 * in remove_next == 1 case.
954 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
966 * If the vma has a ->close operation then the driver probably needs to release
967 * per-vma resources, so we don't attempt to merge those.
969 static inline int is_mergeable_vma(struct vm_area_struct *vma,
970 struct file *file, unsigned long vm_flags,
971 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
974 * VM_SOFTDIRTY should not prevent from VMA merging, if we
975 * match the flags but dirty bit -- the caller should mark
976 * merged VMA as dirty. If dirty bit won't be excluded from
977 * comparison, we increase pressue on the memory system forcing
978 * the kernel to generate new VMAs when old one could be
981 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
983 if (vma->vm_file != file)
985 if (vma->vm_ops && vma->vm_ops->close)
987 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
992 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
993 struct anon_vma *anon_vma2,
994 struct vm_area_struct *vma)
997 * The list_is_singular() test is to avoid merging VMA cloned from
998 * parents. This can improve scalability caused by anon_vma lock.
1000 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1001 list_is_singular(&vma->anon_vma_chain)))
1003 return anon_vma1 == anon_vma2;
1007 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1008 * in front of (at a lower virtual address and file offset than) the vma.
1010 * We cannot merge two vmas if they have differently assigned (non-NULL)
1011 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1013 * We don't check here for the merged mmap wrapping around the end of pagecache
1014 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1015 * wrap, nor mmaps which cover the final page at index -1UL.
1018 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1019 struct anon_vma *anon_vma, struct file *file,
1021 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1023 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
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)
1044 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1045 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1047 vm_pglen = vma_pages(vma);
1048 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1055 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1056 * whether that can be merged with its predecessor or its successor.
1057 * Or both (it neatly fills a hole).
1059 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1060 * certain not to be mapped by the time vma_merge is called; but when
1061 * called for mprotect, it is certain to be already mapped (either at
1062 * an offset within prev, or at the start of next), and the flags of
1063 * this area are about to be changed to vm_flags - and the no-change
1064 * case has already been eliminated.
1066 * The following mprotect cases have to be considered, where AAAA is
1067 * the area passed down from mprotect_fixup, never extending beyond one
1068 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1070 * AAAA AAAA AAAA AAAA
1071 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1072 * cannot merge might become might become might become
1073 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1074 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1075 * mremap move: PPPPXXXXXXXX 8
1077 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1078 * might become case 1 below case 2 below case 3 below
1080 * It is important for case 8 that the the vma NNNN overlapping the
1081 * region AAAA is never going to extended over XXXX. Instead XXXX must
1082 * be extended in region AAAA and NNNN must be removed. This way in
1083 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1084 * rmap_locks, the properties of the merged vma will be already
1085 * correct for the whole merged range. Some of those properties like
1086 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1087 * be correct for the whole merged range immediately after the
1088 * rmap_locks are released. Otherwise if XXXX would be removed and
1089 * NNNN would be extended over the XXXX range, remove_migration_ptes
1090 * or other rmap walkers (if working on addresses beyond the "end"
1091 * parameter) may establish ptes with the wrong permissions of NNNN
1092 * instead of the right permissions of XXXX.
1094 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1095 struct vm_area_struct *prev, unsigned long addr,
1096 unsigned long end, unsigned long vm_flags,
1097 struct anon_vma *anon_vma, struct file *file,
1098 pgoff_t pgoff, struct mempolicy *policy,
1099 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1101 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1102 struct vm_area_struct *area, *next;
1106 * We later require that vma->vm_flags == vm_flags,
1107 * so this tests vma->vm_flags & VM_SPECIAL, too.
1109 if (vm_flags & VM_SPECIAL)
1113 next = prev->vm_next;
1117 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1118 next = next->vm_next;
1120 /* verify some invariant that must be enforced by the caller */
1121 VM_WARN_ON(prev && addr <= prev->vm_start);
1122 VM_WARN_ON(area && end > area->vm_end);
1123 VM_WARN_ON(addr >= end);
1126 * Can it merge with the predecessor?
1128 if (prev && prev->vm_end == addr &&
1129 mpol_equal(vma_policy(prev), policy) &&
1130 can_vma_merge_after(prev, vm_flags,
1131 anon_vma, file, pgoff,
1132 vm_userfaultfd_ctx)) {
1134 * OK, it can. Can we now merge in the successor as well?
1136 if (next && end == next->vm_start &&
1137 mpol_equal(policy, vma_policy(next)) &&
1138 can_vma_merge_before(next, vm_flags,
1141 vm_userfaultfd_ctx) &&
1142 is_mergeable_anon_vma(prev->anon_vma,
1143 next->anon_vma, NULL)) {
1145 err = __vma_adjust(prev, prev->vm_start,
1146 next->vm_end, prev->vm_pgoff, NULL,
1148 } else /* cases 2, 5, 7 */
1149 err = __vma_adjust(prev, prev->vm_start,
1150 end, prev->vm_pgoff, NULL, prev);
1153 khugepaged_enter_vma_merge(prev, vm_flags);
1158 * Can this new request be merged in front of next?
1160 if (next && end == next->vm_start &&
1161 mpol_equal(policy, vma_policy(next)) &&
1162 can_vma_merge_before(next, vm_flags,
1163 anon_vma, file, pgoff+pglen,
1164 vm_userfaultfd_ctx)) {
1165 if (prev && addr < prev->vm_end) /* case 4 */
1166 err = __vma_adjust(prev, prev->vm_start,
1167 addr, prev->vm_pgoff, NULL, next);
1168 else { /* cases 3, 8 */
1169 err = __vma_adjust(area, addr, next->vm_end,
1170 next->vm_pgoff - pglen, NULL, next);
1172 * In case 3 area is already equal to next and
1173 * this is a noop, but in case 8 "area" has
1174 * been removed and next was expanded over it.
1180 khugepaged_enter_vma_merge(area, vm_flags);
1188 * Rough compatbility check to quickly see if it's even worth looking
1189 * at sharing an anon_vma.
1191 * They need to have the same vm_file, and the flags can only differ
1192 * in things that mprotect may change.
1194 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1195 * we can merge the two vma's. For example, we refuse to merge a vma if
1196 * there is a vm_ops->close() function, because that indicates that the
1197 * driver is doing some kind of reference counting. But that doesn't
1198 * really matter for the anon_vma sharing case.
1200 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1202 return a->vm_end == b->vm_start &&
1203 mpol_equal(vma_policy(a), vma_policy(b)) &&
1204 a->vm_file == b->vm_file &&
1205 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1206 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1210 * Do some basic sanity checking to see if we can re-use the anon_vma
1211 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1212 * the same as 'old', the other will be the new one that is trying
1213 * to share the anon_vma.
1215 * NOTE! This runs with mm_sem held for reading, so it is possible that
1216 * the anon_vma of 'old' is concurrently in the process of being set up
1217 * by another page fault trying to merge _that_. But that's ok: if it
1218 * is being set up, that automatically means that it will be a singleton
1219 * acceptable for merging, so we can do all of this optimistically. But
1220 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1222 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1223 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1224 * is to return an anon_vma that is "complex" due to having gone through
1227 * We also make sure that the two vma's are compatible (adjacent,
1228 * and with the same memory policies). That's all stable, even with just
1229 * a read lock on the mm_sem.
1231 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1233 if (anon_vma_compatible(a, b)) {
1234 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1236 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1243 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1244 * neighbouring vmas for a suitable anon_vma, before it goes off
1245 * to allocate a new anon_vma. It checks because a repetitive
1246 * sequence of mprotects and faults may otherwise lead to distinct
1247 * anon_vmas being allocated, preventing vma merge in subsequent
1250 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1252 struct anon_vma *anon_vma;
1253 struct vm_area_struct *near;
1255 near = vma->vm_next;
1259 anon_vma = reusable_anon_vma(near, vma, near);
1263 near = vma->vm_prev;
1267 anon_vma = reusable_anon_vma(near, near, vma);
1272 * There's no absolute need to look only at touching neighbours:
1273 * we could search further afield for "compatible" anon_vmas.
1274 * But it would probably just be a waste of time searching,
1275 * or lead to too many vmas hanging off the same anon_vma.
1276 * We're trying to allow mprotect remerging later on,
1277 * not trying to minimize memory used for anon_vmas.
1283 * If a hint addr is less than mmap_min_addr change hint to be as
1284 * low as possible but still greater than mmap_min_addr
1286 static inline unsigned long round_hint_to_min(unsigned long hint)
1289 if (((void *)hint != NULL) &&
1290 (hint < mmap_min_addr))
1291 return PAGE_ALIGN(mmap_min_addr);
1295 static inline int mlock_future_check(struct mm_struct *mm,
1296 unsigned long flags,
1299 unsigned long locked, lock_limit;
1301 /* mlock MCL_FUTURE? */
1302 if (flags & VM_LOCKED) {
1303 locked = len >> PAGE_SHIFT;
1304 locked += mm->locked_vm;
1305 lock_limit = rlimit(RLIMIT_MEMLOCK);
1306 lock_limit >>= PAGE_SHIFT;
1307 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1313 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1315 if (S_ISREG(inode->i_mode))
1316 return MAX_LFS_FILESIZE;
1318 if (S_ISBLK(inode->i_mode))
1319 return MAX_LFS_FILESIZE;
1321 /* Special "we do even unsigned file positions" case */
1322 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1325 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1329 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1330 unsigned long pgoff, unsigned long len)
1332 u64 maxsize = file_mmap_size_max(file, inode);
1334 if (maxsize && len > maxsize)
1337 if (pgoff > maxsize >> PAGE_SHIFT)
1343 * The caller must hold down_write(¤t->mm->mmap_sem).
1345 unsigned long do_mmap(struct file *file, unsigned long addr,
1346 unsigned long len, unsigned long prot,
1347 unsigned long flags, vm_flags_t vm_flags,
1348 unsigned long pgoff, unsigned long *populate,
1349 struct list_head *uf)
1351 struct mm_struct *mm = current->mm;
1360 * Does the application expect PROT_READ to imply PROT_EXEC?
1362 * (the exception is when the underlying filesystem is noexec
1363 * mounted, in which case we dont add PROT_EXEC.)
1365 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1366 if (!(file && path_noexec(&file->f_path)))
1369 if (!(flags & MAP_FIXED))
1370 addr = round_hint_to_min(addr);
1372 /* Careful about overflows.. */
1373 len = PAGE_ALIGN(len);
1377 /* offset overflow? */
1378 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1381 /* Too many mappings? */
1382 if (mm->map_count > sysctl_max_map_count)
1385 /* Obtain the address to map to. we verify (or select) it and ensure
1386 * that it represents a valid section of the address space.
1388 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1389 if (offset_in_page(addr))
1392 if (prot == PROT_EXEC) {
1393 pkey = execute_only_pkey(mm);
1398 /* Do simple checking here so the lower-level routines won't have
1399 * to. we assume access permissions have been handled by the open
1400 * of the memory object, so we don't do any here.
1402 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1403 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1405 if (flags & MAP_LOCKED)
1406 if (!can_do_mlock())
1409 if (mlock_future_check(mm, vm_flags, len))
1413 struct inode *inode = file_inode(file);
1415 if (!file_mmap_ok(file, inode, pgoff, len))
1418 switch (flags & MAP_TYPE) {
1420 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1424 * Make sure we don't allow writing to an append-only
1427 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1431 * Make sure there are no mandatory locks on the file.
1433 if (locks_verify_locked(file))
1436 vm_flags |= VM_SHARED | VM_MAYSHARE;
1437 if (!(file->f_mode & FMODE_WRITE))
1438 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1442 if (!(file->f_mode & FMODE_READ))
1444 if (path_noexec(&file->f_path)) {
1445 if (vm_flags & VM_EXEC)
1447 vm_flags &= ~VM_MAYEXEC;
1450 if (!file->f_op->mmap)
1452 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1460 switch (flags & MAP_TYPE) {
1462 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1468 vm_flags |= VM_SHARED | VM_MAYSHARE;
1472 * Set pgoff according to addr for anon_vma.
1474 pgoff = addr >> PAGE_SHIFT;
1482 * Set 'VM_NORESERVE' if we should not account for the
1483 * memory use of this mapping.
1485 if (flags & MAP_NORESERVE) {
1486 /* We honor MAP_NORESERVE if allowed to overcommit */
1487 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1488 vm_flags |= VM_NORESERVE;
1490 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1491 if (file && is_file_hugepages(file))
1492 vm_flags |= VM_NORESERVE;
1495 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1496 if (!IS_ERR_VALUE(addr) &&
1497 ((vm_flags & VM_LOCKED) ||
1498 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1503 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1504 unsigned long, prot, unsigned long, flags,
1505 unsigned long, fd, unsigned long, pgoff)
1507 struct file *file = NULL;
1508 unsigned long retval;
1510 if (!(flags & MAP_ANONYMOUS)) {
1511 audit_mmap_fd(fd, flags);
1515 if (is_file_hugepages(file))
1516 len = ALIGN(len, huge_page_size(hstate_file(file)));
1518 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1520 } else if (flags & MAP_HUGETLB) {
1521 struct user_struct *user = NULL;
1524 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1528 len = ALIGN(len, huge_page_size(hs));
1530 * VM_NORESERVE is used because the reservations will be
1531 * taken when vm_ops->mmap() is called
1532 * A dummy user value is used because we are not locking
1533 * memory so no accounting is necessary
1535 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1537 &user, HUGETLB_ANONHUGE_INODE,
1538 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1540 return PTR_ERR(file);
1543 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1545 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1552 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1553 struct mmap_arg_struct {
1557 unsigned long flags;
1559 unsigned long offset;
1562 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1564 struct mmap_arg_struct a;
1566 if (copy_from_user(&a, arg, sizeof(a)))
1568 if (offset_in_page(a.offset))
1571 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1572 a.offset >> PAGE_SHIFT);
1574 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1577 * Some shared mappigns will want the pages marked read-only
1578 * to track write events. If so, we'll downgrade vm_page_prot
1579 * to the private version (using protection_map[] without the
1582 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1584 vm_flags_t vm_flags = vma->vm_flags;
1585 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1587 /* If it was private or non-writable, the write bit is already clear */
1588 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1591 /* The backer wishes to know when pages are first written to? */
1592 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1595 /* The open routine did something to the protections that pgprot_modify
1596 * won't preserve? */
1597 if (pgprot_val(vm_page_prot) !=
1598 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1602 * Do we need to track softdirty? hugetlb does not support softdirty
1605 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY) &&
1606 !is_vm_hugetlb_page(vma))
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,
1636 struct list_head *uf)
1638 struct mm_struct *mm = current->mm;
1639 struct vm_area_struct *vma, *prev;
1641 struct rb_node **rb_link, *rb_parent;
1642 unsigned long charged = 0;
1644 /* Check against address space limit. */
1645 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1646 unsigned long nr_pages;
1649 * MAP_FIXED may remove pages of mappings that intersects with
1650 * requested mapping. Account for the pages it would unmap.
1652 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1654 if (!may_expand_vm(mm, vm_flags,
1655 (len >> PAGE_SHIFT) - nr_pages))
1659 /* Clear old maps */
1660 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1662 if (do_munmap(mm, addr, len, uf))
1667 * Private writable mapping: check memory availability
1669 if (accountable_mapping(file, vm_flags)) {
1670 charged = len >> PAGE_SHIFT;
1671 if (security_vm_enough_memory_mm(mm, charged))
1673 vm_flags |= VM_ACCOUNT;
1677 * Can we just expand an old mapping?
1679 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1680 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1685 * Determine the object being mapped and call the appropriate
1686 * specific mapper. the address has already been validated, but
1687 * not unmapped, but the maps are removed from the list.
1689 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1696 vma->vm_start = addr;
1697 vma->vm_end = addr + len;
1698 vma->vm_flags = vm_flags;
1699 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1700 vma->vm_pgoff = pgoff;
1701 INIT_LIST_HEAD(&vma->anon_vma_chain);
1704 if (vm_flags & VM_DENYWRITE) {
1705 error = deny_write_access(file);
1709 if (vm_flags & VM_SHARED) {
1710 error = mapping_map_writable(file->f_mapping);
1712 goto allow_write_and_free_vma;
1715 /* ->mmap() can change vma->vm_file, but must guarantee that
1716 * vma_link() below can deny write-access if VM_DENYWRITE is set
1717 * and map writably if VM_SHARED is set. This usually means the
1718 * new file must not have been exposed to user-space, yet.
1720 vma->vm_file = get_file(file);
1721 error = call_mmap(file, vma);
1723 goto unmap_and_free_vma;
1725 /* Can addr have changed??
1727 * Answer: Yes, several device drivers can do it in their
1728 * f_op->mmap method. -DaveM
1729 * Bug: If addr is changed, prev, rb_link, rb_parent should
1730 * be updated for vma_link()
1732 WARN_ON_ONCE(addr != vma->vm_start);
1734 addr = vma->vm_start;
1735 vm_flags = vma->vm_flags;
1736 } else if (vm_flags & VM_SHARED) {
1737 error = shmem_zero_setup(vma);
1742 vma_link(mm, vma, prev, rb_link, rb_parent);
1743 /* Once vma denies write, undo our temporary denial count */
1745 if (vm_flags & VM_SHARED)
1746 mapping_unmap_writable(file->f_mapping);
1747 if (vm_flags & VM_DENYWRITE)
1748 allow_write_access(file);
1750 file = vma->vm_file;
1752 perf_event_mmap(vma);
1754 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1755 if (vm_flags & VM_LOCKED) {
1756 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1757 vma == get_gate_vma(current->mm)))
1758 mm->locked_vm += (len >> PAGE_SHIFT);
1760 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1767 * New (or expanded) vma always get soft dirty status.
1768 * Otherwise user-space soft-dirty page tracker won't
1769 * be able to distinguish situation when vma area unmapped,
1770 * then new mapped in-place (which must be aimed as
1771 * a completely new data area).
1773 vma->vm_flags |= VM_SOFTDIRTY;
1775 vma_set_page_prot(vma);
1780 vma->vm_file = NULL;
1783 /* Undo any partial mapping done by a device driver. */
1784 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 info.align_offset = 0;
2041 return vm_unmapped_area(&info);
2046 * This mmap-allocator allocates new areas top-down from below the
2047 * stack's low limit (the base):
2049 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2051 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2052 const unsigned long len, const unsigned long pgoff,
2053 const unsigned long flags)
2055 struct vm_area_struct *vma, *prev;
2056 struct mm_struct *mm = current->mm;
2057 unsigned long addr = addr0;
2058 struct vm_unmapped_area_info info;
2060 /* requested length too big for entire address space */
2061 if (len > TASK_SIZE - mmap_min_addr)
2064 if (flags & MAP_FIXED)
2067 /* requesting a specific address */
2069 addr = PAGE_ALIGN(addr);
2070 vma = find_vma_prev(mm, addr, &prev);
2071 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2072 (!vma || addr + len <= vm_start_gap(vma)) &&
2073 (!prev || addr >= vm_end_gap(prev)))
2077 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2079 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2080 info.high_limit = mm->mmap_base;
2081 info.align_mask = 0;
2082 info.align_offset = 0;
2083 addr = vm_unmapped_area(&info);
2086 * A failed mmap() very likely causes application failure,
2087 * so fall back to the bottom-up function here. This scenario
2088 * can happen with large stack limits and large mmap()
2091 if (offset_in_page(addr)) {
2092 VM_BUG_ON(addr != -ENOMEM);
2094 info.low_limit = TASK_UNMAPPED_BASE;
2095 info.high_limit = TASK_SIZE;
2096 addr = vm_unmapped_area(&info);
2104 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2105 unsigned long pgoff, unsigned long flags)
2107 unsigned long (*get_area)(struct file *, unsigned long,
2108 unsigned long, unsigned long, unsigned long);
2110 unsigned long error = arch_mmap_check(addr, len, flags);
2114 /* Careful about overflows.. */
2115 if (len > TASK_SIZE)
2118 get_area = current->mm->get_unmapped_area;
2120 if (file->f_op->get_unmapped_area)
2121 get_area = file->f_op->get_unmapped_area;
2122 } else if (flags & MAP_SHARED) {
2124 * mmap_region() will call shmem_zero_setup() to create a file,
2125 * so use shmem's get_unmapped_area in case it can be huge.
2126 * do_mmap_pgoff() will clear pgoff, so match alignment.
2129 get_area = shmem_get_unmapped_area;
2132 addr = get_area(file, addr, len, pgoff, flags);
2133 if (IS_ERR_VALUE(addr))
2136 if (addr > TASK_SIZE - len)
2138 if (offset_in_page(addr))
2141 error = security_mmap_addr(addr);
2142 return error ? error : addr;
2145 EXPORT_SYMBOL(get_unmapped_area);
2147 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2148 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2150 struct rb_node *rb_node;
2151 struct vm_area_struct *vma;
2153 /* Check the cache first. */
2154 vma = vmacache_find(mm, addr);
2158 rb_node = mm->mm_rb.rb_node;
2161 struct vm_area_struct *tmp;
2163 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2165 if (tmp->vm_end > addr) {
2167 if (tmp->vm_start <= addr)
2169 rb_node = rb_node->rb_left;
2171 rb_node = rb_node->rb_right;
2175 vmacache_update(addr, vma);
2179 EXPORT_SYMBOL(find_vma);
2182 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2184 struct vm_area_struct *
2185 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2186 struct vm_area_struct **pprev)
2188 struct vm_area_struct *vma;
2190 vma = find_vma(mm, addr);
2192 *pprev = vma->vm_prev;
2194 struct rb_node *rb_node = mm->mm_rb.rb_node;
2197 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2198 rb_node = rb_node->rb_right;
2205 * Verify that the stack growth is acceptable and
2206 * update accounting. This is shared with both the
2207 * grow-up and grow-down cases.
2209 static int acct_stack_growth(struct vm_area_struct *vma,
2210 unsigned long size, unsigned long grow)
2212 struct mm_struct *mm = vma->vm_mm;
2213 unsigned long new_start;
2215 /* address space limit tests */
2216 if (!may_expand_vm(mm, vma->vm_flags, grow))
2219 /* Stack limit test */
2220 if (size > rlimit(RLIMIT_STACK))
2223 /* mlock limit tests */
2224 if (vma->vm_flags & VM_LOCKED) {
2225 unsigned long locked;
2226 unsigned long limit;
2227 locked = mm->locked_vm + grow;
2228 limit = rlimit(RLIMIT_MEMLOCK);
2229 limit >>= PAGE_SHIFT;
2230 if (locked > limit && !capable(CAP_IPC_LOCK))
2234 /* Check to ensure the stack will not grow into a hugetlb-only region */
2235 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2237 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2241 * Overcommit.. This must be the final test, as it will
2242 * update security statistics.
2244 if (security_vm_enough_memory_mm(mm, grow))
2250 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2252 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2253 * vma is the last one with address > vma->vm_end. Have to extend vma.
2255 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2257 struct mm_struct *mm = vma->vm_mm;
2258 struct vm_area_struct *next;
2259 unsigned long gap_addr;
2262 if (!(vma->vm_flags & VM_GROWSUP))
2265 /* Guard against exceeding limits of the address space. */
2266 address &= PAGE_MASK;
2267 if (address >= (TASK_SIZE & PAGE_MASK))
2269 address += PAGE_SIZE;
2271 /* Enforce stack_guard_gap */
2272 gap_addr = address + stack_guard_gap;
2274 /* Guard against overflow */
2275 if (gap_addr < address || gap_addr > TASK_SIZE)
2276 gap_addr = TASK_SIZE;
2278 next = vma->vm_next;
2279 if (next && next->vm_start < gap_addr &&
2280 (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2281 if (!(next->vm_flags & VM_GROWSUP))
2283 /* Check that both stack segments have the same anon_vma? */
2286 /* We must make sure the anon_vma is allocated. */
2287 if (unlikely(anon_vma_prepare(vma)))
2291 * vma->vm_start/vm_end cannot change under us because the caller
2292 * is required to hold the mmap_sem in read mode. We need the
2293 * anon_vma lock to serialize against concurrent expand_stacks.
2295 anon_vma_lock_write(vma->anon_vma);
2297 /* Somebody else might have raced and expanded it already */
2298 if (address > vma->vm_end) {
2299 unsigned long size, grow;
2301 size = address - vma->vm_start;
2302 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2305 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2306 error = acct_stack_growth(vma, size, grow);
2309 * vma_gap_update() doesn't support concurrent
2310 * updates, but we only hold a shared mmap_sem
2311 * lock here, so we need to protect against
2312 * concurrent vma expansions.
2313 * anon_vma_lock_write() doesn't help here, as
2314 * we don't guarantee that all growable vmas
2315 * in a mm share the same root anon vma.
2316 * So, we reuse mm->page_table_lock to guard
2317 * against concurrent vma expansions.
2319 spin_lock(&mm->page_table_lock);
2320 if (vma->vm_flags & VM_LOCKED)
2321 mm->locked_vm += grow;
2322 vm_stat_account(mm, vma->vm_flags, grow);
2323 anon_vma_interval_tree_pre_update_vma(vma);
2324 vma->vm_end = address;
2325 anon_vma_interval_tree_post_update_vma(vma);
2327 vma_gap_update(vma->vm_next);
2329 mm->highest_vm_end = vm_end_gap(vma);
2330 spin_unlock(&mm->page_table_lock);
2332 perf_event_mmap(vma);
2336 anon_vma_unlock_write(vma->anon_vma);
2337 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2341 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2344 * vma is the first one with address < vma->vm_start. Have to extend vma.
2346 int expand_downwards(struct vm_area_struct *vma,
2347 unsigned long address)
2349 struct mm_struct *mm = vma->vm_mm;
2350 struct vm_area_struct *prev;
2353 address &= PAGE_MASK;
2354 if (address < mmap_min_addr)
2357 /* Enforce stack_guard_gap */
2358 prev = vma->vm_prev;
2359 /* Check that both stack segments have the same anon_vma? */
2360 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2361 (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2362 if (address - prev->vm_end < stack_guard_gap)
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;
2527 struct vm_area_struct *cur_vma;
2530 tlb_gather_mmu(&tlb, mm, start, end);
2531 update_hiwater_rss(mm);
2532 unmap_vmas(&tlb, vma, start, end);
2535 * Ensure we have no stale TLB entries by the time this mapping is
2536 * removed from the rmap.
2537 * Note that we don't have to worry about nested flushes here because
2538 * we're holding the mm semaphore for removing the mapping - so any
2539 * concurrent flush in this region has to be coming through the rmap,
2540 * and we synchronize against that using the rmap lock.
2542 for (cur_vma = vma; cur_vma; cur_vma = cur_vma->vm_next) {
2543 if ((cur_vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) != 0) {
2544 tlb_flush_mmu(&tlb);
2549 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2550 next ? next->vm_start : USER_PGTABLES_CEILING);
2551 tlb_finish_mmu(&tlb, start, end);
2555 * Create a list of vma's touched by the unmap, removing them from the mm's
2556 * vma list as we go..
2559 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2560 struct vm_area_struct *prev, unsigned long end)
2562 struct vm_area_struct **insertion_point;
2563 struct vm_area_struct *tail_vma = NULL;
2565 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2566 vma->vm_prev = NULL;
2568 vma_rb_erase(vma, &mm->mm_rb);
2572 } while (vma && vma->vm_start < end);
2573 *insertion_point = vma;
2575 vma->vm_prev = prev;
2576 vma_gap_update(vma);
2578 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2579 tail_vma->vm_next = NULL;
2581 /* Kill the cache */
2582 vmacache_invalidate(mm);
2586 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2587 * has already been checked or doesn't make sense to fail.
2589 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2590 unsigned long addr, int new_below)
2592 struct vm_area_struct *new;
2595 if (vma->vm_ops && vma->vm_ops->split) {
2596 err = vma->vm_ops->split(vma, addr);
2601 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2605 /* most fields are the same, copy all, and then fixup */
2608 INIT_LIST_HEAD(&new->anon_vma_chain);
2613 new->vm_start = addr;
2614 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2617 err = vma_dup_policy(vma, new);
2621 err = anon_vma_clone(new, vma);
2626 get_file(new->vm_file);
2628 if (new->vm_ops && new->vm_ops->open)
2629 new->vm_ops->open(new);
2632 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2633 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2635 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2641 /* Clean everything up if vma_adjust failed. */
2642 if (new->vm_ops && new->vm_ops->close)
2643 new->vm_ops->close(new);
2646 unlink_anon_vmas(new);
2648 mpol_put(vma_policy(new));
2650 kmem_cache_free(vm_area_cachep, new);
2655 * Split a vma into two pieces at address 'addr', a new vma is allocated
2656 * either for the first part or the tail.
2658 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2659 unsigned long addr, int new_below)
2661 if (mm->map_count >= sysctl_max_map_count)
2664 return __split_vma(mm, vma, addr, new_below);
2667 /* Munmap is split into 2 main parts -- this part which finds
2668 * what needs doing, and the areas themselves, which do the
2669 * work. This now handles partial unmappings.
2670 * Jeremy Fitzhardinge <jeremy@goop.org>
2672 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2673 struct list_head *uf)
2676 struct vm_area_struct *vma, *prev, *last;
2678 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2681 len = PAGE_ALIGN(len);
2685 /* Find the first overlapping VMA */
2686 vma = find_vma(mm, start);
2689 prev = vma->vm_prev;
2690 /* we have start < vma->vm_end */
2692 /* if it doesn't overlap, we have nothing.. */
2694 if (vma->vm_start >= end)
2698 * If we need to split any vma, do it now to save pain later.
2700 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2701 * unmapped vm_area_struct will remain in use: so lower split_vma
2702 * places tmp vma above, and higher split_vma places tmp vma below.
2704 if (start > vma->vm_start) {
2708 * Make sure that map_count on return from munmap() will
2709 * not exceed its limit; but let map_count go just above
2710 * its limit temporarily, to help free resources as expected.
2712 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2715 error = __split_vma(mm, vma, start, 0);
2721 /* Does it split the last one? */
2722 last = find_vma(mm, end);
2723 if (last && end > last->vm_start) {
2724 int error = __split_vma(mm, last, end, 1);
2728 vma = prev ? prev->vm_next : mm->mmap;
2732 * If userfaultfd_unmap_prep returns an error the vmas
2733 * will remain splitted, but userland will get a
2734 * highly unexpected error anyway. This is no
2735 * different than the case where the first of the two
2736 * __split_vma fails, but we don't undo the first
2737 * split, despite we could. This is unlikely enough
2738 * failure that it's not worth optimizing it for.
2740 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2746 * unlock any mlock()ed ranges before detaching vmas
2748 if (mm->locked_vm) {
2749 struct vm_area_struct *tmp = vma;
2750 while (tmp && tmp->vm_start < end) {
2751 if (tmp->vm_flags & VM_LOCKED) {
2752 mm->locked_vm -= vma_pages(tmp);
2753 munlock_vma_pages_all(tmp);
2760 * Remove the vma's, and unmap the actual pages
2762 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2763 unmap_region(mm, vma, prev, start, end);
2765 arch_unmap(mm, vma, start, end);
2767 /* Fix up all other VM information */
2768 remove_vma_list(mm, vma);
2773 int vm_munmap(unsigned long start, size_t len)
2776 struct mm_struct *mm = current->mm;
2779 if (down_write_killable(&mm->mmap_sem))
2782 ret = do_munmap(mm, start, len, &uf);
2783 up_write(&mm->mmap_sem);
2784 userfaultfd_unmap_complete(mm, &uf);
2787 EXPORT_SYMBOL(vm_munmap);
2789 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2791 profile_munmap(addr);
2792 return vm_munmap(addr, len);
2797 * Emulation of deprecated remap_file_pages() syscall.
2799 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2800 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2803 struct mm_struct *mm = current->mm;
2804 struct vm_area_struct *vma;
2805 unsigned long populate = 0;
2806 unsigned long ret = -EINVAL;
2809 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2810 current->comm, current->pid);
2814 start = start & PAGE_MASK;
2815 size = size & PAGE_MASK;
2817 if (start + size <= start)
2820 /* Does pgoff wrap? */
2821 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2824 if (down_write_killable(&mm->mmap_sem))
2827 vma = find_vma(mm, start);
2829 if (!vma || !(vma->vm_flags & VM_SHARED))
2832 if (start < vma->vm_start)
2835 if (start + size > vma->vm_end) {
2836 struct vm_area_struct *next;
2838 for (next = vma->vm_next; next; next = next->vm_next) {
2839 /* hole between vmas ? */
2840 if (next->vm_start != next->vm_prev->vm_end)
2843 if (next->vm_file != vma->vm_file)
2846 if (next->vm_flags != vma->vm_flags)
2849 if (start + size <= next->vm_end)
2857 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2858 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2859 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2861 flags &= MAP_NONBLOCK;
2862 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2863 if (vma->vm_flags & VM_LOCKED) {
2864 struct vm_area_struct *tmp;
2865 flags |= MAP_LOCKED;
2867 /* drop PG_Mlocked flag for over-mapped range */
2868 for (tmp = vma; tmp->vm_start >= start + size;
2869 tmp = tmp->vm_next) {
2871 * Split pmd and munlock page on the border
2874 vma_adjust_trans_huge(tmp, start, start + size, 0);
2876 munlock_vma_pages_range(tmp,
2877 max(tmp->vm_start, start),
2878 min(tmp->vm_end, start + size));
2882 file = get_file(vma->vm_file);
2883 ret = do_mmap_pgoff(vma->vm_file, start, size,
2884 prot, flags, pgoff, &populate, NULL);
2887 up_write(&mm->mmap_sem);
2889 mm_populate(ret, populate);
2890 if (!IS_ERR_VALUE(ret))
2895 static inline void verify_mm_writelocked(struct mm_struct *mm)
2897 #ifdef CONFIG_DEBUG_VM
2898 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2900 up_read(&mm->mmap_sem);
2906 * this is really a simplified "do_mmap". it only handles
2907 * anonymous maps. eventually we may be able to do some
2908 * brk-specific accounting here.
2910 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
2912 struct mm_struct *mm = current->mm;
2913 struct vm_area_struct *vma, *prev;
2914 struct rb_node **rb_link, *rb_parent;
2915 pgoff_t pgoff = addr >> PAGE_SHIFT;
2918 /* Until we need other flags, refuse anything except VM_EXEC. */
2919 if ((flags & (~VM_EXEC)) != 0)
2921 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2923 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2924 if (offset_in_page(error))
2927 error = mlock_future_check(mm, mm->def_flags, len);
2932 * mm->mmap_sem is required to protect against another thread
2933 * changing the mappings in case we sleep.
2935 verify_mm_writelocked(mm);
2938 * Clear old maps. this also does some error checking for us
2940 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2942 if (do_munmap(mm, addr, len, uf))
2946 /* Check against address space limits *after* clearing old maps... */
2947 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2950 if (mm->map_count > sysctl_max_map_count)
2953 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2956 /* Can we just expand an old private anonymous mapping? */
2957 vma = vma_merge(mm, prev, addr, addr + len, flags,
2958 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2963 * create a vma struct for an anonymous mapping
2965 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2967 vm_unacct_memory(len >> PAGE_SHIFT);
2971 INIT_LIST_HEAD(&vma->anon_vma_chain);
2973 vma->vm_start = addr;
2974 vma->vm_end = addr + len;
2975 vma->vm_pgoff = pgoff;
2976 vma->vm_flags = flags;
2977 vma->vm_page_prot = vm_get_page_prot(flags);
2978 vma_link(mm, vma, prev, rb_link, rb_parent);
2980 perf_event_mmap(vma);
2981 mm->total_vm += len >> PAGE_SHIFT;
2982 mm->data_vm += len >> PAGE_SHIFT;
2983 if (flags & VM_LOCKED)
2984 mm->locked_vm += (len >> PAGE_SHIFT);
2985 vma->vm_flags |= VM_SOFTDIRTY;
2989 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
2991 struct mm_struct *mm = current->mm;
2997 len = PAGE_ALIGN(request);
3003 if (down_write_killable(&mm->mmap_sem))
3006 ret = do_brk_flags(addr, len, flags, &uf);
3007 populate = ((mm->def_flags & VM_LOCKED) != 0);
3008 up_write(&mm->mmap_sem);
3009 userfaultfd_unmap_complete(mm, &uf);
3010 if (populate && !ret)
3011 mm_populate(addr, len);
3014 EXPORT_SYMBOL(vm_brk_flags);
3016 int vm_brk(unsigned long addr, unsigned long len)
3018 return vm_brk_flags(addr, len, 0);
3020 EXPORT_SYMBOL(vm_brk);
3022 /* Release all mmaps. */
3023 void exit_mmap(struct mm_struct *mm)
3025 struct mmu_gather tlb;
3026 struct vm_area_struct *vma;
3027 unsigned long nr_accounted = 0;
3029 /* mm's last user has gone, and its about to be pulled down */
3030 mmu_notifier_release(mm);
3032 if (unlikely(mm_is_oom_victim(mm))) {
3034 * Manually reap the mm to free as much memory as possible.
3035 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3036 * this mm from further consideration. Taking mm->mmap_sem for
3037 * write after setting MMF_OOM_SKIP will guarantee that the oom
3038 * reaper will not run on this mm again after mmap_sem is
3041 * Nothing can be holding mm->mmap_sem here and the above call
3042 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3043 * __oom_reap_task_mm() will not block.
3045 * This needs to be done before calling munlock_vma_pages_all(),
3046 * which clears VM_LOCKED, otherwise the oom reaper cannot
3049 mutex_lock(&oom_lock);
3050 __oom_reap_task_mm(mm);
3051 mutex_unlock(&oom_lock);
3053 set_bit(MMF_OOM_SKIP, &mm->flags);
3054 down_write(&mm->mmap_sem);
3055 up_write(&mm->mmap_sem);
3058 if (mm->locked_vm) {
3061 if (vma->vm_flags & VM_LOCKED)
3062 munlock_vma_pages_all(vma);
3070 if (!vma) /* Can happen if dup_mmap() received an OOM */
3075 tlb_gather_mmu(&tlb, mm, 0, -1);
3076 /* update_hiwater_rss(mm) here? but nobody should be looking */
3077 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3078 unmap_vmas(&tlb, vma, 0, -1);
3079 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3080 tlb_finish_mmu(&tlb, 0, -1);
3083 * Walk the list again, actually closing and freeing it,
3084 * with preemption enabled, without holding any MM locks.
3087 if (vma->vm_flags & VM_ACCOUNT)
3088 nr_accounted += vma_pages(vma);
3089 vma = remove_vma(vma);
3092 vm_unacct_memory(nr_accounted);
3095 /* Insert vm structure into process list sorted by address
3096 * and into the inode's i_mmap tree. If vm_file is non-NULL
3097 * then i_mmap_rwsem is taken here.
3099 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3101 struct vm_area_struct *prev;
3102 struct rb_node **rb_link, *rb_parent;
3104 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3105 &prev, &rb_link, &rb_parent))
3107 if ((vma->vm_flags & VM_ACCOUNT) &&
3108 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3112 * The vm_pgoff of a purely anonymous vma should be irrelevant
3113 * until its first write fault, when page's anon_vma and index
3114 * are set. But now set the vm_pgoff it will almost certainly
3115 * end up with (unless mremap moves it elsewhere before that
3116 * first wfault), so /proc/pid/maps tells a consistent story.
3118 * By setting it to reflect the virtual start address of the
3119 * vma, merges and splits can happen in a seamless way, just
3120 * using the existing file pgoff checks and manipulations.
3121 * Similarly in do_mmap_pgoff and in do_brk.
3123 if (vma_is_anonymous(vma)) {
3124 BUG_ON(vma->anon_vma);
3125 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3128 vma_link(mm, vma, prev, rb_link, rb_parent);
3133 * Copy the vma structure to a new location in the same mm,
3134 * prior to moving page table entries, to effect an mremap move.
3136 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3137 unsigned long addr, unsigned long len, pgoff_t pgoff,
3138 bool *need_rmap_locks)
3140 struct vm_area_struct *vma = *vmap;
3141 unsigned long vma_start = vma->vm_start;
3142 struct mm_struct *mm = vma->vm_mm;
3143 struct vm_area_struct *new_vma, *prev;
3144 struct rb_node **rb_link, *rb_parent;
3145 bool faulted_in_anon_vma = true;
3148 * If anonymous vma has not yet been faulted, update new pgoff
3149 * to match new location, to increase its chance of merging.
3151 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3152 pgoff = addr >> PAGE_SHIFT;
3153 faulted_in_anon_vma = false;
3156 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3157 return NULL; /* should never get here */
3158 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3159 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3160 vma->vm_userfaultfd_ctx);
3163 * Source vma may have been merged into new_vma
3165 if (unlikely(vma_start >= new_vma->vm_start &&
3166 vma_start < new_vma->vm_end)) {
3168 * The only way we can get a vma_merge with
3169 * self during an mremap is if the vma hasn't
3170 * been faulted in yet and we were allowed to
3171 * reset the dst vma->vm_pgoff to the
3172 * destination address of the mremap to allow
3173 * the merge to happen. mremap must change the
3174 * vm_pgoff linearity between src and dst vmas
3175 * (in turn preventing a vma_merge) to be
3176 * safe. It is only safe to keep the vm_pgoff
3177 * linear if there are no pages mapped yet.
3179 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3180 *vmap = vma = new_vma;
3182 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3184 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3188 new_vma->vm_start = addr;
3189 new_vma->vm_end = addr + len;
3190 new_vma->vm_pgoff = pgoff;
3191 if (vma_dup_policy(vma, new_vma))
3193 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3194 if (anon_vma_clone(new_vma, vma))
3195 goto out_free_mempol;
3196 if (new_vma->vm_file)
3197 get_file(new_vma->vm_file);
3198 if (new_vma->vm_ops && new_vma->vm_ops->open)
3199 new_vma->vm_ops->open(new_vma);
3200 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3201 *need_rmap_locks = false;
3206 mpol_put(vma_policy(new_vma));
3208 kmem_cache_free(vm_area_cachep, new_vma);
3214 * Return true if the calling process may expand its vm space by the passed
3217 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3219 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3222 if (is_data_mapping(flags) &&
3223 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3224 /* Workaround for Valgrind */
3225 if (rlimit(RLIMIT_DATA) == 0 &&
3226 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3228 if (!ignore_rlimit_data) {
3229 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits or use boot option ignore_rlimit_data.\n",
3230 current->comm, current->pid,
3231 (mm->data_vm + npages) << PAGE_SHIFT,
3232 rlimit(RLIMIT_DATA));
3240 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3242 mm->total_vm += npages;
3244 if (is_exec_mapping(flags))
3245 mm->exec_vm += npages;
3246 else if (is_stack_mapping(flags))
3247 mm->stack_vm += npages;
3248 else if (is_data_mapping(flags))
3249 mm->data_vm += npages;
3252 static int special_mapping_fault(struct vm_fault *vmf);
3255 * Having a close hook prevents vma merging regardless of flags.
3257 static void special_mapping_close(struct vm_area_struct *vma)
3261 static const char *special_mapping_name(struct vm_area_struct *vma)
3263 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3266 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3268 struct vm_special_mapping *sm = new_vma->vm_private_data;
3270 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3274 return sm->mremap(sm, new_vma);
3279 static const struct vm_operations_struct special_mapping_vmops = {
3280 .close = special_mapping_close,
3281 .fault = special_mapping_fault,
3282 .mremap = special_mapping_mremap,
3283 .name = special_mapping_name,
3286 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3287 .close = special_mapping_close,
3288 .fault = special_mapping_fault,
3291 static int special_mapping_fault(struct vm_fault *vmf)
3293 struct vm_area_struct *vma = vmf->vma;
3295 struct page **pages;
3297 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3298 pages = vma->vm_private_data;
3300 struct vm_special_mapping *sm = vma->vm_private_data;
3303 return sm->fault(sm, vmf->vma, vmf);
3308 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3312 struct page *page = *pages;
3318 return VM_FAULT_SIGBUS;
3321 static struct vm_area_struct *__install_special_mapping(
3322 struct mm_struct *mm,
3323 unsigned long addr, unsigned long len,
3324 unsigned long vm_flags, void *priv,
3325 const struct vm_operations_struct *ops)
3328 struct vm_area_struct *vma;
3330 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3331 if (unlikely(vma == NULL))
3332 return ERR_PTR(-ENOMEM);
3334 INIT_LIST_HEAD(&vma->anon_vma_chain);
3336 vma->vm_start = addr;
3337 vma->vm_end = addr + len;
3339 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3340 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3343 vma->vm_private_data = priv;
3345 ret = insert_vm_struct(mm, vma);
3349 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3351 perf_event_mmap(vma);
3356 kmem_cache_free(vm_area_cachep, vma);
3357 return ERR_PTR(ret);
3360 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3361 const struct vm_special_mapping *sm)
3363 return vma->vm_private_data == sm &&
3364 (vma->vm_ops == &special_mapping_vmops ||
3365 vma->vm_ops == &legacy_special_mapping_vmops);
3369 * Called with mm->mmap_sem held for writing.
3370 * Insert a new vma covering the given region, with the given flags.
3371 * Its pages are supplied by the given array of struct page *.
3372 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3373 * The region past the last page supplied will always produce SIGBUS.
3374 * The array pointer and the pages it points to are assumed to stay alive
3375 * for as long as this mapping might exist.
3377 struct vm_area_struct *_install_special_mapping(
3378 struct mm_struct *mm,
3379 unsigned long addr, unsigned long len,
3380 unsigned long vm_flags, const struct vm_special_mapping *spec)
3382 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3383 &special_mapping_vmops);
3386 int install_special_mapping(struct mm_struct *mm,
3387 unsigned long addr, unsigned long len,
3388 unsigned long vm_flags, struct page **pages)
3390 struct vm_area_struct *vma = __install_special_mapping(
3391 mm, addr, len, vm_flags, (void *)pages,
3392 &legacy_special_mapping_vmops);
3394 return PTR_ERR_OR_ZERO(vma);
3397 static DEFINE_MUTEX(mm_all_locks_mutex);
3399 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3401 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3403 * The LSB of head.next can't change from under us
3404 * because we hold the mm_all_locks_mutex.
3406 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3408 * We can safely modify head.next after taking the
3409 * anon_vma->root->rwsem. If some other vma in this mm shares
3410 * the same anon_vma we won't take it again.
3412 * No need of atomic instructions here, head.next
3413 * can't change from under us thanks to the
3414 * anon_vma->root->rwsem.
3416 if (__test_and_set_bit(0, (unsigned long *)
3417 &anon_vma->root->rb_root.rb_root.rb_node))
3422 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3424 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3426 * AS_MM_ALL_LOCKS can't change from under us because
3427 * we hold the mm_all_locks_mutex.
3429 * Operations on ->flags have to be atomic because
3430 * even if AS_MM_ALL_LOCKS is stable thanks to the
3431 * mm_all_locks_mutex, there may be other cpus
3432 * changing other bitflags in parallel to us.
3434 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3436 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3441 * This operation locks against the VM for all pte/vma/mm related
3442 * operations that could ever happen on a certain mm. This includes
3443 * vmtruncate, try_to_unmap, and all page faults.
3445 * The caller must take the mmap_sem in write mode before calling
3446 * mm_take_all_locks(). The caller isn't allowed to release the
3447 * mmap_sem until mm_drop_all_locks() returns.
3449 * mmap_sem in write mode is required in order to block all operations
3450 * that could modify pagetables and free pages without need of
3451 * altering the vma layout. It's also needed in write mode to avoid new
3452 * anon_vmas to be associated with existing vmas.
3454 * A single task can't take more than one mm_take_all_locks() in a row
3455 * or it would deadlock.
3457 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3458 * mapping->flags avoid to take the same lock twice, if more than one
3459 * vma in this mm is backed by the same anon_vma or address_space.
3461 * We take locks in following order, accordingly to comment at beginning
3463 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3465 * - all i_mmap_rwsem locks;
3466 * - all anon_vma->rwseml
3468 * We can take all locks within these types randomly because the VM code
3469 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3470 * mm_all_locks_mutex.
3472 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3473 * that may have to take thousand of locks.
3475 * mm_take_all_locks() can fail if it's interrupted by signals.
3477 int mm_take_all_locks(struct mm_struct *mm)
3479 struct vm_area_struct *vma;
3480 struct anon_vma_chain *avc;
3482 BUG_ON(down_read_trylock(&mm->mmap_sem));
3484 mutex_lock(&mm_all_locks_mutex);
3486 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3487 if (signal_pending(current))
3489 if (vma->vm_file && vma->vm_file->f_mapping &&
3490 is_vm_hugetlb_page(vma))
3491 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3494 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3495 if (signal_pending(current))
3497 if (vma->vm_file && vma->vm_file->f_mapping &&
3498 !is_vm_hugetlb_page(vma))
3499 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3502 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3503 if (signal_pending(current))
3506 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3507 vm_lock_anon_vma(mm, avc->anon_vma);
3513 mm_drop_all_locks(mm);
3517 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3519 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3521 * The LSB of head.next can't change to 0 from under
3522 * us because we hold the mm_all_locks_mutex.
3524 * We must however clear the bitflag before unlocking
3525 * the vma so the users using the anon_vma->rb_root will
3526 * never see our bitflag.
3528 * No need of atomic instructions here, head.next
3529 * can't change from under us until we release the
3530 * anon_vma->root->rwsem.
3532 if (!__test_and_clear_bit(0, (unsigned long *)
3533 &anon_vma->root->rb_root.rb_root.rb_node))
3535 anon_vma_unlock_write(anon_vma);
3539 static void vm_unlock_mapping(struct address_space *mapping)
3541 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3543 * AS_MM_ALL_LOCKS can't change to 0 from under us
3544 * because we hold the mm_all_locks_mutex.
3546 i_mmap_unlock_write(mapping);
3547 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3554 * The mmap_sem cannot be released by the caller until
3555 * mm_drop_all_locks() returns.
3557 void mm_drop_all_locks(struct mm_struct *mm)
3559 struct vm_area_struct *vma;
3560 struct anon_vma_chain *avc;
3562 BUG_ON(down_read_trylock(&mm->mmap_sem));
3563 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3565 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3567 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3568 vm_unlock_anon_vma(avc->anon_vma);
3569 if (vma->vm_file && vma->vm_file->f_mapping)
3570 vm_unlock_mapping(vma->vm_file->f_mapping);
3573 mutex_unlock(&mm_all_locks_mutex);
3577 * initialise the percpu counter for VM
3579 void __init mmap_init(void)
3583 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3588 * Initialise sysctl_user_reserve_kbytes.
3590 * This is intended to prevent a user from starting a single memory hogging
3591 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3594 * The default value is min(3% of free memory, 128MB)
3595 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3597 static int init_user_reserve(void)
3599 unsigned long free_kbytes;
3601 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3603 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3606 subsys_initcall(init_user_reserve);
3609 * Initialise sysctl_admin_reserve_kbytes.
3611 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3612 * to log in and kill a memory hogging process.
3614 * Systems with more than 256MB will reserve 8MB, enough to recover
3615 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3616 * only reserve 3% of free pages by default.
3618 static int init_admin_reserve(void)
3620 unsigned long free_kbytes;
3622 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3624 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3627 subsys_initcall(init_admin_reserve);
3630 * Reinititalise user and admin reserves if memory is added or removed.
3632 * The default user reserve max is 128MB, and the default max for the
3633 * admin reserve is 8MB. These are usually, but not always, enough to
3634 * enable recovery from a memory hogging process using login/sshd, a shell,
3635 * and tools like top. It may make sense to increase or even disable the
3636 * reserve depending on the existence of swap or variations in the recovery
3637 * tools. So, the admin may have changed them.
3639 * If memory is added and the reserves have been eliminated or increased above
3640 * the default max, then we'll trust the admin.
3642 * If memory is removed and there isn't enough free memory, then we
3643 * need to reset the reserves.
3645 * Otherwise keep the reserve set by the admin.
3647 static int reserve_mem_notifier(struct notifier_block *nb,
3648 unsigned long action, void *data)
3650 unsigned long tmp, free_kbytes;
3654 /* Default max is 128MB. Leave alone if modified by operator. */
3655 tmp = sysctl_user_reserve_kbytes;
3656 if (0 < tmp && tmp < (1UL << 17))
3657 init_user_reserve();
3659 /* Default max is 8MB. Leave alone if modified by operator. */
3660 tmp = sysctl_admin_reserve_kbytes;
3661 if (0 < tmp && tmp < (1UL << 13))
3662 init_admin_reserve();
3666 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3668 if (sysctl_user_reserve_kbytes > free_kbytes) {
3669 init_user_reserve();
3670 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3671 sysctl_user_reserve_kbytes);
3674 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3675 init_admin_reserve();
3676 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3677 sysctl_admin_reserve_kbytes);
3686 static struct notifier_block reserve_mem_nb = {
3687 .notifier_call = reserve_mem_notifier,
3690 static int __meminit init_reserve_notifier(void)
3692 if (register_hotmemory_notifier(&reserve_mem_nb))
3693 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3697 subsys_initcall(init_reserve_notifier);