6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/shmem_fs.h>
29 #include <linux/profile.h>
30 #include <linux/export.h>
31 #include <linux/mount.h>
32 #include <linux/mempolicy.h>
33 #include <linux/rmap.h>
34 #include <linux/mmu_notifier.h>
35 #include <linux/mmdebug.h>
36 #include <linux/perf_event.h>
37 #include <linux/audit.h>
38 #include <linux/khugepaged.h>
39 #include <linux/uprobes.h>
40 #include <linux/rbtree_augmented.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/moduleparam.h>
46 #include <linux/pkeys.h>
48 #include <asm/uaccess.h>
49 #include <asm/cacheflush.h>
51 #include <asm/mmu_context.h>
55 #ifndef arch_mmap_check
56 #define arch_mmap_check(addr, len, flags) (0)
59 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
60 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
61 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
62 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
64 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
65 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
66 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
67 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
70 static bool ignore_rlimit_data;
71 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
73 static void unmap_region(struct mm_struct *mm,
74 struct vm_area_struct *vma, struct vm_area_struct *prev,
75 unsigned long start, unsigned long end);
77 /* description of effects of mapping type and prot in current implementation.
78 * this is due to the limited x86 page protection hardware. The expected
79 * behavior is in parens:
82 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
83 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
84 * w: (no) no w: (no) no w: (yes) yes w: (no) no
85 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
87 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
88 * w: (no) no w: (no) no w: (copy) copy w: (no) no
89 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
91 pgprot_t protection_map[16] = {
92 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
93 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
96 pgprot_t vm_get_page_prot(unsigned long vm_flags)
98 return __pgprot(pgprot_val(protection_map[vm_flags &
99 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
100 pgprot_val(arch_vm_get_page_prot(vm_flags)));
102 EXPORT_SYMBOL(vm_get_page_prot);
104 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
106 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
109 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
110 void vma_set_page_prot(struct vm_area_struct *vma)
112 unsigned long vm_flags = vma->vm_flags;
113 pgprot_t vm_page_prot;
115 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
116 if (vma_wants_writenotify(vma, vm_page_prot)) {
117 vm_flags &= ~VM_SHARED;
118 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
120 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
121 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
125 * Requires inode->i_mapping->i_mmap_rwsem
127 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
128 struct file *file, struct address_space *mapping)
130 if (vma->vm_flags & VM_DENYWRITE)
131 atomic_inc(&file_inode(file)->i_writecount);
132 if (vma->vm_flags & VM_SHARED)
133 mapping_unmap_writable(mapping);
135 flush_dcache_mmap_lock(mapping);
136 vma_interval_tree_remove(vma, &mapping->i_mmap);
137 flush_dcache_mmap_unlock(mapping);
141 * Unlink a file-based vm structure from its interval tree, to hide
142 * vma from rmap and vmtruncate before freeing its page tables.
144 void unlink_file_vma(struct vm_area_struct *vma)
146 struct file *file = vma->vm_file;
149 struct address_space *mapping = file->f_mapping;
150 i_mmap_lock_write(mapping);
151 __remove_shared_vm_struct(vma, file, mapping);
152 i_mmap_unlock_write(mapping);
157 * Close a vm structure and free it, returning the next.
159 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
161 struct vm_area_struct *next = vma->vm_next;
164 if (vma->vm_ops && vma->vm_ops->close)
165 vma->vm_ops->close(vma);
168 mpol_put(vma_policy(vma));
169 kmem_cache_free(vm_area_cachep, vma);
173 static int do_brk(unsigned long addr, unsigned long len);
175 SYSCALL_DEFINE1(brk, unsigned long, brk)
177 unsigned long retval;
178 unsigned long newbrk, oldbrk;
179 struct mm_struct *mm = current->mm;
180 struct vm_area_struct *next;
181 unsigned long min_brk;
184 if (down_write_killable(&mm->mmap_sem))
187 #ifdef CONFIG_COMPAT_BRK
189 * CONFIG_COMPAT_BRK can still be overridden by setting
190 * randomize_va_space to 2, which will still cause mm->start_brk
191 * to be arbitrarily shifted
193 if (current->brk_randomized)
194 min_brk = mm->start_brk;
196 min_brk = mm->end_data;
198 min_brk = mm->start_brk;
204 * Check against rlimit here. If this check is done later after the test
205 * of oldbrk with newbrk then it can escape the test and let the data
206 * segment grow beyond its set limit the in case where the limit is
207 * not page aligned -Ram Gupta
209 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
210 mm->end_data, mm->start_data))
213 newbrk = PAGE_ALIGN(brk);
214 oldbrk = PAGE_ALIGN(mm->brk);
215 if (oldbrk == newbrk)
218 /* Always allow shrinking brk. */
219 if (brk <= mm->brk) {
220 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
225 /* Check against existing mmap mappings. */
226 next = find_vma(mm, oldbrk);
227 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
230 /* Ok, looks good - let it rip. */
231 if (do_brk(oldbrk, newbrk-oldbrk) < 0)
236 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
237 up_write(&mm->mmap_sem);
239 mm_populate(oldbrk, newbrk - oldbrk);
244 up_write(&mm->mmap_sem);
248 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
250 unsigned long max, prev_end, subtree_gap;
253 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
254 * allow two stack_guard_gaps between them here, and when choosing
255 * an unmapped area; whereas when expanding we only require one.
256 * That's a little inconsistent, but keeps the code here simpler.
258 max = vm_start_gap(vma);
260 prev_end = vm_end_gap(vma->vm_prev);
266 if (vma->vm_rb.rb_left) {
267 subtree_gap = rb_entry(vma->vm_rb.rb_left,
268 struct vm_area_struct, vm_rb)->rb_subtree_gap;
269 if (subtree_gap > max)
272 if (vma->vm_rb.rb_right) {
273 subtree_gap = rb_entry(vma->vm_rb.rb_right,
274 struct vm_area_struct, vm_rb)->rb_subtree_gap;
275 if (subtree_gap > max)
281 #ifdef CONFIG_DEBUG_VM_RB
282 static int browse_rb(struct mm_struct *mm)
284 struct rb_root *root = &mm->mm_rb;
285 int i = 0, j, bug = 0;
286 struct rb_node *nd, *pn = NULL;
287 unsigned long prev = 0, pend = 0;
289 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
290 struct vm_area_struct *vma;
291 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
292 if (vma->vm_start < prev) {
293 pr_emerg("vm_start %lx < prev %lx\n",
294 vma->vm_start, prev);
297 if (vma->vm_start < pend) {
298 pr_emerg("vm_start %lx < pend %lx\n",
299 vma->vm_start, pend);
302 if (vma->vm_start > vma->vm_end) {
303 pr_emerg("vm_start %lx > vm_end %lx\n",
304 vma->vm_start, vma->vm_end);
307 spin_lock(&mm->page_table_lock);
308 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
309 pr_emerg("free gap %lx, correct %lx\n",
311 vma_compute_subtree_gap(vma));
314 spin_unlock(&mm->page_table_lock);
317 prev = vma->vm_start;
321 for (nd = pn; nd; nd = rb_prev(nd))
324 pr_emerg("backwards %d, forwards %d\n", j, i);
330 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
334 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
335 struct vm_area_struct *vma;
336 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
337 VM_BUG_ON_VMA(vma != ignore &&
338 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
343 static void validate_mm(struct mm_struct *mm)
347 unsigned long highest_address = 0;
348 struct vm_area_struct *vma = mm->mmap;
351 struct anon_vma *anon_vma = vma->anon_vma;
352 struct anon_vma_chain *avc;
355 anon_vma_lock_read(anon_vma);
356 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
357 anon_vma_interval_tree_verify(avc);
358 anon_vma_unlock_read(anon_vma);
361 highest_address = vm_end_gap(vma);
365 if (i != mm->map_count) {
366 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
369 if (highest_address != mm->highest_vm_end) {
370 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
371 mm->highest_vm_end, highest_address);
375 if (i != mm->map_count) {
377 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
380 VM_BUG_ON_MM(bug, mm);
383 #define validate_mm_rb(root, ignore) do { } while (0)
384 #define validate_mm(mm) do { } while (0)
387 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
388 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
391 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
392 * vma->vm_prev->vm_end values changed, without modifying the vma's position
395 static void vma_gap_update(struct vm_area_struct *vma)
398 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
399 * function that does exacltly what we want.
401 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
404 static inline void vma_rb_insert(struct vm_area_struct *vma,
405 struct rb_root *root)
407 /* All rb_subtree_gap values must be consistent prior to insertion */
408 validate_mm_rb(root, NULL);
410 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
413 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
416 * Note rb_erase_augmented is a fairly large inline function,
417 * so make sure we instantiate it only once with our desired
418 * augmented rbtree callbacks.
420 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
423 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
424 struct rb_root *root,
425 struct vm_area_struct *ignore)
428 * All rb_subtree_gap values must be consistent prior to erase,
429 * with the possible exception of the "next" vma being erased if
430 * next->vm_start was reduced.
432 validate_mm_rb(root, ignore);
434 __vma_rb_erase(vma, root);
437 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
438 struct rb_root *root)
441 * All rb_subtree_gap values must be consistent prior to erase,
442 * with the possible exception of the vma being erased.
444 validate_mm_rb(root, vma);
446 __vma_rb_erase(vma, root);
450 * vma has some anon_vma assigned, and is already inserted on that
451 * anon_vma's interval trees.
453 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
454 * vma must be removed from the anon_vma's interval trees using
455 * anon_vma_interval_tree_pre_update_vma().
457 * After the update, the vma will be reinserted using
458 * anon_vma_interval_tree_post_update_vma().
460 * The entire update must be protected by exclusive mmap_sem and by
461 * the root anon_vma's mutex.
464 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
466 struct anon_vma_chain *avc;
468 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
469 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
473 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
475 struct anon_vma_chain *avc;
477 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
478 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
481 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
482 unsigned long end, struct vm_area_struct **pprev,
483 struct rb_node ***rb_link, struct rb_node **rb_parent)
485 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
487 __rb_link = &mm->mm_rb.rb_node;
488 rb_prev = __rb_parent = NULL;
491 struct vm_area_struct *vma_tmp;
493 __rb_parent = *__rb_link;
494 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
496 if (vma_tmp->vm_end > addr) {
497 /* Fail if an existing vma overlaps the area */
498 if (vma_tmp->vm_start < end)
500 __rb_link = &__rb_parent->rb_left;
502 rb_prev = __rb_parent;
503 __rb_link = &__rb_parent->rb_right;
509 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
510 *rb_link = __rb_link;
511 *rb_parent = __rb_parent;
515 static unsigned long count_vma_pages_range(struct mm_struct *mm,
516 unsigned long addr, unsigned long end)
518 unsigned long nr_pages = 0;
519 struct vm_area_struct *vma;
521 /* Find first overlaping mapping */
522 vma = find_vma_intersection(mm, addr, end);
526 nr_pages = (min(end, vma->vm_end) -
527 max(addr, vma->vm_start)) >> PAGE_SHIFT;
529 /* Iterate over the rest of the overlaps */
530 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
531 unsigned long overlap_len;
533 if (vma->vm_start > end)
536 overlap_len = min(end, vma->vm_end) - vma->vm_start;
537 nr_pages += overlap_len >> PAGE_SHIFT;
543 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
544 struct rb_node **rb_link, struct rb_node *rb_parent)
546 /* Update tracking information for the gap following the new vma. */
548 vma_gap_update(vma->vm_next);
550 mm->highest_vm_end = vm_end_gap(vma);
553 * vma->vm_prev wasn't known when we followed the rbtree to find the
554 * correct insertion point for that vma. As a result, we could not
555 * update the vma vm_rb parents rb_subtree_gap values on the way down.
556 * So, we first insert the vma with a zero rb_subtree_gap value
557 * (to be consistent with what we did on the way down), and then
558 * immediately update the gap to the correct value. Finally we
559 * rebalance the rbtree after all augmented values have been set.
561 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
562 vma->rb_subtree_gap = 0;
564 vma_rb_insert(vma, &mm->mm_rb);
567 static void __vma_link_file(struct vm_area_struct *vma)
573 struct address_space *mapping = file->f_mapping;
575 if (vma->vm_flags & VM_DENYWRITE)
576 atomic_dec(&file_inode(file)->i_writecount);
577 if (vma->vm_flags & VM_SHARED)
578 atomic_inc(&mapping->i_mmap_writable);
580 flush_dcache_mmap_lock(mapping);
581 vma_interval_tree_insert(vma, &mapping->i_mmap);
582 flush_dcache_mmap_unlock(mapping);
587 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
588 struct vm_area_struct *prev, struct rb_node **rb_link,
589 struct rb_node *rb_parent)
591 __vma_link_list(mm, vma, prev, rb_parent);
592 __vma_link_rb(mm, vma, rb_link, rb_parent);
595 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
596 struct vm_area_struct *prev, struct rb_node **rb_link,
597 struct rb_node *rb_parent)
599 struct address_space *mapping = NULL;
602 mapping = vma->vm_file->f_mapping;
603 i_mmap_lock_write(mapping);
606 __vma_link(mm, vma, prev, rb_link, rb_parent);
607 __vma_link_file(vma);
610 i_mmap_unlock_write(mapping);
617 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
618 * mm's list and rbtree. It has already been inserted into the interval tree.
620 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
622 struct vm_area_struct *prev;
623 struct rb_node **rb_link, *rb_parent;
625 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
626 &prev, &rb_link, &rb_parent))
628 __vma_link(mm, vma, prev, rb_link, rb_parent);
632 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
633 struct vm_area_struct *vma,
634 struct vm_area_struct *prev,
636 struct vm_area_struct *ignore)
638 struct vm_area_struct *next;
640 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
643 prev->vm_next = next;
647 prev->vm_next = next;
652 next->vm_prev = prev;
655 vmacache_invalidate(mm);
658 static inline void __vma_unlink_prev(struct mm_struct *mm,
659 struct vm_area_struct *vma,
660 struct vm_area_struct *prev)
662 __vma_unlink_common(mm, vma, prev, true, vma);
666 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
667 * is already present in an i_mmap tree without adjusting the tree.
668 * The following helper function should be used when such adjustments
669 * are necessary. The "insert" vma (if any) is to be inserted
670 * before we drop the necessary locks.
672 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
673 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
674 struct vm_area_struct *expand)
676 struct mm_struct *mm = vma->vm_mm;
677 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
678 struct address_space *mapping = NULL;
679 struct rb_root *root = NULL;
680 struct anon_vma *anon_vma = NULL;
681 struct file *file = vma->vm_file;
682 bool start_changed = false, end_changed = false;
683 long adjust_next = 0;
686 if (next && !insert) {
687 struct vm_area_struct *exporter = NULL, *importer = NULL;
689 if (end >= next->vm_end) {
691 * vma expands, overlapping all the next, and
692 * perhaps the one after too (mprotect case 6).
693 * The only other cases that gets here are
694 * case 1, case 7 and case 8.
696 if (next == expand) {
698 * The only case where we don't expand "vma"
699 * and we expand "next" instead is case 8.
701 VM_WARN_ON(end != next->vm_end);
703 * remove_next == 3 means we're
704 * removing "vma" and that to do so we
705 * swapped "vma" and "next".
708 VM_WARN_ON(file != next->vm_file);
711 VM_WARN_ON(expand != vma);
713 * case 1, 6, 7, remove_next == 2 is case 6,
714 * remove_next == 1 is case 1 or 7.
716 remove_next = 1 + (end > next->vm_end);
717 VM_WARN_ON(remove_next == 2 &&
718 end != next->vm_next->vm_end);
719 VM_WARN_ON(remove_next == 1 &&
720 end != next->vm_end);
721 /* trim end to next, for case 6 first pass */
729 * If next doesn't have anon_vma, import from vma after
730 * next, if the vma overlaps with it.
732 if (remove_next == 2 && !next->anon_vma)
733 exporter = next->vm_next;
735 } else if (end > next->vm_start) {
737 * vma expands, overlapping part of the next:
738 * mprotect case 5 shifting the boundary up.
740 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
743 VM_WARN_ON(expand != importer);
744 } else if (end < vma->vm_end) {
746 * vma shrinks, and !insert tells it's not
747 * split_vma inserting another: so it must be
748 * mprotect case 4 shifting the boundary down.
750 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
753 VM_WARN_ON(expand != importer);
757 * Easily overlooked: when mprotect shifts the boundary,
758 * make sure the expanding vma has anon_vma set if the
759 * shrinking vma had, to cover any anon pages imported.
761 if (exporter && exporter->anon_vma && !importer->anon_vma) {
764 importer->anon_vma = exporter->anon_vma;
765 error = anon_vma_clone(importer, exporter);
771 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
774 mapping = file->f_mapping;
775 root = &mapping->i_mmap;
776 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
779 uprobe_munmap(next, next->vm_start, next->vm_end);
781 i_mmap_lock_write(mapping);
784 * Put into interval tree now, so instantiated pages
785 * are visible to arm/parisc __flush_dcache_page
786 * throughout; but we cannot insert into address
787 * space until vma start or end is updated.
789 __vma_link_file(insert);
793 anon_vma = vma->anon_vma;
794 if (!anon_vma && adjust_next)
795 anon_vma = next->anon_vma;
797 VM_WARN_ON(adjust_next && next->anon_vma &&
798 anon_vma != next->anon_vma);
799 anon_vma_lock_write(anon_vma);
800 anon_vma_interval_tree_pre_update_vma(vma);
802 anon_vma_interval_tree_pre_update_vma(next);
806 flush_dcache_mmap_lock(mapping);
807 vma_interval_tree_remove(vma, root);
809 vma_interval_tree_remove(next, root);
812 if (start != vma->vm_start) {
813 vma->vm_start = start;
814 start_changed = true;
816 if (end != vma->vm_end) {
820 vma->vm_pgoff = pgoff;
822 next->vm_start += adjust_next << PAGE_SHIFT;
823 next->vm_pgoff += adjust_next;
828 vma_interval_tree_insert(next, root);
829 vma_interval_tree_insert(vma, root);
830 flush_dcache_mmap_unlock(mapping);
835 * vma_merge has merged next into vma, and needs
836 * us to remove next before dropping the locks.
838 if (remove_next != 3)
839 __vma_unlink_prev(mm, next, vma);
842 * vma is not before next if they've been
845 * pre-swap() next->vm_start was reduced so
846 * tell validate_mm_rb to ignore pre-swap()
847 * "next" (which is stored in post-swap()
850 __vma_unlink_common(mm, next, NULL, false, vma);
852 __remove_shared_vm_struct(next, file, mapping);
855 * split_vma has split insert from vma, and needs
856 * us to insert it before dropping the locks
857 * (it may either follow vma or precede it).
859 __insert_vm_struct(mm, insert);
865 mm->highest_vm_end = vm_end_gap(vma);
866 else if (!adjust_next)
867 vma_gap_update(next);
872 anon_vma_interval_tree_post_update_vma(vma);
874 anon_vma_interval_tree_post_update_vma(next);
875 anon_vma_unlock_write(anon_vma);
878 i_mmap_unlock_write(mapping);
889 uprobe_munmap(next, next->vm_start, next->vm_end);
893 anon_vma_merge(vma, next);
895 mpol_put(vma_policy(next));
896 kmem_cache_free(vm_area_cachep, next);
898 * In mprotect's case 6 (see comments on vma_merge),
899 * we must remove another next too. It would clutter
900 * up the code too much to do both in one go.
902 if (remove_next != 3) {
904 * If "next" was removed and vma->vm_end was
905 * expanded (up) over it, in turn
906 * "next->vm_prev->vm_end" changed and the
907 * "vma->vm_next" gap must be updated.
912 * For the scope of the comment "next" and
913 * "vma" considered pre-swap(): if "vma" was
914 * removed, next->vm_start was expanded (down)
915 * over it and the "next" gap must be updated.
916 * Because of the swap() the post-swap() "vma"
917 * actually points to pre-swap() "next"
918 * (post-swap() "next" as opposed is now a
923 if (remove_next == 2) {
929 vma_gap_update(next);
932 * If remove_next == 2 we obviously can't
935 * If remove_next == 3 we can't reach this
936 * path because pre-swap() next is always not
937 * NULL. pre-swap() "next" is not being
938 * removed and its next->vm_end is not altered
939 * (and furthermore "end" already matches
940 * next->vm_end in remove_next == 3).
942 * We reach this only in the remove_next == 1
943 * case if the "next" vma that was removed was
944 * the highest vma of the mm. However in such
945 * case next->vm_end == "end" and the extended
946 * "vma" has vma->vm_end == next->vm_end so
947 * mm->highest_vm_end doesn't need any update
948 * in remove_next == 1 case.
950 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
962 * If the vma has a ->close operation then the driver probably needs to release
963 * per-vma resources, so we don't attempt to merge those.
965 static inline int is_mergeable_vma(struct vm_area_struct *vma,
966 struct file *file, unsigned long vm_flags,
967 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
970 * VM_SOFTDIRTY should not prevent from VMA merging, if we
971 * match the flags but dirty bit -- the caller should mark
972 * merged VMA as dirty. If dirty bit won't be excluded from
973 * comparison, we increase pressue on the memory system forcing
974 * the kernel to generate new VMAs when old one could be
977 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
979 if (vma->vm_file != file)
981 if (vma->vm_ops && vma->vm_ops->close)
983 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
988 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
989 struct anon_vma *anon_vma2,
990 struct vm_area_struct *vma)
993 * The list_is_singular() test is to avoid merging VMA cloned from
994 * parents. This can improve scalability caused by anon_vma lock.
996 if ((!anon_vma1 || !anon_vma2) && (!vma ||
997 list_is_singular(&vma->anon_vma_chain)))
999 return anon_vma1 == anon_vma2;
1003 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1004 * in front of (at a lower virtual address and file offset than) the vma.
1006 * We cannot merge two vmas if they have differently assigned (non-NULL)
1007 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1009 * We don't check here for the merged mmap wrapping around the end of pagecache
1010 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1011 * wrap, nor mmaps which cover the final page at index -1UL.
1014 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1015 struct anon_vma *anon_vma, struct file *file,
1017 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1019 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1020 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1021 if (vma->vm_pgoff == vm_pgoff)
1028 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1029 * beyond (at a higher virtual address and file offset than) the vma.
1031 * We cannot merge two vmas if they have differently assigned (non-NULL)
1032 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1035 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1036 struct anon_vma *anon_vma, struct file *file,
1038 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1040 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1041 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1043 vm_pglen = vma_pages(vma);
1044 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1051 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1052 * whether that can be merged with its predecessor or its successor.
1053 * Or both (it neatly fills a hole).
1055 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1056 * certain not to be mapped by the time vma_merge is called; but when
1057 * called for mprotect, it is certain to be already mapped (either at
1058 * an offset within prev, or at the start of next), and the flags of
1059 * this area are about to be changed to vm_flags - and the no-change
1060 * case has already been eliminated.
1062 * The following mprotect cases have to be considered, where AAAA is
1063 * the area passed down from mprotect_fixup, never extending beyond one
1064 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1066 * AAAA AAAA AAAA AAAA
1067 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1068 * cannot merge might become might become might become
1069 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1070 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1071 * mremap move: PPPPXXXXXXXX 8
1073 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1074 * might become case 1 below case 2 below case 3 below
1076 * It is important for case 8 that the the vma NNNN overlapping the
1077 * region AAAA is never going to extended over XXXX. Instead XXXX must
1078 * be extended in region AAAA and NNNN must be removed. This way in
1079 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1080 * rmap_locks, the properties of the merged vma will be already
1081 * correct for the whole merged range. Some of those properties like
1082 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1083 * be correct for the whole merged range immediately after the
1084 * rmap_locks are released. Otherwise if XXXX would be removed and
1085 * NNNN would be extended over the XXXX range, remove_migration_ptes
1086 * or other rmap walkers (if working on addresses beyond the "end"
1087 * parameter) may establish ptes with the wrong permissions of NNNN
1088 * instead of the right permissions of XXXX.
1090 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1091 struct vm_area_struct *prev, unsigned long addr,
1092 unsigned long end, unsigned long vm_flags,
1093 struct anon_vma *anon_vma, struct file *file,
1094 pgoff_t pgoff, struct mempolicy *policy,
1095 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1097 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1098 struct vm_area_struct *area, *next;
1102 * We later require that vma->vm_flags == vm_flags,
1103 * so this tests vma->vm_flags & VM_SPECIAL, too.
1105 if (vm_flags & VM_SPECIAL)
1109 next = prev->vm_next;
1113 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1114 next = next->vm_next;
1116 /* verify some invariant that must be enforced by the caller */
1117 VM_WARN_ON(prev && addr <= prev->vm_start);
1118 VM_WARN_ON(area && end > area->vm_end);
1119 VM_WARN_ON(addr >= end);
1122 * Can it merge with the predecessor?
1124 if (prev && prev->vm_end == addr &&
1125 mpol_equal(vma_policy(prev), policy) &&
1126 can_vma_merge_after(prev, vm_flags,
1127 anon_vma, file, pgoff,
1128 vm_userfaultfd_ctx)) {
1130 * OK, it can. Can we now merge in the successor as well?
1132 if (next && end == next->vm_start &&
1133 mpol_equal(policy, vma_policy(next)) &&
1134 can_vma_merge_before(next, vm_flags,
1137 vm_userfaultfd_ctx) &&
1138 is_mergeable_anon_vma(prev->anon_vma,
1139 next->anon_vma, NULL)) {
1141 err = __vma_adjust(prev, prev->vm_start,
1142 next->vm_end, prev->vm_pgoff, NULL,
1144 } else /* cases 2, 5, 7 */
1145 err = __vma_adjust(prev, prev->vm_start,
1146 end, prev->vm_pgoff, NULL, prev);
1149 khugepaged_enter_vma_merge(prev, vm_flags);
1154 * Can this new request be merged in front of next?
1156 if (next && end == next->vm_start &&
1157 mpol_equal(policy, vma_policy(next)) &&
1158 can_vma_merge_before(next, vm_flags,
1159 anon_vma, file, pgoff+pglen,
1160 vm_userfaultfd_ctx)) {
1161 if (prev && addr < prev->vm_end) /* case 4 */
1162 err = __vma_adjust(prev, prev->vm_start,
1163 addr, prev->vm_pgoff, NULL, next);
1164 else { /* cases 3, 8 */
1165 err = __vma_adjust(area, addr, next->vm_end,
1166 next->vm_pgoff - pglen, NULL, next);
1168 * In case 3 area is already equal to next and
1169 * this is a noop, but in case 8 "area" has
1170 * been removed and next was expanded over it.
1176 khugepaged_enter_vma_merge(area, vm_flags);
1184 * Rough compatbility check to quickly see if it's even worth looking
1185 * at sharing an anon_vma.
1187 * They need to have the same vm_file, and the flags can only differ
1188 * in things that mprotect may change.
1190 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1191 * we can merge the two vma's. For example, we refuse to merge a vma if
1192 * there is a vm_ops->close() function, because that indicates that the
1193 * driver is doing some kind of reference counting. But that doesn't
1194 * really matter for the anon_vma sharing case.
1196 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1198 return a->vm_end == b->vm_start &&
1199 mpol_equal(vma_policy(a), vma_policy(b)) &&
1200 a->vm_file == b->vm_file &&
1201 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1202 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1206 * Do some basic sanity checking to see if we can re-use the anon_vma
1207 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1208 * the same as 'old', the other will be the new one that is trying
1209 * to share the anon_vma.
1211 * NOTE! This runs with mm_sem held for reading, so it is possible that
1212 * the anon_vma of 'old' is concurrently in the process of being set up
1213 * by another page fault trying to merge _that_. But that's ok: if it
1214 * is being set up, that automatically means that it will be a singleton
1215 * acceptable for merging, so we can do all of this optimistically. But
1216 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1218 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1219 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1220 * is to return an anon_vma that is "complex" due to having gone through
1223 * We also make sure that the two vma's are compatible (adjacent,
1224 * and with the same memory policies). That's all stable, even with just
1225 * a read lock on the mm_sem.
1227 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1229 if (anon_vma_compatible(a, b)) {
1230 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1232 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1239 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1240 * neighbouring vmas for a suitable anon_vma, before it goes off
1241 * to allocate a new anon_vma. It checks because a repetitive
1242 * sequence of mprotects and faults may otherwise lead to distinct
1243 * anon_vmas being allocated, preventing vma merge in subsequent
1246 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1248 struct anon_vma *anon_vma;
1249 struct vm_area_struct *near;
1251 near = vma->vm_next;
1255 anon_vma = reusable_anon_vma(near, vma, near);
1259 near = vma->vm_prev;
1263 anon_vma = reusable_anon_vma(near, near, vma);
1268 * There's no absolute need to look only at touching neighbours:
1269 * we could search further afield for "compatible" anon_vmas.
1270 * But it would probably just be a waste of time searching,
1271 * or lead to too many vmas hanging off the same anon_vma.
1272 * We're trying to allow mprotect remerging later on,
1273 * not trying to minimize memory used for anon_vmas.
1279 * If a hint addr is less than mmap_min_addr change hint to be as
1280 * low as possible but still greater than mmap_min_addr
1282 static inline unsigned long round_hint_to_min(unsigned long hint)
1285 if (((void *)hint != NULL) &&
1286 (hint < mmap_min_addr))
1287 return PAGE_ALIGN(mmap_min_addr);
1291 static inline int mlock_future_check(struct mm_struct *mm,
1292 unsigned long flags,
1295 unsigned long locked, lock_limit;
1297 /* mlock MCL_FUTURE? */
1298 if (flags & VM_LOCKED) {
1299 locked = len >> PAGE_SHIFT;
1300 locked += mm->locked_vm;
1301 lock_limit = rlimit(RLIMIT_MEMLOCK);
1302 lock_limit >>= PAGE_SHIFT;
1303 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1309 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1311 if (S_ISREG(inode->i_mode))
1312 return MAX_LFS_FILESIZE;
1314 if (S_ISBLK(inode->i_mode))
1315 return MAX_LFS_FILESIZE;
1317 /* Special "we do even unsigned file positions" case */
1318 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1321 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1325 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1326 unsigned long pgoff, unsigned long len)
1328 u64 maxsize = file_mmap_size_max(file, inode);
1330 if (maxsize && len > maxsize)
1333 if (pgoff > maxsize >> PAGE_SHIFT)
1339 * The caller must hold down_write(¤t->mm->mmap_sem).
1341 unsigned long do_mmap(struct file *file, unsigned long addr,
1342 unsigned long len, unsigned long prot,
1343 unsigned long flags, vm_flags_t vm_flags,
1344 unsigned long pgoff, unsigned long *populate)
1346 struct mm_struct *mm = current->mm;
1355 * Does the application expect PROT_READ to imply PROT_EXEC?
1357 * (the exception is when the underlying filesystem is noexec
1358 * mounted, in which case we dont add PROT_EXEC.)
1360 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1361 if (!(file && path_noexec(&file->f_path)))
1364 if (!(flags & MAP_FIXED))
1365 addr = round_hint_to_min(addr);
1367 /* Careful about overflows.. */
1368 len = PAGE_ALIGN(len);
1372 /* offset overflow? */
1373 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1376 /* Too many mappings? */
1377 if (mm->map_count > sysctl_max_map_count)
1380 /* Obtain the address to map to. we verify (or select) it and ensure
1381 * that it represents a valid section of the address space.
1383 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1384 if (offset_in_page(addr))
1387 if (prot == PROT_EXEC) {
1388 pkey = execute_only_pkey(mm);
1393 /* Do simple checking here so the lower-level routines won't have
1394 * to. we assume access permissions have been handled by the open
1395 * of the memory object, so we don't do any here.
1397 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1398 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1400 if (flags & MAP_LOCKED)
1401 if (!can_do_mlock())
1404 if (mlock_future_check(mm, vm_flags, len))
1408 struct inode *inode = file_inode(file);
1410 if (!file_mmap_ok(file, inode, pgoff, len))
1413 switch (flags & MAP_TYPE) {
1415 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1419 * Make sure we don't allow writing to an append-only
1422 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1426 * Make sure there are no mandatory locks on the file.
1428 if (locks_verify_locked(file))
1431 vm_flags |= VM_SHARED | VM_MAYSHARE;
1432 if (!(file->f_mode & FMODE_WRITE))
1433 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1437 if (!(file->f_mode & FMODE_READ))
1439 if (path_noexec(&file->f_path)) {
1440 if (vm_flags & VM_EXEC)
1442 vm_flags &= ~VM_MAYEXEC;
1445 if (!file->f_op->mmap)
1447 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1455 switch (flags & MAP_TYPE) {
1457 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1463 vm_flags |= VM_SHARED | VM_MAYSHARE;
1467 * Set pgoff according to addr for anon_vma.
1469 pgoff = addr >> PAGE_SHIFT;
1477 * Set 'VM_NORESERVE' if we should not account for the
1478 * memory use of this mapping.
1480 if (flags & MAP_NORESERVE) {
1481 /* We honor MAP_NORESERVE if allowed to overcommit */
1482 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1483 vm_flags |= VM_NORESERVE;
1485 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1486 if (file && is_file_hugepages(file))
1487 vm_flags |= VM_NORESERVE;
1490 addr = mmap_region(file, addr, len, vm_flags, pgoff);
1491 if (!IS_ERR_VALUE(addr) &&
1492 ((vm_flags & VM_LOCKED) ||
1493 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1498 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1499 unsigned long, prot, unsigned long, flags,
1500 unsigned long, fd, unsigned long, pgoff)
1502 struct file *file = NULL;
1503 unsigned long retval;
1505 if (!(flags & MAP_ANONYMOUS)) {
1506 audit_mmap_fd(fd, flags);
1510 if (is_file_hugepages(file))
1511 len = ALIGN(len, huge_page_size(hstate_file(file)));
1513 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1515 } else if (flags & MAP_HUGETLB) {
1516 struct user_struct *user = NULL;
1519 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1523 len = ALIGN(len, huge_page_size(hs));
1525 * VM_NORESERVE is used because the reservations will be
1526 * taken when vm_ops->mmap() is called
1527 * A dummy user value is used because we are not locking
1528 * memory so no accounting is necessary
1530 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1532 &user, HUGETLB_ANONHUGE_INODE,
1533 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1535 return PTR_ERR(file);
1538 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1540 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1547 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1548 struct mmap_arg_struct {
1552 unsigned long flags;
1554 unsigned long offset;
1557 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1559 struct mmap_arg_struct a;
1561 if (copy_from_user(&a, arg, sizeof(a)))
1563 if (offset_in_page(a.offset))
1566 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1567 a.offset >> PAGE_SHIFT);
1569 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1572 * Some shared mappigns will want the pages marked read-only
1573 * to track write events. If so, we'll downgrade vm_page_prot
1574 * to the private version (using protection_map[] without the
1577 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1579 vm_flags_t vm_flags = vma->vm_flags;
1580 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1582 /* If it was private or non-writable, the write bit is already clear */
1583 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1586 /* The backer wishes to know when pages are first written to? */
1587 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1590 /* The open routine did something to the protections that pgprot_modify
1591 * won't preserve? */
1592 if (pgprot_val(vm_page_prot) !=
1593 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1596 /* Do we need to track softdirty? */
1597 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1600 /* Specialty mapping? */
1601 if (vm_flags & VM_PFNMAP)
1604 /* Can the mapping track the dirty pages? */
1605 return vma->vm_file && vma->vm_file->f_mapping &&
1606 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1610 * We account for memory if it's a private writeable mapping,
1611 * not hugepages and VM_NORESERVE wasn't set.
1613 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1616 * hugetlb has its own accounting separate from the core VM
1617 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1619 if (file && is_file_hugepages(file))
1622 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1625 unsigned long mmap_region(struct file *file, unsigned long addr,
1626 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1628 struct mm_struct *mm = current->mm;
1629 struct vm_area_struct *vma, *prev;
1631 struct rb_node **rb_link, *rb_parent;
1632 unsigned long charged = 0;
1634 /* Check against address space limit. */
1635 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1636 unsigned long nr_pages;
1639 * MAP_FIXED may remove pages of mappings that intersects with
1640 * requested mapping. Account for the pages it would unmap.
1642 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1644 if (!may_expand_vm(mm, vm_flags,
1645 (len >> PAGE_SHIFT) - nr_pages))
1649 /* Clear old maps */
1650 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1652 if (do_munmap(mm, addr, len))
1657 * Private writable mapping: check memory availability
1659 if (accountable_mapping(file, vm_flags)) {
1660 charged = len >> PAGE_SHIFT;
1661 if (security_vm_enough_memory_mm(mm, charged))
1663 vm_flags |= VM_ACCOUNT;
1667 * Can we just expand an old mapping?
1669 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1670 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1675 * Determine the object being mapped and call the appropriate
1676 * specific mapper. the address has already been validated, but
1677 * not unmapped, but the maps are removed from the list.
1679 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1686 vma->vm_start = addr;
1687 vma->vm_end = addr + len;
1688 vma->vm_flags = vm_flags;
1689 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1690 vma->vm_pgoff = pgoff;
1691 INIT_LIST_HEAD(&vma->anon_vma_chain);
1694 if (vm_flags & VM_DENYWRITE) {
1695 error = deny_write_access(file);
1699 if (vm_flags & VM_SHARED) {
1700 error = mapping_map_writable(file->f_mapping);
1702 goto allow_write_and_free_vma;
1705 /* ->mmap() can change vma->vm_file, but must guarantee that
1706 * vma_link() below can deny write-access if VM_DENYWRITE is set
1707 * and map writably if VM_SHARED is set. This usually means the
1708 * new file must not have been exposed to user-space, yet.
1710 vma->vm_file = get_file(file);
1711 error = file->f_op->mmap(file, vma);
1713 goto unmap_and_free_vma;
1715 /* Can addr have changed??
1717 * Answer: Yes, several device drivers can do it in their
1718 * f_op->mmap method. -DaveM
1719 * Bug: If addr is changed, prev, rb_link, rb_parent should
1720 * be updated for vma_link()
1722 WARN_ON_ONCE(addr != vma->vm_start);
1724 addr = vma->vm_start;
1725 vm_flags = vma->vm_flags;
1726 } else if (vm_flags & VM_SHARED) {
1727 error = shmem_zero_setup(vma);
1732 vma_link(mm, vma, prev, rb_link, rb_parent);
1733 /* Once vma denies write, undo our temporary denial count */
1735 if (vm_flags & VM_SHARED)
1736 mapping_unmap_writable(file->f_mapping);
1737 if (vm_flags & VM_DENYWRITE)
1738 allow_write_access(file);
1740 file = vma->vm_file;
1742 perf_event_mmap(vma);
1744 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1745 if (vm_flags & VM_LOCKED) {
1746 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1747 vma == get_gate_vma(current->mm)))
1748 mm->locked_vm += (len >> PAGE_SHIFT);
1750 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1757 * New (or expanded) vma always get soft dirty status.
1758 * Otherwise user-space soft-dirty page tracker won't
1759 * be able to distinguish situation when vma area unmapped,
1760 * then new mapped in-place (which must be aimed as
1761 * a completely new data area).
1763 vma->vm_flags |= VM_SOFTDIRTY;
1765 vma_set_page_prot(vma);
1770 vma->vm_file = NULL;
1773 /* Undo any partial mapping done by a device driver. */
1774 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1776 if (vm_flags & VM_SHARED)
1777 mapping_unmap_writable(file->f_mapping);
1778 allow_write_and_free_vma:
1779 if (vm_flags & VM_DENYWRITE)
1780 allow_write_access(file);
1782 kmem_cache_free(vm_area_cachep, vma);
1785 vm_unacct_memory(charged);
1789 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1792 * We implement the search by looking for an rbtree node that
1793 * immediately follows a suitable gap. That is,
1794 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1795 * - gap_end = vma->vm_start >= info->low_limit + length;
1796 * - gap_end - gap_start >= length
1799 struct mm_struct *mm = current->mm;
1800 struct vm_area_struct *vma;
1801 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1803 /* Adjust search length to account for worst case alignment overhead */
1804 length = info->length + info->align_mask;
1805 if (length < info->length)
1808 /* Adjust search limits by the desired length */
1809 if (info->high_limit < length)
1811 high_limit = info->high_limit - length;
1813 if (info->low_limit > high_limit)
1815 low_limit = info->low_limit + length;
1817 /* Check if rbtree root looks promising */
1818 if (RB_EMPTY_ROOT(&mm->mm_rb))
1820 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1821 if (vma->rb_subtree_gap < length)
1825 /* Visit left subtree if it looks promising */
1826 gap_end = vm_start_gap(vma);
1827 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1828 struct vm_area_struct *left =
1829 rb_entry(vma->vm_rb.rb_left,
1830 struct vm_area_struct, vm_rb);
1831 if (left->rb_subtree_gap >= length) {
1837 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1839 /* Check if current node has a suitable gap */
1840 if (gap_start > high_limit)
1842 if (gap_end >= low_limit &&
1843 gap_end > gap_start && gap_end - gap_start >= length)
1846 /* Visit right subtree if it looks promising */
1847 if (vma->vm_rb.rb_right) {
1848 struct vm_area_struct *right =
1849 rb_entry(vma->vm_rb.rb_right,
1850 struct vm_area_struct, vm_rb);
1851 if (right->rb_subtree_gap >= length) {
1857 /* Go back up the rbtree to find next candidate node */
1859 struct rb_node *prev = &vma->vm_rb;
1860 if (!rb_parent(prev))
1862 vma = rb_entry(rb_parent(prev),
1863 struct vm_area_struct, vm_rb);
1864 if (prev == vma->vm_rb.rb_left) {
1865 gap_start = vm_end_gap(vma->vm_prev);
1866 gap_end = vm_start_gap(vma);
1873 /* Check highest gap, which does not precede any rbtree node */
1874 gap_start = mm->highest_vm_end;
1875 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1876 if (gap_start > high_limit)
1880 /* We found a suitable gap. Clip it with the original low_limit. */
1881 if (gap_start < info->low_limit)
1882 gap_start = info->low_limit;
1884 /* Adjust gap address to the desired alignment */
1885 gap_start += (info->align_offset - gap_start) & info->align_mask;
1887 VM_BUG_ON(gap_start + info->length > info->high_limit);
1888 VM_BUG_ON(gap_start + info->length > gap_end);
1892 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1894 struct mm_struct *mm = current->mm;
1895 struct vm_area_struct *vma;
1896 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1898 /* Adjust search length to account for worst case alignment overhead */
1899 length = info->length + info->align_mask;
1900 if (length < info->length)
1904 * Adjust search limits by the desired length.
1905 * See implementation comment at top of unmapped_area().
1907 gap_end = info->high_limit;
1908 if (gap_end < length)
1910 high_limit = gap_end - length;
1912 if (info->low_limit > high_limit)
1914 low_limit = info->low_limit + length;
1916 /* Check highest gap, which does not precede any rbtree node */
1917 gap_start = mm->highest_vm_end;
1918 if (gap_start <= high_limit)
1921 /* Check if rbtree root looks promising */
1922 if (RB_EMPTY_ROOT(&mm->mm_rb))
1924 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1925 if (vma->rb_subtree_gap < length)
1929 /* Visit right subtree if it looks promising */
1930 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1931 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1932 struct vm_area_struct *right =
1933 rb_entry(vma->vm_rb.rb_right,
1934 struct vm_area_struct, vm_rb);
1935 if (right->rb_subtree_gap >= length) {
1942 /* Check if current node has a suitable gap */
1943 gap_end = vm_start_gap(vma);
1944 if (gap_end < low_limit)
1946 if (gap_start <= high_limit &&
1947 gap_end > gap_start && gap_end - gap_start >= length)
1950 /* Visit left subtree if it looks promising */
1951 if (vma->vm_rb.rb_left) {
1952 struct vm_area_struct *left =
1953 rb_entry(vma->vm_rb.rb_left,
1954 struct vm_area_struct, vm_rb);
1955 if (left->rb_subtree_gap >= length) {
1961 /* Go back up the rbtree to find next candidate node */
1963 struct rb_node *prev = &vma->vm_rb;
1964 if (!rb_parent(prev))
1966 vma = rb_entry(rb_parent(prev),
1967 struct vm_area_struct, vm_rb);
1968 if (prev == vma->vm_rb.rb_right) {
1969 gap_start = vma->vm_prev ?
1970 vm_end_gap(vma->vm_prev) : 0;
1977 /* We found a suitable gap. Clip it with the original high_limit. */
1978 if (gap_end > info->high_limit)
1979 gap_end = info->high_limit;
1982 /* Compute highest gap address at the desired alignment */
1983 gap_end -= info->length;
1984 gap_end -= (gap_end - info->align_offset) & info->align_mask;
1986 VM_BUG_ON(gap_end < info->low_limit);
1987 VM_BUG_ON(gap_end < gap_start);
1991 /* Get an address range which is currently unmapped.
1992 * For shmat() with addr=0.
1994 * Ugly calling convention alert:
1995 * Return value with the low bits set means error value,
1997 * if (ret & ~PAGE_MASK)
2000 * This function "knows" that -ENOMEM has the bits set.
2002 #ifndef HAVE_ARCH_UNMAPPED_AREA
2004 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2005 unsigned long len, unsigned long pgoff, unsigned long flags)
2007 struct mm_struct *mm = current->mm;
2008 struct vm_area_struct *vma, *prev;
2009 struct vm_unmapped_area_info info;
2011 if (len > TASK_SIZE - mmap_min_addr)
2014 if (flags & MAP_FIXED)
2018 addr = PAGE_ALIGN(addr);
2019 vma = find_vma_prev(mm, addr, &prev);
2020 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2021 (!vma || addr + len <= vm_start_gap(vma)) &&
2022 (!prev || addr >= vm_end_gap(prev)))
2028 info.low_limit = mm->mmap_base;
2029 info.high_limit = TASK_SIZE;
2030 info.align_mask = 0;
2031 info.align_offset = 0;
2032 return vm_unmapped_area(&info);
2037 * This mmap-allocator allocates new areas top-down from below the
2038 * stack's low limit (the base):
2040 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2042 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2043 const unsigned long len, const unsigned long pgoff,
2044 const unsigned long flags)
2046 struct vm_area_struct *vma, *prev;
2047 struct mm_struct *mm = current->mm;
2048 unsigned long addr = addr0;
2049 struct vm_unmapped_area_info info;
2051 /* requested length too big for entire address space */
2052 if (len > TASK_SIZE - mmap_min_addr)
2055 if (flags & MAP_FIXED)
2058 /* requesting a specific address */
2060 addr = PAGE_ALIGN(addr);
2061 vma = find_vma_prev(mm, addr, &prev);
2062 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2063 (!vma || addr + len <= vm_start_gap(vma)) &&
2064 (!prev || addr >= vm_end_gap(prev)))
2068 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2070 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2071 info.high_limit = mm->mmap_base;
2072 info.align_mask = 0;
2073 info.align_offset = 0;
2074 addr = vm_unmapped_area(&info);
2077 * A failed mmap() very likely causes application failure,
2078 * so fall back to the bottom-up function here. This scenario
2079 * can happen with large stack limits and large mmap()
2082 if (offset_in_page(addr)) {
2083 VM_BUG_ON(addr != -ENOMEM);
2085 info.low_limit = TASK_UNMAPPED_BASE;
2086 info.high_limit = TASK_SIZE;
2087 addr = vm_unmapped_area(&info);
2095 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2096 unsigned long pgoff, unsigned long flags)
2098 unsigned long (*get_area)(struct file *, unsigned long,
2099 unsigned long, unsigned long, unsigned long);
2101 unsigned long error = arch_mmap_check(addr, len, flags);
2105 /* Careful about overflows.. */
2106 if (len > TASK_SIZE)
2109 get_area = current->mm->get_unmapped_area;
2111 if (file->f_op->get_unmapped_area)
2112 get_area = file->f_op->get_unmapped_area;
2113 } else if (flags & MAP_SHARED) {
2115 * mmap_region() will call shmem_zero_setup() to create a file,
2116 * so use shmem's get_unmapped_area in case it can be huge.
2117 * do_mmap_pgoff() will clear pgoff, so match alignment.
2120 get_area = shmem_get_unmapped_area;
2123 addr = get_area(file, addr, len, pgoff, flags);
2124 if (IS_ERR_VALUE(addr))
2127 if (addr > TASK_SIZE - len)
2129 if (offset_in_page(addr))
2132 error = security_mmap_addr(addr);
2133 return error ? error : addr;
2136 EXPORT_SYMBOL(get_unmapped_area);
2138 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2139 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2141 struct rb_node *rb_node;
2142 struct vm_area_struct *vma;
2144 /* Check the cache first. */
2145 vma = vmacache_find(mm, addr);
2149 rb_node = mm->mm_rb.rb_node;
2152 struct vm_area_struct *tmp;
2154 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2156 if (tmp->vm_end > addr) {
2158 if (tmp->vm_start <= addr)
2160 rb_node = rb_node->rb_left;
2162 rb_node = rb_node->rb_right;
2166 vmacache_update(addr, vma);
2170 EXPORT_SYMBOL(find_vma);
2173 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2175 struct vm_area_struct *
2176 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2177 struct vm_area_struct **pprev)
2179 struct vm_area_struct *vma;
2181 vma = find_vma(mm, addr);
2183 *pprev = vma->vm_prev;
2185 struct rb_node *rb_node = mm->mm_rb.rb_node;
2188 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2189 rb_node = rb_node->rb_right;
2196 * Verify that the stack growth is acceptable and
2197 * update accounting. This is shared with both the
2198 * grow-up and grow-down cases.
2200 static int acct_stack_growth(struct vm_area_struct *vma,
2201 unsigned long size, unsigned long grow)
2203 struct mm_struct *mm = vma->vm_mm;
2204 struct rlimit *rlim = current->signal->rlim;
2205 unsigned long new_start;
2207 /* address space limit tests */
2208 if (!may_expand_vm(mm, vma->vm_flags, grow))
2211 /* Stack limit test */
2212 if (size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2215 /* mlock limit tests */
2216 if (vma->vm_flags & VM_LOCKED) {
2217 unsigned long locked;
2218 unsigned long limit;
2219 locked = mm->locked_vm + grow;
2220 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2221 limit >>= PAGE_SHIFT;
2222 if (locked > limit && !capable(CAP_IPC_LOCK))
2226 /* Check to ensure the stack will not grow into a hugetlb-only region */
2227 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2229 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2233 * Overcommit.. This must be the final test, as it will
2234 * update security statistics.
2236 if (security_vm_enough_memory_mm(mm, grow))
2242 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2244 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2245 * vma is the last one with address > vma->vm_end. Have to extend vma.
2247 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2249 struct mm_struct *mm = vma->vm_mm;
2250 struct vm_area_struct *next;
2251 unsigned long gap_addr;
2254 if (!(vma->vm_flags & VM_GROWSUP))
2257 /* Guard against exceeding limits of the address space. */
2258 address &= PAGE_MASK;
2259 if (address >= (TASK_SIZE & PAGE_MASK))
2261 address += PAGE_SIZE;
2263 /* Enforce stack_guard_gap */
2264 gap_addr = address + stack_guard_gap;
2266 /* Guard against overflow */
2267 if (gap_addr < address || gap_addr > TASK_SIZE)
2268 gap_addr = TASK_SIZE;
2270 next = vma->vm_next;
2271 if (next && next->vm_start < gap_addr &&
2272 (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2273 if (!(next->vm_flags & VM_GROWSUP))
2275 /* Check that both stack segments have the same anon_vma? */
2278 /* We must make sure the anon_vma is allocated. */
2279 if (unlikely(anon_vma_prepare(vma)))
2283 * vma->vm_start/vm_end cannot change under us because the caller
2284 * is required to hold the mmap_sem in read mode. We need the
2285 * anon_vma lock to serialize against concurrent expand_stacks.
2287 anon_vma_lock_write(vma->anon_vma);
2289 /* Somebody else might have raced and expanded it already */
2290 if (address > vma->vm_end) {
2291 unsigned long size, grow;
2293 size = address - vma->vm_start;
2294 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2297 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2298 error = acct_stack_growth(vma, size, grow);
2301 * vma_gap_update() doesn't support concurrent
2302 * updates, but we only hold a shared mmap_sem
2303 * lock here, so we need to protect against
2304 * concurrent vma expansions.
2305 * anon_vma_lock_write() doesn't help here, as
2306 * we don't guarantee that all growable vmas
2307 * in a mm share the same root anon vma.
2308 * So, we reuse mm->page_table_lock to guard
2309 * against concurrent vma expansions.
2311 spin_lock(&mm->page_table_lock);
2312 if (vma->vm_flags & VM_LOCKED)
2313 mm->locked_vm += grow;
2314 vm_stat_account(mm, vma->vm_flags, grow);
2315 anon_vma_interval_tree_pre_update_vma(vma);
2316 vma->vm_end = address;
2317 anon_vma_interval_tree_post_update_vma(vma);
2319 vma_gap_update(vma->vm_next);
2321 mm->highest_vm_end = vm_end_gap(vma);
2322 spin_unlock(&mm->page_table_lock);
2324 perf_event_mmap(vma);
2328 anon_vma_unlock_write(vma->anon_vma);
2329 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2333 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2336 * vma is the first one with address < vma->vm_start. Have to extend vma.
2338 int expand_downwards(struct vm_area_struct *vma,
2339 unsigned long address)
2341 struct mm_struct *mm = vma->vm_mm;
2342 struct vm_area_struct *prev;
2343 unsigned long gap_addr;
2346 address &= PAGE_MASK;
2347 if (address < mmap_min_addr)
2350 /* Enforce stack_guard_gap */
2351 gap_addr = address - stack_guard_gap;
2352 if (gap_addr > address)
2354 prev = vma->vm_prev;
2355 if (prev && prev->vm_end > gap_addr &&
2356 (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2357 if (!(prev->vm_flags & VM_GROWSDOWN))
2359 /* Check that both stack segments have the same anon_vma? */
2362 /* We must make sure the anon_vma is allocated. */
2363 if (unlikely(anon_vma_prepare(vma)))
2367 * vma->vm_start/vm_end cannot change under us because the caller
2368 * is required to hold the mmap_sem in read mode. We need the
2369 * anon_vma lock to serialize against concurrent expand_stacks.
2371 anon_vma_lock_write(vma->anon_vma);
2373 /* Somebody else might have raced and expanded it already */
2374 if (address < vma->vm_start) {
2375 unsigned long size, grow;
2377 size = vma->vm_end - address;
2378 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2381 if (grow <= vma->vm_pgoff) {
2382 error = acct_stack_growth(vma, size, grow);
2385 * vma_gap_update() doesn't support concurrent
2386 * updates, but we only hold a shared mmap_sem
2387 * lock here, so we need to protect against
2388 * concurrent vma expansions.
2389 * anon_vma_lock_write() doesn't help here, as
2390 * we don't guarantee that all growable vmas
2391 * in a mm share the same root anon vma.
2392 * So, we reuse mm->page_table_lock to guard
2393 * against concurrent vma expansions.
2395 spin_lock(&mm->page_table_lock);
2396 if (vma->vm_flags & VM_LOCKED)
2397 mm->locked_vm += grow;
2398 vm_stat_account(mm, vma->vm_flags, grow);
2399 anon_vma_interval_tree_pre_update_vma(vma);
2400 vma->vm_start = address;
2401 vma->vm_pgoff -= grow;
2402 anon_vma_interval_tree_post_update_vma(vma);
2403 vma_gap_update(vma);
2404 spin_unlock(&mm->page_table_lock);
2406 perf_event_mmap(vma);
2410 anon_vma_unlock_write(vma->anon_vma);
2411 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2416 /* enforced gap between the expanding stack and other mappings. */
2417 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2419 static int __init cmdline_parse_stack_guard_gap(char *p)
2424 val = simple_strtoul(p, &endptr, 10);
2426 stack_guard_gap = val << PAGE_SHIFT;
2430 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2432 #ifdef CONFIG_STACK_GROWSUP
2433 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2435 return expand_upwards(vma, address);
2438 struct vm_area_struct *
2439 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2441 struct vm_area_struct *vma, *prev;
2444 vma = find_vma_prev(mm, addr, &prev);
2445 if (vma && (vma->vm_start <= addr))
2447 /* don't alter vm_end if the coredump is running */
2448 if (!prev || !mmget_still_valid(mm) || expand_stack(prev, addr))
2450 if (prev->vm_flags & VM_LOCKED)
2451 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2455 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2457 return expand_downwards(vma, address);
2460 struct vm_area_struct *
2461 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2463 struct vm_area_struct *vma;
2464 unsigned long start;
2467 vma = find_vma(mm, addr);
2470 if (vma->vm_start <= addr)
2472 if (!(vma->vm_flags & VM_GROWSDOWN))
2474 /* don't alter vm_start if the coredump is running */
2475 if (!mmget_still_valid(mm))
2477 start = vma->vm_start;
2478 if (expand_stack(vma, addr))
2480 if (vma->vm_flags & VM_LOCKED)
2481 populate_vma_page_range(vma, addr, start, NULL);
2486 EXPORT_SYMBOL_GPL(find_extend_vma);
2489 * Ok - we have the memory areas we should free on the vma list,
2490 * so release them, and do the vma updates.
2492 * Called with the mm semaphore held.
2494 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2496 unsigned long nr_accounted = 0;
2498 /* Update high watermark before we lower total_vm */
2499 update_hiwater_vm(mm);
2501 long nrpages = vma_pages(vma);
2503 if (vma->vm_flags & VM_ACCOUNT)
2504 nr_accounted += nrpages;
2505 vm_stat_account(mm, vma->vm_flags, -nrpages);
2506 vma = remove_vma(vma);
2508 vm_unacct_memory(nr_accounted);
2513 * Get rid of page table information in the indicated region.
2515 * Called with the mm semaphore held.
2517 static void unmap_region(struct mm_struct *mm,
2518 struct vm_area_struct *vma, struct vm_area_struct *prev,
2519 unsigned long start, unsigned long end)
2521 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2522 struct mmu_gather tlb;
2525 tlb_gather_mmu(&tlb, mm, start, end);
2526 update_hiwater_rss(mm);
2527 unmap_vmas(&tlb, vma, start, end);
2528 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2529 next ? next->vm_start : USER_PGTABLES_CEILING);
2530 tlb_finish_mmu(&tlb, start, end);
2534 * Create a list of vma's touched by the unmap, removing them from the mm's
2535 * vma list as we go..
2538 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2539 struct vm_area_struct *prev, unsigned long end)
2541 struct vm_area_struct **insertion_point;
2542 struct vm_area_struct *tail_vma = NULL;
2544 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2545 vma->vm_prev = NULL;
2547 vma_rb_erase(vma, &mm->mm_rb);
2551 } while (vma && vma->vm_start < end);
2552 *insertion_point = vma;
2554 vma->vm_prev = prev;
2555 vma_gap_update(vma);
2557 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2558 tail_vma->vm_next = NULL;
2560 /* Kill the cache */
2561 vmacache_invalidate(mm);
2565 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2566 * munmap path where it doesn't make sense to fail.
2568 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2569 unsigned long addr, int new_below)
2571 struct vm_area_struct *new;
2574 if (vma->vm_ops && vma->vm_ops->split) {
2575 err = vma->vm_ops->split(vma, addr);
2580 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2584 /* most fields are the same, copy all, and then fixup */
2587 INIT_LIST_HEAD(&new->anon_vma_chain);
2592 new->vm_start = addr;
2593 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2596 err = vma_dup_policy(vma, new);
2600 err = anon_vma_clone(new, vma);
2605 get_file(new->vm_file);
2607 if (new->vm_ops && new->vm_ops->open)
2608 new->vm_ops->open(new);
2611 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2612 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2614 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2620 /* Clean everything up if vma_adjust failed. */
2621 if (new->vm_ops && new->vm_ops->close)
2622 new->vm_ops->close(new);
2625 unlink_anon_vmas(new);
2627 mpol_put(vma_policy(new));
2629 kmem_cache_free(vm_area_cachep, new);
2634 * Split a vma into two pieces at address 'addr', a new vma is allocated
2635 * either for the first part or the tail.
2637 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2638 unsigned long addr, int new_below)
2640 if (mm->map_count >= sysctl_max_map_count)
2643 return __split_vma(mm, vma, addr, new_below);
2646 /* Munmap is split into 2 main parts -- this part which finds
2647 * what needs doing, and the areas themselves, which do the
2648 * work. This now handles partial unmappings.
2649 * Jeremy Fitzhardinge <jeremy@goop.org>
2651 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2654 struct vm_area_struct *vma, *prev, *last;
2656 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2659 len = PAGE_ALIGN(len);
2663 /* Find the first overlapping VMA */
2664 vma = find_vma(mm, start);
2667 prev = vma->vm_prev;
2668 /* we have start < vma->vm_end */
2670 /* if it doesn't overlap, we have nothing.. */
2672 if (vma->vm_start >= end)
2676 * If we need to split any vma, do it now to save pain later.
2678 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2679 * unmapped vm_area_struct will remain in use: so lower split_vma
2680 * places tmp vma above, and higher split_vma places tmp vma below.
2682 if (start > vma->vm_start) {
2686 * Make sure that map_count on return from munmap() will
2687 * not exceed its limit; but let map_count go just above
2688 * its limit temporarily, to help free resources as expected.
2690 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2693 error = __split_vma(mm, vma, start, 0);
2699 /* Does it split the last one? */
2700 last = find_vma(mm, end);
2701 if (last && end > last->vm_start) {
2702 int error = __split_vma(mm, last, end, 1);
2706 vma = prev ? prev->vm_next : mm->mmap;
2709 * unlock any mlock()ed ranges before detaching vmas
2711 if (mm->locked_vm) {
2712 struct vm_area_struct *tmp = vma;
2713 while (tmp && tmp->vm_start < end) {
2714 if (tmp->vm_flags & VM_LOCKED) {
2715 mm->locked_vm -= vma_pages(tmp);
2716 munlock_vma_pages_all(tmp);
2723 * Remove the vma's, and unmap the actual pages
2725 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2726 unmap_region(mm, vma, prev, start, end);
2728 arch_unmap(mm, vma, start, end);
2730 /* Fix up all other VM information */
2731 remove_vma_list(mm, vma);
2736 int vm_munmap(unsigned long start, size_t len)
2739 struct mm_struct *mm = current->mm;
2741 if (down_write_killable(&mm->mmap_sem))
2744 ret = do_munmap(mm, start, len);
2745 up_write(&mm->mmap_sem);
2748 EXPORT_SYMBOL(vm_munmap);
2750 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2753 struct mm_struct *mm = current->mm;
2755 profile_munmap(addr);
2756 if (down_write_killable(&mm->mmap_sem))
2758 ret = do_munmap(mm, addr, len);
2759 up_write(&mm->mmap_sem);
2765 * Emulation of deprecated remap_file_pages() syscall.
2767 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2768 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2771 struct mm_struct *mm = current->mm;
2772 struct vm_area_struct *vma;
2773 unsigned long populate = 0;
2774 unsigned long ret = -EINVAL;
2777 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2778 current->comm, current->pid);
2782 start = start & PAGE_MASK;
2783 size = size & PAGE_MASK;
2785 if (start + size <= start)
2788 /* Does pgoff wrap? */
2789 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2792 if (down_write_killable(&mm->mmap_sem))
2795 vma = find_vma(mm, start);
2797 if (!vma || !(vma->vm_flags & VM_SHARED))
2800 if (start < vma->vm_start)
2803 if (start + size > vma->vm_end) {
2804 struct vm_area_struct *next;
2806 for (next = vma->vm_next; next; next = next->vm_next) {
2807 /* hole between vmas ? */
2808 if (next->vm_start != next->vm_prev->vm_end)
2811 if (next->vm_file != vma->vm_file)
2814 if (next->vm_flags != vma->vm_flags)
2817 if (start + size <= next->vm_end)
2825 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2826 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2827 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2829 flags &= MAP_NONBLOCK;
2830 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2831 if (vma->vm_flags & VM_LOCKED) {
2832 struct vm_area_struct *tmp;
2833 flags |= MAP_LOCKED;
2835 /* drop PG_Mlocked flag for over-mapped range */
2836 for (tmp = vma; tmp->vm_start >= start + size;
2837 tmp = tmp->vm_next) {
2839 * Split pmd and munlock page on the border
2842 vma_adjust_trans_huge(tmp, start, start + size, 0);
2844 munlock_vma_pages_range(tmp,
2845 max(tmp->vm_start, start),
2846 min(tmp->vm_end, start + size));
2850 file = get_file(vma->vm_file);
2851 ret = do_mmap_pgoff(vma->vm_file, start, size,
2852 prot, flags, pgoff, &populate);
2855 up_write(&mm->mmap_sem);
2857 mm_populate(ret, populate);
2858 if (!IS_ERR_VALUE(ret))
2863 static inline void verify_mm_writelocked(struct mm_struct *mm)
2865 #ifdef CONFIG_DEBUG_VM
2866 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2868 up_read(&mm->mmap_sem);
2874 * this is really a simplified "do_mmap". it only handles
2875 * anonymous maps. eventually we may be able to do some
2876 * brk-specific accounting here.
2878 static int do_brk(unsigned long addr, unsigned long len)
2880 struct mm_struct *mm = current->mm;
2881 struct vm_area_struct *vma, *prev;
2882 unsigned long flags;
2883 struct rb_node **rb_link, *rb_parent;
2884 pgoff_t pgoff = addr >> PAGE_SHIFT;
2887 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2889 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2890 if (offset_in_page(error))
2893 error = mlock_future_check(mm, mm->def_flags, len);
2898 * mm->mmap_sem is required to protect against another thread
2899 * changing the mappings in case we sleep.
2901 verify_mm_writelocked(mm);
2904 * Clear old maps. this also does some error checking for us
2906 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2908 if (do_munmap(mm, addr, len))
2912 /* Check against address space limits *after* clearing old maps... */
2913 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2916 if (mm->map_count > sysctl_max_map_count)
2919 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2922 /* Can we just expand an old private anonymous mapping? */
2923 vma = vma_merge(mm, prev, addr, addr + len, flags,
2924 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2929 * create a vma struct for an anonymous mapping
2931 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2933 vm_unacct_memory(len >> PAGE_SHIFT);
2937 INIT_LIST_HEAD(&vma->anon_vma_chain);
2939 vma->vm_start = addr;
2940 vma->vm_end = addr + len;
2941 vma->vm_pgoff = pgoff;
2942 vma->vm_flags = flags;
2943 vma->vm_page_prot = vm_get_page_prot(flags);
2944 vma_link(mm, vma, prev, rb_link, rb_parent);
2946 perf_event_mmap(vma);
2947 mm->total_vm += len >> PAGE_SHIFT;
2948 mm->data_vm += len >> PAGE_SHIFT;
2949 if (flags & VM_LOCKED)
2950 mm->locked_vm += (len >> PAGE_SHIFT);
2951 vma->vm_flags |= VM_SOFTDIRTY;
2955 int vm_brk(unsigned long addr, unsigned long request)
2957 struct mm_struct *mm = current->mm;
2962 len = PAGE_ALIGN(request);
2968 if (down_write_killable(&mm->mmap_sem))
2971 ret = do_brk(addr, len);
2972 populate = ((mm->def_flags & VM_LOCKED) != 0);
2973 up_write(&mm->mmap_sem);
2974 if (populate && !ret)
2975 mm_populate(addr, len);
2978 EXPORT_SYMBOL(vm_brk);
2980 /* Release all mmaps. */
2981 void exit_mmap(struct mm_struct *mm)
2983 struct mmu_gather tlb;
2984 struct vm_area_struct *vma;
2985 unsigned long nr_accounted = 0;
2987 /* mm's last user has gone, and its about to be pulled down */
2988 mmu_notifier_release(mm);
2990 if (mm->locked_vm) {
2993 if (vma->vm_flags & VM_LOCKED)
2994 munlock_vma_pages_all(vma);
3002 if (!vma) /* Can happen if dup_mmap() received an OOM */
3007 tlb_gather_mmu(&tlb, mm, 0, -1);
3008 /* update_hiwater_rss(mm) here? but nobody should be looking */
3009 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3010 unmap_vmas(&tlb, vma, 0, -1);
3012 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3013 tlb_finish_mmu(&tlb, 0, -1);
3016 * Walk the list again, actually closing and freeing it,
3017 * with preemption enabled, without holding any MM locks.
3020 if (vma->vm_flags & VM_ACCOUNT)
3021 nr_accounted += vma_pages(vma);
3022 vma = remove_vma(vma);
3025 vm_unacct_memory(nr_accounted);
3028 /* Insert vm structure into process list sorted by address
3029 * and into the inode's i_mmap tree. If vm_file is non-NULL
3030 * then i_mmap_rwsem is taken here.
3032 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3034 struct vm_area_struct *prev;
3035 struct rb_node **rb_link, *rb_parent;
3037 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3038 &prev, &rb_link, &rb_parent))
3040 if ((vma->vm_flags & VM_ACCOUNT) &&
3041 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3045 * The vm_pgoff of a purely anonymous vma should be irrelevant
3046 * until its first write fault, when page's anon_vma and index
3047 * are set. But now set the vm_pgoff it will almost certainly
3048 * end up with (unless mremap moves it elsewhere before that
3049 * first wfault), so /proc/pid/maps tells a consistent story.
3051 * By setting it to reflect the virtual start address of the
3052 * vma, merges and splits can happen in a seamless way, just
3053 * using the existing file pgoff checks and manipulations.
3054 * Similarly in do_mmap_pgoff and in do_brk.
3056 if (vma_is_anonymous(vma)) {
3057 BUG_ON(vma->anon_vma);
3058 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3061 vma_link(mm, vma, prev, rb_link, rb_parent);
3066 * Copy the vma structure to a new location in the same mm,
3067 * prior to moving page table entries, to effect an mremap move.
3069 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3070 unsigned long addr, unsigned long len, pgoff_t pgoff,
3071 bool *need_rmap_locks)
3073 struct vm_area_struct *vma = *vmap;
3074 unsigned long vma_start = vma->vm_start;
3075 struct mm_struct *mm = vma->vm_mm;
3076 struct vm_area_struct *new_vma, *prev;
3077 struct rb_node **rb_link, *rb_parent;
3078 bool faulted_in_anon_vma = true;
3081 * If anonymous vma has not yet been faulted, update new pgoff
3082 * to match new location, to increase its chance of merging.
3084 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3085 pgoff = addr >> PAGE_SHIFT;
3086 faulted_in_anon_vma = false;
3089 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3090 return NULL; /* should never get here */
3091 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3092 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3093 vma->vm_userfaultfd_ctx);
3096 * Source vma may have been merged into new_vma
3098 if (unlikely(vma_start >= new_vma->vm_start &&
3099 vma_start < new_vma->vm_end)) {
3101 * The only way we can get a vma_merge with
3102 * self during an mremap is if the vma hasn't
3103 * been faulted in yet and we were allowed to
3104 * reset the dst vma->vm_pgoff to the
3105 * destination address of the mremap to allow
3106 * the merge to happen. mremap must change the
3107 * vm_pgoff linearity between src and dst vmas
3108 * (in turn preventing a vma_merge) to be
3109 * safe. It is only safe to keep the vm_pgoff
3110 * linear if there are no pages mapped yet.
3112 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3113 *vmap = vma = new_vma;
3115 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3117 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3121 new_vma->vm_start = addr;
3122 new_vma->vm_end = addr + len;
3123 new_vma->vm_pgoff = pgoff;
3124 if (vma_dup_policy(vma, new_vma))
3126 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3127 if (anon_vma_clone(new_vma, vma))
3128 goto out_free_mempol;
3129 if (new_vma->vm_file)
3130 get_file(new_vma->vm_file);
3131 if (new_vma->vm_ops && new_vma->vm_ops->open)
3132 new_vma->vm_ops->open(new_vma);
3133 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3134 *need_rmap_locks = false;
3139 mpol_put(vma_policy(new_vma));
3141 kmem_cache_free(vm_area_cachep, new_vma);
3147 * Return true if the calling process may expand its vm space by the passed
3150 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3152 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3155 if (is_data_mapping(flags) &&
3156 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3157 /* Workaround for Valgrind */
3158 if (rlimit(RLIMIT_DATA) == 0 &&
3159 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3161 if (!ignore_rlimit_data) {
3162 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits or use boot option ignore_rlimit_data.\n",
3163 current->comm, current->pid,
3164 (mm->data_vm + npages) << PAGE_SHIFT,
3165 rlimit(RLIMIT_DATA));
3173 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3175 mm->total_vm += npages;
3177 if (is_exec_mapping(flags))
3178 mm->exec_vm += npages;
3179 else if (is_stack_mapping(flags))
3180 mm->stack_vm += npages;
3181 else if (is_data_mapping(flags))
3182 mm->data_vm += npages;
3185 static int special_mapping_fault(struct vm_area_struct *vma,
3186 struct vm_fault *vmf);
3189 * Having a close hook prevents vma merging regardless of flags.
3191 static void special_mapping_close(struct vm_area_struct *vma)
3195 static const char *special_mapping_name(struct vm_area_struct *vma)
3197 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3200 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3202 struct vm_special_mapping *sm = new_vma->vm_private_data;
3205 return sm->mremap(sm, new_vma);
3209 static const struct vm_operations_struct special_mapping_vmops = {
3210 .close = special_mapping_close,
3211 .fault = special_mapping_fault,
3212 .mremap = special_mapping_mremap,
3213 .name = special_mapping_name,
3216 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3217 .close = special_mapping_close,
3218 .fault = special_mapping_fault,
3221 static int special_mapping_fault(struct vm_area_struct *vma,
3222 struct vm_fault *vmf)
3225 struct page **pages;
3227 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3228 pages = vma->vm_private_data;
3230 struct vm_special_mapping *sm = vma->vm_private_data;
3233 return sm->fault(sm, vma, vmf);
3238 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3242 struct page *page = *pages;
3248 return VM_FAULT_SIGBUS;
3251 static struct vm_area_struct *__install_special_mapping(
3252 struct mm_struct *mm,
3253 unsigned long addr, unsigned long len,
3254 unsigned long vm_flags, void *priv,
3255 const struct vm_operations_struct *ops)
3258 struct vm_area_struct *vma;
3260 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3261 if (unlikely(vma == NULL))
3262 return ERR_PTR(-ENOMEM);
3264 INIT_LIST_HEAD(&vma->anon_vma_chain);
3266 vma->vm_start = addr;
3267 vma->vm_end = addr + len;
3269 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3270 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3273 vma->vm_private_data = priv;
3275 ret = insert_vm_struct(mm, vma);
3279 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3281 perf_event_mmap(vma);
3286 kmem_cache_free(vm_area_cachep, vma);
3287 return ERR_PTR(ret);
3290 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3291 const struct vm_special_mapping *sm)
3293 return vma->vm_private_data == sm &&
3294 (vma->vm_ops == &special_mapping_vmops ||
3295 vma->vm_ops == &legacy_special_mapping_vmops);
3299 * Called with mm->mmap_sem held for writing.
3300 * Insert a new vma covering the given region, with the given flags.
3301 * Its pages are supplied by the given array of struct page *.
3302 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3303 * The region past the last page supplied will always produce SIGBUS.
3304 * The array pointer and the pages it points to are assumed to stay alive
3305 * for as long as this mapping might exist.
3307 struct vm_area_struct *_install_special_mapping(
3308 struct mm_struct *mm,
3309 unsigned long addr, unsigned long len,
3310 unsigned long vm_flags, const struct vm_special_mapping *spec)
3312 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3313 &special_mapping_vmops);
3316 int install_special_mapping(struct mm_struct *mm,
3317 unsigned long addr, unsigned long len,
3318 unsigned long vm_flags, struct page **pages)
3320 struct vm_area_struct *vma = __install_special_mapping(
3321 mm, addr, len, vm_flags, (void *)pages,
3322 &legacy_special_mapping_vmops);
3324 return PTR_ERR_OR_ZERO(vma);
3327 static DEFINE_MUTEX(mm_all_locks_mutex);
3329 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3331 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3333 * The LSB of head.next can't change from under us
3334 * because we hold the mm_all_locks_mutex.
3336 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3338 * We can safely modify head.next after taking the
3339 * anon_vma->root->rwsem. If some other vma in this mm shares
3340 * the same anon_vma we won't take it again.
3342 * No need of atomic instructions here, head.next
3343 * can't change from under us thanks to the
3344 * anon_vma->root->rwsem.
3346 if (__test_and_set_bit(0, (unsigned long *)
3347 &anon_vma->root->rb_root.rb_node))
3352 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3354 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3356 * AS_MM_ALL_LOCKS can't change from under us because
3357 * we hold the mm_all_locks_mutex.
3359 * Operations on ->flags have to be atomic because
3360 * even if AS_MM_ALL_LOCKS is stable thanks to the
3361 * mm_all_locks_mutex, there may be other cpus
3362 * changing other bitflags in parallel to us.
3364 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3366 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3371 * This operation locks against the VM for all pte/vma/mm related
3372 * operations that could ever happen on a certain mm. This includes
3373 * vmtruncate, try_to_unmap, and all page faults.
3375 * The caller must take the mmap_sem in write mode before calling
3376 * mm_take_all_locks(). The caller isn't allowed to release the
3377 * mmap_sem until mm_drop_all_locks() returns.
3379 * mmap_sem in write mode is required in order to block all operations
3380 * that could modify pagetables and free pages without need of
3381 * altering the vma layout. It's also needed in write mode to avoid new
3382 * anon_vmas to be associated with existing vmas.
3384 * A single task can't take more than one mm_take_all_locks() in a row
3385 * or it would deadlock.
3387 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3388 * mapping->flags avoid to take the same lock twice, if more than one
3389 * vma in this mm is backed by the same anon_vma or address_space.
3391 * We take locks in following order, accordingly to comment at beginning
3393 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3395 * - all i_mmap_rwsem locks;
3396 * - all anon_vma->rwseml
3398 * We can take all locks within these types randomly because the VM code
3399 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3400 * mm_all_locks_mutex.
3402 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3403 * that may have to take thousand of locks.
3405 * mm_take_all_locks() can fail if it's interrupted by signals.
3407 int mm_take_all_locks(struct mm_struct *mm)
3409 struct vm_area_struct *vma;
3410 struct anon_vma_chain *avc;
3412 BUG_ON(down_read_trylock(&mm->mmap_sem));
3414 mutex_lock(&mm_all_locks_mutex);
3416 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3417 if (signal_pending(current))
3419 if (vma->vm_file && vma->vm_file->f_mapping &&
3420 is_vm_hugetlb_page(vma))
3421 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3424 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3425 if (signal_pending(current))
3427 if (vma->vm_file && vma->vm_file->f_mapping &&
3428 !is_vm_hugetlb_page(vma))
3429 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3432 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3433 if (signal_pending(current))
3436 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3437 vm_lock_anon_vma(mm, avc->anon_vma);
3443 mm_drop_all_locks(mm);
3447 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3449 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3451 * The LSB of head.next can't change to 0 from under
3452 * us because we hold the mm_all_locks_mutex.
3454 * We must however clear the bitflag before unlocking
3455 * the vma so the users using the anon_vma->rb_root will
3456 * never see our bitflag.
3458 * No need of atomic instructions here, head.next
3459 * can't change from under us until we release the
3460 * anon_vma->root->rwsem.
3462 if (!__test_and_clear_bit(0, (unsigned long *)
3463 &anon_vma->root->rb_root.rb_node))
3465 anon_vma_unlock_write(anon_vma);
3469 static void vm_unlock_mapping(struct address_space *mapping)
3471 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3473 * AS_MM_ALL_LOCKS can't change to 0 from under us
3474 * because we hold the mm_all_locks_mutex.
3476 i_mmap_unlock_write(mapping);
3477 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3484 * The mmap_sem cannot be released by the caller until
3485 * mm_drop_all_locks() returns.
3487 void mm_drop_all_locks(struct mm_struct *mm)
3489 struct vm_area_struct *vma;
3490 struct anon_vma_chain *avc;
3492 BUG_ON(down_read_trylock(&mm->mmap_sem));
3493 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3495 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3497 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3498 vm_unlock_anon_vma(avc->anon_vma);
3499 if (vma->vm_file && vma->vm_file->f_mapping)
3500 vm_unlock_mapping(vma->vm_file->f_mapping);
3503 mutex_unlock(&mm_all_locks_mutex);
3507 * initialise the VMA slab
3509 void __init mmap_init(void)
3513 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3518 * Initialise sysctl_user_reserve_kbytes.
3520 * This is intended to prevent a user from starting a single memory hogging
3521 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3524 * The default value is min(3% of free memory, 128MB)
3525 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3527 static int init_user_reserve(void)
3529 unsigned long free_kbytes;
3531 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3533 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3536 subsys_initcall(init_user_reserve);
3539 * Initialise sysctl_admin_reserve_kbytes.
3541 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3542 * to log in and kill a memory hogging process.
3544 * Systems with more than 256MB will reserve 8MB, enough to recover
3545 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3546 * only reserve 3% of free pages by default.
3548 static int init_admin_reserve(void)
3550 unsigned long free_kbytes;
3552 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3554 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3557 subsys_initcall(init_admin_reserve);
3560 * Reinititalise user and admin reserves if memory is added or removed.
3562 * The default user reserve max is 128MB, and the default max for the
3563 * admin reserve is 8MB. These are usually, but not always, enough to
3564 * enable recovery from a memory hogging process using login/sshd, a shell,
3565 * and tools like top. It may make sense to increase or even disable the
3566 * reserve depending on the existence of swap or variations in the recovery
3567 * tools. So, the admin may have changed them.
3569 * If memory is added and the reserves have been eliminated or increased above
3570 * the default max, then we'll trust the admin.
3572 * If memory is removed and there isn't enough free memory, then we
3573 * need to reset the reserves.
3575 * Otherwise keep the reserve set by the admin.
3577 static int reserve_mem_notifier(struct notifier_block *nb,
3578 unsigned long action, void *data)
3580 unsigned long tmp, free_kbytes;
3584 /* Default max is 128MB. Leave alone if modified by operator. */
3585 tmp = sysctl_user_reserve_kbytes;
3586 if (0 < tmp && tmp < (1UL << 17))
3587 init_user_reserve();
3589 /* Default max is 8MB. Leave alone if modified by operator. */
3590 tmp = sysctl_admin_reserve_kbytes;
3591 if (0 < tmp && tmp < (1UL << 13))
3592 init_admin_reserve();
3596 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3598 if (sysctl_user_reserve_kbytes > free_kbytes) {
3599 init_user_reserve();
3600 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3601 sysctl_user_reserve_kbytes);
3604 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3605 init_admin_reserve();
3606 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3607 sysctl_admin_reserve_kbytes);
3616 static struct notifier_block reserve_mem_nb = {
3617 .notifier_call = reserve_mem_notifier,
3620 static int __meminit init_reserve_notifier(void)
3622 if (register_hotmemory_notifier(&reserve_mem_nb))
3623 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3627 subsys_initcall(init_reserve_notifier);