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>
49 #include <linux/uaccess.h>
50 #include <asm/cacheflush.h>
52 #include <asm/mmu_context.h>
56 #ifndef arch_mmap_check
57 #define arch_mmap_check(addr, len, flags) (0)
60 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
61 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
62 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
63 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
66 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
67 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
68 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
71 static bool ignore_rlimit_data;
72 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
74 static void unmap_region(struct mm_struct *mm,
75 struct vm_area_struct *vma, struct vm_area_struct *prev,
76 unsigned long start, unsigned long end);
78 /* description of effects of mapping type and prot in current implementation.
79 * this is due to the limited x86 page protection hardware. The expected
80 * behavior is in parens:
83 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
84 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
85 * w: (no) no w: (no) no w: (yes) yes w: (no) no
86 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
88 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
89 * w: (no) no w: (no) no w: (copy) copy w: (no) no
90 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
92 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
98 pgprot_t protection_map[16] __ro_after_init = {
99 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
100 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
103 pgprot_t vm_get_page_prot(unsigned long vm_flags)
105 return __pgprot(pgprot_val(protection_map[vm_flags &
106 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
107 pgprot_val(arch_vm_get_page_prot(vm_flags)));
109 EXPORT_SYMBOL(vm_get_page_prot);
111 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
113 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
116 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
117 void vma_set_page_prot(struct vm_area_struct *vma)
119 unsigned long vm_flags = vma->vm_flags;
120 pgprot_t vm_page_prot;
122 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
123 if (vma_wants_writenotify(vma, vm_page_prot)) {
124 vm_flags &= ~VM_SHARED;
125 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
127 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
128 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
132 * Requires inode->i_mapping->i_mmap_rwsem
134 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
135 struct file *file, struct address_space *mapping)
137 if (vma->vm_flags & VM_DENYWRITE)
138 atomic_inc(&file_inode(file)->i_writecount);
139 if (vma->vm_flags & VM_SHARED)
140 mapping_unmap_writable(mapping);
142 flush_dcache_mmap_lock(mapping);
143 vma_interval_tree_remove(vma, &mapping->i_mmap);
144 flush_dcache_mmap_unlock(mapping);
148 * Unlink a file-based vm structure from its interval tree, to hide
149 * vma from rmap and vmtruncate before freeing its page tables.
151 void unlink_file_vma(struct vm_area_struct *vma)
153 struct file *file = vma->vm_file;
156 struct address_space *mapping = file->f_mapping;
157 i_mmap_lock_write(mapping);
158 __remove_shared_vm_struct(vma, file, mapping);
159 i_mmap_unlock_write(mapping);
164 * Close a vm structure and free it, returning the next.
166 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
168 struct vm_area_struct *next = vma->vm_next;
171 if (vma->vm_ops && vma->vm_ops->close)
172 vma->vm_ops->close(vma);
175 mpol_put(vma_policy(vma));
176 kmem_cache_free(vm_area_cachep, vma);
180 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
181 struct list_head *uf);
182 SYSCALL_DEFINE1(brk, unsigned long, brk)
184 unsigned long retval;
185 unsigned long newbrk, oldbrk;
186 struct mm_struct *mm = current->mm;
187 struct vm_area_struct *next;
188 unsigned long min_brk;
192 if (down_write_killable(&mm->mmap_sem))
195 #ifdef CONFIG_COMPAT_BRK
197 * CONFIG_COMPAT_BRK can still be overridden by setting
198 * randomize_va_space to 2, which will still cause mm->start_brk
199 * to be arbitrarily shifted
201 if (current->brk_randomized)
202 min_brk = mm->start_brk;
204 min_brk = mm->end_data;
206 min_brk = mm->start_brk;
212 * Check against rlimit here. If this check is done later after the test
213 * of oldbrk with newbrk then it can escape the test and let the data
214 * segment grow beyond its set limit the in case where the limit is
215 * not page aligned -Ram Gupta
217 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
218 mm->end_data, mm->start_data))
221 newbrk = PAGE_ALIGN(brk);
222 oldbrk = PAGE_ALIGN(mm->brk);
223 if (oldbrk == newbrk)
226 /* Always allow shrinking brk. */
227 if (brk <= mm->brk) {
228 if (!do_munmap(mm, newbrk, oldbrk-newbrk, &uf))
233 /* Check against existing mmap mappings. */
234 next = find_vma(mm, oldbrk);
235 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
238 /* Ok, looks good - let it rip. */
239 if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
244 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
245 up_write(&mm->mmap_sem);
246 userfaultfd_unmap_complete(mm, &uf);
248 mm_populate(oldbrk, newbrk - oldbrk);
253 up_write(&mm->mmap_sem);
257 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
259 unsigned long max, prev_end, subtree_gap;
262 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
263 * allow two stack_guard_gaps between them here, and when choosing
264 * an unmapped area; whereas when expanding we only require one.
265 * That's a little inconsistent, but keeps the code here simpler.
267 max = vm_start_gap(vma);
269 prev_end = vm_end_gap(vma->vm_prev);
275 if (vma->vm_rb.rb_left) {
276 subtree_gap = rb_entry(vma->vm_rb.rb_left,
277 struct vm_area_struct, vm_rb)->rb_subtree_gap;
278 if (subtree_gap > max)
281 if (vma->vm_rb.rb_right) {
282 subtree_gap = rb_entry(vma->vm_rb.rb_right,
283 struct vm_area_struct, vm_rb)->rb_subtree_gap;
284 if (subtree_gap > max)
290 #ifdef CONFIG_DEBUG_VM_RB
291 static int browse_rb(struct mm_struct *mm)
293 struct rb_root *root = &mm->mm_rb;
294 int i = 0, j, bug = 0;
295 struct rb_node *nd, *pn = NULL;
296 unsigned long prev = 0, pend = 0;
298 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
299 struct vm_area_struct *vma;
300 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
301 if (vma->vm_start < prev) {
302 pr_emerg("vm_start %lx < prev %lx\n",
303 vma->vm_start, prev);
306 if (vma->vm_start < pend) {
307 pr_emerg("vm_start %lx < pend %lx\n",
308 vma->vm_start, pend);
311 if (vma->vm_start > vma->vm_end) {
312 pr_emerg("vm_start %lx > vm_end %lx\n",
313 vma->vm_start, vma->vm_end);
316 spin_lock(&mm->page_table_lock);
317 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
318 pr_emerg("free gap %lx, correct %lx\n",
320 vma_compute_subtree_gap(vma));
323 spin_unlock(&mm->page_table_lock);
326 prev = vma->vm_start;
330 for (nd = pn; nd; nd = rb_prev(nd))
333 pr_emerg("backwards %d, forwards %d\n", j, i);
339 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
343 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
344 struct vm_area_struct *vma;
345 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
346 VM_BUG_ON_VMA(vma != ignore &&
347 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
352 static void validate_mm(struct mm_struct *mm)
356 unsigned long highest_address = 0;
357 struct vm_area_struct *vma = mm->mmap;
360 struct anon_vma *anon_vma = vma->anon_vma;
361 struct anon_vma_chain *avc;
364 anon_vma_lock_read(anon_vma);
365 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
366 anon_vma_interval_tree_verify(avc);
367 anon_vma_unlock_read(anon_vma);
370 highest_address = vm_end_gap(vma);
374 if (i != mm->map_count) {
375 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
378 if (highest_address != mm->highest_vm_end) {
379 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
380 mm->highest_vm_end, highest_address);
384 if (i != mm->map_count) {
386 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
389 VM_BUG_ON_MM(bug, mm);
392 #define validate_mm_rb(root, ignore) do { } while (0)
393 #define validate_mm(mm) do { } while (0)
396 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
397 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
400 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
401 * vma->vm_prev->vm_end values changed, without modifying the vma's position
404 static void vma_gap_update(struct vm_area_struct *vma)
407 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
408 * function that does exacltly what we want.
410 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
413 static inline void vma_rb_insert(struct vm_area_struct *vma,
414 struct rb_root *root)
416 /* All rb_subtree_gap values must be consistent prior to insertion */
417 validate_mm_rb(root, NULL);
419 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
422 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
425 * Note rb_erase_augmented is a fairly large inline function,
426 * so make sure we instantiate it only once with our desired
427 * augmented rbtree callbacks.
429 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
432 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
433 struct rb_root *root,
434 struct vm_area_struct *ignore)
437 * All rb_subtree_gap values must be consistent prior to erase,
438 * with the possible exception of the "next" vma being erased if
439 * next->vm_start was reduced.
441 validate_mm_rb(root, ignore);
443 __vma_rb_erase(vma, root);
446 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
447 struct rb_root *root)
450 * All rb_subtree_gap values must be consistent prior to erase,
451 * with the possible exception of the vma being erased.
453 validate_mm_rb(root, vma);
455 __vma_rb_erase(vma, root);
459 * vma has some anon_vma assigned, and is already inserted on that
460 * anon_vma's interval trees.
462 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
463 * vma must be removed from the anon_vma's interval trees using
464 * anon_vma_interval_tree_pre_update_vma().
466 * After the update, the vma will be reinserted using
467 * anon_vma_interval_tree_post_update_vma().
469 * The entire update must be protected by exclusive mmap_sem and by
470 * the root anon_vma's mutex.
473 anon_vma_interval_tree_pre_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_remove(avc, &avc->anon_vma->rb_root);
482 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
484 struct anon_vma_chain *avc;
486 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
487 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
490 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
491 unsigned long end, struct vm_area_struct **pprev,
492 struct rb_node ***rb_link, struct rb_node **rb_parent)
494 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
496 __rb_link = &mm->mm_rb.rb_node;
497 rb_prev = __rb_parent = NULL;
500 struct vm_area_struct *vma_tmp;
502 __rb_parent = *__rb_link;
503 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
505 if (vma_tmp->vm_end > addr) {
506 /* Fail if an existing vma overlaps the area */
507 if (vma_tmp->vm_start < end)
509 __rb_link = &__rb_parent->rb_left;
511 rb_prev = __rb_parent;
512 __rb_link = &__rb_parent->rb_right;
518 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
519 *rb_link = __rb_link;
520 *rb_parent = __rb_parent;
524 static unsigned long count_vma_pages_range(struct mm_struct *mm,
525 unsigned long addr, unsigned long end)
527 unsigned long nr_pages = 0;
528 struct vm_area_struct *vma;
530 /* Find first overlaping mapping */
531 vma = find_vma_intersection(mm, addr, end);
535 nr_pages = (min(end, vma->vm_end) -
536 max(addr, vma->vm_start)) >> PAGE_SHIFT;
538 /* Iterate over the rest of the overlaps */
539 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
540 unsigned long overlap_len;
542 if (vma->vm_start > end)
545 overlap_len = min(end, vma->vm_end) - vma->vm_start;
546 nr_pages += overlap_len >> PAGE_SHIFT;
552 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
553 struct rb_node **rb_link, struct rb_node *rb_parent)
555 /* Update tracking information for the gap following the new vma. */
557 vma_gap_update(vma->vm_next);
559 mm->highest_vm_end = vm_end_gap(vma);
562 * vma->vm_prev wasn't known when we followed the rbtree to find the
563 * correct insertion point for that vma. As a result, we could not
564 * update the vma vm_rb parents rb_subtree_gap values on the way down.
565 * So, we first insert the vma with a zero rb_subtree_gap value
566 * (to be consistent with what we did on the way down), and then
567 * immediately update the gap to the correct value. Finally we
568 * rebalance the rbtree after all augmented values have been set.
570 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
571 vma->rb_subtree_gap = 0;
573 vma_rb_insert(vma, &mm->mm_rb);
576 static void __vma_link_file(struct vm_area_struct *vma)
582 struct address_space *mapping = file->f_mapping;
584 if (vma->vm_flags & VM_DENYWRITE)
585 atomic_dec(&file_inode(file)->i_writecount);
586 if (vma->vm_flags & VM_SHARED)
587 atomic_inc(&mapping->i_mmap_writable);
589 flush_dcache_mmap_lock(mapping);
590 vma_interval_tree_insert(vma, &mapping->i_mmap);
591 flush_dcache_mmap_unlock(mapping);
596 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
597 struct vm_area_struct *prev, struct rb_node **rb_link,
598 struct rb_node *rb_parent)
600 __vma_link_list(mm, vma, prev, rb_parent);
601 __vma_link_rb(mm, vma, rb_link, rb_parent);
604 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
605 struct vm_area_struct *prev, struct rb_node **rb_link,
606 struct rb_node *rb_parent)
608 struct address_space *mapping = NULL;
611 mapping = vma->vm_file->f_mapping;
612 i_mmap_lock_write(mapping);
615 __vma_link(mm, vma, prev, rb_link, rb_parent);
616 __vma_link_file(vma);
619 i_mmap_unlock_write(mapping);
626 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
627 * mm's list and rbtree. It has already been inserted into the interval tree.
629 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
631 struct vm_area_struct *prev;
632 struct rb_node **rb_link, *rb_parent;
634 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
635 &prev, &rb_link, &rb_parent))
637 __vma_link(mm, vma, prev, rb_link, rb_parent);
641 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
642 struct vm_area_struct *vma,
643 struct vm_area_struct *prev,
645 struct vm_area_struct *ignore)
647 struct vm_area_struct *next;
649 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
652 prev->vm_next = next;
656 prev->vm_next = next;
661 next->vm_prev = prev;
664 vmacache_invalidate(mm);
667 static inline void __vma_unlink_prev(struct mm_struct *mm,
668 struct vm_area_struct *vma,
669 struct vm_area_struct *prev)
671 __vma_unlink_common(mm, vma, prev, true, vma);
675 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
676 * is already present in an i_mmap tree without adjusting the tree.
677 * The following helper function should be used when such adjustments
678 * are necessary. The "insert" vma (if any) is to be inserted
679 * before we drop the necessary locks.
681 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
682 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
683 struct vm_area_struct *expand)
685 struct mm_struct *mm = vma->vm_mm;
686 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
687 struct address_space *mapping = NULL;
688 struct rb_root_cached *root = NULL;
689 struct anon_vma *anon_vma = NULL;
690 struct file *file = vma->vm_file;
691 bool start_changed = false, end_changed = false;
692 long adjust_next = 0;
695 if (next && !insert) {
696 struct vm_area_struct *exporter = NULL, *importer = NULL;
698 if (end >= next->vm_end) {
700 * vma expands, overlapping all the next, and
701 * perhaps the one after too (mprotect case 6).
702 * The only other cases that gets here are
703 * case 1, case 7 and case 8.
705 if (next == expand) {
707 * The only case where we don't expand "vma"
708 * and we expand "next" instead is case 8.
710 VM_WARN_ON(end != next->vm_end);
712 * remove_next == 3 means we're
713 * removing "vma" and that to do so we
714 * swapped "vma" and "next".
717 VM_WARN_ON(file != next->vm_file);
720 VM_WARN_ON(expand != vma);
722 * case 1, 6, 7, remove_next == 2 is case 6,
723 * remove_next == 1 is case 1 or 7.
725 remove_next = 1 + (end > next->vm_end);
726 VM_WARN_ON(remove_next == 2 &&
727 end != next->vm_next->vm_end);
728 VM_WARN_ON(remove_next == 1 &&
729 end != next->vm_end);
730 /* trim end to next, for case 6 first pass */
738 * If next doesn't have anon_vma, import from vma after
739 * next, if the vma overlaps with it.
741 if (remove_next == 2 && !next->anon_vma)
742 exporter = next->vm_next;
744 } else if (end > next->vm_start) {
746 * vma expands, overlapping part of the next:
747 * mprotect case 5 shifting the boundary up.
749 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
752 VM_WARN_ON(expand != importer);
753 } else if (end < vma->vm_end) {
755 * vma shrinks, and !insert tells it's not
756 * split_vma inserting another: so it must be
757 * mprotect case 4 shifting the boundary down.
759 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
762 VM_WARN_ON(expand != importer);
766 * Easily overlooked: when mprotect shifts the boundary,
767 * make sure the expanding vma has anon_vma set if the
768 * shrinking vma had, to cover any anon pages imported.
770 if (exporter && exporter->anon_vma && !importer->anon_vma) {
773 importer->anon_vma = exporter->anon_vma;
774 error = anon_vma_clone(importer, exporter);
780 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
783 mapping = file->f_mapping;
784 root = &mapping->i_mmap;
785 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
788 uprobe_munmap(next, next->vm_start, next->vm_end);
790 i_mmap_lock_write(mapping);
793 * Put into interval tree now, so instantiated pages
794 * are visible to arm/parisc __flush_dcache_page
795 * throughout; but we cannot insert into address
796 * space until vma start or end is updated.
798 __vma_link_file(insert);
802 anon_vma = vma->anon_vma;
803 if (!anon_vma && adjust_next)
804 anon_vma = next->anon_vma;
806 VM_WARN_ON(adjust_next && next->anon_vma &&
807 anon_vma != next->anon_vma);
808 anon_vma_lock_write(anon_vma);
809 anon_vma_interval_tree_pre_update_vma(vma);
811 anon_vma_interval_tree_pre_update_vma(next);
815 flush_dcache_mmap_lock(mapping);
816 vma_interval_tree_remove(vma, root);
818 vma_interval_tree_remove(next, root);
821 if (start != vma->vm_start) {
822 vma->vm_start = start;
823 start_changed = true;
825 if (end != vma->vm_end) {
829 vma->vm_pgoff = pgoff;
831 next->vm_start += adjust_next << PAGE_SHIFT;
832 next->vm_pgoff += adjust_next;
837 vma_interval_tree_insert(next, root);
838 vma_interval_tree_insert(vma, root);
839 flush_dcache_mmap_unlock(mapping);
844 * vma_merge has merged next into vma, and needs
845 * us to remove next before dropping the locks.
847 if (remove_next != 3)
848 __vma_unlink_prev(mm, next, vma);
851 * vma is not before next if they've been
854 * pre-swap() next->vm_start was reduced so
855 * tell validate_mm_rb to ignore pre-swap()
856 * "next" (which is stored in post-swap()
859 __vma_unlink_common(mm, next, NULL, false, vma);
861 __remove_shared_vm_struct(next, file, mapping);
864 * split_vma has split insert from vma, and needs
865 * us to insert it before dropping the locks
866 * (it may either follow vma or precede it).
868 __insert_vm_struct(mm, insert);
874 mm->highest_vm_end = vm_end_gap(vma);
875 else if (!adjust_next)
876 vma_gap_update(next);
881 anon_vma_interval_tree_post_update_vma(vma);
883 anon_vma_interval_tree_post_update_vma(next);
884 anon_vma_unlock_write(anon_vma);
887 i_mmap_unlock_write(mapping);
898 uprobe_munmap(next, next->vm_start, next->vm_end);
902 anon_vma_merge(vma, next);
904 mpol_put(vma_policy(next));
905 kmem_cache_free(vm_area_cachep, next);
907 * In mprotect's case 6 (see comments on vma_merge),
908 * we must remove another next too. It would clutter
909 * up the code too much to do both in one go.
911 if (remove_next != 3) {
913 * If "next" was removed and vma->vm_end was
914 * expanded (up) over it, in turn
915 * "next->vm_prev->vm_end" changed and the
916 * "vma->vm_next" gap must be updated.
921 * For the scope of the comment "next" and
922 * "vma" considered pre-swap(): if "vma" was
923 * removed, next->vm_start was expanded (down)
924 * over it and the "next" gap must be updated.
925 * Because of the swap() the post-swap() "vma"
926 * actually points to pre-swap() "next"
927 * (post-swap() "next" as opposed is now a
932 if (remove_next == 2) {
938 vma_gap_update(next);
941 * If remove_next == 2 we obviously can't
944 * If remove_next == 3 we can't reach this
945 * path because pre-swap() next is always not
946 * NULL. pre-swap() "next" is not being
947 * removed and its next->vm_end is not altered
948 * (and furthermore "end" already matches
949 * next->vm_end in remove_next == 3).
951 * We reach this only in the remove_next == 1
952 * case if the "next" vma that was removed was
953 * the highest vma of the mm. However in such
954 * case next->vm_end == "end" and the extended
955 * "vma" has vma->vm_end == next->vm_end so
956 * mm->highest_vm_end doesn't need any update
957 * in remove_next == 1 case.
959 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
971 * If the vma has a ->close operation then the driver probably needs to release
972 * per-vma resources, so we don't attempt to merge those.
974 static inline int is_mergeable_vma(struct vm_area_struct *vma,
975 struct file *file, unsigned long vm_flags,
976 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
979 * VM_SOFTDIRTY should not prevent from VMA merging, if we
980 * match the flags but dirty bit -- the caller should mark
981 * merged VMA as dirty. If dirty bit won't be excluded from
982 * comparison, we increase pressue on the memory system forcing
983 * the kernel to generate new VMAs when old one could be
986 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
988 if (vma->vm_file != file)
990 if (vma->vm_ops && vma->vm_ops->close)
992 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
997 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
998 struct anon_vma *anon_vma2,
999 struct vm_area_struct *vma)
1002 * The list_is_singular() test is to avoid merging VMA cloned from
1003 * parents. This can improve scalability caused by anon_vma lock.
1005 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1006 list_is_singular(&vma->anon_vma_chain)))
1008 return anon_vma1 == anon_vma2;
1012 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1013 * in front of (at a lower virtual address and file offset than) the vma.
1015 * We cannot merge two vmas if they have differently assigned (non-NULL)
1016 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1018 * We don't check here for the merged mmap wrapping around the end of pagecache
1019 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1020 * wrap, nor mmaps which cover the final page at index -1UL.
1023 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1024 struct anon_vma *anon_vma, struct file *file,
1026 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1028 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1029 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1030 if (vma->vm_pgoff == vm_pgoff)
1037 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1038 * beyond (at a higher virtual address and file offset than) the vma.
1040 * We cannot merge two vmas if they have differently assigned (non-NULL)
1041 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1044 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1045 struct anon_vma *anon_vma, struct file *file,
1047 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1049 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1050 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1052 vm_pglen = vma_pages(vma);
1053 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1060 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1061 * whether that can be merged with its predecessor or its successor.
1062 * Or both (it neatly fills a hole).
1064 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1065 * certain not to be mapped by the time vma_merge is called; but when
1066 * called for mprotect, it is certain to be already mapped (either at
1067 * an offset within prev, or at the start of next), and the flags of
1068 * this area are about to be changed to vm_flags - and the no-change
1069 * case has already been eliminated.
1071 * The following mprotect cases have to be considered, where AAAA is
1072 * the area passed down from mprotect_fixup, never extending beyond one
1073 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1075 * AAAA AAAA AAAA AAAA
1076 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1077 * cannot merge might become might become might become
1078 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1079 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1080 * mremap move: PPPPXXXXXXXX 8
1082 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1083 * might become case 1 below case 2 below case 3 below
1085 * It is important for case 8 that the the vma NNNN overlapping the
1086 * region AAAA is never going to extended over XXXX. Instead XXXX must
1087 * be extended in region AAAA and NNNN must be removed. This way in
1088 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1089 * rmap_locks, the properties of the merged vma will be already
1090 * correct for the whole merged range. Some of those properties like
1091 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1092 * be correct for the whole merged range immediately after the
1093 * rmap_locks are released. Otherwise if XXXX would be removed and
1094 * NNNN would be extended over the XXXX range, remove_migration_ptes
1095 * or other rmap walkers (if working on addresses beyond the "end"
1096 * parameter) may establish ptes with the wrong permissions of NNNN
1097 * instead of the right permissions of XXXX.
1099 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1100 struct vm_area_struct *prev, unsigned long addr,
1101 unsigned long end, unsigned long vm_flags,
1102 struct anon_vma *anon_vma, struct file *file,
1103 pgoff_t pgoff, struct mempolicy *policy,
1104 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1106 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1107 struct vm_area_struct *area, *next;
1111 * We later require that vma->vm_flags == vm_flags,
1112 * so this tests vma->vm_flags & VM_SPECIAL, too.
1114 if (vm_flags & VM_SPECIAL)
1118 next = prev->vm_next;
1122 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1123 next = next->vm_next;
1125 /* verify some invariant that must be enforced by the caller */
1126 VM_WARN_ON(prev && addr <= prev->vm_start);
1127 VM_WARN_ON(area && end > area->vm_end);
1128 VM_WARN_ON(addr >= end);
1131 * Can it merge with the predecessor?
1133 if (prev && prev->vm_end == addr &&
1134 mpol_equal(vma_policy(prev), policy) &&
1135 can_vma_merge_after(prev, vm_flags,
1136 anon_vma, file, pgoff,
1137 vm_userfaultfd_ctx)) {
1139 * OK, it can. Can we now merge in the successor as well?
1141 if (next && end == next->vm_start &&
1142 mpol_equal(policy, vma_policy(next)) &&
1143 can_vma_merge_before(next, vm_flags,
1146 vm_userfaultfd_ctx) &&
1147 is_mergeable_anon_vma(prev->anon_vma,
1148 next->anon_vma, NULL)) {
1150 err = __vma_adjust(prev, prev->vm_start,
1151 next->vm_end, prev->vm_pgoff, NULL,
1153 } else /* cases 2, 5, 7 */
1154 err = __vma_adjust(prev, prev->vm_start,
1155 end, prev->vm_pgoff, NULL, prev);
1158 khugepaged_enter_vma_merge(prev, vm_flags);
1163 * Can this new request be merged in front of next?
1165 if (next && end == next->vm_start &&
1166 mpol_equal(policy, vma_policy(next)) &&
1167 can_vma_merge_before(next, vm_flags,
1168 anon_vma, file, pgoff+pglen,
1169 vm_userfaultfd_ctx)) {
1170 if (prev && addr < prev->vm_end) /* case 4 */
1171 err = __vma_adjust(prev, prev->vm_start,
1172 addr, prev->vm_pgoff, NULL, next);
1173 else { /* cases 3, 8 */
1174 err = __vma_adjust(area, addr, next->vm_end,
1175 next->vm_pgoff - pglen, NULL, next);
1177 * In case 3 area is already equal to next and
1178 * this is a noop, but in case 8 "area" has
1179 * been removed and next was expanded over it.
1185 khugepaged_enter_vma_merge(area, vm_flags);
1193 * Rough compatbility check to quickly see if it's even worth looking
1194 * at sharing an anon_vma.
1196 * They need to have the same vm_file, and the flags can only differ
1197 * in things that mprotect may change.
1199 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1200 * we can merge the two vma's. For example, we refuse to merge a vma if
1201 * there is a vm_ops->close() function, because that indicates that the
1202 * driver is doing some kind of reference counting. But that doesn't
1203 * really matter for the anon_vma sharing case.
1205 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1207 return a->vm_end == b->vm_start &&
1208 mpol_equal(vma_policy(a), vma_policy(b)) &&
1209 a->vm_file == b->vm_file &&
1210 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1211 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1215 * Do some basic sanity checking to see if we can re-use the anon_vma
1216 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1217 * the same as 'old', the other will be the new one that is trying
1218 * to share the anon_vma.
1220 * NOTE! This runs with mm_sem held for reading, so it is possible that
1221 * the anon_vma of 'old' is concurrently in the process of being set up
1222 * by another page fault trying to merge _that_. But that's ok: if it
1223 * is being set up, that automatically means that it will be a singleton
1224 * acceptable for merging, so we can do all of this optimistically. But
1225 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1227 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1228 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1229 * is to return an anon_vma that is "complex" due to having gone through
1232 * We also make sure that the two vma's are compatible (adjacent,
1233 * and with the same memory policies). That's all stable, even with just
1234 * a read lock on the mm_sem.
1236 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1238 if (anon_vma_compatible(a, b)) {
1239 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1241 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1248 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1249 * neighbouring vmas for a suitable anon_vma, before it goes off
1250 * to allocate a new anon_vma. It checks because a repetitive
1251 * sequence of mprotects and faults may otherwise lead to distinct
1252 * anon_vmas being allocated, preventing vma merge in subsequent
1255 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1257 struct anon_vma *anon_vma;
1258 struct vm_area_struct *near;
1260 near = vma->vm_next;
1264 anon_vma = reusable_anon_vma(near, vma, near);
1268 near = vma->vm_prev;
1272 anon_vma = reusable_anon_vma(near, near, vma);
1277 * There's no absolute need to look only at touching neighbours:
1278 * we could search further afield for "compatible" anon_vmas.
1279 * But it would probably just be a waste of time searching,
1280 * or lead to too many vmas hanging off the same anon_vma.
1281 * We're trying to allow mprotect remerging later on,
1282 * not trying to minimize memory used for anon_vmas.
1288 * If a hint addr is less than mmap_min_addr change hint to be as
1289 * low as possible but still greater than mmap_min_addr
1291 static inline unsigned long round_hint_to_min(unsigned long hint)
1294 if (((void *)hint != NULL) &&
1295 (hint < mmap_min_addr))
1296 return PAGE_ALIGN(mmap_min_addr);
1300 static inline int mlock_future_check(struct mm_struct *mm,
1301 unsigned long flags,
1304 unsigned long locked, lock_limit;
1306 /* mlock MCL_FUTURE? */
1307 if (flags & VM_LOCKED) {
1308 locked = len >> PAGE_SHIFT;
1309 locked += mm->locked_vm;
1310 lock_limit = rlimit(RLIMIT_MEMLOCK);
1311 lock_limit >>= PAGE_SHIFT;
1312 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1318 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1320 if (S_ISREG(inode->i_mode))
1321 return MAX_LFS_FILESIZE;
1323 if (S_ISBLK(inode->i_mode))
1324 return MAX_LFS_FILESIZE;
1326 /* Special "we do even unsigned file positions" case */
1327 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1330 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1334 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1335 unsigned long pgoff, unsigned long len)
1337 u64 maxsize = file_mmap_size_max(file, inode);
1339 if (maxsize && len > maxsize)
1342 if (pgoff > maxsize >> PAGE_SHIFT)
1348 * The caller must hold down_write(¤t->mm->mmap_sem).
1350 unsigned long do_mmap(struct file *file, unsigned long addr,
1351 unsigned long len, unsigned long prot,
1352 unsigned long flags, vm_flags_t vm_flags,
1353 unsigned long pgoff, unsigned long *populate,
1354 struct list_head *uf)
1356 struct mm_struct *mm = current->mm;
1365 * Does the application expect PROT_READ to imply PROT_EXEC?
1367 * (the exception is when the underlying filesystem is noexec
1368 * mounted, in which case we dont add PROT_EXEC.)
1370 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1371 if (!(file && path_noexec(&file->f_path)))
1374 if (!(flags & MAP_FIXED))
1375 addr = round_hint_to_min(addr);
1377 /* Careful about overflows.. */
1378 len = PAGE_ALIGN(len);
1382 /* offset overflow? */
1383 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1386 /* Too many mappings? */
1387 if (mm->map_count > sysctl_max_map_count)
1390 /* Obtain the address to map to. we verify (or select) it and ensure
1391 * that it represents a valid section of the address space.
1393 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1394 if (offset_in_page(addr))
1397 if (prot == PROT_EXEC) {
1398 pkey = execute_only_pkey(mm);
1403 /* Do simple checking here so the lower-level routines won't have
1404 * to. we assume access permissions have been handled by the open
1405 * of the memory object, so we don't do any here.
1407 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1408 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1410 if (flags & MAP_LOCKED)
1411 if (!can_do_mlock())
1414 if (mlock_future_check(mm, vm_flags, len))
1418 struct inode *inode = file_inode(file);
1420 if (!file_mmap_ok(file, inode, pgoff, len))
1423 switch (flags & MAP_TYPE) {
1425 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1429 * Make sure we don't allow writing to an append-only
1432 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1436 * Make sure there are no mandatory locks on the file.
1438 if (locks_verify_locked(file))
1441 vm_flags |= VM_SHARED | VM_MAYSHARE;
1442 if (!(file->f_mode & FMODE_WRITE))
1443 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1447 if (!(file->f_mode & FMODE_READ))
1449 if (path_noexec(&file->f_path)) {
1450 if (vm_flags & VM_EXEC)
1452 vm_flags &= ~VM_MAYEXEC;
1455 if (!file->f_op->mmap)
1457 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1465 switch (flags & MAP_TYPE) {
1467 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1473 vm_flags |= VM_SHARED | VM_MAYSHARE;
1477 * Set pgoff according to addr for anon_vma.
1479 pgoff = addr >> PAGE_SHIFT;
1487 * Set 'VM_NORESERVE' if we should not account for the
1488 * memory use of this mapping.
1490 if (flags & MAP_NORESERVE) {
1491 /* We honor MAP_NORESERVE if allowed to overcommit */
1492 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1493 vm_flags |= VM_NORESERVE;
1495 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1496 if (file && is_file_hugepages(file))
1497 vm_flags |= VM_NORESERVE;
1500 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1501 if (!IS_ERR_VALUE(addr) &&
1502 ((vm_flags & VM_LOCKED) ||
1503 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1508 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1509 unsigned long, prot, unsigned long, flags,
1510 unsigned long, fd, unsigned long, pgoff)
1512 struct file *file = NULL;
1513 unsigned long retval;
1515 if (!(flags & MAP_ANONYMOUS)) {
1516 audit_mmap_fd(fd, flags);
1520 if (is_file_hugepages(file))
1521 len = ALIGN(len, huge_page_size(hstate_file(file)));
1523 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1525 } else if (flags & MAP_HUGETLB) {
1526 struct user_struct *user = NULL;
1529 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1533 len = ALIGN(len, huge_page_size(hs));
1535 * VM_NORESERVE is used because the reservations will be
1536 * taken when vm_ops->mmap() is called
1537 * A dummy user value is used because we are not locking
1538 * memory so no accounting is necessary
1540 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1542 &user, HUGETLB_ANONHUGE_INODE,
1543 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1545 return PTR_ERR(file);
1548 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1550 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1557 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1558 struct mmap_arg_struct {
1562 unsigned long flags;
1564 unsigned long offset;
1567 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1569 struct mmap_arg_struct a;
1571 if (copy_from_user(&a, arg, sizeof(a)))
1573 if (offset_in_page(a.offset))
1576 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1577 a.offset >> PAGE_SHIFT);
1579 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1582 * Some shared mappigns will want the pages marked read-only
1583 * to track write events. If so, we'll downgrade vm_page_prot
1584 * to the private version (using protection_map[] without the
1587 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1589 vm_flags_t vm_flags = vma->vm_flags;
1590 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1592 /* If it was private or non-writable, the write bit is already clear */
1593 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1596 /* The backer wishes to know when pages are first written to? */
1597 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1600 /* The open routine did something to the protections that pgprot_modify
1601 * won't preserve? */
1602 if (pgprot_val(vm_page_prot) !=
1603 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1606 /* Do we need to track softdirty? */
1607 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1610 /* Specialty mapping? */
1611 if (vm_flags & VM_PFNMAP)
1614 /* Can the mapping track the dirty pages? */
1615 return vma->vm_file && vma->vm_file->f_mapping &&
1616 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1620 * We account for memory if it's a private writeable mapping,
1621 * not hugepages and VM_NORESERVE wasn't set.
1623 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1626 * hugetlb has its own accounting separate from the core VM
1627 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1629 if (file && is_file_hugepages(file))
1632 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1635 unsigned long mmap_region(struct file *file, unsigned long addr,
1636 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1637 struct list_head *uf)
1639 struct mm_struct *mm = current->mm;
1640 struct vm_area_struct *vma, *prev;
1642 struct rb_node **rb_link, *rb_parent;
1643 unsigned long charged = 0;
1645 /* Check against address space limit. */
1646 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1647 unsigned long nr_pages;
1650 * MAP_FIXED may remove pages of mappings that intersects with
1651 * requested mapping. Account for the pages it would unmap.
1653 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1655 if (!may_expand_vm(mm, vm_flags,
1656 (len >> PAGE_SHIFT) - nr_pages))
1660 /* Clear old maps */
1661 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1663 if (do_munmap(mm, addr, len, uf))
1668 * Private writable mapping: check memory availability
1670 if (accountable_mapping(file, vm_flags)) {
1671 charged = len >> PAGE_SHIFT;
1672 if (security_vm_enough_memory_mm(mm, charged))
1674 vm_flags |= VM_ACCOUNT;
1678 * Can we just expand an old mapping?
1680 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1681 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1686 * Determine the object being mapped and call the appropriate
1687 * specific mapper. the address has already been validated, but
1688 * not unmapped, but the maps are removed from the list.
1690 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1697 vma->vm_start = addr;
1698 vma->vm_end = addr + len;
1699 vma->vm_flags = vm_flags;
1700 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1701 vma->vm_pgoff = pgoff;
1702 INIT_LIST_HEAD(&vma->anon_vma_chain);
1705 if (vm_flags & VM_DENYWRITE) {
1706 error = deny_write_access(file);
1710 if (vm_flags & VM_SHARED) {
1711 error = mapping_map_writable(file->f_mapping);
1713 goto allow_write_and_free_vma;
1716 /* ->mmap() can change vma->vm_file, but must guarantee that
1717 * vma_link() below can deny write-access if VM_DENYWRITE is set
1718 * and map writably if VM_SHARED is set. This usually means the
1719 * new file must not have been exposed to user-space, yet.
1721 vma->vm_file = get_file(file);
1722 error = call_mmap(file, vma);
1724 goto unmap_and_free_vma;
1726 /* Can addr have changed??
1728 * Answer: Yes, several device drivers can do it in their
1729 * f_op->mmap method. -DaveM
1730 * Bug: If addr is changed, prev, rb_link, rb_parent should
1731 * be updated for vma_link()
1733 WARN_ON_ONCE(addr != vma->vm_start);
1735 addr = vma->vm_start;
1736 vm_flags = vma->vm_flags;
1737 } else if (vm_flags & VM_SHARED) {
1738 error = shmem_zero_setup(vma);
1743 vma_link(mm, vma, prev, rb_link, rb_parent);
1744 /* Once vma denies write, undo our temporary denial count */
1746 if (vm_flags & VM_SHARED)
1747 mapping_unmap_writable(file->f_mapping);
1748 if (vm_flags & VM_DENYWRITE)
1749 allow_write_access(file);
1751 file = vma->vm_file;
1753 perf_event_mmap(vma);
1755 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1756 if (vm_flags & VM_LOCKED) {
1757 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1758 vma == get_gate_vma(current->mm)))
1759 mm->locked_vm += (len >> PAGE_SHIFT);
1761 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1768 * New (or expanded) vma always get soft dirty status.
1769 * Otherwise user-space soft-dirty page tracker won't
1770 * be able to distinguish situation when vma area unmapped,
1771 * then new mapped in-place (which must be aimed as
1772 * a completely new data area).
1774 vma->vm_flags |= VM_SOFTDIRTY;
1776 vma_set_page_prot(vma);
1781 vma->vm_file = NULL;
1784 /* Undo any partial mapping done by a device driver. */
1785 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1787 if (vm_flags & VM_SHARED)
1788 mapping_unmap_writable(file->f_mapping);
1789 allow_write_and_free_vma:
1790 if (vm_flags & VM_DENYWRITE)
1791 allow_write_access(file);
1793 kmem_cache_free(vm_area_cachep, vma);
1796 vm_unacct_memory(charged);
1800 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1803 * We implement the search by looking for an rbtree node that
1804 * immediately follows a suitable gap. That is,
1805 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1806 * - gap_end = vma->vm_start >= info->low_limit + length;
1807 * - gap_end - gap_start >= length
1810 struct mm_struct *mm = current->mm;
1811 struct vm_area_struct *vma;
1812 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1814 /* Adjust search length to account for worst case alignment overhead */
1815 length = info->length + info->align_mask;
1816 if (length < info->length)
1819 /* Adjust search limits by the desired length */
1820 if (info->high_limit < length)
1822 high_limit = info->high_limit - length;
1824 if (info->low_limit > high_limit)
1826 low_limit = info->low_limit + length;
1828 /* Check if rbtree root looks promising */
1829 if (RB_EMPTY_ROOT(&mm->mm_rb))
1831 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1832 if (vma->rb_subtree_gap < length)
1836 /* Visit left subtree if it looks promising */
1837 gap_end = vm_start_gap(vma);
1838 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1839 struct vm_area_struct *left =
1840 rb_entry(vma->vm_rb.rb_left,
1841 struct vm_area_struct, vm_rb);
1842 if (left->rb_subtree_gap >= length) {
1848 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1850 /* Check if current node has a suitable gap */
1851 if (gap_start > high_limit)
1853 if (gap_end >= low_limit &&
1854 gap_end > gap_start && gap_end - gap_start >= length)
1857 /* Visit right subtree if it looks promising */
1858 if (vma->vm_rb.rb_right) {
1859 struct vm_area_struct *right =
1860 rb_entry(vma->vm_rb.rb_right,
1861 struct vm_area_struct, vm_rb);
1862 if (right->rb_subtree_gap >= length) {
1868 /* Go back up the rbtree to find next candidate node */
1870 struct rb_node *prev = &vma->vm_rb;
1871 if (!rb_parent(prev))
1873 vma = rb_entry(rb_parent(prev),
1874 struct vm_area_struct, vm_rb);
1875 if (prev == vma->vm_rb.rb_left) {
1876 gap_start = vm_end_gap(vma->vm_prev);
1877 gap_end = vm_start_gap(vma);
1884 /* Check highest gap, which does not precede any rbtree node */
1885 gap_start = mm->highest_vm_end;
1886 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1887 if (gap_start > high_limit)
1891 /* We found a suitable gap. Clip it with the original low_limit. */
1892 if (gap_start < info->low_limit)
1893 gap_start = info->low_limit;
1895 /* Adjust gap address to the desired alignment */
1896 gap_start += (info->align_offset - gap_start) & info->align_mask;
1898 VM_BUG_ON(gap_start + info->length > info->high_limit);
1899 VM_BUG_ON(gap_start + info->length > gap_end);
1903 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1905 struct mm_struct *mm = current->mm;
1906 struct vm_area_struct *vma;
1907 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1909 /* Adjust search length to account for worst case alignment overhead */
1910 length = info->length + info->align_mask;
1911 if (length < info->length)
1915 * Adjust search limits by the desired length.
1916 * See implementation comment at top of unmapped_area().
1918 gap_end = info->high_limit;
1919 if (gap_end < length)
1921 high_limit = gap_end - length;
1923 if (info->low_limit > high_limit)
1925 low_limit = info->low_limit + length;
1927 /* Check highest gap, which does not precede any rbtree node */
1928 gap_start = mm->highest_vm_end;
1929 if (gap_start <= high_limit)
1932 /* Check if rbtree root looks promising */
1933 if (RB_EMPTY_ROOT(&mm->mm_rb))
1935 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1936 if (vma->rb_subtree_gap < length)
1940 /* Visit right subtree if it looks promising */
1941 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1942 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1943 struct vm_area_struct *right =
1944 rb_entry(vma->vm_rb.rb_right,
1945 struct vm_area_struct, vm_rb);
1946 if (right->rb_subtree_gap >= length) {
1953 /* Check if current node has a suitable gap */
1954 gap_end = vm_start_gap(vma);
1955 if (gap_end < low_limit)
1957 if (gap_start <= high_limit &&
1958 gap_end > gap_start && gap_end - gap_start >= length)
1961 /* Visit left subtree if it looks promising */
1962 if (vma->vm_rb.rb_left) {
1963 struct vm_area_struct *left =
1964 rb_entry(vma->vm_rb.rb_left,
1965 struct vm_area_struct, vm_rb);
1966 if (left->rb_subtree_gap >= length) {
1972 /* Go back up the rbtree to find next candidate node */
1974 struct rb_node *prev = &vma->vm_rb;
1975 if (!rb_parent(prev))
1977 vma = rb_entry(rb_parent(prev),
1978 struct vm_area_struct, vm_rb);
1979 if (prev == vma->vm_rb.rb_right) {
1980 gap_start = vma->vm_prev ?
1981 vm_end_gap(vma->vm_prev) : 0;
1988 /* We found a suitable gap. Clip it with the original high_limit. */
1989 if (gap_end > info->high_limit)
1990 gap_end = info->high_limit;
1993 /* Compute highest gap address at the desired alignment */
1994 gap_end -= info->length;
1995 gap_end -= (gap_end - info->align_offset) & info->align_mask;
1997 VM_BUG_ON(gap_end < info->low_limit);
1998 VM_BUG_ON(gap_end < gap_start);
2002 /* Get an address range which is currently unmapped.
2003 * For shmat() with addr=0.
2005 * Ugly calling convention alert:
2006 * Return value with the low bits set means error value,
2008 * if (ret & ~PAGE_MASK)
2011 * This function "knows" that -ENOMEM has the bits set.
2013 #ifndef HAVE_ARCH_UNMAPPED_AREA
2015 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2016 unsigned long len, unsigned long pgoff, unsigned long flags)
2018 struct mm_struct *mm = current->mm;
2019 struct vm_area_struct *vma, *prev;
2020 struct vm_unmapped_area_info info;
2022 if (len > TASK_SIZE - mmap_min_addr)
2025 if (flags & MAP_FIXED)
2029 addr = PAGE_ALIGN(addr);
2030 vma = find_vma_prev(mm, addr, &prev);
2031 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2032 (!vma || addr + len <= vm_start_gap(vma)) &&
2033 (!prev || addr >= vm_end_gap(prev)))
2039 info.low_limit = mm->mmap_base;
2040 info.high_limit = TASK_SIZE;
2041 info.align_mask = 0;
2042 return vm_unmapped_area(&info);
2047 * This mmap-allocator allocates new areas top-down from below the
2048 * stack's low limit (the base):
2050 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2052 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2053 const unsigned long len, const unsigned long pgoff,
2054 const unsigned long flags)
2056 struct vm_area_struct *vma, *prev;
2057 struct mm_struct *mm = current->mm;
2058 unsigned long addr = addr0;
2059 struct vm_unmapped_area_info info;
2061 /* requested length too big for entire address space */
2062 if (len > TASK_SIZE - mmap_min_addr)
2065 if (flags & MAP_FIXED)
2068 /* requesting a specific address */
2070 addr = PAGE_ALIGN(addr);
2071 vma = find_vma_prev(mm, addr, &prev);
2072 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2073 (!vma || addr + len <= vm_start_gap(vma)) &&
2074 (!prev || addr >= vm_end_gap(prev)))
2078 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2080 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2081 info.high_limit = mm->mmap_base;
2082 info.align_mask = 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 error = security_mmap_addr(address);
2358 /* Enforce stack_guard_gap */
2359 prev = vma->vm_prev;
2360 /* Check that both stack segments have the same anon_vma? */
2361 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2362 (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2363 if (address - prev->vm_end < stack_guard_gap)
2367 /* We must make sure the anon_vma is allocated. */
2368 if (unlikely(anon_vma_prepare(vma)))
2372 * vma->vm_start/vm_end cannot change under us because the caller
2373 * is required to hold the mmap_sem in read mode. We need the
2374 * anon_vma lock to serialize against concurrent expand_stacks.
2376 anon_vma_lock_write(vma->anon_vma);
2378 /* Somebody else might have raced and expanded it already */
2379 if (address < vma->vm_start) {
2380 unsigned long size, grow;
2382 size = vma->vm_end - address;
2383 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2386 if (grow <= vma->vm_pgoff) {
2387 error = acct_stack_growth(vma, size, grow);
2390 * vma_gap_update() doesn't support concurrent
2391 * updates, but we only hold a shared mmap_sem
2392 * lock here, so we need to protect against
2393 * concurrent vma expansions.
2394 * anon_vma_lock_write() doesn't help here, as
2395 * we don't guarantee that all growable vmas
2396 * in a mm share the same root anon vma.
2397 * So, we reuse mm->page_table_lock to guard
2398 * against concurrent vma expansions.
2400 spin_lock(&mm->page_table_lock);
2401 if (vma->vm_flags & VM_LOCKED)
2402 mm->locked_vm += grow;
2403 vm_stat_account(mm, vma->vm_flags, grow);
2404 anon_vma_interval_tree_pre_update_vma(vma);
2405 vma->vm_start = address;
2406 vma->vm_pgoff -= grow;
2407 anon_vma_interval_tree_post_update_vma(vma);
2408 vma_gap_update(vma);
2409 spin_unlock(&mm->page_table_lock);
2411 perf_event_mmap(vma);
2415 anon_vma_unlock_write(vma->anon_vma);
2416 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2421 /* enforced gap between the expanding stack and other mappings. */
2422 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2424 static int __init cmdline_parse_stack_guard_gap(char *p)
2429 val = simple_strtoul(p, &endptr, 10);
2431 stack_guard_gap = val << PAGE_SHIFT;
2435 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2437 #ifdef CONFIG_STACK_GROWSUP
2438 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2440 return expand_upwards(vma, address);
2443 struct vm_area_struct *
2444 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2446 struct vm_area_struct *vma, *prev;
2449 vma = find_vma_prev(mm, addr, &prev);
2450 if (vma && (vma->vm_start <= addr))
2452 if (!prev || 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 start = vma->vm_start;
2479 if (expand_stack(vma, addr))
2481 if (vma->vm_flags & VM_LOCKED)
2482 populate_vma_page_range(vma, addr, start, NULL);
2487 EXPORT_SYMBOL_GPL(find_extend_vma);
2490 * Ok - we have the memory areas we should free on the vma list,
2491 * so release them, and do the vma updates.
2493 * Called with the mm semaphore held.
2495 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2497 unsigned long nr_accounted = 0;
2499 /* Update high watermark before we lower total_vm */
2500 update_hiwater_vm(mm);
2502 long nrpages = vma_pages(vma);
2504 if (vma->vm_flags & VM_ACCOUNT)
2505 nr_accounted += nrpages;
2506 vm_stat_account(mm, vma->vm_flags, -nrpages);
2507 vma = remove_vma(vma);
2509 vm_unacct_memory(nr_accounted);
2514 * Get rid of page table information in the indicated region.
2516 * Called with the mm semaphore held.
2518 static void unmap_region(struct mm_struct *mm,
2519 struct vm_area_struct *vma, struct vm_area_struct *prev,
2520 unsigned long start, unsigned long end)
2522 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2523 struct mmu_gather tlb;
2526 tlb_gather_mmu(&tlb, mm, start, end);
2527 update_hiwater_rss(mm);
2528 unmap_vmas(&tlb, vma, start, end);
2529 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2530 next ? next->vm_start : USER_PGTABLES_CEILING);
2531 tlb_finish_mmu(&tlb, start, end);
2535 * Create a list of vma's touched by the unmap, removing them from the mm's
2536 * vma list as we go..
2539 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2540 struct vm_area_struct *prev, unsigned long end)
2542 struct vm_area_struct **insertion_point;
2543 struct vm_area_struct *tail_vma = NULL;
2545 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2546 vma->vm_prev = NULL;
2548 vma_rb_erase(vma, &mm->mm_rb);
2552 } while (vma && vma->vm_start < end);
2553 *insertion_point = vma;
2555 vma->vm_prev = prev;
2556 vma_gap_update(vma);
2558 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2559 tail_vma->vm_next = NULL;
2561 /* Kill the cache */
2562 vmacache_invalidate(mm);
2566 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2567 * has already been checked or doesn't make sense to fail.
2569 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2570 unsigned long addr, int new_below)
2572 struct vm_area_struct *new;
2575 if (vma->vm_ops && vma->vm_ops->split) {
2576 err = vma->vm_ops->split(vma, addr);
2581 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2585 /* most fields are the same, copy all, and then fixup */
2588 INIT_LIST_HEAD(&new->anon_vma_chain);
2593 new->vm_start = addr;
2594 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2597 err = vma_dup_policy(vma, new);
2601 err = anon_vma_clone(new, vma);
2606 get_file(new->vm_file);
2608 if (new->vm_ops && new->vm_ops->open)
2609 new->vm_ops->open(new);
2612 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2613 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2615 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2621 /* Clean everything up if vma_adjust failed. */
2622 if (new->vm_ops && new->vm_ops->close)
2623 new->vm_ops->close(new);
2626 unlink_anon_vmas(new);
2628 mpol_put(vma_policy(new));
2630 kmem_cache_free(vm_area_cachep, new);
2635 * Split a vma into two pieces at address 'addr', a new vma is allocated
2636 * either for the first part or the tail.
2638 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2639 unsigned long addr, int new_below)
2641 if (mm->map_count >= sysctl_max_map_count)
2644 return __split_vma(mm, vma, addr, new_below);
2647 /* Munmap is split into 2 main parts -- this part which finds
2648 * what needs doing, and the areas themselves, which do the
2649 * work. This now handles partial unmappings.
2650 * Jeremy Fitzhardinge <jeremy@goop.org>
2652 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2653 struct list_head *uf)
2656 struct vm_area_struct *vma, *prev, *last;
2658 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2661 len = PAGE_ALIGN(len);
2665 /* Find the first overlapping VMA */
2666 vma = find_vma(mm, start);
2669 prev = vma->vm_prev;
2670 /* we have start < vma->vm_end */
2672 /* if it doesn't overlap, we have nothing.. */
2674 if (vma->vm_start >= end)
2678 * If we need to split any vma, do it now to save pain later.
2680 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2681 * unmapped vm_area_struct will remain in use: so lower split_vma
2682 * places tmp vma above, and higher split_vma places tmp vma below.
2684 if (start > vma->vm_start) {
2688 * Make sure that map_count on return from munmap() will
2689 * not exceed its limit; but let map_count go just above
2690 * its limit temporarily, to help free resources as expected.
2692 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2695 error = __split_vma(mm, vma, start, 0);
2701 /* Does it split the last one? */
2702 last = find_vma(mm, end);
2703 if (last && end > last->vm_start) {
2704 int error = __split_vma(mm, last, end, 1);
2708 vma = prev ? prev->vm_next : mm->mmap;
2712 * If userfaultfd_unmap_prep returns an error the vmas
2713 * will remain splitted, but userland will get a
2714 * highly unexpected error anyway. This is no
2715 * different than the case where the first of the two
2716 * __split_vma fails, but we don't undo the first
2717 * split, despite we could. This is unlikely enough
2718 * failure that it's not worth optimizing it for.
2720 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2726 * unlock any mlock()ed ranges before detaching vmas
2728 if (mm->locked_vm) {
2729 struct vm_area_struct *tmp = vma;
2730 while (tmp && tmp->vm_start < end) {
2731 if (tmp->vm_flags & VM_LOCKED) {
2732 mm->locked_vm -= vma_pages(tmp);
2733 munlock_vma_pages_all(tmp);
2740 * Remove the vma's, and unmap the actual pages
2742 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2743 unmap_region(mm, vma, prev, start, end);
2745 arch_unmap(mm, vma, start, end);
2747 /* Fix up all other VM information */
2748 remove_vma_list(mm, vma);
2753 int vm_munmap(unsigned long start, size_t len)
2756 struct mm_struct *mm = current->mm;
2759 if (down_write_killable(&mm->mmap_sem))
2762 ret = do_munmap(mm, start, len, &uf);
2763 up_write(&mm->mmap_sem);
2764 userfaultfd_unmap_complete(mm, &uf);
2767 EXPORT_SYMBOL(vm_munmap);
2769 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2771 profile_munmap(addr);
2772 return vm_munmap(addr, len);
2777 * Emulation of deprecated remap_file_pages() syscall.
2779 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2780 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2783 struct mm_struct *mm = current->mm;
2784 struct vm_area_struct *vma;
2785 unsigned long populate = 0;
2786 unsigned long ret = -EINVAL;
2789 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2790 current->comm, current->pid);
2794 start = start & PAGE_MASK;
2795 size = size & PAGE_MASK;
2797 if (start + size <= start)
2800 /* Does pgoff wrap? */
2801 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2804 if (down_write_killable(&mm->mmap_sem))
2807 vma = find_vma(mm, start);
2809 if (!vma || !(vma->vm_flags & VM_SHARED))
2812 if (start < vma->vm_start)
2815 if (start + size > vma->vm_end) {
2816 struct vm_area_struct *next;
2818 for (next = vma->vm_next; next; next = next->vm_next) {
2819 /* hole between vmas ? */
2820 if (next->vm_start != next->vm_prev->vm_end)
2823 if (next->vm_file != vma->vm_file)
2826 if (next->vm_flags != vma->vm_flags)
2829 if (start + size <= next->vm_end)
2837 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2838 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2839 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2841 flags &= MAP_NONBLOCK;
2842 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2843 if (vma->vm_flags & VM_LOCKED) {
2844 struct vm_area_struct *tmp;
2845 flags |= MAP_LOCKED;
2847 /* drop PG_Mlocked flag for over-mapped range */
2848 for (tmp = vma; tmp->vm_start >= start + size;
2849 tmp = tmp->vm_next) {
2851 * Split pmd and munlock page on the border
2854 vma_adjust_trans_huge(tmp, start, start + size, 0);
2856 munlock_vma_pages_range(tmp,
2857 max(tmp->vm_start, start),
2858 min(tmp->vm_end, start + size));
2862 file = get_file(vma->vm_file);
2863 ret = do_mmap_pgoff(vma->vm_file, start, size,
2864 prot, flags, pgoff, &populate, NULL);
2867 up_write(&mm->mmap_sem);
2869 mm_populate(ret, populate);
2870 if (!IS_ERR_VALUE(ret))
2875 static inline void verify_mm_writelocked(struct mm_struct *mm)
2877 #ifdef CONFIG_DEBUG_VM
2878 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2880 up_read(&mm->mmap_sem);
2886 * this is really a simplified "do_mmap". it only handles
2887 * anonymous maps. eventually we may be able to do some
2888 * brk-specific accounting here.
2890 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
2892 struct mm_struct *mm = current->mm;
2893 struct vm_area_struct *vma, *prev;
2894 struct rb_node **rb_link, *rb_parent;
2895 pgoff_t pgoff = addr >> PAGE_SHIFT;
2898 /* Until we need other flags, refuse anything except VM_EXEC. */
2899 if ((flags & (~VM_EXEC)) != 0)
2901 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2903 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2904 if (offset_in_page(error))
2907 error = mlock_future_check(mm, mm->def_flags, len);
2912 * mm->mmap_sem is required to protect against another thread
2913 * changing the mappings in case we sleep.
2915 verify_mm_writelocked(mm);
2918 * Clear old maps. this also does some error checking for us
2920 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2922 if (do_munmap(mm, addr, len, uf))
2926 /* Check against address space limits *after* clearing old maps... */
2927 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2930 if (mm->map_count > sysctl_max_map_count)
2933 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2936 /* Can we just expand an old private anonymous mapping? */
2937 vma = vma_merge(mm, prev, addr, addr + len, flags,
2938 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2943 * create a vma struct for an anonymous mapping
2945 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2947 vm_unacct_memory(len >> PAGE_SHIFT);
2951 INIT_LIST_HEAD(&vma->anon_vma_chain);
2953 vma->vm_start = addr;
2954 vma->vm_end = addr + len;
2955 vma->vm_pgoff = pgoff;
2956 vma->vm_flags = flags;
2957 vma->vm_page_prot = vm_get_page_prot(flags);
2958 vma_link(mm, vma, prev, rb_link, rb_parent);
2960 perf_event_mmap(vma);
2961 mm->total_vm += len >> PAGE_SHIFT;
2962 mm->data_vm += len >> PAGE_SHIFT;
2963 if (flags & VM_LOCKED)
2964 mm->locked_vm += (len >> PAGE_SHIFT);
2965 vma->vm_flags |= VM_SOFTDIRTY;
2969 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
2971 struct mm_struct *mm = current->mm;
2977 len = PAGE_ALIGN(request);
2983 if (down_write_killable(&mm->mmap_sem))
2986 ret = do_brk_flags(addr, len, flags, &uf);
2987 populate = ((mm->def_flags & VM_LOCKED) != 0);
2988 up_write(&mm->mmap_sem);
2989 userfaultfd_unmap_complete(mm, &uf);
2990 if (populate && !ret)
2991 mm_populate(addr, len);
2994 EXPORT_SYMBOL(vm_brk_flags);
2996 int vm_brk(unsigned long addr, unsigned long len)
2998 return vm_brk_flags(addr, len, 0);
3000 EXPORT_SYMBOL(vm_brk);
3002 /* Release all mmaps. */
3003 void exit_mmap(struct mm_struct *mm)
3005 struct mmu_gather tlb;
3006 struct vm_area_struct *vma;
3007 unsigned long nr_accounted = 0;
3009 /* mm's last user has gone, and its about to be pulled down */
3010 mmu_notifier_release(mm);
3012 if (unlikely(mm_is_oom_victim(mm))) {
3014 * Manually reap the mm to free as much memory as possible.
3015 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3016 * this mm from further consideration. Taking mm->mmap_sem for
3017 * write after setting MMF_OOM_SKIP will guarantee that the oom
3018 * reaper will not run on this mm again after mmap_sem is
3021 * Nothing can be holding mm->mmap_sem here and the above call
3022 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3023 * __oom_reap_task_mm() will not block.
3025 * This needs to be done before calling munlock_vma_pages_all(),
3026 * which clears VM_LOCKED, otherwise the oom reaper cannot
3029 mutex_lock(&oom_lock);
3030 __oom_reap_task_mm(mm);
3031 mutex_unlock(&oom_lock);
3033 set_bit(MMF_OOM_SKIP, &mm->flags);
3034 down_write(&mm->mmap_sem);
3035 up_write(&mm->mmap_sem);
3038 if (mm->locked_vm) {
3041 if (vma->vm_flags & VM_LOCKED)
3042 munlock_vma_pages_all(vma);
3050 if (!vma) /* Can happen if dup_mmap() received an OOM */
3055 tlb_gather_mmu(&tlb, mm, 0, -1);
3056 /* update_hiwater_rss(mm) here? but nobody should be looking */
3057 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3058 unmap_vmas(&tlb, vma, 0, -1);
3059 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3060 tlb_finish_mmu(&tlb, 0, -1);
3063 * Walk the list again, actually closing and freeing it,
3064 * with preemption enabled, without holding any MM locks.
3067 if (vma->vm_flags & VM_ACCOUNT)
3068 nr_accounted += vma_pages(vma);
3069 vma = remove_vma(vma);
3071 vm_unacct_memory(nr_accounted);
3074 /* Insert vm structure into process list sorted by address
3075 * and into the inode's i_mmap tree. If vm_file is non-NULL
3076 * then i_mmap_rwsem is taken here.
3078 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3080 struct vm_area_struct *prev;
3081 struct rb_node **rb_link, *rb_parent;
3083 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3084 &prev, &rb_link, &rb_parent))
3086 if ((vma->vm_flags & VM_ACCOUNT) &&
3087 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3091 * The vm_pgoff of a purely anonymous vma should be irrelevant
3092 * until its first write fault, when page's anon_vma and index
3093 * are set. But now set the vm_pgoff it will almost certainly
3094 * end up with (unless mremap moves it elsewhere before that
3095 * first wfault), so /proc/pid/maps tells a consistent story.
3097 * By setting it to reflect the virtual start address of the
3098 * vma, merges and splits can happen in a seamless way, just
3099 * using the existing file pgoff checks and manipulations.
3100 * Similarly in do_mmap_pgoff and in do_brk.
3102 if (vma_is_anonymous(vma)) {
3103 BUG_ON(vma->anon_vma);
3104 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3107 vma_link(mm, vma, prev, rb_link, rb_parent);
3112 * Copy the vma structure to a new location in the same mm,
3113 * prior to moving page table entries, to effect an mremap move.
3115 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3116 unsigned long addr, unsigned long len, pgoff_t pgoff,
3117 bool *need_rmap_locks)
3119 struct vm_area_struct *vma = *vmap;
3120 unsigned long vma_start = vma->vm_start;
3121 struct mm_struct *mm = vma->vm_mm;
3122 struct vm_area_struct *new_vma, *prev;
3123 struct rb_node **rb_link, *rb_parent;
3124 bool faulted_in_anon_vma = true;
3127 * If anonymous vma has not yet been faulted, update new pgoff
3128 * to match new location, to increase its chance of merging.
3130 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3131 pgoff = addr >> PAGE_SHIFT;
3132 faulted_in_anon_vma = false;
3135 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3136 return NULL; /* should never get here */
3137 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3138 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3139 vma->vm_userfaultfd_ctx);
3142 * Source vma may have been merged into new_vma
3144 if (unlikely(vma_start >= new_vma->vm_start &&
3145 vma_start < new_vma->vm_end)) {
3147 * The only way we can get a vma_merge with
3148 * self during an mremap is if the vma hasn't
3149 * been faulted in yet and we were allowed to
3150 * reset the dst vma->vm_pgoff to the
3151 * destination address of the mremap to allow
3152 * the merge to happen. mremap must change the
3153 * vm_pgoff linearity between src and dst vmas
3154 * (in turn preventing a vma_merge) to be
3155 * safe. It is only safe to keep the vm_pgoff
3156 * linear if there are no pages mapped yet.
3158 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3159 *vmap = vma = new_vma;
3161 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3163 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3167 new_vma->vm_start = addr;
3168 new_vma->vm_end = addr + len;
3169 new_vma->vm_pgoff = pgoff;
3170 if (vma_dup_policy(vma, new_vma))
3172 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3173 if (anon_vma_clone(new_vma, vma))
3174 goto out_free_mempol;
3175 if (new_vma->vm_file)
3176 get_file(new_vma->vm_file);
3177 if (new_vma->vm_ops && new_vma->vm_ops->open)
3178 new_vma->vm_ops->open(new_vma);
3179 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3180 *need_rmap_locks = false;
3185 mpol_put(vma_policy(new_vma));
3187 kmem_cache_free(vm_area_cachep, new_vma);
3193 * Return true if the calling process may expand its vm space by the passed
3196 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3198 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3201 if (is_data_mapping(flags) &&
3202 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3203 /* Workaround for Valgrind */
3204 if (rlimit(RLIMIT_DATA) == 0 &&
3205 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3207 if (!ignore_rlimit_data) {
3208 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits or use boot option ignore_rlimit_data.\n",
3209 current->comm, current->pid,
3210 (mm->data_vm + npages) << PAGE_SHIFT,
3211 rlimit(RLIMIT_DATA));
3219 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3221 mm->total_vm += npages;
3223 if (is_exec_mapping(flags))
3224 mm->exec_vm += npages;
3225 else if (is_stack_mapping(flags))
3226 mm->stack_vm += npages;
3227 else if (is_data_mapping(flags))
3228 mm->data_vm += npages;
3231 static int special_mapping_fault(struct vm_fault *vmf);
3234 * Having a close hook prevents vma merging regardless of flags.
3236 static void special_mapping_close(struct vm_area_struct *vma)
3240 static const char *special_mapping_name(struct vm_area_struct *vma)
3242 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3245 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3247 struct vm_special_mapping *sm = new_vma->vm_private_data;
3249 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3253 return sm->mremap(sm, new_vma);
3258 static const struct vm_operations_struct special_mapping_vmops = {
3259 .close = special_mapping_close,
3260 .fault = special_mapping_fault,
3261 .mremap = special_mapping_mremap,
3262 .name = special_mapping_name,
3265 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3266 .close = special_mapping_close,
3267 .fault = special_mapping_fault,
3270 static int special_mapping_fault(struct vm_fault *vmf)
3272 struct vm_area_struct *vma = vmf->vma;
3274 struct page **pages;
3276 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3277 pages = vma->vm_private_data;
3279 struct vm_special_mapping *sm = vma->vm_private_data;
3282 return sm->fault(sm, vmf->vma, vmf);
3287 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3291 struct page *page = *pages;
3297 return VM_FAULT_SIGBUS;
3300 static struct vm_area_struct *__install_special_mapping(
3301 struct mm_struct *mm,
3302 unsigned long addr, unsigned long len,
3303 unsigned long vm_flags, void *priv,
3304 const struct vm_operations_struct *ops)
3307 struct vm_area_struct *vma;
3309 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3310 if (unlikely(vma == NULL))
3311 return ERR_PTR(-ENOMEM);
3313 INIT_LIST_HEAD(&vma->anon_vma_chain);
3315 vma->vm_start = addr;
3316 vma->vm_end = addr + len;
3318 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3319 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3322 vma->vm_private_data = priv;
3324 ret = insert_vm_struct(mm, vma);
3328 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3330 perf_event_mmap(vma);
3335 kmem_cache_free(vm_area_cachep, vma);
3336 return ERR_PTR(ret);
3339 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3340 const struct vm_special_mapping *sm)
3342 return vma->vm_private_data == sm &&
3343 (vma->vm_ops == &special_mapping_vmops ||
3344 vma->vm_ops == &legacy_special_mapping_vmops);
3348 * Called with mm->mmap_sem held for writing.
3349 * Insert a new vma covering the given region, with the given flags.
3350 * Its pages are supplied by the given array of struct page *.
3351 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3352 * The region past the last page supplied will always produce SIGBUS.
3353 * The array pointer and the pages it points to are assumed to stay alive
3354 * for as long as this mapping might exist.
3356 struct vm_area_struct *_install_special_mapping(
3357 struct mm_struct *mm,
3358 unsigned long addr, unsigned long len,
3359 unsigned long vm_flags, const struct vm_special_mapping *spec)
3361 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3362 &special_mapping_vmops);
3365 int install_special_mapping(struct mm_struct *mm,
3366 unsigned long addr, unsigned long len,
3367 unsigned long vm_flags, struct page **pages)
3369 struct vm_area_struct *vma = __install_special_mapping(
3370 mm, addr, len, vm_flags, (void *)pages,
3371 &legacy_special_mapping_vmops);
3373 return PTR_ERR_OR_ZERO(vma);
3376 static DEFINE_MUTEX(mm_all_locks_mutex);
3378 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3380 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3382 * The LSB of head.next can't change from under us
3383 * because we hold the mm_all_locks_mutex.
3385 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3387 * We can safely modify head.next after taking the
3388 * anon_vma->root->rwsem. If some other vma in this mm shares
3389 * the same anon_vma we won't take it again.
3391 * No need of atomic instructions here, head.next
3392 * can't change from under us thanks to the
3393 * anon_vma->root->rwsem.
3395 if (__test_and_set_bit(0, (unsigned long *)
3396 &anon_vma->root->rb_root.rb_root.rb_node))
3401 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3403 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3405 * AS_MM_ALL_LOCKS can't change from under us because
3406 * we hold the mm_all_locks_mutex.
3408 * Operations on ->flags have to be atomic because
3409 * even if AS_MM_ALL_LOCKS is stable thanks to the
3410 * mm_all_locks_mutex, there may be other cpus
3411 * changing other bitflags in parallel to us.
3413 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3415 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3420 * This operation locks against the VM for all pte/vma/mm related
3421 * operations that could ever happen on a certain mm. This includes
3422 * vmtruncate, try_to_unmap, and all page faults.
3424 * The caller must take the mmap_sem in write mode before calling
3425 * mm_take_all_locks(). The caller isn't allowed to release the
3426 * mmap_sem until mm_drop_all_locks() returns.
3428 * mmap_sem in write mode is required in order to block all operations
3429 * that could modify pagetables and free pages without need of
3430 * altering the vma layout. It's also needed in write mode to avoid new
3431 * anon_vmas to be associated with existing vmas.
3433 * A single task can't take more than one mm_take_all_locks() in a row
3434 * or it would deadlock.
3436 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3437 * mapping->flags avoid to take the same lock twice, if more than one
3438 * vma in this mm is backed by the same anon_vma or address_space.
3440 * We take locks in following order, accordingly to comment at beginning
3442 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3444 * - all i_mmap_rwsem locks;
3445 * - all anon_vma->rwseml
3447 * We can take all locks within these types randomly because the VM code
3448 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3449 * mm_all_locks_mutex.
3451 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3452 * that may have to take thousand of locks.
3454 * mm_take_all_locks() can fail if it's interrupted by signals.
3456 int mm_take_all_locks(struct mm_struct *mm)
3458 struct vm_area_struct *vma;
3459 struct anon_vma_chain *avc;
3461 BUG_ON(down_read_trylock(&mm->mmap_sem));
3463 mutex_lock(&mm_all_locks_mutex);
3465 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3466 if (signal_pending(current))
3468 if (vma->vm_file && vma->vm_file->f_mapping &&
3469 is_vm_hugetlb_page(vma))
3470 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3473 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3474 if (signal_pending(current))
3476 if (vma->vm_file && vma->vm_file->f_mapping &&
3477 !is_vm_hugetlb_page(vma))
3478 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3481 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3482 if (signal_pending(current))
3485 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3486 vm_lock_anon_vma(mm, avc->anon_vma);
3492 mm_drop_all_locks(mm);
3496 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3498 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3500 * The LSB of head.next can't change to 0 from under
3501 * us because we hold the mm_all_locks_mutex.
3503 * We must however clear the bitflag before unlocking
3504 * the vma so the users using the anon_vma->rb_root will
3505 * never see our bitflag.
3507 * No need of atomic instructions here, head.next
3508 * can't change from under us until we release the
3509 * anon_vma->root->rwsem.
3511 if (!__test_and_clear_bit(0, (unsigned long *)
3512 &anon_vma->root->rb_root.rb_root.rb_node))
3514 anon_vma_unlock_write(anon_vma);
3518 static void vm_unlock_mapping(struct address_space *mapping)
3520 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3522 * AS_MM_ALL_LOCKS can't change to 0 from under us
3523 * because we hold the mm_all_locks_mutex.
3525 i_mmap_unlock_write(mapping);
3526 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3533 * The mmap_sem cannot be released by the caller until
3534 * mm_drop_all_locks() returns.
3536 void mm_drop_all_locks(struct mm_struct *mm)
3538 struct vm_area_struct *vma;
3539 struct anon_vma_chain *avc;
3541 BUG_ON(down_read_trylock(&mm->mmap_sem));
3542 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3544 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3546 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3547 vm_unlock_anon_vma(avc->anon_vma);
3548 if (vma->vm_file && vma->vm_file->f_mapping)
3549 vm_unlock_mapping(vma->vm_file->f_mapping);
3552 mutex_unlock(&mm_all_locks_mutex);
3556 * initialise the percpu counter for VM
3558 void __init mmap_init(void)
3562 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3567 * Initialise sysctl_user_reserve_kbytes.
3569 * This is intended to prevent a user from starting a single memory hogging
3570 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3573 * The default value is min(3% of free memory, 128MB)
3574 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3576 static int init_user_reserve(void)
3578 unsigned long free_kbytes;
3580 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3582 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3585 subsys_initcall(init_user_reserve);
3588 * Initialise sysctl_admin_reserve_kbytes.
3590 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3591 * to log in and kill a memory hogging process.
3593 * Systems with more than 256MB will reserve 8MB, enough to recover
3594 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3595 * only reserve 3% of free pages by default.
3597 static int init_admin_reserve(void)
3599 unsigned long free_kbytes;
3601 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3603 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3606 subsys_initcall(init_admin_reserve);
3609 * Reinititalise user and admin reserves if memory is added or removed.
3611 * The default user reserve max is 128MB, and the default max for the
3612 * admin reserve is 8MB. These are usually, but not always, enough to
3613 * enable recovery from a memory hogging process using login/sshd, a shell,
3614 * and tools like top. It may make sense to increase or even disable the
3615 * reserve depending on the existence of swap or variations in the recovery
3616 * tools. So, the admin may have changed them.
3618 * If memory is added and the reserves have been eliminated or increased above
3619 * the default max, then we'll trust the admin.
3621 * If memory is removed and there isn't enough free memory, then we
3622 * need to reset the reserves.
3624 * Otherwise keep the reserve set by the admin.
3626 static int reserve_mem_notifier(struct notifier_block *nb,
3627 unsigned long action, void *data)
3629 unsigned long tmp, free_kbytes;
3633 /* Default max is 128MB. Leave alone if modified by operator. */
3634 tmp = sysctl_user_reserve_kbytes;
3635 if (0 < tmp && tmp < (1UL << 17))
3636 init_user_reserve();
3638 /* Default max is 8MB. Leave alone if modified by operator. */
3639 tmp = sysctl_admin_reserve_kbytes;
3640 if (0 < tmp && tmp < (1UL << 13))
3641 init_admin_reserve();
3645 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3647 if (sysctl_user_reserve_kbytes > free_kbytes) {
3648 init_user_reserve();
3649 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3650 sysctl_user_reserve_kbytes);
3653 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3654 init_admin_reserve();
3655 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3656 sysctl_admin_reserve_kbytes);
3665 static struct notifier_block reserve_mem_nb = {
3666 .notifier_call = reserve_mem_notifier,
3669 static int __meminit init_reserve_notifier(void)
3671 if (register_hotmemory_notifier(&reserve_mem_nb))
3672 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3676 subsys_initcall(init_reserve_notifier);