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/profile.h>
29 #include <linux/export.h>
30 #include <linux/mount.h>
31 #include <linux/mempolicy.h>
32 #include <linux/rmap.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/mmdebug.h>
35 #include <linux/perf_event.h>
36 #include <linux/audit.h>
37 #include <linux/khugepaged.h>
38 #include <linux/uprobes.h>
39 #include <linux/rbtree_augmented.h>
40 #include <linux/notifier.h>
41 #include <linux/memory.h>
42 #include <linux/printk.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/moduleparam.h>
45 #include <linux/pkeys.h>
47 #include <asm/uaccess.h>
48 #include <asm/cacheflush.h>
50 #include <asm/mmu_context.h>
54 #ifndef arch_mmap_check
55 #define arch_mmap_check(addr, len, flags) (0)
58 #ifndef arch_rebalance_pgtables
59 #define arch_rebalance_pgtables(addr, len) (addr)
62 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
63 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
64 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
65 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
67 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
68 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
69 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
70 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
73 static bool ignore_rlimit_data = true;
74 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
76 static void unmap_region(struct mm_struct *mm,
77 struct vm_area_struct *vma, struct vm_area_struct *prev,
78 unsigned long start, unsigned long end);
80 /* description of effects of mapping type and prot in current implementation.
81 * this is due to the limited x86 page protection hardware. The expected
82 * behavior is in parens:
85 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
86 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
87 * w: (no) no w: (no) no w: (yes) yes w: (no) no
88 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
90 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
91 * w: (no) no w: (no) no w: (copy) copy w: (no) no
92 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
95 pgprot_t protection_map[16] = {
96 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
97 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
100 pgprot_t vm_get_page_prot(unsigned long vm_flags)
102 return __pgprot(pgprot_val(protection_map[vm_flags &
103 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
104 pgprot_val(arch_vm_get_page_prot(vm_flags)));
106 EXPORT_SYMBOL(vm_get_page_prot);
108 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
110 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
113 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
114 void vma_set_page_prot(struct vm_area_struct *vma)
116 unsigned long vm_flags = vma->vm_flags;
118 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
119 if (vma_wants_writenotify(vma)) {
120 vm_flags &= ~VM_SHARED;
121 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot,
127 * Requires inode->i_mapping->i_mmap_rwsem
129 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
130 struct file *file, struct address_space *mapping)
132 if (vma->vm_flags & VM_DENYWRITE)
133 atomic_inc(&file_inode(file)->i_writecount);
134 if (vma->vm_flags & VM_SHARED)
135 mapping_unmap_writable(mapping);
137 flush_dcache_mmap_lock(mapping);
138 vma_interval_tree_remove(vma, &mapping->i_mmap);
139 flush_dcache_mmap_unlock(mapping);
143 * Unlink a file-based vm structure from its interval tree, to hide
144 * vma from rmap and vmtruncate before freeing its page tables.
146 void unlink_file_vma(struct vm_area_struct *vma)
148 struct file *file = vma->vm_file;
151 struct address_space *mapping = file->f_mapping;
152 i_mmap_lock_write(mapping);
153 __remove_shared_vm_struct(vma, file, mapping);
154 i_mmap_unlock_write(mapping);
159 * Close a vm structure and free it, returning the next.
161 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
163 struct vm_area_struct *next = vma->vm_next;
166 if (vma->vm_ops && vma->vm_ops->close)
167 vma->vm_ops->close(vma);
170 mpol_put(vma_policy(vma));
171 kmem_cache_free(vm_area_cachep, vma);
175 static unsigned long do_brk(unsigned long addr, unsigned long len);
177 SYSCALL_DEFINE1(brk, unsigned long, brk)
179 unsigned long retval;
180 unsigned long newbrk, oldbrk;
181 struct mm_struct *mm = current->mm;
182 unsigned long min_brk;
185 down_write(&mm->mmap_sem);
187 #ifdef CONFIG_COMPAT_BRK
189 * CONFIG_COMPAT_BRK can still be overridden by setting
190 * randomize_va_space to 2, which will still cause mm->start_brk
191 * to be arbitrarily shifted
193 if (current->brk_randomized)
194 min_brk = mm->start_brk;
196 min_brk = mm->end_data;
198 min_brk = mm->start_brk;
204 * Check against rlimit here. If this check is done later after the test
205 * of oldbrk with newbrk then it can escape the test and let the data
206 * segment grow beyond its set limit the in case where the limit is
207 * not page aligned -Ram Gupta
209 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
210 mm->end_data, mm->start_data))
213 newbrk = PAGE_ALIGN(brk);
214 oldbrk = PAGE_ALIGN(mm->brk);
215 if (oldbrk == newbrk)
218 /* Always allow shrinking brk. */
219 if (brk <= mm->brk) {
220 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
225 /* Check against existing mmap mappings. */
226 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
229 /* Ok, looks good - let it rip. */
230 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
235 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
236 up_write(&mm->mmap_sem);
238 mm_populate(oldbrk, newbrk - oldbrk);
243 up_write(&mm->mmap_sem);
247 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
249 unsigned long max, subtree_gap;
252 max -= vma->vm_prev->vm_end;
253 if (vma->vm_rb.rb_left) {
254 subtree_gap = rb_entry(vma->vm_rb.rb_left,
255 struct vm_area_struct, vm_rb)->rb_subtree_gap;
256 if (subtree_gap > max)
259 if (vma->vm_rb.rb_right) {
260 subtree_gap = rb_entry(vma->vm_rb.rb_right,
261 struct vm_area_struct, vm_rb)->rb_subtree_gap;
262 if (subtree_gap > max)
268 #ifdef CONFIG_DEBUG_VM_RB
269 static int browse_rb(struct mm_struct *mm)
271 struct rb_root *root = &mm->mm_rb;
272 int i = 0, j, bug = 0;
273 struct rb_node *nd, *pn = NULL;
274 unsigned long prev = 0, pend = 0;
276 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
277 struct vm_area_struct *vma;
278 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
279 if (vma->vm_start < prev) {
280 pr_emerg("vm_start %lx < prev %lx\n",
281 vma->vm_start, prev);
284 if (vma->vm_start < pend) {
285 pr_emerg("vm_start %lx < pend %lx\n",
286 vma->vm_start, pend);
289 if (vma->vm_start > vma->vm_end) {
290 pr_emerg("vm_start %lx > vm_end %lx\n",
291 vma->vm_start, vma->vm_end);
294 spin_lock(&mm->page_table_lock);
295 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
296 pr_emerg("free gap %lx, correct %lx\n",
298 vma_compute_subtree_gap(vma));
301 spin_unlock(&mm->page_table_lock);
304 prev = vma->vm_start;
308 for (nd = pn; nd; nd = rb_prev(nd))
311 pr_emerg("backwards %d, forwards %d\n", j, i);
317 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
321 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
322 struct vm_area_struct *vma;
323 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
324 VM_BUG_ON_VMA(vma != ignore &&
325 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
330 static void validate_mm(struct mm_struct *mm)
334 unsigned long highest_address = 0;
335 struct vm_area_struct *vma = mm->mmap;
338 struct anon_vma *anon_vma = vma->anon_vma;
339 struct anon_vma_chain *avc;
342 anon_vma_lock_read(anon_vma);
343 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
344 anon_vma_interval_tree_verify(avc);
345 anon_vma_unlock_read(anon_vma);
348 highest_address = vma->vm_end;
352 if (i != mm->map_count) {
353 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
356 if (highest_address != mm->highest_vm_end) {
357 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
358 mm->highest_vm_end, highest_address);
362 if (i != mm->map_count) {
364 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
367 VM_BUG_ON_MM(bug, mm);
370 #define validate_mm_rb(root, ignore) do { } while (0)
371 #define validate_mm(mm) do { } while (0)
374 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
375 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
378 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
379 * vma->vm_prev->vm_end values changed, without modifying the vma's position
382 static void vma_gap_update(struct vm_area_struct *vma)
385 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
386 * function that does exacltly what we want.
388 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
391 static inline void vma_rb_insert(struct vm_area_struct *vma,
392 struct rb_root *root)
394 /* All rb_subtree_gap values must be consistent prior to insertion */
395 validate_mm_rb(root, NULL);
397 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
400 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
403 * All rb_subtree_gap values must be consistent prior to erase,
404 * with the possible exception of the vma being erased.
406 validate_mm_rb(root, vma);
409 * Note rb_erase_augmented is a fairly large inline function,
410 * so make sure we instantiate it only once with our desired
411 * augmented rbtree callbacks.
413 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
417 * vma has some anon_vma assigned, and is already inserted on that
418 * anon_vma's interval trees.
420 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
421 * vma must be removed from the anon_vma's interval trees using
422 * anon_vma_interval_tree_pre_update_vma().
424 * After the update, the vma will be reinserted using
425 * anon_vma_interval_tree_post_update_vma().
427 * The entire update must be protected by exclusive mmap_sem and by
428 * the root anon_vma's mutex.
431 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
433 struct anon_vma_chain *avc;
435 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
436 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
440 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
442 struct anon_vma_chain *avc;
444 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
445 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
448 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
449 unsigned long end, struct vm_area_struct **pprev,
450 struct rb_node ***rb_link, struct rb_node **rb_parent)
452 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
454 __rb_link = &mm->mm_rb.rb_node;
455 rb_prev = __rb_parent = NULL;
458 struct vm_area_struct *vma_tmp;
460 __rb_parent = *__rb_link;
461 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
463 if (vma_tmp->vm_end > addr) {
464 /* Fail if an existing vma overlaps the area */
465 if (vma_tmp->vm_start < end)
467 __rb_link = &__rb_parent->rb_left;
469 rb_prev = __rb_parent;
470 __rb_link = &__rb_parent->rb_right;
476 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
477 *rb_link = __rb_link;
478 *rb_parent = __rb_parent;
482 static unsigned long count_vma_pages_range(struct mm_struct *mm,
483 unsigned long addr, unsigned long end)
485 unsigned long nr_pages = 0;
486 struct vm_area_struct *vma;
488 /* Find first overlaping mapping */
489 vma = find_vma_intersection(mm, addr, end);
493 nr_pages = (min(end, vma->vm_end) -
494 max(addr, vma->vm_start)) >> PAGE_SHIFT;
496 /* Iterate over the rest of the overlaps */
497 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
498 unsigned long overlap_len;
500 if (vma->vm_start > end)
503 overlap_len = min(end, vma->vm_end) - vma->vm_start;
504 nr_pages += overlap_len >> PAGE_SHIFT;
510 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
511 struct rb_node **rb_link, struct rb_node *rb_parent)
513 /* Update tracking information for the gap following the new vma. */
515 vma_gap_update(vma->vm_next);
517 mm->highest_vm_end = vma->vm_end;
520 * vma->vm_prev wasn't known when we followed the rbtree to find the
521 * correct insertion point for that vma. As a result, we could not
522 * update the vma vm_rb parents rb_subtree_gap values on the way down.
523 * So, we first insert the vma with a zero rb_subtree_gap value
524 * (to be consistent with what we did on the way down), and then
525 * immediately update the gap to the correct value. Finally we
526 * rebalance the rbtree after all augmented values have been set.
528 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
529 vma->rb_subtree_gap = 0;
531 vma_rb_insert(vma, &mm->mm_rb);
534 static void __vma_link_file(struct vm_area_struct *vma)
540 struct address_space *mapping = file->f_mapping;
542 if (vma->vm_flags & VM_DENYWRITE)
543 atomic_dec(&file_inode(file)->i_writecount);
544 if (vma->vm_flags & VM_SHARED)
545 atomic_inc(&mapping->i_mmap_writable);
547 flush_dcache_mmap_lock(mapping);
548 vma_interval_tree_insert(vma, &mapping->i_mmap);
549 flush_dcache_mmap_unlock(mapping);
554 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
555 struct vm_area_struct *prev, struct rb_node **rb_link,
556 struct rb_node *rb_parent)
558 __vma_link_list(mm, vma, prev, rb_parent);
559 __vma_link_rb(mm, vma, rb_link, rb_parent);
562 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
563 struct vm_area_struct *prev, struct rb_node **rb_link,
564 struct rb_node *rb_parent)
566 struct address_space *mapping = NULL;
569 mapping = vma->vm_file->f_mapping;
570 i_mmap_lock_write(mapping);
573 __vma_link(mm, vma, prev, rb_link, rb_parent);
574 __vma_link_file(vma);
577 i_mmap_unlock_write(mapping);
584 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
585 * mm's list and rbtree. It has already been inserted into the interval tree.
587 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
589 struct vm_area_struct *prev;
590 struct rb_node **rb_link, *rb_parent;
592 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
593 &prev, &rb_link, &rb_parent))
595 __vma_link(mm, vma, prev, rb_link, rb_parent);
600 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
601 struct vm_area_struct *prev)
603 struct vm_area_struct *next;
605 vma_rb_erase(vma, &mm->mm_rb);
606 prev->vm_next = next = vma->vm_next;
608 next->vm_prev = prev;
611 vmacache_invalidate(mm);
615 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
616 * is already present in an i_mmap tree without adjusting the tree.
617 * The following helper function should be used when such adjustments
618 * are necessary. The "insert" vma (if any) is to be inserted
619 * before we drop the necessary locks.
621 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
622 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
624 struct mm_struct *mm = vma->vm_mm;
625 struct vm_area_struct *next = vma->vm_next;
626 struct vm_area_struct *importer = NULL;
627 struct address_space *mapping = NULL;
628 struct rb_root *root = NULL;
629 struct anon_vma *anon_vma = NULL;
630 struct file *file = vma->vm_file;
631 bool start_changed = false, end_changed = false;
632 long adjust_next = 0;
635 if (next && !insert) {
636 struct vm_area_struct *exporter = NULL;
638 if (end >= next->vm_end) {
640 * vma expands, overlapping all the next, and
641 * perhaps the one after too (mprotect case 6).
643 again: remove_next = 1 + (end > next->vm_end);
647 } else if (end > next->vm_start) {
649 * vma expands, overlapping part of the next:
650 * mprotect case 5 shifting the boundary up.
652 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
655 } else if (end < vma->vm_end) {
657 * vma shrinks, and !insert tells it's not
658 * split_vma inserting another: so it must be
659 * mprotect case 4 shifting the boundary down.
661 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
667 * Easily overlooked: when mprotect shifts the boundary,
668 * make sure the expanding vma has anon_vma set if the
669 * shrinking vma had, to cover any anon pages imported.
671 if (exporter && exporter->anon_vma && !importer->anon_vma) {
674 importer->anon_vma = exporter->anon_vma;
675 error = anon_vma_clone(importer, exporter);
682 mapping = file->f_mapping;
683 root = &mapping->i_mmap;
684 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
687 uprobe_munmap(next, next->vm_start, next->vm_end);
689 i_mmap_lock_write(mapping);
692 * Put into interval tree now, so instantiated pages
693 * are visible to arm/parisc __flush_dcache_page
694 * throughout; but we cannot insert into address
695 * space until vma start or end is updated.
697 __vma_link_file(insert);
701 vma_adjust_trans_huge(vma, start, end, adjust_next);
703 anon_vma = vma->anon_vma;
704 if (!anon_vma && adjust_next)
705 anon_vma = next->anon_vma;
707 VM_BUG_ON_VMA(adjust_next && next->anon_vma &&
708 anon_vma != next->anon_vma, next);
709 anon_vma_lock_write(anon_vma);
710 anon_vma_interval_tree_pre_update_vma(vma);
712 anon_vma_interval_tree_pre_update_vma(next);
716 flush_dcache_mmap_lock(mapping);
717 vma_interval_tree_remove(vma, root);
719 vma_interval_tree_remove(next, root);
722 if (start != vma->vm_start) {
723 vma->vm_start = start;
724 start_changed = true;
726 if (end != vma->vm_end) {
730 vma->vm_pgoff = pgoff;
732 next->vm_start += adjust_next << PAGE_SHIFT;
733 next->vm_pgoff += adjust_next;
738 vma_interval_tree_insert(next, root);
739 vma_interval_tree_insert(vma, root);
740 flush_dcache_mmap_unlock(mapping);
745 * vma_merge has merged next into vma, and needs
746 * us to remove next before dropping the locks.
748 __vma_unlink(mm, next, vma);
750 __remove_shared_vm_struct(next, file, mapping);
753 * split_vma has split insert from vma, and needs
754 * us to insert it before dropping the locks
755 * (it may either follow vma or precede it).
757 __insert_vm_struct(mm, insert);
763 mm->highest_vm_end = end;
764 else if (!adjust_next)
765 vma_gap_update(next);
770 anon_vma_interval_tree_post_update_vma(vma);
772 anon_vma_interval_tree_post_update_vma(next);
773 anon_vma_unlock_write(anon_vma);
776 i_mmap_unlock_write(mapping);
787 uprobe_munmap(next, next->vm_start, next->vm_end);
791 anon_vma_merge(vma, next);
793 mpol_put(vma_policy(next));
794 kmem_cache_free(vm_area_cachep, next);
796 * In mprotect's case 6 (see comments on vma_merge),
797 * we must remove another next too. It would clutter
798 * up the code too much to do both in one go.
801 if (remove_next == 2)
804 vma_gap_update(next);
806 mm->highest_vm_end = end;
817 * If the vma has a ->close operation then the driver probably needs to release
818 * per-vma resources, so we don't attempt to merge those.
820 static inline int is_mergeable_vma(struct vm_area_struct *vma,
821 struct file *file, unsigned long vm_flags,
822 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
825 * VM_SOFTDIRTY should not prevent from VMA merging, if we
826 * match the flags but dirty bit -- the caller should mark
827 * merged VMA as dirty. If dirty bit won't be excluded from
828 * comparison, we increase pressue on the memory system forcing
829 * the kernel to generate new VMAs when old one could be
832 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
834 if (vma->vm_file != file)
836 if (vma->vm_ops && vma->vm_ops->close)
838 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
843 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
844 struct anon_vma *anon_vma2,
845 struct vm_area_struct *vma)
848 * The list_is_singular() test is to avoid merging VMA cloned from
849 * parents. This can improve scalability caused by anon_vma lock.
851 if ((!anon_vma1 || !anon_vma2) && (!vma ||
852 list_is_singular(&vma->anon_vma_chain)))
854 return anon_vma1 == anon_vma2;
858 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
859 * in front of (at a lower virtual address and file offset than) the vma.
861 * We cannot merge two vmas if they have differently assigned (non-NULL)
862 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
864 * We don't check here for the merged mmap wrapping around the end of pagecache
865 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
866 * wrap, nor mmaps which cover the final page at index -1UL.
869 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
870 struct anon_vma *anon_vma, struct file *file,
872 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
874 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
875 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
876 if (vma->vm_pgoff == vm_pgoff)
883 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
884 * beyond (at a higher virtual address and file offset than) the vma.
886 * We cannot merge two vmas if they have differently assigned (non-NULL)
887 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
890 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
891 struct anon_vma *anon_vma, struct file *file,
893 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
895 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
896 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
898 vm_pglen = vma_pages(vma);
899 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
906 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
907 * whether that can be merged with its predecessor or its successor.
908 * Or both (it neatly fills a hole).
910 * In most cases - when called for mmap, brk or mremap - [addr,end) is
911 * certain not to be mapped by the time vma_merge is called; but when
912 * called for mprotect, it is certain to be already mapped (either at
913 * an offset within prev, or at the start of next), and the flags of
914 * this area are about to be changed to vm_flags - and the no-change
915 * case has already been eliminated.
917 * The following mprotect cases have to be considered, where AAAA is
918 * the area passed down from mprotect_fixup, never extending beyond one
919 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
921 * AAAA AAAA AAAA AAAA
922 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
923 * cannot merge might become might become might become
924 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
925 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
926 * mremap move: PPPPNNNNNNNN 8
928 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
929 * might become case 1 below case 2 below case 3 below
931 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
932 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
934 struct vm_area_struct *vma_merge(struct mm_struct *mm,
935 struct vm_area_struct *prev, unsigned long addr,
936 unsigned long end, unsigned long vm_flags,
937 struct anon_vma *anon_vma, struct file *file,
938 pgoff_t pgoff, struct mempolicy *policy,
939 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
941 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
942 struct vm_area_struct *area, *next;
946 * We later require that vma->vm_flags == vm_flags,
947 * so this tests vma->vm_flags & VM_SPECIAL, too.
949 if (vm_flags & VM_SPECIAL)
953 next = prev->vm_next;
957 if (next && next->vm_end == end) /* cases 6, 7, 8 */
958 next = next->vm_next;
961 * Can it merge with the predecessor?
963 if (prev && prev->vm_end == addr &&
964 mpol_equal(vma_policy(prev), policy) &&
965 can_vma_merge_after(prev, vm_flags,
966 anon_vma, file, pgoff,
967 vm_userfaultfd_ctx)) {
969 * OK, it can. Can we now merge in the successor as well?
971 if (next && end == next->vm_start &&
972 mpol_equal(policy, vma_policy(next)) &&
973 can_vma_merge_before(next, vm_flags,
976 vm_userfaultfd_ctx) &&
977 is_mergeable_anon_vma(prev->anon_vma,
978 next->anon_vma, NULL)) {
980 err = vma_adjust(prev, prev->vm_start,
981 next->vm_end, prev->vm_pgoff, NULL);
982 } else /* cases 2, 5, 7 */
983 err = vma_adjust(prev, prev->vm_start,
984 end, prev->vm_pgoff, NULL);
987 khugepaged_enter_vma_merge(prev, vm_flags);
992 * Can this new request be merged in front of next?
994 if (next && end == next->vm_start &&
995 mpol_equal(policy, vma_policy(next)) &&
996 can_vma_merge_before(next, vm_flags,
997 anon_vma, file, pgoff+pglen,
998 vm_userfaultfd_ctx)) {
999 if (prev && addr < prev->vm_end) /* case 4 */
1000 err = vma_adjust(prev, prev->vm_start,
1001 addr, prev->vm_pgoff, NULL);
1002 else /* cases 3, 8 */
1003 err = vma_adjust(area, addr, next->vm_end,
1004 next->vm_pgoff - pglen, NULL);
1007 khugepaged_enter_vma_merge(area, vm_flags);
1015 * Rough compatbility check to quickly see if it's even worth looking
1016 * at sharing an anon_vma.
1018 * They need to have the same vm_file, and the flags can only differ
1019 * in things that mprotect may change.
1021 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1022 * we can merge the two vma's. For example, we refuse to merge a vma if
1023 * there is a vm_ops->close() function, because that indicates that the
1024 * driver is doing some kind of reference counting. But that doesn't
1025 * really matter for the anon_vma sharing case.
1027 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1029 return a->vm_end == b->vm_start &&
1030 mpol_equal(vma_policy(a), vma_policy(b)) &&
1031 a->vm_file == b->vm_file &&
1032 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1033 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1037 * Do some basic sanity checking to see if we can re-use the anon_vma
1038 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1039 * the same as 'old', the other will be the new one that is trying
1040 * to share the anon_vma.
1042 * NOTE! This runs with mm_sem held for reading, so it is possible that
1043 * the anon_vma of 'old' is concurrently in the process of being set up
1044 * by another page fault trying to merge _that_. But that's ok: if it
1045 * is being set up, that automatically means that it will be a singleton
1046 * acceptable for merging, so we can do all of this optimistically. But
1047 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1049 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1050 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1051 * is to return an anon_vma that is "complex" due to having gone through
1054 * We also make sure that the two vma's are compatible (adjacent,
1055 * and with the same memory policies). That's all stable, even with just
1056 * a read lock on the mm_sem.
1058 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1060 if (anon_vma_compatible(a, b)) {
1061 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1063 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1070 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1071 * neighbouring vmas for a suitable anon_vma, before it goes off
1072 * to allocate a new anon_vma. It checks because a repetitive
1073 * sequence of mprotects and faults may otherwise lead to distinct
1074 * anon_vmas being allocated, preventing vma merge in subsequent
1077 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1079 struct anon_vma *anon_vma;
1080 struct vm_area_struct *near;
1082 near = vma->vm_next;
1086 anon_vma = reusable_anon_vma(near, vma, near);
1090 near = vma->vm_prev;
1094 anon_vma = reusable_anon_vma(near, near, vma);
1099 * There's no absolute need to look only at touching neighbours:
1100 * we could search further afield for "compatible" anon_vmas.
1101 * But it would probably just be a waste of time searching,
1102 * or lead to too many vmas hanging off the same anon_vma.
1103 * We're trying to allow mprotect remerging later on,
1104 * not trying to minimize memory used for anon_vmas.
1110 * If a hint addr is less than mmap_min_addr change hint to be as
1111 * low as possible but still greater than mmap_min_addr
1113 static inline unsigned long round_hint_to_min(unsigned long hint)
1116 if (((void *)hint != NULL) &&
1117 (hint < mmap_min_addr))
1118 return PAGE_ALIGN(mmap_min_addr);
1122 static inline int mlock_future_check(struct mm_struct *mm,
1123 unsigned long flags,
1126 unsigned long locked, lock_limit;
1128 /* mlock MCL_FUTURE? */
1129 if (flags & VM_LOCKED) {
1130 locked = len >> PAGE_SHIFT;
1131 locked += mm->locked_vm;
1132 lock_limit = rlimit(RLIMIT_MEMLOCK);
1133 lock_limit >>= PAGE_SHIFT;
1134 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1141 * The caller must hold down_write(¤t->mm->mmap_sem).
1143 unsigned long do_mmap(struct file *file, unsigned long addr,
1144 unsigned long len, unsigned long prot,
1145 unsigned long flags, vm_flags_t vm_flags,
1146 unsigned long pgoff, unsigned long *populate)
1148 struct mm_struct *mm = current->mm;
1157 * Does the application expect PROT_READ to imply PROT_EXEC?
1159 * (the exception is when the underlying filesystem is noexec
1160 * mounted, in which case we dont add PROT_EXEC.)
1162 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1163 if (!(file && path_noexec(&file->f_path)))
1166 if (!(flags & MAP_FIXED))
1167 addr = round_hint_to_min(addr);
1169 /* Careful about overflows.. */
1170 len = PAGE_ALIGN(len);
1174 /* offset overflow? */
1175 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1178 /* Too many mappings? */
1179 if (mm->map_count > sysctl_max_map_count)
1182 /* Obtain the address to map to. we verify (or select) it and ensure
1183 * that it represents a valid section of the address space.
1185 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1186 if (offset_in_page(addr))
1189 if (prot == PROT_EXEC) {
1190 pkey = execute_only_pkey(mm);
1195 /* Do simple checking here so the lower-level routines won't have
1196 * to. we assume access permissions have been handled by the open
1197 * of the memory object, so we don't do any here.
1199 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1200 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1202 if (flags & MAP_LOCKED)
1203 if (!can_do_mlock())
1206 if (mlock_future_check(mm, vm_flags, len))
1210 struct inode *inode = file_inode(file);
1212 switch (flags & MAP_TYPE) {
1214 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1218 * Make sure we don't allow writing to an append-only
1221 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1225 * Make sure there are no mandatory locks on the file.
1227 if (locks_verify_locked(file))
1230 vm_flags |= VM_SHARED | VM_MAYSHARE;
1231 if (!(file->f_mode & FMODE_WRITE))
1232 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1236 if (!(file->f_mode & FMODE_READ))
1238 if (path_noexec(&file->f_path)) {
1239 if (vm_flags & VM_EXEC)
1241 vm_flags &= ~VM_MAYEXEC;
1244 if (!file->f_op->mmap)
1246 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1254 switch (flags & MAP_TYPE) {
1256 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1262 vm_flags |= VM_SHARED | VM_MAYSHARE;
1266 * Set pgoff according to addr for anon_vma.
1268 pgoff = addr >> PAGE_SHIFT;
1276 * Set 'VM_NORESERVE' if we should not account for the
1277 * memory use of this mapping.
1279 if (flags & MAP_NORESERVE) {
1280 /* We honor MAP_NORESERVE if allowed to overcommit */
1281 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1282 vm_flags |= VM_NORESERVE;
1284 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1285 if (file && is_file_hugepages(file))
1286 vm_flags |= VM_NORESERVE;
1289 addr = mmap_region(file, addr, len, vm_flags, pgoff);
1290 if (!IS_ERR_VALUE(addr) &&
1291 ((vm_flags & VM_LOCKED) ||
1292 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1297 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1298 unsigned long, prot, unsigned long, flags,
1299 unsigned long, fd, unsigned long, pgoff)
1301 struct file *file = NULL;
1302 unsigned long retval;
1304 if (!(flags & MAP_ANONYMOUS)) {
1305 audit_mmap_fd(fd, flags);
1309 if (is_file_hugepages(file))
1310 len = ALIGN(len, huge_page_size(hstate_file(file)));
1312 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1314 } else if (flags & MAP_HUGETLB) {
1315 struct user_struct *user = NULL;
1318 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1322 len = ALIGN(len, huge_page_size(hs));
1324 * VM_NORESERVE is used because the reservations will be
1325 * taken when vm_ops->mmap() is called
1326 * A dummy user value is used because we are not locking
1327 * memory so no accounting is necessary
1329 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1331 &user, HUGETLB_ANONHUGE_INODE,
1332 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1334 return PTR_ERR(file);
1337 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1339 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1346 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1347 struct mmap_arg_struct {
1351 unsigned long flags;
1353 unsigned long offset;
1356 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1358 struct mmap_arg_struct a;
1360 if (copy_from_user(&a, arg, sizeof(a)))
1362 if (offset_in_page(a.offset))
1365 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1366 a.offset >> PAGE_SHIFT);
1368 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1371 * Some shared mappigns will want the pages marked read-only
1372 * to track write events. If so, we'll downgrade vm_page_prot
1373 * to the private version (using protection_map[] without the
1376 int vma_wants_writenotify(struct vm_area_struct *vma)
1378 vm_flags_t vm_flags = vma->vm_flags;
1379 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1381 /* If it was private or non-writable, the write bit is already clear */
1382 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1385 /* The backer wishes to know when pages are first written to? */
1386 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1389 /* The open routine did something to the protections that pgprot_modify
1390 * won't preserve? */
1391 if (pgprot_val(vma->vm_page_prot) !=
1392 pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags)))
1395 /* Do we need to track softdirty? */
1396 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1399 /* Specialty mapping? */
1400 if (vm_flags & VM_PFNMAP)
1403 /* Can the mapping track the dirty pages? */
1404 return vma->vm_file && vma->vm_file->f_mapping &&
1405 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1409 * We account for memory if it's a private writeable mapping,
1410 * not hugepages and VM_NORESERVE wasn't set.
1412 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1415 * hugetlb has its own accounting separate from the core VM
1416 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1418 if (file && is_file_hugepages(file))
1421 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1424 unsigned long mmap_region(struct file *file, unsigned long addr,
1425 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1427 struct mm_struct *mm = current->mm;
1428 struct vm_area_struct *vma, *prev;
1430 struct rb_node **rb_link, *rb_parent;
1431 unsigned long charged = 0;
1433 /* Check against address space limit. */
1434 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1435 unsigned long nr_pages;
1438 * MAP_FIXED may remove pages of mappings that intersects with
1439 * requested mapping. Account for the pages it would unmap.
1441 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1443 if (!may_expand_vm(mm, vm_flags,
1444 (len >> PAGE_SHIFT) - nr_pages))
1448 /* Clear old maps */
1449 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1451 if (do_munmap(mm, addr, len))
1456 * Private writable mapping: check memory availability
1458 if (accountable_mapping(file, vm_flags)) {
1459 charged = len >> PAGE_SHIFT;
1460 if (security_vm_enough_memory_mm(mm, charged))
1462 vm_flags |= VM_ACCOUNT;
1466 * Can we just expand an old mapping?
1468 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1469 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1474 * Determine the object being mapped and call the appropriate
1475 * specific mapper. the address has already been validated, but
1476 * not unmapped, but the maps are removed from the list.
1478 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1485 vma->vm_start = addr;
1486 vma->vm_end = addr + len;
1487 vma->vm_flags = vm_flags;
1488 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1489 vma->vm_pgoff = pgoff;
1490 INIT_LIST_HEAD(&vma->anon_vma_chain);
1493 if (vm_flags & VM_DENYWRITE) {
1494 error = deny_write_access(file);
1498 if (vm_flags & VM_SHARED) {
1499 error = mapping_map_writable(file->f_mapping);
1501 goto allow_write_and_free_vma;
1504 /* ->mmap() can change vma->vm_file, but must guarantee that
1505 * vma_link() below can deny write-access if VM_DENYWRITE is set
1506 * and map writably if VM_SHARED is set. This usually means the
1507 * new file must not have been exposed to user-space, yet.
1509 vma->vm_file = get_file(file);
1510 error = file->f_op->mmap(file, vma);
1512 goto unmap_and_free_vma;
1514 /* Can addr have changed??
1516 * Answer: Yes, several device drivers can do it in their
1517 * f_op->mmap method. -DaveM
1518 * Bug: If addr is changed, prev, rb_link, rb_parent should
1519 * be updated for vma_link()
1521 WARN_ON_ONCE(addr != vma->vm_start);
1523 addr = vma->vm_start;
1524 vm_flags = vma->vm_flags;
1525 } else if (vm_flags & VM_SHARED) {
1526 error = shmem_zero_setup(vma);
1531 vma_link(mm, vma, prev, rb_link, rb_parent);
1532 /* Once vma denies write, undo our temporary denial count */
1534 if (vm_flags & VM_SHARED)
1535 mapping_unmap_writable(file->f_mapping);
1536 if (vm_flags & VM_DENYWRITE)
1537 allow_write_access(file);
1539 file = vma->vm_file;
1541 perf_event_mmap(vma);
1543 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1544 if (vm_flags & VM_LOCKED) {
1545 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1546 vma == get_gate_vma(current->mm)))
1547 mm->locked_vm += (len >> PAGE_SHIFT);
1549 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1556 * New (or expanded) vma always get soft dirty status.
1557 * Otherwise user-space soft-dirty page tracker won't
1558 * be able to distinguish situation when vma area unmapped,
1559 * then new mapped in-place (which must be aimed as
1560 * a completely new data area).
1562 vma->vm_flags |= VM_SOFTDIRTY;
1564 vma_set_page_prot(vma);
1569 vma->vm_file = NULL;
1572 /* Undo any partial mapping done by a device driver. */
1573 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1575 if (vm_flags & VM_SHARED)
1576 mapping_unmap_writable(file->f_mapping);
1577 allow_write_and_free_vma:
1578 if (vm_flags & VM_DENYWRITE)
1579 allow_write_access(file);
1581 kmem_cache_free(vm_area_cachep, vma);
1584 vm_unacct_memory(charged);
1588 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1591 * We implement the search by looking for an rbtree node that
1592 * immediately follows a suitable gap. That is,
1593 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1594 * - gap_end = vma->vm_start >= info->low_limit + length;
1595 * - gap_end - gap_start >= length
1598 struct mm_struct *mm = current->mm;
1599 struct vm_area_struct *vma;
1600 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1602 /* Adjust search length to account for worst case alignment overhead */
1603 length = info->length + info->align_mask;
1604 if (length < info->length)
1607 /* Adjust search limits by the desired length */
1608 if (info->high_limit < length)
1610 high_limit = info->high_limit - length;
1612 if (info->low_limit > high_limit)
1614 low_limit = info->low_limit + length;
1616 /* Check if rbtree root looks promising */
1617 if (RB_EMPTY_ROOT(&mm->mm_rb))
1619 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1620 if (vma->rb_subtree_gap < length)
1624 /* Visit left subtree if it looks promising */
1625 gap_end = vma->vm_start;
1626 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1627 struct vm_area_struct *left =
1628 rb_entry(vma->vm_rb.rb_left,
1629 struct vm_area_struct, vm_rb);
1630 if (left->rb_subtree_gap >= length) {
1636 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1638 /* Check if current node has a suitable gap */
1639 if (gap_start > high_limit)
1641 if (gap_end >= low_limit && gap_end - gap_start >= length)
1644 /* Visit right subtree if it looks promising */
1645 if (vma->vm_rb.rb_right) {
1646 struct vm_area_struct *right =
1647 rb_entry(vma->vm_rb.rb_right,
1648 struct vm_area_struct, vm_rb);
1649 if (right->rb_subtree_gap >= length) {
1655 /* Go back up the rbtree to find next candidate node */
1657 struct rb_node *prev = &vma->vm_rb;
1658 if (!rb_parent(prev))
1660 vma = rb_entry(rb_parent(prev),
1661 struct vm_area_struct, vm_rb);
1662 if (prev == vma->vm_rb.rb_left) {
1663 gap_start = vma->vm_prev->vm_end;
1664 gap_end = vma->vm_start;
1671 /* Check highest gap, which does not precede any rbtree node */
1672 gap_start = mm->highest_vm_end;
1673 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1674 if (gap_start > high_limit)
1678 /* We found a suitable gap. Clip it with the original low_limit. */
1679 if (gap_start < info->low_limit)
1680 gap_start = info->low_limit;
1682 /* Adjust gap address to the desired alignment */
1683 gap_start += (info->align_offset - gap_start) & info->align_mask;
1685 VM_BUG_ON(gap_start + info->length > info->high_limit);
1686 VM_BUG_ON(gap_start + info->length > gap_end);
1690 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1692 struct mm_struct *mm = current->mm;
1693 struct vm_area_struct *vma;
1694 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1696 /* Adjust search length to account for worst case alignment overhead */
1697 length = info->length + info->align_mask;
1698 if (length < info->length)
1702 * Adjust search limits by the desired length.
1703 * See implementation comment at top of unmapped_area().
1705 gap_end = info->high_limit;
1706 if (gap_end < length)
1708 high_limit = gap_end - length;
1710 if (info->low_limit > high_limit)
1712 low_limit = info->low_limit + length;
1714 /* Check highest gap, which does not precede any rbtree node */
1715 gap_start = mm->highest_vm_end;
1716 if (gap_start <= high_limit)
1719 /* Check if rbtree root looks promising */
1720 if (RB_EMPTY_ROOT(&mm->mm_rb))
1722 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1723 if (vma->rb_subtree_gap < length)
1727 /* Visit right subtree if it looks promising */
1728 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1729 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1730 struct vm_area_struct *right =
1731 rb_entry(vma->vm_rb.rb_right,
1732 struct vm_area_struct, vm_rb);
1733 if (right->rb_subtree_gap >= length) {
1740 /* Check if current node has a suitable gap */
1741 gap_end = vma->vm_start;
1742 if (gap_end < low_limit)
1744 if (gap_start <= high_limit && gap_end - gap_start >= length)
1747 /* Visit left subtree if it looks promising */
1748 if (vma->vm_rb.rb_left) {
1749 struct vm_area_struct *left =
1750 rb_entry(vma->vm_rb.rb_left,
1751 struct vm_area_struct, vm_rb);
1752 if (left->rb_subtree_gap >= length) {
1758 /* Go back up the rbtree to find next candidate node */
1760 struct rb_node *prev = &vma->vm_rb;
1761 if (!rb_parent(prev))
1763 vma = rb_entry(rb_parent(prev),
1764 struct vm_area_struct, vm_rb);
1765 if (prev == vma->vm_rb.rb_right) {
1766 gap_start = vma->vm_prev ?
1767 vma->vm_prev->vm_end : 0;
1774 /* We found a suitable gap. Clip it with the original high_limit. */
1775 if (gap_end > info->high_limit)
1776 gap_end = info->high_limit;
1779 /* Compute highest gap address at the desired alignment */
1780 gap_end -= info->length;
1781 gap_end -= (gap_end - info->align_offset) & info->align_mask;
1783 VM_BUG_ON(gap_end < info->low_limit);
1784 VM_BUG_ON(gap_end < gap_start);
1788 /* Get an address range which is currently unmapped.
1789 * For shmat() with addr=0.
1791 * Ugly calling convention alert:
1792 * Return value with the low bits set means error value,
1794 * if (ret & ~PAGE_MASK)
1797 * This function "knows" that -ENOMEM has the bits set.
1799 #ifndef HAVE_ARCH_UNMAPPED_AREA
1801 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1802 unsigned long len, unsigned long pgoff, unsigned long flags)
1804 struct mm_struct *mm = current->mm;
1805 struct vm_area_struct *vma;
1806 struct vm_unmapped_area_info info;
1808 if (len > TASK_SIZE - mmap_min_addr)
1811 if (flags & MAP_FIXED)
1815 addr = PAGE_ALIGN(addr);
1816 vma = find_vma(mm, addr);
1817 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1818 (!vma || addr + len <= vma->vm_start))
1824 info.low_limit = mm->mmap_base;
1825 info.high_limit = TASK_SIZE;
1826 info.align_mask = 0;
1827 return vm_unmapped_area(&info);
1832 * This mmap-allocator allocates new areas top-down from below the
1833 * stack's low limit (the base):
1835 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1837 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1838 const unsigned long len, const unsigned long pgoff,
1839 const unsigned long flags)
1841 struct vm_area_struct *vma;
1842 struct mm_struct *mm = current->mm;
1843 unsigned long addr = addr0;
1844 struct vm_unmapped_area_info info;
1846 /* requested length too big for entire address space */
1847 if (len > TASK_SIZE - mmap_min_addr)
1850 if (flags & MAP_FIXED)
1853 /* requesting a specific address */
1855 addr = PAGE_ALIGN(addr);
1856 vma = find_vma(mm, addr);
1857 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1858 (!vma || addr + len <= vma->vm_start))
1862 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1864 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1865 info.high_limit = mm->mmap_base;
1866 info.align_mask = 0;
1867 addr = vm_unmapped_area(&info);
1870 * A failed mmap() very likely causes application failure,
1871 * so fall back to the bottom-up function here. This scenario
1872 * can happen with large stack limits and large mmap()
1875 if (offset_in_page(addr)) {
1876 VM_BUG_ON(addr != -ENOMEM);
1878 info.low_limit = TASK_UNMAPPED_BASE;
1879 info.high_limit = TASK_SIZE;
1880 addr = vm_unmapped_area(&info);
1888 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1889 unsigned long pgoff, unsigned long flags)
1891 unsigned long (*get_area)(struct file *, unsigned long,
1892 unsigned long, unsigned long, unsigned long);
1894 unsigned long error = arch_mmap_check(addr, len, flags);
1898 /* Careful about overflows.. */
1899 if (len > TASK_SIZE)
1902 get_area = current->mm->get_unmapped_area;
1903 if (file && file->f_op->get_unmapped_area)
1904 get_area = file->f_op->get_unmapped_area;
1905 addr = get_area(file, addr, len, pgoff, flags);
1906 if (IS_ERR_VALUE(addr))
1909 if (addr > TASK_SIZE - len)
1911 if (offset_in_page(addr))
1914 addr = arch_rebalance_pgtables(addr, len);
1915 error = security_mmap_addr(addr);
1916 return error ? error : addr;
1919 EXPORT_SYMBOL(get_unmapped_area);
1921 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1922 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1924 struct rb_node *rb_node;
1925 struct vm_area_struct *vma;
1927 /* Check the cache first. */
1928 vma = vmacache_find(mm, addr);
1932 rb_node = mm->mm_rb.rb_node;
1935 struct vm_area_struct *tmp;
1937 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1939 if (tmp->vm_end > addr) {
1941 if (tmp->vm_start <= addr)
1943 rb_node = rb_node->rb_left;
1945 rb_node = rb_node->rb_right;
1949 vmacache_update(addr, vma);
1953 EXPORT_SYMBOL(find_vma);
1956 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1958 struct vm_area_struct *
1959 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1960 struct vm_area_struct **pprev)
1962 struct vm_area_struct *vma;
1964 vma = find_vma(mm, addr);
1966 *pprev = vma->vm_prev;
1968 struct rb_node *rb_node = mm->mm_rb.rb_node;
1971 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1972 rb_node = rb_node->rb_right;
1979 * Verify that the stack growth is acceptable and
1980 * update accounting. This is shared with both the
1981 * grow-up and grow-down cases.
1983 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1985 struct mm_struct *mm = vma->vm_mm;
1986 struct rlimit *rlim = current->signal->rlim;
1987 unsigned long new_start, actual_size;
1989 /* address space limit tests */
1990 if (!may_expand_vm(mm, vma->vm_flags, grow))
1993 /* Stack limit test */
1995 if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN)))
1996 actual_size -= PAGE_SIZE;
1997 if (actual_size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2000 /* mlock limit tests */
2001 if (vma->vm_flags & VM_LOCKED) {
2002 unsigned long locked;
2003 unsigned long limit;
2004 locked = mm->locked_vm + grow;
2005 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2006 limit >>= PAGE_SHIFT;
2007 if (locked > limit && !capable(CAP_IPC_LOCK))
2011 /* Check to ensure the stack will not grow into a hugetlb-only region */
2012 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2014 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2018 * Overcommit.. This must be the final test, as it will
2019 * update security statistics.
2021 if (security_vm_enough_memory_mm(mm, grow))
2027 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2029 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2030 * vma is the last one with address > vma->vm_end. Have to extend vma.
2032 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2034 struct mm_struct *mm = vma->vm_mm;
2037 if (!(vma->vm_flags & VM_GROWSUP))
2040 /* Guard against wrapping around to address 0. */
2041 if (address < PAGE_ALIGN(address+4))
2042 address = PAGE_ALIGN(address+4);
2046 /* We must make sure the anon_vma is allocated. */
2047 if (unlikely(anon_vma_prepare(vma)))
2051 * vma->vm_start/vm_end cannot change under us because the caller
2052 * is required to hold the mmap_sem in read mode. We need the
2053 * anon_vma lock to serialize against concurrent expand_stacks.
2055 anon_vma_lock_write(vma->anon_vma);
2057 /* Somebody else might have raced and expanded it already */
2058 if (address > vma->vm_end) {
2059 unsigned long size, grow;
2061 size = address - vma->vm_start;
2062 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2065 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2066 error = acct_stack_growth(vma, size, grow);
2069 * vma_gap_update() doesn't support concurrent
2070 * updates, but we only hold a shared mmap_sem
2071 * lock here, so we need to protect against
2072 * concurrent vma expansions.
2073 * anon_vma_lock_write() doesn't help here, as
2074 * we don't guarantee that all growable vmas
2075 * in a mm share the same root anon vma.
2076 * So, we reuse mm->page_table_lock to guard
2077 * against concurrent vma expansions.
2079 spin_lock(&mm->page_table_lock);
2080 if (vma->vm_flags & VM_LOCKED)
2081 mm->locked_vm += grow;
2082 vm_stat_account(mm, vma->vm_flags, grow);
2083 anon_vma_interval_tree_pre_update_vma(vma);
2084 vma->vm_end = address;
2085 anon_vma_interval_tree_post_update_vma(vma);
2087 vma_gap_update(vma->vm_next);
2089 mm->highest_vm_end = address;
2090 spin_unlock(&mm->page_table_lock);
2092 perf_event_mmap(vma);
2096 anon_vma_unlock_write(vma->anon_vma);
2097 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2101 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2104 * vma is the first one with address < vma->vm_start. Have to extend vma.
2106 int expand_downwards(struct vm_area_struct *vma,
2107 unsigned long address)
2109 struct mm_struct *mm = vma->vm_mm;
2112 address &= PAGE_MASK;
2113 error = security_mmap_addr(address);
2117 /* We must make sure the anon_vma is allocated. */
2118 if (unlikely(anon_vma_prepare(vma)))
2122 * vma->vm_start/vm_end cannot change under us because the caller
2123 * is required to hold the mmap_sem in read mode. We need the
2124 * anon_vma lock to serialize against concurrent expand_stacks.
2126 anon_vma_lock_write(vma->anon_vma);
2128 /* Somebody else might have raced and expanded it already */
2129 if (address < vma->vm_start) {
2130 unsigned long size, grow;
2132 size = vma->vm_end - address;
2133 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2136 if (grow <= vma->vm_pgoff) {
2137 error = acct_stack_growth(vma, size, grow);
2140 * vma_gap_update() doesn't support concurrent
2141 * updates, but we only hold a shared mmap_sem
2142 * lock here, so we need to protect against
2143 * concurrent vma expansions.
2144 * anon_vma_lock_write() doesn't help here, as
2145 * we don't guarantee that all growable vmas
2146 * in a mm share the same root anon vma.
2147 * So, we reuse mm->page_table_lock to guard
2148 * against concurrent vma expansions.
2150 spin_lock(&mm->page_table_lock);
2151 if (vma->vm_flags & VM_LOCKED)
2152 mm->locked_vm += grow;
2153 vm_stat_account(mm, vma->vm_flags, grow);
2154 anon_vma_interval_tree_pre_update_vma(vma);
2155 vma->vm_start = address;
2156 vma->vm_pgoff -= grow;
2157 anon_vma_interval_tree_post_update_vma(vma);
2158 vma_gap_update(vma);
2159 spin_unlock(&mm->page_table_lock);
2161 perf_event_mmap(vma);
2165 anon_vma_unlock_write(vma->anon_vma);
2166 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2172 * Note how expand_stack() refuses to expand the stack all the way to
2173 * abut the next virtual mapping, *unless* that mapping itself is also
2174 * a stack mapping. We want to leave room for a guard page, after all
2175 * (the guard page itself is not added here, that is done by the
2176 * actual page faulting logic)
2178 * This matches the behavior of the guard page logic (see mm/memory.c:
2179 * check_stack_guard_page()), which only allows the guard page to be
2180 * removed under these circumstances.
2182 #ifdef CONFIG_STACK_GROWSUP
2183 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2185 struct vm_area_struct *next;
2187 address &= PAGE_MASK;
2188 next = vma->vm_next;
2189 if (next && next->vm_start == address + PAGE_SIZE) {
2190 if (!(next->vm_flags & VM_GROWSUP))
2193 return expand_upwards(vma, address);
2196 struct vm_area_struct *
2197 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2199 struct vm_area_struct *vma, *prev;
2202 vma = find_vma_prev(mm, addr, &prev);
2203 if (vma && (vma->vm_start <= addr))
2205 if (!prev || expand_stack(prev, addr))
2207 if (prev->vm_flags & VM_LOCKED)
2208 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2212 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2214 struct vm_area_struct *prev;
2216 address &= PAGE_MASK;
2217 prev = vma->vm_prev;
2218 if (prev && prev->vm_end == address) {
2219 if (!(prev->vm_flags & VM_GROWSDOWN))
2222 return expand_downwards(vma, address);
2225 struct vm_area_struct *
2226 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2228 struct vm_area_struct *vma;
2229 unsigned long start;
2232 vma = find_vma(mm, addr);
2235 if (vma->vm_start <= addr)
2237 if (!(vma->vm_flags & VM_GROWSDOWN))
2239 start = vma->vm_start;
2240 if (expand_stack(vma, addr))
2242 if (vma->vm_flags & VM_LOCKED)
2243 populate_vma_page_range(vma, addr, start, NULL);
2248 EXPORT_SYMBOL_GPL(find_extend_vma);
2251 * Ok - we have the memory areas we should free on the vma list,
2252 * so release them, and do the vma updates.
2254 * Called with the mm semaphore held.
2256 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2258 unsigned long nr_accounted = 0;
2260 /* Update high watermark before we lower total_vm */
2261 update_hiwater_vm(mm);
2263 long nrpages = vma_pages(vma);
2265 if (vma->vm_flags & VM_ACCOUNT)
2266 nr_accounted += nrpages;
2267 vm_stat_account(mm, vma->vm_flags, -nrpages);
2268 vma = remove_vma(vma);
2270 vm_unacct_memory(nr_accounted);
2275 * Get rid of page table information in the indicated region.
2277 * Called with the mm semaphore held.
2279 static void unmap_region(struct mm_struct *mm,
2280 struct vm_area_struct *vma, struct vm_area_struct *prev,
2281 unsigned long start, unsigned long end)
2283 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2284 struct mmu_gather tlb;
2287 tlb_gather_mmu(&tlb, mm, start, end);
2288 update_hiwater_rss(mm);
2289 unmap_vmas(&tlb, vma, start, end);
2290 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2291 next ? next->vm_start : USER_PGTABLES_CEILING);
2292 tlb_finish_mmu(&tlb, start, end);
2296 * Create a list of vma's touched by the unmap, removing them from the mm's
2297 * vma list as we go..
2300 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2301 struct vm_area_struct *prev, unsigned long end)
2303 struct vm_area_struct **insertion_point;
2304 struct vm_area_struct *tail_vma = NULL;
2306 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2307 vma->vm_prev = NULL;
2309 vma_rb_erase(vma, &mm->mm_rb);
2313 } while (vma && vma->vm_start < end);
2314 *insertion_point = vma;
2316 vma->vm_prev = prev;
2317 vma_gap_update(vma);
2319 mm->highest_vm_end = prev ? prev->vm_end : 0;
2320 tail_vma->vm_next = NULL;
2322 /* Kill the cache */
2323 vmacache_invalidate(mm);
2327 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2328 * munmap path where it doesn't make sense to fail.
2330 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2331 unsigned long addr, int new_below)
2333 struct vm_area_struct *new;
2336 if (is_vm_hugetlb_page(vma) && (addr &
2337 ~(huge_page_mask(hstate_vma(vma)))))
2340 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2344 /* most fields are the same, copy all, and then fixup */
2347 INIT_LIST_HEAD(&new->anon_vma_chain);
2352 new->vm_start = addr;
2353 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2356 err = vma_dup_policy(vma, new);
2360 err = anon_vma_clone(new, vma);
2365 get_file(new->vm_file);
2367 if (new->vm_ops && new->vm_ops->open)
2368 new->vm_ops->open(new);
2371 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2372 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2374 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2380 /* Clean everything up if vma_adjust failed. */
2381 if (new->vm_ops && new->vm_ops->close)
2382 new->vm_ops->close(new);
2385 unlink_anon_vmas(new);
2387 mpol_put(vma_policy(new));
2389 kmem_cache_free(vm_area_cachep, new);
2394 * Split a vma into two pieces at address 'addr', a new vma is allocated
2395 * either for the first part or the tail.
2397 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2398 unsigned long addr, int new_below)
2400 if (mm->map_count >= sysctl_max_map_count)
2403 return __split_vma(mm, vma, addr, new_below);
2406 /* Munmap is split into 2 main parts -- this part which finds
2407 * what needs doing, and the areas themselves, which do the
2408 * work. This now handles partial unmappings.
2409 * Jeremy Fitzhardinge <jeremy@goop.org>
2411 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2414 struct vm_area_struct *vma, *prev, *last;
2416 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2419 len = PAGE_ALIGN(len);
2423 /* Find the first overlapping VMA */
2424 vma = find_vma(mm, start);
2427 prev = vma->vm_prev;
2428 /* we have start < vma->vm_end */
2430 /* if it doesn't overlap, we have nothing.. */
2432 if (vma->vm_start >= end)
2436 * If we need to split any vma, do it now to save pain later.
2438 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2439 * unmapped vm_area_struct will remain in use: so lower split_vma
2440 * places tmp vma above, and higher split_vma places tmp vma below.
2442 if (start > vma->vm_start) {
2446 * Make sure that map_count on return from munmap() will
2447 * not exceed its limit; but let map_count go just above
2448 * its limit temporarily, to help free resources as expected.
2450 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2453 error = __split_vma(mm, vma, start, 0);
2459 /* Does it split the last one? */
2460 last = find_vma(mm, end);
2461 if (last && end > last->vm_start) {
2462 int error = __split_vma(mm, last, end, 1);
2466 vma = prev ? prev->vm_next : mm->mmap;
2469 * unlock any mlock()ed ranges before detaching vmas
2471 if (mm->locked_vm) {
2472 struct vm_area_struct *tmp = vma;
2473 while (tmp && tmp->vm_start < end) {
2474 if (tmp->vm_flags & VM_LOCKED) {
2475 mm->locked_vm -= vma_pages(tmp);
2476 munlock_vma_pages_all(tmp);
2483 * Remove the vma's, and unmap the actual pages
2485 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2486 unmap_region(mm, vma, prev, start, end);
2488 arch_unmap(mm, vma, start, end);
2490 /* Fix up all other VM information */
2491 remove_vma_list(mm, vma);
2496 int vm_munmap(unsigned long start, size_t len)
2499 struct mm_struct *mm = current->mm;
2501 down_write(&mm->mmap_sem);
2502 ret = do_munmap(mm, start, len);
2503 up_write(&mm->mmap_sem);
2506 EXPORT_SYMBOL(vm_munmap);
2508 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2510 profile_munmap(addr);
2511 return vm_munmap(addr, len);
2516 * Emulation of deprecated remap_file_pages() syscall.
2518 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2519 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2522 struct mm_struct *mm = current->mm;
2523 struct vm_area_struct *vma;
2524 unsigned long populate = 0;
2525 unsigned long ret = -EINVAL;
2528 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2529 current->comm, current->pid);
2533 start = start & PAGE_MASK;
2534 size = size & PAGE_MASK;
2536 if (start + size <= start)
2539 /* Does pgoff wrap? */
2540 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2543 down_write(&mm->mmap_sem);
2544 vma = find_vma(mm, start);
2546 if (!vma || !(vma->vm_flags & VM_SHARED))
2549 if (start < vma->vm_start)
2552 if (start + size > vma->vm_end) {
2553 struct vm_area_struct *next;
2555 for (next = vma->vm_next; next; next = next->vm_next) {
2556 /* hole between vmas ? */
2557 if (next->vm_start != next->vm_prev->vm_end)
2560 if (next->vm_file != vma->vm_file)
2563 if (next->vm_flags != vma->vm_flags)
2566 if (start + size <= next->vm_end)
2574 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2575 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2576 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2578 flags &= MAP_NONBLOCK;
2579 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2580 if (vma->vm_flags & VM_LOCKED) {
2581 struct vm_area_struct *tmp;
2582 flags |= MAP_LOCKED;
2584 /* drop PG_Mlocked flag for over-mapped range */
2585 for (tmp = vma; tmp->vm_start >= start + size;
2586 tmp = tmp->vm_next) {
2587 munlock_vma_pages_range(tmp,
2588 max(tmp->vm_start, start),
2589 min(tmp->vm_end, start + size));
2593 file = get_file(vma->vm_file);
2594 ret = do_mmap_pgoff(vma->vm_file, start, size,
2595 prot, flags, pgoff, &populate);
2598 up_write(&mm->mmap_sem);
2600 mm_populate(ret, populate);
2601 if (!IS_ERR_VALUE(ret))
2606 static inline void verify_mm_writelocked(struct mm_struct *mm)
2608 #ifdef CONFIG_DEBUG_VM
2609 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2611 up_read(&mm->mmap_sem);
2617 * this is really a simplified "do_mmap". it only handles
2618 * anonymous maps. eventually we may be able to do some
2619 * brk-specific accounting here.
2621 static unsigned long do_brk(unsigned long addr, unsigned long len)
2623 struct mm_struct *mm = current->mm;
2624 struct vm_area_struct *vma, *prev;
2625 unsigned long flags;
2626 struct rb_node **rb_link, *rb_parent;
2627 pgoff_t pgoff = addr >> PAGE_SHIFT;
2630 len = PAGE_ALIGN(len);
2634 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2636 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2637 if (offset_in_page(error))
2640 error = mlock_future_check(mm, mm->def_flags, len);
2645 * mm->mmap_sem is required to protect against another thread
2646 * changing the mappings in case we sleep.
2648 verify_mm_writelocked(mm);
2651 * Clear old maps. this also does some error checking for us
2653 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2655 if (do_munmap(mm, addr, len))
2659 /* Check against address space limits *after* clearing old maps... */
2660 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2663 if (mm->map_count > sysctl_max_map_count)
2666 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2669 /* Can we just expand an old private anonymous mapping? */
2670 vma = vma_merge(mm, prev, addr, addr + len, flags,
2671 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2676 * create a vma struct for an anonymous mapping
2678 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2680 vm_unacct_memory(len >> PAGE_SHIFT);
2684 INIT_LIST_HEAD(&vma->anon_vma_chain);
2686 vma->vm_start = addr;
2687 vma->vm_end = addr + len;
2688 vma->vm_pgoff = pgoff;
2689 vma->vm_flags = flags;
2690 vma->vm_page_prot = vm_get_page_prot(flags);
2691 vma_link(mm, vma, prev, rb_link, rb_parent);
2693 perf_event_mmap(vma);
2694 mm->total_vm += len >> PAGE_SHIFT;
2695 mm->data_vm += len >> PAGE_SHIFT;
2696 if (flags & VM_LOCKED)
2697 mm->locked_vm += (len >> PAGE_SHIFT);
2698 vma->vm_flags |= VM_SOFTDIRTY;
2702 unsigned long vm_brk(unsigned long addr, unsigned long len)
2704 struct mm_struct *mm = current->mm;
2708 down_write(&mm->mmap_sem);
2709 ret = do_brk(addr, len);
2710 populate = ((mm->def_flags & VM_LOCKED) != 0);
2711 up_write(&mm->mmap_sem);
2713 mm_populate(addr, len);
2716 EXPORT_SYMBOL(vm_brk);
2718 /* Release all mmaps. */
2719 void exit_mmap(struct mm_struct *mm)
2721 struct mmu_gather tlb;
2722 struct vm_area_struct *vma;
2723 unsigned long nr_accounted = 0;
2725 /* mm's last user has gone, and its about to be pulled down */
2726 mmu_notifier_release(mm);
2728 if (mm->locked_vm) {
2731 if (vma->vm_flags & VM_LOCKED)
2732 munlock_vma_pages_all(vma);
2740 if (!vma) /* Can happen if dup_mmap() received an OOM */
2745 tlb_gather_mmu(&tlb, mm, 0, -1);
2746 /* update_hiwater_rss(mm) here? but nobody should be looking */
2747 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2748 unmap_vmas(&tlb, vma, 0, -1);
2750 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2751 tlb_finish_mmu(&tlb, 0, -1);
2754 * Walk the list again, actually closing and freeing it,
2755 * with preemption enabled, without holding any MM locks.
2758 if (vma->vm_flags & VM_ACCOUNT)
2759 nr_accounted += vma_pages(vma);
2760 vma = remove_vma(vma);
2762 vm_unacct_memory(nr_accounted);
2765 /* Insert vm structure into process list sorted by address
2766 * and into the inode's i_mmap tree. If vm_file is non-NULL
2767 * then i_mmap_rwsem is taken here.
2769 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2771 struct vm_area_struct *prev;
2772 struct rb_node **rb_link, *rb_parent;
2774 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2775 &prev, &rb_link, &rb_parent))
2777 if ((vma->vm_flags & VM_ACCOUNT) &&
2778 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2782 * The vm_pgoff of a purely anonymous vma should be irrelevant
2783 * until its first write fault, when page's anon_vma and index
2784 * are set. But now set the vm_pgoff it will almost certainly
2785 * end up with (unless mremap moves it elsewhere before that
2786 * first wfault), so /proc/pid/maps tells a consistent story.
2788 * By setting it to reflect the virtual start address of the
2789 * vma, merges and splits can happen in a seamless way, just
2790 * using the existing file pgoff checks and manipulations.
2791 * Similarly in do_mmap_pgoff and in do_brk.
2793 if (vma_is_anonymous(vma)) {
2794 BUG_ON(vma->anon_vma);
2795 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2798 vma_link(mm, vma, prev, rb_link, rb_parent);
2803 * Copy the vma structure to a new location in the same mm,
2804 * prior to moving page table entries, to effect an mremap move.
2806 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2807 unsigned long addr, unsigned long len, pgoff_t pgoff,
2808 bool *need_rmap_locks)
2810 struct vm_area_struct *vma = *vmap;
2811 unsigned long vma_start = vma->vm_start;
2812 struct mm_struct *mm = vma->vm_mm;
2813 struct vm_area_struct *new_vma, *prev;
2814 struct rb_node **rb_link, *rb_parent;
2815 bool faulted_in_anon_vma = true;
2818 * If anonymous vma has not yet been faulted, update new pgoff
2819 * to match new location, to increase its chance of merging.
2821 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
2822 pgoff = addr >> PAGE_SHIFT;
2823 faulted_in_anon_vma = false;
2826 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2827 return NULL; /* should never get here */
2828 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2829 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
2830 vma->vm_userfaultfd_ctx);
2833 * Source vma may have been merged into new_vma
2835 if (unlikely(vma_start >= new_vma->vm_start &&
2836 vma_start < new_vma->vm_end)) {
2838 * The only way we can get a vma_merge with
2839 * self during an mremap is if the vma hasn't
2840 * been faulted in yet and we were allowed to
2841 * reset the dst vma->vm_pgoff to the
2842 * destination address of the mremap to allow
2843 * the merge to happen. mremap must change the
2844 * vm_pgoff linearity between src and dst vmas
2845 * (in turn preventing a vma_merge) to be
2846 * safe. It is only safe to keep the vm_pgoff
2847 * linear if there are no pages mapped yet.
2849 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
2850 *vmap = vma = new_vma;
2852 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2854 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2858 new_vma->vm_start = addr;
2859 new_vma->vm_end = addr + len;
2860 new_vma->vm_pgoff = pgoff;
2861 if (vma_dup_policy(vma, new_vma))
2863 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2864 if (anon_vma_clone(new_vma, vma))
2865 goto out_free_mempol;
2866 if (new_vma->vm_file)
2867 get_file(new_vma->vm_file);
2868 if (new_vma->vm_ops && new_vma->vm_ops->open)
2869 new_vma->vm_ops->open(new_vma);
2870 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2871 *need_rmap_locks = false;
2876 mpol_put(vma_policy(new_vma));
2878 kmem_cache_free(vm_area_cachep, new_vma);
2884 * Return true if the calling process may expand its vm space by the passed
2887 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
2889 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
2892 if (is_data_mapping(flags) &&
2893 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
2894 if (ignore_rlimit_data)
2895 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Will be forbidden soon.\n",
2896 current->comm, current->pid,
2897 (mm->data_vm + npages) << PAGE_SHIFT,
2898 rlimit(RLIMIT_DATA));
2906 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
2908 mm->total_vm += npages;
2910 if (is_exec_mapping(flags))
2911 mm->exec_vm += npages;
2912 else if (is_stack_mapping(flags))
2913 mm->stack_vm += npages;
2914 else if (is_data_mapping(flags))
2915 mm->data_vm += npages;
2918 static int special_mapping_fault(struct vm_area_struct *vma,
2919 struct vm_fault *vmf);
2922 * Having a close hook prevents vma merging regardless of flags.
2924 static void special_mapping_close(struct vm_area_struct *vma)
2928 static const char *special_mapping_name(struct vm_area_struct *vma)
2930 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
2933 static const struct vm_operations_struct special_mapping_vmops = {
2934 .close = special_mapping_close,
2935 .fault = special_mapping_fault,
2936 .name = special_mapping_name,
2939 static const struct vm_operations_struct legacy_special_mapping_vmops = {
2940 .close = special_mapping_close,
2941 .fault = special_mapping_fault,
2944 static int special_mapping_fault(struct vm_area_struct *vma,
2945 struct vm_fault *vmf)
2948 struct page **pages;
2950 if (vma->vm_ops == &legacy_special_mapping_vmops) {
2951 pages = vma->vm_private_data;
2953 struct vm_special_mapping *sm = vma->vm_private_data;
2956 return sm->fault(sm, vma, vmf);
2961 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
2965 struct page *page = *pages;
2971 return VM_FAULT_SIGBUS;
2974 static struct vm_area_struct *__install_special_mapping(
2975 struct mm_struct *mm,
2976 unsigned long addr, unsigned long len,
2977 unsigned long vm_flags, void *priv,
2978 const struct vm_operations_struct *ops)
2981 struct vm_area_struct *vma;
2983 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2984 if (unlikely(vma == NULL))
2985 return ERR_PTR(-ENOMEM);
2987 INIT_LIST_HEAD(&vma->anon_vma_chain);
2989 vma->vm_start = addr;
2990 vma->vm_end = addr + len;
2992 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
2993 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2996 vma->vm_private_data = priv;
2998 ret = insert_vm_struct(mm, vma);
3002 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3004 perf_event_mmap(vma);
3009 kmem_cache_free(vm_area_cachep, vma);
3010 return ERR_PTR(ret);
3014 * Called with mm->mmap_sem held for writing.
3015 * Insert a new vma covering the given region, with the given flags.
3016 * Its pages are supplied by the given array of struct page *.
3017 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3018 * The region past the last page supplied will always produce SIGBUS.
3019 * The array pointer and the pages it points to are assumed to stay alive
3020 * for as long as this mapping might exist.
3022 struct vm_area_struct *_install_special_mapping(
3023 struct mm_struct *mm,
3024 unsigned long addr, unsigned long len,
3025 unsigned long vm_flags, const struct vm_special_mapping *spec)
3027 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3028 &special_mapping_vmops);
3031 int install_special_mapping(struct mm_struct *mm,
3032 unsigned long addr, unsigned long len,
3033 unsigned long vm_flags, struct page **pages)
3035 struct vm_area_struct *vma = __install_special_mapping(
3036 mm, addr, len, vm_flags, (void *)pages,
3037 &legacy_special_mapping_vmops);
3039 return PTR_ERR_OR_ZERO(vma);
3042 static DEFINE_MUTEX(mm_all_locks_mutex);
3044 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3046 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3048 * The LSB of head.next can't change from under us
3049 * because we hold the mm_all_locks_mutex.
3051 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3053 * We can safely modify head.next after taking the
3054 * anon_vma->root->rwsem. If some other vma in this mm shares
3055 * the same anon_vma we won't take it again.
3057 * No need of atomic instructions here, head.next
3058 * can't change from under us thanks to the
3059 * anon_vma->root->rwsem.
3061 if (__test_and_set_bit(0, (unsigned long *)
3062 &anon_vma->root->rb_root.rb_node))
3067 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3069 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3071 * AS_MM_ALL_LOCKS can't change from under us because
3072 * we hold the mm_all_locks_mutex.
3074 * Operations on ->flags have to be atomic because
3075 * even if AS_MM_ALL_LOCKS is stable thanks to the
3076 * mm_all_locks_mutex, there may be other cpus
3077 * changing other bitflags in parallel to us.
3079 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3081 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3086 * This operation locks against the VM for all pte/vma/mm related
3087 * operations that could ever happen on a certain mm. This includes
3088 * vmtruncate, try_to_unmap, and all page faults.
3090 * The caller must take the mmap_sem in write mode before calling
3091 * mm_take_all_locks(). The caller isn't allowed to release the
3092 * mmap_sem until mm_drop_all_locks() returns.
3094 * mmap_sem in write mode is required in order to block all operations
3095 * that could modify pagetables and free pages without need of
3096 * altering the vma layout. It's also needed in write mode to avoid new
3097 * anon_vmas to be associated with existing vmas.
3099 * A single task can't take more than one mm_take_all_locks() in a row
3100 * or it would deadlock.
3102 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3103 * mapping->flags avoid to take the same lock twice, if more than one
3104 * vma in this mm is backed by the same anon_vma or address_space.
3106 * We take locks in following order, accordingly to comment at beginning
3108 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3110 * - all i_mmap_rwsem locks;
3111 * - all anon_vma->rwseml
3113 * We can take all locks within these types randomly because the VM code
3114 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3115 * mm_all_locks_mutex.
3117 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3118 * that may have to take thousand of locks.
3120 * mm_take_all_locks() can fail if it's interrupted by signals.
3122 int mm_take_all_locks(struct mm_struct *mm)
3124 struct vm_area_struct *vma;
3125 struct anon_vma_chain *avc;
3127 BUG_ON(down_read_trylock(&mm->mmap_sem));
3129 mutex_lock(&mm_all_locks_mutex);
3131 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3132 if (signal_pending(current))
3134 if (vma->vm_file && vma->vm_file->f_mapping &&
3135 is_vm_hugetlb_page(vma))
3136 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3139 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3140 if (signal_pending(current))
3142 if (vma->vm_file && vma->vm_file->f_mapping &&
3143 !is_vm_hugetlb_page(vma))
3144 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3147 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3148 if (signal_pending(current))
3151 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3152 vm_lock_anon_vma(mm, avc->anon_vma);
3158 mm_drop_all_locks(mm);
3162 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3164 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3166 * The LSB of head.next can't change to 0 from under
3167 * us because we hold the mm_all_locks_mutex.
3169 * We must however clear the bitflag before unlocking
3170 * the vma so the users using the anon_vma->rb_root will
3171 * never see our bitflag.
3173 * No need of atomic instructions here, head.next
3174 * can't change from under us until we release the
3175 * anon_vma->root->rwsem.
3177 if (!__test_and_clear_bit(0, (unsigned long *)
3178 &anon_vma->root->rb_root.rb_node))
3180 anon_vma_unlock_write(anon_vma);
3184 static void vm_unlock_mapping(struct address_space *mapping)
3186 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3188 * AS_MM_ALL_LOCKS can't change to 0 from under us
3189 * because we hold the mm_all_locks_mutex.
3191 i_mmap_unlock_write(mapping);
3192 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3199 * The mmap_sem cannot be released by the caller until
3200 * mm_drop_all_locks() returns.
3202 void mm_drop_all_locks(struct mm_struct *mm)
3204 struct vm_area_struct *vma;
3205 struct anon_vma_chain *avc;
3207 BUG_ON(down_read_trylock(&mm->mmap_sem));
3208 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3210 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3212 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3213 vm_unlock_anon_vma(avc->anon_vma);
3214 if (vma->vm_file && vma->vm_file->f_mapping)
3215 vm_unlock_mapping(vma->vm_file->f_mapping);
3218 mutex_unlock(&mm_all_locks_mutex);
3222 * initialise the VMA slab
3224 void __init mmap_init(void)
3228 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3233 * Initialise sysctl_user_reserve_kbytes.
3235 * This is intended to prevent a user from starting a single memory hogging
3236 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3239 * The default value is min(3% of free memory, 128MB)
3240 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3242 static int init_user_reserve(void)
3244 unsigned long free_kbytes;
3246 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3248 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3251 subsys_initcall(init_user_reserve);
3254 * Initialise sysctl_admin_reserve_kbytes.
3256 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3257 * to log in and kill a memory hogging process.
3259 * Systems with more than 256MB will reserve 8MB, enough to recover
3260 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3261 * only reserve 3% of free pages by default.
3263 static int init_admin_reserve(void)
3265 unsigned long free_kbytes;
3267 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3269 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3272 subsys_initcall(init_admin_reserve);
3275 * Reinititalise user and admin reserves if memory is added or removed.
3277 * The default user reserve max is 128MB, and the default max for the
3278 * admin reserve is 8MB. These are usually, but not always, enough to
3279 * enable recovery from a memory hogging process using login/sshd, a shell,
3280 * and tools like top. It may make sense to increase or even disable the
3281 * reserve depending on the existence of swap or variations in the recovery
3282 * tools. So, the admin may have changed them.
3284 * If memory is added and the reserves have been eliminated or increased above
3285 * the default max, then we'll trust the admin.
3287 * If memory is removed and there isn't enough free memory, then we
3288 * need to reset the reserves.
3290 * Otherwise keep the reserve set by the admin.
3292 static int reserve_mem_notifier(struct notifier_block *nb,
3293 unsigned long action, void *data)
3295 unsigned long tmp, free_kbytes;
3299 /* Default max is 128MB. Leave alone if modified by operator. */
3300 tmp = sysctl_user_reserve_kbytes;
3301 if (0 < tmp && tmp < (1UL << 17))
3302 init_user_reserve();
3304 /* Default max is 8MB. Leave alone if modified by operator. */
3305 tmp = sysctl_admin_reserve_kbytes;
3306 if (0 < tmp && tmp < (1UL << 13))
3307 init_admin_reserve();
3311 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3313 if (sysctl_user_reserve_kbytes > free_kbytes) {
3314 init_user_reserve();
3315 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3316 sysctl_user_reserve_kbytes);
3319 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3320 init_admin_reserve();
3321 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3322 sysctl_admin_reserve_kbytes);
3331 static struct notifier_block reserve_mem_nb = {
3332 .notifier_call = reserve_mem_notifier,
3335 static int __meminit init_reserve_notifier(void)
3337 if (register_hotmemory_notifier(&reserve_mem_nb))
3338 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3342 subsys_initcall(init_reserve_notifier);
projects / linux-2.6-block.git / blob