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/sched/sysctl.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>
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 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */
128 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
129 unsigned long sysctl_overcommit_kbytes __read_mostly;
130 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
131 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
132 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
134 * Make sure vm_committed_as in one cacheline and not cacheline shared with
135 * other variables. It can be updated by several CPUs frequently.
137 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
140 * The global memory commitment made in the system can be a metric
141 * that can be used to drive ballooning decisions when Linux is hosted
142 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
143 * balancing memory across competing virtual machines that are hosted.
144 * Several metrics drive this policy engine including the guest reported
147 unsigned long vm_memory_committed(void)
149 return percpu_counter_read_positive(&vm_committed_as);
151 EXPORT_SYMBOL_GPL(vm_memory_committed);
154 * Check that a process has enough memory to allocate a new virtual
155 * mapping. 0 means there is enough memory for the allocation to
156 * succeed and -ENOMEM implies there is not.
158 * We currently support three overcommit policies, which are set via the
159 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
161 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
162 * Additional code 2002 Jul 20 by Robert Love.
164 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
166 * Note this is a helper function intended to be used by LSMs which
167 * wish to use this logic.
169 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
171 long free, allowed, reserve;
173 VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) <
174 -(s64)vm_committed_as_batch * num_online_cpus(),
175 "memory commitment underflow");
177 vm_acct_memory(pages);
180 * Sometimes we want to use more memory than we have
182 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
185 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
186 free = global_page_state(NR_FREE_PAGES);
187 free += global_page_state(NR_FILE_PAGES);
190 * shmem pages shouldn't be counted as free in this
191 * case, they can't be purged, only swapped out, and
192 * that won't affect the overall amount of available
193 * memory in the system.
195 free -= global_page_state(NR_SHMEM);
197 free += get_nr_swap_pages();
200 * Any slabs which are created with the
201 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
202 * which are reclaimable, under pressure. The dentry
203 * cache and most inode caches should fall into this
205 free += global_page_state(NR_SLAB_RECLAIMABLE);
208 * Leave reserved pages. The pages are not for anonymous pages.
210 if (free <= totalreserve_pages)
213 free -= totalreserve_pages;
216 * Reserve some for root
219 free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
227 allowed = vm_commit_limit();
229 * Reserve some for root
232 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
235 * Don't let a single process grow so big a user can't recover
238 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
239 allowed -= min_t(long, mm->total_vm / 32, reserve);
242 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
245 vm_unacct_memory(pages);
251 * Requires inode->i_mapping->i_mmap_rwsem
253 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
254 struct file *file, struct address_space *mapping)
256 if (vma->vm_flags & VM_DENYWRITE)
257 atomic_inc(&file_inode(file)->i_writecount);
258 if (vma->vm_flags & VM_SHARED)
259 mapping_unmap_writable(mapping);
261 flush_dcache_mmap_lock(mapping);
262 vma_interval_tree_remove(vma, &mapping->i_mmap);
263 flush_dcache_mmap_unlock(mapping);
267 * Unlink a file-based vm structure from its interval tree, to hide
268 * vma from rmap and vmtruncate before freeing its page tables.
270 void unlink_file_vma(struct vm_area_struct *vma)
272 struct file *file = vma->vm_file;
275 struct address_space *mapping = file->f_mapping;
276 i_mmap_lock_write(mapping);
277 __remove_shared_vm_struct(vma, file, mapping);
278 i_mmap_unlock_write(mapping);
283 * Close a vm structure and free it, returning the next.
285 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
287 struct vm_area_struct *next = vma->vm_next;
290 if (vma->vm_ops && vma->vm_ops->close)
291 vma->vm_ops->close(vma);
294 mpol_put(vma_policy(vma));
295 kmem_cache_free(vm_area_cachep, vma);
299 static unsigned long do_brk(unsigned long addr, unsigned long len);
301 SYSCALL_DEFINE1(brk, unsigned long, brk)
303 unsigned long retval;
304 unsigned long newbrk, oldbrk;
305 struct mm_struct *mm = current->mm;
306 unsigned long min_brk;
309 down_write(&mm->mmap_sem);
311 #ifdef CONFIG_COMPAT_BRK
313 * CONFIG_COMPAT_BRK can still be overridden by setting
314 * randomize_va_space to 2, which will still cause mm->start_brk
315 * to be arbitrarily shifted
317 if (current->brk_randomized)
318 min_brk = mm->start_brk;
320 min_brk = mm->end_data;
322 min_brk = mm->start_brk;
328 * Check against rlimit here. If this check is done later after the test
329 * of oldbrk with newbrk then it can escape the test and let the data
330 * segment grow beyond its set limit the in case where the limit is
331 * not page aligned -Ram Gupta
333 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
334 mm->end_data, mm->start_data))
337 newbrk = PAGE_ALIGN(brk);
338 oldbrk = PAGE_ALIGN(mm->brk);
339 if (oldbrk == newbrk)
342 /* Always allow shrinking brk. */
343 if (brk <= mm->brk) {
344 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
349 /* Check against existing mmap mappings. */
350 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
353 /* Ok, looks good - let it rip. */
354 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
359 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
360 up_write(&mm->mmap_sem);
362 mm_populate(oldbrk, newbrk - oldbrk);
367 up_write(&mm->mmap_sem);
371 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
373 unsigned long max, subtree_gap;
376 max -= vma->vm_prev->vm_end;
377 if (vma->vm_rb.rb_left) {
378 subtree_gap = rb_entry(vma->vm_rb.rb_left,
379 struct vm_area_struct, vm_rb)->rb_subtree_gap;
380 if (subtree_gap > max)
383 if (vma->vm_rb.rb_right) {
384 subtree_gap = rb_entry(vma->vm_rb.rb_right,
385 struct vm_area_struct, vm_rb)->rb_subtree_gap;
386 if (subtree_gap > max)
392 #ifdef CONFIG_DEBUG_VM_RB
393 static int browse_rb(struct mm_struct *mm)
395 struct rb_root *root = &mm->mm_rb;
396 int i = 0, j, bug = 0;
397 struct rb_node *nd, *pn = NULL;
398 unsigned long prev = 0, pend = 0;
400 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
401 struct vm_area_struct *vma;
402 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
403 if (vma->vm_start < prev) {
404 pr_emerg("vm_start %lx < prev %lx\n",
405 vma->vm_start, prev);
408 if (vma->vm_start < pend) {
409 pr_emerg("vm_start %lx < pend %lx\n",
410 vma->vm_start, pend);
413 if (vma->vm_start > vma->vm_end) {
414 pr_emerg("vm_start %lx > vm_end %lx\n",
415 vma->vm_start, vma->vm_end);
418 spin_lock(&mm->page_table_lock);
419 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
420 pr_emerg("free gap %lx, correct %lx\n",
422 vma_compute_subtree_gap(vma));
425 spin_unlock(&mm->page_table_lock);
428 prev = vma->vm_start;
432 for (nd = pn; nd; nd = rb_prev(nd))
435 pr_emerg("backwards %d, forwards %d\n", j, i);
441 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
445 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
446 struct vm_area_struct *vma;
447 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
448 VM_BUG_ON_VMA(vma != ignore &&
449 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
454 static void validate_mm(struct mm_struct *mm)
458 unsigned long highest_address = 0;
459 struct vm_area_struct *vma = mm->mmap;
462 struct anon_vma_chain *avc;
464 vma_lock_anon_vma(vma);
465 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
466 anon_vma_interval_tree_verify(avc);
467 vma_unlock_anon_vma(vma);
468 highest_address = vma->vm_end;
472 if (i != mm->map_count) {
473 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
476 if (highest_address != mm->highest_vm_end) {
477 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
478 mm->highest_vm_end, highest_address);
482 if (i != mm->map_count) {
484 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
487 VM_BUG_ON_MM(bug, mm);
490 #define validate_mm_rb(root, ignore) do { } while (0)
491 #define validate_mm(mm) do { } while (0)
494 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
495 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
498 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
499 * vma->vm_prev->vm_end values changed, without modifying the vma's position
502 static void vma_gap_update(struct vm_area_struct *vma)
505 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
506 * function that does exacltly what we want.
508 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
511 static inline void vma_rb_insert(struct vm_area_struct *vma,
512 struct rb_root *root)
514 /* All rb_subtree_gap values must be consistent prior to insertion */
515 validate_mm_rb(root, NULL);
517 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
520 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
523 * All rb_subtree_gap values must be consistent prior to erase,
524 * with the possible exception of the vma being erased.
526 validate_mm_rb(root, vma);
529 * Note rb_erase_augmented is a fairly large inline function,
530 * so make sure we instantiate it only once with our desired
531 * augmented rbtree callbacks.
533 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
537 * vma has some anon_vma assigned, and is already inserted on that
538 * anon_vma's interval trees.
540 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
541 * vma must be removed from the anon_vma's interval trees using
542 * anon_vma_interval_tree_pre_update_vma().
544 * After the update, the vma will be reinserted using
545 * anon_vma_interval_tree_post_update_vma().
547 * The entire update must be protected by exclusive mmap_sem and by
548 * the root anon_vma's mutex.
551 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
553 struct anon_vma_chain *avc;
555 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
556 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
560 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
562 struct anon_vma_chain *avc;
564 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
565 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
568 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
569 unsigned long end, struct vm_area_struct **pprev,
570 struct rb_node ***rb_link, struct rb_node **rb_parent)
572 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
574 __rb_link = &mm->mm_rb.rb_node;
575 rb_prev = __rb_parent = NULL;
578 struct vm_area_struct *vma_tmp;
580 __rb_parent = *__rb_link;
581 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
583 if (vma_tmp->vm_end > addr) {
584 /* Fail if an existing vma overlaps the area */
585 if (vma_tmp->vm_start < end)
587 __rb_link = &__rb_parent->rb_left;
589 rb_prev = __rb_parent;
590 __rb_link = &__rb_parent->rb_right;
596 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
597 *rb_link = __rb_link;
598 *rb_parent = __rb_parent;
602 static unsigned long count_vma_pages_range(struct mm_struct *mm,
603 unsigned long addr, unsigned long end)
605 unsigned long nr_pages = 0;
606 struct vm_area_struct *vma;
608 /* Find first overlaping mapping */
609 vma = find_vma_intersection(mm, addr, end);
613 nr_pages = (min(end, vma->vm_end) -
614 max(addr, vma->vm_start)) >> PAGE_SHIFT;
616 /* Iterate over the rest of the overlaps */
617 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
618 unsigned long overlap_len;
620 if (vma->vm_start > end)
623 overlap_len = min(end, vma->vm_end) - vma->vm_start;
624 nr_pages += overlap_len >> PAGE_SHIFT;
630 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
631 struct rb_node **rb_link, struct rb_node *rb_parent)
633 /* Update tracking information for the gap following the new vma. */
635 vma_gap_update(vma->vm_next);
637 mm->highest_vm_end = vma->vm_end;
640 * vma->vm_prev wasn't known when we followed the rbtree to find the
641 * correct insertion point for that vma. As a result, we could not
642 * update the vma vm_rb parents rb_subtree_gap values on the way down.
643 * So, we first insert the vma with a zero rb_subtree_gap value
644 * (to be consistent with what we did on the way down), and then
645 * immediately update the gap to the correct value. Finally we
646 * rebalance the rbtree after all augmented values have been set.
648 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
649 vma->rb_subtree_gap = 0;
651 vma_rb_insert(vma, &mm->mm_rb);
654 static void __vma_link_file(struct vm_area_struct *vma)
660 struct address_space *mapping = file->f_mapping;
662 if (vma->vm_flags & VM_DENYWRITE)
663 atomic_dec(&file_inode(file)->i_writecount);
664 if (vma->vm_flags & VM_SHARED)
665 atomic_inc(&mapping->i_mmap_writable);
667 flush_dcache_mmap_lock(mapping);
668 vma_interval_tree_insert(vma, &mapping->i_mmap);
669 flush_dcache_mmap_unlock(mapping);
674 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
675 struct vm_area_struct *prev, struct rb_node **rb_link,
676 struct rb_node *rb_parent)
678 __vma_link_list(mm, vma, prev, rb_parent);
679 __vma_link_rb(mm, vma, rb_link, rb_parent);
682 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
683 struct vm_area_struct *prev, struct rb_node **rb_link,
684 struct rb_node *rb_parent)
686 struct address_space *mapping = NULL;
689 mapping = vma->vm_file->f_mapping;
690 i_mmap_lock_write(mapping);
693 __vma_link(mm, vma, prev, rb_link, rb_parent);
694 __vma_link_file(vma);
697 i_mmap_unlock_write(mapping);
704 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
705 * mm's list and rbtree. It has already been inserted into the interval tree.
707 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
709 struct vm_area_struct *prev;
710 struct rb_node **rb_link, *rb_parent;
712 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
713 &prev, &rb_link, &rb_parent))
715 __vma_link(mm, vma, prev, rb_link, rb_parent);
720 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
721 struct vm_area_struct *prev)
723 struct vm_area_struct *next;
725 vma_rb_erase(vma, &mm->mm_rb);
726 prev->vm_next = next = vma->vm_next;
728 next->vm_prev = prev;
731 vmacache_invalidate(mm);
735 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
736 * is already present in an i_mmap tree without adjusting the tree.
737 * The following helper function should be used when such adjustments
738 * are necessary. The "insert" vma (if any) is to be inserted
739 * before we drop the necessary locks.
741 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
742 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
744 struct mm_struct *mm = vma->vm_mm;
745 struct vm_area_struct *next = vma->vm_next;
746 struct vm_area_struct *importer = NULL;
747 struct address_space *mapping = NULL;
748 struct rb_root *root = NULL;
749 struct anon_vma *anon_vma = NULL;
750 struct file *file = vma->vm_file;
751 bool start_changed = false, end_changed = false;
752 long adjust_next = 0;
755 if (next && !insert) {
756 struct vm_area_struct *exporter = NULL;
758 if (end >= next->vm_end) {
760 * vma expands, overlapping all the next, and
761 * perhaps the one after too (mprotect case 6).
763 again: remove_next = 1 + (end > next->vm_end);
767 } else if (end > next->vm_start) {
769 * vma expands, overlapping part of the next:
770 * mprotect case 5 shifting the boundary up.
772 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
775 } else if (end < vma->vm_end) {
777 * vma shrinks, and !insert tells it's not
778 * split_vma inserting another: so it must be
779 * mprotect case 4 shifting the boundary down.
781 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
787 * Easily overlooked: when mprotect shifts the boundary,
788 * make sure the expanding vma has anon_vma set if the
789 * shrinking vma had, to cover any anon pages imported.
791 if (exporter && exporter->anon_vma && !importer->anon_vma) {
794 importer->anon_vma = exporter->anon_vma;
795 error = anon_vma_clone(importer, exporter);
802 mapping = file->f_mapping;
803 root = &mapping->i_mmap;
804 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
807 uprobe_munmap(next, next->vm_start, next->vm_end);
809 i_mmap_lock_write(mapping);
812 * Put into interval tree now, so instantiated pages
813 * are visible to arm/parisc __flush_dcache_page
814 * throughout; but we cannot insert into address
815 * space until vma start or end is updated.
817 __vma_link_file(insert);
821 vma_adjust_trans_huge(vma, start, end, adjust_next);
823 anon_vma = vma->anon_vma;
824 if (!anon_vma && adjust_next)
825 anon_vma = next->anon_vma;
827 VM_BUG_ON_VMA(adjust_next && next->anon_vma &&
828 anon_vma != next->anon_vma, next);
829 anon_vma_lock_write(anon_vma);
830 anon_vma_interval_tree_pre_update_vma(vma);
832 anon_vma_interval_tree_pre_update_vma(next);
836 flush_dcache_mmap_lock(mapping);
837 vma_interval_tree_remove(vma, root);
839 vma_interval_tree_remove(next, root);
842 if (start != vma->vm_start) {
843 vma->vm_start = start;
844 start_changed = true;
846 if (end != vma->vm_end) {
850 vma->vm_pgoff = pgoff;
852 next->vm_start += adjust_next << PAGE_SHIFT;
853 next->vm_pgoff += adjust_next;
858 vma_interval_tree_insert(next, root);
859 vma_interval_tree_insert(vma, root);
860 flush_dcache_mmap_unlock(mapping);
865 * vma_merge has merged next into vma, and needs
866 * us to remove next before dropping the locks.
868 __vma_unlink(mm, next, vma);
870 __remove_shared_vm_struct(next, file, mapping);
873 * split_vma has split insert from vma, and needs
874 * us to insert it before dropping the locks
875 * (it may either follow vma or precede it).
877 __insert_vm_struct(mm, insert);
883 mm->highest_vm_end = end;
884 else if (!adjust_next)
885 vma_gap_update(next);
890 anon_vma_interval_tree_post_update_vma(vma);
892 anon_vma_interval_tree_post_update_vma(next);
893 anon_vma_unlock_write(anon_vma);
896 i_mmap_unlock_write(mapping);
907 uprobe_munmap(next, next->vm_start, next->vm_end);
911 anon_vma_merge(vma, next);
913 mpol_put(vma_policy(next));
914 kmem_cache_free(vm_area_cachep, next);
916 * In mprotect's case 6 (see comments on vma_merge),
917 * we must remove another next too. It would clutter
918 * up the code too much to do both in one go.
921 if (remove_next == 2)
924 vma_gap_update(next);
926 mm->highest_vm_end = end;
937 * If the vma has a ->close operation then the driver probably needs to release
938 * per-vma resources, so we don't attempt to merge those.
940 static inline int is_mergeable_vma(struct vm_area_struct *vma,
941 struct file *file, unsigned long vm_flags,
942 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
945 * VM_SOFTDIRTY should not prevent from VMA merging, if we
946 * match the flags but dirty bit -- the caller should mark
947 * merged VMA as dirty. If dirty bit won't be excluded from
948 * comparison, we increase pressue on the memory system forcing
949 * the kernel to generate new VMAs when old one could be
952 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
954 if (vma->vm_file != file)
956 if (vma->vm_ops && vma->vm_ops->close)
958 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
963 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
964 struct anon_vma *anon_vma2,
965 struct vm_area_struct *vma)
968 * The list_is_singular() test is to avoid merging VMA cloned from
969 * parents. This can improve scalability caused by anon_vma lock.
971 if ((!anon_vma1 || !anon_vma2) && (!vma ||
972 list_is_singular(&vma->anon_vma_chain)))
974 return anon_vma1 == anon_vma2;
978 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
979 * in front of (at a lower virtual address and file offset than) the vma.
981 * We cannot merge two vmas if they have differently assigned (non-NULL)
982 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
984 * We don't check here for the merged mmap wrapping around the end of pagecache
985 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
986 * wrap, nor mmaps which cover the final page at index -1UL.
989 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
990 struct anon_vma *anon_vma, struct file *file,
992 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
994 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
995 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
996 if (vma->vm_pgoff == vm_pgoff)
1003 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1004 * beyond (at a higher virtual address and file offset than) the vma.
1006 * We cannot merge two vmas if they have differently assigned (non-NULL)
1007 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1010 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1011 struct anon_vma *anon_vma, struct file *file,
1013 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1015 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1016 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1018 vm_pglen = vma_pages(vma);
1019 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1026 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1027 * whether that can be merged with its predecessor or its successor.
1028 * Or both (it neatly fills a hole).
1030 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1031 * certain not to be mapped by the time vma_merge is called; but when
1032 * called for mprotect, it is certain to be already mapped (either at
1033 * an offset within prev, or at the start of next), and the flags of
1034 * this area are about to be changed to vm_flags - and the no-change
1035 * case has already been eliminated.
1037 * The following mprotect cases have to be considered, where AAAA is
1038 * the area passed down from mprotect_fixup, never extending beyond one
1039 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1041 * AAAA AAAA AAAA AAAA
1042 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1043 * cannot merge might become might become might become
1044 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1045 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1046 * mremap move: PPPPNNNNNNNN 8
1048 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1049 * might become case 1 below case 2 below case 3 below
1051 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1052 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1054 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1055 struct vm_area_struct *prev, unsigned long addr,
1056 unsigned long end, unsigned long vm_flags,
1057 struct anon_vma *anon_vma, struct file *file,
1058 pgoff_t pgoff, struct mempolicy *policy,
1059 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1061 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1062 struct vm_area_struct *area, *next;
1066 * We later require that vma->vm_flags == vm_flags,
1067 * so this tests vma->vm_flags & VM_SPECIAL, too.
1069 if (vm_flags & VM_SPECIAL)
1073 next = prev->vm_next;
1077 if (next && next->vm_end == end) /* cases 6, 7, 8 */
1078 next = next->vm_next;
1081 * Can it merge with the predecessor?
1083 if (prev && prev->vm_end == addr &&
1084 mpol_equal(vma_policy(prev), policy) &&
1085 can_vma_merge_after(prev, vm_flags,
1086 anon_vma, file, pgoff,
1087 vm_userfaultfd_ctx)) {
1089 * OK, it can. Can we now merge in the successor as well?
1091 if (next && end == next->vm_start &&
1092 mpol_equal(policy, vma_policy(next)) &&
1093 can_vma_merge_before(next, vm_flags,
1096 vm_userfaultfd_ctx) &&
1097 is_mergeable_anon_vma(prev->anon_vma,
1098 next->anon_vma, NULL)) {
1100 err = vma_adjust(prev, prev->vm_start,
1101 next->vm_end, prev->vm_pgoff, NULL);
1102 } else /* cases 2, 5, 7 */
1103 err = vma_adjust(prev, prev->vm_start,
1104 end, prev->vm_pgoff, NULL);
1107 khugepaged_enter_vma_merge(prev, vm_flags);
1112 * Can this new request be merged in front of next?
1114 if (next && end == next->vm_start &&
1115 mpol_equal(policy, vma_policy(next)) &&
1116 can_vma_merge_before(next, vm_flags,
1117 anon_vma, file, pgoff+pglen,
1118 vm_userfaultfd_ctx)) {
1119 if (prev && addr < prev->vm_end) /* case 4 */
1120 err = vma_adjust(prev, prev->vm_start,
1121 addr, prev->vm_pgoff, NULL);
1122 else /* cases 3, 8 */
1123 err = vma_adjust(area, addr, next->vm_end,
1124 next->vm_pgoff - pglen, NULL);
1127 khugepaged_enter_vma_merge(area, vm_flags);
1135 * Rough compatbility check to quickly see if it's even worth looking
1136 * at sharing an anon_vma.
1138 * They need to have the same vm_file, and the flags can only differ
1139 * in things that mprotect may change.
1141 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1142 * we can merge the two vma's. For example, we refuse to merge a vma if
1143 * there is a vm_ops->close() function, because that indicates that the
1144 * driver is doing some kind of reference counting. But that doesn't
1145 * really matter for the anon_vma sharing case.
1147 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1149 return a->vm_end == b->vm_start &&
1150 mpol_equal(vma_policy(a), vma_policy(b)) &&
1151 a->vm_file == b->vm_file &&
1152 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1153 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1157 * Do some basic sanity checking to see if we can re-use the anon_vma
1158 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1159 * the same as 'old', the other will be the new one that is trying
1160 * to share the anon_vma.
1162 * NOTE! This runs with mm_sem held for reading, so it is possible that
1163 * the anon_vma of 'old' is concurrently in the process of being set up
1164 * by another page fault trying to merge _that_. But that's ok: if it
1165 * is being set up, that automatically means that it will be a singleton
1166 * acceptable for merging, so we can do all of this optimistically. But
1167 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1169 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1170 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1171 * is to return an anon_vma that is "complex" due to having gone through
1174 * We also make sure that the two vma's are compatible (adjacent,
1175 * and with the same memory policies). That's all stable, even with just
1176 * a read lock on the mm_sem.
1178 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1180 if (anon_vma_compatible(a, b)) {
1181 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1183 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1190 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1191 * neighbouring vmas for a suitable anon_vma, before it goes off
1192 * to allocate a new anon_vma. It checks because a repetitive
1193 * sequence of mprotects and faults may otherwise lead to distinct
1194 * anon_vmas being allocated, preventing vma merge in subsequent
1197 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1199 struct anon_vma *anon_vma;
1200 struct vm_area_struct *near;
1202 near = vma->vm_next;
1206 anon_vma = reusable_anon_vma(near, vma, near);
1210 near = vma->vm_prev;
1214 anon_vma = reusable_anon_vma(near, near, vma);
1219 * There's no absolute need to look only at touching neighbours:
1220 * we could search further afield for "compatible" anon_vmas.
1221 * But it would probably just be a waste of time searching,
1222 * or lead to too many vmas hanging off the same anon_vma.
1223 * We're trying to allow mprotect remerging later on,
1224 * not trying to minimize memory used for anon_vmas.
1230 * If a hint addr is less than mmap_min_addr change hint to be as
1231 * low as possible but still greater than mmap_min_addr
1233 static inline unsigned long round_hint_to_min(unsigned long hint)
1236 if (((void *)hint != NULL) &&
1237 (hint < mmap_min_addr))
1238 return PAGE_ALIGN(mmap_min_addr);
1242 static inline int mlock_future_check(struct mm_struct *mm,
1243 unsigned long flags,
1246 unsigned long locked, lock_limit;
1248 /* mlock MCL_FUTURE? */
1249 if (flags & VM_LOCKED) {
1250 locked = len >> PAGE_SHIFT;
1251 locked += mm->locked_vm;
1252 lock_limit = rlimit(RLIMIT_MEMLOCK);
1253 lock_limit >>= PAGE_SHIFT;
1254 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1261 * The caller must hold down_write(¤t->mm->mmap_sem).
1263 unsigned long do_mmap(struct file *file, unsigned long addr,
1264 unsigned long len, unsigned long prot,
1265 unsigned long flags, vm_flags_t vm_flags,
1266 unsigned long pgoff, unsigned long *populate)
1268 struct mm_struct *mm = current->mm;
1276 * Does the application expect PROT_READ to imply PROT_EXEC?
1278 * (the exception is when the underlying filesystem is noexec
1279 * mounted, in which case we dont add PROT_EXEC.)
1281 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1282 if (!(file && path_noexec(&file->f_path)))
1285 if (!(flags & MAP_FIXED))
1286 addr = round_hint_to_min(addr);
1288 /* Careful about overflows.. */
1289 len = PAGE_ALIGN(len);
1293 /* offset overflow? */
1294 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1297 /* Too many mappings? */
1298 if (mm->map_count > sysctl_max_map_count)
1301 /* Obtain the address to map to. we verify (or select) it and ensure
1302 * that it represents a valid section of the address space.
1304 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1305 if (offset_in_page(addr))
1308 /* Do simple checking here so the lower-level routines won't have
1309 * to. we assume access permissions have been handled by the open
1310 * of the memory object, so we don't do any here.
1312 vm_flags |= calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1313 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1315 if (flags & MAP_LOCKED)
1316 if (!can_do_mlock())
1319 if (mlock_future_check(mm, vm_flags, len))
1323 struct inode *inode = file_inode(file);
1325 switch (flags & MAP_TYPE) {
1327 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1331 * Make sure we don't allow writing to an append-only
1334 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1338 * Make sure there are no mandatory locks on the file.
1340 if (locks_verify_locked(file))
1343 vm_flags |= VM_SHARED | VM_MAYSHARE;
1344 if (!(file->f_mode & FMODE_WRITE))
1345 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1349 if (!(file->f_mode & FMODE_READ))
1351 if (path_noexec(&file->f_path)) {
1352 if (vm_flags & VM_EXEC)
1354 vm_flags &= ~VM_MAYEXEC;
1357 if (!file->f_op->mmap)
1359 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1367 switch (flags & MAP_TYPE) {
1369 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1375 vm_flags |= VM_SHARED | VM_MAYSHARE;
1379 * Set pgoff according to addr for anon_vma.
1381 pgoff = addr >> PAGE_SHIFT;
1389 * Set 'VM_NORESERVE' if we should not account for the
1390 * memory use of this mapping.
1392 if (flags & MAP_NORESERVE) {
1393 /* We honor MAP_NORESERVE if allowed to overcommit */
1394 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1395 vm_flags |= VM_NORESERVE;
1397 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1398 if (file && is_file_hugepages(file))
1399 vm_flags |= VM_NORESERVE;
1402 addr = mmap_region(file, addr, len, vm_flags, pgoff);
1403 if (!IS_ERR_VALUE(addr) &&
1404 ((vm_flags & VM_LOCKED) ||
1405 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1410 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1411 unsigned long, prot, unsigned long, flags,
1412 unsigned long, fd, unsigned long, pgoff)
1414 struct file *file = NULL;
1415 unsigned long retval;
1417 if (!(flags & MAP_ANONYMOUS)) {
1418 audit_mmap_fd(fd, flags);
1422 if (is_file_hugepages(file))
1423 len = ALIGN(len, huge_page_size(hstate_file(file)));
1425 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1427 } else if (flags & MAP_HUGETLB) {
1428 struct user_struct *user = NULL;
1431 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1435 len = ALIGN(len, huge_page_size(hs));
1437 * VM_NORESERVE is used because the reservations will be
1438 * taken when vm_ops->mmap() is called
1439 * A dummy user value is used because we are not locking
1440 * memory so no accounting is necessary
1442 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1444 &user, HUGETLB_ANONHUGE_INODE,
1445 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1447 return PTR_ERR(file);
1450 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1452 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1459 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1460 struct mmap_arg_struct {
1464 unsigned long flags;
1466 unsigned long offset;
1469 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1471 struct mmap_arg_struct a;
1473 if (copy_from_user(&a, arg, sizeof(a)))
1475 if (offset_in_page(a.offset))
1478 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1479 a.offset >> PAGE_SHIFT);
1481 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1484 * Some shared mappigns will want the pages marked read-only
1485 * to track write events. If so, we'll downgrade vm_page_prot
1486 * to the private version (using protection_map[] without the
1489 int vma_wants_writenotify(struct vm_area_struct *vma)
1491 vm_flags_t vm_flags = vma->vm_flags;
1492 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1494 /* If it was private or non-writable, the write bit is already clear */
1495 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1498 /* The backer wishes to know when pages are first written to? */
1499 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1502 /* The open routine did something to the protections that pgprot_modify
1503 * won't preserve? */
1504 if (pgprot_val(vma->vm_page_prot) !=
1505 pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags)))
1508 /* Do we need to track softdirty? */
1509 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1512 /* Specialty mapping? */
1513 if (vm_flags & VM_PFNMAP)
1516 /* Can the mapping track the dirty pages? */
1517 return vma->vm_file && vma->vm_file->f_mapping &&
1518 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1522 * We account for memory if it's a private writeable mapping,
1523 * not hugepages and VM_NORESERVE wasn't set.
1525 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1528 * hugetlb has its own accounting separate from the core VM
1529 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1531 if (file && is_file_hugepages(file))
1534 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1537 unsigned long mmap_region(struct file *file, unsigned long addr,
1538 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1540 struct mm_struct *mm = current->mm;
1541 struct vm_area_struct *vma, *prev;
1543 struct rb_node **rb_link, *rb_parent;
1544 unsigned long charged = 0;
1546 /* Check against address space limit. */
1547 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1548 unsigned long nr_pages;
1551 * MAP_FIXED may remove pages of mappings that intersects with
1552 * requested mapping. Account for the pages it would unmap.
1554 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1556 if (!may_expand_vm(mm, vm_flags,
1557 (len >> PAGE_SHIFT) - nr_pages))
1561 /* Clear old maps */
1562 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1564 if (do_munmap(mm, addr, len))
1569 * Private writable mapping: check memory availability
1571 if (accountable_mapping(file, vm_flags)) {
1572 charged = len >> PAGE_SHIFT;
1573 if (security_vm_enough_memory_mm(mm, charged))
1575 vm_flags |= VM_ACCOUNT;
1579 * Can we just expand an old mapping?
1581 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1582 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1587 * Determine the object being mapped and call the appropriate
1588 * specific mapper. the address has already been validated, but
1589 * not unmapped, but the maps are removed from the list.
1591 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1598 vma->vm_start = addr;
1599 vma->vm_end = addr + len;
1600 vma->vm_flags = vm_flags;
1601 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1602 vma->vm_pgoff = pgoff;
1603 INIT_LIST_HEAD(&vma->anon_vma_chain);
1606 if (vm_flags & VM_DENYWRITE) {
1607 error = deny_write_access(file);
1611 if (vm_flags & VM_SHARED) {
1612 error = mapping_map_writable(file->f_mapping);
1614 goto allow_write_and_free_vma;
1617 /* ->mmap() can change vma->vm_file, but must guarantee that
1618 * vma_link() below can deny write-access if VM_DENYWRITE is set
1619 * and map writably if VM_SHARED is set. This usually means the
1620 * new file must not have been exposed to user-space, yet.
1622 vma->vm_file = get_file(file);
1623 error = file->f_op->mmap(file, vma);
1625 goto unmap_and_free_vma;
1627 /* Can addr have changed??
1629 * Answer: Yes, several device drivers can do it in their
1630 * f_op->mmap method. -DaveM
1631 * Bug: If addr is changed, prev, rb_link, rb_parent should
1632 * be updated for vma_link()
1634 WARN_ON_ONCE(addr != vma->vm_start);
1636 addr = vma->vm_start;
1637 vm_flags = vma->vm_flags;
1638 } else if (vm_flags & VM_SHARED) {
1639 error = shmem_zero_setup(vma);
1644 vma_link(mm, vma, prev, rb_link, rb_parent);
1645 /* Once vma denies write, undo our temporary denial count */
1647 if (vm_flags & VM_SHARED)
1648 mapping_unmap_writable(file->f_mapping);
1649 if (vm_flags & VM_DENYWRITE)
1650 allow_write_access(file);
1652 file = vma->vm_file;
1654 perf_event_mmap(vma);
1656 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1657 if (vm_flags & VM_LOCKED) {
1658 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1659 vma == get_gate_vma(current->mm)))
1660 mm->locked_vm += (len >> PAGE_SHIFT);
1662 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1669 * New (or expanded) vma always get soft dirty status.
1670 * Otherwise user-space soft-dirty page tracker won't
1671 * be able to distinguish situation when vma area unmapped,
1672 * then new mapped in-place (which must be aimed as
1673 * a completely new data area).
1675 vma->vm_flags |= VM_SOFTDIRTY;
1677 vma_set_page_prot(vma);
1682 vma->vm_file = NULL;
1685 /* Undo any partial mapping done by a device driver. */
1686 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1688 if (vm_flags & VM_SHARED)
1689 mapping_unmap_writable(file->f_mapping);
1690 allow_write_and_free_vma:
1691 if (vm_flags & VM_DENYWRITE)
1692 allow_write_access(file);
1694 kmem_cache_free(vm_area_cachep, vma);
1697 vm_unacct_memory(charged);
1701 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1704 * We implement the search by looking for an rbtree node that
1705 * immediately follows a suitable gap. That is,
1706 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1707 * - gap_end = vma->vm_start >= info->low_limit + length;
1708 * - gap_end - gap_start >= length
1711 struct mm_struct *mm = current->mm;
1712 struct vm_area_struct *vma;
1713 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1715 /* Adjust search length to account for worst case alignment overhead */
1716 length = info->length + info->align_mask;
1717 if (length < info->length)
1720 /* Adjust search limits by the desired length */
1721 if (info->high_limit < length)
1723 high_limit = info->high_limit - length;
1725 if (info->low_limit > high_limit)
1727 low_limit = info->low_limit + length;
1729 /* Check if rbtree root looks promising */
1730 if (RB_EMPTY_ROOT(&mm->mm_rb))
1732 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1733 if (vma->rb_subtree_gap < length)
1737 /* Visit left subtree if it looks promising */
1738 gap_end = vma->vm_start;
1739 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1740 struct vm_area_struct *left =
1741 rb_entry(vma->vm_rb.rb_left,
1742 struct vm_area_struct, vm_rb);
1743 if (left->rb_subtree_gap >= length) {
1749 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1751 /* Check if current node has a suitable gap */
1752 if (gap_start > high_limit)
1754 if (gap_end >= low_limit && gap_end - gap_start >= length)
1757 /* Visit right subtree if it looks promising */
1758 if (vma->vm_rb.rb_right) {
1759 struct vm_area_struct *right =
1760 rb_entry(vma->vm_rb.rb_right,
1761 struct vm_area_struct, vm_rb);
1762 if (right->rb_subtree_gap >= length) {
1768 /* Go back up the rbtree to find next candidate node */
1770 struct rb_node *prev = &vma->vm_rb;
1771 if (!rb_parent(prev))
1773 vma = rb_entry(rb_parent(prev),
1774 struct vm_area_struct, vm_rb);
1775 if (prev == vma->vm_rb.rb_left) {
1776 gap_start = vma->vm_prev->vm_end;
1777 gap_end = vma->vm_start;
1784 /* Check highest gap, which does not precede any rbtree node */
1785 gap_start = mm->highest_vm_end;
1786 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1787 if (gap_start > high_limit)
1791 /* We found a suitable gap. Clip it with the original low_limit. */
1792 if (gap_start < info->low_limit)
1793 gap_start = info->low_limit;
1795 /* Adjust gap address to the desired alignment */
1796 gap_start += (info->align_offset - gap_start) & info->align_mask;
1798 VM_BUG_ON(gap_start + info->length > info->high_limit);
1799 VM_BUG_ON(gap_start + info->length > gap_end);
1803 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1805 struct mm_struct *mm = current->mm;
1806 struct vm_area_struct *vma;
1807 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1809 /* Adjust search length to account for worst case alignment overhead */
1810 length = info->length + info->align_mask;
1811 if (length < info->length)
1815 * Adjust search limits by the desired length.
1816 * See implementation comment at top of unmapped_area().
1818 gap_end = info->high_limit;
1819 if (gap_end < length)
1821 high_limit = gap_end - length;
1823 if (info->low_limit > high_limit)
1825 low_limit = info->low_limit + length;
1827 /* Check highest gap, which does not precede any rbtree node */
1828 gap_start = mm->highest_vm_end;
1829 if (gap_start <= high_limit)
1832 /* Check if rbtree root looks promising */
1833 if (RB_EMPTY_ROOT(&mm->mm_rb))
1835 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1836 if (vma->rb_subtree_gap < length)
1840 /* Visit right subtree if it looks promising */
1841 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1842 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1843 struct vm_area_struct *right =
1844 rb_entry(vma->vm_rb.rb_right,
1845 struct vm_area_struct, vm_rb);
1846 if (right->rb_subtree_gap >= length) {
1853 /* Check if current node has a suitable gap */
1854 gap_end = vma->vm_start;
1855 if (gap_end < low_limit)
1857 if (gap_start <= high_limit && gap_end - gap_start >= length)
1860 /* Visit left subtree if it looks promising */
1861 if (vma->vm_rb.rb_left) {
1862 struct vm_area_struct *left =
1863 rb_entry(vma->vm_rb.rb_left,
1864 struct vm_area_struct, vm_rb);
1865 if (left->rb_subtree_gap >= length) {
1871 /* Go back up the rbtree to find next candidate node */
1873 struct rb_node *prev = &vma->vm_rb;
1874 if (!rb_parent(prev))
1876 vma = rb_entry(rb_parent(prev),
1877 struct vm_area_struct, vm_rb);
1878 if (prev == vma->vm_rb.rb_right) {
1879 gap_start = vma->vm_prev ?
1880 vma->vm_prev->vm_end : 0;
1887 /* We found a suitable gap. Clip it with the original high_limit. */
1888 if (gap_end > info->high_limit)
1889 gap_end = info->high_limit;
1892 /* Compute highest gap address at the desired alignment */
1893 gap_end -= info->length;
1894 gap_end -= (gap_end - info->align_offset) & info->align_mask;
1896 VM_BUG_ON(gap_end < info->low_limit);
1897 VM_BUG_ON(gap_end < gap_start);
1901 /* Get an address range which is currently unmapped.
1902 * For shmat() with addr=0.
1904 * Ugly calling convention alert:
1905 * Return value with the low bits set means error value,
1907 * if (ret & ~PAGE_MASK)
1910 * This function "knows" that -ENOMEM has the bits set.
1912 #ifndef HAVE_ARCH_UNMAPPED_AREA
1914 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1915 unsigned long len, unsigned long pgoff, unsigned long flags)
1917 struct mm_struct *mm = current->mm;
1918 struct vm_area_struct *vma;
1919 struct vm_unmapped_area_info info;
1921 if (len > TASK_SIZE - mmap_min_addr)
1924 if (flags & MAP_FIXED)
1928 addr = PAGE_ALIGN(addr);
1929 vma = find_vma(mm, addr);
1930 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1931 (!vma || addr + len <= vma->vm_start))
1937 info.low_limit = mm->mmap_base;
1938 info.high_limit = TASK_SIZE;
1939 info.align_mask = 0;
1940 return vm_unmapped_area(&info);
1945 * This mmap-allocator allocates new areas top-down from below the
1946 * stack's low limit (the base):
1948 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1950 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1951 const unsigned long len, const unsigned long pgoff,
1952 const unsigned long flags)
1954 struct vm_area_struct *vma;
1955 struct mm_struct *mm = current->mm;
1956 unsigned long addr = addr0;
1957 struct vm_unmapped_area_info info;
1959 /* requested length too big for entire address space */
1960 if (len > TASK_SIZE - mmap_min_addr)
1963 if (flags & MAP_FIXED)
1966 /* requesting a specific address */
1968 addr = PAGE_ALIGN(addr);
1969 vma = find_vma(mm, addr);
1970 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1971 (!vma || addr + len <= vma->vm_start))
1975 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1977 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1978 info.high_limit = mm->mmap_base;
1979 info.align_mask = 0;
1980 addr = vm_unmapped_area(&info);
1983 * A failed mmap() very likely causes application failure,
1984 * so fall back to the bottom-up function here. This scenario
1985 * can happen with large stack limits and large mmap()
1988 if (offset_in_page(addr)) {
1989 VM_BUG_ON(addr != -ENOMEM);
1991 info.low_limit = TASK_UNMAPPED_BASE;
1992 info.high_limit = TASK_SIZE;
1993 addr = vm_unmapped_area(&info);
2001 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2002 unsigned long pgoff, unsigned long flags)
2004 unsigned long (*get_area)(struct file *, unsigned long,
2005 unsigned long, unsigned long, unsigned long);
2007 unsigned long error = arch_mmap_check(addr, len, flags);
2011 /* Careful about overflows.. */
2012 if (len > TASK_SIZE)
2015 get_area = current->mm->get_unmapped_area;
2016 if (file && file->f_op->get_unmapped_area)
2017 get_area = file->f_op->get_unmapped_area;
2018 addr = get_area(file, addr, len, pgoff, flags);
2019 if (IS_ERR_VALUE(addr))
2022 if (addr > TASK_SIZE - len)
2024 if (offset_in_page(addr))
2027 addr = arch_rebalance_pgtables(addr, len);
2028 error = security_mmap_addr(addr);
2029 return error ? error : addr;
2032 EXPORT_SYMBOL(get_unmapped_area);
2034 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2035 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2037 struct rb_node *rb_node;
2038 struct vm_area_struct *vma;
2040 /* Check the cache first. */
2041 vma = vmacache_find(mm, addr);
2045 rb_node = mm->mm_rb.rb_node;
2048 struct vm_area_struct *tmp;
2050 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2052 if (tmp->vm_end > addr) {
2054 if (tmp->vm_start <= addr)
2056 rb_node = rb_node->rb_left;
2058 rb_node = rb_node->rb_right;
2062 vmacache_update(addr, vma);
2066 EXPORT_SYMBOL(find_vma);
2069 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2071 struct vm_area_struct *
2072 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2073 struct vm_area_struct **pprev)
2075 struct vm_area_struct *vma;
2077 vma = find_vma(mm, addr);
2079 *pprev = vma->vm_prev;
2081 struct rb_node *rb_node = mm->mm_rb.rb_node;
2084 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2085 rb_node = rb_node->rb_right;
2092 * Verify that the stack growth is acceptable and
2093 * update accounting. This is shared with both the
2094 * grow-up and grow-down cases.
2096 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2098 struct mm_struct *mm = vma->vm_mm;
2099 struct rlimit *rlim = current->signal->rlim;
2100 unsigned long new_start, actual_size;
2102 /* address space limit tests */
2103 if (!may_expand_vm(mm, vma->vm_flags, grow))
2106 /* Stack limit test */
2108 if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN)))
2109 actual_size -= PAGE_SIZE;
2110 if (actual_size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2113 /* mlock limit tests */
2114 if (vma->vm_flags & VM_LOCKED) {
2115 unsigned long locked;
2116 unsigned long limit;
2117 locked = mm->locked_vm + grow;
2118 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2119 limit >>= PAGE_SHIFT;
2120 if (locked > limit && !capable(CAP_IPC_LOCK))
2124 /* Check to ensure the stack will not grow into a hugetlb-only region */
2125 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2127 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2131 * Overcommit.. This must be the final test, as it will
2132 * update security statistics.
2134 if (security_vm_enough_memory_mm(mm, grow))
2140 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2142 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2143 * vma is the last one with address > vma->vm_end. Have to extend vma.
2145 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2147 struct mm_struct *mm = vma->vm_mm;
2150 if (!(vma->vm_flags & VM_GROWSUP))
2154 * We must make sure the anon_vma is allocated
2155 * so that the anon_vma locking is not a noop.
2157 if (unlikely(anon_vma_prepare(vma)))
2159 vma_lock_anon_vma(vma);
2162 * vma->vm_start/vm_end cannot change under us because the caller
2163 * is required to hold the mmap_sem in read mode. We need the
2164 * anon_vma lock to serialize against concurrent expand_stacks.
2165 * Also guard against wrapping around to address 0.
2167 if (address < PAGE_ALIGN(address+4))
2168 address = PAGE_ALIGN(address+4);
2170 vma_unlock_anon_vma(vma);
2175 /* Somebody else might have raced and expanded it already */
2176 if (address > vma->vm_end) {
2177 unsigned long size, grow;
2179 size = address - vma->vm_start;
2180 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2183 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2184 error = acct_stack_growth(vma, size, grow);
2187 * vma_gap_update() doesn't support concurrent
2188 * updates, but we only hold a shared mmap_sem
2189 * lock here, so we need to protect against
2190 * concurrent vma expansions.
2191 * vma_lock_anon_vma() doesn't help here, as
2192 * we don't guarantee that all growable vmas
2193 * in a mm share the same root anon vma.
2194 * So, we reuse mm->page_table_lock to guard
2195 * against concurrent vma expansions.
2197 spin_lock(&mm->page_table_lock);
2198 if (vma->vm_flags & VM_LOCKED)
2199 mm->locked_vm += grow;
2200 vm_stat_account(mm, vma->vm_flags, grow);
2201 anon_vma_interval_tree_pre_update_vma(vma);
2202 vma->vm_end = address;
2203 anon_vma_interval_tree_post_update_vma(vma);
2205 vma_gap_update(vma->vm_next);
2207 mm->highest_vm_end = address;
2208 spin_unlock(&mm->page_table_lock);
2210 perf_event_mmap(vma);
2214 vma_unlock_anon_vma(vma);
2215 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2219 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2222 * vma is the first one with address < vma->vm_start. Have to extend vma.
2224 int expand_downwards(struct vm_area_struct *vma,
2225 unsigned long address)
2227 struct mm_struct *mm = vma->vm_mm;
2231 * We must make sure the anon_vma is allocated
2232 * so that the anon_vma locking is not a noop.
2234 if (unlikely(anon_vma_prepare(vma)))
2237 address &= PAGE_MASK;
2238 error = security_mmap_addr(address);
2242 vma_lock_anon_vma(vma);
2245 * vma->vm_start/vm_end cannot change under us because the caller
2246 * is required to hold the mmap_sem in read mode. We need the
2247 * anon_vma lock to serialize against concurrent expand_stacks.
2250 /* Somebody else might have raced and expanded it already */
2251 if (address < vma->vm_start) {
2252 unsigned long size, grow;
2254 size = vma->vm_end - address;
2255 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2258 if (grow <= vma->vm_pgoff) {
2259 error = acct_stack_growth(vma, size, grow);
2262 * vma_gap_update() doesn't support concurrent
2263 * updates, but we only hold a shared mmap_sem
2264 * lock here, so we need to protect against
2265 * concurrent vma expansions.
2266 * vma_lock_anon_vma() doesn't help here, as
2267 * we don't guarantee that all growable vmas
2268 * in a mm share the same root anon vma.
2269 * So, we reuse mm->page_table_lock to guard
2270 * against concurrent vma expansions.
2272 spin_lock(&mm->page_table_lock);
2273 if (vma->vm_flags & VM_LOCKED)
2274 mm->locked_vm += grow;
2275 vm_stat_account(mm, vma->vm_flags, grow);
2276 anon_vma_interval_tree_pre_update_vma(vma);
2277 vma->vm_start = address;
2278 vma->vm_pgoff -= grow;
2279 anon_vma_interval_tree_post_update_vma(vma);
2280 vma_gap_update(vma);
2281 spin_unlock(&mm->page_table_lock);
2283 perf_event_mmap(vma);
2287 vma_unlock_anon_vma(vma);
2288 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2294 * Note how expand_stack() refuses to expand the stack all the way to
2295 * abut the next virtual mapping, *unless* that mapping itself is also
2296 * a stack mapping. We want to leave room for a guard page, after all
2297 * (the guard page itself is not added here, that is done by the
2298 * actual page faulting logic)
2300 * This matches the behavior of the guard page logic (see mm/memory.c:
2301 * check_stack_guard_page()), which only allows the guard page to be
2302 * removed under these circumstances.
2304 #ifdef CONFIG_STACK_GROWSUP
2305 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2307 struct vm_area_struct *next;
2309 address &= PAGE_MASK;
2310 next = vma->vm_next;
2311 if (next && next->vm_start == address + PAGE_SIZE) {
2312 if (!(next->vm_flags & VM_GROWSUP))
2315 return expand_upwards(vma, address);
2318 struct vm_area_struct *
2319 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2321 struct vm_area_struct *vma, *prev;
2324 vma = find_vma_prev(mm, addr, &prev);
2325 if (vma && (vma->vm_start <= addr))
2327 if (!prev || expand_stack(prev, addr))
2329 if (prev->vm_flags & VM_LOCKED)
2330 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2334 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2336 struct vm_area_struct *prev;
2338 address &= PAGE_MASK;
2339 prev = vma->vm_prev;
2340 if (prev && prev->vm_end == address) {
2341 if (!(prev->vm_flags & VM_GROWSDOWN))
2344 return expand_downwards(vma, address);
2347 struct vm_area_struct *
2348 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2350 struct vm_area_struct *vma;
2351 unsigned long start;
2354 vma = find_vma(mm, addr);
2357 if (vma->vm_start <= addr)
2359 if (!(vma->vm_flags & VM_GROWSDOWN))
2361 start = vma->vm_start;
2362 if (expand_stack(vma, addr))
2364 if (vma->vm_flags & VM_LOCKED)
2365 populate_vma_page_range(vma, addr, start, NULL);
2370 EXPORT_SYMBOL_GPL(find_extend_vma);
2373 * Ok - we have the memory areas we should free on the vma list,
2374 * so release them, and do the vma updates.
2376 * Called with the mm semaphore held.
2378 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2380 unsigned long nr_accounted = 0;
2382 /* Update high watermark before we lower total_vm */
2383 update_hiwater_vm(mm);
2385 long nrpages = vma_pages(vma);
2387 if (vma->vm_flags & VM_ACCOUNT)
2388 nr_accounted += nrpages;
2389 vm_stat_account(mm, vma->vm_flags, -nrpages);
2390 vma = remove_vma(vma);
2392 vm_unacct_memory(nr_accounted);
2397 * Get rid of page table information in the indicated region.
2399 * Called with the mm semaphore held.
2401 static void unmap_region(struct mm_struct *mm,
2402 struct vm_area_struct *vma, struct vm_area_struct *prev,
2403 unsigned long start, unsigned long end)
2405 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2406 struct mmu_gather tlb;
2409 tlb_gather_mmu(&tlb, mm, start, end);
2410 update_hiwater_rss(mm);
2411 unmap_vmas(&tlb, vma, start, end);
2412 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2413 next ? next->vm_start : USER_PGTABLES_CEILING);
2414 tlb_finish_mmu(&tlb, start, end);
2418 * Create a list of vma's touched by the unmap, removing them from the mm's
2419 * vma list as we go..
2422 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2423 struct vm_area_struct *prev, unsigned long end)
2425 struct vm_area_struct **insertion_point;
2426 struct vm_area_struct *tail_vma = NULL;
2428 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2429 vma->vm_prev = NULL;
2431 vma_rb_erase(vma, &mm->mm_rb);
2435 } while (vma && vma->vm_start < end);
2436 *insertion_point = vma;
2438 vma->vm_prev = prev;
2439 vma_gap_update(vma);
2441 mm->highest_vm_end = prev ? prev->vm_end : 0;
2442 tail_vma->vm_next = NULL;
2444 /* Kill the cache */
2445 vmacache_invalidate(mm);
2449 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2450 * munmap path where it doesn't make sense to fail.
2452 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2453 unsigned long addr, int new_below)
2455 struct vm_area_struct *new;
2458 if (is_vm_hugetlb_page(vma) && (addr &
2459 ~(huge_page_mask(hstate_vma(vma)))))
2462 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2466 /* most fields are the same, copy all, and then fixup */
2469 INIT_LIST_HEAD(&new->anon_vma_chain);
2474 new->vm_start = addr;
2475 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2478 err = vma_dup_policy(vma, new);
2482 err = anon_vma_clone(new, vma);
2487 get_file(new->vm_file);
2489 if (new->vm_ops && new->vm_ops->open)
2490 new->vm_ops->open(new);
2493 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2494 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2496 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2502 /* Clean everything up if vma_adjust failed. */
2503 if (new->vm_ops && new->vm_ops->close)
2504 new->vm_ops->close(new);
2507 unlink_anon_vmas(new);
2509 mpol_put(vma_policy(new));
2511 kmem_cache_free(vm_area_cachep, new);
2516 * Split a vma into two pieces at address 'addr', a new vma is allocated
2517 * either for the first part or the tail.
2519 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2520 unsigned long addr, int new_below)
2522 if (mm->map_count >= sysctl_max_map_count)
2525 return __split_vma(mm, vma, addr, new_below);
2528 /* Munmap is split into 2 main parts -- this part which finds
2529 * what needs doing, and the areas themselves, which do the
2530 * work. This now handles partial unmappings.
2531 * Jeremy Fitzhardinge <jeremy@goop.org>
2533 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2536 struct vm_area_struct *vma, *prev, *last;
2538 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2541 len = PAGE_ALIGN(len);
2545 /* Find the first overlapping VMA */
2546 vma = find_vma(mm, start);
2549 prev = vma->vm_prev;
2550 /* we have start < vma->vm_end */
2552 /* if it doesn't overlap, we have nothing.. */
2554 if (vma->vm_start >= end)
2558 * If we need to split any vma, do it now to save pain later.
2560 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2561 * unmapped vm_area_struct will remain in use: so lower split_vma
2562 * places tmp vma above, and higher split_vma places tmp vma below.
2564 if (start > vma->vm_start) {
2568 * Make sure that map_count on return from munmap() will
2569 * not exceed its limit; but let map_count go just above
2570 * its limit temporarily, to help free resources as expected.
2572 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2575 error = __split_vma(mm, vma, start, 0);
2581 /* Does it split the last one? */
2582 last = find_vma(mm, end);
2583 if (last && end > last->vm_start) {
2584 int error = __split_vma(mm, last, end, 1);
2588 vma = prev ? prev->vm_next : mm->mmap;
2591 * unlock any mlock()ed ranges before detaching vmas
2593 if (mm->locked_vm) {
2594 struct vm_area_struct *tmp = vma;
2595 while (tmp && tmp->vm_start < end) {
2596 if (tmp->vm_flags & VM_LOCKED) {
2597 mm->locked_vm -= vma_pages(tmp);
2598 munlock_vma_pages_all(tmp);
2605 * Remove the vma's, and unmap the actual pages
2607 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2608 unmap_region(mm, vma, prev, start, end);
2610 arch_unmap(mm, vma, start, end);
2612 /* Fix up all other VM information */
2613 remove_vma_list(mm, vma);
2618 int vm_munmap(unsigned long start, size_t len)
2621 struct mm_struct *mm = current->mm;
2623 down_write(&mm->mmap_sem);
2624 ret = do_munmap(mm, start, len);
2625 up_write(&mm->mmap_sem);
2628 EXPORT_SYMBOL(vm_munmap);
2630 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2632 profile_munmap(addr);
2633 return vm_munmap(addr, len);
2638 * Emulation of deprecated remap_file_pages() syscall.
2640 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2641 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2644 struct mm_struct *mm = current->mm;
2645 struct vm_area_struct *vma;
2646 unsigned long populate = 0;
2647 unsigned long ret = -EINVAL;
2650 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. "
2651 "See Documentation/vm/remap_file_pages.txt.\n",
2652 current->comm, current->pid);
2656 start = start & PAGE_MASK;
2657 size = size & PAGE_MASK;
2659 if (start + size <= start)
2662 /* Does pgoff wrap? */
2663 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2666 down_write(&mm->mmap_sem);
2667 vma = find_vma(mm, start);
2669 if (!vma || !(vma->vm_flags & VM_SHARED))
2672 if (start < vma->vm_start || start + size > vma->vm_end)
2675 if (pgoff == linear_page_index(vma, start)) {
2680 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2681 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2682 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2684 flags &= MAP_NONBLOCK;
2685 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2686 if (vma->vm_flags & VM_LOCKED) {
2687 flags |= MAP_LOCKED;
2688 /* drop PG_Mlocked flag for over-mapped range */
2689 munlock_vma_pages_range(vma, start, start + size);
2692 file = get_file(vma->vm_file);
2693 ret = do_mmap_pgoff(vma->vm_file, start, size,
2694 prot, flags, pgoff, &populate);
2697 up_write(&mm->mmap_sem);
2699 mm_populate(ret, populate);
2700 if (!IS_ERR_VALUE(ret))
2705 static inline void verify_mm_writelocked(struct mm_struct *mm)
2707 #ifdef CONFIG_DEBUG_VM
2708 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2710 up_read(&mm->mmap_sem);
2716 * this is really a simplified "do_mmap". it only handles
2717 * anonymous maps. eventually we may be able to do some
2718 * brk-specific accounting here.
2720 static unsigned long do_brk(unsigned long addr, unsigned long len)
2722 struct mm_struct *mm = current->mm;
2723 struct vm_area_struct *vma, *prev;
2724 unsigned long flags;
2725 struct rb_node **rb_link, *rb_parent;
2726 pgoff_t pgoff = addr >> PAGE_SHIFT;
2729 len = PAGE_ALIGN(len);
2733 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2735 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2736 if (offset_in_page(error))
2739 error = mlock_future_check(mm, mm->def_flags, len);
2744 * mm->mmap_sem is required to protect against another thread
2745 * changing the mappings in case we sleep.
2747 verify_mm_writelocked(mm);
2750 * Clear old maps. this also does some error checking for us
2752 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2754 if (do_munmap(mm, addr, len))
2758 /* Check against address space limits *after* clearing old maps... */
2759 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2762 if (mm->map_count > sysctl_max_map_count)
2765 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2768 /* Can we just expand an old private anonymous mapping? */
2769 vma = vma_merge(mm, prev, addr, addr + len, flags,
2770 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2775 * create a vma struct for an anonymous mapping
2777 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2779 vm_unacct_memory(len >> PAGE_SHIFT);
2783 INIT_LIST_HEAD(&vma->anon_vma_chain);
2785 vma->vm_start = addr;
2786 vma->vm_end = addr + len;
2787 vma->vm_pgoff = pgoff;
2788 vma->vm_flags = flags;
2789 vma->vm_page_prot = vm_get_page_prot(flags);
2790 vma_link(mm, vma, prev, rb_link, rb_parent);
2792 perf_event_mmap(vma);
2793 mm->total_vm += len >> PAGE_SHIFT;
2794 mm->data_vm += len >> PAGE_SHIFT;
2795 if (flags & VM_LOCKED)
2796 mm->locked_vm += (len >> PAGE_SHIFT);
2797 vma->vm_flags |= VM_SOFTDIRTY;
2801 unsigned long vm_brk(unsigned long addr, unsigned long len)
2803 struct mm_struct *mm = current->mm;
2807 down_write(&mm->mmap_sem);
2808 ret = do_brk(addr, len);
2809 populate = ((mm->def_flags & VM_LOCKED) != 0);
2810 up_write(&mm->mmap_sem);
2812 mm_populate(addr, len);
2815 EXPORT_SYMBOL(vm_brk);
2817 /* Release all mmaps. */
2818 void exit_mmap(struct mm_struct *mm)
2820 struct mmu_gather tlb;
2821 struct vm_area_struct *vma;
2822 unsigned long nr_accounted = 0;
2824 /* mm's last user has gone, and its about to be pulled down */
2825 mmu_notifier_release(mm);
2827 if (mm->locked_vm) {
2830 if (vma->vm_flags & VM_LOCKED)
2831 munlock_vma_pages_all(vma);
2839 if (!vma) /* Can happen if dup_mmap() received an OOM */
2844 tlb_gather_mmu(&tlb, mm, 0, -1);
2845 /* update_hiwater_rss(mm) here? but nobody should be looking */
2846 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2847 unmap_vmas(&tlb, vma, 0, -1);
2849 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2850 tlb_finish_mmu(&tlb, 0, -1);
2853 * Walk the list again, actually closing and freeing it,
2854 * with preemption enabled, without holding any MM locks.
2857 if (vma->vm_flags & VM_ACCOUNT)
2858 nr_accounted += vma_pages(vma);
2859 vma = remove_vma(vma);
2861 vm_unacct_memory(nr_accounted);
2864 /* Insert vm structure into process list sorted by address
2865 * and into the inode's i_mmap tree. If vm_file is non-NULL
2866 * then i_mmap_rwsem is taken here.
2868 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2870 struct vm_area_struct *prev;
2871 struct rb_node **rb_link, *rb_parent;
2873 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2874 &prev, &rb_link, &rb_parent))
2876 if ((vma->vm_flags & VM_ACCOUNT) &&
2877 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2881 * The vm_pgoff of a purely anonymous vma should be irrelevant
2882 * until its first write fault, when page's anon_vma and index
2883 * are set. But now set the vm_pgoff it will almost certainly
2884 * end up with (unless mremap moves it elsewhere before that
2885 * first wfault), so /proc/pid/maps tells a consistent story.
2887 * By setting it to reflect the virtual start address of the
2888 * vma, merges and splits can happen in a seamless way, just
2889 * using the existing file pgoff checks and manipulations.
2890 * Similarly in do_mmap_pgoff and in do_brk.
2892 if (vma_is_anonymous(vma)) {
2893 BUG_ON(vma->anon_vma);
2894 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2897 vma_link(mm, vma, prev, rb_link, rb_parent);
2902 * Copy the vma structure to a new location in the same mm,
2903 * prior to moving page table entries, to effect an mremap move.
2905 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2906 unsigned long addr, unsigned long len, pgoff_t pgoff,
2907 bool *need_rmap_locks)
2909 struct vm_area_struct *vma = *vmap;
2910 unsigned long vma_start = vma->vm_start;
2911 struct mm_struct *mm = vma->vm_mm;
2912 struct vm_area_struct *new_vma, *prev;
2913 struct rb_node **rb_link, *rb_parent;
2914 bool faulted_in_anon_vma = true;
2917 * If anonymous vma has not yet been faulted, update new pgoff
2918 * to match new location, to increase its chance of merging.
2920 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
2921 pgoff = addr >> PAGE_SHIFT;
2922 faulted_in_anon_vma = false;
2925 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2926 return NULL; /* should never get here */
2927 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2928 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
2929 vma->vm_userfaultfd_ctx);
2932 * Source vma may have been merged into new_vma
2934 if (unlikely(vma_start >= new_vma->vm_start &&
2935 vma_start < new_vma->vm_end)) {
2937 * The only way we can get a vma_merge with
2938 * self during an mremap is if the vma hasn't
2939 * been faulted in yet and we were allowed to
2940 * reset the dst vma->vm_pgoff to the
2941 * destination address of the mremap to allow
2942 * the merge to happen. mremap must change the
2943 * vm_pgoff linearity between src and dst vmas
2944 * (in turn preventing a vma_merge) to be
2945 * safe. It is only safe to keep the vm_pgoff
2946 * linear if there are no pages mapped yet.
2948 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
2949 *vmap = vma = new_vma;
2951 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2953 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2957 new_vma->vm_start = addr;
2958 new_vma->vm_end = addr + len;
2959 new_vma->vm_pgoff = pgoff;
2960 if (vma_dup_policy(vma, new_vma))
2962 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2963 if (anon_vma_clone(new_vma, vma))
2964 goto out_free_mempol;
2965 if (new_vma->vm_file)
2966 get_file(new_vma->vm_file);
2967 if (new_vma->vm_ops && new_vma->vm_ops->open)
2968 new_vma->vm_ops->open(new_vma);
2969 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2970 *need_rmap_locks = false;
2975 mpol_put(vma_policy(new_vma));
2977 kmem_cache_free(vm_area_cachep, new_vma);
2983 * Return true if the calling process may expand its vm space by the passed
2986 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
2988 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
2991 if (is_data_mapping(flags) &&
2992 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
2993 if (ignore_rlimit_data)
2994 pr_warn_once("%s (%d): VmData %lu exceed data ulimit "
2995 "%lu. Will be forbidden soon.\n",
2996 current->comm, current->pid,
2997 (mm->data_vm + npages) << PAGE_SHIFT,
2998 rlimit(RLIMIT_DATA));
3006 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3008 mm->total_vm += npages;
3010 if (is_exec_mapping(flags))
3011 mm->exec_vm += npages;
3012 else if (is_stack_mapping(flags))
3013 mm->stack_vm += npages;
3014 else if (is_data_mapping(flags))
3015 mm->data_vm += npages;
3018 static int special_mapping_fault(struct vm_area_struct *vma,
3019 struct vm_fault *vmf);
3022 * Having a close hook prevents vma merging regardless of flags.
3024 static void special_mapping_close(struct vm_area_struct *vma)
3028 static const char *special_mapping_name(struct vm_area_struct *vma)
3030 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3033 static const struct vm_operations_struct special_mapping_vmops = {
3034 .close = special_mapping_close,
3035 .fault = special_mapping_fault,
3036 .name = special_mapping_name,
3039 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3040 .close = special_mapping_close,
3041 .fault = special_mapping_fault,
3044 static int special_mapping_fault(struct vm_area_struct *vma,
3045 struct vm_fault *vmf)
3048 struct page **pages;
3050 if (vma->vm_ops == &legacy_special_mapping_vmops)
3051 pages = vma->vm_private_data;
3053 pages = ((struct vm_special_mapping *)vma->vm_private_data)->
3056 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3060 struct page *page = *pages;
3066 return VM_FAULT_SIGBUS;
3069 static struct vm_area_struct *__install_special_mapping(
3070 struct mm_struct *mm,
3071 unsigned long addr, unsigned long len,
3072 unsigned long vm_flags, void *priv,
3073 const struct vm_operations_struct *ops)
3076 struct vm_area_struct *vma;
3078 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3079 if (unlikely(vma == NULL))
3080 return ERR_PTR(-ENOMEM);
3082 INIT_LIST_HEAD(&vma->anon_vma_chain);
3084 vma->vm_start = addr;
3085 vma->vm_end = addr + len;
3087 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3088 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3091 vma->vm_private_data = priv;
3093 ret = insert_vm_struct(mm, vma);
3097 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3099 perf_event_mmap(vma);
3104 kmem_cache_free(vm_area_cachep, vma);
3105 return ERR_PTR(ret);
3109 * Called with mm->mmap_sem held for writing.
3110 * Insert a new vma covering the given region, with the given flags.
3111 * Its pages are supplied by the given array of struct page *.
3112 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3113 * The region past the last page supplied will always produce SIGBUS.
3114 * The array pointer and the pages it points to are assumed to stay alive
3115 * for as long as this mapping might exist.
3117 struct vm_area_struct *_install_special_mapping(
3118 struct mm_struct *mm,
3119 unsigned long addr, unsigned long len,
3120 unsigned long vm_flags, const struct vm_special_mapping *spec)
3122 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3123 &special_mapping_vmops);
3126 int install_special_mapping(struct mm_struct *mm,
3127 unsigned long addr, unsigned long len,
3128 unsigned long vm_flags, struct page **pages)
3130 struct vm_area_struct *vma = __install_special_mapping(
3131 mm, addr, len, vm_flags, (void *)pages,
3132 &legacy_special_mapping_vmops);
3134 return PTR_ERR_OR_ZERO(vma);
3137 static DEFINE_MUTEX(mm_all_locks_mutex);
3139 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3141 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3143 * The LSB of head.next can't change from under us
3144 * because we hold the mm_all_locks_mutex.
3146 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3148 * We can safely modify head.next after taking the
3149 * anon_vma->root->rwsem. If some other vma in this mm shares
3150 * the same anon_vma we won't take it again.
3152 * No need of atomic instructions here, head.next
3153 * can't change from under us thanks to the
3154 * anon_vma->root->rwsem.
3156 if (__test_and_set_bit(0, (unsigned long *)
3157 &anon_vma->root->rb_root.rb_node))
3162 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3164 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3166 * AS_MM_ALL_LOCKS can't change from under us because
3167 * we hold the mm_all_locks_mutex.
3169 * Operations on ->flags have to be atomic because
3170 * even if AS_MM_ALL_LOCKS is stable thanks to the
3171 * mm_all_locks_mutex, there may be other cpus
3172 * changing other bitflags in parallel to us.
3174 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3176 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3181 * This operation locks against the VM for all pte/vma/mm related
3182 * operations that could ever happen on a certain mm. This includes
3183 * vmtruncate, try_to_unmap, and all page faults.
3185 * The caller must take the mmap_sem in write mode before calling
3186 * mm_take_all_locks(). The caller isn't allowed to release the
3187 * mmap_sem until mm_drop_all_locks() returns.
3189 * mmap_sem in write mode is required in order to block all operations
3190 * that could modify pagetables and free pages without need of
3191 * altering the vma layout. It's also needed in write mode to avoid new
3192 * anon_vmas to be associated with existing vmas.
3194 * A single task can't take more than one mm_take_all_locks() in a row
3195 * or it would deadlock.
3197 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3198 * mapping->flags avoid to take the same lock twice, if more than one
3199 * vma in this mm is backed by the same anon_vma or address_space.
3201 * We take locks in following order, accordingly to comment at beginning
3203 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3205 * - all i_mmap_rwsem locks;
3206 * - all anon_vma->rwseml
3208 * We can take all locks within these types randomly because the VM code
3209 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3210 * mm_all_locks_mutex.
3212 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3213 * that may have to take thousand of locks.
3215 * mm_take_all_locks() can fail if it's interrupted by signals.
3217 int mm_take_all_locks(struct mm_struct *mm)
3219 struct vm_area_struct *vma;
3220 struct anon_vma_chain *avc;
3222 BUG_ON(down_read_trylock(&mm->mmap_sem));
3224 mutex_lock(&mm_all_locks_mutex);
3226 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3227 if (signal_pending(current))
3229 if (vma->vm_file && vma->vm_file->f_mapping &&
3230 is_vm_hugetlb_page(vma))
3231 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3234 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3235 if (signal_pending(current))
3237 if (vma->vm_file && vma->vm_file->f_mapping &&
3238 !is_vm_hugetlb_page(vma))
3239 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3242 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3243 if (signal_pending(current))
3246 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3247 vm_lock_anon_vma(mm, avc->anon_vma);
3253 mm_drop_all_locks(mm);
3257 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3259 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3261 * The LSB of head.next can't change to 0 from under
3262 * us because we hold the mm_all_locks_mutex.
3264 * We must however clear the bitflag before unlocking
3265 * the vma so the users using the anon_vma->rb_root will
3266 * never see our bitflag.
3268 * No need of atomic instructions here, head.next
3269 * can't change from under us until we release the
3270 * anon_vma->root->rwsem.
3272 if (!__test_and_clear_bit(0, (unsigned long *)
3273 &anon_vma->root->rb_root.rb_node))
3275 anon_vma_unlock_write(anon_vma);
3279 static void vm_unlock_mapping(struct address_space *mapping)
3281 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3283 * AS_MM_ALL_LOCKS can't change to 0 from under us
3284 * because we hold the mm_all_locks_mutex.
3286 i_mmap_unlock_write(mapping);
3287 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3294 * The mmap_sem cannot be released by the caller until
3295 * mm_drop_all_locks() returns.
3297 void mm_drop_all_locks(struct mm_struct *mm)
3299 struct vm_area_struct *vma;
3300 struct anon_vma_chain *avc;
3302 BUG_ON(down_read_trylock(&mm->mmap_sem));
3303 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3305 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3307 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3308 vm_unlock_anon_vma(avc->anon_vma);
3309 if (vma->vm_file && vma->vm_file->f_mapping)
3310 vm_unlock_mapping(vma->vm_file->f_mapping);
3313 mutex_unlock(&mm_all_locks_mutex);
3317 * initialise the VMA slab
3319 void __init mmap_init(void)
3323 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3328 * Initialise sysctl_user_reserve_kbytes.
3330 * This is intended to prevent a user from starting a single memory hogging
3331 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3334 * The default value is min(3% of free memory, 128MB)
3335 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3337 static int init_user_reserve(void)
3339 unsigned long free_kbytes;
3341 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3343 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3346 subsys_initcall(init_user_reserve);
3349 * Initialise sysctl_admin_reserve_kbytes.
3351 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3352 * to log in and kill a memory hogging process.
3354 * Systems with more than 256MB will reserve 8MB, enough to recover
3355 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3356 * only reserve 3% of free pages by default.
3358 static int init_admin_reserve(void)
3360 unsigned long free_kbytes;
3362 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3364 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3367 subsys_initcall(init_admin_reserve);
3370 * Reinititalise user and admin reserves if memory is added or removed.
3372 * The default user reserve max is 128MB, and the default max for the
3373 * admin reserve is 8MB. These are usually, but not always, enough to
3374 * enable recovery from a memory hogging process using login/sshd, a shell,
3375 * and tools like top. It may make sense to increase or even disable the
3376 * reserve depending on the existence of swap or variations in the recovery
3377 * tools. So, the admin may have changed them.
3379 * If memory is added and the reserves have been eliminated or increased above
3380 * the default max, then we'll trust the admin.
3382 * If memory is removed and there isn't enough free memory, then we
3383 * need to reset the reserves.
3385 * Otherwise keep the reserve set by the admin.
3387 static int reserve_mem_notifier(struct notifier_block *nb,
3388 unsigned long action, void *data)
3390 unsigned long tmp, free_kbytes;
3394 /* Default max is 128MB. Leave alone if modified by operator. */
3395 tmp = sysctl_user_reserve_kbytes;
3396 if (0 < tmp && tmp < (1UL << 17))
3397 init_user_reserve();
3399 /* Default max is 8MB. Leave alone if modified by operator. */
3400 tmp = sysctl_admin_reserve_kbytes;
3401 if (0 < tmp && tmp < (1UL << 13))
3402 init_admin_reserve();
3406 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3408 if (sysctl_user_reserve_kbytes > free_kbytes) {
3409 init_user_reserve();
3410 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3411 sysctl_user_reserve_kbytes);
3414 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3415 init_admin_reserve();
3416 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3417 sysctl_admin_reserve_kbytes);
3426 static struct notifier_block reserve_mem_nb = {
3427 .notifier_call = reserve_mem_notifier,
3430 static int __meminit init_reserve_notifier(void)
3432 if (register_hotmemory_notifier(&reserve_mem_nb))
3433 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3437 subsys_initcall(init_reserve_notifier);