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>
46 #include <asm/uaccess.h>
47 #include <asm/cacheflush.h>
49 #include <asm/mmu_context.h>
53 #ifndef arch_mmap_check
54 #define arch_mmap_check(addr, len, flags) (0)
57 #ifndef arch_rebalance_pgtables
58 #define arch_rebalance_pgtables(addr, len) (addr)
61 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
62 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
63 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
64 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
66 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
67 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
68 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
69 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
73 static void unmap_region(struct mm_struct *mm,
74 struct vm_area_struct *vma, struct vm_area_struct *prev,
75 unsigned long start, unsigned long end);
77 /* description of effects of mapping type and prot in current implementation.
78 * this is due to the limited x86 page protection hardware. The expected
79 * behavior is in parens:
82 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
83 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
84 * w: (no) no w: (no) no w: (yes) yes w: (no) no
85 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
87 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
88 * w: (no) no w: (no) no w: (copy) copy w: (no) no
89 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
92 pgprot_t protection_map[16] = {
93 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
94 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
97 pgprot_t vm_get_page_prot(unsigned long vm_flags)
99 return __pgprot(pgprot_val(protection_map[vm_flags &
100 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
101 pgprot_val(arch_vm_get_page_prot(vm_flags)));
103 EXPORT_SYMBOL(vm_get_page_prot);
105 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
107 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
110 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
111 void vma_set_page_prot(struct vm_area_struct *vma)
113 unsigned long vm_flags = vma->vm_flags;
115 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
116 if (vma_wants_writenotify(vma)) {
117 vm_flags &= ~VM_SHARED;
118 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot,
124 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */
125 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
126 unsigned long sysctl_overcommit_kbytes __read_mostly;
127 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
128 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
129 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
131 * Make sure vm_committed_as in one cacheline and not cacheline shared with
132 * other variables. It can be updated by several CPUs frequently.
134 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
137 * The global memory commitment made in the system can be a metric
138 * that can be used to drive ballooning decisions when Linux is hosted
139 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
140 * balancing memory across competing virtual machines that are hosted.
141 * Several metrics drive this policy engine including the guest reported
144 unsigned long vm_memory_committed(void)
146 return percpu_counter_read_positive(&vm_committed_as);
148 EXPORT_SYMBOL_GPL(vm_memory_committed);
151 * Check that a process has enough memory to allocate a new virtual
152 * mapping. 0 means there is enough memory for the allocation to
153 * succeed and -ENOMEM implies there is not.
155 * We currently support three overcommit policies, which are set via the
156 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
158 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
159 * Additional code 2002 Jul 20 by Robert Love.
161 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
163 * Note this is a helper function intended to be used by LSMs which
164 * wish to use this logic.
166 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
168 long free, allowed, reserve;
170 VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) <
171 -(s64)vm_committed_as_batch * num_online_cpus(),
172 "memory commitment underflow");
174 vm_acct_memory(pages);
177 * Sometimes we want to use more memory than we have
179 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
182 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
183 free = global_page_state(NR_FREE_PAGES);
184 free += global_page_state(NR_FILE_PAGES);
187 * shmem pages shouldn't be counted as free in this
188 * case, they can't be purged, only swapped out, and
189 * that won't affect the overall amount of available
190 * memory in the system.
192 free -= global_page_state(NR_SHMEM);
194 free += get_nr_swap_pages();
197 * Any slabs which are created with the
198 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
199 * which are reclaimable, under pressure. The dentry
200 * cache and most inode caches should fall into this
202 free += global_page_state(NR_SLAB_RECLAIMABLE);
205 * Leave reserved pages. The pages are not for anonymous pages.
207 if (free <= totalreserve_pages)
210 free -= totalreserve_pages;
213 * Reserve some for root
216 free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
224 allowed = vm_commit_limit();
226 * Reserve some for root
229 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
232 * Don't let a single process grow so big a user can't recover
235 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
236 allowed -= min_t(long, mm->total_vm / 32, reserve);
239 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
242 vm_unacct_memory(pages);
248 * Requires inode->i_mapping->i_mmap_rwsem
250 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
251 struct file *file, struct address_space *mapping)
253 if (vma->vm_flags & VM_DENYWRITE)
254 atomic_inc(&file_inode(file)->i_writecount);
255 if (vma->vm_flags & VM_SHARED)
256 mapping_unmap_writable(mapping);
258 flush_dcache_mmap_lock(mapping);
259 vma_interval_tree_remove(vma, &mapping->i_mmap);
260 flush_dcache_mmap_unlock(mapping);
264 * Unlink a file-based vm structure from its interval tree, to hide
265 * vma from rmap and vmtruncate before freeing its page tables.
267 void unlink_file_vma(struct vm_area_struct *vma)
269 struct file *file = vma->vm_file;
272 struct address_space *mapping = file->f_mapping;
273 i_mmap_lock_write(mapping);
274 __remove_shared_vm_struct(vma, file, mapping);
275 i_mmap_unlock_write(mapping);
280 * Close a vm structure and free it, returning the next.
282 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
284 struct vm_area_struct *next = vma->vm_next;
287 if (vma->vm_ops && vma->vm_ops->close)
288 vma->vm_ops->close(vma);
291 mpol_put(vma_policy(vma));
292 kmem_cache_free(vm_area_cachep, vma);
296 static unsigned long do_brk(unsigned long addr, unsigned long len);
298 SYSCALL_DEFINE1(brk, unsigned long, brk)
300 unsigned long retval;
301 unsigned long newbrk, oldbrk;
302 struct mm_struct *mm = current->mm;
303 struct vm_area_struct *next;
304 unsigned long min_brk;
307 down_write(&mm->mmap_sem);
309 #ifdef CONFIG_COMPAT_BRK
311 * CONFIG_COMPAT_BRK can still be overridden by setting
312 * randomize_va_space to 2, which will still cause mm->start_brk
313 * to be arbitrarily shifted
315 if (current->brk_randomized)
316 min_brk = mm->start_brk;
318 min_brk = mm->end_data;
320 min_brk = mm->start_brk;
326 * Check against rlimit here. If this check is done later after the test
327 * of oldbrk with newbrk then it can escape the test and let the data
328 * segment grow beyond its set limit the in case where the limit is
329 * not page aligned -Ram Gupta
331 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
332 mm->end_data, mm->start_data))
335 newbrk = PAGE_ALIGN(brk);
336 oldbrk = PAGE_ALIGN(mm->brk);
337 if (oldbrk == newbrk)
340 /* Always allow shrinking brk. */
341 if (brk <= mm->brk) {
342 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
347 /* Check against existing mmap mappings. */
348 next = find_vma(mm, oldbrk);
349 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
352 /* Ok, looks good - let it rip. */
353 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
358 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
359 up_write(&mm->mmap_sem);
361 mm_populate(oldbrk, newbrk - oldbrk);
366 up_write(&mm->mmap_sem);
370 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
372 unsigned long max, prev_end, subtree_gap;
375 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
376 * allow two stack_guard_gaps between them here, and when choosing
377 * an unmapped area; whereas when expanding we only require one.
378 * That's a little inconsistent, but keeps the code here simpler.
380 max = vm_start_gap(vma);
382 prev_end = vm_end_gap(vma->vm_prev);
388 if (vma->vm_rb.rb_left) {
389 subtree_gap = rb_entry(vma->vm_rb.rb_left,
390 struct vm_area_struct, vm_rb)->rb_subtree_gap;
391 if (subtree_gap > max)
394 if (vma->vm_rb.rb_right) {
395 subtree_gap = rb_entry(vma->vm_rb.rb_right,
396 struct vm_area_struct, vm_rb)->rb_subtree_gap;
397 if (subtree_gap > max)
403 #ifdef CONFIG_DEBUG_VM_RB
404 static int browse_rb(struct rb_root *root)
406 int i = 0, j, bug = 0;
407 struct rb_node *nd, *pn = NULL;
408 unsigned long prev = 0, pend = 0;
410 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
411 struct vm_area_struct *vma;
412 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
413 if (vma->vm_start < prev) {
414 pr_emerg("vm_start %lx < prev %lx\n",
415 vma->vm_start, prev);
418 if (vma->vm_start < pend) {
419 pr_emerg("vm_start %lx < pend %lx\n",
420 vma->vm_start, pend);
423 if (vma->vm_start > vma->vm_end) {
424 pr_emerg("vm_start %lx > vm_end %lx\n",
425 vma->vm_start, vma->vm_end);
428 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
429 pr_emerg("free gap %lx, correct %lx\n",
431 vma_compute_subtree_gap(vma));
436 prev = vma->vm_start;
440 for (nd = pn; nd; nd = rb_prev(nd))
443 pr_emerg("backwards %d, forwards %d\n", j, i);
449 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
453 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
454 struct vm_area_struct *vma;
455 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
456 VM_BUG_ON_VMA(vma != ignore &&
457 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
462 static void validate_mm(struct mm_struct *mm)
466 unsigned long highest_address = 0;
467 struct vm_area_struct *vma = mm->mmap;
470 struct anon_vma *anon_vma = vma->anon_vma;
471 struct anon_vma_chain *avc;
474 anon_vma_lock_read(anon_vma);
475 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
476 anon_vma_interval_tree_verify(avc);
477 anon_vma_unlock_read(anon_vma);
480 highest_address = vm_end_gap(vma);
484 if (i != mm->map_count) {
485 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
488 if (highest_address != mm->highest_vm_end) {
489 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
490 mm->highest_vm_end, highest_address);
493 i = browse_rb(&mm->mm_rb);
494 if (i != mm->map_count) {
496 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
499 VM_BUG_ON_MM(bug, mm);
502 #define validate_mm_rb(root, ignore) do { } while (0)
503 #define validate_mm(mm) do { } while (0)
506 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
507 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
510 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
511 * vma->vm_prev->vm_end values changed, without modifying the vma's position
514 static void vma_gap_update(struct vm_area_struct *vma)
517 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
518 * function that does exacltly what we want.
520 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
523 static inline void vma_rb_insert(struct vm_area_struct *vma,
524 struct rb_root *root)
526 /* All rb_subtree_gap values must be consistent prior to insertion */
527 validate_mm_rb(root, NULL);
529 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
532 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
535 * All rb_subtree_gap values must be consistent prior to erase,
536 * with the possible exception of the vma being erased.
538 validate_mm_rb(root, vma);
541 * Note rb_erase_augmented is a fairly large inline function,
542 * so make sure we instantiate it only once with our desired
543 * augmented rbtree callbacks.
545 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
549 * vma has some anon_vma assigned, and is already inserted on that
550 * anon_vma's interval trees.
552 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
553 * vma must be removed from the anon_vma's interval trees using
554 * anon_vma_interval_tree_pre_update_vma().
556 * After the update, the vma will be reinserted using
557 * anon_vma_interval_tree_post_update_vma().
559 * The entire update must be protected by exclusive mmap_sem and by
560 * the root anon_vma's mutex.
563 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
565 struct anon_vma_chain *avc;
567 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
568 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
572 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
574 struct anon_vma_chain *avc;
576 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
577 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
580 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
581 unsigned long end, struct vm_area_struct **pprev,
582 struct rb_node ***rb_link, struct rb_node **rb_parent)
584 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
586 __rb_link = &mm->mm_rb.rb_node;
587 rb_prev = __rb_parent = NULL;
590 struct vm_area_struct *vma_tmp;
592 __rb_parent = *__rb_link;
593 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
595 if (vma_tmp->vm_end > addr) {
596 /* Fail if an existing vma overlaps the area */
597 if (vma_tmp->vm_start < end)
599 __rb_link = &__rb_parent->rb_left;
601 rb_prev = __rb_parent;
602 __rb_link = &__rb_parent->rb_right;
608 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
609 *rb_link = __rb_link;
610 *rb_parent = __rb_parent;
614 static unsigned long count_vma_pages_range(struct mm_struct *mm,
615 unsigned long addr, unsigned long end)
617 unsigned long nr_pages = 0;
618 struct vm_area_struct *vma;
620 /* Find first overlaping mapping */
621 vma = find_vma_intersection(mm, addr, end);
625 nr_pages = (min(end, vma->vm_end) -
626 max(addr, vma->vm_start)) >> PAGE_SHIFT;
628 /* Iterate over the rest of the overlaps */
629 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
630 unsigned long overlap_len;
632 if (vma->vm_start > end)
635 overlap_len = min(end, vma->vm_end) - vma->vm_start;
636 nr_pages += overlap_len >> PAGE_SHIFT;
642 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
643 struct rb_node **rb_link, struct rb_node *rb_parent)
645 /* Update tracking information for the gap following the new vma. */
647 vma_gap_update(vma->vm_next);
649 mm->highest_vm_end = vm_end_gap(vma);
652 * vma->vm_prev wasn't known when we followed the rbtree to find the
653 * correct insertion point for that vma. As a result, we could not
654 * update the vma vm_rb parents rb_subtree_gap values on the way down.
655 * So, we first insert the vma with a zero rb_subtree_gap value
656 * (to be consistent with what we did on the way down), and then
657 * immediately update the gap to the correct value. Finally we
658 * rebalance the rbtree after all augmented values have been set.
660 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
661 vma->rb_subtree_gap = 0;
663 vma_rb_insert(vma, &mm->mm_rb);
666 static void __vma_link_file(struct vm_area_struct *vma)
672 struct address_space *mapping = file->f_mapping;
674 if (vma->vm_flags & VM_DENYWRITE)
675 atomic_dec(&file_inode(file)->i_writecount);
676 if (vma->vm_flags & VM_SHARED)
677 atomic_inc(&mapping->i_mmap_writable);
679 flush_dcache_mmap_lock(mapping);
680 vma_interval_tree_insert(vma, &mapping->i_mmap);
681 flush_dcache_mmap_unlock(mapping);
686 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
687 struct vm_area_struct *prev, struct rb_node **rb_link,
688 struct rb_node *rb_parent)
690 __vma_link_list(mm, vma, prev, rb_parent);
691 __vma_link_rb(mm, vma, rb_link, rb_parent);
694 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
695 struct vm_area_struct *prev, struct rb_node **rb_link,
696 struct rb_node *rb_parent)
698 struct address_space *mapping = NULL;
701 mapping = vma->vm_file->f_mapping;
702 i_mmap_lock_write(mapping);
705 __vma_link(mm, vma, prev, rb_link, rb_parent);
706 __vma_link_file(vma);
709 i_mmap_unlock_write(mapping);
716 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
717 * mm's list and rbtree. It has already been inserted into the interval tree.
719 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
721 struct vm_area_struct *prev;
722 struct rb_node **rb_link, *rb_parent;
724 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
725 &prev, &rb_link, &rb_parent))
727 __vma_link(mm, vma, prev, rb_link, rb_parent);
732 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
733 struct vm_area_struct *prev)
735 struct vm_area_struct *next;
737 vma_rb_erase(vma, &mm->mm_rb);
738 prev->vm_next = next = vma->vm_next;
740 next->vm_prev = prev;
743 vmacache_invalidate(mm);
747 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
748 * is already present in an i_mmap tree without adjusting the tree.
749 * The following helper function should be used when such adjustments
750 * are necessary. The "insert" vma (if any) is to be inserted
751 * before we drop the necessary locks.
753 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
754 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
756 struct mm_struct *mm = vma->vm_mm;
757 struct vm_area_struct *next = vma->vm_next;
758 struct vm_area_struct *importer = NULL;
759 struct address_space *mapping = NULL;
760 struct rb_root *root = NULL;
761 struct anon_vma *anon_vma = NULL;
762 struct file *file = vma->vm_file;
763 bool start_changed = false, end_changed = false;
764 long adjust_next = 0;
767 if (next && !insert) {
768 struct vm_area_struct *exporter = NULL;
770 if (end >= next->vm_end) {
772 * vma expands, overlapping all the next, and
773 * perhaps the one after too (mprotect case 6).
775 again: remove_next = 1 + (end > next->vm_end);
779 } else if (end > next->vm_start) {
781 * vma expands, overlapping part of the next:
782 * mprotect case 5 shifting the boundary up.
784 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
787 } else if (end < vma->vm_end) {
789 * vma shrinks, and !insert tells it's not
790 * split_vma inserting another: so it must be
791 * mprotect case 4 shifting the boundary down.
793 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
799 * Easily overlooked: when mprotect shifts the boundary,
800 * make sure the expanding vma has anon_vma set if the
801 * shrinking vma had, to cover any anon pages imported.
803 if (exporter && exporter->anon_vma && !importer->anon_vma) {
806 importer->anon_vma = exporter->anon_vma;
807 error = anon_vma_clone(importer, exporter);
814 mapping = file->f_mapping;
815 root = &mapping->i_mmap;
816 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
819 uprobe_munmap(next, next->vm_start, next->vm_end);
821 i_mmap_lock_write(mapping);
824 * Put into interval tree now, so instantiated pages
825 * are visible to arm/parisc __flush_dcache_page
826 * throughout; but we cannot insert into address
827 * space until vma start or end is updated.
829 __vma_link_file(insert);
833 vma_adjust_trans_huge(vma, start, end, adjust_next);
835 anon_vma = vma->anon_vma;
836 if (!anon_vma && adjust_next)
837 anon_vma = next->anon_vma;
839 VM_BUG_ON_VMA(adjust_next && next->anon_vma &&
840 anon_vma != next->anon_vma, next);
841 anon_vma_lock_write(anon_vma);
842 anon_vma_interval_tree_pre_update_vma(vma);
844 anon_vma_interval_tree_pre_update_vma(next);
848 flush_dcache_mmap_lock(mapping);
849 vma_interval_tree_remove(vma, root);
851 vma_interval_tree_remove(next, root);
854 if (start != vma->vm_start) {
855 vma->vm_start = start;
856 start_changed = true;
858 if (end != vma->vm_end) {
862 vma->vm_pgoff = pgoff;
864 next->vm_start += adjust_next << PAGE_SHIFT;
865 next->vm_pgoff += adjust_next;
870 vma_interval_tree_insert(next, root);
871 vma_interval_tree_insert(vma, root);
872 flush_dcache_mmap_unlock(mapping);
877 * vma_merge has merged next into vma, and needs
878 * us to remove next before dropping the locks.
880 __vma_unlink(mm, next, vma);
882 __remove_shared_vm_struct(next, file, mapping);
885 * split_vma has split insert from vma, and needs
886 * us to insert it before dropping the locks
887 * (it may either follow vma or precede it).
889 __insert_vm_struct(mm, insert);
895 mm->highest_vm_end = vm_end_gap(vma);
896 else if (!adjust_next)
897 vma_gap_update(next);
902 anon_vma_interval_tree_post_update_vma(vma);
904 anon_vma_interval_tree_post_update_vma(next);
905 anon_vma_unlock_write(anon_vma);
908 i_mmap_unlock_write(mapping);
919 uprobe_munmap(next, next->vm_start, next->vm_end);
923 anon_vma_merge(vma, next);
925 mpol_put(vma_policy(next));
926 kmem_cache_free(vm_area_cachep, next);
928 * In mprotect's case 6 (see comments on vma_merge),
929 * we must remove another next too. It would clutter
930 * up the code too much to do both in one go.
933 if (remove_next == 2)
936 vma_gap_update(next);
938 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
949 * If the vma has a ->close operation then the driver probably needs to release
950 * per-vma resources, so we don't attempt to merge those.
952 static inline int is_mergeable_vma(struct vm_area_struct *vma,
953 struct file *file, unsigned long vm_flags,
954 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
955 const char __user *anon_name)
958 * VM_SOFTDIRTY should not prevent from VMA merging, if we
959 * match the flags but dirty bit -- the caller should mark
960 * merged VMA as dirty. If dirty bit won't be excluded from
961 * comparison, we increase pressue on the memory system forcing
962 * the kernel to generate new VMAs when old one could be
965 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
967 if (vma->vm_file != file)
969 if (vma->vm_ops && vma->vm_ops->close)
971 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
973 if (vma_get_anon_name(vma) != anon_name)
978 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
979 struct anon_vma *anon_vma2,
980 struct vm_area_struct *vma)
983 * The list_is_singular() test is to avoid merging VMA cloned from
984 * parents. This can improve scalability caused by anon_vma lock.
986 if ((!anon_vma1 || !anon_vma2) && (!vma ||
987 list_is_singular(&vma->anon_vma_chain)))
989 return anon_vma1 == anon_vma2;
993 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
994 * in front of (at a lower virtual address and file offset than) the vma.
996 * We cannot merge two vmas if they have differently assigned (non-NULL)
997 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
999 * We don't check here for the merged mmap wrapping around the end of pagecache
1000 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1001 * wrap, nor mmaps which cover the final page at index -1UL.
1004 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1005 struct anon_vma *anon_vma, struct file *file,
1007 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1008 const char __user *anon_name)
1010 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
1011 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1012 if (vma->vm_pgoff == vm_pgoff)
1019 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1020 * beyond (at a higher virtual address and file offset than) the vma.
1022 * We cannot merge two vmas if they have differently assigned (non-NULL)
1023 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1026 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1027 struct anon_vma *anon_vma, struct file *file,
1029 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1030 const char __user *anon_name)
1032 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
1033 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1035 vm_pglen = vma_pages(vma);
1036 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1043 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
1044 * figure out whether that can be merged with its predecessor or its
1045 * successor. Or both (it neatly fills a hole).
1047 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1048 * certain not to be mapped by the time vma_merge is called; but when
1049 * called for mprotect, it is certain to be already mapped (either at
1050 * an offset within prev, or at the start of next), and the flags of
1051 * this area are about to be changed to vm_flags - and the no-change
1052 * case has already been eliminated.
1054 * The following mprotect cases have to be considered, where AAAA is
1055 * the area passed down from mprotect_fixup, never extending beyond one
1056 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1058 * AAAA AAAA AAAA AAAA
1059 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1060 * cannot merge might become might become might become
1061 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1062 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1063 * mremap move: PPPPNNNNNNNN 8
1065 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1066 * might become case 1 below case 2 below case 3 below
1068 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1069 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1071 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1072 struct vm_area_struct *prev, unsigned long addr,
1073 unsigned long end, unsigned long vm_flags,
1074 struct anon_vma *anon_vma, struct file *file,
1075 pgoff_t pgoff, struct mempolicy *policy,
1076 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1077 const char __user *anon_name)
1079 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1080 struct vm_area_struct *area, *next;
1084 * We later require that vma->vm_flags == vm_flags,
1085 * so this tests vma->vm_flags & VM_SPECIAL, too.
1087 if (vm_flags & VM_SPECIAL)
1091 next = prev->vm_next;
1095 if (next && next->vm_end == end) /* cases 6, 7, 8 */
1096 next = next->vm_next;
1099 * Can it merge with the predecessor?
1101 if (prev && prev->vm_end == addr &&
1102 mpol_equal(vma_policy(prev), policy) &&
1103 can_vma_merge_after(prev, vm_flags,
1104 anon_vma, file, pgoff,
1108 * OK, it can. Can we now merge in the successor as well?
1110 if (next && end == next->vm_start &&
1111 mpol_equal(policy, vma_policy(next)) &&
1112 can_vma_merge_before(next, vm_flags,
1117 is_mergeable_anon_vma(prev->anon_vma,
1118 next->anon_vma, NULL)) {
1120 err = vma_adjust(prev, prev->vm_start,
1121 next->vm_end, prev->vm_pgoff, NULL);
1122 } else /* cases 2, 5, 7 */
1123 err = vma_adjust(prev, prev->vm_start,
1124 end, prev->vm_pgoff, NULL);
1127 khugepaged_enter_vma_merge(prev, vm_flags);
1132 * Can this new request be merged in front of next?
1134 if (next && end == next->vm_start &&
1135 mpol_equal(policy, vma_policy(next)) &&
1136 can_vma_merge_before(next, vm_flags,
1137 anon_vma, file, pgoff+pglen,
1140 if (prev && addr < prev->vm_end) /* case 4 */
1141 err = vma_adjust(prev, prev->vm_start,
1142 addr, prev->vm_pgoff, NULL);
1143 else /* cases 3, 8 */
1144 err = vma_adjust(area, addr, next->vm_end,
1145 next->vm_pgoff - pglen, NULL);
1148 khugepaged_enter_vma_merge(area, vm_flags);
1156 * Rough compatbility check to quickly see if it's even worth looking
1157 * at sharing an anon_vma.
1159 * They need to have the same vm_file, and the flags can only differ
1160 * in things that mprotect may change.
1162 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1163 * we can merge the two vma's. For example, we refuse to merge a vma if
1164 * there is a vm_ops->close() function, because that indicates that the
1165 * driver is doing some kind of reference counting. But that doesn't
1166 * really matter for the anon_vma sharing case.
1168 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1170 return a->vm_end == b->vm_start &&
1171 mpol_equal(vma_policy(a), vma_policy(b)) &&
1172 a->vm_file == b->vm_file &&
1173 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1174 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1178 * Do some basic sanity checking to see if we can re-use the anon_vma
1179 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1180 * the same as 'old', the other will be the new one that is trying
1181 * to share the anon_vma.
1183 * NOTE! This runs with mm_sem held for reading, so it is possible that
1184 * the anon_vma of 'old' is concurrently in the process of being set up
1185 * by another page fault trying to merge _that_. But that's ok: if it
1186 * is being set up, that automatically means that it will be a singleton
1187 * acceptable for merging, so we can do all of this optimistically. But
1188 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1190 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1191 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1192 * is to return an anon_vma that is "complex" due to having gone through
1195 * We also make sure that the two vma's are compatible (adjacent,
1196 * and with the same memory policies). That's all stable, even with just
1197 * a read lock on the mm_sem.
1199 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1201 if (anon_vma_compatible(a, b)) {
1202 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1204 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1211 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1212 * neighbouring vmas for a suitable anon_vma, before it goes off
1213 * to allocate a new anon_vma. It checks because a repetitive
1214 * sequence of mprotects and faults may otherwise lead to distinct
1215 * anon_vmas being allocated, preventing vma merge in subsequent
1218 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1220 struct anon_vma *anon_vma;
1221 struct vm_area_struct *near;
1223 near = vma->vm_next;
1227 anon_vma = reusable_anon_vma(near, vma, near);
1231 near = vma->vm_prev;
1235 anon_vma = reusable_anon_vma(near, near, vma);
1240 * There's no absolute need to look only at touching neighbours:
1241 * we could search further afield for "compatible" anon_vmas.
1242 * But it would probably just be a waste of time searching,
1243 * or lead to too many vmas hanging off the same anon_vma.
1244 * We're trying to allow mprotect remerging later on,
1245 * not trying to minimize memory used for anon_vmas.
1250 #ifdef CONFIG_PROC_FS
1251 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1252 struct file *file, long pages)
1254 const unsigned long stack_flags
1255 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1257 mm->total_vm += pages;
1260 mm->shared_vm += pages;
1261 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1262 mm->exec_vm += pages;
1263 } else if (flags & stack_flags)
1264 mm->stack_vm += pages;
1266 #endif /* CONFIG_PROC_FS */
1269 * If a hint addr is less than mmap_min_addr change hint to be as
1270 * low as possible but still greater than mmap_min_addr
1272 static inline unsigned long round_hint_to_min(unsigned long hint)
1275 if (((void *)hint != NULL) &&
1276 (hint < mmap_min_addr))
1277 return PAGE_ALIGN(mmap_min_addr);
1281 static inline int mlock_future_check(struct mm_struct *mm,
1282 unsigned long flags,
1285 unsigned long locked, lock_limit;
1287 /* mlock MCL_FUTURE? */
1288 if (flags & VM_LOCKED) {
1289 locked = len >> PAGE_SHIFT;
1290 locked += mm->locked_vm;
1291 lock_limit = rlimit(RLIMIT_MEMLOCK);
1292 lock_limit >>= PAGE_SHIFT;
1293 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1300 * The caller must hold down_write(¤t->mm->mmap_sem).
1302 unsigned long do_mmap(struct file *file, unsigned long addr,
1303 unsigned long len, unsigned long prot,
1304 unsigned long flags, vm_flags_t vm_flags,
1305 unsigned long pgoff, unsigned long *populate)
1307 struct mm_struct *mm = current->mm;
1315 * Does the application expect PROT_READ to imply PROT_EXEC?
1317 * (the exception is when the underlying filesystem is noexec
1318 * mounted, in which case we dont add PROT_EXEC.)
1320 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1321 if (!(file && path_noexec(&file->f_path)))
1324 if (!(flags & MAP_FIXED))
1325 addr = round_hint_to_min(addr);
1327 /* Careful about overflows.. */
1328 len = PAGE_ALIGN(len);
1332 /* offset overflow? */
1333 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1336 /* Too many mappings? */
1337 if (mm->map_count > sysctl_max_map_count)
1340 /* Obtain the address to map to. we verify (or select) it and ensure
1341 * that it represents a valid section of the address space.
1343 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1344 if (offset_in_page(addr))
1347 /* Do simple checking here so the lower-level routines won't have
1348 * to. we assume access permissions have been handled by the open
1349 * of the memory object, so we don't do any here.
1351 vm_flags |= calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1352 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1354 if (flags & MAP_LOCKED)
1355 if (!can_do_mlock())
1358 if (mlock_future_check(mm, vm_flags, len))
1362 struct inode *inode = file_inode(file);
1364 switch (flags & MAP_TYPE) {
1366 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1370 * Make sure we don't allow writing to an append-only
1373 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1377 * Make sure there are no mandatory locks on the file.
1379 if (locks_verify_locked(file))
1382 vm_flags |= VM_SHARED | VM_MAYSHARE;
1383 if (!(file->f_mode & FMODE_WRITE))
1384 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1388 if (!(file->f_mode & FMODE_READ))
1390 if (path_noexec(&file->f_path)) {
1391 if (vm_flags & VM_EXEC)
1393 vm_flags &= ~VM_MAYEXEC;
1396 if (!file->f_op->mmap)
1398 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1406 switch (flags & MAP_TYPE) {
1408 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1414 vm_flags |= VM_SHARED | VM_MAYSHARE;
1418 * Set pgoff according to addr for anon_vma.
1420 pgoff = addr >> PAGE_SHIFT;
1428 * Set 'VM_NORESERVE' if we should not account for the
1429 * memory use of this mapping.
1431 if (flags & MAP_NORESERVE) {
1432 /* We honor MAP_NORESERVE if allowed to overcommit */
1433 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1434 vm_flags |= VM_NORESERVE;
1436 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1437 if (file && is_file_hugepages(file))
1438 vm_flags |= VM_NORESERVE;
1441 addr = mmap_region(file, addr, len, vm_flags, pgoff);
1442 if (!IS_ERR_VALUE(addr) &&
1443 ((vm_flags & VM_LOCKED) ||
1444 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1449 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1450 unsigned long, prot, unsigned long, flags,
1451 unsigned long, fd, unsigned long, pgoff)
1453 struct file *file = NULL;
1454 unsigned long retval;
1456 if (!(flags & MAP_ANONYMOUS)) {
1457 audit_mmap_fd(fd, flags);
1461 if (is_file_hugepages(file))
1462 len = ALIGN(len, huge_page_size(hstate_file(file)));
1464 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1466 } else if (flags & MAP_HUGETLB) {
1467 struct user_struct *user = NULL;
1470 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1474 len = ALIGN(len, huge_page_size(hs));
1476 * VM_NORESERVE is used because the reservations will be
1477 * taken when vm_ops->mmap() is called
1478 * A dummy user value is used because we are not locking
1479 * memory so no accounting is necessary
1481 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1483 &user, HUGETLB_ANONHUGE_INODE,
1484 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1486 return PTR_ERR(file);
1489 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1491 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1498 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1499 struct mmap_arg_struct {
1503 unsigned long flags;
1505 unsigned long offset;
1508 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1510 struct mmap_arg_struct a;
1512 if (copy_from_user(&a, arg, sizeof(a)))
1514 if (offset_in_page(a.offset))
1517 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1518 a.offset >> PAGE_SHIFT);
1520 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1523 * Some shared mappigns will want the pages marked read-only
1524 * to track write events. If so, we'll downgrade vm_page_prot
1525 * to the private version (using protection_map[] without the
1528 int vma_wants_writenotify(struct vm_area_struct *vma)
1530 vm_flags_t vm_flags = vma->vm_flags;
1531 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1533 /* If it was private or non-writable, the write bit is already clear */
1534 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1537 /* The backer wishes to know when pages are first written to? */
1538 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1541 /* The open routine did something to the protections that pgprot_modify
1542 * won't preserve? */
1543 if (pgprot_val(vma->vm_page_prot) !=
1544 pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags)))
1547 /* Do we need to track softdirty? */
1548 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1551 /* Specialty mapping? */
1552 if (vm_flags & VM_PFNMAP)
1555 /* Can the mapping track the dirty pages? */
1556 return vma->vm_file && vma->vm_file->f_mapping &&
1557 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1561 * We account for memory if it's a private writeable mapping,
1562 * not hugepages and VM_NORESERVE wasn't set.
1564 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1567 * hugetlb has its own accounting separate from the core VM
1568 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1570 if (file && is_file_hugepages(file))
1573 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1576 unsigned long mmap_region(struct file *file, unsigned long addr,
1577 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1579 struct mm_struct *mm = current->mm;
1580 struct vm_area_struct *vma, *prev;
1582 struct rb_node **rb_link, *rb_parent;
1583 unsigned long charged = 0;
1585 /* Check against address space limit. */
1586 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) {
1587 unsigned long nr_pages;
1590 * MAP_FIXED may remove pages of mappings that intersects with
1591 * requested mapping. Account for the pages it would unmap.
1593 if (!(vm_flags & MAP_FIXED))
1596 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1598 if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages))
1602 /* Clear old maps */
1603 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1605 if (do_munmap(mm, addr, len))
1610 * Private writable mapping: check memory availability
1612 if (accountable_mapping(file, vm_flags)) {
1613 charged = len >> PAGE_SHIFT;
1614 if (security_vm_enough_memory_mm(mm, charged))
1616 vm_flags |= VM_ACCOUNT;
1620 * Can we just expand an old mapping?
1622 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1623 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
1628 * Determine the object being mapped and call the appropriate
1629 * specific mapper. the address has already been validated, but
1630 * not unmapped, but the maps are removed from the list.
1632 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1639 vma->vm_start = addr;
1640 vma->vm_end = addr + len;
1641 vma->vm_flags = vm_flags;
1642 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1643 vma->vm_pgoff = pgoff;
1644 INIT_LIST_HEAD(&vma->anon_vma_chain);
1647 if (vm_flags & VM_DENYWRITE) {
1648 error = deny_write_access(file);
1652 if (vm_flags & VM_SHARED) {
1653 error = mapping_map_writable(file->f_mapping);
1655 goto allow_write_and_free_vma;
1658 /* ->mmap() can change vma->vm_file, but must guarantee that
1659 * vma_link() below can deny write-access if VM_DENYWRITE is set
1660 * and map writably if VM_SHARED is set. This usually means the
1661 * new file must not have been exposed to user-space, yet.
1663 vma->vm_file = get_file(file);
1664 error = file->f_op->mmap(file, vma);
1666 goto unmap_and_free_vma;
1668 /* Can addr have changed??
1670 * Answer: Yes, several device drivers can do it in their
1671 * f_op->mmap method. -DaveM
1672 * Bug: If addr is changed, prev, rb_link, rb_parent should
1673 * be updated for vma_link()
1675 WARN_ON_ONCE(addr != vma->vm_start);
1677 addr = vma->vm_start;
1678 vm_flags = vma->vm_flags;
1679 } else if (vm_flags & VM_SHARED) {
1680 error = shmem_zero_setup(vma);
1685 vma_link(mm, vma, prev, rb_link, rb_parent);
1686 /* Once vma denies write, undo our temporary denial count */
1688 if (vm_flags & VM_SHARED)
1689 mapping_unmap_writable(file->f_mapping);
1690 if (vm_flags & VM_DENYWRITE)
1691 allow_write_access(file);
1693 file = vma->vm_file;
1695 perf_event_mmap(vma);
1697 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1698 if (vm_flags & VM_LOCKED) {
1699 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1700 vma == get_gate_vma(current->mm)))
1701 mm->locked_vm += (len >> PAGE_SHIFT);
1703 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1710 * New (or expanded) vma always get soft dirty status.
1711 * Otherwise user-space soft-dirty page tracker won't
1712 * be able to distinguish situation when vma area unmapped,
1713 * then new mapped in-place (which must be aimed as
1714 * a completely new data area).
1716 vma->vm_flags |= VM_SOFTDIRTY;
1718 vma_set_page_prot(vma);
1723 vma->vm_file = NULL;
1726 /* Undo any partial mapping done by a device driver. */
1727 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1729 if (vm_flags & VM_SHARED)
1730 mapping_unmap_writable(file->f_mapping);
1731 allow_write_and_free_vma:
1732 if (vm_flags & VM_DENYWRITE)
1733 allow_write_access(file);
1735 kmem_cache_free(vm_area_cachep, vma);
1738 vm_unacct_memory(charged);
1742 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1745 * We implement the search by looking for an rbtree node that
1746 * immediately follows a suitable gap. That is,
1747 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1748 * - gap_end = vma->vm_start >= info->low_limit + length;
1749 * - gap_end - gap_start >= length
1752 struct mm_struct *mm = current->mm;
1753 struct vm_area_struct *vma;
1754 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1756 /* Adjust search length to account for worst case alignment overhead */
1757 length = info->length + info->align_mask;
1758 if (length < info->length)
1761 /* Adjust search limits by the desired length */
1762 if (info->high_limit < length)
1764 high_limit = info->high_limit - length;
1766 if (info->low_limit > high_limit)
1768 low_limit = info->low_limit + length;
1770 /* Check if rbtree root looks promising */
1771 if (RB_EMPTY_ROOT(&mm->mm_rb))
1773 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1774 if (vma->rb_subtree_gap < length)
1778 /* Visit left subtree if it looks promising */
1779 gap_end = vm_start_gap(vma);
1780 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1781 struct vm_area_struct *left =
1782 rb_entry(vma->vm_rb.rb_left,
1783 struct vm_area_struct, vm_rb);
1784 if (left->rb_subtree_gap >= length) {
1790 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1792 /* Check if current node has a suitable gap */
1793 if (gap_start > high_limit)
1795 if (gap_end >= low_limit &&
1796 gap_end > gap_start && gap_end - gap_start >= length)
1799 /* Visit right subtree if it looks promising */
1800 if (vma->vm_rb.rb_right) {
1801 struct vm_area_struct *right =
1802 rb_entry(vma->vm_rb.rb_right,
1803 struct vm_area_struct, vm_rb);
1804 if (right->rb_subtree_gap >= length) {
1810 /* Go back up the rbtree to find next candidate node */
1812 struct rb_node *prev = &vma->vm_rb;
1813 if (!rb_parent(prev))
1815 vma = rb_entry(rb_parent(prev),
1816 struct vm_area_struct, vm_rb);
1817 if (prev == vma->vm_rb.rb_left) {
1818 gap_start = vm_end_gap(vma->vm_prev);
1819 gap_end = vm_start_gap(vma);
1826 /* Check highest gap, which does not precede any rbtree node */
1827 gap_start = mm->highest_vm_end;
1828 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1829 if (gap_start > high_limit)
1833 /* We found a suitable gap. Clip it with the original low_limit. */
1834 if (gap_start < info->low_limit)
1835 gap_start = info->low_limit;
1837 /* Adjust gap address to the desired alignment */
1838 gap_start += (info->align_offset - gap_start) & info->align_mask;
1840 VM_BUG_ON(gap_start + info->length > info->high_limit);
1841 VM_BUG_ON(gap_start + info->length > gap_end);
1845 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1847 struct mm_struct *mm = current->mm;
1848 struct vm_area_struct *vma;
1849 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1851 /* Adjust search length to account for worst case alignment overhead */
1852 length = info->length + info->align_mask;
1853 if (length < info->length)
1857 * Adjust search limits by the desired length.
1858 * See implementation comment at top of unmapped_area().
1860 gap_end = info->high_limit;
1861 if (gap_end < length)
1863 high_limit = gap_end - length;
1865 if (info->low_limit > high_limit)
1867 low_limit = info->low_limit + length;
1869 /* Check highest gap, which does not precede any rbtree node */
1870 gap_start = mm->highest_vm_end;
1871 if (gap_start <= high_limit)
1874 /* Check if rbtree root looks promising */
1875 if (RB_EMPTY_ROOT(&mm->mm_rb))
1877 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1878 if (vma->rb_subtree_gap < length)
1882 /* Visit right subtree if it looks promising */
1883 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1884 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1885 struct vm_area_struct *right =
1886 rb_entry(vma->vm_rb.rb_right,
1887 struct vm_area_struct, vm_rb);
1888 if (right->rb_subtree_gap >= length) {
1895 /* Check if current node has a suitable gap */
1896 gap_end = vm_start_gap(vma);
1897 if (gap_end < low_limit)
1899 if (gap_start <= high_limit &&
1900 gap_end > gap_start && gap_end - gap_start >= length)
1903 /* Visit left subtree if it looks promising */
1904 if (vma->vm_rb.rb_left) {
1905 struct vm_area_struct *left =
1906 rb_entry(vma->vm_rb.rb_left,
1907 struct vm_area_struct, vm_rb);
1908 if (left->rb_subtree_gap >= length) {
1914 /* Go back up the rbtree to find next candidate node */
1916 struct rb_node *prev = &vma->vm_rb;
1917 if (!rb_parent(prev))
1919 vma = rb_entry(rb_parent(prev),
1920 struct vm_area_struct, vm_rb);
1921 if (prev == vma->vm_rb.rb_right) {
1922 gap_start = vma->vm_prev ?
1923 vm_end_gap(vma->vm_prev) : 0;
1930 /* We found a suitable gap. Clip it with the original high_limit. */
1931 if (gap_end > info->high_limit)
1932 gap_end = info->high_limit;
1935 /* Compute highest gap address at the desired alignment */
1936 gap_end -= info->length;
1937 gap_end -= (gap_end - info->align_offset) & info->align_mask;
1939 VM_BUG_ON(gap_end < info->low_limit);
1940 VM_BUG_ON(gap_end < gap_start);
1944 /* Get an address range which is currently unmapped.
1945 * For shmat() with addr=0.
1947 * Ugly calling convention alert:
1948 * Return value with the low bits set means error value,
1950 * if (ret & ~PAGE_MASK)
1953 * This function "knows" that -ENOMEM has the bits set.
1955 #ifndef HAVE_ARCH_UNMAPPED_AREA
1957 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1958 unsigned long len, unsigned long pgoff, unsigned long flags)
1960 struct mm_struct *mm = current->mm;
1961 struct vm_area_struct *vma, *prev;
1962 struct vm_unmapped_area_info info;
1964 if (len > TASK_SIZE - mmap_min_addr)
1967 if (flags & MAP_FIXED)
1971 addr = PAGE_ALIGN(addr);
1972 vma = find_vma_prev(mm, addr, &prev);
1973 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1974 (!vma || addr + len <= vm_start_gap(vma)) &&
1975 (!prev || addr >= vm_end_gap(prev)))
1981 info.low_limit = mm->mmap_base;
1982 info.high_limit = TASK_SIZE;
1983 info.align_mask = 0;
1984 return vm_unmapped_area(&info);
1989 * This mmap-allocator allocates new areas top-down from below the
1990 * stack's low limit (the base):
1992 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1994 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1995 const unsigned long len, const unsigned long pgoff,
1996 const unsigned long flags)
1998 struct vm_area_struct *vma, *prev;
1999 struct mm_struct *mm = current->mm;
2000 unsigned long addr = addr0;
2001 struct vm_unmapped_area_info info;
2003 /* requested length too big for entire address space */
2004 if (len > TASK_SIZE - mmap_min_addr)
2007 if (flags & MAP_FIXED)
2010 /* requesting a specific address */
2012 addr = PAGE_ALIGN(addr);
2013 vma = find_vma_prev(mm, addr, &prev);
2014 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2015 (!vma || addr + len <= vm_start_gap(vma)) &&
2016 (!prev || addr >= vm_end_gap(prev)))
2020 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2022 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2023 info.high_limit = mm->mmap_base;
2024 info.align_mask = 0;
2025 addr = vm_unmapped_area(&info);
2028 * A failed mmap() very likely causes application failure,
2029 * so fall back to the bottom-up function here. This scenario
2030 * can happen with large stack limits and large mmap()
2033 if (offset_in_page(addr)) {
2034 VM_BUG_ON(addr != -ENOMEM);
2036 info.low_limit = TASK_UNMAPPED_BASE;
2037 info.high_limit = TASK_SIZE;
2038 addr = vm_unmapped_area(&info);
2046 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2047 unsigned long pgoff, unsigned long flags)
2049 unsigned long (*get_area)(struct file *, unsigned long,
2050 unsigned long, unsigned long, unsigned long);
2052 unsigned long error = arch_mmap_check(addr, len, flags);
2056 /* Careful about overflows.. */
2057 if (len > TASK_SIZE)
2060 get_area = current->mm->get_unmapped_area;
2061 if (file && file->f_op->get_unmapped_area)
2062 get_area = file->f_op->get_unmapped_area;
2063 addr = get_area(file, addr, len, pgoff, flags);
2064 if (IS_ERR_VALUE(addr))
2067 if (addr > TASK_SIZE - len)
2069 if (offset_in_page(addr))
2072 addr = arch_rebalance_pgtables(addr, len);
2073 error = security_mmap_addr(addr);
2074 return error ? error : addr;
2077 EXPORT_SYMBOL(get_unmapped_area);
2079 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2080 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2082 struct rb_node *rb_node;
2083 struct vm_area_struct *vma;
2085 /* Check the cache first. */
2086 vma = vmacache_find(mm, addr);
2090 rb_node = mm->mm_rb.rb_node;
2093 struct vm_area_struct *tmp;
2095 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2097 if (tmp->vm_end > addr) {
2099 if (tmp->vm_start <= addr)
2101 rb_node = rb_node->rb_left;
2103 rb_node = rb_node->rb_right;
2107 vmacache_update(addr, vma);
2111 EXPORT_SYMBOL(find_vma);
2114 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2116 struct vm_area_struct *
2117 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2118 struct vm_area_struct **pprev)
2120 struct vm_area_struct *vma;
2122 vma = find_vma(mm, addr);
2124 *pprev = vma->vm_prev;
2126 struct rb_node *rb_node = mm->mm_rb.rb_node;
2129 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2130 rb_node = rb_node->rb_right;
2137 * Verify that the stack growth is acceptable and
2138 * update accounting. This is shared with both the
2139 * grow-up and grow-down cases.
2141 static int acct_stack_growth(struct vm_area_struct *vma,
2142 unsigned long size, unsigned long grow)
2144 struct mm_struct *mm = vma->vm_mm;
2145 struct rlimit *rlim = current->signal->rlim;
2146 unsigned long new_start;
2148 /* address space limit tests */
2149 if (!may_expand_vm(mm, grow))
2152 /* Stack limit test */
2153 if (size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2156 /* mlock limit tests */
2157 if (vma->vm_flags & VM_LOCKED) {
2158 unsigned long locked;
2159 unsigned long limit;
2160 locked = mm->locked_vm + grow;
2161 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2162 limit >>= PAGE_SHIFT;
2163 if (locked > limit && !capable(CAP_IPC_LOCK))
2167 /* Check to ensure the stack will not grow into a hugetlb-only region */
2168 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2170 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2174 * Overcommit.. This must be the final test, as it will
2175 * update security statistics.
2177 if (security_vm_enough_memory_mm(mm, grow))
2183 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2185 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2186 * vma is the last one with address > vma->vm_end. Have to extend vma.
2188 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2190 struct mm_struct *mm = vma->vm_mm;
2191 struct vm_area_struct *next;
2192 unsigned long gap_addr;
2195 if (!(vma->vm_flags & VM_GROWSUP))
2198 /* Guard against exceeding limits of the address space. */
2199 address &= PAGE_MASK;
2200 if (address >= TASK_SIZE)
2202 address += PAGE_SIZE;
2204 /* Enforce stack_guard_gap */
2205 gap_addr = address + stack_guard_gap;
2207 /* Guard against overflow */
2208 if (gap_addr < address || gap_addr > TASK_SIZE)
2209 gap_addr = TASK_SIZE;
2211 next = vma->vm_next;
2212 if (next && next->vm_start < gap_addr) {
2213 if (!(next->vm_flags & VM_GROWSUP))
2215 /* Check that both stack segments have the same anon_vma? */
2218 /* We must make sure the anon_vma is allocated. */
2219 if (unlikely(anon_vma_prepare(vma)))
2223 * vma->vm_start/vm_end cannot change under us because the caller
2224 * is required to hold the mmap_sem in read mode. We need the
2225 * anon_vma lock to serialize against concurrent expand_stacks.
2227 anon_vma_lock_write(vma->anon_vma);
2229 /* Somebody else might have raced and expanded it already */
2230 if (address > vma->vm_end) {
2231 unsigned long size, grow;
2233 size = address - vma->vm_start;
2234 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2237 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2238 error = acct_stack_growth(vma, size, grow);
2241 * vma_gap_update() doesn't support concurrent
2242 * updates, but we only hold a shared mmap_sem
2243 * lock here, so we need to protect against
2244 * concurrent vma expansions.
2245 * anon_vma_lock_write() doesn't help here, as
2246 * we don't guarantee that all growable vmas
2247 * in a mm share the same root anon vma.
2248 * So, we reuse mm->page_table_lock to guard
2249 * against concurrent vma expansions.
2251 spin_lock(&mm->page_table_lock);
2252 if (vma->vm_flags & VM_LOCKED)
2253 mm->locked_vm += grow;
2254 vm_stat_account(mm, vma->vm_flags,
2255 vma->vm_file, grow);
2256 anon_vma_interval_tree_pre_update_vma(vma);
2257 vma->vm_end = address;
2258 anon_vma_interval_tree_post_update_vma(vma);
2260 vma_gap_update(vma->vm_next);
2262 mm->highest_vm_end = vm_end_gap(vma);
2263 spin_unlock(&mm->page_table_lock);
2265 perf_event_mmap(vma);
2269 anon_vma_unlock_write(vma->anon_vma);
2270 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2274 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2277 * vma is the first one with address < vma->vm_start. Have to extend vma.
2279 int expand_downwards(struct vm_area_struct *vma,
2280 unsigned long address)
2282 struct mm_struct *mm = vma->vm_mm;
2283 struct vm_area_struct *prev;
2284 unsigned long gap_addr;
2287 address &= PAGE_MASK;
2288 error = security_mmap_addr(address);
2292 /* Enforce stack_guard_gap */
2293 gap_addr = address - stack_guard_gap;
2294 if (gap_addr > address)
2296 prev = vma->vm_prev;
2297 if (prev && prev->vm_end > gap_addr) {
2298 if (!(prev->vm_flags & VM_GROWSDOWN))
2300 /* Check that both stack segments have the same anon_vma? */
2303 /* We must make sure the anon_vma is allocated. */
2304 if (unlikely(anon_vma_prepare(vma)))
2308 * vma->vm_start/vm_end cannot change under us because the caller
2309 * is required to hold the mmap_sem in read mode. We need the
2310 * anon_vma lock to serialize against concurrent expand_stacks.
2312 anon_vma_lock_write(vma->anon_vma);
2314 /* Somebody else might have raced and expanded it already */
2315 if (address < vma->vm_start) {
2316 unsigned long size, grow;
2318 size = vma->vm_end - address;
2319 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2322 if (grow <= vma->vm_pgoff) {
2323 error = acct_stack_growth(vma, size, grow);
2326 * vma_gap_update() doesn't support concurrent
2327 * updates, but we only hold a shared mmap_sem
2328 * lock here, so we need to protect against
2329 * concurrent vma expansions.
2330 * anon_vma_lock_write() doesn't help here, as
2331 * we don't guarantee that all growable vmas
2332 * in a mm share the same root anon vma.
2333 * So, we reuse mm->page_table_lock to guard
2334 * against concurrent vma expansions.
2336 spin_lock(&mm->page_table_lock);
2337 if (vma->vm_flags & VM_LOCKED)
2338 mm->locked_vm += grow;
2339 vm_stat_account(mm, vma->vm_flags,
2340 vma->vm_file, grow);
2341 anon_vma_interval_tree_pre_update_vma(vma);
2342 vma->vm_start = address;
2343 vma->vm_pgoff -= grow;
2344 anon_vma_interval_tree_post_update_vma(vma);
2345 vma_gap_update(vma);
2346 spin_unlock(&mm->page_table_lock);
2348 perf_event_mmap(vma);
2352 anon_vma_unlock_write(vma->anon_vma);
2353 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2358 /* enforced gap between the expanding stack and other mappings. */
2359 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2361 static int __init cmdline_parse_stack_guard_gap(char *p)
2366 val = simple_strtoul(p, &endptr, 10);
2368 stack_guard_gap = val << PAGE_SHIFT;
2372 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2374 #ifdef CONFIG_STACK_GROWSUP
2375 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2377 return expand_upwards(vma, address);
2380 struct vm_area_struct *
2381 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2383 struct vm_area_struct *vma, *prev;
2386 vma = find_vma_prev(mm, addr, &prev);
2387 if (vma && (vma->vm_start <= addr))
2389 if (!prev || expand_stack(prev, addr))
2391 if (prev->vm_flags & VM_LOCKED)
2392 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2396 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2398 return expand_downwards(vma, address);
2401 struct vm_area_struct *
2402 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2404 struct vm_area_struct *vma;
2405 unsigned long start;
2408 vma = find_vma(mm, addr);
2411 if (vma->vm_start <= addr)
2413 if (!(vma->vm_flags & VM_GROWSDOWN))
2415 start = vma->vm_start;
2416 if (expand_stack(vma, addr))
2418 if (vma->vm_flags & VM_LOCKED)
2419 populate_vma_page_range(vma, addr, start, NULL);
2424 EXPORT_SYMBOL_GPL(find_extend_vma);
2427 * Ok - we have the memory areas we should free on the vma list,
2428 * so release them, and do the vma updates.
2430 * Called with the mm semaphore held.
2432 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2434 unsigned long nr_accounted = 0;
2436 /* Update high watermark before we lower total_vm */
2437 update_hiwater_vm(mm);
2439 long nrpages = vma_pages(vma);
2441 if (vma->vm_flags & VM_ACCOUNT)
2442 nr_accounted += nrpages;
2443 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2444 vma = remove_vma(vma);
2446 vm_unacct_memory(nr_accounted);
2451 * Get rid of page table information in the indicated region.
2453 * Called with the mm semaphore held.
2455 static void unmap_region(struct mm_struct *mm,
2456 struct vm_area_struct *vma, struct vm_area_struct *prev,
2457 unsigned long start, unsigned long end)
2459 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2460 struct mmu_gather tlb;
2463 tlb_gather_mmu(&tlb, mm, start, end);
2464 update_hiwater_rss(mm);
2465 unmap_vmas(&tlb, vma, start, end);
2466 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2467 next ? next->vm_start : USER_PGTABLES_CEILING);
2468 tlb_finish_mmu(&tlb, start, end);
2472 * Create a list of vma's touched by the unmap, removing them from the mm's
2473 * vma list as we go..
2476 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2477 struct vm_area_struct *prev, unsigned long end)
2479 struct vm_area_struct **insertion_point;
2480 struct vm_area_struct *tail_vma = NULL;
2482 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2483 vma->vm_prev = NULL;
2485 vma_rb_erase(vma, &mm->mm_rb);
2489 } while (vma && vma->vm_start < end);
2490 *insertion_point = vma;
2492 vma->vm_prev = prev;
2493 vma_gap_update(vma);
2495 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2496 tail_vma->vm_next = NULL;
2498 /* Kill the cache */
2499 vmacache_invalidate(mm);
2503 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2504 * munmap path where it doesn't make sense to fail.
2506 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2507 unsigned long addr, int new_below)
2509 struct vm_area_struct *new;
2512 if (is_vm_hugetlb_page(vma) && (addr &
2513 ~(huge_page_mask(hstate_vma(vma)))))
2516 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2520 /* most fields are the same, copy all, and then fixup */
2523 INIT_LIST_HEAD(&new->anon_vma_chain);
2528 new->vm_start = addr;
2529 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2532 err = vma_dup_policy(vma, new);
2536 err = anon_vma_clone(new, vma);
2541 get_file(new->vm_file);
2543 if (new->vm_ops && new->vm_ops->open)
2544 new->vm_ops->open(new);
2547 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2548 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2550 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2556 /* Clean everything up if vma_adjust failed. */
2557 if (new->vm_ops && new->vm_ops->close)
2558 new->vm_ops->close(new);
2561 unlink_anon_vmas(new);
2563 mpol_put(vma_policy(new));
2565 kmem_cache_free(vm_area_cachep, new);
2570 * Split a vma into two pieces at address 'addr', a new vma is allocated
2571 * either for the first part or the tail.
2573 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2574 unsigned long addr, int new_below)
2576 if (mm->map_count >= sysctl_max_map_count)
2579 return __split_vma(mm, vma, addr, new_below);
2582 /* Munmap is split into 2 main parts -- this part which finds
2583 * what needs doing, and the areas themselves, which do the
2584 * work. This now handles partial unmappings.
2585 * Jeremy Fitzhardinge <jeremy@goop.org>
2587 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2590 struct vm_area_struct *vma, *prev, *last;
2592 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2595 len = PAGE_ALIGN(len);
2599 /* Find the first overlapping VMA */
2600 vma = find_vma(mm, start);
2603 prev = vma->vm_prev;
2604 /* we have start < vma->vm_end */
2606 /* if it doesn't overlap, we have nothing.. */
2608 if (vma->vm_start >= end)
2612 * If we need to split any vma, do it now to save pain later.
2614 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2615 * unmapped vm_area_struct will remain in use: so lower split_vma
2616 * places tmp vma above, and higher split_vma places tmp vma below.
2618 if (start > vma->vm_start) {
2622 * Make sure that map_count on return from munmap() will
2623 * not exceed its limit; but let map_count go just above
2624 * its limit temporarily, to help free resources as expected.
2626 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2629 error = __split_vma(mm, vma, start, 0);
2635 /* Does it split the last one? */
2636 last = find_vma(mm, end);
2637 if (last && end > last->vm_start) {
2638 int error = __split_vma(mm, last, end, 1);
2642 vma = prev ? prev->vm_next : mm->mmap;
2645 * unlock any mlock()ed ranges before detaching vmas
2647 if (mm->locked_vm) {
2648 struct vm_area_struct *tmp = vma;
2649 while (tmp && tmp->vm_start < end) {
2650 if (tmp->vm_flags & VM_LOCKED) {
2651 mm->locked_vm -= vma_pages(tmp);
2652 munlock_vma_pages_all(tmp);
2659 * Remove the vma's, and unmap the actual pages
2661 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2662 unmap_region(mm, vma, prev, start, end);
2664 arch_unmap(mm, vma, start, end);
2666 /* Fix up all other VM information */
2667 remove_vma_list(mm, vma);
2671 EXPORT_SYMBOL(do_munmap);
2673 int vm_munmap(unsigned long start, size_t len)
2676 struct mm_struct *mm = current->mm;
2678 down_write(&mm->mmap_sem);
2679 ret = do_munmap(mm, start, len);
2680 up_write(&mm->mmap_sem);
2683 EXPORT_SYMBOL(vm_munmap);
2685 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2687 profile_munmap(addr);
2688 return vm_munmap(addr, len);
2693 * Emulation of deprecated remap_file_pages() syscall.
2695 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2696 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2699 struct mm_struct *mm = current->mm;
2700 struct vm_area_struct *vma;
2701 unsigned long populate = 0;
2702 unsigned long ret = -EINVAL;
2705 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. "
2706 "See Documentation/vm/remap_file_pages.txt.\n",
2707 current->comm, current->pid);
2711 start = start & PAGE_MASK;
2712 size = size & PAGE_MASK;
2714 if (start + size <= start)
2717 /* Does pgoff wrap? */
2718 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2721 down_write(&mm->mmap_sem);
2722 vma = find_vma(mm, start);
2724 if (!vma || !(vma->vm_flags & VM_SHARED))
2727 if (start < vma->vm_start)
2730 if (start + size > vma->vm_end) {
2731 struct vm_area_struct *next;
2733 for (next = vma->vm_next; next; next = next->vm_next) {
2734 /* hole between vmas ? */
2735 if (next->vm_start != next->vm_prev->vm_end)
2738 if (next->vm_file != vma->vm_file)
2741 if (next->vm_flags != vma->vm_flags)
2744 if (start + size <= next->vm_end)
2752 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2753 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2754 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2756 flags &= MAP_NONBLOCK;
2757 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2758 if (vma->vm_flags & VM_LOCKED) {
2759 struct vm_area_struct *tmp;
2760 flags |= MAP_LOCKED;
2762 /* drop PG_Mlocked flag for over-mapped range */
2763 for (tmp = vma; tmp->vm_start >= start + size;
2764 tmp = tmp->vm_next) {
2765 munlock_vma_pages_range(tmp,
2766 max(tmp->vm_start, start),
2767 min(tmp->vm_end, start + size));
2771 file = get_file(vma->vm_file);
2772 ret = do_mmap_pgoff(vma->vm_file, start, size,
2773 prot, flags, pgoff, &populate);
2776 up_write(&mm->mmap_sem);
2778 mm_populate(ret, populate);
2779 if (!IS_ERR_VALUE(ret))
2784 static inline void verify_mm_writelocked(struct mm_struct *mm)
2786 #ifdef CONFIG_DEBUG_VM
2787 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2789 up_read(&mm->mmap_sem);
2795 * this is really a simplified "do_mmap". it only handles
2796 * anonymous maps. eventually we may be able to do some
2797 * brk-specific accounting here.
2799 static unsigned long do_brk(unsigned long addr, unsigned long len)
2801 struct mm_struct *mm = current->mm;
2802 struct vm_area_struct *vma, *prev;
2803 unsigned long flags;
2804 struct rb_node **rb_link, *rb_parent;
2805 pgoff_t pgoff = addr >> PAGE_SHIFT;
2808 len = PAGE_ALIGN(len);
2812 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2814 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2815 if (offset_in_page(error))
2818 error = mlock_future_check(mm, mm->def_flags, len);
2823 * mm->mmap_sem is required to protect against another thread
2824 * changing the mappings in case we sleep.
2826 verify_mm_writelocked(mm);
2829 * Clear old maps. this also does some error checking for us
2831 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2833 if (do_munmap(mm, addr, len))
2837 /* Check against address space limits *after* clearing old maps... */
2838 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2841 if (mm->map_count > sysctl_max_map_count)
2844 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2847 /* Can we just expand an old private anonymous mapping? */
2848 vma = vma_merge(mm, prev, addr, addr + len, flags,
2849 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
2854 * create a vma struct for an anonymous mapping
2856 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2858 vm_unacct_memory(len >> PAGE_SHIFT);
2862 INIT_LIST_HEAD(&vma->anon_vma_chain);
2864 vma->vm_start = addr;
2865 vma->vm_end = addr + len;
2866 vma->vm_pgoff = pgoff;
2867 vma->vm_flags = flags;
2868 vma->vm_page_prot = vm_get_page_prot(flags);
2869 vma_link(mm, vma, prev, rb_link, rb_parent);
2871 perf_event_mmap(vma);
2872 mm->total_vm += len >> PAGE_SHIFT;
2873 if (flags & VM_LOCKED)
2874 mm->locked_vm += (len >> PAGE_SHIFT);
2875 vma->vm_flags |= VM_SOFTDIRTY;
2879 unsigned long vm_brk(unsigned long addr, unsigned long len)
2881 struct mm_struct *mm = current->mm;
2885 down_write(&mm->mmap_sem);
2886 ret = do_brk(addr, len);
2887 populate = ((mm->def_flags & VM_LOCKED) != 0);
2888 up_write(&mm->mmap_sem);
2890 mm_populate(addr, len);
2893 EXPORT_SYMBOL(vm_brk);
2895 /* Release all mmaps. */
2896 void exit_mmap(struct mm_struct *mm)
2898 struct mmu_gather tlb;
2899 struct vm_area_struct *vma;
2900 unsigned long nr_accounted = 0;
2902 /* mm's last user has gone, and its about to be pulled down */
2903 mmu_notifier_release(mm);
2905 if (mm->locked_vm) {
2908 if (vma->vm_flags & VM_LOCKED)
2909 munlock_vma_pages_all(vma);
2917 if (!vma) /* Can happen if dup_mmap() received an OOM */
2922 tlb_gather_mmu(&tlb, mm, 0, -1);
2923 /* update_hiwater_rss(mm) here? but nobody should be looking */
2924 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2925 unmap_vmas(&tlb, vma, 0, -1);
2927 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2928 tlb_finish_mmu(&tlb, 0, -1);
2931 * Walk the list again, actually closing and freeing it,
2932 * with preemption enabled, without holding any MM locks.
2935 if (vma->vm_flags & VM_ACCOUNT)
2936 nr_accounted += vma_pages(vma);
2937 vma = remove_vma(vma);
2939 vm_unacct_memory(nr_accounted);
2942 /* Insert vm structure into process list sorted by address
2943 * and into the inode's i_mmap tree. If vm_file is non-NULL
2944 * then i_mmap_rwsem is taken here.
2946 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2948 struct vm_area_struct *prev;
2949 struct rb_node **rb_link, *rb_parent;
2951 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2952 &prev, &rb_link, &rb_parent))
2954 if ((vma->vm_flags & VM_ACCOUNT) &&
2955 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2959 * The vm_pgoff of a purely anonymous vma should be irrelevant
2960 * until its first write fault, when page's anon_vma and index
2961 * are set. But now set the vm_pgoff it will almost certainly
2962 * end up with (unless mremap moves it elsewhere before that
2963 * first wfault), so /proc/pid/maps tells a consistent story.
2965 * By setting it to reflect the virtual start address of the
2966 * vma, merges and splits can happen in a seamless way, just
2967 * using the existing file pgoff checks and manipulations.
2968 * Similarly in do_mmap_pgoff and in do_brk.
2970 if (vma_is_anonymous(vma)) {
2971 BUG_ON(vma->anon_vma);
2972 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2975 vma_link(mm, vma, prev, rb_link, rb_parent);
2980 * Copy the vma structure to a new location in the same mm,
2981 * prior to moving page table entries, to effect an mremap move.
2983 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2984 unsigned long addr, unsigned long len, pgoff_t pgoff,
2985 bool *need_rmap_locks)
2987 struct vm_area_struct *vma = *vmap;
2988 unsigned long vma_start = vma->vm_start;
2989 struct mm_struct *mm = vma->vm_mm;
2990 struct vm_area_struct *new_vma, *prev;
2991 struct rb_node **rb_link, *rb_parent;
2992 bool faulted_in_anon_vma = true;
2995 * If anonymous vma has not yet been faulted, update new pgoff
2996 * to match new location, to increase its chance of merging.
2998 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
2999 pgoff = addr >> PAGE_SHIFT;
3000 faulted_in_anon_vma = false;
3003 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3004 return NULL; /* should never get here */
3005 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3006 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3007 vma->vm_userfaultfd_ctx, vma_get_anon_name(vma));
3010 * Source vma may have been merged into new_vma
3012 if (unlikely(vma_start >= new_vma->vm_start &&
3013 vma_start < new_vma->vm_end)) {
3015 * The only way we can get a vma_merge with
3016 * self during an mremap is if the vma hasn't
3017 * been faulted in yet and we were allowed to
3018 * reset the dst vma->vm_pgoff to the
3019 * destination address of the mremap to allow
3020 * the merge to happen. mremap must change the
3021 * vm_pgoff linearity between src and dst vmas
3022 * (in turn preventing a vma_merge) to be
3023 * safe. It is only safe to keep the vm_pgoff
3024 * linear if there are no pages mapped yet.
3026 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3027 *vmap = vma = new_vma;
3029 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3031 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3035 new_vma->vm_start = addr;
3036 new_vma->vm_end = addr + len;
3037 new_vma->vm_pgoff = pgoff;
3038 if (vma_dup_policy(vma, new_vma))
3040 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3041 if (anon_vma_clone(new_vma, vma))
3042 goto out_free_mempol;
3043 if (new_vma->vm_file)
3044 get_file(new_vma->vm_file);
3045 if (new_vma->vm_ops && new_vma->vm_ops->open)
3046 new_vma->vm_ops->open(new_vma);
3047 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3048 *need_rmap_locks = false;
3053 mpol_put(vma_policy(new_vma));
3055 kmem_cache_free(vm_area_cachep, new_vma);
3061 * Return true if the calling process may expand its vm space by the passed
3064 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
3066 unsigned long cur = mm->total_vm; /* pages */
3069 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
3071 if (cur + npages > lim)
3076 static int special_mapping_fault(struct vm_area_struct *vma,
3077 struct vm_fault *vmf);
3080 * Having a close hook prevents vma merging regardless of flags.
3082 static void special_mapping_close(struct vm_area_struct *vma)
3086 static const char *special_mapping_name(struct vm_area_struct *vma)
3088 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3091 static const struct vm_operations_struct special_mapping_vmops = {
3092 .close = special_mapping_close,
3093 .fault = special_mapping_fault,
3094 .name = special_mapping_name,
3097 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3098 .close = special_mapping_close,
3099 .fault = special_mapping_fault,
3102 static int special_mapping_fault(struct vm_area_struct *vma,
3103 struct vm_fault *vmf)
3106 struct page **pages;
3108 if (vma->vm_ops == &legacy_special_mapping_vmops)
3109 pages = vma->vm_private_data;
3111 pages = ((struct vm_special_mapping *)vma->vm_private_data)->
3114 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3118 struct page *page = *pages;
3124 return VM_FAULT_SIGBUS;
3127 static struct vm_area_struct *__install_special_mapping(
3128 struct mm_struct *mm,
3129 unsigned long addr, unsigned long len,
3130 unsigned long vm_flags, void *priv,
3131 const struct vm_operations_struct *ops)
3134 struct vm_area_struct *vma;
3136 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3137 if (unlikely(vma == NULL))
3138 return ERR_PTR(-ENOMEM);
3140 INIT_LIST_HEAD(&vma->anon_vma_chain);
3142 vma->vm_start = addr;
3143 vma->vm_end = addr + len;
3145 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3146 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3149 vma->vm_private_data = priv;
3151 ret = insert_vm_struct(mm, vma);
3155 mm->total_vm += len >> PAGE_SHIFT;
3157 perf_event_mmap(vma);
3162 kmem_cache_free(vm_area_cachep, vma);
3163 return ERR_PTR(ret);
3167 * Called with mm->mmap_sem held for writing.
3168 * Insert a new vma covering the given region, with the given flags.
3169 * Its pages are supplied by the given array of struct page *.
3170 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3171 * The region past the last page supplied will always produce SIGBUS.
3172 * The array pointer and the pages it points to are assumed to stay alive
3173 * for as long as this mapping might exist.
3175 struct vm_area_struct *_install_special_mapping(
3176 struct mm_struct *mm,
3177 unsigned long addr, unsigned long len,
3178 unsigned long vm_flags, const struct vm_special_mapping *spec)
3180 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3181 &special_mapping_vmops);
3184 int install_special_mapping(struct mm_struct *mm,
3185 unsigned long addr, unsigned long len,
3186 unsigned long vm_flags, struct page **pages)
3188 struct vm_area_struct *vma = __install_special_mapping(
3189 mm, addr, len, vm_flags, (void *)pages,
3190 &legacy_special_mapping_vmops);
3192 return PTR_ERR_OR_ZERO(vma);
3195 static DEFINE_MUTEX(mm_all_locks_mutex);
3197 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3199 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3201 * The LSB of head.next can't change from under us
3202 * because we hold the mm_all_locks_mutex.
3204 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3206 * We can safely modify head.next after taking the
3207 * anon_vma->root->rwsem. If some other vma in this mm shares
3208 * the same anon_vma we won't take it again.
3210 * No need of atomic instructions here, head.next
3211 * can't change from under us thanks to the
3212 * anon_vma->root->rwsem.
3214 if (__test_and_set_bit(0, (unsigned long *)
3215 &anon_vma->root->rb_root.rb_node))
3220 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3222 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3224 * AS_MM_ALL_LOCKS can't change from under us because
3225 * we hold the mm_all_locks_mutex.
3227 * Operations on ->flags have to be atomic because
3228 * even if AS_MM_ALL_LOCKS is stable thanks to the
3229 * mm_all_locks_mutex, there may be other cpus
3230 * changing other bitflags in parallel to us.
3232 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3234 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3239 * This operation locks against the VM for all pte/vma/mm related
3240 * operations that could ever happen on a certain mm. This includes
3241 * vmtruncate, try_to_unmap, and all page faults.
3243 * The caller must take the mmap_sem in write mode before calling
3244 * mm_take_all_locks(). The caller isn't allowed to release the
3245 * mmap_sem until mm_drop_all_locks() returns.
3247 * mmap_sem in write mode is required in order to block all operations
3248 * that could modify pagetables and free pages without need of
3249 * altering the vma layout. It's also needed in write mode to avoid new
3250 * anon_vmas to be associated with existing vmas.
3252 * A single task can't take more than one mm_take_all_locks() in a row
3253 * or it would deadlock.
3255 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3256 * mapping->flags avoid to take the same lock twice, if more than one
3257 * vma in this mm is backed by the same anon_vma or address_space.
3259 * We can take all the locks in random order because the VM code
3260 * taking i_mmap_rwsem or anon_vma->rwsem outside the mmap_sem never
3261 * takes more than one of them in a row. Secondly we're protected
3262 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3264 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3265 * that may have to take thousand of locks.
3267 * mm_take_all_locks() can fail if it's interrupted by signals.
3269 int mm_take_all_locks(struct mm_struct *mm)
3271 struct vm_area_struct *vma;
3272 struct anon_vma_chain *avc;
3274 BUG_ON(down_read_trylock(&mm->mmap_sem));
3276 mutex_lock(&mm_all_locks_mutex);
3278 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3279 if (signal_pending(current))
3281 if (vma->vm_file && vma->vm_file->f_mapping)
3282 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3285 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3286 if (signal_pending(current))
3289 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3290 vm_lock_anon_vma(mm, avc->anon_vma);
3296 mm_drop_all_locks(mm);
3300 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3302 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3304 * The LSB of head.next can't change to 0 from under
3305 * us because we hold the mm_all_locks_mutex.
3307 * We must however clear the bitflag before unlocking
3308 * the vma so the users using the anon_vma->rb_root will
3309 * never see our bitflag.
3311 * No need of atomic instructions here, head.next
3312 * can't change from under us until we release the
3313 * anon_vma->root->rwsem.
3315 if (!__test_and_clear_bit(0, (unsigned long *)
3316 &anon_vma->root->rb_root.rb_node))
3318 anon_vma_unlock_write(anon_vma);
3322 static void vm_unlock_mapping(struct address_space *mapping)
3324 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3326 * AS_MM_ALL_LOCKS can't change to 0 from under us
3327 * because we hold the mm_all_locks_mutex.
3329 i_mmap_unlock_write(mapping);
3330 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3337 * The mmap_sem cannot be released by the caller until
3338 * mm_drop_all_locks() returns.
3340 void mm_drop_all_locks(struct mm_struct *mm)
3342 struct vm_area_struct *vma;
3343 struct anon_vma_chain *avc;
3345 BUG_ON(down_read_trylock(&mm->mmap_sem));
3346 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3348 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3350 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3351 vm_unlock_anon_vma(avc->anon_vma);
3352 if (vma->vm_file && vma->vm_file->f_mapping)
3353 vm_unlock_mapping(vma->vm_file->f_mapping);
3356 mutex_unlock(&mm_all_locks_mutex);
3360 * initialise the VMA slab
3362 void __init mmap_init(void)
3366 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3371 * Initialise sysctl_user_reserve_kbytes.
3373 * This is intended to prevent a user from starting a single memory hogging
3374 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3377 * The default value is min(3% of free memory, 128MB)
3378 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3380 static int init_user_reserve(void)
3382 unsigned long free_kbytes;
3384 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3386 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3389 subsys_initcall(init_user_reserve);
3392 * Initialise sysctl_admin_reserve_kbytes.
3394 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3395 * to log in and kill a memory hogging process.
3397 * Systems with more than 256MB will reserve 8MB, enough to recover
3398 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3399 * only reserve 3% of free pages by default.
3401 static int init_admin_reserve(void)
3403 unsigned long free_kbytes;
3405 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3407 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3410 subsys_initcall(init_admin_reserve);
3413 * Reinititalise user and admin reserves if memory is added or removed.
3415 * The default user reserve max is 128MB, and the default max for the
3416 * admin reserve is 8MB. These are usually, but not always, enough to
3417 * enable recovery from a memory hogging process using login/sshd, a shell,
3418 * and tools like top. It may make sense to increase or even disable the
3419 * reserve depending on the existence of swap or variations in the recovery
3420 * tools. So, the admin may have changed them.
3422 * If memory is added and the reserves have been eliminated or increased above
3423 * the default max, then we'll trust the admin.
3425 * If memory is removed and there isn't enough free memory, then we
3426 * need to reset the reserves.
3428 * Otherwise keep the reserve set by the admin.
3430 static int reserve_mem_notifier(struct notifier_block *nb,
3431 unsigned long action, void *data)
3433 unsigned long tmp, free_kbytes;
3437 /* Default max is 128MB. Leave alone if modified by operator. */
3438 tmp = sysctl_user_reserve_kbytes;
3439 if (0 < tmp && tmp < (1UL << 17))
3440 init_user_reserve();
3442 /* Default max is 8MB. Leave alone if modified by operator. */
3443 tmp = sysctl_admin_reserve_kbytes;
3444 if (0 < tmp && tmp < (1UL << 13))
3445 init_admin_reserve();
3449 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3451 if (sysctl_user_reserve_kbytes > free_kbytes) {
3452 init_user_reserve();
3453 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3454 sysctl_user_reserve_kbytes);
3457 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3458 init_admin_reserve();
3459 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3460 sysctl_admin_reserve_kbytes);
3469 static struct notifier_block reserve_mem_nb = {
3470 .notifier_call = reserve_mem_notifier,
3473 static int __meminit init_reserve_notifier(void)
3475 if (register_hotmemory_notifier(&reserve_mem_nb))
3476 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3480 subsys_initcall(init_reserve_notifier);