1 // SPDX-License-Identifier: GPL-2.0-only
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/backing-dev.h>
16 #include <linux/mm_inline.h>
17 #include <linux/shm.h>
18 #include <linux/mman.h>
19 #include <linux/pagemap.h>
20 #include <linux/swap.h>
21 #include <linux/syscalls.h>
22 #include <linux/capability.h>
23 #include <linux/init.h>
24 #include <linux/file.h>
26 #include <linux/personality.h>
27 #include <linux/security.h>
28 #include <linux/hugetlb.h>
29 #include <linux/shmem_fs.h>
30 #include <linux/profile.h>
31 #include <linux/export.h>
32 #include <linux/mount.h>
33 #include <linux/mempolicy.h>
34 #include <linux/rmap.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/mmdebug.h>
37 #include <linux/perf_event.h>
38 #include <linux/audit.h>
39 #include <linux/khugepaged.h>
40 #include <linux/uprobes.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/moduleparam.h>
46 #include <linux/pkeys.h>
47 #include <linux/oom.h>
48 #include <linux/sched/mm.h>
50 #include <linux/uaccess.h>
51 #include <asm/cacheflush.h>
53 #include <asm/mmu_context.h>
55 #define CREATE_TRACE_POINTS
56 #include <trace/events/mmap.h>
60 #ifndef arch_mmap_check
61 #define arch_mmap_check(addr, len, flags) (0)
64 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
65 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
66 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
67 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
69 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
70 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
71 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
72 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
75 static bool ignore_rlimit_data;
76 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
78 static void unmap_region(struct mm_struct *mm, struct maple_tree *mt,
79 struct vm_area_struct *vma, struct vm_area_struct *prev,
80 struct vm_area_struct *next, unsigned long start,
83 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
85 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
88 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
89 void vma_set_page_prot(struct vm_area_struct *vma)
91 unsigned long vm_flags = vma->vm_flags;
92 pgprot_t vm_page_prot;
94 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
95 if (vma_wants_writenotify(vma, vm_page_prot)) {
96 vm_flags &= ~VM_SHARED;
97 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
99 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
100 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
104 * Requires inode->i_mapping->i_mmap_rwsem
106 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
107 struct file *file, struct address_space *mapping)
109 if (vma->vm_flags & VM_SHARED)
110 mapping_unmap_writable(mapping);
112 flush_dcache_mmap_lock(mapping);
113 vma_interval_tree_remove(vma, &mapping->i_mmap);
114 flush_dcache_mmap_unlock(mapping);
118 * Unlink a file-based vm structure from its interval tree, to hide
119 * vma from rmap and vmtruncate before freeing its page tables.
121 void unlink_file_vma(struct vm_area_struct *vma)
123 struct file *file = vma->vm_file;
126 struct address_space *mapping = file->f_mapping;
127 i_mmap_lock_write(mapping);
128 __remove_shared_vm_struct(vma, file, mapping);
129 i_mmap_unlock_write(mapping);
134 * Close a vm structure and free it.
136 static void remove_vma(struct vm_area_struct *vma)
139 if (vma->vm_ops && vma->vm_ops->close)
140 vma->vm_ops->close(vma);
143 mpol_put(vma_policy(vma));
148 * check_brk_limits() - Use platform specific check of range & verify mlock
150 * @addr: The address to check
151 * @len: The size of increase.
153 * Return: 0 on success.
155 static int check_brk_limits(unsigned long addr, unsigned long len)
157 unsigned long mapped_addr;
159 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
160 if (IS_ERR_VALUE(mapped_addr))
163 return mlock_future_check(current->mm, current->mm->def_flags, len);
165 static int do_brk_munmap(struct ma_state *mas, struct vm_area_struct *vma,
166 unsigned long newbrk, unsigned long oldbrk,
167 struct list_head *uf);
168 static int do_brk_flags(struct ma_state *mas, struct vm_area_struct *brkvma,
169 unsigned long addr, unsigned long request, unsigned long flags);
170 SYSCALL_DEFINE1(brk, unsigned long, brk)
172 unsigned long newbrk, oldbrk, origbrk;
173 struct mm_struct *mm = current->mm;
174 struct vm_area_struct *brkvma, *next = NULL;
175 unsigned long min_brk;
177 bool downgraded = false;
179 MA_STATE(mas, &mm->mm_mt, 0, 0);
181 if (mmap_write_lock_killable(mm))
186 #ifdef CONFIG_COMPAT_BRK
188 * CONFIG_COMPAT_BRK can still be overridden by setting
189 * randomize_va_space to 2, which will still cause mm->start_brk
190 * to be arbitrarily shifted
192 if (current->brk_randomized)
193 min_brk = mm->start_brk;
195 min_brk = mm->end_data;
197 min_brk = mm->start_brk;
203 * Check against rlimit here. If this check is done later after the test
204 * of oldbrk with newbrk then it can escape the test and let the data
205 * segment grow beyond its set limit the in case where the limit is
206 * not page aligned -Ram Gupta
208 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
209 mm->end_data, mm->start_data))
212 newbrk = PAGE_ALIGN(brk);
213 oldbrk = PAGE_ALIGN(mm->brk);
214 if (oldbrk == newbrk) {
220 * Always allow shrinking brk.
221 * do_brk_munmap() may downgrade mmap_lock to read.
223 if (brk <= mm->brk) {
226 /* Search one past newbrk */
227 mas_set(&mas, newbrk);
228 brkvma = mas_find(&mas, oldbrk);
229 if (!brkvma || brkvma->vm_start >= oldbrk)
230 goto out; /* mapping intersects with an existing non-brk vma. */
232 * mm->brk must be protected by write mmap_lock.
233 * do_brk_munmap() may downgrade the lock, so update it
234 * before calling do_brk_munmap().
237 ret = do_brk_munmap(&mas, brkvma, newbrk, oldbrk, &uf);
248 if (check_brk_limits(oldbrk, newbrk - oldbrk))
252 * Only check if the next VMA is within the stack_guard_gap of the
255 mas_set(&mas, oldbrk);
256 next = mas_find(&mas, newbrk - 1 + PAGE_SIZE + stack_guard_gap);
257 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
260 brkvma = mas_prev(&mas, mm->start_brk);
261 /* Ok, looks good - let it rip. */
262 if (do_brk_flags(&mas, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
268 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
270 mmap_read_unlock(mm);
272 mmap_write_unlock(mm);
273 userfaultfd_unmap_complete(mm, &uf);
275 mm_populate(oldbrk, newbrk - oldbrk);
279 mmap_write_unlock(mm);
283 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
284 extern void mt_validate(struct maple_tree *mt);
285 extern void mt_dump(const struct maple_tree *mt);
287 /* Validate the maple tree */
288 static void validate_mm_mt(struct mm_struct *mm)
290 struct maple_tree *mt = &mm->mm_mt;
291 struct vm_area_struct *vma_mt;
293 MA_STATE(mas, mt, 0, 0);
295 mt_validate(&mm->mm_mt);
296 mas_for_each(&mas, vma_mt, ULONG_MAX) {
297 if ((vma_mt->vm_start != mas.index) ||
298 (vma_mt->vm_end - 1 != mas.last)) {
299 pr_emerg("issue in %s\n", current->comm);
302 pr_emerg("mt piv: %p %lu - %lu\n", vma_mt,
303 mas.index, mas.last);
304 pr_emerg("mt vma: %p %lu - %lu\n", vma_mt,
305 vma_mt->vm_start, vma_mt->vm_end);
308 if (vma_mt->vm_end != mas.last + 1) {
309 pr_err("vma: %p vma_mt %lu-%lu\tmt %lu-%lu\n",
310 mm, vma_mt->vm_start, vma_mt->vm_end,
311 mas.index, mas.last);
314 VM_BUG_ON_MM(vma_mt->vm_end != mas.last + 1, mm);
315 if (vma_mt->vm_start != mas.index) {
316 pr_err("vma: %p vma_mt %p %lu - %lu doesn't match\n",
317 mm, vma_mt, vma_mt->vm_start, vma_mt->vm_end);
320 VM_BUG_ON_MM(vma_mt->vm_start != mas.index, mm);
325 static void validate_mm(struct mm_struct *mm)
329 struct vm_area_struct *vma;
330 MA_STATE(mas, &mm->mm_mt, 0, 0);
334 mas_for_each(&mas, vma, ULONG_MAX) {
335 #ifdef CONFIG_DEBUG_VM_RB
336 struct anon_vma *anon_vma = vma->anon_vma;
337 struct anon_vma_chain *avc;
340 anon_vma_lock_read(anon_vma);
341 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
342 anon_vma_interval_tree_verify(avc);
343 anon_vma_unlock_read(anon_vma);
348 if (i != mm->map_count) {
349 pr_emerg("map_count %d mas_for_each %d\n", mm->map_count, i);
352 VM_BUG_ON_MM(bug, mm);
355 #else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
356 #define validate_mm_mt(root) do { } while (0)
357 #define validate_mm(mm) do { } while (0)
358 #endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
361 * vma has some anon_vma assigned, and is already inserted on that
362 * anon_vma's interval trees.
364 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
365 * vma must be removed from the anon_vma's interval trees using
366 * anon_vma_interval_tree_pre_update_vma().
368 * After the update, the vma will be reinserted using
369 * anon_vma_interval_tree_post_update_vma().
371 * The entire update must be protected by exclusive mmap_lock and by
372 * the root anon_vma's mutex.
375 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
377 struct anon_vma_chain *avc;
379 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
380 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
384 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
386 struct anon_vma_chain *avc;
388 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
389 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
392 static unsigned long count_vma_pages_range(struct mm_struct *mm,
393 unsigned long addr, unsigned long end)
395 VMA_ITERATOR(vmi, mm, addr);
396 struct vm_area_struct *vma;
397 unsigned long nr_pages = 0;
399 for_each_vma_range(vmi, vma, end) {
400 unsigned long vm_start = max(addr, vma->vm_start);
401 unsigned long vm_end = min(end, vma->vm_end);
403 nr_pages += PHYS_PFN(vm_end - vm_start);
409 static void __vma_link_file(struct vm_area_struct *vma,
410 struct address_space *mapping)
412 if (vma->vm_flags & VM_SHARED)
413 mapping_allow_writable(mapping);
415 flush_dcache_mmap_lock(mapping);
416 vma_interval_tree_insert(vma, &mapping->i_mmap);
417 flush_dcache_mmap_unlock(mapping);
421 * vma_mas_store() - Store a VMA in the maple tree.
422 * @vma: The vm_area_struct
423 * @mas: The maple state
425 * Efficient way to store a VMA in the maple tree when the @mas has already
426 * walked to the correct location.
428 * Note: the end address is inclusive in the maple tree.
430 void vma_mas_store(struct vm_area_struct *vma, struct ma_state *mas)
432 trace_vma_store(mas->tree, vma);
433 mas_set_range(mas, vma->vm_start, vma->vm_end - 1);
434 mas_store_prealloc(mas, vma);
438 * vma_mas_remove() - Remove a VMA from the maple tree.
439 * @vma: The vm_area_struct
440 * @mas: The maple state
442 * Efficient way to remove a VMA from the maple tree when the @mas has already
443 * been established and points to the correct location.
444 * Note: the end address is inclusive in the maple tree.
446 void vma_mas_remove(struct vm_area_struct *vma, struct ma_state *mas)
448 trace_vma_mas_szero(mas->tree, vma->vm_start, vma->vm_end - 1);
449 mas->index = vma->vm_start;
450 mas->last = vma->vm_end - 1;
451 mas_store_prealloc(mas, NULL);
455 * vma_mas_szero() - Set a given range to zero. Used when modifying a
456 * vm_area_struct start or end.
458 * @mas: The maple tree ma_state
459 * @start: The start address to zero
460 * @end: The end address to zero.
462 static inline void vma_mas_szero(struct ma_state *mas, unsigned long start,
465 trace_vma_mas_szero(mas->tree, start, end - 1);
466 mas_set_range(mas, start, end - 1);
467 mas_store_prealloc(mas, NULL);
470 static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
472 MA_STATE(mas, &mm->mm_mt, 0, 0);
473 struct address_space *mapping = NULL;
475 if (mas_preallocate(&mas, vma, GFP_KERNEL))
479 mapping = vma->vm_file->f_mapping;
480 i_mmap_lock_write(mapping);
483 vma_mas_store(vma, &mas);
486 __vma_link_file(vma, mapping);
487 i_mmap_unlock_write(mapping);
496 * vma_expand - Expand an existing VMA
498 * @mas: The maple state
499 * @vma: The vma to expand
500 * @start: The start of the vma
501 * @end: The exclusive end of the vma
502 * @pgoff: The page offset of vma
503 * @next: The current of next vma.
505 * Expand @vma to @start and @end. Can expand off the start and end. Will
506 * expand over @next if it's different from @vma and @end == @next->vm_end.
507 * Checking if the @vma can expand and merge with @next needs to be handled by
510 * Returns: 0 on success
512 inline int vma_expand(struct ma_state *mas, struct vm_area_struct *vma,
513 unsigned long start, unsigned long end, pgoff_t pgoff,
514 struct vm_area_struct *next)
516 struct mm_struct *mm = vma->vm_mm;
517 struct address_space *mapping = NULL;
518 struct rb_root_cached *root = NULL;
519 struct anon_vma *anon_vma = vma->anon_vma;
520 struct file *file = vma->vm_file;
521 bool remove_next = false;
523 if (next && (vma != next) && (end == next->vm_end)) {
525 if (next->anon_vma && !vma->anon_vma) {
528 anon_vma = next->anon_vma;
529 vma->anon_vma = anon_vma;
530 error = anon_vma_clone(vma, next);
536 /* Not merging but overwriting any part of next is not handled. */
537 VM_BUG_ON(next && !remove_next && next != vma && end > next->vm_start);
538 /* Only handles expanding */
539 VM_BUG_ON(vma->vm_start < start || vma->vm_end > end);
541 if (mas_preallocate(mas, vma, GFP_KERNEL))
544 vma_adjust_trans_huge(vma, start, end, 0);
547 mapping = file->f_mapping;
548 root = &mapping->i_mmap;
549 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
550 i_mmap_lock_write(mapping);
554 anon_vma_lock_write(anon_vma);
555 anon_vma_interval_tree_pre_update_vma(vma);
559 flush_dcache_mmap_lock(mapping);
560 vma_interval_tree_remove(vma, root);
563 vma->vm_start = start;
565 vma->vm_pgoff = pgoff;
566 /* Note: mas must be pointing to the expanding VMA */
567 vma_mas_store(vma, mas);
570 vma_interval_tree_insert(vma, root);
571 flush_dcache_mmap_unlock(mapping);
574 /* Expanding over the next vma */
575 if (remove_next && file) {
576 __remove_shared_vm_struct(next, file, mapping);
580 anon_vma_interval_tree_post_update_vma(vma);
581 anon_vma_unlock_write(anon_vma);
585 i_mmap_unlock_write(mapping);
591 uprobe_munmap(next, next->vm_start, next->vm_end);
595 anon_vma_merge(vma, next);
597 mpol_put(vma_policy(next));
609 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
610 * is already present in an i_mmap tree without adjusting the tree.
611 * The following helper function should be used when such adjustments
612 * are necessary. The "insert" vma (if any) is to be inserted
613 * before we drop the necessary locks.
615 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
616 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
617 struct vm_area_struct *expand)
619 struct mm_struct *mm = vma->vm_mm;
620 struct vm_area_struct *next_next = NULL; /* uninit var warning */
621 struct vm_area_struct *next = find_vma(mm, vma->vm_end);
622 struct vm_area_struct *orig_vma = vma;
623 struct address_space *mapping = NULL;
624 struct rb_root_cached *root = NULL;
625 struct anon_vma *anon_vma = NULL;
626 struct file *file = vma->vm_file;
627 bool vma_changed = false;
628 long adjust_next = 0;
630 MA_STATE(mas, &mm->mm_mt, 0, 0);
631 struct vm_area_struct *exporter = NULL, *importer = NULL;
633 if (next && !insert) {
634 if (end >= next->vm_end) {
636 * vma expands, overlapping all the next, and
637 * perhaps the one after too (mprotect case 6).
638 * The only other cases that gets here are
639 * case 1, case 7 and case 8.
641 if (next == expand) {
643 * The only case where we don't expand "vma"
644 * and we expand "next" instead is case 8.
646 VM_WARN_ON(end != next->vm_end);
648 * remove_next == 3 means we're
649 * removing "vma" and that to do so we
650 * swapped "vma" and "next".
653 VM_WARN_ON(file != next->vm_file);
656 VM_WARN_ON(expand != vma);
658 * case 1, 6, 7, remove_next == 2 is case 6,
659 * remove_next == 1 is case 1 or 7.
661 remove_next = 1 + (end > next->vm_end);
662 if (remove_next == 2)
663 next_next = find_vma(mm, next->vm_end);
665 VM_WARN_ON(remove_next == 2 &&
666 end != next_next->vm_end);
673 * If next doesn't have anon_vma, import from vma after
674 * next, if the vma overlaps with it.
676 if (remove_next == 2 && !next->anon_vma)
677 exporter = next_next;
679 } else if (end > next->vm_start) {
681 * vma expands, overlapping part of the next:
682 * mprotect case 5 shifting the boundary up.
684 adjust_next = (end - next->vm_start);
687 VM_WARN_ON(expand != importer);
688 } else if (end < vma->vm_end) {
690 * vma shrinks, and !insert tells it's not
691 * split_vma inserting another: so it must be
692 * mprotect case 4 shifting the boundary down.
694 adjust_next = -(vma->vm_end - end);
697 VM_WARN_ON(expand != importer);
701 * Easily overlooked: when mprotect shifts the boundary,
702 * make sure the expanding vma has anon_vma set if the
703 * shrinking vma had, to cover any anon pages imported.
705 if (exporter && exporter->anon_vma && !importer->anon_vma) {
708 importer->anon_vma = exporter->anon_vma;
709 error = anon_vma_clone(importer, exporter);
715 if (mas_preallocate(&mas, vma, GFP_KERNEL))
718 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
720 mapping = file->f_mapping;
721 root = &mapping->i_mmap;
722 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
725 uprobe_munmap(next, next->vm_start, next->vm_end);
727 i_mmap_lock_write(mapping);
728 if (insert && insert->vm_file) {
730 * Put into interval tree now, so instantiated pages
731 * are visible to arm/parisc __flush_dcache_page
732 * throughout; but we cannot insert into address
733 * space until vma start or end is updated.
735 __vma_link_file(insert, insert->vm_file->f_mapping);
739 anon_vma = vma->anon_vma;
740 if (!anon_vma && adjust_next)
741 anon_vma = next->anon_vma;
743 VM_WARN_ON(adjust_next && next->anon_vma &&
744 anon_vma != next->anon_vma);
745 anon_vma_lock_write(anon_vma);
746 anon_vma_interval_tree_pre_update_vma(vma);
748 anon_vma_interval_tree_pre_update_vma(next);
752 flush_dcache_mmap_lock(mapping);
753 vma_interval_tree_remove(vma, root);
755 vma_interval_tree_remove(next, root);
758 if (start != vma->vm_start) {
759 if ((vma->vm_start < start) &&
760 (!insert || (insert->vm_end != start))) {
761 vma_mas_szero(&mas, vma->vm_start, start);
762 VM_WARN_ON(insert && insert->vm_start > vma->vm_start);
766 vma->vm_start = start;
768 if (end != vma->vm_end) {
769 if (vma->vm_end > end) {
770 if ((vma->vm_end + adjust_next != end) &&
771 (!insert || (insert->vm_start != end))) {
772 vma_mas_szero(&mas, end, vma->vm_end);
775 insert->vm_end < vma->vm_end);
784 vma_mas_store(vma, &mas);
786 vma->vm_pgoff = pgoff;
788 next->vm_start += adjust_next;
789 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
790 vma_mas_store(next, &mas);
795 vma_interval_tree_insert(next, root);
796 vma_interval_tree_insert(vma, root);
797 flush_dcache_mmap_unlock(mapping);
800 if (remove_next && file) {
801 __remove_shared_vm_struct(next, file, mapping);
802 if (remove_next == 2)
803 __remove_shared_vm_struct(next_next, file, mapping);
806 * split_vma has split insert from vma, and needs
807 * us to insert it before dropping the locks
808 * (it may either follow vma or precede it).
811 vma_mas_store(insert, &mas);
816 anon_vma_interval_tree_post_update_vma(vma);
818 anon_vma_interval_tree_post_update_vma(next);
819 anon_vma_unlock_write(anon_vma);
823 i_mmap_unlock_write(mapping);
833 uprobe_munmap(next, next->vm_start, next->vm_end);
837 anon_vma_merge(vma, next);
839 mpol_put(vma_policy(next));
840 if (remove_next != 2)
841 BUG_ON(vma->vm_end < next->vm_end);
845 * In mprotect's case 6 (see comments on vma_merge),
846 * we must remove next_next too.
848 if (remove_next == 2) {
864 * If the vma has a ->close operation then the driver probably needs to release
865 * per-vma resources, so we don't attempt to merge those.
867 static inline int is_mergeable_vma(struct vm_area_struct *vma,
868 struct file *file, unsigned long vm_flags,
869 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
870 struct anon_vma_name *anon_name)
873 * VM_SOFTDIRTY should not prevent from VMA merging, if we
874 * match the flags but dirty bit -- the caller should mark
875 * merged VMA as dirty. If dirty bit won't be excluded from
876 * comparison, we increase pressure on the memory system forcing
877 * the kernel to generate new VMAs when old one could be
880 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
882 if (vma->vm_file != file)
884 if (vma->vm_ops && vma->vm_ops->close)
886 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
888 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
893 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
894 struct anon_vma *anon_vma2,
895 struct vm_area_struct *vma)
898 * The list_is_singular() test is to avoid merging VMA cloned from
899 * parents. This can improve scalability caused by anon_vma lock.
901 if ((!anon_vma1 || !anon_vma2) && (!vma ||
902 list_is_singular(&vma->anon_vma_chain)))
904 return anon_vma1 == anon_vma2;
908 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
909 * in front of (at a lower virtual address and file offset than) the vma.
911 * We cannot merge two vmas if they have differently assigned (non-NULL)
912 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
914 * We don't check here for the merged mmap wrapping around the end of pagecache
915 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
916 * wrap, nor mmaps which cover the final page at index -1UL.
919 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
920 struct anon_vma *anon_vma, struct file *file,
922 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
923 struct anon_vma_name *anon_name)
925 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
926 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
927 if (vma->vm_pgoff == vm_pgoff)
934 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
935 * beyond (at a higher virtual address and file offset than) the vma.
937 * We cannot merge two vmas if they have differently assigned (non-NULL)
938 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
941 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
942 struct anon_vma *anon_vma, struct file *file,
944 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
945 struct anon_vma_name *anon_name)
947 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
948 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
950 vm_pglen = vma_pages(vma);
951 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
958 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
959 * figure out whether that can be merged with its predecessor or its
960 * successor. Or both (it neatly fills a hole).
962 * In most cases - when called for mmap, brk or mremap - [addr,end) is
963 * certain not to be mapped by the time vma_merge is called; but when
964 * called for mprotect, it is certain to be already mapped (either at
965 * an offset within prev, or at the start of next), and the flags of
966 * this area are about to be changed to vm_flags - and the no-change
967 * case has already been eliminated.
969 * The following mprotect cases have to be considered, where AAAA is
970 * the area passed down from mprotect_fixup, never extending beyond one
971 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
974 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
975 * cannot merge might become might become
976 * PPNNNNNNNNNN PPPPPPPPPPNN
977 * mmap, brk or case 4 below case 5 below
980 * PPPP NNNN PPPPNNNNXXXX
981 * might become might become
982 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
983 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
984 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
986 * It is important for case 8 that the vma NNNN overlapping the
987 * region AAAA is never going to extended over XXXX. Instead XXXX must
988 * be extended in region AAAA and NNNN must be removed. This way in
989 * all cases where vma_merge succeeds, the moment vma_adjust drops the
990 * rmap_locks, the properties of the merged vma will be already
991 * correct for the whole merged range. Some of those properties like
992 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
993 * be correct for the whole merged range immediately after the
994 * rmap_locks are released. Otherwise if XXXX would be removed and
995 * NNNN would be extended over the XXXX range, remove_migration_ptes
996 * or other rmap walkers (if working on addresses beyond the "end"
997 * parameter) may establish ptes with the wrong permissions of NNNN
998 * instead of the right permissions of XXXX.
1000 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1001 struct vm_area_struct *prev, unsigned long addr,
1002 unsigned long end, unsigned long vm_flags,
1003 struct anon_vma *anon_vma, struct file *file,
1004 pgoff_t pgoff, struct mempolicy *policy,
1005 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1006 struct anon_vma_name *anon_name)
1008 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1009 struct vm_area_struct *mid, *next, *res;
1011 bool merge_prev = false;
1012 bool merge_next = false;
1015 * We later require that vma->vm_flags == vm_flags,
1016 * so this tests vma->vm_flags & VM_SPECIAL, too.
1018 if (vm_flags & VM_SPECIAL)
1021 next = find_vma(mm, prev ? prev->vm_end : 0);
1023 if (next && next->vm_end == end) /* cases 6, 7, 8 */
1024 next = find_vma(mm, next->vm_end);
1026 /* verify some invariant that must be enforced by the caller */
1027 VM_WARN_ON(prev && addr <= prev->vm_start);
1028 VM_WARN_ON(mid && end > mid->vm_end);
1029 VM_WARN_ON(addr >= end);
1031 /* Can we merge the predecessor? */
1032 if (prev && prev->vm_end == addr &&
1033 mpol_equal(vma_policy(prev), policy) &&
1034 can_vma_merge_after(prev, vm_flags,
1035 anon_vma, file, pgoff,
1036 vm_userfaultfd_ctx, anon_name)) {
1039 /* Can we merge the successor? */
1040 if (next && end == next->vm_start &&
1041 mpol_equal(policy, vma_policy(next)) &&
1042 can_vma_merge_before(next, vm_flags,
1043 anon_vma, file, pgoff+pglen,
1044 vm_userfaultfd_ctx, anon_name)) {
1047 /* Can we merge both the predecessor and the successor? */
1048 if (merge_prev && merge_next &&
1049 is_mergeable_anon_vma(prev->anon_vma,
1050 next->anon_vma, NULL)) { /* cases 1, 6 */
1051 err = __vma_adjust(prev, prev->vm_start,
1052 next->vm_end, prev->vm_pgoff, NULL,
1055 } else if (merge_prev) { /* cases 2, 5, 7 */
1056 err = __vma_adjust(prev, prev->vm_start,
1057 end, prev->vm_pgoff, NULL, prev);
1059 } else if (merge_next) {
1060 if (prev && addr < prev->vm_end) /* case 4 */
1061 err = __vma_adjust(prev, prev->vm_start,
1062 addr, prev->vm_pgoff, NULL, next);
1063 else /* cases 3, 8 */
1064 err = __vma_adjust(mid, addr, next->vm_end,
1065 next->vm_pgoff - pglen, NULL, next);
1070 * Cannot merge with predecessor or successor or error in __vma_adjust?
1074 khugepaged_enter_vma(res, vm_flags);
1079 * Rough compatibility check to quickly see if it's even worth looking
1080 * at sharing an anon_vma.
1082 * They need to have the same vm_file, and the flags can only differ
1083 * in things that mprotect may change.
1085 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1086 * we can merge the two vma's. For example, we refuse to merge a vma if
1087 * there is a vm_ops->close() function, because that indicates that the
1088 * driver is doing some kind of reference counting. But that doesn't
1089 * really matter for the anon_vma sharing case.
1091 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1093 return a->vm_end == b->vm_start &&
1094 mpol_equal(vma_policy(a), vma_policy(b)) &&
1095 a->vm_file == b->vm_file &&
1096 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1097 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1101 * Do some basic sanity checking to see if we can re-use the anon_vma
1102 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1103 * the same as 'old', the other will be the new one that is trying
1104 * to share the anon_vma.
1106 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1107 * the anon_vma of 'old' is concurrently in the process of being set up
1108 * by another page fault trying to merge _that_. But that's ok: if it
1109 * is being set up, that automatically means that it will be a singleton
1110 * acceptable for merging, so we can do all of this optimistically. But
1111 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1113 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1114 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1115 * is to return an anon_vma that is "complex" due to having gone through
1118 * We also make sure that the two vma's are compatible (adjacent,
1119 * and with the same memory policies). That's all stable, even with just
1120 * a read lock on the mmap_lock.
1122 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1124 if (anon_vma_compatible(a, b)) {
1125 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1127 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1134 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1135 * neighbouring vmas for a suitable anon_vma, before it goes off
1136 * to allocate a new anon_vma. It checks because a repetitive
1137 * sequence of mprotects and faults may otherwise lead to distinct
1138 * anon_vmas being allocated, preventing vma merge in subsequent
1141 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1143 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1144 struct anon_vma *anon_vma = NULL;
1145 struct vm_area_struct *prev, *next;
1147 /* Try next first. */
1148 next = mas_walk(&mas);
1150 anon_vma = reusable_anon_vma(next, vma, next);
1155 prev = mas_prev(&mas, 0);
1156 VM_BUG_ON_VMA(prev != vma, vma);
1157 prev = mas_prev(&mas, 0);
1158 /* Try prev next. */
1160 anon_vma = reusable_anon_vma(prev, prev, vma);
1163 * We might reach here with anon_vma == NULL if we can't find
1164 * any reusable anon_vma.
1165 * There's no absolute need to look only at touching neighbours:
1166 * we could search further afield for "compatible" anon_vmas.
1167 * But it would probably just be a waste of time searching,
1168 * or lead to too many vmas hanging off the same anon_vma.
1169 * We're trying to allow mprotect remerging later on,
1170 * not trying to minimize memory used for anon_vmas.
1176 * If a hint addr is less than mmap_min_addr change hint to be as
1177 * low as possible but still greater than mmap_min_addr
1179 static inline unsigned long round_hint_to_min(unsigned long hint)
1182 if (((void *)hint != NULL) &&
1183 (hint < mmap_min_addr))
1184 return PAGE_ALIGN(mmap_min_addr);
1188 int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1191 unsigned long locked, lock_limit;
1193 /* mlock MCL_FUTURE? */
1194 if (flags & VM_LOCKED) {
1195 locked = len >> PAGE_SHIFT;
1196 locked += mm->locked_vm;
1197 lock_limit = rlimit(RLIMIT_MEMLOCK);
1198 lock_limit >>= PAGE_SHIFT;
1199 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1205 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1207 if (S_ISREG(inode->i_mode))
1208 return MAX_LFS_FILESIZE;
1210 if (S_ISBLK(inode->i_mode))
1211 return MAX_LFS_FILESIZE;
1213 if (S_ISSOCK(inode->i_mode))
1214 return MAX_LFS_FILESIZE;
1216 /* Special "we do even unsigned file positions" case */
1217 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1220 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1224 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1225 unsigned long pgoff, unsigned long len)
1227 u64 maxsize = file_mmap_size_max(file, inode);
1229 if (maxsize && len > maxsize)
1232 if (pgoff > maxsize >> PAGE_SHIFT)
1238 * The caller must write-lock current->mm->mmap_lock.
1240 unsigned long do_mmap(struct file *file, unsigned long addr,
1241 unsigned long len, unsigned long prot,
1242 unsigned long flags, unsigned long pgoff,
1243 unsigned long *populate, struct list_head *uf)
1245 struct mm_struct *mm = current->mm;
1246 vm_flags_t vm_flags;
1256 * Does the application expect PROT_READ to imply PROT_EXEC?
1258 * (the exception is when the underlying filesystem is noexec
1259 * mounted, in which case we dont add PROT_EXEC.)
1261 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1262 if (!(file && path_noexec(&file->f_path)))
1265 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1266 if (flags & MAP_FIXED_NOREPLACE)
1269 if (!(flags & MAP_FIXED))
1270 addr = round_hint_to_min(addr);
1272 /* Careful about overflows.. */
1273 len = PAGE_ALIGN(len);
1277 /* offset overflow? */
1278 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1281 /* Too many mappings? */
1282 if (mm->map_count > sysctl_max_map_count)
1285 /* Obtain the address to map to. we verify (or select) it and ensure
1286 * that it represents a valid section of the address space.
1288 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1289 if (IS_ERR_VALUE(addr))
1292 if (flags & MAP_FIXED_NOREPLACE) {
1293 if (find_vma_intersection(mm, addr, addr + len))
1297 if (prot == PROT_EXEC) {
1298 pkey = execute_only_pkey(mm);
1303 /* Do simple checking here so the lower-level routines won't have
1304 * to. we assume access permissions have been handled by the open
1305 * of the memory object, so we don't do any here.
1307 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1308 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1310 if (flags & MAP_LOCKED)
1311 if (!can_do_mlock())
1314 if (mlock_future_check(mm, vm_flags, len))
1318 struct inode *inode = file_inode(file);
1319 unsigned long flags_mask;
1321 if (!file_mmap_ok(file, inode, pgoff, len))
1324 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1326 switch (flags & MAP_TYPE) {
1329 * Force use of MAP_SHARED_VALIDATE with non-legacy
1330 * flags. E.g. MAP_SYNC is dangerous to use with
1331 * MAP_SHARED as you don't know which consistency model
1332 * you will get. We silently ignore unsupported flags
1333 * with MAP_SHARED to preserve backward compatibility.
1335 flags &= LEGACY_MAP_MASK;
1337 case MAP_SHARED_VALIDATE:
1338 if (flags & ~flags_mask)
1340 if (prot & PROT_WRITE) {
1341 if (!(file->f_mode & FMODE_WRITE))
1343 if (IS_SWAPFILE(file->f_mapping->host))
1348 * Make sure we don't allow writing to an append-only
1351 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1354 vm_flags |= VM_SHARED | VM_MAYSHARE;
1355 if (!(file->f_mode & FMODE_WRITE))
1356 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1359 if (!(file->f_mode & FMODE_READ))
1361 if (path_noexec(&file->f_path)) {
1362 if (vm_flags & VM_EXEC)
1364 vm_flags &= ~VM_MAYEXEC;
1367 if (!file->f_op->mmap)
1369 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1377 switch (flags & MAP_TYPE) {
1379 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1385 vm_flags |= VM_SHARED | VM_MAYSHARE;
1389 * Set pgoff according to addr for anon_vma.
1391 pgoff = addr >> PAGE_SHIFT;
1399 * Set 'VM_NORESERVE' if we should not account for the
1400 * memory use of this mapping.
1402 if (flags & MAP_NORESERVE) {
1403 /* We honor MAP_NORESERVE if allowed to overcommit */
1404 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1405 vm_flags |= VM_NORESERVE;
1407 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1408 if (file && is_file_hugepages(file))
1409 vm_flags |= VM_NORESERVE;
1412 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1413 if (!IS_ERR_VALUE(addr) &&
1414 ((vm_flags & VM_LOCKED) ||
1415 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1420 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1421 unsigned long prot, unsigned long flags,
1422 unsigned long fd, unsigned long pgoff)
1424 struct file *file = NULL;
1425 unsigned long retval;
1427 if (!(flags & MAP_ANONYMOUS)) {
1428 audit_mmap_fd(fd, flags);
1432 if (is_file_hugepages(file)) {
1433 len = ALIGN(len, huge_page_size(hstate_file(file)));
1434 } else if (unlikely(flags & MAP_HUGETLB)) {
1438 } else if (flags & MAP_HUGETLB) {
1441 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1445 len = ALIGN(len, huge_page_size(hs));
1447 * VM_NORESERVE is used because the reservations will be
1448 * taken when vm_ops->mmap() is called
1450 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1452 HUGETLB_ANONHUGE_INODE,
1453 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1455 return PTR_ERR(file);
1458 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1465 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1466 unsigned long, prot, unsigned long, flags,
1467 unsigned long, fd, unsigned long, pgoff)
1469 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1472 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1473 struct mmap_arg_struct {
1477 unsigned long flags;
1479 unsigned long offset;
1482 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1484 struct mmap_arg_struct a;
1486 if (copy_from_user(&a, arg, sizeof(a)))
1488 if (offset_in_page(a.offset))
1491 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1492 a.offset >> PAGE_SHIFT);
1494 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1497 * Some shared mappings will want the pages marked read-only
1498 * to track write events. If so, we'll downgrade vm_page_prot
1499 * to the private version (using protection_map[] without the
1502 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1504 vm_flags_t vm_flags = vma->vm_flags;
1505 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1507 /* If it was private or non-writable, the write bit is already clear */
1508 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1511 /* The backer wishes to know when pages are first written to? */
1512 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1515 /* The open routine did something to the protections that pgprot_modify
1516 * won't preserve? */
1517 if (pgprot_val(vm_page_prot) !=
1518 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1522 * Do we need to track softdirty? hugetlb does not support softdirty
1525 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1528 /* Do we need write faults for uffd-wp tracking? */
1529 if (userfaultfd_wp(vma))
1532 /* Specialty mapping? */
1533 if (vm_flags & VM_PFNMAP)
1536 /* Can the mapping track the dirty pages? */
1537 return vma->vm_file && vma->vm_file->f_mapping &&
1538 mapping_can_writeback(vma->vm_file->f_mapping);
1542 * We account for memory if it's a private writeable mapping,
1543 * not hugepages and VM_NORESERVE wasn't set.
1545 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1548 * hugetlb has its own accounting separate from the core VM
1549 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1551 if (file && is_file_hugepages(file))
1554 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1558 * unmapped_area() - Find an area between the low_limit and the high_limit with
1559 * the correct alignment and offset, all from @info. Note: current->mm is used
1562 * @info: The unmapped area information including the range (low_limit -
1563 * hight_limit), the alignment offset and mask.
1565 * Return: A memory address or -ENOMEM.
1567 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1569 unsigned long length, gap, low_limit;
1570 struct vm_area_struct *tmp;
1572 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1574 /* Adjust search length to account for worst case alignment overhead */
1575 length = info->length + info->align_mask;
1576 if (length < info->length)
1579 low_limit = info->low_limit;
1581 if (mas_empty_area(&mas, low_limit, info->high_limit - 1, length))
1585 gap += (info->align_offset - gap) & info->align_mask;
1586 tmp = mas_next(&mas, ULONG_MAX);
1587 if (tmp && (tmp->vm_flags & VM_GROWSDOWN)) { /* Avoid prev check if possible */
1588 if (vm_start_gap(tmp) < gap + length - 1) {
1589 low_limit = tmp->vm_end;
1594 tmp = mas_prev(&mas, 0);
1595 if (tmp && vm_end_gap(tmp) > gap) {
1596 low_limit = vm_end_gap(tmp);
1606 * unmapped_area_topdown() - Find an area between the low_limit and the
1607 * high_limit with * the correct alignment and offset at the highest available
1608 * address, all from @info. Note: current->mm is used for the search.
1610 * @info: The unmapped area information including the range (low_limit -
1611 * hight_limit), the alignment offset and mask.
1613 * Return: A memory address or -ENOMEM.
1615 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1617 unsigned long length, gap, high_limit, gap_end;
1618 struct vm_area_struct *tmp;
1620 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1621 /* Adjust search length to account for worst case alignment overhead */
1622 length = info->length + info->align_mask;
1623 if (length < info->length)
1626 high_limit = info->high_limit;
1628 if (mas_empty_area_rev(&mas, info->low_limit, high_limit - 1,
1632 gap = mas.last + 1 - info->length;
1633 gap -= (gap - info->align_offset) & info->align_mask;
1635 tmp = mas_next(&mas, ULONG_MAX);
1636 if (tmp && (tmp->vm_flags & VM_GROWSDOWN)) { /* Avoid prev check if possible */
1637 if (vm_start_gap(tmp) <= gap_end) {
1638 high_limit = vm_start_gap(tmp);
1643 tmp = mas_prev(&mas, 0);
1644 if (tmp && vm_end_gap(tmp) > gap) {
1645 high_limit = tmp->vm_start;
1655 * Search for an unmapped address range.
1657 * We are looking for a range that:
1658 * - does not intersect with any VMA;
1659 * - is contained within the [low_limit, high_limit) interval;
1660 * - is at least the desired size.
1661 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1663 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1667 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1668 addr = unmapped_area_topdown(info);
1670 addr = unmapped_area(info);
1672 trace_vm_unmapped_area(addr, info);
1676 /* Get an address range which is currently unmapped.
1677 * For shmat() with addr=0.
1679 * Ugly calling convention alert:
1680 * Return value with the low bits set means error value,
1682 * if (ret & ~PAGE_MASK)
1685 * This function "knows" that -ENOMEM has the bits set.
1688 generic_get_unmapped_area(struct file *filp, unsigned long addr,
1689 unsigned long len, unsigned long pgoff,
1690 unsigned long flags)
1692 struct mm_struct *mm = current->mm;
1693 struct vm_area_struct *vma, *prev;
1694 struct vm_unmapped_area_info info;
1695 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1697 if (len > mmap_end - mmap_min_addr)
1700 if (flags & MAP_FIXED)
1704 addr = PAGE_ALIGN(addr);
1705 vma = find_vma_prev(mm, addr, &prev);
1706 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1707 (!vma || addr + len <= vm_start_gap(vma)) &&
1708 (!prev || addr >= vm_end_gap(prev)))
1714 info.low_limit = mm->mmap_base;
1715 info.high_limit = mmap_end;
1716 info.align_mask = 0;
1717 info.align_offset = 0;
1718 return vm_unmapped_area(&info);
1721 #ifndef HAVE_ARCH_UNMAPPED_AREA
1723 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1724 unsigned long len, unsigned long pgoff,
1725 unsigned long flags)
1727 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1732 * This mmap-allocator allocates new areas top-down from below the
1733 * stack's low limit (the base):
1736 generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1737 unsigned long len, unsigned long pgoff,
1738 unsigned long flags)
1740 struct vm_area_struct *vma, *prev;
1741 struct mm_struct *mm = current->mm;
1742 struct vm_unmapped_area_info info;
1743 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1745 /* requested length too big for entire address space */
1746 if (len > mmap_end - mmap_min_addr)
1749 if (flags & MAP_FIXED)
1752 /* requesting a specific address */
1754 addr = PAGE_ALIGN(addr);
1755 vma = find_vma_prev(mm, addr, &prev);
1756 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1757 (!vma || addr + len <= vm_start_gap(vma)) &&
1758 (!prev || addr >= vm_end_gap(prev)))
1762 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1764 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1765 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1766 info.align_mask = 0;
1767 info.align_offset = 0;
1768 addr = vm_unmapped_area(&info);
1771 * A failed mmap() very likely causes application failure,
1772 * so fall back to the bottom-up function here. This scenario
1773 * can happen with large stack limits and large mmap()
1776 if (offset_in_page(addr)) {
1777 VM_BUG_ON(addr != -ENOMEM);
1779 info.low_limit = TASK_UNMAPPED_BASE;
1780 info.high_limit = mmap_end;
1781 addr = vm_unmapped_area(&info);
1787 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1789 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1790 unsigned long len, unsigned long pgoff,
1791 unsigned long flags)
1793 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1798 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1799 unsigned long pgoff, unsigned long flags)
1801 unsigned long (*get_area)(struct file *, unsigned long,
1802 unsigned long, unsigned long, unsigned long);
1804 unsigned long error = arch_mmap_check(addr, len, flags);
1808 /* Careful about overflows.. */
1809 if (len > TASK_SIZE)
1812 get_area = current->mm->get_unmapped_area;
1814 if (file->f_op->get_unmapped_area)
1815 get_area = file->f_op->get_unmapped_area;
1816 } else if (flags & MAP_SHARED) {
1818 * mmap_region() will call shmem_zero_setup() to create a file,
1819 * so use shmem's get_unmapped_area in case it can be huge.
1820 * do_mmap() will clear pgoff, so match alignment.
1823 get_area = shmem_get_unmapped_area;
1826 addr = get_area(file, addr, len, pgoff, flags);
1827 if (IS_ERR_VALUE(addr))
1830 if (addr > TASK_SIZE - len)
1832 if (offset_in_page(addr))
1835 error = security_mmap_addr(addr);
1836 return error ? error : addr;
1839 EXPORT_SYMBOL(get_unmapped_area);
1842 * find_vma_intersection() - Look up the first VMA which intersects the interval
1843 * @mm: The process address space.
1844 * @start_addr: The inclusive start user address.
1845 * @end_addr: The exclusive end user address.
1847 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1848 * start_addr < end_addr.
1850 struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1851 unsigned long start_addr,
1852 unsigned long end_addr)
1854 unsigned long index = start_addr;
1856 mmap_assert_locked(mm);
1857 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1859 EXPORT_SYMBOL(find_vma_intersection);
1862 * find_vma() - Find the VMA for a given address, or the next VMA.
1863 * @mm: The mm_struct to check
1864 * @addr: The address
1866 * Returns: The VMA associated with addr, or the next VMA.
1867 * May return %NULL in the case of no VMA at addr or above.
1869 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1871 unsigned long index = addr;
1873 mmap_assert_locked(mm);
1874 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1876 EXPORT_SYMBOL(find_vma);
1879 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1880 * set %pprev to the previous VMA, if any.
1881 * @mm: The mm_struct to check
1882 * @addr: The address
1883 * @pprev: The pointer to set to the previous VMA
1885 * Note that RCU lock is missing here since the external mmap_lock() is used
1888 * Returns: The VMA associated with @addr, or the next vma.
1889 * May return %NULL in the case of no vma at addr or above.
1891 struct vm_area_struct *
1892 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1893 struct vm_area_struct **pprev)
1895 struct vm_area_struct *vma;
1896 MA_STATE(mas, &mm->mm_mt, addr, addr);
1898 vma = mas_walk(&mas);
1899 *pprev = mas_prev(&mas, 0);
1901 vma = mas_next(&mas, ULONG_MAX);
1906 * Verify that the stack growth is acceptable and
1907 * update accounting. This is shared with both the
1908 * grow-up and grow-down cases.
1910 static int acct_stack_growth(struct vm_area_struct *vma,
1911 unsigned long size, unsigned long grow)
1913 struct mm_struct *mm = vma->vm_mm;
1914 unsigned long new_start;
1916 /* address space limit tests */
1917 if (!may_expand_vm(mm, vma->vm_flags, grow))
1920 /* Stack limit test */
1921 if (size > rlimit(RLIMIT_STACK))
1924 /* mlock limit tests */
1925 if (mlock_future_check(mm, vma->vm_flags, grow << PAGE_SHIFT))
1928 /* Check to ensure the stack will not grow into a hugetlb-only region */
1929 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1931 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1935 * Overcommit.. This must be the final test, as it will
1936 * update security statistics.
1938 if (security_vm_enough_memory_mm(mm, grow))
1944 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1946 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1947 * vma is the last one with address > vma->vm_end. Have to extend vma.
1949 static int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1951 struct mm_struct *mm = vma->vm_mm;
1952 struct vm_area_struct *next;
1953 unsigned long gap_addr;
1955 MA_STATE(mas, &mm->mm_mt, 0, 0);
1957 if (!(vma->vm_flags & VM_GROWSUP))
1960 /* Guard against exceeding limits of the address space. */
1961 address &= PAGE_MASK;
1962 if (address >= (TASK_SIZE & PAGE_MASK))
1964 address += PAGE_SIZE;
1966 /* Enforce stack_guard_gap */
1967 gap_addr = address + stack_guard_gap;
1969 /* Guard against overflow */
1970 if (gap_addr < address || gap_addr > TASK_SIZE)
1971 gap_addr = TASK_SIZE;
1973 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1974 if (next && vma_is_accessible(next)) {
1975 if (!(next->vm_flags & VM_GROWSUP))
1977 /* Check that both stack segments have the same anon_vma? */
1980 if (mas_preallocate(&mas, vma, GFP_KERNEL))
1983 /* We must make sure the anon_vma is allocated. */
1984 if (unlikely(anon_vma_prepare(vma))) {
1990 * vma->vm_start/vm_end cannot change under us because the caller
1991 * is required to hold the mmap_lock in read mode. We need the
1992 * anon_vma lock to serialize against concurrent expand_stacks.
1994 anon_vma_lock_write(vma->anon_vma);
1996 /* Somebody else might have raced and expanded it already */
1997 if (address > vma->vm_end) {
1998 unsigned long size, grow;
2000 size = address - vma->vm_start;
2001 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2004 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2005 error = acct_stack_growth(vma, size, grow);
2008 * We only hold a shared mmap_lock lock here, so
2009 * we need to protect against concurrent vma
2010 * expansions. anon_vma_lock_write() doesn't
2011 * help here, as we don't guarantee that all
2012 * growable vmas in a mm share the same root
2013 * anon vma. So, we reuse mm->page_table_lock
2014 * to guard against concurrent vma expansions.
2016 spin_lock(&mm->page_table_lock);
2017 if (vma->vm_flags & VM_LOCKED)
2018 mm->locked_vm += grow;
2019 vm_stat_account(mm, vma->vm_flags, grow);
2020 anon_vma_interval_tree_pre_update_vma(vma);
2021 vma->vm_end = address;
2022 /* Overwrite old entry in mtree. */
2023 vma_mas_store(vma, &mas);
2024 anon_vma_interval_tree_post_update_vma(vma);
2025 spin_unlock(&mm->page_table_lock);
2027 perf_event_mmap(vma);
2031 anon_vma_unlock_write(vma->anon_vma);
2032 khugepaged_enter_vma(vma, vma->vm_flags);
2036 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2039 * vma is the first one with address < vma->vm_start. Have to extend vma.
2040 * mmap_lock held for writing.
2042 int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2044 struct mm_struct *mm = vma->vm_mm;
2045 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2046 struct vm_area_struct *prev;
2049 if (!(vma->vm_flags & VM_GROWSDOWN))
2052 address &= PAGE_MASK;
2053 if (address < mmap_min_addr || address < FIRST_USER_ADDRESS)
2056 /* Enforce stack_guard_gap */
2057 prev = mas_prev(&mas, 0);
2058 /* Check that both stack segments have the same anon_vma? */
2060 if (!(prev->vm_flags & VM_GROWSDOWN) &&
2061 vma_is_accessible(prev) &&
2062 (address - prev->vm_end < stack_guard_gap))
2066 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2069 /* We must make sure the anon_vma is allocated. */
2070 if (unlikely(anon_vma_prepare(vma))) {
2076 * vma->vm_start/vm_end cannot change under us because the caller
2077 * is required to hold the mmap_lock in read mode. We need the
2078 * anon_vma lock to serialize against concurrent expand_stacks.
2080 anon_vma_lock_write(vma->anon_vma);
2082 /* Somebody else might have raced and expanded it already */
2083 if (address < vma->vm_start) {
2084 unsigned long size, grow;
2086 size = vma->vm_end - address;
2087 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2090 if (grow <= vma->vm_pgoff) {
2091 error = acct_stack_growth(vma, size, grow);
2094 * We only hold a shared mmap_lock lock here, so
2095 * we need to protect against concurrent vma
2096 * expansions. anon_vma_lock_write() doesn't
2097 * help here, as we don't guarantee that all
2098 * growable vmas in a mm share the same root
2099 * anon vma. So, we reuse mm->page_table_lock
2100 * to guard against concurrent vma expansions.
2102 spin_lock(&mm->page_table_lock);
2103 if (vma->vm_flags & VM_LOCKED)
2104 mm->locked_vm += grow;
2105 vm_stat_account(mm, vma->vm_flags, grow);
2106 anon_vma_interval_tree_pre_update_vma(vma);
2107 vma->vm_start = address;
2108 vma->vm_pgoff -= grow;
2109 /* Overwrite old entry in mtree. */
2110 vma_mas_store(vma, &mas);
2111 anon_vma_interval_tree_post_update_vma(vma);
2112 spin_unlock(&mm->page_table_lock);
2114 perf_event_mmap(vma);
2118 anon_vma_unlock_write(vma->anon_vma);
2119 khugepaged_enter_vma(vma, vma->vm_flags);
2124 /* enforced gap between the expanding stack and other mappings. */
2125 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2127 static int __init cmdline_parse_stack_guard_gap(char *p)
2132 val = simple_strtoul(p, &endptr, 10);
2134 stack_guard_gap = val << PAGE_SHIFT;
2138 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2140 #ifdef CONFIG_STACK_GROWSUP
2141 int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2143 return expand_upwards(vma, address);
2146 struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2148 struct vm_area_struct *vma, *prev;
2151 vma = find_vma_prev(mm, addr, &prev);
2152 if (vma && (vma->vm_start <= addr))
2156 if (expand_stack_locked(prev, addr))
2158 if (prev->vm_flags & VM_LOCKED)
2159 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2163 int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2165 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN)))
2167 return expand_downwards(vma, address);
2170 struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2172 struct vm_area_struct *vma;
2173 unsigned long start;
2176 vma = find_vma(mm, addr);
2179 if (vma->vm_start <= addr)
2181 start = vma->vm_start;
2182 if (expand_stack_locked(vma, addr))
2184 if (vma->vm_flags & VM_LOCKED)
2185 populate_vma_page_range(vma, addr, start, NULL);
2191 * IA64 has some horrid mapping rules: it can expand both up and down,
2192 * but with various special rules.
2194 * We'll get rid of this architecture eventually, so the ugliness is
2198 static inline bool vma_expand_ok(struct vm_area_struct *vma, unsigned long addr)
2200 return REGION_NUMBER(addr) == REGION_NUMBER(vma->vm_start) &&
2201 REGION_OFFSET(addr) < RGN_MAP_LIMIT;
2205 * IA64 stacks grow down, but there's a special register backing store
2206 * that can grow up. Only sequentially, though, so the new address must
2209 static inline int vma_expand_up(struct vm_area_struct *vma, unsigned long addr)
2211 if (!vma_expand_ok(vma, addr))
2213 if (vma->vm_end != (addr & PAGE_MASK))
2215 return expand_upwards(vma, addr);
2218 static inline bool vma_expand_down(struct vm_area_struct *vma, unsigned long addr)
2220 if (!vma_expand_ok(vma, addr))
2222 return expand_downwards(vma, addr);
2225 #elif defined(CONFIG_STACK_GROWSUP)
2227 #define vma_expand_up(vma,addr) expand_upwards(vma, addr)
2228 #define vma_expand_down(vma, addr) (-EFAULT)
2232 #define vma_expand_up(vma,addr) (-EFAULT)
2233 #define vma_expand_down(vma, addr) expand_downwards(vma, addr)
2238 * expand_stack(): legacy interface for page faulting. Don't use unless
2241 * This is called with the mm locked for reading, drops the lock, takes
2242 * the lock for writing, tries to look up a vma again, expands it if
2243 * necessary, and downgrades the lock to reading again.
2245 * If no vma is found or it can't be expanded, it returns NULL and has
2248 struct vm_area_struct *expand_stack(struct mm_struct *mm, unsigned long addr)
2250 struct vm_area_struct *vma, *prev;
2252 mmap_read_unlock(mm);
2253 if (mmap_write_lock_killable(mm))
2256 vma = find_vma_prev(mm, addr, &prev);
2257 if (vma && vma->vm_start <= addr)
2260 if (prev && !vma_expand_up(prev, addr)) {
2265 if (vma && !vma_expand_down(vma, addr))
2268 mmap_write_unlock(mm);
2272 mmap_write_downgrade(mm);
2277 * Ok - we have the memory areas we should free on a maple tree so release them,
2278 * and do the vma updates.
2280 * Called with the mm semaphore held.
2282 static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2284 unsigned long nr_accounted = 0;
2285 struct vm_area_struct *vma;
2287 /* Update high watermark before we lower total_vm */
2288 update_hiwater_vm(mm);
2289 mas_for_each(mas, vma, ULONG_MAX) {
2290 long nrpages = vma_pages(vma);
2292 if (vma->vm_flags & VM_ACCOUNT)
2293 nr_accounted += nrpages;
2294 vm_stat_account(mm, vma->vm_flags, -nrpages);
2297 vm_unacct_memory(nr_accounted);
2302 * Get rid of page table information in the indicated region.
2304 * Called with the mm semaphore held.
2306 static void unmap_region(struct mm_struct *mm, struct maple_tree *mt,
2307 struct vm_area_struct *vma, struct vm_area_struct *prev,
2308 struct vm_area_struct *next,
2309 unsigned long start, unsigned long end)
2311 struct mmu_gather tlb;
2314 tlb_gather_mmu(&tlb, mm);
2315 update_hiwater_rss(mm);
2316 unmap_vmas(&tlb, mt, vma, start, end);
2317 free_pgtables(&tlb, mt, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2318 next ? next->vm_start : USER_PGTABLES_CEILING);
2319 tlb_finish_mmu(&tlb);
2323 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2324 * has already been checked or doesn't make sense to fail.
2326 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2327 unsigned long addr, int new_below)
2329 struct vm_area_struct *new;
2333 if (vma->vm_ops && vma->vm_ops->may_split) {
2334 err = vma->vm_ops->may_split(vma, addr);
2339 new = vm_area_dup(vma);
2346 new->vm_start = addr;
2347 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2350 err = vma_dup_policy(vma, new);
2354 err = anon_vma_clone(new, vma);
2359 get_file(new->vm_file);
2361 if (new->vm_ops && new->vm_ops->open)
2362 new->vm_ops->open(new);
2365 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2366 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2368 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2374 /* Avoid vm accounting in close() operation */
2375 new->vm_start = new->vm_end;
2377 /* Clean everything up if vma_adjust failed. */
2378 if (new->vm_ops && new->vm_ops->close)
2379 new->vm_ops->close(new);
2382 unlink_anon_vmas(new);
2384 mpol_put(vma_policy(new));
2392 * Split a vma into two pieces at address 'addr', a new vma is allocated
2393 * either for the first part or the tail.
2395 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2396 unsigned long addr, int new_below)
2398 if (mm->map_count >= sysctl_max_map_count)
2401 return __split_vma(mm, vma, addr, new_below);
2405 * do_mas_align_munmap() - munmap the aligned region from @start to @end.
2406 * @mas: The maple_state, ideally set up to alter the correct tree location.
2407 * @vma: The starting vm_area_struct
2408 * @mm: The mm_struct
2409 * @start: The aligned start address to munmap.
2410 * @end: The aligned end address to munmap.
2411 * @uf: The userfaultfd list_head
2412 * @downgrade: Set to true to attempt a write downgrade of the mmap_sem
2414 * If @downgrade is true, check return code for potential release of the lock.
2417 do_mas_align_munmap(struct ma_state *mas, struct vm_area_struct *vma,
2418 struct mm_struct *mm, unsigned long start,
2419 unsigned long end, struct list_head *uf, bool downgrade)
2421 struct vm_area_struct *prev, *next = NULL;
2422 struct maple_tree mt_detach;
2424 int error = -ENOMEM;
2425 unsigned long locked_vm = 0;
2426 MA_STATE(mas_detach, &mt_detach, 0, 0);
2427 mt_init_flags(&mt_detach, mas->tree->ma_flags & MT_FLAGS_LOCK_MASK);
2428 mt_set_external_lock(&mt_detach, &mm->mmap_lock);
2430 if (mas_preallocate(mas, vma, GFP_KERNEL))
2433 mas->last = end - 1;
2435 * If we need to split any vma, do it now to save pain later.
2437 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2438 * unmapped vm_area_struct will remain in use: so lower split_vma
2439 * places tmp vma above, and higher split_vma places tmp vma below.
2442 /* Does it split the first one? */
2443 if (start > vma->vm_start) {
2446 * Make sure that map_count on return from munmap() will
2447 * not exceed its limit; but let map_count go just above
2448 * its limit temporarily, to help free resources as expected.
2450 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2451 goto map_count_exceeded;
2454 * mas_pause() is not needed since mas->index needs to be set
2455 * differently than vma->vm_end anyways.
2457 error = __split_vma(mm, vma, start, 0);
2459 goto start_split_failed;
2461 mas_set(mas, start);
2462 vma = mas_walk(mas);
2465 prev = mas_prev(mas, 0);
2466 if (unlikely((!prev)))
2467 mas_set(mas, start);
2470 * Detach a range of VMAs from the mm. Using next as a temp variable as
2471 * it is always overwritten.
2473 mas_for_each(mas, next, end - 1) {
2474 /* Does it split the end? */
2475 if (next->vm_end > end) {
2476 struct vm_area_struct *split;
2478 error = __split_vma(mm, next, end, 1);
2480 goto end_split_failed;
2483 split = mas_prev(mas, 0);
2484 mas_set_range(&mas_detach, split->vm_start, split->vm_end - 1);
2485 error = mas_store_gfp(&mas_detach, split, GFP_KERNEL);
2487 goto munmap_gather_failed;
2488 if (split->vm_flags & VM_LOCKED)
2489 locked_vm += vma_pages(split);
2496 mas_set_range(&mas_detach, next->vm_start, next->vm_end - 1);
2497 error = mas_store_gfp(&mas_detach, next, GFP_KERNEL);
2499 goto munmap_gather_failed;
2500 if (next->vm_flags & VM_LOCKED)
2501 locked_vm += vma_pages(next);
2504 #ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2505 BUG_ON(next->vm_start < start);
2506 BUG_ON(next->vm_start > end);
2511 next = mas_next(mas, ULONG_MAX);
2515 * If userfaultfd_unmap_prep returns an error the vmas
2516 * will remain split, but userland will get a
2517 * highly unexpected error anyway. This is no
2518 * different than the case where the first of the two
2519 * __split_vma fails, but we don't undo the first
2520 * split, despite we could. This is unlikely enough
2521 * failure that it's not worth optimizing it for.
2523 error = userfaultfd_unmap_prep(mm, start, end, uf);
2526 goto userfaultfd_error;
2529 /* Point of no return */
2530 mas_set_range(mas, start, end - 1);
2531 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2532 /* Make sure no VMAs are about to be lost. */
2534 MA_STATE(test, &mt_detach, start, end - 1);
2535 struct vm_area_struct *vma_mas, *vma_test;
2539 vma_test = mas_find(&test, end - 1);
2540 mas_for_each(mas, vma_mas, end - 1) {
2541 BUG_ON(vma_mas != vma_test);
2543 vma_test = mas_next(&test, end - 1);
2546 BUG_ON(count != test_count);
2547 mas_set_range(mas, start, end - 1);
2550 /* Point of no return */
2551 mas_store_prealloc(mas, NULL);
2553 mm->locked_vm -= locked_vm;
2554 mm->map_count -= count;
2556 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2557 * VM_GROWSUP VMA. Such VMAs can change their size under
2558 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2561 if (next && (next->vm_flags & VM_GROWSDOWN))
2563 else if (prev && (prev->vm_flags & VM_GROWSUP))
2566 mmap_write_downgrade(mm);
2569 unmap_region(mm, &mt_detach, vma, prev, next, start, end);
2570 /* Statistics and freeing VMAs */
2571 mas_set(&mas_detach, start);
2572 remove_mt(mm, &mas_detach);
2573 __mt_destroy(&mt_detach);
2577 return downgrade ? 1 : 0;
2580 munmap_gather_failed:
2582 __mt_destroy(&mt_detach);
2590 * do_mas_munmap() - munmap a given range.
2591 * @mas: The maple state
2592 * @mm: The mm_struct
2593 * @start: The start address to munmap
2594 * @len: The length of the range to munmap
2595 * @uf: The userfaultfd list_head
2596 * @downgrade: set to true if the user wants to attempt to write_downgrade the
2599 * This function takes a @mas that is either pointing to the previous VMA or set
2600 * to MA_START and sets it up to remove the mapping(s). The @len will be
2601 * aligned and any arch_unmap work will be preformed.
2603 * Returns: -EINVAL on failure, 1 on success and unlock, 0 otherwise.
2605 int do_mas_munmap(struct ma_state *mas, struct mm_struct *mm,
2606 unsigned long start, size_t len, struct list_head *uf,
2610 struct vm_area_struct *vma;
2612 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2615 end = start + PAGE_ALIGN(len);
2619 /* arch_unmap() might do unmaps itself. */
2620 arch_unmap(mm, start, end);
2622 /* Find the first overlapping VMA */
2623 vma = mas_find(mas, end - 1);
2627 return do_mas_align_munmap(mas, vma, mm, start, end, uf, downgrade);
2630 /* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2631 * @mm: The mm_struct
2632 * @start: The start address to munmap
2633 * @len: The length to be munmapped.
2634 * @uf: The userfaultfd list_head
2636 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2637 struct list_head *uf)
2639 MA_STATE(mas, &mm->mm_mt, start, start);
2641 return do_mas_munmap(&mas, mm, start, len, uf, false);
2644 unsigned long mmap_region(struct file *file, unsigned long addr,
2645 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2646 struct list_head *uf)
2648 struct mm_struct *mm = current->mm;
2649 struct vm_area_struct *vma = NULL;
2650 struct vm_area_struct *next, *prev, *merge;
2651 pgoff_t pglen = len >> PAGE_SHIFT;
2652 unsigned long charged = 0;
2653 unsigned long end = addr + len;
2654 unsigned long merge_start = addr, merge_end = end;
2657 MA_STATE(mas, &mm->mm_mt, addr, end - 1);
2659 /* Check against address space limit. */
2660 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2661 unsigned long nr_pages;
2664 * MAP_FIXED may remove pages of mappings that intersects with
2665 * requested mapping. Account for the pages it would unmap.
2667 nr_pages = count_vma_pages_range(mm, addr, end);
2669 if (!may_expand_vm(mm, vm_flags,
2670 (len >> PAGE_SHIFT) - nr_pages))
2674 /* Unmap any existing mapping in the area */
2675 if (do_mas_munmap(&mas, mm, addr, len, uf, false))
2679 * Private writable mapping: check memory availability
2681 if (accountable_mapping(file, vm_flags)) {
2682 charged = len >> PAGE_SHIFT;
2683 if (security_vm_enough_memory_mm(mm, charged))
2685 vm_flags |= VM_ACCOUNT;
2688 next = mas_next(&mas, ULONG_MAX);
2689 prev = mas_prev(&mas, 0);
2690 if (vm_flags & VM_SPECIAL)
2693 /* Attempt to expand an old mapping */
2695 if (next && next->vm_start == end && !vma_policy(next) &&
2696 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2697 NULL_VM_UFFD_CTX, NULL)) {
2698 merge_end = next->vm_end;
2700 vm_pgoff = next->vm_pgoff - pglen;
2704 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2705 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2706 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2707 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2708 NULL_VM_UFFD_CTX, NULL))) {
2709 merge_start = prev->vm_start;
2711 vm_pgoff = prev->vm_pgoff;
2715 /* Actually expand, if possible */
2717 !vma_expand(&mas, vma, merge_start, merge_end, vm_pgoff, next)) {
2718 khugepaged_enter_vma(vma, vm_flags);
2726 * Determine the object being mapped and call the appropriate
2727 * specific mapper. the address has already been validated, but
2728 * not unmapped, but the maps are removed from the list.
2730 vma = vm_area_alloc(mm);
2736 vma->vm_start = addr;
2738 vma->vm_flags = vm_flags;
2739 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2740 vma->vm_pgoff = pgoff;
2743 if (vm_flags & VM_SHARED) {
2744 error = mapping_map_writable(file->f_mapping);
2749 vma->vm_file = get_file(file);
2750 error = call_mmap(file, vma);
2752 goto unmap_and_free_vma;
2755 * Expansion is handled above, merging is handled below.
2756 * Drivers should not alter the address of the VMA.
2758 if (WARN_ON((addr != vma->vm_start))) {
2760 goto close_and_free_vma;
2765 * If vm_flags changed after call_mmap(), we should try merge
2766 * vma again as we may succeed this time.
2768 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2769 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
2770 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
2773 * ->mmap() can change vma->vm_file and fput
2774 * the original file. So fput the vma->vm_file
2775 * here or we would add an extra fput for file
2776 * and cause general protection fault
2782 /* Update vm_flags to pick up the change. */
2783 vm_flags = vma->vm_flags;
2784 goto unmap_writable;
2788 vm_flags = vma->vm_flags;
2789 } else if (vm_flags & VM_SHARED) {
2790 error = shmem_zero_setup(vma);
2794 vma_set_anonymous(vma);
2797 /* Allow architectures to sanity-check the vm_flags */
2798 if (!arch_validate_flags(vma->vm_flags)) {
2801 goto close_and_free_vma;
2802 else if (vma->vm_file)
2803 goto unmap_and_free_vma;
2808 if (mas_preallocate(&mas, vma, GFP_KERNEL)) {
2811 goto close_and_free_vma;
2812 else if (vma->vm_file)
2813 goto unmap_and_free_vma;
2819 i_mmap_lock_write(vma->vm_file->f_mapping);
2821 vma_mas_store(vma, &mas);
2824 if (vma->vm_flags & VM_SHARED)
2825 mapping_allow_writable(vma->vm_file->f_mapping);
2827 flush_dcache_mmap_lock(vma->vm_file->f_mapping);
2828 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
2829 flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
2830 i_mmap_unlock_write(vma->vm_file->f_mapping);
2834 * vma_merge() calls khugepaged_enter_vma() either, the below
2835 * call covers the non-merge case.
2837 khugepaged_enter_vma(vma, vma->vm_flags);
2839 /* Once vma denies write, undo our temporary denial count */
2841 if (file && vm_flags & VM_SHARED)
2842 mapping_unmap_writable(file->f_mapping);
2843 file = vma->vm_file;
2845 perf_event_mmap(vma);
2847 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2848 if (vm_flags & VM_LOCKED) {
2849 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2850 is_vm_hugetlb_page(vma) ||
2851 vma == get_gate_vma(current->mm))
2852 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
2854 mm->locked_vm += (len >> PAGE_SHIFT);
2861 * New (or expanded) vma always get soft dirty status.
2862 * Otherwise user-space soft-dirty page tracker won't
2863 * be able to distinguish situation when vma area unmapped,
2864 * then new mapped in-place (which must be aimed as
2865 * a completely new data area).
2867 vma->vm_flags |= VM_SOFTDIRTY;
2869 vma_set_page_prot(vma);
2875 if (vma->vm_ops && vma->vm_ops->close)
2876 vma->vm_ops->close(vma);
2879 vma->vm_file = NULL;
2881 /* Undo any partial mapping done by a device driver. */
2882 unmap_region(mm, mas.tree, vma, prev, next, vma->vm_start, vma->vm_end);
2883 if (file && (vm_flags & VM_SHARED))
2884 mapping_unmap_writable(file->f_mapping);
2889 vm_unacct_memory(charged);
2894 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2897 struct mm_struct *mm = current->mm;
2899 MA_STATE(mas, &mm->mm_mt, start, start);
2901 if (mmap_write_lock_killable(mm))
2904 ret = do_mas_munmap(&mas, mm, start, len, &uf, downgrade);
2906 * Returning 1 indicates mmap_lock is downgraded.
2907 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2908 * it to 0 before return.
2911 mmap_read_unlock(mm);
2914 mmap_write_unlock(mm);
2916 userfaultfd_unmap_complete(mm, &uf);
2920 int vm_munmap(unsigned long start, size_t len)
2922 return __vm_munmap(start, len, false);
2924 EXPORT_SYMBOL(vm_munmap);
2926 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2928 addr = untagged_addr(addr);
2929 return __vm_munmap(addr, len, true);
2934 * Emulation of deprecated remap_file_pages() syscall.
2936 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2937 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2940 struct mm_struct *mm = current->mm;
2941 struct vm_area_struct *vma;
2942 unsigned long populate = 0;
2943 unsigned long ret = -EINVAL;
2946 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
2947 current->comm, current->pid);
2951 start = start & PAGE_MASK;
2952 size = size & PAGE_MASK;
2954 if (start + size <= start)
2957 /* Does pgoff wrap? */
2958 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2961 if (mmap_write_lock_killable(mm))
2964 vma = vma_lookup(mm, start);
2966 if (!vma || !(vma->vm_flags & VM_SHARED))
2969 if (start + size > vma->vm_end) {
2970 VMA_ITERATOR(vmi, mm, vma->vm_end);
2971 struct vm_area_struct *next, *prev = vma;
2973 for_each_vma_range(vmi, next, start + size) {
2974 /* hole between vmas ? */
2975 if (next->vm_start != prev->vm_end)
2978 if (next->vm_file != vma->vm_file)
2981 if (next->vm_flags != vma->vm_flags)
2984 if (start + size <= next->vm_end)
2994 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2995 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2996 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2998 flags &= MAP_NONBLOCK;
2999 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3000 if (vma->vm_flags & VM_LOCKED)
3001 flags |= MAP_LOCKED;
3003 file = get_file(vma->vm_file);
3004 ret = do_mmap(vma->vm_file, start, size,
3005 prot, flags, pgoff, &populate, NULL);
3008 mmap_write_unlock(mm);
3010 mm_populate(ret, populate);
3011 if (!IS_ERR_VALUE(ret))
3017 * brk_munmap() - Unmap a parital vma.
3018 * @mas: The maple tree state.
3019 * @vma: The vma to be modified
3020 * @newbrk: the start of the address to unmap
3021 * @oldbrk: The end of the address to unmap
3022 * @uf: The userfaultfd list_head
3024 * Returns: 1 on success.
3025 * unmaps a partial VMA mapping. Does not handle alignment, downgrades lock if
3028 static int do_brk_munmap(struct ma_state *mas, struct vm_area_struct *vma,
3029 unsigned long newbrk, unsigned long oldbrk,
3030 struct list_head *uf)
3032 struct mm_struct *mm = vma->vm_mm;
3035 arch_unmap(mm, newbrk, oldbrk);
3036 ret = do_mas_align_munmap(mas, vma, mm, newbrk, oldbrk, uf, true);
3042 * do_brk_flags() - Increase the brk vma if the flags match.
3043 * @mas: The maple tree state.
3044 * @addr: The start address
3045 * @len: The length of the increase
3047 * @flags: The VMA Flags
3049 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
3050 * do not match then create a new anonymous VMA. Eventually we may be able to
3051 * do some brk-specific accounting here.
3053 static int do_brk_flags(struct ma_state *mas, struct vm_area_struct *vma,
3054 unsigned long addr, unsigned long len, unsigned long flags)
3056 struct mm_struct *mm = current->mm;
3060 * Check against address space limits by the changed size
3061 * Note: This happens *after* clearing old mappings in some code paths.
3063 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3064 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3067 if (mm->map_count > sysctl_max_map_count)
3070 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3074 * Expand the existing vma if possible; Note that singular lists do not
3075 * occur after forking, so the expand will only happen on new VMAs.
3077 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
3078 can_vma_merge_after(vma, flags, NULL, NULL,
3079 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
3080 mas_set_range(mas, vma->vm_start, addr + len - 1);
3081 if (mas_preallocate(mas, vma, GFP_KERNEL))
3084 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
3085 if (vma->anon_vma) {
3086 anon_vma_lock_write(vma->anon_vma);
3087 anon_vma_interval_tree_pre_update_vma(vma);
3089 vma->vm_end = addr + len;
3090 vma->vm_flags |= VM_SOFTDIRTY;
3091 mas_store_prealloc(mas, vma);
3093 if (vma->anon_vma) {
3094 anon_vma_interval_tree_post_update_vma(vma);
3095 anon_vma_unlock_write(vma->anon_vma);
3097 khugepaged_enter_vma(vma, flags);
3101 /* create a vma struct for an anonymous mapping */
3102 vma = vm_area_alloc(mm);
3104 goto vma_alloc_fail;
3106 vma_set_anonymous(vma);
3107 vma->vm_start = addr;
3108 vma->vm_end = addr + len;
3109 vma->vm_pgoff = addr >> PAGE_SHIFT;
3110 vma->vm_flags = flags;
3111 vma->vm_page_prot = vm_get_page_prot(flags);
3112 mas_set_range(mas, vma->vm_start, addr + len - 1);
3113 if (mas_store_gfp(mas, vma, GFP_KERNEL))
3114 goto mas_store_fail;
3118 perf_event_mmap(vma);
3119 mm->total_vm += len >> PAGE_SHIFT;
3120 mm->data_vm += len >> PAGE_SHIFT;
3121 if (flags & VM_LOCKED)
3122 mm->locked_vm += (len >> PAGE_SHIFT);
3123 vma->vm_flags |= VM_SOFTDIRTY;
3130 vm_unacct_memory(len >> PAGE_SHIFT);
3134 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3136 struct mm_struct *mm = current->mm;
3137 struct vm_area_struct *vma = NULL;
3142 MA_STATE(mas, &mm->mm_mt, addr, addr);
3144 len = PAGE_ALIGN(request);
3150 if (mmap_write_lock_killable(mm))
3153 /* Until we need other flags, refuse anything except VM_EXEC. */
3154 if ((flags & (~VM_EXEC)) != 0)
3157 ret = check_brk_limits(addr, len);
3161 ret = do_mas_munmap(&mas, mm, addr, len, &uf, 0);
3165 vma = mas_prev(&mas, 0);
3166 ret = do_brk_flags(&mas, vma, addr, len, flags);
3167 populate = ((mm->def_flags & VM_LOCKED) != 0);
3168 mmap_write_unlock(mm);
3169 userfaultfd_unmap_complete(mm, &uf);
3170 if (populate && !ret)
3171 mm_populate(addr, len);
3176 mmap_write_unlock(mm);
3179 EXPORT_SYMBOL(vm_brk_flags);
3181 int vm_brk(unsigned long addr, unsigned long len)
3183 return vm_brk_flags(addr, len, 0);
3185 EXPORT_SYMBOL(vm_brk);
3187 /* Release all mmaps. */
3188 void exit_mmap(struct mm_struct *mm)
3190 struct mmu_gather tlb;
3191 struct vm_area_struct *vma;
3192 unsigned long nr_accounted = 0;
3193 MA_STATE(mas, &mm->mm_mt, 0, 0);
3196 /* mm's last user has gone, and its about to be pulled down */
3197 mmu_notifier_release(mm);
3202 vma = mas_find(&mas, ULONG_MAX);
3204 /* Can happen if dup_mmap() received an OOM */
3205 mmap_read_unlock(mm);
3211 tlb_gather_mmu_fullmm(&tlb, mm);
3212 /* update_hiwater_rss(mm) here? but nobody should be looking */
3213 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3214 unmap_vmas(&tlb, &mm->mm_mt, vma, 0, ULONG_MAX);
3215 mmap_read_unlock(mm);
3218 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3219 * because the memory has been already freed.
3221 set_bit(MMF_OOM_SKIP, &mm->flags);
3222 mmap_write_lock(mm);
3223 mt_clear_in_rcu(&mm->mm_mt);
3224 free_pgtables(&tlb, &mm->mm_mt, vma, FIRST_USER_ADDRESS,
3225 USER_PGTABLES_CEILING);
3226 tlb_finish_mmu(&tlb);
3229 * Walk the list again, actually closing and freeing it, with preemption
3230 * enabled, without holding any MM locks besides the unreachable
3234 if (vma->vm_flags & VM_ACCOUNT)
3235 nr_accounted += vma_pages(vma);
3239 } while ((vma = mas_find(&mas, ULONG_MAX)) != NULL);
3241 BUG_ON(count != mm->map_count);
3243 trace_exit_mmap(mm);
3244 __mt_destroy(&mm->mm_mt);
3245 mmap_write_unlock(mm);
3246 vm_unacct_memory(nr_accounted);
3249 /* Insert vm structure into process list sorted by address
3250 * and into the inode's i_mmap tree. If vm_file is non-NULL
3251 * then i_mmap_rwsem is taken here.
3253 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3255 unsigned long charged = vma_pages(vma);
3258 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3261 if ((vma->vm_flags & VM_ACCOUNT) &&
3262 security_vm_enough_memory_mm(mm, charged))
3266 * The vm_pgoff of a purely anonymous vma should be irrelevant
3267 * until its first write fault, when page's anon_vma and index
3268 * are set. But now set the vm_pgoff it will almost certainly
3269 * end up with (unless mremap moves it elsewhere before that
3270 * first wfault), so /proc/pid/maps tells a consistent story.
3272 * By setting it to reflect the virtual start address of the
3273 * vma, merges and splits can happen in a seamless way, just
3274 * using the existing file pgoff checks and manipulations.
3275 * Similarly in do_mmap and in do_brk_flags.
3277 if (vma_is_anonymous(vma)) {
3278 BUG_ON(vma->anon_vma);
3279 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3282 if (vma_link(mm, vma)) {
3283 vm_unacct_memory(charged);
3291 * Copy the vma structure to a new location in the same mm,
3292 * prior to moving page table entries, to effect an mremap move.
3294 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3295 unsigned long addr, unsigned long len, pgoff_t pgoff,
3296 bool *need_rmap_locks)
3298 struct vm_area_struct *vma = *vmap;
3299 unsigned long vma_start = vma->vm_start;
3300 struct mm_struct *mm = vma->vm_mm;
3301 struct vm_area_struct *new_vma, *prev;
3302 bool faulted_in_anon_vma = true;
3306 * If anonymous vma has not yet been faulted, update new pgoff
3307 * to match new location, to increase its chance of merging.
3309 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3310 pgoff = addr >> PAGE_SHIFT;
3311 faulted_in_anon_vma = false;
3314 new_vma = find_vma_prev(mm, addr, &prev);
3315 if (new_vma && new_vma->vm_start < addr + len)
3316 return NULL; /* should never get here */
3318 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3319 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3320 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
3323 * Source vma may have been merged into new_vma
3325 if (unlikely(vma_start >= new_vma->vm_start &&
3326 vma_start < new_vma->vm_end)) {
3328 * The only way we can get a vma_merge with
3329 * self during an mremap is if the vma hasn't
3330 * been faulted in yet and we were allowed to
3331 * reset the dst vma->vm_pgoff to the
3332 * destination address of the mremap to allow
3333 * the merge to happen. mremap must change the
3334 * vm_pgoff linearity between src and dst vmas
3335 * (in turn preventing a vma_merge) to be
3336 * safe. It is only safe to keep the vm_pgoff
3337 * linear if there are no pages mapped yet.
3339 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3340 *vmap = vma = new_vma;
3342 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3344 new_vma = vm_area_dup(vma);
3347 new_vma->vm_start = addr;
3348 new_vma->vm_end = addr + len;
3349 new_vma->vm_pgoff = pgoff;
3350 if (vma_dup_policy(vma, new_vma))
3352 if (anon_vma_clone(new_vma, vma))
3353 goto out_free_mempol;
3354 if (new_vma->vm_file)
3355 get_file(new_vma->vm_file);
3356 if (new_vma->vm_ops && new_vma->vm_ops->open)
3357 new_vma->vm_ops->open(new_vma);
3358 if (vma_link(mm, new_vma))
3360 *need_rmap_locks = false;
3366 if (new_vma->vm_ops && new_vma->vm_ops->close)
3367 new_vma->vm_ops->close(new_vma);
3369 if (new_vma->vm_file)
3370 fput(new_vma->vm_file);
3372 unlink_anon_vmas(new_vma);
3374 mpol_put(vma_policy(new_vma));
3376 vm_area_free(new_vma);
3383 * Return true if the calling process may expand its vm space by the passed
3386 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3388 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3391 if (is_data_mapping(flags) &&
3392 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3393 /* Workaround for Valgrind */
3394 if (rlimit(RLIMIT_DATA) == 0 &&
3395 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3398 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3399 current->comm, current->pid,
3400 (mm->data_vm + npages) << PAGE_SHIFT,
3401 rlimit(RLIMIT_DATA),
3402 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3404 if (!ignore_rlimit_data)
3411 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3413 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3415 if (is_exec_mapping(flags))
3416 mm->exec_vm += npages;
3417 else if (is_stack_mapping(flags))
3418 mm->stack_vm += npages;
3419 else if (is_data_mapping(flags))
3420 mm->data_vm += npages;
3423 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3426 * Having a close hook prevents vma merging regardless of flags.
3428 static void special_mapping_close(struct vm_area_struct *vma)
3432 static const char *special_mapping_name(struct vm_area_struct *vma)
3434 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3437 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3439 struct vm_special_mapping *sm = new_vma->vm_private_data;
3441 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3445 return sm->mremap(sm, new_vma);
3450 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3453 * Forbid splitting special mappings - kernel has expectations over
3454 * the number of pages in mapping. Together with VM_DONTEXPAND
3455 * the size of vma should stay the same over the special mapping's
3461 static const struct vm_operations_struct special_mapping_vmops = {
3462 .close = special_mapping_close,
3463 .fault = special_mapping_fault,
3464 .mremap = special_mapping_mremap,
3465 .name = special_mapping_name,
3466 /* vDSO code relies that VVAR can't be accessed remotely */
3468 .may_split = special_mapping_split,
3471 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3472 .close = special_mapping_close,
3473 .fault = special_mapping_fault,
3476 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3478 struct vm_area_struct *vma = vmf->vma;
3480 struct page **pages;
3482 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3483 pages = vma->vm_private_data;
3485 struct vm_special_mapping *sm = vma->vm_private_data;
3488 return sm->fault(sm, vmf->vma, vmf);
3493 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3497 struct page *page = *pages;
3503 return VM_FAULT_SIGBUS;
3506 static struct vm_area_struct *__install_special_mapping(
3507 struct mm_struct *mm,
3508 unsigned long addr, unsigned long len,
3509 unsigned long vm_flags, void *priv,
3510 const struct vm_operations_struct *ops)
3513 struct vm_area_struct *vma;
3516 vma = vm_area_alloc(mm);
3517 if (unlikely(vma == NULL))
3518 return ERR_PTR(-ENOMEM);
3520 vma->vm_start = addr;
3521 vma->vm_end = addr + len;
3523 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3524 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
3525 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3528 vma->vm_private_data = priv;
3530 ret = insert_vm_struct(mm, vma);
3534 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3536 perf_event_mmap(vma);
3544 return ERR_PTR(ret);
3547 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3548 const struct vm_special_mapping *sm)
3550 return vma->vm_private_data == sm &&
3551 (vma->vm_ops == &special_mapping_vmops ||
3552 vma->vm_ops == &legacy_special_mapping_vmops);
3556 * Called with mm->mmap_lock held for writing.
3557 * Insert a new vma covering the given region, with the given flags.
3558 * Its pages are supplied by the given array of struct page *.
3559 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3560 * The region past the last page supplied will always produce SIGBUS.
3561 * The array pointer and the pages it points to are assumed to stay alive
3562 * for as long as this mapping might exist.
3564 struct vm_area_struct *_install_special_mapping(
3565 struct mm_struct *mm,
3566 unsigned long addr, unsigned long len,
3567 unsigned long vm_flags, const struct vm_special_mapping *spec)
3569 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3570 &special_mapping_vmops);
3573 int install_special_mapping(struct mm_struct *mm,
3574 unsigned long addr, unsigned long len,
3575 unsigned long vm_flags, struct page **pages)
3577 struct vm_area_struct *vma = __install_special_mapping(
3578 mm, addr, len, vm_flags, (void *)pages,
3579 &legacy_special_mapping_vmops);
3581 return PTR_ERR_OR_ZERO(vma);
3584 static DEFINE_MUTEX(mm_all_locks_mutex);
3586 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3588 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3590 * The LSB of head.next can't change from under us
3591 * because we hold the mm_all_locks_mutex.
3593 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3595 * We can safely modify head.next after taking the
3596 * anon_vma->root->rwsem. If some other vma in this mm shares
3597 * the same anon_vma we won't take it again.
3599 * No need of atomic instructions here, head.next
3600 * can't change from under us thanks to the
3601 * anon_vma->root->rwsem.
3603 if (__test_and_set_bit(0, (unsigned long *)
3604 &anon_vma->root->rb_root.rb_root.rb_node))
3609 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3611 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3613 * AS_MM_ALL_LOCKS can't change from under us because
3614 * we hold the mm_all_locks_mutex.
3616 * Operations on ->flags have to be atomic because
3617 * even if AS_MM_ALL_LOCKS is stable thanks to the
3618 * mm_all_locks_mutex, there may be other cpus
3619 * changing other bitflags in parallel to us.
3621 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3623 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3628 * This operation locks against the VM for all pte/vma/mm related
3629 * operations that could ever happen on a certain mm. This includes
3630 * vmtruncate, try_to_unmap, and all page faults.
3632 * The caller must take the mmap_lock in write mode before calling
3633 * mm_take_all_locks(). The caller isn't allowed to release the
3634 * mmap_lock until mm_drop_all_locks() returns.
3636 * mmap_lock in write mode is required in order to block all operations
3637 * that could modify pagetables and free pages without need of
3638 * altering the vma layout. It's also needed in write mode to avoid new
3639 * anon_vmas to be associated with existing vmas.
3641 * A single task can't take more than one mm_take_all_locks() in a row
3642 * or it would deadlock.
3644 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3645 * mapping->flags avoid to take the same lock twice, if more than one
3646 * vma in this mm is backed by the same anon_vma or address_space.
3648 * We take locks in following order, accordingly to comment at beginning
3650 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3652 * - all i_mmap_rwsem locks;
3653 * - all anon_vma->rwseml
3655 * We can take all locks within these types randomly because the VM code
3656 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3657 * mm_all_locks_mutex.
3659 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3660 * that may have to take thousand of locks.
3662 * mm_take_all_locks() can fail if it's interrupted by signals.
3664 int mm_take_all_locks(struct mm_struct *mm)
3666 struct vm_area_struct *vma;
3667 struct anon_vma_chain *avc;
3668 MA_STATE(mas, &mm->mm_mt, 0, 0);
3670 mmap_assert_write_locked(mm);
3672 mutex_lock(&mm_all_locks_mutex);
3674 mas_for_each(&mas, vma, ULONG_MAX) {
3675 if (signal_pending(current))
3677 if (vma->vm_file && vma->vm_file->f_mapping &&
3678 is_vm_hugetlb_page(vma))
3679 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3683 mas_for_each(&mas, vma, ULONG_MAX) {
3684 if (signal_pending(current))
3686 if (vma->vm_file && vma->vm_file->f_mapping &&
3687 !is_vm_hugetlb_page(vma))
3688 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3692 mas_for_each(&mas, vma, ULONG_MAX) {
3693 if (signal_pending(current))
3696 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3697 vm_lock_anon_vma(mm, avc->anon_vma);
3703 mm_drop_all_locks(mm);
3707 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3709 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3711 * The LSB of head.next can't change to 0 from under
3712 * us because we hold the mm_all_locks_mutex.
3714 * We must however clear the bitflag before unlocking
3715 * the vma so the users using the anon_vma->rb_root will
3716 * never see our bitflag.
3718 * No need of atomic instructions here, head.next
3719 * can't change from under us until we release the
3720 * anon_vma->root->rwsem.
3722 if (!__test_and_clear_bit(0, (unsigned long *)
3723 &anon_vma->root->rb_root.rb_root.rb_node))
3725 anon_vma_unlock_write(anon_vma);
3729 static void vm_unlock_mapping(struct address_space *mapping)
3731 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3733 * AS_MM_ALL_LOCKS can't change to 0 from under us
3734 * because we hold the mm_all_locks_mutex.
3736 i_mmap_unlock_write(mapping);
3737 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3744 * The mmap_lock cannot be released by the caller until
3745 * mm_drop_all_locks() returns.
3747 void mm_drop_all_locks(struct mm_struct *mm)
3749 struct vm_area_struct *vma;
3750 struct anon_vma_chain *avc;
3751 MA_STATE(mas, &mm->mm_mt, 0, 0);
3753 mmap_assert_write_locked(mm);
3754 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3756 mas_for_each(&mas, vma, ULONG_MAX) {
3758 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3759 vm_unlock_anon_vma(avc->anon_vma);
3760 if (vma->vm_file && vma->vm_file->f_mapping)
3761 vm_unlock_mapping(vma->vm_file->f_mapping);
3764 mutex_unlock(&mm_all_locks_mutex);
3768 * initialise the percpu counter for VM
3770 void __init mmap_init(void)
3774 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3779 * Initialise sysctl_user_reserve_kbytes.
3781 * This is intended to prevent a user from starting a single memory hogging
3782 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3785 * The default value is min(3% of free memory, 128MB)
3786 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3788 static int init_user_reserve(void)
3790 unsigned long free_kbytes;
3792 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3794 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3797 subsys_initcall(init_user_reserve);
3800 * Initialise sysctl_admin_reserve_kbytes.
3802 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3803 * to log in and kill a memory hogging process.
3805 * Systems with more than 256MB will reserve 8MB, enough to recover
3806 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3807 * only reserve 3% of free pages by default.
3809 static int init_admin_reserve(void)
3811 unsigned long free_kbytes;
3813 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3815 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3818 subsys_initcall(init_admin_reserve);
3821 * Reinititalise user and admin reserves if memory is added or removed.
3823 * The default user reserve max is 128MB, and the default max for the
3824 * admin reserve is 8MB. These are usually, but not always, enough to
3825 * enable recovery from a memory hogging process using login/sshd, a shell,
3826 * and tools like top. It may make sense to increase or even disable the
3827 * reserve depending on the existence of swap or variations in the recovery
3828 * tools. So, the admin may have changed them.
3830 * If memory is added and the reserves have been eliminated or increased above
3831 * the default max, then we'll trust the admin.
3833 * If memory is removed and there isn't enough free memory, then we
3834 * need to reset the reserves.
3836 * Otherwise keep the reserve set by the admin.
3838 static int reserve_mem_notifier(struct notifier_block *nb,
3839 unsigned long action, void *data)
3841 unsigned long tmp, free_kbytes;
3845 /* Default max is 128MB. Leave alone if modified by operator. */
3846 tmp = sysctl_user_reserve_kbytes;
3847 if (0 < tmp && tmp < (1UL << 17))
3848 init_user_reserve();
3850 /* Default max is 8MB. Leave alone if modified by operator. */
3851 tmp = sysctl_admin_reserve_kbytes;
3852 if (0 < tmp && tmp < (1UL << 13))
3853 init_admin_reserve();
3857 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3859 if (sysctl_user_reserve_kbytes > free_kbytes) {
3860 init_user_reserve();
3861 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3862 sysctl_user_reserve_kbytes);
3865 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3866 init_admin_reserve();
3867 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3868 sysctl_admin_reserve_kbytes);
3877 static struct notifier_block reserve_mem_nb = {
3878 .notifier_call = reserve_mem_notifier,
3881 static int __meminit init_reserve_notifier(void)
3883 if (register_hotmemory_notifier(&reserve_mem_nb))
3884 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3888 subsys_initcall(init_reserve_notifier);