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,
81 unsigned long end, bool mm_wr_locked);
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));
147 static inline struct vm_area_struct *vma_prev_limit(struct vma_iterator *vmi,
150 return mas_prev(&vmi->mas, min);
153 static inline int vma_iter_clear_gfp(struct vma_iterator *vmi,
154 unsigned long start, unsigned long end, gfp_t gfp)
156 vmi->mas.index = start;
157 vmi->mas.last = end - 1;
158 mas_store_gfp(&vmi->mas, NULL, gfp);
159 if (unlikely(mas_is_err(&vmi->mas)))
166 * check_brk_limits() - Use platform specific check of range & verify mlock
168 * @addr: The address to check
169 * @len: The size of increase.
171 * Return: 0 on success.
173 static int check_brk_limits(unsigned long addr, unsigned long len)
175 unsigned long mapped_addr;
177 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
178 if (IS_ERR_VALUE(mapped_addr))
181 return mlock_future_check(current->mm, current->mm->def_flags, len);
183 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma,
184 unsigned long addr, unsigned long request, unsigned long flags);
185 SYSCALL_DEFINE1(brk, unsigned long, brk)
187 unsigned long newbrk, oldbrk, origbrk;
188 struct mm_struct *mm = current->mm;
189 struct vm_area_struct *brkvma, *next = NULL;
190 unsigned long min_brk;
192 bool downgraded = false;
194 struct vma_iterator vmi;
196 if (mmap_write_lock_killable(mm))
201 #ifdef CONFIG_COMPAT_BRK
203 * CONFIG_COMPAT_BRK can still be overridden by setting
204 * randomize_va_space to 2, which will still cause mm->start_brk
205 * to be arbitrarily shifted
207 if (current->brk_randomized)
208 min_brk = mm->start_brk;
210 min_brk = mm->end_data;
212 min_brk = mm->start_brk;
218 * Check against rlimit here. If this check is done later after the test
219 * of oldbrk with newbrk then it can escape the test and let the data
220 * segment grow beyond its set limit the in case where the limit is
221 * not page aligned -Ram Gupta
223 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
224 mm->end_data, mm->start_data))
227 newbrk = PAGE_ALIGN(brk);
228 oldbrk = PAGE_ALIGN(mm->brk);
229 if (oldbrk == newbrk) {
235 * Always allow shrinking brk.
236 * do_vma_munmap() may downgrade mmap_lock to read.
238 if (brk <= mm->brk) {
241 /* Search one past newbrk */
242 vma_iter_init(&vmi, mm, newbrk);
243 brkvma = vma_find(&vmi, oldbrk);
244 if (!brkvma || brkvma->vm_start >= oldbrk)
245 goto out; /* mapping intersects with an existing non-brk vma. */
247 * mm->brk must be protected by write mmap_lock.
248 * do_vma_munmap() may downgrade the lock, so update it
249 * before calling do_vma_munmap().
252 ret = do_vma_munmap(&vmi, brkvma, newbrk, oldbrk, &uf, true);
263 if (check_brk_limits(oldbrk, newbrk - oldbrk))
267 * Only check if the next VMA is within the stack_guard_gap of the
270 vma_iter_init(&vmi, mm, oldbrk);
271 next = vma_find(&vmi, newbrk + PAGE_SIZE + stack_guard_gap);
272 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
275 brkvma = vma_prev_limit(&vmi, mm->start_brk);
276 /* Ok, looks good - let it rip. */
277 if (do_brk_flags(&vmi, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
283 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
285 mmap_read_unlock(mm);
287 mmap_write_unlock(mm);
288 userfaultfd_unmap_complete(mm, &uf);
290 mm_populate(oldbrk, newbrk - oldbrk);
294 mmap_write_unlock(mm);
298 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
299 extern void mt_validate(struct maple_tree *mt);
300 extern void mt_dump(const struct maple_tree *mt);
302 /* Validate the maple tree */
303 static void validate_mm_mt(struct mm_struct *mm)
305 struct maple_tree *mt = &mm->mm_mt;
306 struct vm_area_struct *vma_mt;
308 MA_STATE(mas, mt, 0, 0);
310 mt_validate(&mm->mm_mt);
311 mas_for_each(&mas, vma_mt, ULONG_MAX) {
312 if ((vma_mt->vm_start != mas.index) ||
313 (vma_mt->vm_end - 1 != mas.last)) {
314 pr_emerg("issue in %s\n", current->comm);
317 pr_emerg("mt piv: %p %lu - %lu\n", vma_mt,
318 mas.index, mas.last);
319 pr_emerg("mt vma: %p %lu - %lu\n", vma_mt,
320 vma_mt->vm_start, vma_mt->vm_end);
323 if (vma_mt->vm_end != mas.last + 1) {
324 pr_err("vma: %p vma_mt %lu-%lu\tmt %lu-%lu\n",
325 mm, vma_mt->vm_start, vma_mt->vm_end,
326 mas.index, mas.last);
329 VM_BUG_ON_MM(vma_mt->vm_end != mas.last + 1, mm);
330 if (vma_mt->vm_start != mas.index) {
331 pr_err("vma: %p vma_mt %p %lu - %lu doesn't match\n",
332 mm, vma_mt, vma_mt->vm_start, vma_mt->vm_end);
335 VM_BUG_ON_MM(vma_mt->vm_start != mas.index, mm);
340 static void validate_mm(struct mm_struct *mm)
344 struct vm_area_struct *vma;
345 MA_STATE(mas, &mm->mm_mt, 0, 0);
349 mas_for_each(&mas, vma, ULONG_MAX) {
350 #ifdef CONFIG_DEBUG_VM_RB
351 struct anon_vma *anon_vma = vma->anon_vma;
352 struct anon_vma_chain *avc;
355 anon_vma_lock_read(anon_vma);
356 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
357 anon_vma_interval_tree_verify(avc);
358 anon_vma_unlock_read(anon_vma);
363 if (i != mm->map_count) {
364 pr_emerg("map_count %d mas_for_each %d\n", mm->map_count, i);
367 VM_BUG_ON_MM(bug, mm);
370 #else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
371 #define validate_mm_mt(root) do { } while (0)
372 #define validate_mm(mm) do { } while (0)
373 #endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
376 * vma has some anon_vma assigned, and is already inserted on that
377 * anon_vma's interval trees.
379 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
380 * vma must be removed from the anon_vma's interval trees using
381 * anon_vma_interval_tree_pre_update_vma().
383 * After the update, the vma will be reinserted using
384 * anon_vma_interval_tree_post_update_vma().
386 * The entire update must be protected by exclusive mmap_lock and by
387 * the root anon_vma's mutex.
390 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
392 struct anon_vma_chain *avc;
394 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
395 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
399 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
401 struct anon_vma_chain *avc;
403 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
404 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
407 static unsigned long count_vma_pages_range(struct mm_struct *mm,
408 unsigned long addr, unsigned long end)
410 VMA_ITERATOR(vmi, mm, addr);
411 struct vm_area_struct *vma;
412 unsigned long nr_pages = 0;
414 for_each_vma_range(vmi, vma, end) {
415 unsigned long vm_start = max(addr, vma->vm_start);
416 unsigned long vm_end = min(end, vma->vm_end);
418 nr_pages += PHYS_PFN(vm_end - vm_start);
424 static void __vma_link_file(struct vm_area_struct *vma,
425 struct address_space *mapping)
427 if (vma->vm_flags & VM_SHARED)
428 mapping_allow_writable(mapping);
430 flush_dcache_mmap_lock(mapping);
431 vma_interval_tree_insert(vma, &mapping->i_mmap);
432 flush_dcache_mmap_unlock(mapping);
435 static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
437 VMA_ITERATOR(vmi, mm, 0);
438 struct address_space *mapping = NULL;
440 if (vma_iter_prealloc(&vmi))
444 mapping = vma->vm_file->f_mapping;
445 i_mmap_lock_write(mapping);
448 vma_iter_store(&vmi, vma);
451 __vma_link_file(vma, mapping);
452 i_mmap_unlock_write(mapping);
461 * init_multi_vma_prep() - Initializer for struct vma_prepare
462 * @vp: The vma_prepare struct
463 * @vma: The vma that will be altered once locked
464 * @next: The next vma if it is to be adjusted
465 * @remove: The first vma to be removed
466 * @remove2: The second vma to be removed
468 static inline void init_multi_vma_prep(struct vma_prepare *vp,
469 struct vm_area_struct *vma, struct vm_area_struct *next,
470 struct vm_area_struct *remove, struct vm_area_struct *remove2)
472 memset(vp, 0, sizeof(struct vma_prepare));
474 vp->anon_vma = vma->anon_vma;
476 vp->remove2 = remove2;
478 if (!vp->anon_vma && next)
479 vp->anon_vma = next->anon_vma;
481 vp->file = vma->vm_file;
483 vp->mapping = vma->vm_file->f_mapping;
488 * init_vma_prep() - Initializer wrapper for vma_prepare struct
489 * @vp: The vma_prepare struct
490 * @vma: The vma that will be altered once locked
492 static inline void init_vma_prep(struct vma_prepare *vp,
493 struct vm_area_struct *vma)
495 init_multi_vma_prep(vp, vma, NULL, NULL, NULL);
500 * vma_prepare() - Helper function for handling locking VMAs prior to altering
501 * @vp: The initialized vma_prepare struct
503 static inline void vma_prepare(struct vma_prepare *vp)
506 uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end);
509 uprobe_munmap(vp->adj_next, vp->adj_next->vm_start,
510 vp->adj_next->vm_end);
512 i_mmap_lock_write(vp->mapping);
513 if (vp->insert && vp->insert->vm_file) {
515 * Put into interval tree now, so instantiated pages
516 * are visible to arm/parisc __flush_dcache_page
517 * throughout; but we cannot insert into address
518 * space until vma start or end is updated.
520 __vma_link_file(vp->insert,
521 vp->insert->vm_file->f_mapping);
526 anon_vma_lock_write(vp->anon_vma);
527 anon_vma_interval_tree_pre_update_vma(vp->vma);
529 anon_vma_interval_tree_pre_update_vma(vp->adj_next);
533 flush_dcache_mmap_lock(vp->mapping);
534 vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap);
536 vma_interval_tree_remove(vp->adj_next,
537 &vp->mapping->i_mmap);
543 * vma_complete- Helper function for handling the unlocking after altering VMAs,
544 * or for inserting a VMA.
546 * @vp: The vma_prepare struct
547 * @vmi: The vma iterator
550 static inline void vma_complete(struct vma_prepare *vp,
551 struct vma_iterator *vmi, struct mm_struct *mm)
555 vma_interval_tree_insert(vp->adj_next,
556 &vp->mapping->i_mmap);
557 vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap);
558 flush_dcache_mmap_unlock(vp->mapping);
561 if (vp->remove && vp->file) {
562 __remove_shared_vm_struct(vp->remove, vp->file, vp->mapping);
564 __remove_shared_vm_struct(vp->remove2, vp->file,
566 } else if (vp->insert) {
568 * split_vma has split insert from vma, and needs
569 * us to insert it before dropping the locks
570 * (it may either follow vma or precede it).
572 vma_iter_store(vmi, vp->insert);
577 anon_vma_interval_tree_post_update_vma(vp->vma);
579 anon_vma_interval_tree_post_update_vma(vp->adj_next);
580 anon_vma_unlock_write(vp->anon_vma);
584 i_mmap_unlock_write(vp->mapping);
585 uprobe_mmap(vp->vma);
588 uprobe_mmap(vp->adj_next);
594 uprobe_munmap(vp->remove, vp->remove->vm_start,
598 if (vp->remove->anon_vma)
599 anon_vma_merge(vp->vma, vp->remove);
601 mpol_put(vma_policy(vp->remove));
603 WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end);
604 vm_area_free(vp->remove);
607 * In mprotect's case 6 (see comments on vma_merge),
608 * we are removing both mid and next vmas
611 vp->remove = vp->remove2;
616 if (vp->insert && vp->file)
617 uprobe_mmap(vp->insert);
621 * dup_anon_vma() - Helper function to duplicate anon_vma
622 * @dst: The destination VMA
623 * @src: The source VMA
625 * Returns: 0 on success.
627 static inline int dup_anon_vma(struct vm_area_struct *dst,
628 struct vm_area_struct *src)
631 * Easily overlooked: when mprotect shifts the boundary, make sure the
632 * expanding vma has anon_vma set if the shrinking vma had, to cover any
633 * anon pages imported.
635 if (src->anon_vma && !dst->anon_vma) {
636 dst->anon_vma = src->anon_vma;
637 return anon_vma_clone(dst, src);
644 * vma_expand - Expand an existing VMA
646 * @vmi: The vma iterator
647 * @vma: The vma to expand
648 * @start: The start of the vma
649 * @end: The exclusive end of the vma
650 * @pgoff: The page offset of vma
651 * @next: The current of next vma.
653 * Expand @vma to @start and @end. Can expand off the start and end. Will
654 * expand over @next if it's different from @vma and @end == @next->vm_end.
655 * Checking if the @vma can expand and merge with @next needs to be handled by
658 * Returns: 0 on success
660 int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
661 unsigned long start, unsigned long end, pgoff_t pgoff,
662 struct vm_area_struct *next)
664 bool remove_next = false;
665 struct vma_prepare vp;
667 if (next && (vma != next) && (end == next->vm_end)) {
671 ret = dup_anon_vma(vma, next);
676 init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL);
677 /* Not merging but overwriting any part of next is not handled. */
678 VM_WARN_ON(next && !vp.remove &&
679 next != vma && end > next->vm_start);
680 /* Only handles expanding */
681 VM_WARN_ON(vma->vm_start < start || vma->vm_end > end);
683 if (vma_iter_prealloc(vmi))
686 vma_adjust_trans_huge(vma, start, end, 0);
687 /* VMA iterator points to previous, so set to start if necessary */
688 if (vma_iter_addr(vmi) != start)
689 vma_iter_set(vmi, start);
692 vma->vm_start = start;
694 vma->vm_pgoff = pgoff;
695 /* Note: mas must be pointing to the expanding VMA */
696 vma_iter_store(vmi, vma);
698 vma_complete(&vp, vmi, vma->vm_mm);
699 validate_mm(vma->vm_mm);
707 * vma_shrink() - Reduce an existing VMAs memory area
708 * @vmi: The vma iterator
709 * @vma: The VMA to modify
710 * @start: The new start
713 * Returns: 0 on success, -ENOMEM otherwise
715 int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
716 unsigned long start, unsigned long end, pgoff_t pgoff)
718 struct vma_prepare vp;
720 WARN_ON((vma->vm_start != start) && (vma->vm_end != end));
722 if (vma_iter_prealloc(vmi))
725 init_vma_prep(&vp, vma);
726 vma_adjust_trans_huge(vma, start, end, 0);
729 if (vma->vm_start < start)
730 vma_iter_clear(vmi, vma->vm_start, start);
732 if (vma->vm_end > end)
733 vma_iter_clear(vmi, end, vma->vm_end);
735 vma->vm_start = start;
737 vma->vm_pgoff = pgoff;
738 vma_complete(&vp, vmi, vma->vm_mm);
739 validate_mm(vma->vm_mm);
744 * If the vma has a ->close operation then the driver probably needs to release
745 * per-vma resources, so we don't attempt to merge those.
747 static inline bool is_mergeable_vma(struct vm_area_struct *vma,
748 struct file *file, unsigned long vm_flags,
749 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
750 struct anon_vma_name *anon_name)
753 * VM_SOFTDIRTY should not prevent from VMA merging, if we
754 * match the flags but dirty bit -- the caller should mark
755 * merged VMA as dirty. If dirty bit won't be excluded from
756 * comparison, we increase pressure on the memory system forcing
757 * the kernel to generate new VMAs when old one could be
760 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
762 if (vma->vm_file != file)
764 if (vma->vm_ops && vma->vm_ops->close)
766 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
768 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
773 static inline bool is_mergeable_anon_vma(struct anon_vma *anon_vma1,
774 struct anon_vma *anon_vma2, struct vm_area_struct *vma)
777 * The list_is_singular() test is to avoid merging VMA cloned from
778 * parents. This can improve scalability caused by anon_vma lock.
780 if ((!anon_vma1 || !anon_vma2) && (!vma ||
781 list_is_singular(&vma->anon_vma_chain)))
783 return anon_vma1 == anon_vma2;
787 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
788 * in front of (at a lower virtual address and file offset than) the vma.
790 * We cannot merge two vmas if they have differently assigned (non-NULL)
791 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
793 * We don't check here for the merged mmap wrapping around the end of pagecache
794 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
795 * wrap, nor mmaps which cover the final page at index -1UL.
798 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
799 struct anon_vma *anon_vma, struct file *file,
800 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
801 struct anon_vma_name *anon_name)
803 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
804 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
805 if (vma->vm_pgoff == vm_pgoff)
812 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
813 * beyond (at a higher virtual address and file offset than) the vma.
815 * We cannot merge two vmas if they have differently assigned (non-NULL)
816 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
819 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
820 struct anon_vma *anon_vma, struct file *file,
821 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
822 struct anon_vma_name *anon_name)
824 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
825 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
827 vm_pglen = vma_pages(vma);
828 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
835 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
836 * figure out whether that can be merged with its predecessor or its
837 * successor. Or both (it neatly fills a hole).
839 * In most cases - when called for mmap, brk or mremap - [addr,end) is
840 * certain not to be mapped by the time vma_merge is called; but when
841 * called for mprotect, it is certain to be already mapped (either at
842 * an offset within prev, or at the start of next), and the flags of
843 * this area are about to be changed to vm_flags - and the no-change
844 * case has already been eliminated.
846 * The following mprotect cases have to be considered, where AAAA is
847 * the area passed down from mprotect_fixup, never extending beyond one
848 * vma, PPPP is the previous vma, NNNN is a vma that starts at the same
849 * address as AAAA and is of the same or larger span, and XXXX the next
853 * PPPPPPXXXXXX PPPPPPXXXXXX PPPPPPNNNNNN
854 * cannot merge might become might become
855 * PPXXXXXXXXXX PPPPPPPPPPNN
856 * mmap, brk or case 4 below case 5 below
859 * PPPP XXXX PPPPNNNNXXXX
860 * might become might become
861 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
862 * PPPPPPPPXXXX 2 or PPPPPPPPXXXX 7 or
863 * PPPPXXXXXXXX 3 PPPPXXXXXXXX 8
865 * It is important for case 8 that the vma NNNN overlapping the
866 * region AAAA is never going to extended over XXXX. Instead XXXX must
867 * be extended in region AAAA and NNNN must be removed. This way in
868 * all cases where vma_merge succeeds, the moment vma_merge drops the
869 * rmap_locks, the properties of the merged vma will be already
870 * correct for the whole merged range. Some of those properties like
871 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
872 * be correct for the whole merged range immediately after the
873 * rmap_locks are released. Otherwise if XXXX would be removed and
874 * NNNN would be extended over the XXXX range, remove_migration_ptes
875 * or other rmap walkers (if working on addresses beyond the "end"
876 * parameter) may establish ptes with the wrong permissions of NNNN
877 * instead of the right permissions of XXXX.
880 * PPPP is represented by *prev
881 * NNNN is represented by *mid or not represented at all (NULL)
882 * XXXX is represented by *next or not represented at all (NULL)
883 * AAAA is not represented - it will be merged and the vma containing the
884 * area is returned, or the function will return NULL
886 struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
887 struct vm_area_struct *prev, unsigned long addr,
888 unsigned long end, unsigned long vm_flags,
889 struct anon_vma *anon_vma, struct file *file,
890 pgoff_t pgoff, struct mempolicy *policy,
891 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
892 struct anon_vma_name *anon_name)
894 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
896 struct vm_area_struct *mid, *next, *res = NULL;
897 struct vm_area_struct *vma, *adjust, *remove, *remove2;
899 bool merge_prev = false;
900 bool merge_next = false;
901 bool vma_expanded = false;
902 struct vma_prepare vp;
903 unsigned long vma_end = end;
905 unsigned long vma_start = addr;
909 * We later require that vma->vm_flags == vm_flags,
910 * so this tests vma->vm_flags & VM_SPECIAL, too.
912 if (vm_flags & VM_SPECIAL)
915 mid = find_vma(mm, prev ? prev->vm_end : 0);
916 if (mid && mid->vm_end == end) /* cases 6, 7, 8 */
917 next = find_vma(mm, mid->vm_end);
921 /* In cases 1 - 4 there's no NNNN vma */
922 if (mid && end <= mid->vm_start)
925 /* verify some invariant that must be enforced by the caller */
926 VM_WARN_ON(prev && addr <= prev->vm_start);
927 VM_WARN_ON(mid && end > mid->vm_end);
928 VM_WARN_ON(addr >= end);
933 vma_start = prev->vm_start;
934 vma_pgoff = prev->vm_pgoff;
935 /* Can we merge the predecessor? */
936 if (prev->vm_end == addr && mpol_equal(vma_policy(prev), policy)
937 && can_vma_merge_after(prev, vm_flags, anon_vma, file,
938 pgoff, vm_userfaultfd_ctx, anon_name)) {
943 /* Can we merge the successor? */
944 if (next && end == next->vm_start &&
945 mpol_equal(policy, vma_policy(next)) &&
946 can_vma_merge_before(next, vm_flags,
947 anon_vma, file, pgoff+pglen,
948 vm_userfaultfd_ctx, anon_name)) {
952 remove = remove2 = adjust = NULL;
953 /* Can we merge both the predecessor and the successor? */
954 if (merge_prev && merge_next &&
955 is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) {
956 remove = next; /* case 1 */
957 vma_end = next->vm_end;
958 err = dup_anon_vma(prev, next);
959 if (mid) { /* case 6 */
963 err = dup_anon_vma(prev, mid);
965 } else if (merge_prev) {
966 err = 0; /* case 2 */
968 err = dup_anon_vma(prev, mid);
969 if (end == mid->vm_end) { /* case 7 */
971 } else { /* case 5 */
973 adj_start = (end - mid->vm_start);
976 } else if (merge_next) {
978 if (prev && addr < prev->vm_end) { /* case 4 */
981 adj_start = -(prev->vm_end - addr);
982 err = dup_anon_vma(next, prev);
984 vma = next; /* case 3 */
986 vma_end = next->vm_end;
987 vma_pgoff = next->vm_pgoff;
989 if (mid) { /* case 8 */
990 vma_pgoff = mid->vm_pgoff;
992 err = dup_anon_vma(next, mid);
997 /* Cannot merge or error in anon_vma clone */
1001 if (vma_iter_prealloc(vmi))
1004 vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start);
1005 init_multi_vma_prep(&vp, vma, adjust, remove, remove2);
1006 VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
1007 vp.anon_vma != adjust->anon_vma);
1010 if (vma_start < vma->vm_start || vma_end > vma->vm_end)
1011 vma_expanded = true;
1013 vma->vm_start = vma_start;
1014 vma->vm_end = vma_end;
1015 vma->vm_pgoff = vma_pgoff;
1018 vma_iter_store(vmi, vma);
1021 adjust->vm_start += adj_start;
1022 adjust->vm_pgoff += adj_start >> PAGE_SHIFT;
1023 if (adj_start < 0) {
1024 WARN_ON(vma_expanded);
1025 vma_iter_store(vmi, next);
1029 vma_complete(&vp, vmi, mm);
1032 khugepaged_enter_vma(res, vm_flags);
1038 * Rough compatibility check to quickly see if it's even worth looking
1039 * at sharing an anon_vma.
1041 * They need to have the same vm_file, and the flags can only differ
1042 * in things that mprotect may change.
1044 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1045 * we can merge the two vma's. For example, we refuse to merge a vma if
1046 * there is a vm_ops->close() function, because that indicates that the
1047 * driver is doing some kind of reference counting. But that doesn't
1048 * really matter for the anon_vma sharing case.
1050 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1052 return a->vm_end == b->vm_start &&
1053 mpol_equal(vma_policy(a), vma_policy(b)) &&
1054 a->vm_file == b->vm_file &&
1055 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1056 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1060 * Do some basic sanity checking to see if we can re-use the anon_vma
1061 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1062 * the same as 'old', the other will be the new one that is trying
1063 * to share the anon_vma.
1065 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1066 * the anon_vma of 'old' is concurrently in the process of being set up
1067 * by another page fault trying to merge _that_. But that's ok: if it
1068 * is being set up, that automatically means that it will be a singleton
1069 * acceptable for merging, so we can do all of this optimistically. But
1070 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1072 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1073 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1074 * is to return an anon_vma that is "complex" due to having gone through
1077 * We also make sure that the two vma's are compatible (adjacent,
1078 * and with the same memory policies). That's all stable, even with just
1079 * a read lock on the mmap_lock.
1081 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1083 if (anon_vma_compatible(a, b)) {
1084 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1086 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1093 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1094 * neighbouring vmas for a suitable anon_vma, before it goes off
1095 * to allocate a new anon_vma. It checks because a repetitive
1096 * sequence of mprotects and faults may otherwise lead to distinct
1097 * anon_vmas being allocated, preventing vma merge in subsequent
1100 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1102 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1103 struct anon_vma *anon_vma = NULL;
1104 struct vm_area_struct *prev, *next;
1106 /* Try next first. */
1107 next = mas_walk(&mas);
1109 anon_vma = reusable_anon_vma(next, vma, next);
1114 prev = mas_prev(&mas, 0);
1115 VM_BUG_ON_VMA(prev != vma, vma);
1116 prev = mas_prev(&mas, 0);
1117 /* Try prev next. */
1119 anon_vma = reusable_anon_vma(prev, prev, vma);
1122 * We might reach here with anon_vma == NULL if we can't find
1123 * any reusable anon_vma.
1124 * There's no absolute need to look only at touching neighbours:
1125 * we could search further afield for "compatible" anon_vmas.
1126 * But it would probably just be a waste of time searching,
1127 * or lead to too many vmas hanging off the same anon_vma.
1128 * We're trying to allow mprotect remerging later on,
1129 * not trying to minimize memory used for anon_vmas.
1135 * If a hint addr is less than mmap_min_addr change hint to be as
1136 * low as possible but still greater than mmap_min_addr
1138 static inline unsigned long round_hint_to_min(unsigned long hint)
1141 if (((void *)hint != NULL) &&
1142 (hint < mmap_min_addr))
1143 return PAGE_ALIGN(mmap_min_addr);
1147 int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1150 unsigned long locked, lock_limit;
1152 /* mlock MCL_FUTURE? */
1153 if (flags & VM_LOCKED) {
1154 locked = len >> PAGE_SHIFT;
1155 locked += mm->locked_vm;
1156 lock_limit = rlimit(RLIMIT_MEMLOCK);
1157 lock_limit >>= PAGE_SHIFT;
1158 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1164 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1166 if (S_ISREG(inode->i_mode))
1167 return MAX_LFS_FILESIZE;
1169 if (S_ISBLK(inode->i_mode))
1170 return MAX_LFS_FILESIZE;
1172 if (S_ISSOCK(inode->i_mode))
1173 return MAX_LFS_FILESIZE;
1175 /* Special "we do even unsigned file positions" case */
1176 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1179 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1183 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1184 unsigned long pgoff, unsigned long len)
1186 u64 maxsize = file_mmap_size_max(file, inode);
1188 if (maxsize && len > maxsize)
1191 if (pgoff > maxsize >> PAGE_SHIFT)
1197 * The caller must write-lock current->mm->mmap_lock.
1199 unsigned long do_mmap(struct file *file, unsigned long addr,
1200 unsigned long len, unsigned long prot,
1201 unsigned long flags, unsigned long pgoff,
1202 unsigned long *populate, struct list_head *uf)
1204 struct mm_struct *mm = current->mm;
1205 vm_flags_t vm_flags;
1215 * Does the application expect PROT_READ to imply PROT_EXEC?
1217 * (the exception is when the underlying filesystem is noexec
1218 * mounted, in which case we dont add PROT_EXEC.)
1220 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1221 if (!(file && path_noexec(&file->f_path)))
1224 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1225 if (flags & MAP_FIXED_NOREPLACE)
1228 if (!(flags & MAP_FIXED))
1229 addr = round_hint_to_min(addr);
1231 /* Careful about overflows.. */
1232 len = PAGE_ALIGN(len);
1236 /* offset overflow? */
1237 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1240 /* Too many mappings? */
1241 if (mm->map_count > sysctl_max_map_count)
1244 /* Obtain the address to map to. we verify (or select) it and ensure
1245 * that it represents a valid section of the address space.
1247 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1248 if (IS_ERR_VALUE(addr))
1251 if (flags & MAP_FIXED_NOREPLACE) {
1252 if (find_vma_intersection(mm, addr, addr + len))
1256 if (prot == PROT_EXEC) {
1257 pkey = execute_only_pkey(mm);
1262 /* Do simple checking here so the lower-level routines won't have
1263 * to. we assume access permissions have been handled by the open
1264 * of the memory object, so we don't do any here.
1266 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1267 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1269 if (flags & MAP_LOCKED)
1270 if (!can_do_mlock())
1273 if (mlock_future_check(mm, vm_flags, len))
1277 struct inode *inode = file_inode(file);
1278 unsigned long flags_mask;
1280 if (!file_mmap_ok(file, inode, pgoff, len))
1283 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1285 switch (flags & MAP_TYPE) {
1288 * Force use of MAP_SHARED_VALIDATE with non-legacy
1289 * flags. E.g. MAP_SYNC is dangerous to use with
1290 * MAP_SHARED as you don't know which consistency model
1291 * you will get. We silently ignore unsupported flags
1292 * with MAP_SHARED to preserve backward compatibility.
1294 flags &= LEGACY_MAP_MASK;
1296 case MAP_SHARED_VALIDATE:
1297 if (flags & ~flags_mask)
1299 if (prot & PROT_WRITE) {
1300 if (!(file->f_mode & FMODE_WRITE))
1302 if (IS_SWAPFILE(file->f_mapping->host))
1307 * Make sure we don't allow writing to an append-only
1310 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1313 vm_flags |= VM_SHARED | VM_MAYSHARE;
1314 if (!(file->f_mode & FMODE_WRITE))
1315 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1318 if (!(file->f_mode & FMODE_READ))
1320 if (path_noexec(&file->f_path)) {
1321 if (vm_flags & VM_EXEC)
1323 vm_flags &= ~VM_MAYEXEC;
1326 if (!file->f_op->mmap)
1328 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1336 switch (flags & MAP_TYPE) {
1338 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1344 vm_flags |= VM_SHARED | VM_MAYSHARE;
1348 * Set pgoff according to addr for anon_vma.
1350 pgoff = addr >> PAGE_SHIFT;
1358 * Set 'VM_NORESERVE' if we should not account for the
1359 * memory use of this mapping.
1361 if (flags & MAP_NORESERVE) {
1362 /* We honor MAP_NORESERVE if allowed to overcommit */
1363 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1364 vm_flags |= VM_NORESERVE;
1366 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1367 if (file && is_file_hugepages(file))
1368 vm_flags |= VM_NORESERVE;
1371 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1372 if (!IS_ERR_VALUE(addr) &&
1373 ((vm_flags & VM_LOCKED) ||
1374 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1379 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1380 unsigned long prot, unsigned long flags,
1381 unsigned long fd, unsigned long pgoff)
1383 struct file *file = NULL;
1384 unsigned long retval;
1386 if (!(flags & MAP_ANONYMOUS)) {
1387 audit_mmap_fd(fd, flags);
1391 if (is_file_hugepages(file)) {
1392 len = ALIGN(len, huge_page_size(hstate_file(file)));
1393 } else if (unlikely(flags & MAP_HUGETLB)) {
1397 } else if (flags & MAP_HUGETLB) {
1400 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1404 len = ALIGN(len, huge_page_size(hs));
1406 * VM_NORESERVE is used because the reservations will be
1407 * taken when vm_ops->mmap() is called
1409 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1411 HUGETLB_ANONHUGE_INODE,
1412 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1414 return PTR_ERR(file);
1417 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1424 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1425 unsigned long, prot, unsigned long, flags,
1426 unsigned long, fd, unsigned long, pgoff)
1428 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1431 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1432 struct mmap_arg_struct {
1436 unsigned long flags;
1438 unsigned long offset;
1441 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1443 struct mmap_arg_struct a;
1445 if (copy_from_user(&a, arg, sizeof(a)))
1447 if (offset_in_page(a.offset))
1450 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1451 a.offset >> PAGE_SHIFT);
1453 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1456 * Some shared mappings will want the pages marked read-only
1457 * to track write events. If so, we'll downgrade vm_page_prot
1458 * to the private version (using protection_map[] without the
1461 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1463 vm_flags_t vm_flags = vma->vm_flags;
1464 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1466 /* If it was private or non-writable, the write bit is already clear */
1467 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1470 /* The backer wishes to know when pages are first written to? */
1471 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1474 /* The open routine did something to the protections that pgprot_modify
1475 * won't preserve? */
1476 if (pgprot_val(vm_page_prot) !=
1477 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1481 * Do we need to track softdirty? hugetlb does not support softdirty
1484 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1487 /* Do we need write faults for uffd-wp tracking? */
1488 if (userfaultfd_wp(vma))
1491 /* Specialty mapping? */
1492 if (vm_flags & VM_PFNMAP)
1495 /* Can the mapping track the dirty pages? */
1496 return vma->vm_file && vma->vm_file->f_mapping &&
1497 mapping_can_writeback(vma->vm_file->f_mapping);
1501 * We account for memory if it's a private writeable mapping,
1502 * not hugepages and VM_NORESERVE wasn't set.
1504 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1507 * hugetlb has its own accounting separate from the core VM
1508 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1510 if (file && is_file_hugepages(file))
1513 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1517 * unmapped_area() - Find an area between the low_limit and the high_limit with
1518 * the correct alignment and offset, all from @info. Note: current->mm is used
1521 * @info: The unmapped area information including the range [low_limit -
1522 * high_limit), the alignment offset and mask.
1524 * Return: A memory address or -ENOMEM.
1526 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1528 unsigned long length, gap;
1530 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1532 /* Adjust search length to account for worst case alignment overhead */
1533 length = info->length + info->align_mask;
1534 if (length < info->length)
1537 if (mas_empty_area(&mas, info->low_limit, info->high_limit - 1,
1542 gap += (info->align_offset - gap) & info->align_mask;
1547 * unmapped_area_topdown() - Find an area between the low_limit and the
1548 * high_limit with the correct alignment and offset at the highest available
1549 * address, all from @info. Note: current->mm is used for the search.
1551 * @info: The unmapped area information including the range [low_limit -
1552 * high_limit), the alignment offset and mask.
1554 * Return: A memory address or -ENOMEM.
1556 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1558 unsigned long length, gap;
1560 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1561 /* Adjust search length to account for worst case alignment overhead */
1562 length = info->length + info->align_mask;
1563 if (length < info->length)
1566 if (mas_empty_area_rev(&mas, info->low_limit, info->high_limit - 1,
1570 gap = mas.last + 1 - info->length;
1571 gap -= (gap - info->align_offset) & info->align_mask;
1576 * Search for an unmapped address range.
1578 * We are looking for a range that:
1579 * - does not intersect with any VMA;
1580 * - is contained within the [low_limit, high_limit) interval;
1581 * - is at least the desired size.
1582 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1584 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1588 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1589 addr = unmapped_area_topdown(info);
1591 addr = unmapped_area(info);
1593 trace_vm_unmapped_area(addr, info);
1597 /* Get an address range which is currently unmapped.
1598 * For shmat() with addr=0.
1600 * Ugly calling convention alert:
1601 * Return value with the low bits set means error value,
1603 * if (ret & ~PAGE_MASK)
1606 * This function "knows" that -ENOMEM has the bits set.
1609 generic_get_unmapped_area(struct file *filp, unsigned long addr,
1610 unsigned long len, unsigned long pgoff,
1611 unsigned long flags)
1613 struct mm_struct *mm = current->mm;
1614 struct vm_area_struct *vma, *prev;
1615 struct vm_unmapped_area_info info;
1616 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1618 if (len > mmap_end - mmap_min_addr)
1621 if (flags & MAP_FIXED)
1625 addr = PAGE_ALIGN(addr);
1626 vma = find_vma_prev(mm, addr, &prev);
1627 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1628 (!vma || addr + len <= vm_start_gap(vma)) &&
1629 (!prev || addr >= vm_end_gap(prev)))
1635 info.low_limit = mm->mmap_base;
1636 info.high_limit = mmap_end;
1637 info.align_mask = 0;
1638 info.align_offset = 0;
1639 return vm_unmapped_area(&info);
1642 #ifndef HAVE_ARCH_UNMAPPED_AREA
1644 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1645 unsigned long len, unsigned long pgoff,
1646 unsigned long flags)
1648 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1653 * This mmap-allocator allocates new areas top-down from below the
1654 * stack's low limit (the base):
1657 generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1658 unsigned long len, unsigned long pgoff,
1659 unsigned long flags)
1661 struct vm_area_struct *vma, *prev;
1662 struct mm_struct *mm = current->mm;
1663 struct vm_unmapped_area_info info;
1664 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1666 /* requested length too big for entire address space */
1667 if (len > mmap_end - mmap_min_addr)
1670 if (flags & MAP_FIXED)
1673 /* requesting a specific address */
1675 addr = PAGE_ALIGN(addr);
1676 vma = find_vma_prev(mm, addr, &prev);
1677 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1678 (!vma || addr + len <= vm_start_gap(vma)) &&
1679 (!prev || addr >= vm_end_gap(prev)))
1683 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1685 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1686 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1687 info.align_mask = 0;
1688 info.align_offset = 0;
1689 addr = vm_unmapped_area(&info);
1692 * A failed mmap() very likely causes application failure,
1693 * so fall back to the bottom-up function here. This scenario
1694 * can happen with large stack limits and large mmap()
1697 if (offset_in_page(addr)) {
1698 VM_BUG_ON(addr != -ENOMEM);
1700 info.low_limit = TASK_UNMAPPED_BASE;
1701 info.high_limit = mmap_end;
1702 addr = vm_unmapped_area(&info);
1708 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1710 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1711 unsigned long len, unsigned long pgoff,
1712 unsigned long flags)
1714 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1719 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1720 unsigned long pgoff, unsigned long flags)
1722 unsigned long (*get_area)(struct file *, unsigned long,
1723 unsigned long, unsigned long, unsigned long);
1725 unsigned long error = arch_mmap_check(addr, len, flags);
1729 /* Careful about overflows.. */
1730 if (len > TASK_SIZE)
1733 get_area = current->mm->get_unmapped_area;
1735 if (file->f_op->get_unmapped_area)
1736 get_area = file->f_op->get_unmapped_area;
1737 } else if (flags & MAP_SHARED) {
1739 * mmap_region() will call shmem_zero_setup() to create a file,
1740 * so use shmem's get_unmapped_area in case it can be huge.
1741 * do_mmap() will clear pgoff, so match alignment.
1744 get_area = shmem_get_unmapped_area;
1747 addr = get_area(file, addr, len, pgoff, flags);
1748 if (IS_ERR_VALUE(addr))
1751 if (addr > TASK_SIZE - len)
1753 if (offset_in_page(addr))
1756 error = security_mmap_addr(addr);
1757 return error ? error : addr;
1760 EXPORT_SYMBOL(get_unmapped_area);
1763 * find_vma_intersection() - Look up the first VMA which intersects the interval
1764 * @mm: The process address space.
1765 * @start_addr: The inclusive start user address.
1766 * @end_addr: The exclusive end user address.
1768 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1769 * start_addr < end_addr.
1771 struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1772 unsigned long start_addr,
1773 unsigned long end_addr)
1775 unsigned long index = start_addr;
1777 mmap_assert_locked(mm);
1778 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1780 EXPORT_SYMBOL(find_vma_intersection);
1783 * find_vma() - Find the VMA for a given address, or the next VMA.
1784 * @mm: The mm_struct to check
1785 * @addr: The address
1787 * Returns: The VMA associated with addr, or the next VMA.
1788 * May return %NULL in the case of no VMA at addr or above.
1790 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1792 unsigned long index = addr;
1794 mmap_assert_locked(mm);
1795 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1797 EXPORT_SYMBOL(find_vma);
1800 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1801 * set %pprev to the previous VMA, if any.
1802 * @mm: The mm_struct to check
1803 * @addr: The address
1804 * @pprev: The pointer to set to the previous VMA
1806 * Note that RCU lock is missing here since the external mmap_lock() is used
1809 * Returns: The VMA associated with @addr, or the next vma.
1810 * May return %NULL in the case of no vma at addr or above.
1812 struct vm_area_struct *
1813 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1814 struct vm_area_struct **pprev)
1816 struct vm_area_struct *vma;
1817 MA_STATE(mas, &mm->mm_mt, addr, addr);
1819 vma = mas_walk(&mas);
1820 *pprev = mas_prev(&mas, 0);
1822 vma = mas_next(&mas, ULONG_MAX);
1827 * Verify that the stack growth is acceptable and
1828 * update accounting. This is shared with both the
1829 * grow-up and grow-down cases.
1831 static int acct_stack_growth(struct vm_area_struct *vma,
1832 unsigned long size, unsigned long grow)
1834 struct mm_struct *mm = vma->vm_mm;
1835 unsigned long new_start;
1837 /* address space limit tests */
1838 if (!may_expand_vm(mm, vma->vm_flags, grow))
1841 /* Stack limit test */
1842 if (size > rlimit(RLIMIT_STACK))
1845 /* mlock limit tests */
1846 if (mlock_future_check(mm, vma->vm_flags, grow << PAGE_SHIFT))
1849 /* Check to ensure the stack will not grow into a hugetlb-only region */
1850 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1852 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1856 * Overcommit.. This must be the final test, as it will
1857 * update security statistics.
1859 if (security_vm_enough_memory_mm(mm, grow))
1865 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1867 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1868 * vma is the last one with address > vma->vm_end. Have to extend vma.
1870 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1872 struct mm_struct *mm = vma->vm_mm;
1873 struct vm_area_struct *next;
1874 unsigned long gap_addr;
1876 MA_STATE(mas, &mm->mm_mt, 0, 0);
1878 if (!(vma->vm_flags & VM_GROWSUP))
1881 /* Guard against exceeding limits of the address space. */
1882 address &= PAGE_MASK;
1883 if (address >= (TASK_SIZE & PAGE_MASK))
1885 address += PAGE_SIZE;
1887 /* Enforce stack_guard_gap */
1888 gap_addr = address + stack_guard_gap;
1890 /* Guard against overflow */
1891 if (gap_addr < address || gap_addr > TASK_SIZE)
1892 gap_addr = TASK_SIZE;
1894 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1895 if (next && vma_is_accessible(next)) {
1896 if (!(next->vm_flags & VM_GROWSUP))
1898 /* Check that both stack segments have the same anon_vma? */
1901 if (mas_preallocate(&mas, GFP_KERNEL))
1904 /* We must make sure the anon_vma is allocated. */
1905 if (unlikely(anon_vma_prepare(vma))) {
1911 * vma->vm_start/vm_end cannot change under us because the caller
1912 * is required to hold the mmap_lock in read mode. We need the
1913 * anon_vma lock to serialize against concurrent expand_stacks.
1915 anon_vma_lock_write(vma->anon_vma);
1917 /* Somebody else might have raced and expanded it already */
1918 if (address > vma->vm_end) {
1919 unsigned long size, grow;
1921 size = address - vma->vm_start;
1922 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1925 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1926 error = acct_stack_growth(vma, size, grow);
1929 * We only hold a shared mmap_lock lock here, so
1930 * we need to protect against concurrent vma
1931 * expansions. anon_vma_lock_write() doesn't
1932 * help here, as we don't guarantee that all
1933 * growable vmas in a mm share the same root
1934 * anon vma. So, we reuse mm->page_table_lock
1935 * to guard against concurrent vma expansions.
1937 spin_lock(&mm->page_table_lock);
1938 if (vma->vm_flags & VM_LOCKED)
1939 mm->locked_vm += grow;
1940 vm_stat_account(mm, vma->vm_flags, grow);
1941 anon_vma_interval_tree_pre_update_vma(vma);
1942 vma->vm_end = address;
1943 /* Overwrite old entry in mtree. */
1944 mas_set_range(&mas, vma->vm_start, address - 1);
1945 mas_store_prealloc(&mas, vma);
1946 anon_vma_interval_tree_post_update_vma(vma);
1947 spin_unlock(&mm->page_table_lock);
1949 perf_event_mmap(vma);
1953 anon_vma_unlock_write(vma->anon_vma);
1954 khugepaged_enter_vma(vma, vma->vm_flags);
1958 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1961 * vma is the first one with address < vma->vm_start. Have to extend vma.
1963 int expand_downwards(struct vm_area_struct *vma, unsigned long address)
1965 struct mm_struct *mm = vma->vm_mm;
1966 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
1967 struct vm_area_struct *prev;
1970 address &= PAGE_MASK;
1971 if (address < mmap_min_addr)
1974 /* Enforce stack_guard_gap */
1975 prev = mas_prev(&mas, 0);
1976 /* Check that both stack segments have the same anon_vma? */
1977 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
1978 vma_is_accessible(prev)) {
1979 if (address - prev->vm_end < stack_guard_gap)
1983 if (mas_preallocate(&mas, GFP_KERNEL))
1986 /* We must make sure the anon_vma is allocated. */
1987 if (unlikely(anon_vma_prepare(vma))) {
1993 * vma->vm_start/vm_end cannot change under us because the caller
1994 * is required to hold the mmap_lock in read mode. We need the
1995 * anon_vma lock to serialize against concurrent expand_stacks.
1997 anon_vma_lock_write(vma->anon_vma);
1999 /* Somebody else might have raced and expanded it already */
2000 if (address < vma->vm_start) {
2001 unsigned long size, grow;
2003 size = vma->vm_end - address;
2004 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2007 if (grow <= vma->vm_pgoff) {
2008 error = acct_stack_growth(vma, size, grow);
2011 * We only hold a shared mmap_lock lock here, so
2012 * we need to protect against concurrent vma
2013 * expansions. anon_vma_lock_write() doesn't
2014 * help here, as we don't guarantee that all
2015 * growable vmas in a mm share the same root
2016 * anon vma. So, we reuse mm->page_table_lock
2017 * to guard against concurrent vma expansions.
2019 spin_lock(&mm->page_table_lock);
2020 if (vma->vm_flags & VM_LOCKED)
2021 mm->locked_vm += grow;
2022 vm_stat_account(mm, vma->vm_flags, grow);
2023 anon_vma_interval_tree_pre_update_vma(vma);
2024 vma->vm_start = address;
2025 vma->vm_pgoff -= grow;
2026 /* Overwrite old entry in mtree. */
2027 mas_set_range(&mas, address, vma->vm_end - 1);
2028 mas_store_prealloc(&mas, vma);
2029 anon_vma_interval_tree_post_update_vma(vma);
2030 spin_unlock(&mm->page_table_lock);
2032 perf_event_mmap(vma);
2036 anon_vma_unlock_write(vma->anon_vma);
2037 khugepaged_enter_vma(vma, vma->vm_flags);
2042 /* enforced gap between the expanding stack and other mappings. */
2043 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2045 static int __init cmdline_parse_stack_guard_gap(char *p)
2050 val = simple_strtoul(p, &endptr, 10);
2052 stack_guard_gap = val << PAGE_SHIFT;
2056 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2058 #ifdef CONFIG_STACK_GROWSUP
2059 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2061 return expand_upwards(vma, address);
2064 struct vm_area_struct *
2065 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2067 struct vm_area_struct *vma, *prev;
2070 vma = find_vma_prev(mm, addr, &prev);
2071 if (vma && (vma->vm_start <= addr))
2073 if (!prev || expand_stack(prev, addr))
2075 if (prev->vm_flags & VM_LOCKED)
2076 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2080 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2082 return expand_downwards(vma, address);
2085 struct vm_area_struct *
2086 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2088 struct vm_area_struct *vma;
2089 unsigned long start;
2092 vma = find_vma(mm, addr);
2095 if (vma->vm_start <= addr)
2097 if (!(vma->vm_flags & VM_GROWSDOWN))
2099 start = vma->vm_start;
2100 if (expand_stack(vma, addr))
2102 if (vma->vm_flags & VM_LOCKED)
2103 populate_vma_page_range(vma, addr, start, NULL);
2108 EXPORT_SYMBOL_GPL(find_extend_vma);
2111 * Ok - we have the memory areas we should free on a maple tree so release them,
2112 * and do the vma updates.
2114 * Called with the mm semaphore held.
2116 static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2118 unsigned long nr_accounted = 0;
2119 struct vm_area_struct *vma;
2121 /* Update high watermark before we lower total_vm */
2122 update_hiwater_vm(mm);
2123 mas_for_each(mas, vma, ULONG_MAX) {
2124 long nrpages = vma_pages(vma);
2126 if (vma->vm_flags & VM_ACCOUNT)
2127 nr_accounted += nrpages;
2128 vm_stat_account(mm, vma->vm_flags, -nrpages);
2131 vm_unacct_memory(nr_accounted);
2136 * Get rid of page table information in the indicated region.
2138 * Called with the mm semaphore held.
2140 static void unmap_region(struct mm_struct *mm, struct maple_tree *mt,
2141 struct vm_area_struct *vma, struct vm_area_struct *prev,
2142 struct vm_area_struct *next,
2143 unsigned long start, unsigned long end, bool mm_wr_locked)
2145 struct mmu_gather tlb;
2148 tlb_gather_mmu(&tlb, mm);
2149 update_hiwater_rss(mm);
2150 unmap_vmas(&tlb, mt, vma, start, end, mm_wr_locked);
2151 free_pgtables(&tlb, mt, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2152 next ? next->vm_start : USER_PGTABLES_CEILING);
2153 tlb_finish_mmu(&tlb);
2157 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2158 * has already been checked or doesn't make sense to fail.
2159 * VMA Iterator will point to the end VMA.
2161 int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2162 unsigned long addr, int new_below)
2164 struct vma_prepare vp;
2165 struct vm_area_struct *new;
2168 validate_mm_mt(vma->vm_mm);
2170 WARN_ON(vma->vm_start >= addr);
2171 WARN_ON(vma->vm_end <= addr);
2173 if (vma->vm_ops && vma->vm_ops->may_split) {
2174 err = vma->vm_ops->may_split(vma, addr);
2179 new = vm_area_dup(vma);
2184 if (vma_iter_prealloc(vmi))
2190 new->vm_start = addr;
2191 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2194 err = vma_dup_policy(vma, new);
2198 err = anon_vma_clone(new, vma);
2203 get_file(new->vm_file);
2205 if (new->vm_ops && new->vm_ops->open)
2206 new->vm_ops->open(new);
2208 vma_adjust_trans_huge(vma, vma->vm_start, addr, 0);
2209 init_vma_prep(&vp, vma);
2214 vma->vm_start = addr;
2215 vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT;
2220 /* vma_complete stores the new vma */
2221 vma_complete(&vp, vmi, vma->vm_mm);
2226 validate_mm_mt(vma->vm_mm);
2230 mpol_put(vma_policy(new));
2235 validate_mm_mt(vma->vm_mm);
2240 * Split a vma into two pieces at address 'addr', a new vma is allocated
2241 * either for the first part or the tail.
2243 int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2244 unsigned long addr, int new_below)
2246 if (vma->vm_mm->map_count >= sysctl_max_map_count)
2249 return __split_vma(vmi, vma, addr, new_below);
2252 static inline int munmap_sidetree(struct vm_area_struct *vma,
2253 struct ma_state *mas_detach)
2255 mas_set_range(mas_detach, vma->vm_start, vma->vm_end - 1);
2256 if (mas_store_gfp(mas_detach, vma, GFP_KERNEL))
2259 if (vma->vm_flags & VM_LOCKED)
2260 vma->vm_mm->locked_vm -= vma_pages(vma);
2266 * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
2267 * @vmi: The vma iterator
2268 * @vma: The starting vm_area_struct
2269 * @mm: The mm_struct
2270 * @start: The aligned start address to munmap.
2271 * @end: The aligned end address to munmap.
2272 * @uf: The userfaultfd list_head
2273 * @downgrade: Set to true to attempt a write downgrade of the mmap_lock
2275 * If @downgrade is true, check return code for potential release of the lock.
2278 do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
2279 struct mm_struct *mm, unsigned long start,
2280 unsigned long end, struct list_head *uf, bool downgrade)
2282 struct vm_area_struct *prev, *next = NULL;
2283 struct maple_tree mt_detach;
2285 int error = -ENOMEM;
2286 MA_STATE(mas_detach, &mt_detach, 0, 0);
2287 mt_init_flags(&mt_detach, MT_FLAGS_LOCK_EXTERN);
2288 mt_set_external_lock(&mt_detach, &mm->mmap_lock);
2291 * If we need to split any vma, do it now to save pain later.
2293 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2294 * unmapped vm_area_struct will remain in use: so lower split_vma
2295 * places tmp vma above, and higher split_vma places tmp vma below.
2298 /* Does it split the first one? */
2299 if (start > vma->vm_start) {
2302 * Make sure that map_count on return from munmap() will
2303 * not exceed its limit; but let map_count go just above
2304 * its limit temporarily, to help free resources as expected.
2306 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2307 goto map_count_exceeded;
2309 error = __split_vma(vmi, vma, start, 0);
2311 goto start_split_failed;
2313 vma = vma_iter_load(vmi);
2316 prev = vma_prev(vmi);
2317 if (unlikely((!prev)))
2318 vma_iter_set(vmi, start);
2321 * Detach a range of VMAs from the mm. Using next as a temp variable as
2322 * it is always overwritten.
2324 for_each_vma_range(*vmi, next, end) {
2325 /* Does it split the end? */
2326 if (next->vm_end > end) {
2327 error = __split_vma(vmi, next, end, 0);
2329 goto end_split_failed;
2331 error = munmap_sidetree(next, &mas_detach);
2333 goto munmap_sidetree_failed;
2336 #ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2337 BUG_ON(next->vm_start < start);
2338 BUG_ON(next->vm_start > end);
2342 next = vma_next(vmi);
2345 * If userfaultfd_unmap_prep returns an error the vmas
2346 * will remain split, but userland will get a
2347 * highly unexpected error anyway. This is no
2348 * different than the case where the first of the two
2349 * __split_vma fails, but we don't undo the first
2350 * split, despite we could. This is unlikely enough
2351 * failure that it's not worth optimizing it for.
2353 error = userfaultfd_unmap_prep(mm, start, end, uf);
2356 goto userfaultfd_error;
2359 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2360 /* Make sure no VMAs are about to be lost. */
2362 MA_STATE(test, &mt_detach, start, end - 1);
2363 struct vm_area_struct *vma_mas, *vma_test;
2366 vma_iter_set(vmi, start);
2368 vma_test = mas_find(&test, end - 1);
2369 for_each_vma_range(*vmi, vma_mas, end) {
2370 BUG_ON(vma_mas != vma_test);
2372 vma_test = mas_next(&test, end - 1);
2375 BUG_ON(count != test_count);
2378 /* Point of no return */
2379 vma_iter_set(vmi, start);
2380 if (vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL))
2383 mm->map_count -= count;
2385 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2386 * VM_GROWSUP VMA. Such VMAs can change their size under
2387 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2390 if (next && (next->vm_flags & VM_GROWSDOWN))
2392 else if (prev && (prev->vm_flags & VM_GROWSUP))
2395 mmap_write_downgrade(mm);
2399 * We can free page tables without write-locking mmap_lock because VMAs
2400 * were isolated before we downgraded mmap_lock.
2402 unmap_region(mm, &mt_detach, vma, prev, next, start, end, !downgrade);
2403 /* Statistics and freeing VMAs */
2404 mas_set(&mas_detach, start);
2405 remove_mt(mm, &mas_detach);
2406 __mt_destroy(&mt_detach);
2410 return downgrade ? 1 : 0;
2413 munmap_sidetree_failed:
2415 __mt_destroy(&mt_detach);
2422 * do_vmi_munmap() - munmap a given range.
2423 * @vmi: The vma iterator
2424 * @mm: The mm_struct
2425 * @start: The start address to munmap
2426 * @len: The length of the range to munmap
2427 * @uf: The userfaultfd list_head
2428 * @downgrade: set to true if the user wants to attempt to write_downgrade the
2431 * This function takes a @mas that is either pointing to the previous VMA or set
2432 * to MA_START and sets it up to remove the mapping(s). The @len will be
2433 * aligned and any arch_unmap work will be preformed.
2435 * Returns: -EINVAL on failure, 1 on success and unlock, 0 otherwise.
2437 int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm,
2438 unsigned long start, size_t len, struct list_head *uf,
2442 struct vm_area_struct *vma;
2444 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2447 end = start + PAGE_ALIGN(len);
2451 /* arch_unmap() might do unmaps itself. */
2452 arch_unmap(mm, start, end);
2454 /* Find the first overlapping VMA */
2455 vma = vma_find(vmi, end);
2459 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, downgrade);
2462 /* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2463 * @mm: The mm_struct
2464 * @start: The start address to munmap
2465 * @len: The length to be munmapped.
2466 * @uf: The userfaultfd list_head
2468 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2469 struct list_head *uf)
2471 VMA_ITERATOR(vmi, mm, start);
2473 return do_vmi_munmap(&vmi, mm, start, len, uf, false);
2476 unsigned long mmap_region(struct file *file, unsigned long addr,
2477 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2478 struct list_head *uf)
2480 struct mm_struct *mm = current->mm;
2481 struct vm_area_struct *vma = NULL;
2482 struct vm_area_struct *next, *prev, *merge;
2483 pgoff_t pglen = len >> PAGE_SHIFT;
2484 unsigned long charged = 0;
2485 unsigned long end = addr + len;
2486 unsigned long merge_start = addr, merge_end = end;
2489 VMA_ITERATOR(vmi, mm, addr);
2491 /* Check against address space limit. */
2492 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2493 unsigned long nr_pages;
2496 * MAP_FIXED may remove pages of mappings that intersects with
2497 * requested mapping. Account for the pages it would unmap.
2499 nr_pages = count_vma_pages_range(mm, addr, end);
2501 if (!may_expand_vm(mm, vm_flags,
2502 (len >> PAGE_SHIFT) - nr_pages))
2506 /* Unmap any existing mapping in the area */
2507 if (do_vmi_munmap(&vmi, mm, addr, len, uf, false))
2511 * Private writable mapping: check memory availability
2513 if (accountable_mapping(file, vm_flags)) {
2514 charged = len >> PAGE_SHIFT;
2515 if (security_vm_enough_memory_mm(mm, charged))
2517 vm_flags |= VM_ACCOUNT;
2520 next = vma_next(&vmi);
2521 prev = vma_prev(&vmi);
2522 if (vm_flags & VM_SPECIAL)
2525 /* Attempt to expand an old mapping */
2527 if (next && next->vm_start == end && !vma_policy(next) &&
2528 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2529 NULL_VM_UFFD_CTX, NULL)) {
2530 merge_end = next->vm_end;
2532 vm_pgoff = next->vm_pgoff - pglen;
2536 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2537 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2538 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2539 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2540 NULL_VM_UFFD_CTX, NULL))) {
2541 merge_start = prev->vm_start;
2543 vm_pgoff = prev->vm_pgoff;
2547 /* Actually expand, if possible */
2549 !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) {
2550 khugepaged_enter_vma(vma, vm_flags);
2556 * Determine the object being mapped and call the appropriate
2557 * specific mapper. the address has already been validated, but
2558 * not unmapped, but the maps are removed from the list.
2560 vma = vm_area_alloc(mm);
2566 vma_iter_set(&vmi, addr);
2567 vma->vm_start = addr;
2569 vm_flags_init(vma, vm_flags);
2570 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2571 vma->vm_pgoff = pgoff;
2574 if (vm_flags & VM_SHARED) {
2575 error = mapping_map_writable(file->f_mapping);
2580 vma->vm_file = get_file(file);
2581 error = call_mmap(file, vma);
2583 goto unmap_and_free_vma;
2586 * Expansion is handled above, merging is handled below.
2587 * Drivers should not alter the address of the VMA.
2590 if (WARN_ON((addr != vma->vm_start)))
2591 goto close_and_free_vma;
2593 vma_iter_set(&vmi, addr);
2595 * If vm_flags changed after call_mmap(), we should try merge
2596 * vma again as we may succeed this time.
2598 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2599 merge = vma_merge(&vmi, mm, prev, vma->vm_start,
2600 vma->vm_end, vma->vm_flags, NULL,
2601 vma->vm_file, vma->vm_pgoff, NULL,
2602 NULL_VM_UFFD_CTX, NULL);
2605 * ->mmap() can change vma->vm_file and fput
2606 * the original file. So fput the vma->vm_file
2607 * here or we would add an extra fput for file
2608 * and cause general protection fault
2614 /* Update vm_flags to pick up the change. */
2615 vm_flags = vma->vm_flags;
2616 goto unmap_writable;
2620 vm_flags = vma->vm_flags;
2621 } else if (vm_flags & VM_SHARED) {
2622 error = shmem_zero_setup(vma);
2626 vma_set_anonymous(vma);
2629 if (map_deny_write_exec(vma, vma->vm_flags)) {
2631 goto close_and_free_vma;
2634 /* Allow architectures to sanity-check the vm_flags */
2636 if (!arch_validate_flags(vma->vm_flags))
2637 goto close_and_free_vma;
2640 if (vma_iter_prealloc(&vmi))
2641 goto close_and_free_vma;
2644 i_mmap_lock_write(vma->vm_file->f_mapping);
2646 vma_iter_store(&vmi, vma);
2649 if (vma->vm_flags & VM_SHARED)
2650 mapping_allow_writable(vma->vm_file->f_mapping);
2652 flush_dcache_mmap_lock(vma->vm_file->f_mapping);
2653 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
2654 flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
2655 i_mmap_unlock_write(vma->vm_file->f_mapping);
2659 * vma_merge() calls khugepaged_enter_vma() either, the below
2660 * call covers the non-merge case.
2662 khugepaged_enter_vma(vma, vma->vm_flags);
2664 /* Once vma denies write, undo our temporary denial count */
2666 if (file && vm_flags & VM_SHARED)
2667 mapping_unmap_writable(file->f_mapping);
2668 file = vma->vm_file;
2670 perf_event_mmap(vma);
2672 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2673 if (vm_flags & VM_LOCKED) {
2674 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2675 is_vm_hugetlb_page(vma) ||
2676 vma == get_gate_vma(current->mm))
2677 vm_flags_clear(vma, VM_LOCKED_MASK);
2679 mm->locked_vm += (len >> PAGE_SHIFT);
2686 * New (or expanded) vma always get soft dirty status.
2687 * Otherwise user-space soft-dirty page tracker won't
2688 * be able to distinguish situation when vma area unmapped,
2689 * then new mapped in-place (which must be aimed as
2690 * a completely new data area).
2692 vm_flags_set(vma, VM_SOFTDIRTY);
2694 vma_set_page_prot(vma);
2700 if (file && vma->vm_ops && vma->vm_ops->close)
2701 vma->vm_ops->close(vma);
2703 if (file || vma->vm_file) {
2706 vma->vm_file = NULL;
2708 /* Undo any partial mapping done by a device driver. */
2709 unmap_region(mm, &mm->mm_mt, vma, prev, next, vma->vm_start,
2712 if (file && (vm_flags & VM_SHARED))
2713 mapping_unmap_writable(file->f_mapping);
2718 vm_unacct_memory(charged);
2723 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2726 struct mm_struct *mm = current->mm;
2728 VMA_ITERATOR(vmi, mm, start);
2730 if (mmap_write_lock_killable(mm))
2733 ret = do_vmi_munmap(&vmi, mm, start, len, &uf, downgrade);
2735 * Returning 1 indicates mmap_lock is downgraded.
2736 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2737 * it to 0 before return.
2740 mmap_read_unlock(mm);
2743 mmap_write_unlock(mm);
2745 userfaultfd_unmap_complete(mm, &uf);
2749 int vm_munmap(unsigned long start, size_t len)
2751 return __vm_munmap(start, len, false);
2753 EXPORT_SYMBOL(vm_munmap);
2755 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2757 addr = untagged_addr(addr);
2758 return __vm_munmap(addr, len, true);
2763 * Emulation of deprecated remap_file_pages() syscall.
2765 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2766 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2769 struct mm_struct *mm = current->mm;
2770 struct vm_area_struct *vma;
2771 unsigned long populate = 0;
2772 unsigned long ret = -EINVAL;
2775 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
2776 current->comm, current->pid);
2780 start = start & PAGE_MASK;
2781 size = size & PAGE_MASK;
2783 if (start + size <= start)
2786 /* Does pgoff wrap? */
2787 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2790 if (mmap_write_lock_killable(mm))
2793 vma = vma_lookup(mm, start);
2795 if (!vma || !(vma->vm_flags & VM_SHARED))
2798 if (start + size > vma->vm_end) {
2799 VMA_ITERATOR(vmi, mm, vma->vm_end);
2800 struct vm_area_struct *next, *prev = vma;
2802 for_each_vma_range(vmi, next, start + size) {
2803 /* hole between vmas ? */
2804 if (next->vm_start != prev->vm_end)
2807 if (next->vm_file != vma->vm_file)
2810 if (next->vm_flags != vma->vm_flags)
2813 if (start + size <= next->vm_end)
2823 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2824 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2825 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2827 flags &= MAP_NONBLOCK;
2828 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2829 if (vma->vm_flags & VM_LOCKED)
2830 flags |= MAP_LOCKED;
2832 file = get_file(vma->vm_file);
2833 ret = do_mmap(vma->vm_file, start, size,
2834 prot, flags, pgoff, &populate, NULL);
2837 mmap_write_unlock(mm);
2839 mm_populate(ret, populate);
2840 if (!IS_ERR_VALUE(ret))
2846 * do_vma_munmap() - Unmap a full or partial vma.
2847 * @vmi: The vma iterator pointing at the vma
2848 * @vma: The first vma to be munmapped
2849 * @start: the start of the address to unmap
2850 * @end: The end of the address to unmap
2851 * @uf: The userfaultfd list_head
2852 * @downgrade: Attempt to downgrade or not
2854 * Returns: 0 on success and not downgraded, 1 on success and downgraded.
2855 * unmaps a VMA mapping when the vma iterator is already in position.
2856 * Does not handle alignment.
2858 int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
2859 unsigned long start, unsigned long end,
2860 struct list_head *uf, bool downgrade)
2862 struct mm_struct *mm = vma->vm_mm;
2865 arch_unmap(mm, start, end);
2866 ret = do_vmi_align_munmap(vmi, vma, mm, start, end, uf, downgrade);
2872 * do_brk_flags() - Increase the brk vma if the flags match.
2873 * @vmi: The vma iterator
2874 * @addr: The start address
2875 * @len: The length of the increase
2877 * @flags: The VMA Flags
2879 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
2880 * do not match then create a new anonymous VMA. Eventually we may be able to
2881 * do some brk-specific accounting here.
2883 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
2884 unsigned long addr, unsigned long len, unsigned long flags)
2886 struct mm_struct *mm = current->mm;
2887 struct vma_prepare vp;
2891 * Check against address space limits by the changed size
2892 * Note: This happens *after* clearing old mappings in some code paths.
2894 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2895 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2898 if (mm->map_count > sysctl_max_map_count)
2901 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2905 * Expand the existing vma if possible; Note that singular lists do not
2906 * occur after forking, so the expand will only happen on new VMAs.
2908 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
2909 can_vma_merge_after(vma, flags, NULL, NULL,
2910 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
2911 if (vma_iter_prealloc(vmi))
2914 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
2915 init_vma_prep(&vp, vma);
2917 vma->vm_end = addr + len;
2918 vm_flags_set(vma, VM_SOFTDIRTY);
2919 vma_iter_store(vmi, vma);
2921 vma_complete(&vp, vmi, mm);
2922 khugepaged_enter_vma(vma, flags);
2926 /* create a vma struct for an anonymous mapping */
2927 vma = vm_area_alloc(mm);
2931 vma_set_anonymous(vma);
2932 vma->vm_start = addr;
2933 vma->vm_end = addr + len;
2934 vma->vm_pgoff = addr >> PAGE_SHIFT;
2935 vm_flags_init(vma, flags);
2936 vma->vm_page_prot = vm_get_page_prot(flags);
2937 if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
2938 goto mas_store_fail;
2942 perf_event_mmap(vma);
2943 mm->total_vm += len >> PAGE_SHIFT;
2944 mm->data_vm += len >> PAGE_SHIFT;
2945 if (flags & VM_LOCKED)
2946 mm->locked_vm += (len >> PAGE_SHIFT);
2947 vm_flags_set(vma, VM_SOFTDIRTY);
2954 vm_unacct_memory(len >> PAGE_SHIFT);
2958 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
2960 struct mm_struct *mm = current->mm;
2961 struct vm_area_struct *vma = NULL;
2966 VMA_ITERATOR(vmi, mm, addr);
2968 len = PAGE_ALIGN(request);
2974 if (mmap_write_lock_killable(mm))
2977 /* Until we need other flags, refuse anything except VM_EXEC. */
2978 if ((flags & (~VM_EXEC)) != 0)
2981 ret = check_brk_limits(addr, len);
2985 ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0);
2989 vma = vma_prev(&vmi);
2990 ret = do_brk_flags(&vmi, vma, addr, len, flags);
2991 populate = ((mm->def_flags & VM_LOCKED) != 0);
2992 mmap_write_unlock(mm);
2993 userfaultfd_unmap_complete(mm, &uf);
2994 if (populate && !ret)
2995 mm_populate(addr, len);
3000 mmap_write_unlock(mm);
3003 EXPORT_SYMBOL(vm_brk_flags);
3005 int vm_brk(unsigned long addr, unsigned long len)
3007 return vm_brk_flags(addr, len, 0);
3009 EXPORT_SYMBOL(vm_brk);
3011 /* Release all mmaps. */
3012 void exit_mmap(struct mm_struct *mm)
3014 struct mmu_gather tlb;
3015 struct vm_area_struct *vma;
3016 unsigned long nr_accounted = 0;
3017 MA_STATE(mas, &mm->mm_mt, 0, 0);
3020 /* mm's last user has gone, and its about to be pulled down */
3021 mmu_notifier_release(mm);
3026 vma = mas_find(&mas, ULONG_MAX);
3028 /* Can happen if dup_mmap() received an OOM */
3029 mmap_read_unlock(mm);
3035 tlb_gather_mmu_fullmm(&tlb, mm);
3036 /* update_hiwater_rss(mm) here? but nobody should be looking */
3037 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3038 unmap_vmas(&tlb, &mm->mm_mt, vma, 0, ULONG_MAX, false);
3039 mmap_read_unlock(mm);
3042 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3043 * because the memory has been already freed.
3045 set_bit(MMF_OOM_SKIP, &mm->flags);
3046 mmap_write_lock(mm);
3047 free_pgtables(&tlb, &mm->mm_mt, vma, FIRST_USER_ADDRESS,
3048 USER_PGTABLES_CEILING);
3049 tlb_finish_mmu(&tlb);
3052 * Walk the list again, actually closing and freeing it, with preemption
3053 * enabled, without holding any MM locks besides the unreachable
3057 if (vma->vm_flags & VM_ACCOUNT)
3058 nr_accounted += vma_pages(vma);
3062 } while ((vma = mas_find(&mas, ULONG_MAX)) != NULL);
3064 BUG_ON(count != mm->map_count);
3066 trace_exit_mmap(mm);
3067 __mt_destroy(&mm->mm_mt);
3068 mmap_write_unlock(mm);
3069 vm_unacct_memory(nr_accounted);
3072 /* Insert vm structure into process list sorted by address
3073 * and into the inode's i_mmap tree. If vm_file is non-NULL
3074 * then i_mmap_rwsem is taken here.
3076 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3078 unsigned long charged = vma_pages(vma);
3081 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3084 if ((vma->vm_flags & VM_ACCOUNT) &&
3085 security_vm_enough_memory_mm(mm, charged))
3089 * The vm_pgoff of a purely anonymous vma should be irrelevant
3090 * until its first write fault, when page's anon_vma and index
3091 * are set. But now set the vm_pgoff it will almost certainly
3092 * end up with (unless mremap moves it elsewhere before that
3093 * first wfault), so /proc/pid/maps tells a consistent story.
3095 * By setting it to reflect the virtual start address of the
3096 * vma, merges and splits can happen in a seamless way, just
3097 * using the existing file pgoff checks and manipulations.
3098 * Similarly in do_mmap and in do_brk_flags.
3100 if (vma_is_anonymous(vma)) {
3101 BUG_ON(vma->anon_vma);
3102 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3105 if (vma_link(mm, vma)) {
3106 vm_unacct_memory(charged);
3114 * Copy the vma structure to a new location in the same mm,
3115 * prior to moving page table entries, to effect an mremap move.
3117 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3118 unsigned long addr, unsigned long len, pgoff_t pgoff,
3119 bool *need_rmap_locks)
3121 struct vm_area_struct *vma = *vmap;
3122 unsigned long vma_start = vma->vm_start;
3123 struct mm_struct *mm = vma->vm_mm;
3124 struct vm_area_struct *new_vma, *prev;
3125 bool faulted_in_anon_vma = true;
3126 VMA_ITERATOR(vmi, mm, addr);
3130 * If anonymous vma has not yet been faulted, update new pgoff
3131 * to match new location, to increase its chance of merging.
3133 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3134 pgoff = addr >> PAGE_SHIFT;
3135 faulted_in_anon_vma = false;
3138 new_vma = find_vma_prev(mm, addr, &prev);
3139 if (new_vma && new_vma->vm_start < addr + len)
3140 return NULL; /* should never get here */
3142 new_vma = vma_merge(&vmi, mm, prev, addr, addr + len, vma->vm_flags,
3143 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3144 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
3147 * Source vma may have been merged into new_vma
3149 if (unlikely(vma_start >= new_vma->vm_start &&
3150 vma_start < new_vma->vm_end)) {
3152 * The only way we can get a vma_merge with
3153 * self during an mremap is if the vma hasn't
3154 * been faulted in yet and we were allowed to
3155 * reset the dst vma->vm_pgoff to the
3156 * destination address of the mremap to allow
3157 * the merge to happen. mremap must change the
3158 * vm_pgoff linearity between src and dst vmas
3159 * (in turn preventing a vma_merge) to be
3160 * safe. It is only safe to keep the vm_pgoff
3161 * linear if there are no pages mapped yet.
3163 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3164 *vmap = vma = new_vma;
3166 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3168 new_vma = vm_area_dup(vma);
3171 new_vma->vm_start = addr;
3172 new_vma->vm_end = addr + len;
3173 new_vma->vm_pgoff = pgoff;
3174 if (vma_dup_policy(vma, new_vma))
3176 if (anon_vma_clone(new_vma, vma))
3177 goto out_free_mempol;
3178 if (new_vma->vm_file)
3179 get_file(new_vma->vm_file);
3180 if (new_vma->vm_ops && new_vma->vm_ops->open)
3181 new_vma->vm_ops->open(new_vma);
3182 if (vma_link(mm, new_vma))
3184 *need_rmap_locks = false;
3190 if (new_vma->vm_ops && new_vma->vm_ops->close)
3191 new_vma->vm_ops->close(new_vma);
3193 if (new_vma->vm_file)
3194 fput(new_vma->vm_file);
3196 unlink_anon_vmas(new_vma);
3198 mpol_put(vma_policy(new_vma));
3200 vm_area_free(new_vma);
3207 * Return true if the calling process may expand its vm space by the passed
3210 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3212 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3215 if (is_data_mapping(flags) &&
3216 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3217 /* Workaround for Valgrind */
3218 if (rlimit(RLIMIT_DATA) == 0 &&
3219 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3222 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3223 current->comm, current->pid,
3224 (mm->data_vm + npages) << PAGE_SHIFT,
3225 rlimit(RLIMIT_DATA),
3226 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3228 if (!ignore_rlimit_data)
3235 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3237 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3239 if (is_exec_mapping(flags))
3240 mm->exec_vm += npages;
3241 else if (is_stack_mapping(flags))
3242 mm->stack_vm += npages;
3243 else if (is_data_mapping(flags))
3244 mm->data_vm += npages;
3247 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3250 * Having a close hook prevents vma merging regardless of flags.
3252 static void special_mapping_close(struct vm_area_struct *vma)
3256 static const char *special_mapping_name(struct vm_area_struct *vma)
3258 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3261 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3263 struct vm_special_mapping *sm = new_vma->vm_private_data;
3265 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3269 return sm->mremap(sm, new_vma);
3274 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3277 * Forbid splitting special mappings - kernel has expectations over
3278 * the number of pages in mapping. Together with VM_DONTEXPAND
3279 * the size of vma should stay the same over the special mapping's
3285 static const struct vm_operations_struct special_mapping_vmops = {
3286 .close = special_mapping_close,
3287 .fault = special_mapping_fault,
3288 .mremap = special_mapping_mremap,
3289 .name = special_mapping_name,
3290 /* vDSO code relies that VVAR can't be accessed remotely */
3292 .may_split = special_mapping_split,
3295 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3296 .close = special_mapping_close,
3297 .fault = special_mapping_fault,
3300 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3302 struct vm_area_struct *vma = vmf->vma;
3304 struct page **pages;
3306 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3307 pages = vma->vm_private_data;
3309 struct vm_special_mapping *sm = vma->vm_private_data;
3312 return sm->fault(sm, vmf->vma, vmf);
3317 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3321 struct page *page = *pages;
3327 return VM_FAULT_SIGBUS;
3330 static struct vm_area_struct *__install_special_mapping(
3331 struct mm_struct *mm,
3332 unsigned long addr, unsigned long len,
3333 unsigned long vm_flags, void *priv,
3334 const struct vm_operations_struct *ops)
3337 struct vm_area_struct *vma;
3340 vma = vm_area_alloc(mm);
3341 if (unlikely(vma == NULL))
3342 return ERR_PTR(-ENOMEM);
3344 vma->vm_start = addr;
3345 vma->vm_end = addr + len;
3347 vm_flags_init(vma, (vm_flags | mm->def_flags |
3348 VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK);
3349 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3352 vma->vm_private_data = priv;
3354 ret = insert_vm_struct(mm, vma);
3358 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3360 perf_event_mmap(vma);
3368 return ERR_PTR(ret);
3371 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3372 const struct vm_special_mapping *sm)
3374 return vma->vm_private_data == sm &&
3375 (vma->vm_ops == &special_mapping_vmops ||
3376 vma->vm_ops == &legacy_special_mapping_vmops);
3380 * Called with mm->mmap_lock held for writing.
3381 * Insert a new vma covering the given region, with the given flags.
3382 * Its pages are supplied by the given array of struct page *.
3383 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3384 * The region past the last page supplied will always produce SIGBUS.
3385 * The array pointer and the pages it points to are assumed to stay alive
3386 * for as long as this mapping might exist.
3388 struct vm_area_struct *_install_special_mapping(
3389 struct mm_struct *mm,
3390 unsigned long addr, unsigned long len,
3391 unsigned long vm_flags, const struct vm_special_mapping *spec)
3393 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3394 &special_mapping_vmops);
3397 int install_special_mapping(struct mm_struct *mm,
3398 unsigned long addr, unsigned long len,
3399 unsigned long vm_flags, struct page **pages)
3401 struct vm_area_struct *vma = __install_special_mapping(
3402 mm, addr, len, vm_flags, (void *)pages,
3403 &legacy_special_mapping_vmops);
3405 return PTR_ERR_OR_ZERO(vma);
3408 static DEFINE_MUTEX(mm_all_locks_mutex);
3410 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3412 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3414 * The LSB of head.next can't change from under us
3415 * because we hold the mm_all_locks_mutex.
3417 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3419 * We can safely modify head.next after taking the
3420 * anon_vma->root->rwsem. If some other vma in this mm shares
3421 * the same anon_vma we won't take it again.
3423 * No need of atomic instructions here, head.next
3424 * can't change from under us thanks to the
3425 * anon_vma->root->rwsem.
3427 if (__test_and_set_bit(0, (unsigned long *)
3428 &anon_vma->root->rb_root.rb_root.rb_node))
3433 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3435 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3437 * AS_MM_ALL_LOCKS can't change from under us because
3438 * we hold the mm_all_locks_mutex.
3440 * Operations on ->flags have to be atomic because
3441 * even if AS_MM_ALL_LOCKS is stable thanks to the
3442 * mm_all_locks_mutex, there may be other cpus
3443 * changing other bitflags in parallel to us.
3445 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3447 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3452 * This operation locks against the VM for all pte/vma/mm related
3453 * operations that could ever happen on a certain mm. This includes
3454 * vmtruncate, try_to_unmap, and all page faults.
3456 * The caller must take the mmap_lock in write mode before calling
3457 * mm_take_all_locks(). The caller isn't allowed to release the
3458 * mmap_lock until mm_drop_all_locks() returns.
3460 * mmap_lock in write mode is required in order to block all operations
3461 * that could modify pagetables and free pages without need of
3462 * altering the vma layout. It's also needed in write mode to avoid new
3463 * anon_vmas to be associated with existing vmas.
3465 * A single task can't take more than one mm_take_all_locks() in a row
3466 * or it would deadlock.
3468 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3469 * mapping->flags avoid to take the same lock twice, if more than one
3470 * vma in this mm is backed by the same anon_vma or address_space.
3472 * We take locks in following order, accordingly to comment at beginning
3474 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3476 * - all i_mmap_rwsem locks;
3477 * - all anon_vma->rwseml
3479 * We can take all locks within these types randomly because the VM code
3480 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3481 * mm_all_locks_mutex.
3483 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3484 * that may have to take thousand of locks.
3486 * mm_take_all_locks() can fail if it's interrupted by signals.
3488 int mm_take_all_locks(struct mm_struct *mm)
3490 struct vm_area_struct *vma;
3491 struct anon_vma_chain *avc;
3492 MA_STATE(mas, &mm->mm_mt, 0, 0);
3494 mmap_assert_write_locked(mm);
3496 mutex_lock(&mm_all_locks_mutex);
3498 mas_for_each(&mas, vma, ULONG_MAX) {
3499 if (signal_pending(current))
3501 if (vma->vm_file && vma->vm_file->f_mapping &&
3502 is_vm_hugetlb_page(vma))
3503 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3507 mas_for_each(&mas, vma, ULONG_MAX) {
3508 if (signal_pending(current))
3510 if (vma->vm_file && vma->vm_file->f_mapping &&
3511 !is_vm_hugetlb_page(vma))
3512 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3516 mas_for_each(&mas, vma, ULONG_MAX) {
3517 if (signal_pending(current))
3520 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3521 vm_lock_anon_vma(mm, avc->anon_vma);
3527 mm_drop_all_locks(mm);
3531 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3533 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3535 * The LSB of head.next can't change to 0 from under
3536 * us because we hold the mm_all_locks_mutex.
3538 * We must however clear the bitflag before unlocking
3539 * the vma so the users using the anon_vma->rb_root will
3540 * never see our bitflag.
3542 * No need of atomic instructions here, head.next
3543 * can't change from under us until we release the
3544 * anon_vma->root->rwsem.
3546 if (!__test_and_clear_bit(0, (unsigned long *)
3547 &anon_vma->root->rb_root.rb_root.rb_node))
3549 anon_vma_unlock_write(anon_vma);
3553 static void vm_unlock_mapping(struct address_space *mapping)
3555 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3557 * AS_MM_ALL_LOCKS can't change to 0 from under us
3558 * because we hold the mm_all_locks_mutex.
3560 i_mmap_unlock_write(mapping);
3561 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3568 * The mmap_lock cannot be released by the caller until
3569 * mm_drop_all_locks() returns.
3571 void mm_drop_all_locks(struct mm_struct *mm)
3573 struct vm_area_struct *vma;
3574 struct anon_vma_chain *avc;
3575 MA_STATE(mas, &mm->mm_mt, 0, 0);
3577 mmap_assert_write_locked(mm);
3578 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3580 mas_for_each(&mas, vma, ULONG_MAX) {
3582 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3583 vm_unlock_anon_vma(avc->anon_vma);
3584 if (vma->vm_file && vma->vm_file->f_mapping)
3585 vm_unlock_mapping(vma->vm_file->f_mapping);
3588 mutex_unlock(&mm_all_locks_mutex);
3592 * initialise the percpu counter for VM
3594 void __init mmap_init(void)
3598 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3603 * Initialise sysctl_user_reserve_kbytes.
3605 * This is intended to prevent a user from starting a single memory hogging
3606 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3609 * The default value is min(3% of free memory, 128MB)
3610 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3612 static int init_user_reserve(void)
3614 unsigned long free_kbytes;
3616 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3618 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3621 subsys_initcall(init_user_reserve);
3624 * Initialise sysctl_admin_reserve_kbytes.
3626 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3627 * to log in and kill a memory hogging process.
3629 * Systems with more than 256MB will reserve 8MB, enough to recover
3630 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3631 * only reserve 3% of free pages by default.
3633 static int init_admin_reserve(void)
3635 unsigned long free_kbytes;
3637 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3639 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3642 subsys_initcall(init_admin_reserve);
3645 * Reinititalise user and admin reserves if memory is added or removed.
3647 * The default user reserve max is 128MB, and the default max for the
3648 * admin reserve is 8MB. These are usually, but not always, enough to
3649 * enable recovery from a memory hogging process using login/sshd, a shell,
3650 * and tools like top. It may make sense to increase or even disable the
3651 * reserve depending on the existence of swap or variations in the recovery
3652 * tools. So, the admin may have changed them.
3654 * If memory is added and the reserves have been eliminated or increased above
3655 * the default max, then we'll trust the admin.
3657 * If memory is removed and there isn't enough free memory, then we
3658 * need to reset the reserves.
3660 * Otherwise keep the reserve set by the admin.
3662 static int reserve_mem_notifier(struct notifier_block *nb,
3663 unsigned long action, void *data)
3665 unsigned long tmp, free_kbytes;
3669 /* Default max is 128MB. Leave alone if modified by operator. */
3670 tmp = sysctl_user_reserve_kbytes;
3671 if (0 < tmp && tmp < (1UL << 17))
3672 init_user_reserve();
3674 /* Default max is 8MB. Leave alone if modified by operator. */
3675 tmp = sysctl_admin_reserve_kbytes;
3676 if (0 < tmp && tmp < (1UL << 13))
3677 init_admin_reserve();
3681 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3683 if (sysctl_user_reserve_kbytes > free_kbytes) {
3684 init_user_reserve();
3685 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3686 sysctl_user_reserve_kbytes);
3689 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3690 init_admin_reserve();
3691 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3692 sysctl_admin_reserve_kbytes);
3701 static int __meminit init_reserve_notifier(void)
3703 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3704 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3708 subsys_initcall(init_reserve_notifier);