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>
49 #include <linux/ksm.h>
51 #include <linux/uaccess.h>
52 #include <asm/cacheflush.h>
54 #include <asm/mmu_context.h>
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/mmap.h>
61 #ifndef arch_mmap_check
62 #define arch_mmap_check(addr, len, flags) (0)
65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
66 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
67 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
68 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
70 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
71 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
72 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
73 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
76 static bool ignore_rlimit_data;
77 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
79 static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
80 struct vm_area_struct *vma, struct vm_area_struct *prev,
81 struct vm_area_struct *next, unsigned long start,
82 unsigned long end, unsigned long tree_end, bool mm_wr_locked);
84 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
86 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
89 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
90 void vma_set_page_prot(struct vm_area_struct *vma)
92 unsigned long vm_flags = vma->vm_flags;
93 pgprot_t vm_page_prot;
95 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
96 if (vma_wants_writenotify(vma, vm_page_prot)) {
97 vm_flags &= ~VM_SHARED;
98 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
100 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
101 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
105 * Requires inode->i_mapping->i_mmap_rwsem
107 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
108 struct file *file, struct address_space *mapping)
110 if (vma->vm_flags & VM_SHARED)
111 mapping_unmap_writable(mapping);
113 flush_dcache_mmap_lock(mapping);
114 vma_interval_tree_remove(vma, &mapping->i_mmap);
115 flush_dcache_mmap_unlock(mapping);
119 * Unlink a file-based vm structure from its interval tree, to hide
120 * vma from rmap and vmtruncate before freeing its page tables.
122 void unlink_file_vma(struct vm_area_struct *vma)
124 struct file *file = vma->vm_file;
127 struct address_space *mapping = file->f_mapping;
128 i_mmap_lock_write(mapping);
129 __remove_shared_vm_struct(vma, file, mapping);
130 i_mmap_unlock_write(mapping);
135 * Close a vm structure and free it.
137 static void remove_vma(struct vm_area_struct *vma, bool unreachable)
140 if (vma->vm_ops && vma->vm_ops->close)
141 vma->vm_ops->close(vma);
144 mpol_put(vma_policy(vma));
151 static inline struct vm_area_struct *vma_prev_limit(struct vma_iterator *vmi,
154 return mas_prev(&vmi->mas, min);
158 * check_brk_limits() - Use platform specific check of range & verify mlock
160 * @addr: The address to check
161 * @len: The size of increase.
163 * Return: 0 on success.
165 static int check_brk_limits(unsigned long addr, unsigned long len)
167 unsigned long mapped_addr;
169 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
170 if (IS_ERR_VALUE(mapped_addr))
173 return mlock_future_ok(current->mm, current->mm->def_flags, len)
176 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma,
177 unsigned long addr, unsigned long request, unsigned long flags);
178 SYSCALL_DEFINE1(brk, unsigned long, brk)
180 unsigned long newbrk, oldbrk, origbrk;
181 struct mm_struct *mm = current->mm;
182 struct vm_area_struct *brkvma, *next = NULL;
183 unsigned long min_brk;
184 bool populate = false;
186 struct vma_iterator vmi;
188 if (mmap_write_lock_killable(mm))
193 #ifdef CONFIG_COMPAT_BRK
195 * CONFIG_COMPAT_BRK can still be overridden by setting
196 * randomize_va_space to 2, which will still cause mm->start_brk
197 * to be arbitrarily shifted
199 if (current->brk_randomized)
200 min_brk = mm->start_brk;
202 min_brk = mm->end_data;
204 min_brk = mm->start_brk;
210 * Check against rlimit here. If this check is done later after the test
211 * of oldbrk with newbrk then it can escape the test and let the data
212 * segment grow beyond its set limit the in case where the limit is
213 * not page aligned -Ram Gupta
215 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
216 mm->end_data, mm->start_data))
219 newbrk = PAGE_ALIGN(brk);
220 oldbrk = PAGE_ALIGN(mm->brk);
221 if (oldbrk == newbrk) {
226 /* Always allow shrinking brk. */
227 if (brk <= mm->brk) {
228 /* Search one past newbrk */
229 vma_iter_init(&vmi, mm, newbrk);
230 brkvma = vma_find(&vmi, oldbrk);
231 if (!brkvma || brkvma->vm_start >= oldbrk)
232 goto out; /* mapping intersects with an existing non-brk vma. */
234 * mm->brk must be protected by write mmap_lock.
235 * do_vma_munmap() will drop the lock on success, so update it
236 * before calling do_vma_munmap().
239 if (do_vma_munmap(&vmi, brkvma, newbrk, oldbrk, &uf, true))
242 goto success_unlocked;
245 if (check_brk_limits(oldbrk, newbrk - oldbrk))
249 * Only check if the next VMA is within the stack_guard_gap of the
252 vma_iter_init(&vmi, mm, oldbrk);
253 next = vma_find(&vmi, newbrk + PAGE_SIZE + stack_guard_gap);
254 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
257 brkvma = vma_prev_limit(&vmi, mm->start_brk);
258 /* Ok, looks good - let it rip. */
259 if (do_brk_flags(&vmi, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
263 if (mm->def_flags & VM_LOCKED)
267 mmap_write_unlock(mm);
269 userfaultfd_unmap_complete(mm, &uf);
271 mm_populate(oldbrk, newbrk - oldbrk);
276 mmap_write_unlock(mm);
280 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
281 static void validate_mm(struct mm_struct *mm)
285 struct vm_area_struct *vma;
286 VMA_ITERATOR(vmi, mm, 0);
288 mt_validate(&mm->mm_mt);
289 for_each_vma(vmi, vma) {
290 #ifdef CONFIG_DEBUG_VM_RB
291 struct anon_vma *anon_vma = vma->anon_vma;
292 struct anon_vma_chain *avc;
294 unsigned long vmi_start, vmi_end;
297 vmi_start = vma_iter_addr(&vmi);
298 vmi_end = vma_iter_end(&vmi);
299 if (VM_WARN_ON_ONCE_MM(vma->vm_end != vmi_end, mm))
302 if (VM_WARN_ON_ONCE_MM(vma->vm_start != vmi_start, mm))
306 pr_emerg("issue in %s\n", current->comm);
309 pr_emerg("tree range: %px start %lx end %lx\n", vma,
310 vmi_start, vmi_end - 1);
311 vma_iter_dump_tree(&vmi);
314 #ifdef CONFIG_DEBUG_VM_RB
316 anon_vma_lock_read(anon_vma);
317 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
318 anon_vma_interval_tree_verify(avc);
319 anon_vma_unlock_read(anon_vma);
324 if (i != mm->map_count) {
325 pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);
328 VM_BUG_ON_MM(bug, mm);
331 #else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
332 #define validate_mm(mm) do { } while (0)
333 #endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
336 * vma has some anon_vma assigned, and is already inserted on that
337 * anon_vma's interval trees.
339 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
340 * vma must be removed from the anon_vma's interval trees using
341 * anon_vma_interval_tree_pre_update_vma().
343 * After the update, the vma will be reinserted using
344 * anon_vma_interval_tree_post_update_vma().
346 * The entire update must be protected by exclusive mmap_lock and by
347 * the root anon_vma's mutex.
350 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
352 struct anon_vma_chain *avc;
354 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
355 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
359 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
361 struct anon_vma_chain *avc;
363 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
364 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
367 static unsigned long count_vma_pages_range(struct mm_struct *mm,
368 unsigned long addr, unsigned long end)
370 VMA_ITERATOR(vmi, mm, addr);
371 struct vm_area_struct *vma;
372 unsigned long nr_pages = 0;
374 for_each_vma_range(vmi, vma, end) {
375 unsigned long vm_start = max(addr, vma->vm_start);
376 unsigned long vm_end = min(end, vma->vm_end);
378 nr_pages += PHYS_PFN(vm_end - vm_start);
384 static void __vma_link_file(struct vm_area_struct *vma,
385 struct address_space *mapping)
387 if (vma->vm_flags & VM_SHARED)
388 mapping_allow_writable(mapping);
390 flush_dcache_mmap_lock(mapping);
391 vma_interval_tree_insert(vma, &mapping->i_mmap);
392 flush_dcache_mmap_unlock(mapping);
395 static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
397 VMA_ITERATOR(vmi, mm, 0);
398 struct address_space *mapping = NULL;
400 vma_iter_config(&vmi, vma->vm_start, vma->vm_end);
401 if (vma_iter_prealloc(&vmi, vma))
404 vma_start_write(vma);
406 vma_iter_store(&vmi, vma);
409 mapping = vma->vm_file->f_mapping;
410 i_mmap_lock_write(mapping);
411 __vma_link_file(vma, mapping);
412 i_mmap_unlock_write(mapping);
421 * init_multi_vma_prep() - Initializer for struct vma_prepare
422 * @vp: The vma_prepare struct
423 * @vma: The vma that will be altered once locked
424 * @next: The next vma if it is to be adjusted
425 * @remove: The first vma to be removed
426 * @remove2: The second vma to be removed
428 static inline void init_multi_vma_prep(struct vma_prepare *vp,
429 struct vm_area_struct *vma, struct vm_area_struct *next,
430 struct vm_area_struct *remove, struct vm_area_struct *remove2)
432 memset(vp, 0, sizeof(struct vma_prepare));
434 vp->anon_vma = vma->anon_vma;
436 vp->remove2 = remove2;
438 if (!vp->anon_vma && next)
439 vp->anon_vma = next->anon_vma;
441 vp->file = vma->vm_file;
443 vp->mapping = vma->vm_file->f_mapping;
448 * init_vma_prep() - Initializer wrapper for vma_prepare struct
449 * @vp: The vma_prepare struct
450 * @vma: The vma that will be altered once locked
452 static inline void init_vma_prep(struct vma_prepare *vp,
453 struct vm_area_struct *vma)
455 init_multi_vma_prep(vp, vma, NULL, NULL, NULL);
460 * vma_prepare() - Helper function for handling locking VMAs prior to altering
461 * @vp: The initialized vma_prepare struct
463 static inline void vma_prepare(struct vma_prepare *vp)
466 uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end);
469 uprobe_munmap(vp->adj_next, vp->adj_next->vm_start,
470 vp->adj_next->vm_end);
472 i_mmap_lock_write(vp->mapping);
473 if (vp->insert && vp->insert->vm_file) {
475 * Put into interval tree now, so instantiated pages
476 * are visible to arm/parisc __flush_dcache_page
477 * throughout; but we cannot insert into address
478 * space until vma start or end is updated.
480 __vma_link_file(vp->insert,
481 vp->insert->vm_file->f_mapping);
486 anon_vma_lock_write(vp->anon_vma);
487 anon_vma_interval_tree_pre_update_vma(vp->vma);
489 anon_vma_interval_tree_pre_update_vma(vp->adj_next);
493 flush_dcache_mmap_lock(vp->mapping);
494 vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap);
496 vma_interval_tree_remove(vp->adj_next,
497 &vp->mapping->i_mmap);
503 * vma_complete- Helper function for handling the unlocking after altering VMAs,
504 * or for inserting a VMA.
506 * @vp: The vma_prepare struct
507 * @vmi: The vma iterator
510 static inline void vma_complete(struct vma_prepare *vp,
511 struct vma_iterator *vmi, struct mm_struct *mm)
515 vma_interval_tree_insert(vp->adj_next,
516 &vp->mapping->i_mmap);
517 vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap);
518 flush_dcache_mmap_unlock(vp->mapping);
521 if (vp->remove && vp->file) {
522 __remove_shared_vm_struct(vp->remove, vp->file, vp->mapping);
524 __remove_shared_vm_struct(vp->remove2, vp->file,
526 } else if (vp->insert) {
528 * split_vma has split insert from vma, and needs
529 * us to insert it before dropping the locks
530 * (it may either follow vma or precede it).
532 vma_iter_store(vmi, vp->insert);
537 anon_vma_interval_tree_post_update_vma(vp->vma);
539 anon_vma_interval_tree_post_update_vma(vp->adj_next);
540 anon_vma_unlock_write(vp->anon_vma);
544 i_mmap_unlock_write(vp->mapping);
545 uprobe_mmap(vp->vma);
548 uprobe_mmap(vp->adj_next);
553 vma_mark_detached(vp->remove, true);
555 uprobe_munmap(vp->remove, vp->remove->vm_start,
559 if (vp->remove->anon_vma)
560 anon_vma_merge(vp->vma, vp->remove);
562 mpol_put(vma_policy(vp->remove));
564 WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end);
565 vm_area_free(vp->remove);
568 * In mprotect's case 6 (see comments on vma_merge),
569 * we are removing both mid and next vmas
572 vp->remove = vp->remove2;
577 if (vp->insert && vp->file)
578 uprobe_mmap(vp->insert);
583 * dup_anon_vma() - Helper function to duplicate anon_vma
584 * @dst: The destination VMA
585 * @src: The source VMA
587 * Returns: 0 on success.
589 static inline int dup_anon_vma(struct vm_area_struct *dst,
590 struct vm_area_struct *src)
593 * Easily overlooked: when mprotect shifts the boundary, make sure the
594 * expanding vma has anon_vma set if the shrinking vma had, to cover any
595 * anon pages imported.
597 if (src->anon_vma && !dst->anon_vma) {
598 vma_assert_write_locked(dst);
599 dst->anon_vma = src->anon_vma;
600 return anon_vma_clone(dst, src);
607 * vma_expand - Expand an existing VMA
609 * @vmi: The vma iterator
610 * @vma: The vma to expand
611 * @start: The start of the vma
612 * @end: The exclusive end of the vma
613 * @pgoff: The page offset of vma
614 * @next: The current of next vma.
616 * Expand @vma to @start and @end. Can expand off the start and end. Will
617 * expand over @next if it's different from @vma and @end == @next->vm_end.
618 * Checking if the @vma can expand and merge with @next needs to be handled by
621 * Returns: 0 on success
623 int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
624 unsigned long start, unsigned long end, pgoff_t pgoff,
625 struct vm_area_struct *next)
627 bool remove_next = false;
628 struct vma_prepare vp;
630 vma_start_write(vma);
631 if (next && (vma != next) && (end == next->vm_end)) {
635 vma_start_write(next);
636 ret = dup_anon_vma(vma, next);
641 init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL);
642 /* Not merging but overwriting any part of next is not handled. */
643 VM_WARN_ON(next && !vp.remove &&
644 next != vma && end > next->vm_start);
645 /* Only handles expanding */
646 VM_WARN_ON(vma->vm_start < start || vma->vm_end > end);
648 /* Note: vma iterator must be pointing to 'start' */
649 vma_iter_config(vmi, start, end);
650 if (vma_iter_prealloc(vmi, vma))
654 vma_adjust_trans_huge(vma, start, end, 0);
655 vma->vm_start = start;
657 vma->vm_pgoff = pgoff;
658 vma_iter_store(vmi, vma);
660 vma_complete(&vp, vmi, vma->vm_mm);
668 * vma_shrink() - Reduce an existing VMAs memory area
669 * @vmi: The vma iterator
670 * @vma: The VMA to modify
671 * @start: The new start
674 * Returns: 0 on success, -ENOMEM otherwise
676 int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
677 unsigned long start, unsigned long end, pgoff_t pgoff)
679 struct vma_prepare vp;
681 WARN_ON((vma->vm_start != start) && (vma->vm_end != end));
683 if (vma->vm_start < start)
684 vma_iter_config(vmi, vma->vm_start, start);
686 vma_iter_config(vmi, end, vma->vm_end);
688 if (vma_iter_prealloc(vmi, NULL))
691 vma_start_write(vma);
693 init_vma_prep(&vp, vma);
695 vma_adjust_trans_huge(vma, start, end, 0);
698 vma->vm_start = start;
700 vma->vm_pgoff = pgoff;
701 vma_complete(&vp, vmi, vma->vm_mm);
706 * If the vma has a ->close operation then the driver probably needs to release
707 * per-vma resources, so we don't attempt to merge those if the caller indicates
708 * the current vma may be removed as part of the merge.
710 static inline bool is_mergeable_vma(struct vm_area_struct *vma,
711 struct file *file, unsigned long vm_flags,
712 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
713 struct anon_vma_name *anon_name, bool may_remove_vma)
716 * VM_SOFTDIRTY should not prevent from VMA merging, if we
717 * match the flags but dirty bit -- the caller should mark
718 * merged VMA as dirty. If dirty bit won't be excluded from
719 * comparison, we increase pressure on the memory system forcing
720 * the kernel to generate new VMAs when old one could be
723 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
725 if (vma->vm_file != file)
727 if (may_remove_vma && vma->vm_ops && vma->vm_ops->close)
729 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
731 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
736 static inline bool is_mergeable_anon_vma(struct anon_vma *anon_vma1,
737 struct anon_vma *anon_vma2, struct vm_area_struct *vma)
740 * The list_is_singular() test is to avoid merging VMA cloned from
741 * parents. This can improve scalability caused by anon_vma lock.
743 if ((!anon_vma1 || !anon_vma2) && (!vma ||
744 list_is_singular(&vma->anon_vma_chain)))
746 return anon_vma1 == anon_vma2;
750 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
751 * in front of (at a lower virtual address and file offset than) the vma.
753 * We cannot merge two vmas if they have differently assigned (non-NULL)
754 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
756 * We don't check here for the merged mmap wrapping around the end of pagecache
757 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
758 * wrap, nor mmaps which cover the final page at index -1UL.
760 * We assume the vma may be removed as part of the merge.
763 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
764 struct anon_vma *anon_vma, struct file *file,
765 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
766 struct anon_vma_name *anon_name)
768 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, true) &&
769 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
770 if (vma->vm_pgoff == vm_pgoff)
777 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
778 * beyond (at a higher virtual address and file offset than) the vma.
780 * We cannot merge two vmas if they have differently assigned (non-NULL)
781 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
783 * We assume that vma is not removed as part of the merge.
786 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
787 struct anon_vma *anon_vma, struct file *file,
788 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
789 struct anon_vma_name *anon_name)
791 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, false) &&
792 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
794 vm_pglen = vma_pages(vma);
795 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
802 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
803 * figure out whether that can be merged with its predecessor or its
804 * successor. Or both (it neatly fills a hole).
806 * In most cases - when called for mmap, brk or mremap - [addr,end) is
807 * certain not to be mapped by the time vma_merge is called; but when
808 * called for mprotect, it is certain to be already mapped (either at
809 * an offset within prev, or at the start of next), and the flags of
810 * this area are about to be changed to vm_flags - and the no-change
811 * case has already been eliminated.
813 * The following mprotect cases have to be considered, where **** is
814 * the area passed down from mprotect_fixup, never extending beyond one
815 * vma, PPPP is the previous vma, CCCC is a concurrent vma that starts
816 * at the same address as **** and is of the same or larger span, and
817 * NNNN the next vma after ****:
820 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPCCCCCC
821 * cannot merge might become might become
822 * PPNNNNNNNNNN PPPPPPPPPPCC
823 * mmap, brk or case 4 below case 5 below
826 * PPPP NNNN PPPPCCCCNNNN
827 * might become might become
828 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
829 * PPPPPPPPNNNN 2 or PPPPPPPPNNNN 7 or
830 * PPPPNNNNNNNN 3 PPPPNNNNNNNN 8
832 * It is important for case 8 that the vma CCCC overlapping the
833 * region **** is never going to extended over NNNN. Instead NNNN must
834 * be extended in region **** and CCCC must be removed. This way in
835 * all cases where vma_merge succeeds, the moment vma_merge drops the
836 * rmap_locks, the properties of the merged vma will be already
837 * correct for the whole merged range. Some of those properties like
838 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
839 * be correct for the whole merged range immediately after the
840 * rmap_locks are released. Otherwise if NNNN would be removed and
841 * CCCC would be extended over the NNNN range, remove_migration_ptes
842 * or other rmap walkers (if working on addresses beyond the "end"
843 * parameter) may establish ptes with the wrong permissions of CCCC
844 * instead of the right permissions of NNNN.
847 * PPPP is represented by *prev
848 * CCCC is represented by *curr or not represented at all (NULL)
849 * NNNN is represented by *next or not represented at all (NULL)
850 * **** is not represented - it will be merged and the vma containing the
851 * area is returned, or the function will return NULL
853 struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
854 struct vm_area_struct *prev, unsigned long addr,
855 unsigned long end, unsigned long vm_flags,
856 struct anon_vma *anon_vma, struct file *file,
857 pgoff_t pgoff, struct mempolicy *policy,
858 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
859 struct anon_vma_name *anon_name)
861 struct vm_area_struct *curr, *next, *res;
862 struct vm_area_struct *vma, *adjust, *remove, *remove2;
863 struct vma_prepare vp;
866 bool merge_prev = false;
867 bool merge_next = false;
868 bool vma_expanded = false;
869 unsigned long vma_start = addr;
870 unsigned long vma_end = end;
871 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
875 * We later require that vma->vm_flags == vm_flags,
876 * so this tests vma->vm_flags & VM_SPECIAL, too.
878 if (vm_flags & VM_SPECIAL)
881 /* Does the input range span an existing VMA? (cases 5 - 8) */
882 curr = find_vma_intersection(mm, prev ? prev->vm_end : 0, end);
884 if (!curr || /* cases 1 - 4 */
885 end == curr->vm_end) /* cases 6 - 8, adjacent VMA */
886 next = vma_lookup(mm, end);
888 next = NULL; /* case 5 */
891 vma_start = prev->vm_start;
892 vma_pgoff = prev->vm_pgoff;
894 /* Can we merge the predecessor? */
895 if (addr == prev->vm_end && mpol_equal(vma_policy(prev), policy)
896 && can_vma_merge_after(prev, vm_flags, anon_vma, file,
897 pgoff, vm_userfaultfd_ctx, anon_name)) {
903 /* Can we merge the successor? */
904 if (next && mpol_equal(policy, vma_policy(next)) &&
905 can_vma_merge_before(next, vm_flags, anon_vma, file, pgoff+pglen,
906 vm_userfaultfd_ctx, anon_name)) {
910 /* Verify some invariant that must be enforced by the caller. */
911 VM_WARN_ON(prev && addr <= prev->vm_start);
912 VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
913 VM_WARN_ON(addr >= end);
915 if (!merge_prev && !merge_next)
916 return NULL; /* Not mergeable. */
919 vma_start_write(prev);
922 remove = remove2 = adjust = NULL;
924 /* Can we merge both the predecessor and the successor? */
925 if (merge_prev && merge_next &&
926 is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) {
927 vma_start_write(next);
928 remove = next; /* case 1 */
929 vma_end = next->vm_end;
930 err = dup_anon_vma(prev, next);
931 if (curr) { /* case 6 */
932 vma_start_write(curr);
936 err = dup_anon_vma(prev, curr);
938 } else if (merge_prev) { /* case 2 */
940 vma_start_write(curr);
941 err = dup_anon_vma(prev, curr);
942 if (end == curr->vm_end) { /* case 7 */
944 } else { /* case 5 */
946 adj_start = (end - curr->vm_start);
949 } else { /* merge_next */
950 vma_start_write(next);
952 if (prev && addr < prev->vm_end) { /* case 4 */
953 vma_start_write(prev);
956 adj_start = -(prev->vm_end - addr);
957 err = dup_anon_vma(next, prev);
960 * Note that cases 3 and 8 are the ONLY ones where prev
961 * is permitted to be (but is not necessarily) NULL.
963 vma = next; /* case 3 */
965 vma_end = next->vm_end;
966 vma_pgoff = next->vm_pgoff - pglen;
967 if (curr) { /* case 8 */
968 vma_pgoff = curr->vm_pgoff;
969 vma_start_write(curr);
971 err = dup_anon_vma(next, curr);
976 /* Error in anon_vma clone. */
980 if (vma_start < vma->vm_start || vma_end > vma->vm_end)
984 vma_iter_config(vmi, vma_start, vma_end);
986 vma_iter_config(vmi, adjust->vm_start + adj_start,
990 if (vma_iter_prealloc(vmi, vma))
993 init_multi_vma_prep(&vp, vma, adjust, remove, remove2);
994 VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
995 vp.anon_vma != adjust->anon_vma);
998 vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start);
1000 vma->vm_start = vma_start;
1001 vma->vm_end = vma_end;
1002 vma->vm_pgoff = vma_pgoff;
1005 vma_iter_store(vmi, vma);
1008 adjust->vm_start += adj_start;
1009 adjust->vm_pgoff += adj_start >> PAGE_SHIFT;
1010 if (adj_start < 0) {
1011 WARN_ON(vma_expanded);
1012 vma_iter_store(vmi, next);
1016 vma_complete(&vp, vmi, mm);
1017 khugepaged_enter_vma(res, vm_flags);
1022 vma_iter_set(vmi, addr);
1028 * Rough compatibility check to quickly see if it's even worth looking
1029 * at sharing an anon_vma.
1031 * They need to have the same vm_file, and the flags can only differ
1032 * in things that mprotect may change.
1034 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1035 * we can merge the two vma's. For example, we refuse to merge a vma if
1036 * there is a vm_ops->close() function, because that indicates that the
1037 * driver is doing some kind of reference counting. But that doesn't
1038 * really matter for the anon_vma sharing case.
1040 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1042 return a->vm_end == b->vm_start &&
1043 mpol_equal(vma_policy(a), vma_policy(b)) &&
1044 a->vm_file == b->vm_file &&
1045 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1046 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1050 * Do some basic sanity checking to see if we can re-use the anon_vma
1051 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1052 * the same as 'old', the other will be the new one that is trying
1053 * to share the anon_vma.
1055 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1056 * the anon_vma of 'old' is concurrently in the process of being set up
1057 * by another page fault trying to merge _that_. But that's ok: if it
1058 * is being set up, that automatically means that it will be a singleton
1059 * acceptable for merging, so we can do all of this optimistically. But
1060 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1062 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1063 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1064 * is to return an anon_vma that is "complex" due to having gone through
1067 * We also make sure that the two vma's are compatible (adjacent,
1068 * and with the same memory policies). That's all stable, even with just
1069 * a read lock on the mmap_lock.
1071 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1073 if (anon_vma_compatible(a, b)) {
1074 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1076 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1083 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1084 * neighbouring vmas for a suitable anon_vma, before it goes off
1085 * to allocate a new anon_vma. It checks because a repetitive
1086 * sequence of mprotects and faults may otherwise lead to distinct
1087 * anon_vmas being allocated, preventing vma merge in subsequent
1090 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1092 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1093 struct anon_vma *anon_vma = NULL;
1094 struct vm_area_struct *prev, *next;
1096 /* Try next first. */
1097 next = mas_walk(&mas);
1099 anon_vma = reusable_anon_vma(next, vma, next);
1104 prev = mas_prev(&mas, 0);
1105 VM_BUG_ON_VMA(prev != vma, vma);
1106 prev = mas_prev(&mas, 0);
1107 /* Try prev next. */
1109 anon_vma = reusable_anon_vma(prev, prev, vma);
1112 * We might reach here with anon_vma == NULL if we can't find
1113 * any reusable anon_vma.
1114 * There's no absolute need to look only at touching neighbours:
1115 * we could search further afield for "compatible" anon_vmas.
1116 * But it would probably just be a waste of time searching,
1117 * or lead to too many vmas hanging off the same anon_vma.
1118 * We're trying to allow mprotect remerging later on,
1119 * not trying to minimize memory used for anon_vmas.
1125 * If a hint addr is less than mmap_min_addr change hint to be as
1126 * low as possible but still greater than mmap_min_addr
1128 static inline unsigned long round_hint_to_min(unsigned long hint)
1131 if (((void *)hint != NULL) &&
1132 (hint < mmap_min_addr))
1133 return PAGE_ALIGN(mmap_min_addr);
1137 bool mlock_future_ok(struct mm_struct *mm, unsigned long flags,
1138 unsigned long bytes)
1140 unsigned long locked_pages, limit_pages;
1142 if (!(flags & VM_LOCKED) || capable(CAP_IPC_LOCK))
1145 locked_pages = bytes >> PAGE_SHIFT;
1146 locked_pages += mm->locked_vm;
1148 limit_pages = rlimit(RLIMIT_MEMLOCK);
1149 limit_pages >>= PAGE_SHIFT;
1151 return locked_pages <= limit_pages;
1154 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1156 if (S_ISREG(inode->i_mode))
1157 return MAX_LFS_FILESIZE;
1159 if (S_ISBLK(inode->i_mode))
1160 return MAX_LFS_FILESIZE;
1162 if (S_ISSOCK(inode->i_mode))
1163 return MAX_LFS_FILESIZE;
1165 /* Special "we do even unsigned file positions" case */
1166 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1169 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1173 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1174 unsigned long pgoff, unsigned long len)
1176 u64 maxsize = file_mmap_size_max(file, inode);
1178 if (maxsize && len > maxsize)
1181 if (pgoff > maxsize >> PAGE_SHIFT)
1187 * The caller must write-lock current->mm->mmap_lock.
1189 unsigned long do_mmap(struct file *file, unsigned long addr,
1190 unsigned long len, unsigned long prot,
1191 unsigned long flags, vm_flags_t vm_flags,
1192 unsigned long pgoff, unsigned long *populate,
1193 struct list_head *uf)
1195 struct mm_struct *mm = current->mm;
1204 * Does the application expect PROT_READ to imply PROT_EXEC?
1206 * (the exception is when the underlying filesystem is noexec
1207 * mounted, in which case we dont add PROT_EXEC.)
1209 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1210 if (!(file && path_noexec(&file->f_path)))
1213 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1214 if (flags & MAP_FIXED_NOREPLACE)
1217 if (!(flags & MAP_FIXED))
1218 addr = round_hint_to_min(addr);
1220 /* Careful about overflows.. */
1221 len = PAGE_ALIGN(len);
1225 /* offset overflow? */
1226 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1229 /* Too many mappings? */
1230 if (mm->map_count > sysctl_max_map_count)
1233 /* Obtain the address to map to. we verify (or select) it and ensure
1234 * that it represents a valid section of the address space.
1236 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1237 if (IS_ERR_VALUE(addr))
1240 if (flags & MAP_FIXED_NOREPLACE) {
1241 if (find_vma_intersection(mm, addr, addr + len))
1245 if (prot == PROT_EXEC) {
1246 pkey = execute_only_pkey(mm);
1251 /* Do simple checking here so the lower-level routines won't have
1252 * to. we assume access permissions have been handled by the open
1253 * of the memory object, so we don't do any here.
1255 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1256 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1258 if (flags & MAP_LOCKED)
1259 if (!can_do_mlock())
1262 if (!mlock_future_ok(mm, vm_flags, len))
1266 struct inode *inode = file_inode(file);
1267 unsigned long flags_mask;
1269 if (!file_mmap_ok(file, inode, pgoff, len))
1272 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1274 switch (flags & MAP_TYPE) {
1277 * Force use of MAP_SHARED_VALIDATE with non-legacy
1278 * flags. E.g. MAP_SYNC is dangerous to use with
1279 * MAP_SHARED as you don't know which consistency model
1280 * you will get. We silently ignore unsupported flags
1281 * with MAP_SHARED to preserve backward compatibility.
1283 flags &= LEGACY_MAP_MASK;
1285 case MAP_SHARED_VALIDATE:
1286 if (flags & ~flags_mask)
1288 if (prot & PROT_WRITE) {
1289 if (!(file->f_mode & FMODE_WRITE))
1291 if (IS_SWAPFILE(file->f_mapping->host))
1296 * Make sure we don't allow writing to an append-only
1299 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1302 vm_flags |= VM_SHARED | VM_MAYSHARE;
1303 if (!(file->f_mode & FMODE_WRITE))
1304 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1307 if (!(file->f_mode & FMODE_READ))
1309 if (path_noexec(&file->f_path)) {
1310 if (vm_flags & VM_EXEC)
1312 vm_flags &= ~VM_MAYEXEC;
1315 if (!file->f_op->mmap)
1317 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1325 switch (flags & MAP_TYPE) {
1327 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1333 vm_flags |= VM_SHARED | VM_MAYSHARE;
1337 * Set pgoff according to addr for anon_vma.
1339 pgoff = addr >> PAGE_SHIFT;
1347 * Set 'VM_NORESERVE' if we should not account for the
1348 * memory use of this mapping.
1350 if (flags & MAP_NORESERVE) {
1351 /* We honor MAP_NORESERVE if allowed to overcommit */
1352 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1353 vm_flags |= VM_NORESERVE;
1355 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1356 if (file && is_file_hugepages(file))
1357 vm_flags |= VM_NORESERVE;
1360 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1361 if (!IS_ERR_VALUE(addr) &&
1362 ((vm_flags & VM_LOCKED) ||
1363 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1368 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1369 unsigned long prot, unsigned long flags,
1370 unsigned long fd, unsigned long pgoff)
1372 struct file *file = NULL;
1373 unsigned long retval;
1375 if (!(flags & MAP_ANONYMOUS)) {
1376 audit_mmap_fd(fd, flags);
1380 if (is_file_hugepages(file)) {
1381 len = ALIGN(len, huge_page_size(hstate_file(file)));
1382 } else if (unlikely(flags & MAP_HUGETLB)) {
1386 } else if (flags & MAP_HUGETLB) {
1389 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1393 len = ALIGN(len, huge_page_size(hs));
1395 * VM_NORESERVE is used because the reservations will be
1396 * taken when vm_ops->mmap() is called
1398 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1400 HUGETLB_ANONHUGE_INODE,
1401 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1403 return PTR_ERR(file);
1406 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1413 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1414 unsigned long, prot, unsigned long, flags,
1415 unsigned long, fd, unsigned long, pgoff)
1417 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1420 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1421 struct mmap_arg_struct {
1425 unsigned long flags;
1427 unsigned long offset;
1430 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1432 struct mmap_arg_struct a;
1434 if (copy_from_user(&a, arg, sizeof(a)))
1436 if (offset_in_page(a.offset))
1439 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1440 a.offset >> PAGE_SHIFT);
1442 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1444 static bool vm_ops_needs_writenotify(const struct vm_operations_struct *vm_ops)
1446 return vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite);
1449 static bool vma_is_shared_writable(struct vm_area_struct *vma)
1451 return (vma->vm_flags & (VM_WRITE | VM_SHARED)) ==
1452 (VM_WRITE | VM_SHARED);
1455 static bool vma_fs_can_writeback(struct vm_area_struct *vma)
1457 /* No managed pages to writeback. */
1458 if (vma->vm_flags & VM_PFNMAP)
1461 return vma->vm_file && vma->vm_file->f_mapping &&
1462 mapping_can_writeback(vma->vm_file->f_mapping);
1466 * Does this VMA require the underlying folios to have their dirty state
1469 bool vma_needs_dirty_tracking(struct vm_area_struct *vma)
1471 /* Only shared, writable VMAs require dirty tracking. */
1472 if (!vma_is_shared_writable(vma))
1475 /* Does the filesystem need to be notified? */
1476 if (vm_ops_needs_writenotify(vma->vm_ops))
1480 * Even if the filesystem doesn't indicate a need for writenotify, if it
1481 * can writeback, dirty tracking is still required.
1483 return vma_fs_can_writeback(vma);
1487 * Some shared mappings will want the pages marked read-only
1488 * to track write events. If so, we'll downgrade vm_page_prot
1489 * to the private version (using protection_map[] without the
1492 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1494 /* If it was private or non-writable, the write bit is already clear */
1495 if (!vma_is_shared_writable(vma))
1498 /* The backer wishes to know when pages are first written to? */
1499 if (vm_ops_needs_writenotify(vma->vm_ops))
1502 /* The open routine did something to the protections that pgprot_modify
1503 * won't preserve? */
1504 if (pgprot_val(vm_page_prot) !=
1505 pgprot_val(vm_pgprot_modify(vm_page_prot, vma->vm_flags)))
1509 * Do we need to track softdirty? hugetlb does not support softdirty
1512 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1515 /* Do we need write faults for uffd-wp tracking? */
1516 if (userfaultfd_wp(vma))
1519 /* Can the mapping track the dirty pages? */
1520 return vma_fs_can_writeback(vma);
1524 * We account for memory if it's a private writeable mapping,
1525 * not hugepages and VM_NORESERVE wasn't set.
1527 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1530 * hugetlb has its own accounting separate from the core VM
1531 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1533 if (file && is_file_hugepages(file))
1536 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1540 * unmapped_area() - Find an area between the low_limit and the high_limit with
1541 * the correct alignment and offset, all from @info. Note: current->mm is used
1544 * @info: The unmapped area information including the range [low_limit -
1545 * high_limit), the alignment offset and mask.
1547 * Return: A memory address or -ENOMEM.
1549 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1551 unsigned long length, gap;
1552 unsigned long low_limit, high_limit;
1553 struct vm_area_struct *tmp;
1555 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1557 /* Adjust search length to account for worst case alignment overhead */
1558 length = info->length + info->align_mask;
1559 if (length < info->length)
1562 low_limit = info->low_limit;
1563 if (low_limit < mmap_min_addr)
1564 low_limit = mmap_min_addr;
1565 high_limit = info->high_limit;
1567 if (mas_empty_area(&mas, low_limit, high_limit - 1, length))
1571 gap += (info->align_offset - gap) & info->align_mask;
1572 tmp = mas_next(&mas, ULONG_MAX);
1573 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1574 if (vm_start_gap(tmp) < gap + length - 1) {
1575 low_limit = tmp->vm_end;
1580 tmp = mas_prev(&mas, 0);
1581 if (tmp && vm_end_gap(tmp) > gap) {
1582 low_limit = vm_end_gap(tmp);
1592 * unmapped_area_topdown() - Find an area between the low_limit and the
1593 * high_limit with the correct alignment and offset at the highest available
1594 * address, all from @info. Note: current->mm is used for the search.
1596 * @info: The unmapped area information including the range [low_limit -
1597 * high_limit), the alignment offset and mask.
1599 * Return: A memory address or -ENOMEM.
1601 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1603 unsigned long length, gap, gap_end;
1604 unsigned long low_limit, high_limit;
1605 struct vm_area_struct *tmp;
1607 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1608 /* Adjust search length to account for worst case alignment overhead */
1609 length = info->length + info->align_mask;
1610 if (length < info->length)
1613 low_limit = info->low_limit;
1614 if (low_limit < mmap_min_addr)
1615 low_limit = mmap_min_addr;
1616 high_limit = info->high_limit;
1618 if (mas_empty_area_rev(&mas, low_limit, high_limit - 1, length))
1621 gap = mas.last + 1 - info->length;
1622 gap -= (gap - info->align_offset) & info->align_mask;
1624 tmp = mas_next(&mas, ULONG_MAX);
1625 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1626 if (vm_start_gap(tmp) <= gap_end) {
1627 high_limit = vm_start_gap(tmp);
1632 tmp = mas_prev(&mas, 0);
1633 if (tmp && vm_end_gap(tmp) > gap) {
1634 high_limit = tmp->vm_start;
1644 * Search for an unmapped address range.
1646 * We are looking for a range that:
1647 * - does not intersect with any VMA;
1648 * - is contained within the [low_limit, high_limit) interval;
1649 * - is at least the desired size.
1650 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1652 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1656 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1657 addr = unmapped_area_topdown(info);
1659 addr = unmapped_area(info);
1661 trace_vm_unmapped_area(addr, info);
1665 /* Get an address range which is currently unmapped.
1666 * For shmat() with addr=0.
1668 * Ugly calling convention alert:
1669 * Return value with the low bits set means error value,
1671 * if (ret & ~PAGE_MASK)
1674 * This function "knows" that -ENOMEM has the bits set.
1677 generic_get_unmapped_area(struct file *filp, unsigned long addr,
1678 unsigned long len, unsigned long pgoff,
1679 unsigned long flags)
1681 struct mm_struct *mm = current->mm;
1682 struct vm_area_struct *vma, *prev;
1683 struct vm_unmapped_area_info info;
1684 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1686 if (len > mmap_end - mmap_min_addr)
1689 if (flags & MAP_FIXED)
1693 addr = PAGE_ALIGN(addr);
1694 vma = find_vma_prev(mm, addr, &prev);
1695 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1696 (!vma || addr + len <= vm_start_gap(vma)) &&
1697 (!prev || addr >= vm_end_gap(prev)))
1703 info.low_limit = mm->mmap_base;
1704 info.high_limit = mmap_end;
1705 info.align_mask = 0;
1706 info.align_offset = 0;
1707 return vm_unmapped_area(&info);
1710 #ifndef HAVE_ARCH_UNMAPPED_AREA
1712 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1713 unsigned long len, unsigned long pgoff,
1714 unsigned long flags)
1716 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1721 * This mmap-allocator allocates new areas top-down from below the
1722 * stack's low limit (the base):
1725 generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1726 unsigned long len, unsigned long pgoff,
1727 unsigned long flags)
1729 struct vm_area_struct *vma, *prev;
1730 struct mm_struct *mm = current->mm;
1731 struct vm_unmapped_area_info info;
1732 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1734 /* requested length too big for entire address space */
1735 if (len > mmap_end - mmap_min_addr)
1738 if (flags & MAP_FIXED)
1741 /* requesting a specific address */
1743 addr = PAGE_ALIGN(addr);
1744 vma = find_vma_prev(mm, addr, &prev);
1745 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1746 (!vma || addr + len <= vm_start_gap(vma)) &&
1747 (!prev || addr >= vm_end_gap(prev)))
1751 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1753 info.low_limit = PAGE_SIZE;
1754 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1755 info.align_mask = 0;
1756 info.align_offset = 0;
1757 addr = vm_unmapped_area(&info);
1760 * A failed mmap() very likely causes application failure,
1761 * so fall back to the bottom-up function here. This scenario
1762 * can happen with large stack limits and large mmap()
1765 if (offset_in_page(addr)) {
1766 VM_BUG_ON(addr != -ENOMEM);
1768 info.low_limit = TASK_UNMAPPED_BASE;
1769 info.high_limit = mmap_end;
1770 addr = vm_unmapped_area(&info);
1776 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1778 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1779 unsigned long len, unsigned long pgoff,
1780 unsigned long flags)
1782 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1787 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1788 unsigned long pgoff, unsigned long flags)
1790 unsigned long (*get_area)(struct file *, unsigned long,
1791 unsigned long, unsigned long, unsigned long);
1793 unsigned long error = arch_mmap_check(addr, len, flags);
1797 /* Careful about overflows.. */
1798 if (len > TASK_SIZE)
1801 get_area = current->mm->get_unmapped_area;
1803 if (file->f_op->get_unmapped_area)
1804 get_area = file->f_op->get_unmapped_area;
1805 } else if (flags & MAP_SHARED) {
1807 * mmap_region() will call shmem_zero_setup() to create a file,
1808 * so use shmem's get_unmapped_area in case it can be huge.
1809 * do_mmap() will clear pgoff, so match alignment.
1812 get_area = shmem_get_unmapped_area;
1815 addr = get_area(file, addr, len, pgoff, flags);
1816 if (IS_ERR_VALUE(addr))
1819 if (addr > TASK_SIZE - len)
1821 if (offset_in_page(addr))
1824 error = security_mmap_addr(addr);
1825 return error ? error : addr;
1828 EXPORT_SYMBOL(get_unmapped_area);
1831 * find_vma_intersection() - Look up the first VMA which intersects the interval
1832 * @mm: The process address space.
1833 * @start_addr: The inclusive start user address.
1834 * @end_addr: The exclusive end user address.
1836 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1837 * start_addr < end_addr.
1839 struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1840 unsigned long start_addr,
1841 unsigned long end_addr)
1843 unsigned long index = start_addr;
1845 mmap_assert_locked(mm);
1846 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1848 EXPORT_SYMBOL(find_vma_intersection);
1851 * find_vma() - Find the VMA for a given address, or the next VMA.
1852 * @mm: The mm_struct to check
1853 * @addr: The address
1855 * Returns: The VMA associated with addr, or the next VMA.
1856 * May return %NULL in the case of no VMA at addr or above.
1858 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1860 unsigned long index = addr;
1862 mmap_assert_locked(mm);
1863 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1865 EXPORT_SYMBOL(find_vma);
1868 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1869 * set %pprev to the previous VMA, if any.
1870 * @mm: The mm_struct to check
1871 * @addr: The address
1872 * @pprev: The pointer to set to the previous VMA
1874 * Note that RCU lock is missing here since the external mmap_lock() is used
1877 * Returns: The VMA associated with @addr, or the next vma.
1878 * May return %NULL in the case of no vma at addr or above.
1880 struct vm_area_struct *
1881 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1882 struct vm_area_struct **pprev)
1884 struct vm_area_struct *vma;
1885 MA_STATE(mas, &mm->mm_mt, addr, addr);
1887 vma = mas_walk(&mas);
1888 *pprev = mas_prev(&mas, 0);
1890 vma = mas_next(&mas, ULONG_MAX);
1895 * Verify that the stack growth is acceptable and
1896 * update accounting. This is shared with both the
1897 * grow-up and grow-down cases.
1899 static int acct_stack_growth(struct vm_area_struct *vma,
1900 unsigned long size, unsigned long grow)
1902 struct mm_struct *mm = vma->vm_mm;
1903 unsigned long new_start;
1905 /* address space limit tests */
1906 if (!may_expand_vm(mm, vma->vm_flags, grow))
1909 /* Stack limit test */
1910 if (size > rlimit(RLIMIT_STACK))
1913 /* mlock limit tests */
1914 if (!mlock_future_ok(mm, vma->vm_flags, grow << PAGE_SHIFT))
1917 /* Check to ensure the stack will not grow into a hugetlb-only region */
1918 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1920 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1924 * Overcommit.. This must be the final test, as it will
1925 * update security statistics.
1927 if (security_vm_enough_memory_mm(mm, grow))
1933 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1935 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1936 * vma is the last one with address > vma->vm_end. Have to extend vma.
1938 static int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1940 struct mm_struct *mm = vma->vm_mm;
1941 struct vm_area_struct *next;
1942 unsigned long gap_addr;
1944 MA_STATE(mas, &mm->mm_mt, vma->vm_start, address);
1946 if (!(vma->vm_flags & VM_GROWSUP))
1949 /* Guard against exceeding limits of the address space. */
1950 address &= PAGE_MASK;
1951 if (address >= (TASK_SIZE & PAGE_MASK))
1953 address += PAGE_SIZE;
1955 /* Enforce stack_guard_gap */
1956 gap_addr = address + stack_guard_gap;
1958 /* Guard against overflow */
1959 if (gap_addr < address || gap_addr > TASK_SIZE)
1960 gap_addr = TASK_SIZE;
1962 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1963 if (next && vma_is_accessible(next)) {
1964 if (!(next->vm_flags & VM_GROWSUP))
1966 /* Check that both stack segments have the same anon_vma? */
1970 mas_prev_range(&mas, address);
1972 __mas_set_range(&mas, vma->vm_start, address - 1);
1973 if (mas_preallocate(&mas, vma, GFP_KERNEL))
1976 /* We must make sure the anon_vma is allocated. */
1977 if (unlikely(anon_vma_prepare(vma))) {
1982 /* Lock the VMA before expanding to prevent concurrent page faults */
1983 vma_start_write(vma);
1985 * vma->vm_start/vm_end cannot change under us because the caller
1986 * is required to hold the mmap_lock in read mode. We need the
1987 * anon_vma lock to serialize against concurrent expand_stacks.
1989 anon_vma_lock_write(vma->anon_vma);
1991 /* Somebody else might have raced and expanded it already */
1992 if (address > vma->vm_end) {
1993 unsigned long size, grow;
1995 size = address - vma->vm_start;
1996 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1999 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2000 error = acct_stack_growth(vma, size, grow);
2003 * We only hold a shared mmap_lock lock here, so
2004 * we need to protect against concurrent vma
2005 * expansions. anon_vma_lock_write() doesn't
2006 * help here, as we don't guarantee that all
2007 * growable vmas in a mm share the same root
2008 * anon vma. So, we reuse mm->page_table_lock
2009 * to guard against concurrent vma expansions.
2011 spin_lock(&mm->page_table_lock);
2012 if (vma->vm_flags & VM_LOCKED)
2013 mm->locked_vm += grow;
2014 vm_stat_account(mm, vma->vm_flags, grow);
2015 anon_vma_interval_tree_pre_update_vma(vma);
2016 vma->vm_end = address;
2017 /* Overwrite old entry in mtree. */
2018 mas_store_prealloc(&mas, vma);
2019 anon_vma_interval_tree_post_update_vma(vma);
2020 spin_unlock(&mm->page_table_lock);
2022 perf_event_mmap(vma);
2026 anon_vma_unlock_write(vma->anon_vma);
2027 khugepaged_enter_vma(vma, vma->vm_flags);
2032 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2035 * vma is the first one with address < vma->vm_start. Have to extend vma.
2036 * mmap_lock held for writing.
2038 int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2040 struct mm_struct *mm = vma->vm_mm;
2041 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2042 struct vm_area_struct *prev;
2045 if (!(vma->vm_flags & VM_GROWSDOWN))
2048 address &= PAGE_MASK;
2049 if (address < mmap_min_addr || address < FIRST_USER_ADDRESS)
2052 /* Enforce stack_guard_gap */
2053 prev = mas_prev(&mas, 0);
2054 /* Check that both stack segments have the same anon_vma? */
2056 if (!(prev->vm_flags & VM_GROWSDOWN) &&
2057 vma_is_accessible(prev) &&
2058 (address - prev->vm_end < stack_guard_gap))
2063 mas_next_range(&mas, vma->vm_start);
2065 __mas_set_range(&mas, address, vma->vm_end - 1);
2066 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2069 /* We must make sure the anon_vma is allocated. */
2070 if (unlikely(anon_vma_prepare(vma))) {
2075 /* Lock the VMA before expanding to prevent concurrent page faults */
2076 vma_start_write(vma);
2078 * vma->vm_start/vm_end cannot change under us because the caller
2079 * is required to hold the mmap_lock in read mode. We need the
2080 * anon_vma lock to serialize against concurrent expand_stacks.
2082 anon_vma_lock_write(vma->anon_vma);
2084 /* Somebody else might have raced and expanded it already */
2085 if (address < vma->vm_start) {
2086 unsigned long size, grow;
2088 size = vma->vm_end - address;
2089 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2092 if (grow <= vma->vm_pgoff) {
2093 error = acct_stack_growth(vma, size, grow);
2096 * We only hold a shared mmap_lock lock here, so
2097 * we need to protect against concurrent vma
2098 * expansions. anon_vma_lock_write() doesn't
2099 * help here, as we don't guarantee that all
2100 * growable vmas in a mm share the same root
2101 * anon vma. So, we reuse mm->page_table_lock
2102 * to guard against concurrent vma expansions.
2104 spin_lock(&mm->page_table_lock);
2105 if (vma->vm_flags & VM_LOCKED)
2106 mm->locked_vm += grow;
2107 vm_stat_account(mm, vma->vm_flags, grow);
2108 anon_vma_interval_tree_pre_update_vma(vma);
2109 vma->vm_start = address;
2110 vma->vm_pgoff -= grow;
2111 /* Overwrite old entry in mtree. */
2112 mas_store_prealloc(&mas, vma);
2113 anon_vma_interval_tree_post_update_vma(vma);
2114 spin_unlock(&mm->page_table_lock);
2116 perf_event_mmap(vma);
2120 anon_vma_unlock_write(vma->anon_vma);
2121 khugepaged_enter_vma(vma, vma->vm_flags);
2127 /* enforced gap between the expanding stack and other mappings. */
2128 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2130 static int __init cmdline_parse_stack_guard_gap(char *p)
2135 val = simple_strtoul(p, &endptr, 10);
2137 stack_guard_gap = val << PAGE_SHIFT;
2141 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2143 #ifdef CONFIG_STACK_GROWSUP
2144 int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2146 return expand_upwards(vma, address);
2149 struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2151 struct vm_area_struct *vma, *prev;
2154 vma = find_vma_prev(mm, addr, &prev);
2155 if (vma && (vma->vm_start <= addr))
2159 if (expand_stack_locked(prev, addr))
2161 if (prev->vm_flags & VM_LOCKED)
2162 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2166 int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2168 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN)))
2170 return expand_downwards(vma, address);
2173 struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2175 struct vm_area_struct *vma;
2176 unsigned long start;
2179 vma = find_vma(mm, addr);
2182 if (vma->vm_start <= addr)
2184 start = vma->vm_start;
2185 if (expand_stack_locked(vma, addr))
2187 if (vma->vm_flags & VM_LOCKED)
2188 populate_vma_page_range(vma, addr, start, NULL);
2194 * IA64 has some horrid mapping rules: it can expand both up and down,
2195 * but with various special rules.
2197 * We'll get rid of this architecture eventually, so the ugliness is
2201 static inline bool vma_expand_ok(struct vm_area_struct *vma, unsigned long addr)
2203 return REGION_NUMBER(addr) == REGION_NUMBER(vma->vm_start) &&
2204 REGION_OFFSET(addr) < RGN_MAP_LIMIT;
2208 * IA64 stacks grow down, but there's a special register backing store
2209 * that can grow up. Only sequentially, though, so the new address must
2212 static inline int vma_expand_up(struct vm_area_struct *vma, unsigned long addr)
2214 if (!vma_expand_ok(vma, addr))
2216 if (vma->vm_end != (addr & PAGE_MASK))
2218 return expand_upwards(vma, addr);
2221 static inline bool vma_expand_down(struct vm_area_struct *vma, unsigned long addr)
2223 if (!vma_expand_ok(vma, addr))
2225 return expand_downwards(vma, addr);
2228 #elif defined(CONFIG_STACK_GROWSUP)
2230 #define vma_expand_up(vma,addr) expand_upwards(vma, addr)
2231 #define vma_expand_down(vma, addr) (-EFAULT)
2235 #define vma_expand_up(vma,addr) (-EFAULT)
2236 #define vma_expand_down(vma, addr) expand_downwards(vma, addr)
2241 * expand_stack(): legacy interface for page faulting. Don't use unless
2244 * This is called with the mm locked for reading, drops the lock, takes
2245 * the lock for writing, tries to look up a vma again, expands it if
2246 * necessary, and downgrades the lock to reading again.
2248 * If no vma is found or it can't be expanded, it returns NULL and has
2251 struct vm_area_struct *expand_stack(struct mm_struct *mm, unsigned long addr)
2253 struct vm_area_struct *vma, *prev;
2255 mmap_read_unlock(mm);
2256 if (mmap_write_lock_killable(mm))
2259 vma = find_vma_prev(mm, addr, &prev);
2260 if (vma && vma->vm_start <= addr)
2263 if (prev && !vma_expand_up(prev, addr)) {
2268 if (vma && !vma_expand_down(vma, addr))
2271 mmap_write_unlock(mm);
2275 mmap_write_downgrade(mm);
2280 * Ok - we have the memory areas we should free on a maple tree so release them,
2281 * and do the vma updates.
2283 * Called with the mm semaphore held.
2285 static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2287 unsigned long nr_accounted = 0;
2288 struct vm_area_struct *vma;
2290 /* Update high watermark before we lower total_vm */
2291 update_hiwater_vm(mm);
2292 mas_for_each(mas, vma, ULONG_MAX) {
2293 long nrpages = vma_pages(vma);
2295 if (vma->vm_flags & VM_ACCOUNT)
2296 nr_accounted += nrpages;
2297 vm_stat_account(mm, vma->vm_flags, -nrpages);
2298 remove_vma(vma, false);
2300 vm_unacct_memory(nr_accounted);
2304 * Get rid of page table information in the indicated region.
2306 * Called with the mm semaphore held.
2308 static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
2309 struct vm_area_struct *vma, struct vm_area_struct *prev,
2310 struct vm_area_struct *next, unsigned long start,
2311 unsigned long end, unsigned long tree_end, bool mm_wr_locked)
2313 struct mmu_gather tlb;
2314 unsigned long mt_start = mas->index;
2317 tlb_gather_mmu(&tlb, mm);
2318 update_hiwater_rss(mm);
2319 unmap_vmas(&tlb, mas, vma, start, end, tree_end, mm_wr_locked);
2320 mas_set(mas, mt_start);
2321 free_pgtables(&tlb, mas, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2322 next ? next->vm_start : USER_PGTABLES_CEILING,
2324 tlb_finish_mmu(&tlb);
2328 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2329 * has already been checked or doesn't make sense to fail.
2330 * VMA Iterator will point to the end VMA.
2332 int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2333 unsigned long addr, int new_below)
2335 struct vma_prepare vp;
2336 struct vm_area_struct *new;
2339 WARN_ON(vma->vm_start >= addr);
2340 WARN_ON(vma->vm_end <= addr);
2342 if (vma->vm_ops && vma->vm_ops->may_split) {
2343 err = vma->vm_ops->may_split(vma, addr);
2348 new = vm_area_dup(vma);
2355 new->vm_start = addr;
2356 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2360 vma_iter_config(vmi, new->vm_start, new->vm_end);
2361 if (vma_iter_prealloc(vmi, new))
2364 err = vma_dup_policy(vma, new);
2368 err = anon_vma_clone(new, vma);
2373 get_file(new->vm_file);
2375 if (new->vm_ops && new->vm_ops->open)
2376 new->vm_ops->open(new);
2378 vma_start_write(vma);
2379 vma_start_write(new);
2381 init_vma_prep(&vp, vma);
2384 vma_adjust_trans_huge(vma, vma->vm_start, addr, 0);
2387 vma->vm_start = addr;
2388 vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT;
2393 /* vma_complete stores the new vma */
2394 vma_complete(&vp, vmi, vma->vm_mm);
2402 mpol_put(vma_policy(new));
2411 * Split a vma into two pieces at address 'addr', a new vma is allocated
2412 * either for the first part or the tail.
2414 int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2415 unsigned long addr, int new_below)
2417 if (vma->vm_mm->map_count >= sysctl_max_map_count)
2420 return __split_vma(vmi, vma, addr, new_below);
2424 * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
2425 * @vmi: The vma iterator
2426 * @vma: The starting vm_area_struct
2427 * @mm: The mm_struct
2428 * @start: The aligned start address to munmap.
2429 * @end: The aligned end address to munmap.
2430 * @uf: The userfaultfd list_head
2431 * @unlock: Set to true to drop the mmap_lock. unlocking only happens on
2434 * Return: 0 on success and drops the lock if so directed, error and leaves the
2435 * lock held otherwise.
2438 do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
2439 struct mm_struct *mm, unsigned long start,
2440 unsigned long end, struct list_head *uf, bool unlock)
2442 struct vm_area_struct *prev, *next = NULL;
2443 struct maple_tree mt_detach;
2445 int error = -ENOMEM;
2446 unsigned long locked_vm = 0;
2447 MA_STATE(mas_detach, &mt_detach, 0, 0);
2448 mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK);
2449 mt_on_stack(mt_detach);
2452 * If we need to split any vma, do it now to save pain later.
2454 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2455 * unmapped vm_area_struct will remain in use: so lower split_vma
2456 * places tmp vma above, and higher split_vma places tmp vma below.
2459 /* Does it split the first one? */
2460 if (start > vma->vm_start) {
2463 * Make sure that map_count on return from munmap() will
2464 * not exceed its limit; but let map_count go just above
2465 * its limit temporarily, to help free resources as expected.
2467 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2468 goto map_count_exceeded;
2470 error = __split_vma(vmi, vma, start, 1);
2472 goto start_split_failed;
2476 * Detach a range of VMAs from the mm. Using next as a temp variable as
2477 * it is always overwritten.
2481 /* Does it split the end? */
2482 if (next->vm_end > end) {
2483 error = __split_vma(vmi, next, end, 0);
2485 goto end_split_failed;
2487 vma_start_write(next);
2488 mas_set(&mas_detach, count);
2489 error = mas_store_gfp(&mas_detach, next, GFP_KERNEL);
2491 goto munmap_gather_failed;
2492 vma_mark_detached(next, true);
2493 if (next->vm_flags & VM_LOCKED)
2494 locked_vm += vma_pages(next);
2499 * If userfaultfd_unmap_prep returns an error the vmas
2500 * will remain split, but userland will get a
2501 * highly unexpected error anyway. This is no
2502 * different than the case where the first of the two
2503 * __split_vma fails, but we don't undo the first
2504 * split, despite we could. This is unlikely enough
2505 * failure that it's not worth optimizing it for.
2507 error = userfaultfd_unmap_prep(next, start, end, uf);
2510 goto userfaultfd_error;
2512 #ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2513 BUG_ON(next->vm_start < start);
2514 BUG_ON(next->vm_start > end);
2516 } for_each_vma_range(*vmi, next, end);
2518 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2519 /* Make sure no VMAs are about to be lost. */
2521 MA_STATE(test, &mt_detach, 0, 0);
2522 struct vm_area_struct *vma_mas, *vma_test;
2525 vma_iter_set(vmi, start);
2527 vma_test = mas_find(&test, count - 1);
2528 for_each_vma_range(*vmi, vma_mas, end) {
2529 BUG_ON(vma_mas != vma_test);
2531 vma_test = mas_next(&test, count - 1);
2534 BUG_ON(count != test_count);
2538 while (vma_iter_addr(vmi) > start)
2539 vma_iter_prev_range(vmi);
2541 error = vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL);
2543 goto clear_tree_failed;
2545 /* Point of no return */
2546 mm->locked_vm -= locked_vm;
2547 mm->map_count -= count;
2549 mmap_write_downgrade(mm);
2551 prev = vma_iter_prev_range(vmi);
2552 next = vma_next(vmi);
2554 vma_iter_prev_range(vmi);
2557 * We can free page tables without write-locking mmap_lock because VMAs
2558 * were isolated before we downgraded mmap_lock.
2560 mas_set(&mas_detach, 1);
2561 unmap_region(mm, &mas_detach, vma, prev, next, start, end, count,
2563 /* Statistics and freeing VMAs */
2564 mas_set(&mas_detach, 0);
2565 remove_mt(mm, &mas_detach);
2568 mmap_read_unlock(mm);
2570 __mt_destroy(&mt_detach);
2575 munmap_gather_failed:
2577 mas_set(&mas_detach, 0);
2578 mas_for_each(&mas_detach, next, end)
2579 vma_mark_detached(next, false);
2581 __mt_destroy(&mt_detach);
2589 * do_vmi_munmap() - munmap a given range.
2590 * @vmi: The vma iterator
2591 * @mm: The mm_struct
2592 * @start: The start address to munmap
2593 * @len: The length of the range to munmap
2594 * @uf: The userfaultfd list_head
2595 * @unlock: set to true if the user wants to drop the mmap_lock on success
2597 * This function takes a @mas that is either pointing to the previous VMA or set
2598 * to MA_START and sets it up to remove the mapping(s). The @len will be
2599 * aligned and any arch_unmap work will be preformed.
2601 * Return: 0 on success and drops the lock if so directed, error and leaves the
2602 * lock held otherwise.
2604 int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm,
2605 unsigned long start, size_t len, struct list_head *uf,
2609 struct vm_area_struct *vma;
2611 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2614 end = start + PAGE_ALIGN(len);
2618 /* arch_unmap() might do unmaps itself. */
2619 arch_unmap(mm, start, end);
2621 /* Find the first overlapping VMA */
2622 vma = vma_find(vmi, end);
2625 mmap_write_unlock(mm);
2629 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
2632 /* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2633 * @mm: The mm_struct
2634 * @start: The start address to munmap
2635 * @len: The length to be munmapped.
2636 * @uf: The userfaultfd list_head
2638 * Return: 0 on success, error otherwise.
2640 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2641 struct list_head *uf)
2643 VMA_ITERATOR(vmi, mm, start);
2645 return do_vmi_munmap(&vmi, mm, start, len, uf, false);
2648 unsigned long mmap_region(struct file *file, unsigned long addr,
2649 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2650 struct list_head *uf)
2652 struct mm_struct *mm = current->mm;
2653 struct vm_area_struct *vma = NULL;
2654 struct vm_area_struct *next, *prev, *merge;
2655 pgoff_t pglen = len >> PAGE_SHIFT;
2656 unsigned long charged = 0;
2657 unsigned long end = addr + len;
2658 unsigned long merge_start = addr, merge_end = end;
2661 VMA_ITERATOR(vmi, mm, addr);
2663 /* Check against address space limit. */
2664 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2665 unsigned long nr_pages;
2668 * MAP_FIXED may remove pages of mappings that intersects with
2669 * requested mapping. Account for the pages it would unmap.
2671 nr_pages = count_vma_pages_range(mm, addr, end);
2673 if (!may_expand_vm(mm, vm_flags,
2674 (len >> PAGE_SHIFT) - nr_pages))
2678 /* Unmap any existing mapping in the area */
2679 if (do_vmi_munmap(&vmi, mm, addr, len, uf, false))
2683 * Private writable mapping: check memory availability
2685 if (accountable_mapping(file, vm_flags)) {
2686 charged = len >> PAGE_SHIFT;
2687 if (security_vm_enough_memory_mm(mm, charged))
2689 vm_flags |= VM_ACCOUNT;
2692 next = vma_next(&vmi);
2693 prev = vma_prev(&vmi);
2694 if (vm_flags & VM_SPECIAL) {
2696 vma_iter_next_range(&vmi);
2700 /* Attempt to expand an old mapping */
2702 if (next && next->vm_start == end && !vma_policy(next) &&
2703 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2704 NULL_VM_UFFD_CTX, NULL)) {
2705 merge_end = next->vm_end;
2707 vm_pgoff = next->vm_pgoff - pglen;
2711 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2712 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2713 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2714 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2715 NULL_VM_UFFD_CTX, NULL))) {
2716 merge_start = prev->vm_start;
2718 vm_pgoff = prev->vm_pgoff;
2720 vma_iter_next_range(&vmi);
2723 /* Actually expand, if possible */
2725 !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) {
2726 khugepaged_enter_vma(vma, vm_flags);
2731 vma_iter_set(&vmi, addr);
2735 * Determine the object being mapped and call the appropriate
2736 * specific mapper. the address has already been validated, but
2737 * not unmapped, but the maps are removed from the list.
2739 vma = vm_area_alloc(mm);
2745 vma_iter_config(&vmi, addr, end);
2746 vma->vm_start = addr;
2748 vm_flags_init(vma, vm_flags);
2749 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2750 vma->vm_pgoff = pgoff;
2753 if (vm_flags & VM_SHARED) {
2754 error = mapping_map_writable(file->f_mapping);
2759 vma->vm_file = get_file(file);
2760 error = call_mmap(file, vma);
2762 goto unmap_and_free_vma;
2765 * Expansion is handled above, merging is handled below.
2766 * Drivers should not alter the address of the VMA.
2769 if (WARN_ON((addr != vma->vm_start)))
2770 goto close_and_free_vma;
2772 vma_iter_config(&vmi, addr, end);
2774 * If vm_flags changed after call_mmap(), we should try merge
2775 * vma again as we may succeed this time.
2777 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2778 merge = vma_merge(&vmi, mm, prev, vma->vm_start,
2779 vma->vm_end, vma->vm_flags, NULL,
2780 vma->vm_file, vma->vm_pgoff, NULL,
2781 NULL_VM_UFFD_CTX, NULL);
2784 * ->mmap() can change vma->vm_file and fput
2785 * the original file. So fput the vma->vm_file
2786 * here or we would add an extra fput for file
2787 * and cause general protection fault
2793 /* Update vm_flags to pick up the change. */
2794 vm_flags = vma->vm_flags;
2795 goto unmap_writable;
2799 vm_flags = vma->vm_flags;
2800 } else if (vm_flags & VM_SHARED) {
2801 error = shmem_zero_setup(vma);
2805 vma_set_anonymous(vma);
2808 if (map_deny_write_exec(vma, vma->vm_flags)) {
2810 goto close_and_free_vma;
2813 /* Allow architectures to sanity-check the vm_flags */
2815 if (!arch_validate_flags(vma->vm_flags))
2816 goto close_and_free_vma;
2819 if (vma_iter_prealloc(&vmi, vma))
2820 goto close_and_free_vma;
2822 /* Lock the VMA since it is modified after insertion into VMA tree */
2823 vma_start_write(vma);
2824 vma_iter_store(&vmi, vma);
2827 i_mmap_lock_write(vma->vm_file->f_mapping);
2828 if (vma->vm_flags & VM_SHARED)
2829 mapping_allow_writable(vma->vm_file->f_mapping);
2831 flush_dcache_mmap_lock(vma->vm_file->f_mapping);
2832 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
2833 flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
2834 i_mmap_unlock_write(vma->vm_file->f_mapping);
2838 * vma_merge() calls khugepaged_enter_vma() either, the below
2839 * call covers the non-merge case.
2841 khugepaged_enter_vma(vma, vma->vm_flags);
2843 /* Once vma denies write, undo our temporary denial count */
2845 if (file && vm_flags & VM_SHARED)
2846 mapping_unmap_writable(file->f_mapping);
2847 file = vma->vm_file;
2850 perf_event_mmap(vma);
2852 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2853 if (vm_flags & VM_LOCKED) {
2854 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2855 is_vm_hugetlb_page(vma) ||
2856 vma == get_gate_vma(current->mm))
2857 vm_flags_clear(vma, VM_LOCKED_MASK);
2859 mm->locked_vm += (len >> PAGE_SHIFT);
2866 * New (or expanded) vma always get soft dirty status.
2867 * Otherwise user-space soft-dirty page tracker won't
2868 * be able to distinguish situation when vma area unmapped,
2869 * then new mapped in-place (which must be aimed as
2870 * a completely new data area).
2872 vm_flags_set(vma, VM_SOFTDIRTY);
2874 vma_set_page_prot(vma);
2880 if (file && vma->vm_ops && vma->vm_ops->close)
2881 vma->vm_ops->close(vma);
2883 if (file || vma->vm_file) {
2886 vma->vm_file = NULL;
2888 vma_iter_set(&vmi, vma->vm_end);
2889 /* Undo any partial mapping done by a device driver. */
2890 unmap_region(mm, &vmi.mas, vma, prev, next, vma->vm_start,
2891 vma->vm_end, vma->vm_end, true);
2893 if (file && (vm_flags & VM_SHARED))
2894 mapping_unmap_writable(file->f_mapping);
2899 vm_unacct_memory(charged);
2904 static int __vm_munmap(unsigned long start, size_t len, bool unlock)
2907 struct mm_struct *mm = current->mm;
2909 VMA_ITERATOR(vmi, mm, start);
2911 if (mmap_write_lock_killable(mm))
2914 ret = do_vmi_munmap(&vmi, mm, start, len, &uf, unlock);
2916 mmap_write_unlock(mm);
2918 userfaultfd_unmap_complete(mm, &uf);
2922 int vm_munmap(unsigned long start, size_t len)
2924 return __vm_munmap(start, len, false);
2926 EXPORT_SYMBOL(vm_munmap);
2928 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2930 addr = untagged_addr(addr);
2931 return __vm_munmap(addr, len, true);
2936 * Emulation of deprecated remap_file_pages() syscall.
2938 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2939 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2942 struct mm_struct *mm = current->mm;
2943 struct vm_area_struct *vma;
2944 unsigned long populate = 0;
2945 unsigned long ret = -EINVAL;
2948 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
2949 current->comm, current->pid);
2953 start = start & PAGE_MASK;
2954 size = size & PAGE_MASK;
2956 if (start + size <= start)
2959 /* Does pgoff wrap? */
2960 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2963 if (mmap_write_lock_killable(mm))
2966 vma = vma_lookup(mm, start);
2968 if (!vma || !(vma->vm_flags & VM_SHARED))
2971 if (start + size > vma->vm_end) {
2972 VMA_ITERATOR(vmi, mm, vma->vm_end);
2973 struct vm_area_struct *next, *prev = vma;
2975 for_each_vma_range(vmi, next, start + size) {
2976 /* hole between vmas ? */
2977 if (next->vm_start != prev->vm_end)
2980 if (next->vm_file != vma->vm_file)
2983 if (next->vm_flags != vma->vm_flags)
2986 if (start + size <= next->vm_end)
2996 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2997 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2998 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3000 flags &= MAP_NONBLOCK;
3001 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3002 if (vma->vm_flags & VM_LOCKED)
3003 flags |= MAP_LOCKED;
3005 file = get_file(vma->vm_file);
3006 ret = do_mmap(vma->vm_file, start, size,
3007 prot, flags, 0, pgoff, &populate, NULL);
3010 mmap_write_unlock(mm);
3012 mm_populate(ret, populate);
3013 if (!IS_ERR_VALUE(ret))
3019 * do_vma_munmap() - Unmap a full or partial vma.
3020 * @vmi: The vma iterator pointing at the vma
3021 * @vma: The first vma to be munmapped
3022 * @start: the start of the address to unmap
3023 * @end: The end of the address to unmap
3024 * @uf: The userfaultfd list_head
3025 * @unlock: Drop the lock on success
3027 * unmaps a VMA mapping when the vma iterator is already in position.
3028 * Does not handle alignment.
3030 * Return: 0 on success drops the lock of so directed, error on failure and will
3031 * still hold the lock.
3033 int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
3034 unsigned long start, unsigned long end, struct list_head *uf,
3037 struct mm_struct *mm = vma->vm_mm;
3039 arch_unmap(mm, start, end);
3040 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
3044 * do_brk_flags() - Increase the brk vma if the flags match.
3045 * @vmi: The vma iterator
3046 * @addr: The start address
3047 * @len: The length of the increase
3049 * @flags: The VMA Flags
3051 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
3052 * do not match then create a new anonymous VMA. Eventually we may be able to
3053 * do some brk-specific accounting here.
3055 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
3056 unsigned long addr, unsigned long len, unsigned long flags)
3058 struct mm_struct *mm = current->mm;
3059 struct vma_prepare vp;
3062 * Check against address space limits by the changed size
3063 * Note: This happens *after* clearing old mappings in some code paths.
3065 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3066 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3069 if (mm->map_count > sysctl_max_map_count)
3072 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3076 * Expand the existing vma if possible; Note that singular lists do not
3077 * occur after forking, so the expand will only happen on new VMAs.
3079 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
3080 can_vma_merge_after(vma, flags, NULL, NULL,
3081 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
3082 vma_iter_config(vmi, vma->vm_start, addr + len);
3083 if (vma_iter_prealloc(vmi, vma))
3086 vma_start_write(vma);
3088 init_vma_prep(&vp, vma);
3090 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
3091 vma->vm_end = addr + len;
3092 vm_flags_set(vma, VM_SOFTDIRTY);
3093 vma_iter_store(vmi, vma);
3095 vma_complete(&vp, vmi, mm);
3096 khugepaged_enter_vma(vma, flags);
3101 vma_iter_next_range(vmi);
3102 /* create a vma struct for an anonymous mapping */
3103 vma = vm_area_alloc(mm);
3107 vma_set_anonymous(vma);
3108 vma->vm_start = addr;
3109 vma->vm_end = addr + len;
3110 vma->vm_pgoff = addr >> PAGE_SHIFT;
3111 vm_flags_init(vma, flags);
3112 vma->vm_page_prot = vm_get_page_prot(flags);
3113 vma_start_write(vma);
3114 if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
3115 goto mas_store_fail;
3121 perf_event_mmap(vma);
3122 mm->total_vm += len >> PAGE_SHIFT;
3123 mm->data_vm += len >> PAGE_SHIFT;
3124 if (flags & VM_LOCKED)
3125 mm->locked_vm += (len >> PAGE_SHIFT);
3126 vm_flags_set(vma, VM_SOFTDIRTY);
3132 vm_unacct_memory(len >> PAGE_SHIFT);
3136 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3138 struct mm_struct *mm = current->mm;
3139 struct vm_area_struct *vma = NULL;
3144 VMA_ITERATOR(vmi, mm, addr);
3146 len = PAGE_ALIGN(request);
3152 /* Until we need other flags, refuse anything except VM_EXEC. */
3153 if ((flags & (~VM_EXEC)) != 0)
3156 if (mmap_write_lock_killable(mm))
3159 ret = check_brk_limits(addr, len);
3163 ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0);
3167 vma = vma_prev(&vmi);
3168 ret = do_brk_flags(&vmi, vma, addr, len, flags);
3169 populate = ((mm->def_flags & VM_LOCKED) != 0);
3170 mmap_write_unlock(mm);
3171 userfaultfd_unmap_complete(mm, &uf);
3172 if (populate && !ret)
3173 mm_populate(addr, len);
3178 mmap_write_unlock(mm);
3181 EXPORT_SYMBOL(vm_brk_flags);
3183 int vm_brk(unsigned long addr, unsigned long len)
3185 return vm_brk_flags(addr, len, 0);
3187 EXPORT_SYMBOL(vm_brk);
3189 /* Release all mmaps. */
3190 void exit_mmap(struct mm_struct *mm)
3192 struct mmu_gather tlb;
3193 struct vm_area_struct *vma;
3194 unsigned long nr_accounted = 0;
3195 MA_STATE(mas, &mm->mm_mt, 0, 0);
3198 /* mm's last user has gone, and its about to be pulled down */
3199 mmu_notifier_release(mm);
3204 vma = mas_find(&mas, ULONG_MAX);
3206 /* Can happen if dup_mmap() received an OOM */
3207 mmap_read_unlock(mm);
3213 tlb_gather_mmu_fullmm(&tlb, mm);
3214 /* update_hiwater_rss(mm) here? but nobody should be looking */
3215 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3216 unmap_vmas(&tlb, &mas, vma, 0, ULONG_MAX, ULONG_MAX, false);
3217 mmap_read_unlock(mm);
3220 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3221 * because the memory has been already freed.
3223 set_bit(MMF_OOM_SKIP, &mm->flags);
3224 mmap_write_lock(mm);
3225 mt_clear_in_rcu(&mm->mm_mt);
3226 mas_set(&mas, vma->vm_end);
3227 free_pgtables(&tlb, &mas, vma, FIRST_USER_ADDRESS,
3228 USER_PGTABLES_CEILING, true);
3229 tlb_finish_mmu(&tlb);
3232 * Walk the list again, actually closing and freeing it, with preemption
3233 * enabled, without holding any MM locks besides the unreachable
3236 mas_set(&mas, vma->vm_end);
3238 if (vma->vm_flags & VM_ACCOUNT)
3239 nr_accounted += vma_pages(vma);
3240 remove_vma(vma, true);
3243 } while ((vma = mas_find(&mas, ULONG_MAX)) != NULL);
3245 BUG_ON(count != mm->map_count);
3247 trace_exit_mmap(mm);
3248 __mt_destroy(&mm->mm_mt);
3249 mmap_write_unlock(mm);
3250 vm_unacct_memory(nr_accounted);
3253 /* Insert vm structure into process list sorted by address
3254 * and into the inode's i_mmap tree. If vm_file is non-NULL
3255 * then i_mmap_rwsem is taken here.
3257 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3259 unsigned long charged = vma_pages(vma);
3262 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3265 if ((vma->vm_flags & VM_ACCOUNT) &&
3266 security_vm_enough_memory_mm(mm, charged))
3270 * The vm_pgoff of a purely anonymous vma should be irrelevant
3271 * until its first write fault, when page's anon_vma and index
3272 * are set. But now set the vm_pgoff it will almost certainly
3273 * end up with (unless mremap moves it elsewhere before that
3274 * first wfault), so /proc/pid/maps tells a consistent story.
3276 * By setting it to reflect the virtual start address of the
3277 * vma, merges and splits can happen in a seamless way, just
3278 * using the existing file pgoff checks and manipulations.
3279 * Similarly in do_mmap and in do_brk_flags.
3281 if (vma_is_anonymous(vma)) {
3282 BUG_ON(vma->anon_vma);
3283 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3286 if (vma_link(mm, vma)) {
3287 vm_unacct_memory(charged);
3295 * Copy the vma structure to a new location in the same mm,
3296 * prior to moving page table entries, to effect an mremap move.
3298 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3299 unsigned long addr, unsigned long len, pgoff_t pgoff,
3300 bool *need_rmap_locks)
3302 struct vm_area_struct *vma = *vmap;
3303 unsigned long vma_start = vma->vm_start;
3304 struct mm_struct *mm = vma->vm_mm;
3305 struct vm_area_struct *new_vma, *prev;
3306 bool faulted_in_anon_vma = true;
3307 VMA_ITERATOR(vmi, mm, addr);
3310 * If anonymous vma has not yet been faulted, update new pgoff
3311 * to match new location, to increase its chance of merging.
3313 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3314 pgoff = addr >> PAGE_SHIFT;
3315 faulted_in_anon_vma = false;
3318 new_vma = find_vma_prev(mm, addr, &prev);
3319 if (new_vma && new_vma->vm_start < addr + len)
3320 return NULL; /* should never get here */
3322 new_vma = vma_merge(&vmi, mm, prev, addr, addr + len, vma->vm_flags,
3323 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3324 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
3327 * Source vma may have been merged into new_vma
3329 if (unlikely(vma_start >= new_vma->vm_start &&
3330 vma_start < new_vma->vm_end)) {
3332 * The only way we can get a vma_merge with
3333 * self during an mremap is if the vma hasn't
3334 * been faulted in yet and we were allowed to
3335 * reset the dst vma->vm_pgoff to the
3336 * destination address of the mremap to allow
3337 * the merge to happen. mremap must change the
3338 * vm_pgoff linearity between src and dst vmas
3339 * (in turn preventing a vma_merge) to be
3340 * safe. It is only safe to keep the vm_pgoff
3341 * linear if there are no pages mapped yet.
3343 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3344 *vmap = vma = new_vma;
3346 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3348 new_vma = vm_area_dup(vma);
3351 new_vma->vm_start = addr;
3352 new_vma->vm_end = addr + len;
3353 new_vma->vm_pgoff = pgoff;
3354 if (vma_dup_policy(vma, new_vma))
3356 if (anon_vma_clone(new_vma, vma))
3357 goto out_free_mempol;
3358 if (new_vma->vm_file)
3359 get_file(new_vma->vm_file);
3360 if (new_vma->vm_ops && new_vma->vm_ops->open)
3361 new_vma->vm_ops->open(new_vma);
3362 if (vma_link(mm, new_vma))
3364 *need_rmap_locks = false;
3369 if (new_vma->vm_ops && new_vma->vm_ops->close)
3370 new_vma->vm_ops->close(new_vma);
3372 if (new_vma->vm_file)
3373 fput(new_vma->vm_file);
3375 unlink_anon_vmas(new_vma);
3377 mpol_put(vma_policy(new_vma));
3379 vm_area_free(new_vma);
3385 * Return true if the calling process may expand its vm space by the passed
3388 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3390 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3393 if (is_data_mapping(flags) &&
3394 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3395 /* Workaround for Valgrind */
3396 if (rlimit(RLIMIT_DATA) == 0 &&
3397 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3400 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3401 current->comm, current->pid,
3402 (mm->data_vm + npages) << PAGE_SHIFT,
3403 rlimit(RLIMIT_DATA),
3404 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3406 if (!ignore_rlimit_data)
3413 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3415 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3417 if (is_exec_mapping(flags))
3418 mm->exec_vm += npages;
3419 else if (is_stack_mapping(flags))
3420 mm->stack_vm += npages;
3421 else if (is_data_mapping(flags))
3422 mm->data_vm += npages;
3425 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3428 * Having a close hook prevents vma merging regardless of flags.
3430 static void special_mapping_close(struct vm_area_struct *vma)
3434 static const char *special_mapping_name(struct vm_area_struct *vma)
3436 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3439 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3441 struct vm_special_mapping *sm = new_vma->vm_private_data;
3443 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3447 return sm->mremap(sm, new_vma);
3452 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3455 * Forbid splitting special mappings - kernel has expectations over
3456 * the number of pages in mapping. Together with VM_DONTEXPAND
3457 * the size of vma should stay the same over the special mapping's
3463 static const struct vm_operations_struct special_mapping_vmops = {
3464 .close = special_mapping_close,
3465 .fault = special_mapping_fault,
3466 .mremap = special_mapping_mremap,
3467 .name = special_mapping_name,
3468 /* vDSO code relies that VVAR can't be accessed remotely */
3470 .may_split = special_mapping_split,
3473 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3474 .close = special_mapping_close,
3475 .fault = special_mapping_fault,
3478 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3480 struct vm_area_struct *vma = vmf->vma;
3482 struct page **pages;
3484 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3485 pages = vma->vm_private_data;
3487 struct vm_special_mapping *sm = vma->vm_private_data;
3490 return sm->fault(sm, vmf->vma, vmf);
3495 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3499 struct page *page = *pages;
3505 return VM_FAULT_SIGBUS;
3508 static struct vm_area_struct *__install_special_mapping(
3509 struct mm_struct *mm,
3510 unsigned long addr, unsigned long len,
3511 unsigned long vm_flags, void *priv,
3512 const struct vm_operations_struct *ops)
3515 struct vm_area_struct *vma;
3517 vma = vm_area_alloc(mm);
3518 if (unlikely(vma == NULL))
3519 return ERR_PTR(-ENOMEM);
3521 vma->vm_start = addr;
3522 vma->vm_end = addr + len;
3524 vm_flags_init(vma, (vm_flags | mm->def_flags |
3525 VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK);
3526 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3529 vma->vm_private_data = priv;
3531 ret = insert_vm_struct(mm, vma);
3535 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3537 perf_event_mmap(vma);
3543 return ERR_PTR(ret);
3546 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3547 const struct vm_special_mapping *sm)
3549 return vma->vm_private_data == sm &&
3550 (vma->vm_ops == &special_mapping_vmops ||
3551 vma->vm_ops == &legacy_special_mapping_vmops);
3555 * Called with mm->mmap_lock held for writing.
3556 * Insert a new vma covering the given region, with the given flags.
3557 * Its pages are supplied by the given array of struct page *.
3558 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3559 * The region past the last page supplied will always produce SIGBUS.
3560 * The array pointer and the pages it points to are assumed to stay alive
3561 * for as long as this mapping might exist.
3563 struct vm_area_struct *_install_special_mapping(
3564 struct mm_struct *mm,
3565 unsigned long addr, unsigned long len,
3566 unsigned long vm_flags, const struct vm_special_mapping *spec)
3568 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3569 &special_mapping_vmops);
3572 int install_special_mapping(struct mm_struct *mm,
3573 unsigned long addr, unsigned long len,
3574 unsigned long vm_flags, struct page **pages)
3576 struct vm_area_struct *vma = __install_special_mapping(
3577 mm, addr, len, vm_flags, (void *)pages,
3578 &legacy_special_mapping_vmops);
3580 return PTR_ERR_OR_ZERO(vma);
3583 static DEFINE_MUTEX(mm_all_locks_mutex);
3585 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3587 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3589 * The LSB of head.next can't change from under us
3590 * because we hold the mm_all_locks_mutex.
3592 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3594 * We can safely modify head.next after taking the
3595 * anon_vma->root->rwsem. If some other vma in this mm shares
3596 * the same anon_vma we won't take it again.
3598 * No need of atomic instructions here, head.next
3599 * can't change from under us thanks to the
3600 * anon_vma->root->rwsem.
3602 if (__test_and_set_bit(0, (unsigned long *)
3603 &anon_vma->root->rb_root.rb_root.rb_node))
3608 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3610 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3612 * AS_MM_ALL_LOCKS can't change from under us because
3613 * we hold the mm_all_locks_mutex.
3615 * Operations on ->flags have to be atomic because
3616 * even if AS_MM_ALL_LOCKS is stable thanks to the
3617 * mm_all_locks_mutex, there may be other cpus
3618 * changing other bitflags in parallel to us.
3620 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3622 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3627 * This operation locks against the VM for all pte/vma/mm related
3628 * operations that could ever happen on a certain mm. This includes
3629 * vmtruncate, try_to_unmap, and all page faults.
3631 * The caller must take the mmap_lock in write mode before calling
3632 * mm_take_all_locks(). The caller isn't allowed to release the
3633 * mmap_lock until mm_drop_all_locks() returns.
3635 * mmap_lock in write mode is required in order to block all operations
3636 * that could modify pagetables and free pages without need of
3637 * altering the vma layout. It's also needed in write mode to avoid new
3638 * anon_vmas to be associated with existing vmas.
3640 * A single task can't take more than one mm_take_all_locks() in a row
3641 * or it would deadlock.
3643 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3644 * mapping->flags avoid to take the same lock twice, if more than one
3645 * vma in this mm is backed by the same anon_vma or address_space.
3647 * We take locks in following order, accordingly to comment at beginning
3649 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3651 * - all vmas marked locked
3652 * - all i_mmap_rwsem locks;
3653 * - all anon_vma->rwseml
3655 * We can take all locks within these types randomly because the VM code
3656 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3657 * mm_all_locks_mutex.
3659 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3660 * that may have to take thousand of locks.
3662 * mm_take_all_locks() can fail if it's interrupted by signals.
3664 int mm_take_all_locks(struct mm_struct *mm)
3666 struct vm_area_struct *vma;
3667 struct anon_vma_chain *avc;
3668 MA_STATE(mas, &mm->mm_mt, 0, 0);
3670 mmap_assert_write_locked(mm);
3672 mutex_lock(&mm_all_locks_mutex);
3675 * vma_start_write() does not have a complement in mm_drop_all_locks()
3676 * because vma_start_write() is always asymmetrical; it marks a VMA as
3677 * being written to until mmap_write_unlock() or mmap_write_downgrade()
3680 mas_for_each(&mas, vma, ULONG_MAX) {
3681 if (signal_pending(current))
3683 vma_start_write(vma);
3687 mas_for_each(&mas, vma, ULONG_MAX) {
3688 if (signal_pending(current))
3690 if (vma->vm_file && vma->vm_file->f_mapping &&
3691 is_vm_hugetlb_page(vma))
3692 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3696 mas_for_each(&mas, vma, ULONG_MAX) {
3697 if (signal_pending(current))
3699 if (vma->vm_file && vma->vm_file->f_mapping &&
3700 !is_vm_hugetlb_page(vma))
3701 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3705 mas_for_each(&mas, vma, ULONG_MAX) {
3706 if (signal_pending(current))
3709 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3710 vm_lock_anon_vma(mm, avc->anon_vma);
3716 mm_drop_all_locks(mm);
3720 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3722 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3724 * The LSB of head.next can't change to 0 from under
3725 * us because we hold the mm_all_locks_mutex.
3727 * We must however clear the bitflag before unlocking
3728 * the vma so the users using the anon_vma->rb_root will
3729 * never see our bitflag.
3731 * No need of atomic instructions here, head.next
3732 * can't change from under us until we release the
3733 * anon_vma->root->rwsem.
3735 if (!__test_and_clear_bit(0, (unsigned long *)
3736 &anon_vma->root->rb_root.rb_root.rb_node))
3738 anon_vma_unlock_write(anon_vma);
3742 static void vm_unlock_mapping(struct address_space *mapping)
3744 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3746 * AS_MM_ALL_LOCKS can't change to 0 from under us
3747 * because we hold the mm_all_locks_mutex.
3749 i_mmap_unlock_write(mapping);
3750 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3757 * The mmap_lock cannot be released by the caller until
3758 * mm_drop_all_locks() returns.
3760 void mm_drop_all_locks(struct mm_struct *mm)
3762 struct vm_area_struct *vma;
3763 struct anon_vma_chain *avc;
3764 MA_STATE(mas, &mm->mm_mt, 0, 0);
3766 mmap_assert_write_locked(mm);
3767 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3769 mas_for_each(&mas, vma, ULONG_MAX) {
3771 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3772 vm_unlock_anon_vma(avc->anon_vma);
3773 if (vma->vm_file && vma->vm_file->f_mapping)
3774 vm_unlock_mapping(vma->vm_file->f_mapping);
3777 mutex_unlock(&mm_all_locks_mutex);
3781 * initialise the percpu counter for VM
3783 void __init mmap_init(void)
3787 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3792 * Initialise sysctl_user_reserve_kbytes.
3794 * This is intended to prevent a user from starting a single memory hogging
3795 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3798 * The default value is min(3% of free memory, 128MB)
3799 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3801 static int init_user_reserve(void)
3803 unsigned long free_kbytes;
3805 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3807 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3810 subsys_initcall(init_user_reserve);
3813 * Initialise sysctl_admin_reserve_kbytes.
3815 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3816 * to log in and kill a memory hogging process.
3818 * Systems with more than 256MB will reserve 8MB, enough to recover
3819 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3820 * only reserve 3% of free pages by default.
3822 static int init_admin_reserve(void)
3824 unsigned long free_kbytes;
3826 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3828 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3831 subsys_initcall(init_admin_reserve);
3834 * Reinititalise user and admin reserves if memory is added or removed.
3836 * The default user reserve max is 128MB, and the default max for the
3837 * admin reserve is 8MB. These are usually, but not always, enough to
3838 * enable recovery from a memory hogging process using login/sshd, a shell,
3839 * and tools like top. It may make sense to increase or even disable the
3840 * reserve depending on the existence of swap or variations in the recovery
3841 * tools. So, the admin may have changed them.
3843 * If memory is added and the reserves have been eliminated or increased above
3844 * the default max, then we'll trust the admin.
3846 * If memory is removed and there isn't enough free memory, then we
3847 * need to reset the reserves.
3849 * Otherwise keep the reserve set by the admin.
3851 static int reserve_mem_notifier(struct notifier_block *nb,
3852 unsigned long action, void *data)
3854 unsigned long tmp, free_kbytes;
3858 /* Default max is 128MB. Leave alone if modified by operator. */
3859 tmp = sysctl_user_reserve_kbytes;
3860 if (0 < tmp && tmp < (1UL << 17))
3861 init_user_reserve();
3863 /* Default max is 8MB. Leave alone if modified by operator. */
3864 tmp = sysctl_admin_reserve_kbytes;
3865 if (0 < tmp && tmp < (1UL << 13))
3866 init_admin_reserve();
3870 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3872 if (sysctl_user_reserve_kbytes > free_kbytes) {
3873 init_user_reserve();
3874 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3875 sysctl_user_reserve_kbytes);
3878 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3879 init_admin_reserve();
3880 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3881 sysctl_admin_reserve_kbytes);
3890 static int __meminit init_reserve_notifier(void)
3892 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3893 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3897 subsys_initcall(init_reserve_notifier);