6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/shmem_fs.h>
29 #include <linux/profile.h>
30 #include <linux/export.h>
31 #include <linux/mount.h>
32 #include <linux/mempolicy.h>
33 #include <linux/rmap.h>
34 #include <linux/mmu_notifier.h>
35 #include <linux/mmdebug.h>
36 #include <linux/perf_event.h>
37 #include <linux/audit.h>
38 #include <linux/khugepaged.h>
39 #include <linux/uprobes.h>
40 #include <linux/rbtree_augmented.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>
49 #include <linux/uaccess.h>
50 #include <asm/cacheflush.h>
52 #include <asm/mmu_context.h>
56 #ifndef arch_mmap_check
57 #define arch_mmap_check(addr, len, flags) (0)
60 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
61 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
62 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
63 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
66 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
67 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
68 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
71 static bool ignore_rlimit_data;
72 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
74 static void unmap_region(struct mm_struct *mm,
75 struct vm_area_struct *vma, struct vm_area_struct *prev,
76 unsigned long start, unsigned long end);
78 /* description of effects of mapping type and prot in current implementation.
79 * this is due to the limited x86 page protection hardware. The expected
80 * behavior is in parens:
83 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
84 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
85 * w: (no) no w: (no) no w: (yes) yes w: (no) no
86 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
88 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
89 * w: (no) no w: (no) no w: (copy) copy w: (no) no
90 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
92 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
98 pgprot_t protection_map[16] __ro_after_init = {
99 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
100 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
103 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
104 static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
110 pgprot_t vm_get_page_prot(unsigned long vm_flags)
112 pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
113 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
114 pgprot_val(arch_vm_get_page_prot(vm_flags)));
116 return arch_filter_pgprot(ret);
118 EXPORT_SYMBOL(vm_get_page_prot);
120 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
122 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
125 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
126 void vma_set_page_prot(struct vm_area_struct *vma)
128 unsigned long vm_flags = vma->vm_flags;
129 pgprot_t vm_page_prot;
131 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
132 if (vma_wants_writenotify(vma, vm_page_prot)) {
133 vm_flags &= ~VM_SHARED;
134 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
136 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
137 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
141 * Requires inode->i_mapping->i_mmap_rwsem
143 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
144 struct file *file, struct address_space *mapping)
146 if (vma->vm_flags & VM_DENYWRITE)
147 atomic_inc(&file_inode(file)->i_writecount);
148 if (vma->vm_flags & VM_SHARED)
149 mapping_unmap_writable(mapping);
151 flush_dcache_mmap_lock(mapping);
152 vma_interval_tree_remove(vma, &mapping->i_mmap);
153 flush_dcache_mmap_unlock(mapping);
157 * Unlink a file-based vm structure from its interval tree, to hide
158 * vma from rmap and vmtruncate before freeing its page tables.
160 void unlink_file_vma(struct vm_area_struct *vma)
162 struct file *file = vma->vm_file;
165 struct address_space *mapping = file->f_mapping;
166 i_mmap_lock_write(mapping);
167 __remove_shared_vm_struct(vma, file, mapping);
168 i_mmap_unlock_write(mapping);
173 * Close a vm structure and free it, returning the next.
175 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
177 struct vm_area_struct *next = vma->vm_next;
180 if (vma->vm_ops && vma->vm_ops->close)
181 vma->vm_ops->close(vma);
184 mpol_put(vma_policy(vma));
185 kmem_cache_free(vm_area_cachep, vma);
189 static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf);
191 SYSCALL_DEFINE1(brk, unsigned long, brk)
193 unsigned long retval;
194 unsigned long newbrk, oldbrk;
195 struct mm_struct *mm = current->mm;
196 struct vm_area_struct *next;
197 unsigned long min_brk;
201 if (down_write_killable(&mm->mmap_sem))
204 #ifdef CONFIG_COMPAT_BRK
206 * CONFIG_COMPAT_BRK can still be overridden by setting
207 * randomize_va_space to 2, which will still cause mm->start_brk
208 * to be arbitrarily shifted
210 if (current->brk_randomized)
211 min_brk = mm->start_brk;
213 min_brk = mm->end_data;
215 min_brk = mm->start_brk;
221 * Check against rlimit here. If this check is done later after the test
222 * of oldbrk with newbrk then it can escape the test and let the data
223 * segment grow beyond its set limit the in case where the limit is
224 * not page aligned -Ram Gupta
226 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
227 mm->end_data, mm->start_data))
230 newbrk = PAGE_ALIGN(brk);
231 oldbrk = PAGE_ALIGN(mm->brk);
232 if (oldbrk == newbrk)
235 /* Always allow shrinking brk. */
236 if (brk <= mm->brk) {
237 if (!do_munmap(mm, newbrk, oldbrk-newbrk, &uf))
242 /* Check against existing mmap mappings. */
243 next = find_vma(mm, oldbrk);
244 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
247 /* Ok, looks good - let it rip. */
248 if (do_brk(oldbrk, newbrk-oldbrk, &uf) < 0)
253 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
254 up_write(&mm->mmap_sem);
255 userfaultfd_unmap_complete(mm, &uf);
257 mm_populate(oldbrk, newbrk - oldbrk);
262 up_write(&mm->mmap_sem);
266 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
268 unsigned long max, prev_end, subtree_gap;
271 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
272 * allow two stack_guard_gaps between them here, and when choosing
273 * an unmapped area; whereas when expanding we only require one.
274 * That's a little inconsistent, but keeps the code here simpler.
276 max = vm_start_gap(vma);
278 prev_end = vm_end_gap(vma->vm_prev);
284 if (vma->vm_rb.rb_left) {
285 subtree_gap = rb_entry(vma->vm_rb.rb_left,
286 struct vm_area_struct, vm_rb)->rb_subtree_gap;
287 if (subtree_gap > max)
290 if (vma->vm_rb.rb_right) {
291 subtree_gap = rb_entry(vma->vm_rb.rb_right,
292 struct vm_area_struct, vm_rb)->rb_subtree_gap;
293 if (subtree_gap > max)
299 #ifdef CONFIG_DEBUG_VM_RB
300 static int browse_rb(struct mm_struct *mm)
302 struct rb_root *root = &mm->mm_rb;
303 int i = 0, j, bug = 0;
304 struct rb_node *nd, *pn = NULL;
305 unsigned long prev = 0, pend = 0;
307 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
308 struct vm_area_struct *vma;
309 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
310 if (vma->vm_start < prev) {
311 pr_emerg("vm_start %lx < prev %lx\n",
312 vma->vm_start, prev);
315 if (vma->vm_start < pend) {
316 pr_emerg("vm_start %lx < pend %lx\n",
317 vma->vm_start, pend);
320 if (vma->vm_start > vma->vm_end) {
321 pr_emerg("vm_start %lx > vm_end %lx\n",
322 vma->vm_start, vma->vm_end);
325 spin_lock(&mm->page_table_lock);
326 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
327 pr_emerg("free gap %lx, correct %lx\n",
329 vma_compute_subtree_gap(vma));
332 spin_unlock(&mm->page_table_lock);
335 prev = vma->vm_start;
339 for (nd = pn; nd; nd = rb_prev(nd))
342 pr_emerg("backwards %d, forwards %d\n", j, i);
348 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
352 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
353 struct vm_area_struct *vma;
354 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
355 VM_BUG_ON_VMA(vma != ignore &&
356 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
361 static void validate_mm(struct mm_struct *mm)
365 unsigned long highest_address = 0;
366 struct vm_area_struct *vma = mm->mmap;
369 struct anon_vma *anon_vma = vma->anon_vma;
370 struct anon_vma_chain *avc;
373 anon_vma_lock_read(anon_vma);
374 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
375 anon_vma_interval_tree_verify(avc);
376 anon_vma_unlock_read(anon_vma);
379 highest_address = vm_end_gap(vma);
383 if (i != mm->map_count) {
384 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
387 if (highest_address != mm->highest_vm_end) {
388 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
389 mm->highest_vm_end, highest_address);
393 if (i != mm->map_count) {
395 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
398 VM_BUG_ON_MM(bug, mm);
401 #define validate_mm_rb(root, ignore) do { } while (0)
402 #define validate_mm(mm) do { } while (0)
405 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
406 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
409 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
410 * vma->vm_prev->vm_end values changed, without modifying the vma's position
413 static void vma_gap_update(struct vm_area_struct *vma)
416 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
417 * function that does exacltly what we want.
419 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
422 static inline void vma_rb_insert(struct vm_area_struct *vma,
423 struct rb_root *root)
425 /* All rb_subtree_gap values must be consistent prior to insertion */
426 validate_mm_rb(root, NULL);
428 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
431 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
434 * Note rb_erase_augmented is a fairly large inline function,
435 * so make sure we instantiate it only once with our desired
436 * augmented rbtree callbacks.
438 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
441 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
442 struct rb_root *root,
443 struct vm_area_struct *ignore)
446 * All rb_subtree_gap values must be consistent prior to erase,
447 * with the possible exception of the "next" vma being erased if
448 * next->vm_start was reduced.
450 validate_mm_rb(root, ignore);
452 __vma_rb_erase(vma, root);
455 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
456 struct rb_root *root)
459 * All rb_subtree_gap values must be consistent prior to erase,
460 * with the possible exception of the vma being erased.
462 validate_mm_rb(root, vma);
464 __vma_rb_erase(vma, root);
468 * vma has some anon_vma assigned, and is already inserted on that
469 * anon_vma's interval trees.
471 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
472 * vma must be removed from the anon_vma's interval trees using
473 * anon_vma_interval_tree_pre_update_vma().
475 * After the update, the vma will be reinserted using
476 * anon_vma_interval_tree_post_update_vma().
478 * The entire update must be protected by exclusive mmap_sem and by
479 * the root anon_vma's mutex.
482 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
484 struct anon_vma_chain *avc;
486 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
487 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
491 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
493 struct anon_vma_chain *avc;
495 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
496 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
499 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
500 unsigned long end, struct vm_area_struct **pprev,
501 struct rb_node ***rb_link, struct rb_node **rb_parent)
503 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
505 __rb_link = &mm->mm_rb.rb_node;
506 rb_prev = __rb_parent = NULL;
509 struct vm_area_struct *vma_tmp;
511 __rb_parent = *__rb_link;
512 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
514 if (vma_tmp->vm_end > addr) {
515 /* Fail if an existing vma overlaps the area */
516 if (vma_tmp->vm_start < end)
518 __rb_link = &__rb_parent->rb_left;
520 rb_prev = __rb_parent;
521 __rb_link = &__rb_parent->rb_right;
527 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
528 *rb_link = __rb_link;
529 *rb_parent = __rb_parent;
533 static unsigned long count_vma_pages_range(struct mm_struct *mm,
534 unsigned long addr, unsigned long end)
536 unsigned long nr_pages = 0;
537 struct vm_area_struct *vma;
539 /* Find first overlaping mapping */
540 vma = find_vma_intersection(mm, addr, end);
544 nr_pages = (min(end, vma->vm_end) -
545 max(addr, vma->vm_start)) >> PAGE_SHIFT;
547 /* Iterate over the rest of the overlaps */
548 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
549 unsigned long overlap_len;
551 if (vma->vm_start > end)
554 overlap_len = min(end, vma->vm_end) - vma->vm_start;
555 nr_pages += overlap_len >> PAGE_SHIFT;
561 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
562 struct rb_node **rb_link, struct rb_node *rb_parent)
564 /* Update tracking information for the gap following the new vma. */
566 vma_gap_update(vma->vm_next);
568 mm->highest_vm_end = vm_end_gap(vma);
571 * vma->vm_prev wasn't known when we followed the rbtree to find the
572 * correct insertion point for that vma. As a result, we could not
573 * update the vma vm_rb parents rb_subtree_gap values on the way down.
574 * So, we first insert the vma with a zero rb_subtree_gap value
575 * (to be consistent with what we did on the way down), and then
576 * immediately update the gap to the correct value. Finally we
577 * rebalance the rbtree after all augmented values have been set.
579 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
580 vma->rb_subtree_gap = 0;
582 vma_rb_insert(vma, &mm->mm_rb);
585 static void __vma_link_file(struct vm_area_struct *vma)
591 struct address_space *mapping = file->f_mapping;
593 if (vma->vm_flags & VM_DENYWRITE)
594 atomic_dec(&file_inode(file)->i_writecount);
595 if (vma->vm_flags & VM_SHARED)
596 atomic_inc(&mapping->i_mmap_writable);
598 flush_dcache_mmap_lock(mapping);
599 vma_interval_tree_insert(vma, &mapping->i_mmap);
600 flush_dcache_mmap_unlock(mapping);
605 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
606 struct vm_area_struct *prev, struct rb_node **rb_link,
607 struct rb_node *rb_parent)
609 __vma_link_list(mm, vma, prev, rb_parent);
610 __vma_link_rb(mm, vma, rb_link, rb_parent);
613 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
614 struct vm_area_struct *prev, struct rb_node **rb_link,
615 struct rb_node *rb_parent)
617 struct address_space *mapping = NULL;
620 mapping = vma->vm_file->f_mapping;
621 i_mmap_lock_write(mapping);
624 __vma_link(mm, vma, prev, rb_link, rb_parent);
625 __vma_link_file(vma);
628 i_mmap_unlock_write(mapping);
635 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
636 * mm's list and rbtree. It has already been inserted into the interval tree.
638 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
640 struct vm_area_struct *prev;
641 struct rb_node **rb_link, *rb_parent;
643 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
644 &prev, &rb_link, &rb_parent))
646 __vma_link(mm, vma, prev, rb_link, rb_parent);
650 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
651 struct vm_area_struct *vma,
652 struct vm_area_struct *prev,
654 struct vm_area_struct *ignore)
656 struct vm_area_struct *next;
658 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
661 prev->vm_next = next;
665 prev->vm_next = next;
670 next->vm_prev = prev;
673 vmacache_invalidate(mm);
676 static inline void __vma_unlink_prev(struct mm_struct *mm,
677 struct vm_area_struct *vma,
678 struct vm_area_struct *prev)
680 __vma_unlink_common(mm, vma, prev, true, vma);
684 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
685 * is already present in an i_mmap tree without adjusting the tree.
686 * The following helper function should be used when such adjustments
687 * are necessary. The "insert" vma (if any) is to be inserted
688 * before we drop the necessary locks.
690 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
691 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
692 struct vm_area_struct *expand)
694 struct mm_struct *mm = vma->vm_mm;
695 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
696 struct address_space *mapping = NULL;
697 struct rb_root_cached *root = NULL;
698 struct anon_vma *anon_vma = NULL;
699 struct file *file = vma->vm_file;
700 bool start_changed = false, end_changed = false;
701 long adjust_next = 0;
704 if (next && !insert) {
705 struct vm_area_struct *exporter = NULL, *importer = NULL;
707 if (end >= next->vm_end) {
709 * vma expands, overlapping all the next, and
710 * perhaps the one after too (mprotect case 6).
711 * The only other cases that gets here are
712 * case 1, case 7 and case 8.
714 if (next == expand) {
716 * The only case where we don't expand "vma"
717 * and we expand "next" instead is case 8.
719 VM_WARN_ON(end != next->vm_end);
721 * remove_next == 3 means we're
722 * removing "vma" and that to do so we
723 * swapped "vma" and "next".
726 VM_WARN_ON(file != next->vm_file);
729 VM_WARN_ON(expand != vma);
731 * case 1, 6, 7, remove_next == 2 is case 6,
732 * remove_next == 1 is case 1 or 7.
734 remove_next = 1 + (end > next->vm_end);
735 VM_WARN_ON(remove_next == 2 &&
736 end != next->vm_next->vm_end);
737 VM_WARN_ON(remove_next == 1 &&
738 end != next->vm_end);
739 /* trim end to next, for case 6 first pass */
747 * If next doesn't have anon_vma, import from vma after
748 * next, if the vma overlaps with it.
750 if (remove_next == 2 && !next->anon_vma)
751 exporter = next->vm_next;
753 } else if (end > next->vm_start) {
755 * vma expands, overlapping part of the next:
756 * mprotect case 5 shifting the boundary up.
758 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
761 VM_WARN_ON(expand != importer);
762 } else if (end < vma->vm_end) {
764 * vma shrinks, and !insert tells it's not
765 * split_vma inserting another: so it must be
766 * mprotect case 4 shifting the boundary down.
768 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
771 VM_WARN_ON(expand != importer);
775 * Easily overlooked: when mprotect shifts the boundary,
776 * make sure the expanding vma has anon_vma set if the
777 * shrinking vma had, to cover any anon pages imported.
779 if (exporter && exporter->anon_vma && !importer->anon_vma) {
782 importer->anon_vma = exporter->anon_vma;
783 error = anon_vma_clone(importer, exporter);
789 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
792 mapping = file->f_mapping;
793 root = &mapping->i_mmap;
794 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
797 uprobe_munmap(next, next->vm_start, next->vm_end);
799 i_mmap_lock_write(mapping);
802 * Put into interval tree now, so instantiated pages
803 * are visible to arm/parisc __flush_dcache_page
804 * throughout; but we cannot insert into address
805 * space until vma start or end is updated.
807 __vma_link_file(insert);
811 anon_vma = vma->anon_vma;
812 if (!anon_vma && adjust_next)
813 anon_vma = next->anon_vma;
815 VM_WARN_ON(adjust_next && next->anon_vma &&
816 anon_vma != next->anon_vma);
817 anon_vma_lock_write(anon_vma);
818 anon_vma_interval_tree_pre_update_vma(vma);
820 anon_vma_interval_tree_pre_update_vma(next);
824 flush_dcache_mmap_lock(mapping);
825 vma_interval_tree_remove(vma, root);
827 vma_interval_tree_remove(next, root);
830 if (start != vma->vm_start) {
831 vma->vm_start = start;
832 start_changed = true;
834 if (end != vma->vm_end) {
838 vma->vm_pgoff = pgoff;
840 next->vm_start += adjust_next << PAGE_SHIFT;
841 next->vm_pgoff += adjust_next;
846 vma_interval_tree_insert(next, root);
847 vma_interval_tree_insert(vma, root);
848 flush_dcache_mmap_unlock(mapping);
853 * vma_merge has merged next into vma, and needs
854 * us to remove next before dropping the locks.
856 if (remove_next != 3)
857 __vma_unlink_prev(mm, next, vma);
860 * vma is not before next if they've been
863 * pre-swap() next->vm_start was reduced so
864 * tell validate_mm_rb to ignore pre-swap()
865 * "next" (which is stored in post-swap()
868 __vma_unlink_common(mm, next, NULL, false, vma);
870 __remove_shared_vm_struct(next, file, mapping);
873 * split_vma has split insert from vma, and needs
874 * us to insert it before dropping the locks
875 * (it may either follow vma or precede it).
877 __insert_vm_struct(mm, insert);
883 mm->highest_vm_end = vm_end_gap(vma);
884 else if (!adjust_next)
885 vma_gap_update(next);
890 anon_vma_interval_tree_post_update_vma(vma);
892 anon_vma_interval_tree_post_update_vma(next);
893 anon_vma_unlock_write(anon_vma);
896 i_mmap_unlock_write(mapping);
907 uprobe_munmap(next, next->vm_start, next->vm_end);
911 anon_vma_merge(vma, next);
913 mpol_put(vma_policy(next));
914 kmem_cache_free(vm_area_cachep, next);
916 * In mprotect's case 6 (see comments on vma_merge),
917 * we must remove another next too. It would clutter
918 * up the code too much to do both in one go.
920 if (remove_next != 3) {
922 * If "next" was removed and vma->vm_end was
923 * expanded (up) over it, in turn
924 * "next->vm_prev->vm_end" changed and the
925 * "vma->vm_next" gap must be updated.
930 * For the scope of the comment "next" and
931 * "vma" considered pre-swap(): if "vma" was
932 * removed, next->vm_start was expanded (down)
933 * over it and the "next" gap must be updated.
934 * Because of the swap() the post-swap() "vma"
935 * actually points to pre-swap() "next"
936 * (post-swap() "next" as opposed is now a
941 if (remove_next == 2) {
947 vma_gap_update(next);
950 * If remove_next == 2 we obviously can't
953 * If remove_next == 3 we can't reach this
954 * path because pre-swap() next is always not
955 * NULL. pre-swap() "next" is not being
956 * removed and its next->vm_end is not altered
957 * (and furthermore "end" already matches
958 * next->vm_end in remove_next == 3).
960 * We reach this only in the remove_next == 1
961 * case if the "next" vma that was removed was
962 * the highest vma of the mm. However in such
963 * case next->vm_end == "end" and the extended
964 * "vma" has vma->vm_end == next->vm_end so
965 * mm->highest_vm_end doesn't need any update
966 * in remove_next == 1 case.
968 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
980 * If the vma has a ->close operation then the driver probably needs to release
981 * per-vma resources, so we don't attempt to merge those.
983 static inline int is_mergeable_vma(struct vm_area_struct *vma,
984 struct file *file, unsigned long vm_flags,
985 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
988 * VM_SOFTDIRTY should not prevent from VMA merging, if we
989 * match the flags but dirty bit -- the caller should mark
990 * merged VMA as dirty. If dirty bit won't be excluded from
991 * comparison, we increase pressue on the memory system forcing
992 * the kernel to generate new VMAs when old one could be
995 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
997 if (vma->vm_file != file)
999 if (vma->vm_ops && vma->vm_ops->close)
1001 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1006 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1007 struct anon_vma *anon_vma2,
1008 struct vm_area_struct *vma)
1011 * The list_is_singular() test is to avoid merging VMA cloned from
1012 * parents. This can improve scalability caused by anon_vma lock.
1014 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1015 list_is_singular(&vma->anon_vma_chain)))
1017 return anon_vma1 == anon_vma2;
1021 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1022 * in front of (at a lower virtual address and file offset than) the vma.
1024 * We cannot merge two vmas if they have differently assigned (non-NULL)
1025 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1027 * We don't check here for the merged mmap wrapping around the end of pagecache
1028 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1029 * wrap, nor mmaps which cover the final page at index -1UL.
1032 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1033 struct anon_vma *anon_vma, struct file *file,
1035 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1037 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1038 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1039 if (vma->vm_pgoff == vm_pgoff)
1046 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1047 * beyond (at a higher virtual address and file offset than) the vma.
1049 * We cannot merge two vmas if they have differently assigned (non-NULL)
1050 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1053 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1054 struct anon_vma *anon_vma, struct file *file,
1056 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1058 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1059 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1061 vm_pglen = vma_pages(vma);
1062 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1069 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1070 * whether that can be merged with its predecessor or its successor.
1071 * Or both (it neatly fills a hole).
1073 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1074 * certain not to be mapped by the time vma_merge is called; but when
1075 * called for mprotect, it is certain to be already mapped (either at
1076 * an offset within prev, or at the start of next), and the flags of
1077 * this area are about to be changed to vm_flags - and the no-change
1078 * case has already been eliminated.
1080 * The following mprotect cases have to be considered, where AAAA is
1081 * the area passed down from mprotect_fixup, never extending beyond one
1082 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1084 * AAAA AAAA AAAA AAAA
1085 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1086 * cannot merge might become might become might become
1087 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1088 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1089 * mremap move: PPPPXXXXXXXX 8
1091 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1092 * might become case 1 below case 2 below case 3 below
1094 * It is important for case 8 that the the vma NNNN overlapping the
1095 * region AAAA is never going to extended over XXXX. Instead XXXX must
1096 * be extended in region AAAA and NNNN must be removed. This way in
1097 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1098 * rmap_locks, the properties of the merged vma will be already
1099 * correct for the whole merged range. Some of those properties like
1100 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1101 * be correct for the whole merged range immediately after the
1102 * rmap_locks are released. Otherwise if XXXX would be removed and
1103 * NNNN would be extended over the XXXX range, remove_migration_ptes
1104 * or other rmap walkers (if working on addresses beyond the "end"
1105 * parameter) may establish ptes with the wrong permissions of NNNN
1106 * instead of the right permissions of XXXX.
1108 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1109 struct vm_area_struct *prev, unsigned long addr,
1110 unsigned long end, unsigned long vm_flags,
1111 struct anon_vma *anon_vma, struct file *file,
1112 pgoff_t pgoff, struct mempolicy *policy,
1113 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1115 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1116 struct vm_area_struct *area, *next;
1120 * We later require that vma->vm_flags == vm_flags,
1121 * so this tests vma->vm_flags & VM_SPECIAL, too.
1123 if (vm_flags & VM_SPECIAL)
1127 next = prev->vm_next;
1131 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1132 next = next->vm_next;
1134 /* verify some invariant that must be enforced by the caller */
1135 VM_WARN_ON(prev && addr <= prev->vm_start);
1136 VM_WARN_ON(area && end > area->vm_end);
1137 VM_WARN_ON(addr >= end);
1140 * Can it merge with the predecessor?
1142 if (prev && prev->vm_end == addr &&
1143 mpol_equal(vma_policy(prev), policy) &&
1144 can_vma_merge_after(prev, vm_flags,
1145 anon_vma, file, pgoff,
1146 vm_userfaultfd_ctx)) {
1148 * OK, it can. Can we now merge in the successor as well?
1150 if (next && end == next->vm_start &&
1151 mpol_equal(policy, vma_policy(next)) &&
1152 can_vma_merge_before(next, vm_flags,
1155 vm_userfaultfd_ctx) &&
1156 is_mergeable_anon_vma(prev->anon_vma,
1157 next->anon_vma, NULL)) {
1159 err = __vma_adjust(prev, prev->vm_start,
1160 next->vm_end, prev->vm_pgoff, NULL,
1162 } else /* cases 2, 5, 7 */
1163 err = __vma_adjust(prev, prev->vm_start,
1164 end, prev->vm_pgoff, NULL, prev);
1167 khugepaged_enter_vma_merge(prev, vm_flags);
1172 * Can this new request be merged in front of next?
1174 if (next && end == next->vm_start &&
1175 mpol_equal(policy, vma_policy(next)) &&
1176 can_vma_merge_before(next, vm_flags,
1177 anon_vma, file, pgoff+pglen,
1178 vm_userfaultfd_ctx)) {
1179 if (prev && addr < prev->vm_end) /* case 4 */
1180 err = __vma_adjust(prev, prev->vm_start,
1181 addr, prev->vm_pgoff, NULL, next);
1182 else { /* cases 3, 8 */
1183 err = __vma_adjust(area, addr, next->vm_end,
1184 next->vm_pgoff - pglen, NULL, next);
1186 * In case 3 area is already equal to next and
1187 * this is a noop, but in case 8 "area" has
1188 * been removed and next was expanded over it.
1194 khugepaged_enter_vma_merge(area, vm_flags);
1202 * Rough compatbility check to quickly see if it's even worth looking
1203 * at sharing an anon_vma.
1205 * They need to have the same vm_file, and the flags can only differ
1206 * in things that mprotect may change.
1208 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1209 * we can merge the two vma's. For example, we refuse to merge a vma if
1210 * there is a vm_ops->close() function, because that indicates that the
1211 * driver is doing some kind of reference counting. But that doesn't
1212 * really matter for the anon_vma sharing case.
1214 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1216 return a->vm_end == b->vm_start &&
1217 mpol_equal(vma_policy(a), vma_policy(b)) &&
1218 a->vm_file == b->vm_file &&
1219 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1220 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1224 * Do some basic sanity checking to see if we can re-use the anon_vma
1225 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1226 * the same as 'old', the other will be the new one that is trying
1227 * to share the anon_vma.
1229 * NOTE! This runs with mm_sem held for reading, so it is possible that
1230 * the anon_vma of 'old' is concurrently in the process of being set up
1231 * by another page fault trying to merge _that_. But that's ok: if it
1232 * is being set up, that automatically means that it will be a singleton
1233 * acceptable for merging, so we can do all of this optimistically. But
1234 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1236 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1237 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1238 * is to return an anon_vma that is "complex" due to having gone through
1241 * We also make sure that the two vma's are compatible (adjacent,
1242 * and with the same memory policies). That's all stable, even with just
1243 * a read lock on the mm_sem.
1245 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1247 if (anon_vma_compatible(a, b)) {
1248 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1250 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1257 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1258 * neighbouring vmas for a suitable anon_vma, before it goes off
1259 * to allocate a new anon_vma. It checks because a repetitive
1260 * sequence of mprotects and faults may otherwise lead to distinct
1261 * anon_vmas being allocated, preventing vma merge in subsequent
1264 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1266 struct anon_vma *anon_vma;
1267 struct vm_area_struct *near;
1269 near = vma->vm_next;
1273 anon_vma = reusable_anon_vma(near, vma, near);
1277 near = vma->vm_prev;
1281 anon_vma = reusable_anon_vma(near, near, vma);
1286 * There's no absolute need to look only at touching neighbours:
1287 * we could search further afield for "compatible" anon_vmas.
1288 * But it would probably just be a waste of time searching,
1289 * or lead to too many vmas hanging off the same anon_vma.
1290 * We're trying to allow mprotect remerging later on,
1291 * not trying to minimize memory used for anon_vmas.
1297 * If a hint addr is less than mmap_min_addr change hint to be as
1298 * low as possible but still greater than mmap_min_addr
1300 static inline unsigned long round_hint_to_min(unsigned long hint)
1303 if (((void *)hint != NULL) &&
1304 (hint < mmap_min_addr))
1305 return PAGE_ALIGN(mmap_min_addr);
1309 static inline int mlock_future_check(struct mm_struct *mm,
1310 unsigned long flags,
1313 unsigned long locked, lock_limit;
1315 /* mlock MCL_FUTURE? */
1316 if (flags & VM_LOCKED) {
1317 locked = len >> PAGE_SHIFT;
1318 locked += mm->locked_vm;
1319 lock_limit = rlimit(RLIMIT_MEMLOCK);
1320 lock_limit >>= PAGE_SHIFT;
1321 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1327 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1329 if (S_ISREG(inode->i_mode))
1330 return inode->i_sb->s_maxbytes;
1332 if (S_ISBLK(inode->i_mode))
1333 return MAX_LFS_FILESIZE;
1335 /* Special "we do even unsigned file positions" case */
1336 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1339 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1343 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1344 unsigned long pgoff, unsigned long len)
1346 u64 maxsize = file_mmap_size_max(file, inode);
1348 if (maxsize && len > maxsize)
1351 if (pgoff > maxsize >> PAGE_SHIFT)
1357 * The caller must hold down_write(¤t->mm->mmap_sem).
1359 unsigned long do_mmap(struct file *file, unsigned long addr,
1360 unsigned long len, unsigned long prot,
1361 unsigned long flags, vm_flags_t vm_flags,
1362 unsigned long pgoff, unsigned long *populate,
1363 struct list_head *uf)
1365 struct mm_struct *mm = current->mm;
1374 * Does the application expect PROT_READ to imply PROT_EXEC?
1376 * (the exception is when the underlying filesystem is noexec
1377 * mounted, in which case we dont add PROT_EXEC.)
1379 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1380 if (!(file && path_noexec(&file->f_path)))
1383 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1384 if (flags & MAP_FIXED_NOREPLACE)
1387 if (!(flags & MAP_FIXED))
1388 addr = round_hint_to_min(addr);
1390 /* Careful about overflows.. */
1391 len = PAGE_ALIGN(len);
1395 /* offset overflow? */
1396 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1399 /* Too many mappings? */
1400 if (mm->map_count > sysctl_max_map_count)
1403 /* Obtain the address to map to. we verify (or select) it and ensure
1404 * that it represents a valid section of the address space.
1406 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1407 if (offset_in_page(addr))
1410 if (flags & MAP_FIXED_NOREPLACE) {
1411 struct vm_area_struct *vma = find_vma(mm, addr);
1413 if (vma && vma->vm_start <= addr)
1417 if (prot == PROT_EXEC) {
1418 pkey = execute_only_pkey(mm);
1423 /* Do simple checking here so the lower-level routines won't have
1424 * to. we assume access permissions have been handled by the open
1425 * of the memory object, so we don't do any here.
1427 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1428 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1430 if (flags & MAP_LOCKED)
1431 if (!can_do_mlock())
1434 if (mlock_future_check(mm, vm_flags, len))
1438 struct inode *inode = file_inode(file);
1439 unsigned long flags_mask;
1441 if (!file_mmap_ok(file, inode, pgoff, len))
1444 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1446 switch (flags & MAP_TYPE) {
1449 * Force use of MAP_SHARED_VALIDATE with non-legacy
1450 * flags. E.g. MAP_SYNC is dangerous to use with
1451 * MAP_SHARED as you don't know which consistency model
1452 * you will get. We silently ignore unsupported flags
1453 * with MAP_SHARED to preserve backward compatibility.
1455 flags &= LEGACY_MAP_MASK;
1457 case MAP_SHARED_VALIDATE:
1458 if (flags & ~flags_mask)
1460 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1464 * Make sure we don't allow writing to an append-only
1467 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1471 * Make sure there are no mandatory locks on the file.
1473 if (locks_verify_locked(file))
1476 vm_flags |= VM_SHARED | VM_MAYSHARE;
1477 if (!(file->f_mode & FMODE_WRITE))
1478 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1482 if (!(file->f_mode & FMODE_READ))
1484 if (path_noexec(&file->f_path)) {
1485 if (vm_flags & VM_EXEC)
1487 vm_flags &= ~VM_MAYEXEC;
1490 if (!file->f_op->mmap)
1492 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1500 switch (flags & MAP_TYPE) {
1502 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1508 vm_flags |= VM_SHARED | VM_MAYSHARE;
1512 * Set pgoff according to addr for anon_vma.
1514 pgoff = addr >> PAGE_SHIFT;
1522 * Set 'VM_NORESERVE' if we should not account for the
1523 * memory use of this mapping.
1525 if (flags & MAP_NORESERVE) {
1526 /* We honor MAP_NORESERVE if allowed to overcommit */
1527 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1528 vm_flags |= VM_NORESERVE;
1530 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1531 if (file && is_file_hugepages(file))
1532 vm_flags |= VM_NORESERVE;
1535 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1536 if (!IS_ERR_VALUE(addr) &&
1537 ((vm_flags & VM_LOCKED) ||
1538 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1543 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1544 unsigned long prot, unsigned long flags,
1545 unsigned long fd, unsigned long pgoff)
1547 struct file *file = NULL;
1548 unsigned long retval;
1550 if (!(flags & MAP_ANONYMOUS)) {
1551 audit_mmap_fd(fd, flags);
1555 if (is_file_hugepages(file))
1556 len = ALIGN(len, huge_page_size(hstate_file(file)));
1558 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1560 } else if (flags & MAP_HUGETLB) {
1561 struct user_struct *user = NULL;
1564 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1568 len = ALIGN(len, huge_page_size(hs));
1570 * VM_NORESERVE is used because the reservations will be
1571 * taken when vm_ops->mmap() is called
1572 * A dummy user value is used because we are not locking
1573 * memory so no accounting is necessary
1575 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1577 &user, HUGETLB_ANONHUGE_INODE,
1578 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1580 return PTR_ERR(file);
1583 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1585 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1592 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1593 unsigned long, prot, unsigned long, flags,
1594 unsigned long, fd, unsigned long, pgoff)
1596 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1599 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1600 struct mmap_arg_struct {
1604 unsigned long flags;
1606 unsigned long offset;
1609 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1611 struct mmap_arg_struct a;
1613 if (copy_from_user(&a, arg, sizeof(a)))
1615 if (offset_in_page(a.offset))
1618 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1619 a.offset >> PAGE_SHIFT);
1621 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1624 * Some shared mappigns will want the pages marked read-only
1625 * to track write events. If so, we'll downgrade vm_page_prot
1626 * to the private version (using protection_map[] without the
1629 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1631 vm_flags_t vm_flags = vma->vm_flags;
1632 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1634 /* If it was private or non-writable, the write bit is already clear */
1635 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1638 /* The backer wishes to know when pages are first written to? */
1639 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1642 /* The open routine did something to the protections that pgprot_modify
1643 * won't preserve? */
1644 if (pgprot_val(vm_page_prot) !=
1645 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1648 /* Do we need to track softdirty? */
1649 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1652 /* Specialty mapping? */
1653 if (vm_flags & VM_PFNMAP)
1656 /* Can the mapping track the dirty pages? */
1657 return vma->vm_file && vma->vm_file->f_mapping &&
1658 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1662 * We account for memory if it's a private writeable mapping,
1663 * not hugepages and VM_NORESERVE wasn't set.
1665 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1668 * hugetlb has its own accounting separate from the core VM
1669 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1671 if (file && is_file_hugepages(file))
1674 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1677 unsigned long mmap_region(struct file *file, unsigned long addr,
1678 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1679 struct list_head *uf)
1681 struct mm_struct *mm = current->mm;
1682 struct vm_area_struct *vma, *prev;
1684 struct rb_node **rb_link, *rb_parent;
1685 unsigned long charged = 0;
1687 /* Check against address space limit. */
1688 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1689 unsigned long nr_pages;
1692 * MAP_FIXED may remove pages of mappings that intersects with
1693 * requested mapping. Account for the pages it would unmap.
1695 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1697 if (!may_expand_vm(mm, vm_flags,
1698 (len >> PAGE_SHIFT) - nr_pages))
1702 /* Clear old maps */
1703 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1705 if (do_munmap(mm, addr, len, uf))
1710 * Private writable mapping: check memory availability
1712 if (accountable_mapping(file, vm_flags)) {
1713 charged = len >> PAGE_SHIFT;
1714 if (security_vm_enough_memory_mm(mm, charged))
1716 vm_flags |= VM_ACCOUNT;
1720 * Can we just expand an old mapping?
1722 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1723 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1728 * Determine the object being mapped and call the appropriate
1729 * specific mapper. the address has already been validated, but
1730 * not unmapped, but the maps are removed from the list.
1732 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1739 vma->vm_start = addr;
1740 vma->vm_end = addr + len;
1741 vma->vm_flags = vm_flags;
1742 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1743 vma->vm_pgoff = pgoff;
1744 INIT_LIST_HEAD(&vma->anon_vma_chain);
1747 if (vm_flags & VM_DENYWRITE) {
1748 error = deny_write_access(file);
1752 if (vm_flags & VM_SHARED) {
1753 error = mapping_map_writable(file->f_mapping);
1755 goto allow_write_and_free_vma;
1758 /* ->mmap() can change vma->vm_file, but must guarantee that
1759 * vma_link() below can deny write-access if VM_DENYWRITE is set
1760 * and map writably if VM_SHARED is set. This usually means the
1761 * new file must not have been exposed to user-space, yet.
1763 vma->vm_file = get_file(file);
1764 error = call_mmap(file, vma);
1766 goto unmap_and_free_vma;
1768 /* Can addr have changed??
1770 * Answer: Yes, several device drivers can do it in their
1771 * f_op->mmap method. -DaveM
1772 * Bug: If addr is changed, prev, rb_link, rb_parent should
1773 * be updated for vma_link()
1775 WARN_ON_ONCE(addr != vma->vm_start);
1777 addr = vma->vm_start;
1778 vm_flags = vma->vm_flags;
1779 } else if (vm_flags & VM_SHARED) {
1780 error = shmem_zero_setup(vma);
1785 vma_link(mm, vma, prev, rb_link, rb_parent);
1786 /* Once vma denies write, undo our temporary denial count */
1788 if (vm_flags & VM_SHARED)
1789 mapping_unmap_writable(file->f_mapping);
1790 if (vm_flags & VM_DENYWRITE)
1791 allow_write_access(file);
1793 file = vma->vm_file;
1795 perf_event_mmap(vma);
1797 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1798 if (vm_flags & VM_LOCKED) {
1799 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1800 vma == get_gate_vma(current->mm)))
1801 mm->locked_vm += (len >> PAGE_SHIFT);
1803 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1810 * New (or expanded) vma always get soft dirty status.
1811 * Otherwise user-space soft-dirty page tracker won't
1812 * be able to distinguish situation when vma area unmapped,
1813 * then new mapped in-place (which must be aimed as
1814 * a completely new data area).
1816 vma->vm_flags |= VM_SOFTDIRTY;
1818 vma_set_page_prot(vma);
1823 vma->vm_file = NULL;
1826 /* Undo any partial mapping done by a device driver. */
1827 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1829 if (vm_flags & VM_SHARED)
1830 mapping_unmap_writable(file->f_mapping);
1831 allow_write_and_free_vma:
1832 if (vm_flags & VM_DENYWRITE)
1833 allow_write_access(file);
1835 kmem_cache_free(vm_area_cachep, vma);
1838 vm_unacct_memory(charged);
1842 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1845 * We implement the search by looking for an rbtree node that
1846 * immediately follows a suitable gap. That is,
1847 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1848 * - gap_end = vma->vm_start >= info->low_limit + length;
1849 * - gap_end - gap_start >= length
1852 struct mm_struct *mm = current->mm;
1853 struct vm_area_struct *vma;
1854 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1856 /* Adjust search length to account for worst case alignment overhead */
1857 length = info->length + info->align_mask;
1858 if (length < info->length)
1861 /* Adjust search limits by the desired length */
1862 if (info->high_limit < length)
1864 high_limit = info->high_limit - length;
1866 if (info->low_limit > high_limit)
1868 low_limit = info->low_limit + length;
1870 /* Check if rbtree root looks promising */
1871 if (RB_EMPTY_ROOT(&mm->mm_rb))
1873 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1874 if (vma->rb_subtree_gap < length)
1878 /* Visit left subtree if it looks promising */
1879 gap_end = vm_start_gap(vma);
1880 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1881 struct vm_area_struct *left =
1882 rb_entry(vma->vm_rb.rb_left,
1883 struct vm_area_struct, vm_rb);
1884 if (left->rb_subtree_gap >= length) {
1890 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1892 /* Check if current node has a suitable gap */
1893 if (gap_start > high_limit)
1895 if (gap_end >= low_limit &&
1896 gap_end > gap_start && gap_end - gap_start >= length)
1899 /* Visit right subtree if it looks promising */
1900 if (vma->vm_rb.rb_right) {
1901 struct vm_area_struct *right =
1902 rb_entry(vma->vm_rb.rb_right,
1903 struct vm_area_struct, vm_rb);
1904 if (right->rb_subtree_gap >= length) {
1910 /* Go back up the rbtree to find next candidate node */
1912 struct rb_node *prev = &vma->vm_rb;
1913 if (!rb_parent(prev))
1915 vma = rb_entry(rb_parent(prev),
1916 struct vm_area_struct, vm_rb);
1917 if (prev == vma->vm_rb.rb_left) {
1918 gap_start = vm_end_gap(vma->vm_prev);
1919 gap_end = vm_start_gap(vma);
1926 /* Check highest gap, which does not precede any rbtree node */
1927 gap_start = mm->highest_vm_end;
1928 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1929 if (gap_start > high_limit)
1933 /* We found a suitable gap. Clip it with the original low_limit. */
1934 if (gap_start < info->low_limit)
1935 gap_start = info->low_limit;
1937 /* Adjust gap address to the desired alignment */
1938 gap_start += (info->align_offset - gap_start) & info->align_mask;
1940 VM_BUG_ON(gap_start + info->length > info->high_limit);
1941 VM_BUG_ON(gap_start + info->length > gap_end);
1945 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1947 struct mm_struct *mm = current->mm;
1948 struct vm_area_struct *vma;
1949 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1951 /* Adjust search length to account for worst case alignment overhead */
1952 length = info->length + info->align_mask;
1953 if (length < info->length)
1957 * Adjust search limits by the desired length.
1958 * See implementation comment at top of unmapped_area().
1960 gap_end = info->high_limit;
1961 if (gap_end < length)
1963 high_limit = gap_end - length;
1965 if (info->low_limit > high_limit)
1967 low_limit = info->low_limit + length;
1969 /* Check highest gap, which does not precede any rbtree node */
1970 gap_start = mm->highest_vm_end;
1971 if (gap_start <= high_limit)
1974 /* Check if rbtree root looks promising */
1975 if (RB_EMPTY_ROOT(&mm->mm_rb))
1977 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1978 if (vma->rb_subtree_gap < length)
1982 /* Visit right subtree if it looks promising */
1983 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1984 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1985 struct vm_area_struct *right =
1986 rb_entry(vma->vm_rb.rb_right,
1987 struct vm_area_struct, vm_rb);
1988 if (right->rb_subtree_gap >= length) {
1995 /* Check if current node has a suitable gap */
1996 gap_end = vm_start_gap(vma);
1997 if (gap_end < low_limit)
1999 if (gap_start <= high_limit &&
2000 gap_end > gap_start && gap_end - gap_start >= length)
2003 /* Visit left subtree if it looks promising */
2004 if (vma->vm_rb.rb_left) {
2005 struct vm_area_struct *left =
2006 rb_entry(vma->vm_rb.rb_left,
2007 struct vm_area_struct, vm_rb);
2008 if (left->rb_subtree_gap >= length) {
2014 /* Go back up the rbtree to find next candidate node */
2016 struct rb_node *prev = &vma->vm_rb;
2017 if (!rb_parent(prev))
2019 vma = rb_entry(rb_parent(prev),
2020 struct vm_area_struct, vm_rb);
2021 if (prev == vma->vm_rb.rb_right) {
2022 gap_start = vma->vm_prev ?
2023 vm_end_gap(vma->vm_prev) : 0;
2030 /* We found a suitable gap. Clip it with the original high_limit. */
2031 if (gap_end > info->high_limit)
2032 gap_end = info->high_limit;
2035 /* Compute highest gap address at the desired alignment */
2036 gap_end -= info->length;
2037 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2039 VM_BUG_ON(gap_end < info->low_limit);
2040 VM_BUG_ON(gap_end < gap_start);
2044 /* Get an address range which is currently unmapped.
2045 * For shmat() with addr=0.
2047 * Ugly calling convention alert:
2048 * Return value with the low bits set means error value,
2050 * if (ret & ~PAGE_MASK)
2053 * This function "knows" that -ENOMEM has the bits set.
2055 #ifndef HAVE_ARCH_UNMAPPED_AREA
2057 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2058 unsigned long len, unsigned long pgoff, unsigned long flags)
2060 struct mm_struct *mm = current->mm;
2061 struct vm_area_struct *vma, *prev;
2062 struct vm_unmapped_area_info info;
2064 if (len > TASK_SIZE - mmap_min_addr)
2067 if (flags & MAP_FIXED)
2071 addr = PAGE_ALIGN(addr);
2072 vma = find_vma_prev(mm, addr, &prev);
2073 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2074 (!vma || addr + len <= vm_start_gap(vma)) &&
2075 (!prev || addr >= vm_end_gap(prev)))
2081 info.low_limit = mm->mmap_base;
2082 info.high_limit = TASK_SIZE;
2083 info.align_mask = 0;
2084 return vm_unmapped_area(&info);
2089 * This mmap-allocator allocates new areas top-down from below the
2090 * stack's low limit (the base):
2092 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2094 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2095 const unsigned long len, const unsigned long pgoff,
2096 const unsigned long flags)
2098 struct vm_area_struct *vma, *prev;
2099 struct mm_struct *mm = current->mm;
2100 unsigned long addr = addr0;
2101 struct vm_unmapped_area_info info;
2103 /* requested length too big for entire address space */
2104 if (len > TASK_SIZE - mmap_min_addr)
2107 if (flags & MAP_FIXED)
2110 /* requesting a specific address */
2112 addr = PAGE_ALIGN(addr);
2113 vma = find_vma_prev(mm, addr, &prev);
2114 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2115 (!vma || addr + len <= vm_start_gap(vma)) &&
2116 (!prev || addr >= vm_end_gap(prev)))
2120 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2122 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2123 info.high_limit = mm->mmap_base;
2124 info.align_mask = 0;
2125 addr = vm_unmapped_area(&info);
2128 * A failed mmap() very likely causes application failure,
2129 * so fall back to the bottom-up function here. This scenario
2130 * can happen with large stack limits and large mmap()
2133 if (offset_in_page(addr)) {
2134 VM_BUG_ON(addr != -ENOMEM);
2136 info.low_limit = TASK_UNMAPPED_BASE;
2137 info.high_limit = TASK_SIZE;
2138 addr = vm_unmapped_area(&info);
2146 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2147 unsigned long pgoff, unsigned long flags)
2149 unsigned long (*get_area)(struct file *, unsigned long,
2150 unsigned long, unsigned long, unsigned long);
2152 unsigned long error = arch_mmap_check(addr, len, flags);
2156 /* Careful about overflows.. */
2157 if (len > TASK_SIZE)
2160 get_area = current->mm->get_unmapped_area;
2162 if (file->f_op->get_unmapped_area)
2163 get_area = file->f_op->get_unmapped_area;
2164 } else if (flags & MAP_SHARED) {
2166 * mmap_region() will call shmem_zero_setup() to create a file,
2167 * so use shmem's get_unmapped_area in case it can be huge.
2168 * do_mmap_pgoff() will clear pgoff, so match alignment.
2171 get_area = shmem_get_unmapped_area;
2174 addr = get_area(file, addr, len, pgoff, flags);
2175 if (IS_ERR_VALUE(addr))
2178 if (addr > TASK_SIZE - len)
2180 if (offset_in_page(addr))
2183 error = security_mmap_addr(addr);
2184 return error ? error : addr;
2187 EXPORT_SYMBOL(get_unmapped_area);
2189 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2190 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2192 struct rb_node *rb_node;
2193 struct vm_area_struct *vma;
2195 /* Check the cache first. */
2196 vma = vmacache_find(mm, addr);
2200 rb_node = mm->mm_rb.rb_node;
2203 struct vm_area_struct *tmp;
2205 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2207 if (tmp->vm_end > addr) {
2209 if (tmp->vm_start <= addr)
2211 rb_node = rb_node->rb_left;
2213 rb_node = rb_node->rb_right;
2217 vmacache_update(addr, vma);
2221 EXPORT_SYMBOL(find_vma);
2224 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2226 struct vm_area_struct *
2227 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2228 struct vm_area_struct **pprev)
2230 struct vm_area_struct *vma;
2232 vma = find_vma(mm, addr);
2234 *pprev = vma->vm_prev;
2236 struct rb_node *rb_node = mm->mm_rb.rb_node;
2239 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2240 rb_node = rb_node->rb_right;
2247 * Verify that the stack growth is acceptable and
2248 * update accounting. This is shared with both the
2249 * grow-up and grow-down cases.
2251 static int acct_stack_growth(struct vm_area_struct *vma,
2252 unsigned long size, unsigned long grow)
2254 struct mm_struct *mm = vma->vm_mm;
2255 unsigned long new_start;
2257 /* address space limit tests */
2258 if (!may_expand_vm(mm, vma->vm_flags, grow))
2261 /* Stack limit test */
2262 if (size > rlimit(RLIMIT_STACK))
2265 /* mlock limit tests */
2266 if (vma->vm_flags & VM_LOCKED) {
2267 unsigned long locked;
2268 unsigned long limit;
2269 locked = mm->locked_vm + grow;
2270 limit = rlimit(RLIMIT_MEMLOCK);
2271 limit >>= PAGE_SHIFT;
2272 if (locked > limit && !capable(CAP_IPC_LOCK))
2276 /* Check to ensure the stack will not grow into a hugetlb-only region */
2277 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2279 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2283 * Overcommit.. This must be the final test, as it will
2284 * update security statistics.
2286 if (security_vm_enough_memory_mm(mm, grow))
2292 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2294 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2295 * vma is the last one with address > vma->vm_end. Have to extend vma.
2297 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2299 struct mm_struct *mm = vma->vm_mm;
2300 struct vm_area_struct *next;
2301 unsigned long gap_addr;
2304 if (!(vma->vm_flags & VM_GROWSUP))
2307 /* Guard against exceeding limits of the address space. */
2308 address &= PAGE_MASK;
2309 if (address >= (TASK_SIZE & PAGE_MASK))
2311 address += PAGE_SIZE;
2313 /* Enforce stack_guard_gap */
2314 gap_addr = address + stack_guard_gap;
2316 /* Guard against overflow */
2317 if (gap_addr < address || gap_addr > TASK_SIZE)
2318 gap_addr = TASK_SIZE;
2320 next = vma->vm_next;
2321 if (next && next->vm_start < gap_addr &&
2322 (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2323 if (!(next->vm_flags & VM_GROWSUP))
2325 /* Check that both stack segments have the same anon_vma? */
2328 /* We must make sure the anon_vma is allocated. */
2329 if (unlikely(anon_vma_prepare(vma)))
2333 * vma->vm_start/vm_end cannot change under us because the caller
2334 * is required to hold the mmap_sem in read mode. We need the
2335 * anon_vma lock to serialize against concurrent expand_stacks.
2337 anon_vma_lock_write(vma->anon_vma);
2339 /* Somebody else might have raced and expanded it already */
2340 if (address > vma->vm_end) {
2341 unsigned long size, grow;
2343 size = address - vma->vm_start;
2344 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2347 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2348 error = acct_stack_growth(vma, size, grow);
2351 * vma_gap_update() doesn't support concurrent
2352 * updates, but we only hold a shared mmap_sem
2353 * lock here, so we need to protect against
2354 * concurrent vma expansions.
2355 * anon_vma_lock_write() doesn't help here, as
2356 * we don't guarantee that all growable vmas
2357 * in a mm share the same root anon vma.
2358 * So, we reuse mm->page_table_lock to guard
2359 * against concurrent vma expansions.
2361 spin_lock(&mm->page_table_lock);
2362 if (vma->vm_flags & VM_LOCKED)
2363 mm->locked_vm += grow;
2364 vm_stat_account(mm, vma->vm_flags, grow);
2365 anon_vma_interval_tree_pre_update_vma(vma);
2366 vma->vm_end = address;
2367 anon_vma_interval_tree_post_update_vma(vma);
2369 vma_gap_update(vma->vm_next);
2371 mm->highest_vm_end = vm_end_gap(vma);
2372 spin_unlock(&mm->page_table_lock);
2374 perf_event_mmap(vma);
2378 anon_vma_unlock_write(vma->anon_vma);
2379 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2383 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2386 * vma is the first one with address < vma->vm_start. Have to extend vma.
2388 int expand_downwards(struct vm_area_struct *vma,
2389 unsigned long address)
2391 struct mm_struct *mm = vma->vm_mm;
2392 struct vm_area_struct *prev;
2395 address &= PAGE_MASK;
2396 error = security_mmap_addr(address);
2400 /* Enforce stack_guard_gap */
2401 prev = vma->vm_prev;
2402 /* Check that both stack segments have the same anon_vma? */
2403 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2404 (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2405 if (address - prev->vm_end < stack_guard_gap)
2409 /* We must make sure the anon_vma is allocated. */
2410 if (unlikely(anon_vma_prepare(vma)))
2414 * vma->vm_start/vm_end cannot change under us because the caller
2415 * is required to hold the mmap_sem in read mode. We need the
2416 * anon_vma lock to serialize against concurrent expand_stacks.
2418 anon_vma_lock_write(vma->anon_vma);
2420 /* Somebody else might have raced and expanded it already */
2421 if (address < vma->vm_start) {
2422 unsigned long size, grow;
2424 size = vma->vm_end - address;
2425 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2428 if (grow <= vma->vm_pgoff) {
2429 error = acct_stack_growth(vma, size, grow);
2432 * vma_gap_update() doesn't support concurrent
2433 * updates, but we only hold a shared mmap_sem
2434 * lock here, so we need to protect against
2435 * concurrent vma expansions.
2436 * anon_vma_lock_write() doesn't help here, as
2437 * we don't guarantee that all growable vmas
2438 * in a mm share the same root anon vma.
2439 * So, we reuse mm->page_table_lock to guard
2440 * against concurrent vma expansions.
2442 spin_lock(&mm->page_table_lock);
2443 if (vma->vm_flags & VM_LOCKED)
2444 mm->locked_vm += grow;
2445 vm_stat_account(mm, vma->vm_flags, grow);
2446 anon_vma_interval_tree_pre_update_vma(vma);
2447 vma->vm_start = address;
2448 vma->vm_pgoff -= grow;
2449 anon_vma_interval_tree_post_update_vma(vma);
2450 vma_gap_update(vma);
2451 spin_unlock(&mm->page_table_lock);
2453 perf_event_mmap(vma);
2457 anon_vma_unlock_write(vma->anon_vma);
2458 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2463 /* enforced gap between the expanding stack and other mappings. */
2464 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2466 static int __init cmdline_parse_stack_guard_gap(char *p)
2471 val = simple_strtoul(p, &endptr, 10);
2473 stack_guard_gap = val << PAGE_SHIFT;
2477 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2479 #ifdef CONFIG_STACK_GROWSUP
2480 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2482 return expand_upwards(vma, address);
2485 struct vm_area_struct *
2486 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2488 struct vm_area_struct *vma, *prev;
2491 vma = find_vma_prev(mm, addr, &prev);
2492 if (vma && (vma->vm_start <= addr))
2494 if (!prev || expand_stack(prev, addr))
2496 if (prev->vm_flags & VM_LOCKED)
2497 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2501 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2503 return expand_downwards(vma, address);
2506 struct vm_area_struct *
2507 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2509 struct vm_area_struct *vma;
2510 unsigned long start;
2513 vma = find_vma(mm, addr);
2516 if (vma->vm_start <= addr)
2518 if (!(vma->vm_flags & VM_GROWSDOWN))
2520 start = vma->vm_start;
2521 if (expand_stack(vma, addr))
2523 if (vma->vm_flags & VM_LOCKED)
2524 populate_vma_page_range(vma, addr, start, NULL);
2529 EXPORT_SYMBOL_GPL(find_extend_vma);
2532 * Ok - we have the memory areas we should free on the vma list,
2533 * so release them, and do the vma updates.
2535 * Called with the mm semaphore held.
2537 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2539 unsigned long nr_accounted = 0;
2541 /* Update high watermark before we lower total_vm */
2542 update_hiwater_vm(mm);
2544 long nrpages = vma_pages(vma);
2546 if (vma->vm_flags & VM_ACCOUNT)
2547 nr_accounted += nrpages;
2548 vm_stat_account(mm, vma->vm_flags, -nrpages);
2549 vma = remove_vma(vma);
2551 vm_unacct_memory(nr_accounted);
2556 * Get rid of page table information in the indicated region.
2558 * Called with the mm semaphore held.
2560 static void unmap_region(struct mm_struct *mm,
2561 struct vm_area_struct *vma, struct vm_area_struct *prev,
2562 unsigned long start, unsigned long end)
2564 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2565 struct mmu_gather tlb;
2568 tlb_gather_mmu(&tlb, mm, start, end);
2569 update_hiwater_rss(mm);
2570 unmap_vmas(&tlb, vma, start, end);
2571 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2572 next ? next->vm_start : USER_PGTABLES_CEILING);
2573 tlb_finish_mmu(&tlb, start, end);
2577 * Create a list of vma's touched by the unmap, removing them from the mm's
2578 * vma list as we go..
2581 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2582 struct vm_area_struct *prev, unsigned long end)
2584 struct vm_area_struct **insertion_point;
2585 struct vm_area_struct *tail_vma = NULL;
2587 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2588 vma->vm_prev = NULL;
2590 vma_rb_erase(vma, &mm->mm_rb);
2594 } while (vma && vma->vm_start < end);
2595 *insertion_point = vma;
2597 vma->vm_prev = prev;
2598 vma_gap_update(vma);
2600 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2601 tail_vma->vm_next = NULL;
2603 /* Kill the cache */
2604 vmacache_invalidate(mm);
2608 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2609 * has already been checked or doesn't make sense to fail.
2611 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2612 unsigned long addr, int new_below)
2614 struct vm_area_struct *new;
2617 if (vma->vm_ops && vma->vm_ops->split) {
2618 err = vma->vm_ops->split(vma, addr);
2623 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2627 /* most fields are the same, copy all, and then fixup */
2630 INIT_LIST_HEAD(&new->anon_vma_chain);
2635 new->vm_start = addr;
2636 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2639 err = vma_dup_policy(vma, new);
2643 err = anon_vma_clone(new, vma);
2648 get_file(new->vm_file);
2650 if (new->vm_ops && new->vm_ops->open)
2651 new->vm_ops->open(new);
2654 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2655 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2657 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2663 /* Clean everything up if vma_adjust failed. */
2664 if (new->vm_ops && new->vm_ops->close)
2665 new->vm_ops->close(new);
2668 unlink_anon_vmas(new);
2670 mpol_put(vma_policy(new));
2672 kmem_cache_free(vm_area_cachep, new);
2677 * Split a vma into two pieces at address 'addr', a new vma is allocated
2678 * either for the first part or the tail.
2680 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2681 unsigned long addr, int new_below)
2683 if (mm->map_count >= sysctl_max_map_count)
2686 return __split_vma(mm, vma, addr, new_below);
2689 /* Munmap is split into 2 main parts -- this part which finds
2690 * what needs doing, and the areas themselves, which do the
2691 * work. This now handles partial unmappings.
2692 * Jeremy Fitzhardinge <jeremy@goop.org>
2694 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2695 struct list_head *uf)
2698 struct vm_area_struct *vma, *prev, *last;
2700 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2703 len = PAGE_ALIGN(len);
2707 /* Find the first overlapping VMA */
2708 vma = find_vma(mm, start);
2711 prev = vma->vm_prev;
2712 /* we have start < vma->vm_end */
2714 /* if it doesn't overlap, we have nothing.. */
2716 if (vma->vm_start >= end)
2720 * If we need to split any vma, do it now to save pain later.
2722 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2723 * unmapped vm_area_struct will remain in use: so lower split_vma
2724 * places tmp vma above, and higher split_vma places tmp vma below.
2726 if (start > vma->vm_start) {
2730 * Make sure that map_count on return from munmap() will
2731 * not exceed its limit; but let map_count go just above
2732 * its limit temporarily, to help free resources as expected.
2734 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2737 error = __split_vma(mm, vma, start, 0);
2743 /* Does it split the last one? */
2744 last = find_vma(mm, end);
2745 if (last && end > last->vm_start) {
2746 int error = __split_vma(mm, last, end, 1);
2750 vma = prev ? prev->vm_next : mm->mmap;
2754 * If userfaultfd_unmap_prep returns an error the vmas
2755 * will remain splitted, but userland will get a
2756 * highly unexpected error anyway. This is no
2757 * different than the case where the first of the two
2758 * __split_vma fails, but we don't undo the first
2759 * split, despite we could. This is unlikely enough
2760 * failure that it's not worth optimizing it for.
2762 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2768 * unlock any mlock()ed ranges before detaching vmas
2770 if (mm->locked_vm) {
2771 struct vm_area_struct *tmp = vma;
2772 while (tmp && tmp->vm_start < end) {
2773 if (tmp->vm_flags & VM_LOCKED) {
2774 mm->locked_vm -= vma_pages(tmp);
2775 munlock_vma_pages_all(tmp);
2782 * Remove the vma's, and unmap the actual pages
2784 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2785 unmap_region(mm, vma, prev, start, end);
2787 arch_unmap(mm, vma, start, end);
2789 /* Fix up all other VM information */
2790 remove_vma_list(mm, vma);
2795 int vm_munmap(unsigned long start, size_t len)
2798 struct mm_struct *mm = current->mm;
2801 if (down_write_killable(&mm->mmap_sem))
2804 ret = do_munmap(mm, start, len, &uf);
2805 up_write(&mm->mmap_sem);
2806 userfaultfd_unmap_complete(mm, &uf);
2809 EXPORT_SYMBOL(vm_munmap);
2811 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2813 profile_munmap(addr);
2814 return vm_munmap(addr, len);
2819 * Emulation of deprecated remap_file_pages() syscall.
2821 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2822 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2825 struct mm_struct *mm = current->mm;
2826 struct vm_area_struct *vma;
2827 unsigned long populate = 0;
2828 unsigned long ret = -EINVAL;
2831 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2832 current->comm, current->pid);
2836 start = start & PAGE_MASK;
2837 size = size & PAGE_MASK;
2839 if (start + size <= start)
2842 /* Does pgoff wrap? */
2843 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2846 if (down_write_killable(&mm->mmap_sem))
2849 vma = find_vma(mm, start);
2851 if (!vma || !(vma->vm_flags & VM_SHARED))
2854 if (start < vma->vm_start)
2857 if (start + size > vma->vm_end) {
2858 struct vm_area_struct *next;
2860 for (next = vma->vm_next; next; next = next->vm_next) {
2861 /* hole between vmas ? */
2862 if (next->vm_start != next->vm_prev->vm_end)
2865 if (next->vm_file != vma->vm_file)
2868 if (next->vm_flags != vma->vm_flags)
2871 if (start + size <= next->vm_end)
2879 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2880 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2881 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2883 flags &= MAP_NONBLOCK;
2884 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2885 if (vma->vm_flags & VM_LOCKED) {
2886 struct vm_area_struct *tmp;
2887 flags |= MAP_LOCKED;
2889 /* drop PG_Mlocked flag for over-mapped range */
2890 for (tmp = vma; tmp->vm_start >= start + size;
2891 tmp = tmp->vm_next) {
2893 * Split pmd and munlock page on the border
2896 vma_adjust_trans_huge(tmp, start, start + size, 0);
2898 munlock_vma_pages_range(tmp,
2899 max(tmp->vm_start, start),
2900 min(tmp->vm_end, start + size));
2904 file = get_file(vma->vm_file);
2905 ret = do_mmap_pgoff(vma->vm_file, start, size,
2906 prot, flags, pgoff, &populate, NULL);
2909 up_write(&mm->mmap_sem);
2911 mm_populate(ret, populate);
2912 if (!IS_ERR_VALUE(ret))
2917 static inline void verify_mm_writelocked(struct mm_struct *mm)
2919 #ifdef CONFIG_DEBUG_VM
2920 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2922 up_read(&mm->mmap_sem);
2928 * this is really a simplified "do_mmap". it only handles
2929 * anonymous maps. eventually we may be able to do some
2930 * brk-specific accounting here.
2932 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags, struct list_head *uf)
2934 struct mm_struct *mm = current->mm;
2935 struct vm_area_struct *vma, *prev;
2937 struct rb_node **rb_link, *rb_parent;
2938 pgoff_t pgoff = addr >> PAGE_SHIFT;
2941 len = PAGE_ALIGN(request);
2947 /* Until we need other flags, refuse anything except VM_EXEC. */
2948 if ((flags & (~VM_EXEC)) != 0)
2950 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2952 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2953 if (offset_in_page(error))
2956 error = mlock_future_check(mm, mm->def_flags, len);
2961 * mm->mmap_sem is required to protect against another thread
2962 * changing the mappings in case we sleep.
2964 verify_mm_writelocked(mm);
2967 * Clear old maps. this also does some error checking for us
2969 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2971 if (do_munmap(mm, addr, len, uf))
2975 /* Check against address space limits *after* clearing old maps... */
2976 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2979 if (mm->map_count > sysctl_max_map_count)
2982 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2985 /* Can we just expand an old private anonymous mapping? */
2986 vma = vma_merge(mm, prev, addr, addr + len, flags,
2987 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2992 * create a vma struct for an anonymous mapping
2994 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2996 vm_unacct_memory(len >> PAGE_SHIFT);
3000 INIT_LIST_HEAD(&vma->anon_vma_chain);
3002 vma->vm_start = addr;
3003 vma->vm_end = addr + len;
3004 vma->vm_pgoff = pgoff;
3005 vma->vm_flags = flags;
3006 vma->vm_page_prot = vm_get_page_prot(flags);
3007 vma_link(mm, vma, prev, rb_link, rb_parent);
3009 perf_event_mmap(vma);
3010 mm->total_vm += len >> PAGE_SHIFT;
3011 mm->data_vm += len >> PAGE_SHIFT;
3012 if (flags & VM_LOCKED)
3013 mm->locked_vm += (len >> PAGE_SHIFT);
3014 vma->vm_flags |= VM_SOFTDIRTY;
3018 static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf)
3020 return do_brk_flags(addr, len, 0, uf);
3023 int vm_brk_flags(unsigned long addr, unsigned long len, unsigned long flags)
3025 struct mm_struct *mm = current->mm;
3030 if (down_write_killable(&mm->mmap_sem))
3033 ret = do_brk_flags(addr, len, flags, &uf);
3034 populate = ((mm->def_flags & VM_LOCKED) != 0);
3035 up_write(&mm->mmap_sem);
3036 userfaultfd_unmap_complete(mm, &uf);
3037 if (populate && !ret)
3038 mm_populate(addr, len);
3041 EXPORT_SYMBOL(vm_brk_flags);
3043 int vm_brk(unsigned long addr, unsigned long len)
3045 return vm_brk_flags(addr, len, 0);
3047 EXPORT_SYMBOL(vm_brk);
3049 /* Release all mmaps. */
3050 void exit_mmap(struct mm_struct *mm)
3052 struct mmu_gather tlb;
3053 struct vm_area_struct *vma;
3054 unsigned long nr_accounted = 0;
3056 /* mm's last user has gone, and its about to be pulled down */
3057 mmu_notifier_release(mm);
3059 if (unlikely(mm_is_oom_victim(mm))) {
3061 * Manually reap the mm to free as much memory as possible.
3062 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3063 * this mm from further consideration. Taking mm->mmap_sem for
3064 * write after setting MMF_OOM_SKIP will guarantee that the oom
3065 * reaper will not run on this mm again after mmap_sem is
3068 * Nothing can be holding mm->mmap_sem here and the above call
3069 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3070 * __oom_reap_task_mm() will not block.
3072 * This needs to be done before calling munlock_vma_pages_all(),
3073 * which clears VM_LOCKED, otherwise the oom reaper cannot
3076 mutex_lock(&oom_lock);
3077 __oom_reap_task_mm(mm);
3078 mutex_unlock(&oom_lock);
3080 set_bit(MMF_OOM_SKIP, &mm->flags);
3081 down_write(&mm->mmap_sem);
3082 up_write(&mm->mmap_sem);
3085 if (mm->locked_vm) {
3088 if (vma->vm_flags & VM_LOCKED)
3089 munlock_vma_pages_all(vma);
3097 if (!vma) /* Can happen if dup_mmap() received an OOM */
3102 tlb_gather_mmu(&tlb, mm, 0, -1);
3103 /* update_hiwater_rss(mm) here? but nobody should be looking */
3104 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3105 unmap_vmas(&tlb, vma, 0, -1);
3106 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3107 tlb_finish_mmu(&tlb, 0, -1);
3110 * Walk the list again, actually closing and freeing it,
3111 * with preemption enabled, without holding any MM locks.
3114 if (vma->vm_flags & VM_ACCOUNT)
3115 nr_accounted += vma_pages(vma);
3116 vma = remove_vma(vma);
3118 vm_unacct_memory(nr_accounted);
3121 /* Insert vm structure into process list sorted by address
3122 * and into the inode's i_mmap tree. If vm_file is non-NULL
3123 * then i_mmap_rwsem is taken here.
3125 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3127 struct vm_area_struct *prev;
3128 struct rb_node **rb_link, *rb_parent;
3130 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3131 &prev, &rb_link, &rb_parent))
3133 if ((vma->vm_flags & VM_ACCOUNT) &&
3134 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3138 * The vm_pgoff of a purely anonymous vma should be irrelevant
3139 * until its first write fault, when page's anon_vma and index
3140 * are set. But now set the vm_pgoff it will almost certainly
3141 * end up with (unless mremap moves it elsewhere before that
3142 * first wfault), so /proc/pid/maps tells a consistent story.
3144 * By setting it to reflect the virtual start address of the
3145 * vma, merges and splits can happen in a seamless way, just
3146 * using the existing file pgoff checks and manipulations.
3147 * Similarly in do_mmap_pgoff and in do_brk.
3149 if (vma_is_anonymous(vma)) {
3150 BUG_ON(vma->anon_vma);
3151 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3154 vma_link(mm, vma, prev, rb_link, rb_parent);
3159 * Copy the vma structure to a new location in the same mm,
3160 * prior to moving page table entries, to effect an mremap move.
3162 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3163 unsigned long addr, unsigned long len, pgoff_t pgoff,
3164 bool *need_rmap_locks)
3166 struct vm_area_struct *vma = *vmap;
3167 unsigned long vma_start = vma->vm_start;
3168 struct mm_struct *mm = vma->vm_mm;
3169 struct vm_area_struct *new_vma, *prev;
3170 struct rb_node **rb_link, *rb_parent;
3171 bool faulted_in_anon_vma = true;
3174 * If anonymous vma has not yet been faulted, update new pgoff
3175 * to match new location, to increase its chance of merging.
3177 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3178 pgoff = addr >> PAGE_SHIFT;
3179 faulted_in_anon_vma = false;
3182 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3183 return NULL; /* should never get here */
3184 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3185 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3186 vma->vm_userfaultfd_ctx);
3189 * Source vma may have been merged into new_vma
3191 if (unlikely(vma_start >= new_vma->vm_start &&
3192 vma_start < new_vma->vm_end)) {
3194 * The only way we can get a vma_merge with
3195 * self during an mremap is if the vma hasn't
3196 * been faulted in yet and we were allowed to
3197 * reset the dst vma->vm_pgoff to the
3198 * destination address of the mremap to allow
3199 * the merge to happen. mremap must change the
3200 * vm_pgoff linearity between src and dst vmas
3201 * (in turn preventing a vma_merge) to be
3202 * safe. It is only safe to keep the vm_pgoff
3203 * linear if there are no pages mapped yet.
3205 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3206 *vmap = vma = new_vma;
3208 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3210 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3214 new_vma->vm_start = addr;
3215 new_vma->vm_end = addr + len;
3216 new_vma->vm_pgoff = pgoff;
3217 if (vma_dup_policy(vma, new_vma))
3219 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3220 if (anon_vma_clone(new_vma, vma))
3221 goto out_free_mempol;
3222 if (new_vma->vm_file)
3223 get_file(new_vma->vm_file);
3224 if (new_vma->vm_ops && new_vma->vm_ops->open)
3225 new_vma->vm_ops->open(new_vma);
3226 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3227 *need_rmap_locks = false;
3232 mpol_put(vma_policy(new_vma));
3234 kmem_cache_free(vm_area_cachep, new_vma);
3240 * Return true if the calling process may expand its vm space by the passed
3243 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3245 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3248 if (is_data_mapping(flags) &&
3249 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3250 /* Workaround for Valgrind */
3251 if (rlimit(RLIMIT_DATA) == 0 &&
3252 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3255 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3256 current->comm, current->pid,
3257 (mm->data_vm + npages) << PAGE_SHIFT,
3258 rlimit(RLIMIT_DATA),
3259 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3261 if (!ignore_rlimit_data)
3268 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3270 mm->total_vm += npages;
3272 if (is_exec_mapping(flags))
3273 mm->exec_vm += npages;
3274 else if (is_stack_mapping(flags))
3275 mm->stack_vm += npages;
3276 else if (is_data_mapping(flags))
3277 mm->data_vm += npages;
3280 static int special_mapping_fault(struct vm_fault *vmf);
3283 * Having a close hook prevents vma merging regardless of flags.
3285 static void special_mapping_close(struct vm_area_struct *vma)
3289 static const char *special_mapping_name(struct vm_area_struct *vma)
3291 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3294 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3296 struct vm_special_mapping *sm = new_vma->vm_private_data;
3298 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3302 return sm->mremap(sm, new_vma);
3307 static const struct vm_operations_struct special_mapping_vmops = {
3308 .close = special_mapping_close,
3309 .fault = special_mapping_fault,
3310 .mremap = special_mapping_mremap,
3311 .name = special_mapping_name,
3314 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3315 .close = special_mapping_close,
3316 .fault = special_mapping_fault,
3319 static int special_mapping_fault(struct vm_fault *vmf)
3321 struct vm_area_struct *vma = vmf->vma;
3323 struct page **pages;
3325 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3326 pages = vma->vm_private_data;
3328 struct vm_special_mapping *sm = vma->vm_private_data;
3331 return sm->fault(sm, vmf->vma, vmf);
3336 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3340 struct page *page = *pages;
3346 return VM_FAULT_SIGBUS;
3349 static struct vm_area_struct *__install_special_mapping(
3350 struct mm_struct *mm,
3351 unsigned long addr, unsigned long len,
3352 unsigned long vm_flags, void *priv,
3353 const struct vm_operations_struct *ops)
3356 struct vm_area_struct *vma;
3358 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3359 if (unlikely(vma == NULL))
3360 return ERR_PTR(-ENOMEM);
3362 INIT_LIST_HEAD(&vma->anon_vma_chain);
3364 vma->vm_start = addr;
3365 vma->vm_end = addr + len;
3367 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3368 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3371 vma->vm_private_data = priv;
3373 ret = insert_vm_struct(mm, vma);
3377 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3379 perf_event_mmap(vma);
3384 kmem_cache_free(vm_area_cachep, vma);
3385 return ERR_PTR(ret);
3388 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3389 const struct vm_special_mapping *sm)
3391 return vma->vm_private_data == sm &&
3392 (vma->vm_ops == &special_mapping_vmops ||
3393 vma->vm_ops == &legacy_special_mapping_vmops);
3397 * Called with mm->mmap_sem held for writing.
3398 * Insert a new vma covering the given region, with the given flags.
3399 * Its pages are supplied by the given array of struct page *.
3400 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3401 * The region past the last page supplied will always produce SIGBUS.
3402 * The array pointer and the pages it points to are assumed to stay alive
3403 * for as long as this mapping might exist.
3405 struct vm_area_struct *_install_special_mapping(
3406 struct mm_struct *mm,
3407 unsigned long addr, unsigned long len,
3408 unsigned long vm_flags, const struct vm_special_mapping *spec)
3410 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3411 &special_mapping_vmops);
3414 int install_special_mapping(struct mm_struct *mm,
3415 unsigned long addr, unsigned long len,
3416 unsigned long vm_flags, struct page **pages)
3418 struct vm_area_struct *vma = __install_special_mapping(
3419 mm, addr, len, vm_flags, (void *)pages,
3420 &legacy_special_mapping_vmops);
3422 return PTR_ERR_OR_ZERO(vma);
3425 static DEFINE_MUTEX(mm_all_locks_mutex);
3427 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3429 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3431 * The LSB of head.next can't change from under us
3432 * because we hold the mm_all_locks_mutex.
3434 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3436 * We can safely modify head.next after taking the
3437 * anon_vma->root->rwsem. If some other vma in this mm shares
3438 * the same anon_vma we won't take it again.
3440 * No need of atomic instructions here, head.next
3441 * can't change from under us thanks to the
3442 * anon_vma->root->rwsem.
3444 if (__test_and_set_bit(0, (unsigned long *)
3445 &anon_vma->root->rb_root.rb_root.rb_node))
3450 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3452 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3454 * AS_MM_ALL_LOCKS can't change from under us because
3455 * we hold the mm_all_locks_mutex.
3457 * Operations on ->flags have to be atomic because
3458 * even if AS_MM_ALL_LOCKS is stable thanks to the
3459 * mm_all_locks_mutex, there may be other cpus
3460 * changing other bitflags in parallel to us.
3462 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3464 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3469 * This operation locks against the VM for all pte/vma/mm related
3470 * operations that could ever happen on a certain mm. This includes
3471 * vmtruncate, try_to_unmap, and all page faults.
3473 * The caller must take the mmap_sem in write mode before calling
3474 * mm_take_all_locks(). The caller isn't allowed to release the
3475 * mmap_sem until mm_drop_all_locks() returns.
3477 * mmap_sem in write mode is required in order to block all operations
3478 * that could modify pagetables and free pages without need of
3479 * altering the vma layout. It's also needed in write mode to avoid new
3480 * anon_vmas to be associated with existing vmas.
3482 * A single task can't take more than one mm_take_all_locks() in a row
3483 * or it would deadlock.
3485 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3486 * mapping->flags avoid to take the same lock twice, if more than one
3487 * vma in this mm is backed by the same anon_vma or address_space.
3489 * We take locks in following order, accordingly to comment at beginning
3491 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3493 * - all i_mmap_rwsem locks;
3494 * - all anon_vma->rwseml
3496 * We can take all locks within these types randomly because the VM code
3497 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3498 * mm_all_locks_mutex.
3500 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3501 * that may have to take thousand of locks.
3503 * mm_take_all_locks() can fail if it's interrupted by signals.
3505 int mm_take_all_locks(struct mm_struct *mm)
3507 struct vm_area_struct *vma;
3508 struct anon_vma_chain *avc;
3510 BUG_ON(down_read_trylock(&mm->mmap_sem));
3512 mutex_lock(&mm_all_locks_mutex);
3514 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3515 if (signal_pending(current))
3517 if (vma->vm_file && vma->vm_file->f_mapping &&
3518 is_vm_hugetlb_page(vma))
3519 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3522 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3523 if (signal_pending(current))
3525 if (vma->vm_file && vma->vm_file->f_mapping &&
3526 !is_vm_hugetlb_page(vma))
3527 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3530 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3531 if (signal_pending(current))
3534 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3535 vm_lock_anon_vma(mm, avc->anon_vma);
3541 mm_drop_all_locks(mm);
3545 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3547 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3549 * The LSB of head.next can't change to 0 from under
3550 * us because we hold the mm_all_locks_mutex.
3552 * We must however clear the bitflag before unlocking
3553 * the vma so the users using the anon_vma->rb_root will
3554 * never see our bitflag.
3556 * No need of atomic instructions here, head.next
3557 * can't change from under us until we release the
3558 * anon_vma->root->rwsem.
3560 if (!__test_and_clear_bit(0, (unsigned long *)
3561 &anon_vma->root->rb_root.rb_root.rb_node))
3563 anon_vma_unlock_write(anon_vma);
3567 static void vm_unlock_mapping(struct address_space *mapping)
3569 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3571 * AS_MM_ALL_LOCKS can't change to 0 from under us
3572 * because we hold the mm_all_locks_mutex.
3574 i_mmap_unlock_write(mapping);
3575 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3582 * The mmap_sem cannot be released by the caller until
3583 * mm_drop_all_locks() returns.
3585 void mm_drop_all_locks(struct mm_struct *mm)
3587 struct vm_area_struct *vma;
3588 struct anon_vma_chain *avc;
3590 BUG_ON(down_read_trylock(&mm->mmap_sem));
3591 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3593 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3595 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3596 vm_unlock_anon_vma(avc->anon_vma);
3597 if (vma->vm_file && vma->vm_file->f_mapping)
3598 vm_unlock_mapping(vma->vm_file->f_mapping);
3601 mutex_unlock(&mm_all_locks_mutex);
3605 * initialise the percpu counter for VM
3607 void __init mmap_init(void)
3611 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3616 * Initialise sysctl_user_reserve_kbytes.
3618 * This is intended to prevent a user from starting a single memory hogging
3619 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3622 * The default value is min(3% of free memory, 128MB)
3623 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3625 static int init_user_reserve(void)
3627 unsigned long free_kbytes;
3629 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3631 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3634 subsys_initcall(init_user_reserve);
3637 * Initialise sysctl_admin_reserve_kbytes.
3639 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3640 * to log in and kill a memory hogging process.
3642 * Systems with more than 256MB will reserve 8MB, enough to recover
3643 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3644 * only reserve 3% of free pages by default.
3646 static int init_admin_reserve(void)
3648 unsigned long free_kbytes;
3650 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3652 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3655 subsys_initcall(init_admin_reserve);
3658 * Reinititalise user and admin reserves if memory is added or removed.
3660 * The default user reserve max is 128MB, and the default max for the
3661 * admin reserve is 8MB. These are usually, but not always, enough to
3662 * enable recovery from a memory hogging process using login/sshd, a shell,
3663 * and tools like top. It may make sense to increase or even disable the
3664 * reserve depending on the existence of swap or variations in the recovery
3665 * tools. So, the admin may have changed them.
3667 * If memory is added and the reserves have been eliminated or increased above
3668 * the default max, then we'll trust the admin.
3670 * If memory is removed and there isn't enough free memory, then we
3671 * need to reset the reserves.
3673 * Otherwise keep the reserve set by the admin.
3675 static int reserve_mem_notifier(struct notifier_block *nb,
3676 unsigned long action, void *data)
3678 unsigned long tmp, free_kbytes;
3682 /* Default max is 128MB. Leave alone if modified by operator. */
3683 tmp = sysctl_user_reserve_kbytes;
3684 if (0 < tmp && tmp < (1UL << 17))
3685 init_user_reserve();
3687 /* Default max is 8MB. Leave alone if modified by operator. */
3688 tmp = sysctl_admin_reserve_kbytes;
3689 if (0 < tmp && tmp < (1UL << 13))
3690 init_admin_reserve();
3694 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3696 if (sysctl_user_reserve_kbytes > free_kbytes) {
3697 init_user_reserve();
3698 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3699 sysctl_user_reserve_kbytes);
3702 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3703 init_admin_reserve();
3704 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3705 sysctl_admin_reserve_kbytes);
3714 static struct notifier_block reserve_mem_nb = {
3715 .notifier_call = reserve_mem_notifier,
3718 static int __meminit init_reserve_notifier(void)
3720 if (register_hotmemory_notifier(&reserve_mem_nb))
3721 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3725 subsys_initcall(init_reserve_notifier);