Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/mmap.c
4 *
5 * Written by obz.
6 *
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
8 */
9
10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12#include <linux/kernel.h>
13#include <linux/slab.h>
14#include <linux/backing-dev.h>
15#include <linux/mm.h>
16#include <linux/mm_inline.h>
17#include <linux/shm.h>
18#include <linux/mman.h>
19#include <linux/pagemap.h>
20#include <linux/swap.h>
21#include <linux/syscalls.h>
22#include <linux/capability.h>
23#include <linux/init.h>
24#include <linux/file.h>
25#include <linux/fs.h>
26#include <linux/personality.h>
27#include <linux/security.h>
28#include <linux/hugetlb.h>
29#include <linux/shmem_fs.h>
30#include <linux/profile.h>
31#include <linux/export.h>
32#include <linux/mount.h>
33#include <linux/mempolicy.h>
34#include <linux/rmap.h>
35#include <linux/mmu_notifier.h>
36#include <linux/mmdebug.h>
37#include <linux/perf_event.h>
38#include <linux/audit.h>
39#include <linux/khugepaged.h>
40#include <linux/uprobes.h>
41#include <linux/notifier.h>
42#include <linux/memory.h>
43#include <linux/printk.h>
44#include <linux/userfaultfd_k.h>
45#include <linux/moduleparam.h>
46#include <linux/pkeys.h>
47#include <linux/oom.h>
48#include <linux/sched/mm.h>
49#include <linux/ksm.h>
50
51#include <linux/uaccess.h>
52#include <asm/cacheflush.h>
53#include <asm/tlb.h>
54#include <asm/mmu_context.h>
55
56#define CREATE_TRACE_POINTS
57#include <trace/events/mmap.h>
58
59#include "internal.h"
60
61#ifndef arch_mmap_check
62#define arch_mmap_check(addr, len, flags) (0)
63#endif
64
65#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
66const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
67const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
68int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
69#endif
70#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
71const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
72const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
73int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
74#endif
75
76static bool ignore_rlimit_data;
77core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
78
79static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
80 struct vm_area_struct *vma, struct vm_area_struct *prev,
81 struct vm_area_struct *next, unsigned long start,
82 unsigned long end, unsigned long tree_end, bool mm_wr_locked);
83
84static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
85{
86 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
87}
88
89/* Update vma->vm_page_prot to reflect vma->vm_flags. */
90void vma_set_page_prot(struct vm_area_struct *vma)
91{
92 unsigned long vm_flags = vma->vm_flags;
93 pgprot_t vm_page_prot;
94
95 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
96 if (vma_wants_writenotify(vma, vm_page_prot)) {
97 vm_flags &= ~VM_SHARED;
98 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
99 }
100 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
101 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
102}
103
104/*
105 * Requires inode->i_mapping->i_mmap_rwsem
106 */
107static void __remove_shared_vm_struct(struct vm_area_struct *vma,
108 struct file *file, struct address_space *mapping)
109{
110 if (vma_is_shared_maywrite(vma))
111 mapping_unmap_writable(mapping);
112
113 flush_dcache_mmap_lock(mapping);
114 vma_interval_tree_remove(vma, &mapping->i_mmap);
115 flush_dcache_mmap_unlock(mapping);
116}
117
118/*
119 * Unlink a file-based vm structure from its interval tree, to hide
120 * vma from rmap and vmtruncate before freeing its page tables.
121 */
122void unlink_file_vma(struct vm_area_struct *vma)
123{
124 struct file *file = vma->vm_file;
125
126 if (file) {
127 struct address_space *mapping = file->f_mapping;
128 i_mmap_lock_write(mapping);
129 __remove_shared_vm_struct(vma, file, mapping);
130 i_mmap_unlock_write(mapping);
131 }
132}
133
134/*
135 * Close a vm structure and free it.
136 */
137static void remove_vma(struct vm_area_struct *vma, bool unreachable)
138{
139 might_sleep();
140 if (vma->vm_ops && vma->vm_ops->close)
141 vma->vm_ops->close(vma);
142 if (vma->vm_file)
143 fput(vma->vm_file);
144 mpol_put(vma_policy(vma));
145 if (unreachable)
146 __vm_area_free(vma);
147 else
148 vm_area_free(vma);
149}
150
151static inline struct vm_area_struct *vma_prev_limit(struct vma_iterator *vmi,
152 unsigned long min)
153{
154 return mas_prev(&vmi->mas, min);
155}
156
157/*
158 * check_brk_limits() - Use platform specific check of range & verify mlock
159 * limits.
160 * @addr: The address to check
161 * @len: The size of increase.
162 *
163 * Return: 0 on success.
164 */
165static int check_brk_limits(unsigned long addr, unsigned long len)
166{
167 unsigned long mapped_addr;
168
169 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
170 if (IS_ERR_VALUE(mapped_addr))
171 return mapped_addr;
172
173 return mlock_future_ok(current->mm, current->mm->def_flags, len)
174 ? 0 : -EAGAIN;
175}
176static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma,
177 unsigned long addr, unsigned long request, unsigned long flags);
178SYSCALL_DEFINE1(brk, unsigned long, brk)
179{
180 unsigned long newbrk, oldbrk, origbrk;
181 struct mm_struct *mm = current->mm;
182 struct vm_area_struct *brkvma, *next = NULL;
183 unsigned long min_brk;
184 bool populate = false;
185 LIST_HEAD(uf);
186 struct vma_iterator vmi;
187
188 if (mmap_write_lock_killable(mm))
189 return -EINTR;
190
191 origbrk = mm->brk;
192
193#ifdef CONFIG_COMPAT_BRK
194 /*
195 * CONFIG_COMPAT_BRK can still be overridden by setting
196 * randomize_va_space to 2, which will still cause mm->start_brk
197 * to be arbitrarily shifted
198 */
199 if (current->brk_randomized)
200 min_brk = mm->start_brk;
201 else
202 min_brk = mm->end_data;
203#else
204 min_brk = mm->start_brk;
205#endif
206 if (brk < min_brk)
207 goto out;
208
209 /*
210 * Check against rlimit here. If this check is done later after the test
211 * of oldbrk with newbrk then it can escape the test and let the data
212 * segment grow beyond its set limit the in case where the limit is
213 * not page aligned -Ram Gupta
214 */
215 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
216 mm->end_data, mm->start_data))
217 goto out;
218
219 newbrk = PAGE_ALIGN(brk);
220 oldbrk = PAGE_ALIGN(mm->brk);
221 if (oldbrk == newbrk) {
222 mm->brk = brk;
223 goto success;
224 }
225
226 /* Always allow shrinking brk. */
227 if (brk <= mm->brk) {
228 /* Search one past newbrk */
229 vma_iter_init(&vmi, mm, newbrk);
230 brkvma = vma_find(&vmi, oldbrk);
231 if (!brkvma || brkvma->vm_start >= oldbrk)
232 goto out; /* mapping intersects with an existing non-brk vma. */
233 /*
234 * mm->brk must be protected by write mmap_lock.
235 * do_vma_munmap() will drop the lock on success, so update it
236 * before calling do_vma_munmap().
237 */
238 mm->brk = brk;
239 if (do_vma_munmap(&vmi, brkvma, newbrk, oldbrk, &uf, true))
240 goto out;
241
242 goto success_unlocked;
243 }
244
245 if (check_brk_limits(oldbrk, newbrk - oldbrk))
246 goto out;
247
248 /*
249 * Only check if the next VMA is within the stack_guard_gap of the
250 * expansion area
251 */
252 vma_iter_init(&vmi, mm, oldbrk);
253 next = vma_find(&vmi, newbrk + PAGE_SIZE + stack_guard_gap);
254 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
255 goto out;
256
257 brkvma = vma_prev_limit(&vmi, mm->start_brk);
258 /* Ok, looks good - let it rip. */
259 if (do_brk_flags(&vmi, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
260 goto out;
261
262 mm->brk = brk;
263 if (mm->def_flags & VM_LOCKED)
264 populate = true;
265
266success:
267 mmap_write_unlock(mm);
268success_unlocked:
269 userfaultfd_unmap_complete(mm, &uf);
270 if (populate)
271 mm_populate(oldbrk, newbrk - oldbrk);
272 return brk;
273
274out:
275 mm->brk = origbrk;
276 mmap_write_unlock(mm);
277 return origbrk;
278}
279
280#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
281static void validate_mm(struct mm_struct *mm)
282{
283 int bug = 0;
284 int i = 0;
285 struct vm_area_struct *vma;
286 VMA_ITERATOR(vmi, mm, 0);
287
288 mt_validate(&mm->mm_mt);
289 for_each_vma(vmi, vma) {
290#ifdef CONFIG_DEBUG_VM_RB
291 struct anon_vma *anon_vma = vma->anon_vma;
292 struct anon_vma_chain *avc;
293#endif
294 unsigned long vmi_start, vmi_end;
295 bool warn = 0;
296
297 vmi_start = vma_iter_addr(&vmi);
298 vmi_end = vma_iter_end(&vmi);
299 if (VM_WARN_ON_ONCE_MM(vma->vm_end != vmi_end, mm))
300 warn = 1;
301
302 if (VM_WARN_ON_ONCE_MM(vma->vm_start != vmi_start, mm))
303 warn = 1;
304
305 if (warn) {
306 pr_emerg("issue in %s\n", current->comm);
307 dump_stack();
308 dump_vma(vma);
309 pr_emerg("tree range: %px start %lx end %lx\n", vma,
310 vmi_start, vmi_end - 1);
311 vma_iter_dump_tree(&vmi);
312 }
313
314#ifdef CONFIG_DEBUG_VM_RB
315 if (anon_vma) {
316 anon_vma_lock_read(anon_vma);
317 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
318 anon_vma_interval_tree_verify(avc);
319 anon_vma_unlock_read(anon_vma);
320 }
321#endif
322 i++;
323 }
324 if (i != mm->map_count) {
325 pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);
326 bug = 1;
327 }
328 VM_BUG_ON_MM(bug, mm);
329}
330
331#else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
332#define validate_mm(mm) do { } while (0)
333#endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
334
335/*
336 * vma has some anon_vma assigned, and is already inserted on that
337 * anon_vma's interval trees.
338 *
339 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
340 * vma must be removed from the anon_vma's interval trees using
341 * anon_vma_interval_tree_pre_update_vma().
342 *
343 * After the update, the vma will be reinserted using
344 * anon_vma_interval_tree_post_update_vma().
345 *
346 * The entire update must be protected by exclusive mmap_lock and by
347 * the root anon_vma's mutex.
348 */
349static inline void
350anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
351{
352 struct anon_vma_chain *avc;
353
354 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
355 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
356}
357
358static inline void
359anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
360{
361 struct anon_vma_chain *avc;
362
363 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
364 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
365}
366
367static unsigned long count_vma_pages_range(struct mm_struct *mm,
368 unsigned long addr, unsigned long end)
369{
370 VMA_ITERATOR(vmi, mm, addr);
371 struct vm_area_struct *vma;
372 unsigned long nr_pages = 0;
373
374 for_each_vma_range(vmi, vma, end) {
375 unsigned long vm_start = max(addr, vma->vm_start);
376 unsigned long vm_end = min(end, vma->vm_end);
377
378 nr_pages += PHYS_PFN(vm_end - vm_start);
379 }
380
381 return nr_pages;
382}
383
384static void __vma_link_file(struct vm_area_struct *vma,
385 struct address_space *mapping)
386{
387 if (vma_is_shared_maywrite(vma))
388 mapping_allow_writable(mapping);
389
390 flush_dcache_mmap_lock(mapping);
391 vma_interval_tree_insert(vma, &mapping->i_mmap);
392 flush_dcache_mmap_unlock(mapping);
393}
394
395static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
396{
397 VMA_ITERATOR(vmi, mm, 0);
398 struct address_space *mapping = NULL;
399
400 vma_iter_config(&vmi, vma->vm_start, vma->vm_end);
401 if (vma_iter_prealloc(&vmi, vma))
402 return -ENOMEM;
403
404 vma_start_write(vma);
405
406 vma_iter_store(&vmi, vma);
407
408 if (vma->vm_file) {
409 mapping = vma->vm_file->f_mapping;
410 i_mmap_lock_write(mapping);
411 __vma_link_file(vma, mapping);
412 i_mmap_unlock_write(mapping);
413 }
414
415 mm->map_count++;
416 validate_mm(mm);
417 return 0;
418}
419
420/*
421 * init_multi_vma_prep() - Initializer for struct vma_prepare
422 * @vp: The vma_prepare struct
423 * @vma: The vma that will be altered once locked
424 * @next: The next vma if it is to be adjusted
425 * @remove: The first vma to be removed
426 * @remove2: The second vma to be removed
427 */
428static inline void init_multi_vma_prep(struct vma_prepare *vp,
429 struct vm_area_struct *vma, struct vm_area_struct *next,
430 struct vm_area_struct *remove, struct vm_area_struct *remove2)
431{
432 memset(vp, 0, sizeof(struct vma_prepare));
433 vp->vma = vma;
434 vp->anon_vma = vma->anon_vma;
435 vp->remove = remove;
436 vp->remove2 = remove2;
437 vp->adj_next = next;
438 if (!vp->anon_vma && next)
439 vp->anon_vma = next->anon_vma;
440
441 vp->file = vma->vm_file;
442 if (vp->file)
443 vp->mapping = vma->vm_file->f_mapping;
444
445}
446
447/*
448 * init_vma_prep() - Initializer wrapper for vma_prepare struct
449 * @vp: The vma_prepare struct
450 * @vma: The vma that will be altered once locked
451 */
452static inline void init_vma_prep(struct vma_prepare *vp,
453 struct vm_area_struct *vma)
454{
455 init_multi_vma_prep(vp, vma, NULL, NULL, NULL);
456}
457
458
459/*
460 * vma_prepare() - Helper function for handling locking VMAs prior to altering
461 * @vp: The initialized vma_prepare struct
462 */
463static inline void vma_prepare(struct vma_prepare *vp)
464{
465 if (vp->file) {
466 uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end);
467
468 if (vp->adj_next)
469 uprobe_munmap(vp->adj_next, vp->adj_next->vm_start,
470 vp->adj_next->vm_end);
471
472 i_mmap_lock_write(vp->mapping);
473 if (vp->insert && vp->insert->vm_file) {
474 /*
475 * Put into interval tree now, so instantiated pages
476 * are visible to arm/parisc __flush_dcache_page
477 * throughout; but we cannot insert into address
478 * space until vma start or end is updated.
479 */
480 __vma_link_file(vp->insert,
481 vp->insert->vm_file->f_mapping);
482 }
483 }
484
485 if (vp->anon_vma) {
486 anon_vma_lock_write(vp->anon_vma);
487 anon_vma_interval_tree_pre_update_vma(vp->vma);
488 if (vp->adj_next)
489 anon_vma_interval_tree_pre_update_vma(vp->adj_next);
490 }
491
492 if (vp->file) {
493 flush_dcache_mmap_lock(vp->mapping);
494 vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap);
495 if (vp->adj_next)
496 vma_interval_tree_remove(vp->adj_next,
497 &vp->mapping->i_mmap);
498 }
499
500}
501
502/*
503 * vma_complete- Helper function for handling the unlocking after altering VMAs,
504 * or for inserting a VMA.
505 *
506 * @vp: The vma_prepare struct
507 * @vmi: The vma iterator
508 * @mm: The mm_struct
509 */
510static inline void vma_complete(struct vma_prepare *vp,
511 struct vma_iterator *vmi, struct mm_struct *mm)
512{
513 if (vp->file) {
514 if (vp->adj_next)
515 vma_interval_tree_insert(vp->adj_next,
516 &vp->mapping->i_mmap);
517 vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap);
518 flush_dcache_mmap_unlock(vp->mapping);
519 }
520
521 if (vp->remove && vp->file) {
522 __remove_shared_vm_struct(vp->remove, vp->file, vp->mapping);
523 if (vp->remove2)
524 __remove_shared_vm_struct(vp->remove2, vp->file,
525 vp->mapping);
526 } else if (vp->insert) {
527 /*
528 * split_vma has split insert from vma, and needs
529 * us to insert it before dropping the locks
530 * (it may either follow vma or precede it).
531 */
532 vma_iter_store(vmi, vp->insert);
533 mm->map_count++;
534 }
535
536 if (vp->anon_vma) {
537 anon_vma_interval_tree_post_update_vma(vp->vma);
538 if (vp->adj_next)
539 anon_vma_interval_tree_post_update_vma(vp->adj_next);
540 anon_vma_unlock_write(vp->anon_vma);
541 }
542
543 if (vp->file) {
544 i_mmap_unlock_write(vp->mapping);
545 uprobe_mmap(vp->vma);
546
547 if (vp->adj_next)
548 uprobe_mmap(vp->adj_next);
549 }
550
551 if (vp->remove) {
552again:
553 vma_mark_detached(vp->remove, true);
554 if (vp->file) {
555 uprobe_munmap(vp->remove, vp->remove->vm_start,
556 vp->remove->vm_end);
557 fput(vp->file);
558 }
559 if (vp->remove->anon_vma)
560 anon_vma_merge(vp->vma, vp->remove);
561 mm->map_count--;
562 mpol_put(vma_policy(vp->remove));
563 if (!vp->remove2)
564 WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end);
565 vm_area_free(vp->remove);
566
567 /*
568 * In mprotect's case 6 (see comments on vma_merge),
569 * we are removing both mid and next vmas
570 */
571 if (vp->remove2) {
572 vp->remove = vp->remove2;
573 vp->remove2 = NULL;
574 goto again;
575 }
576 }
577 if (vp->insert && vp->file)
578 uprobe_mmap(vp->insert);
579 validate_mm(mm);
580}
581
582/*
583 * dup_anon_vma() - Helper function to duplicate anon_vma
584 * @dst: The destination VMA
585 * @src: The source VMA
586 * @dup: Pointer to the destination VMA when successful.
587 *
588 * Returns: 0 on success.
589 */
590static inline int dup_anon_vma(struct vm_area_struct *dst,
591 struct vm_area_struct *src, struct vm_area_struct **dup)
592{
593 /*
594 * Easily overlooked: when mprotect shifts the boundary, make sure the
595 * expanding vma has anon_vma set if the shrinking vma had, to cover any
596 * anon pages imported.
597 */
598 if (src->anon_vma && !dst->anon_vma) {
599 int ret;
600
601 vma_assert_write_locked(dst);
602 dst->anon_vma = src->anon_vma;
603 ret = anon_vma_clone(dst, src);
604 if (ret)
605 return ret;
606
607 *dup = dst;
608 }
609
610 return 0;
611}
612
613/*
614 * vma_expand - Expand an existing VMA
615 *
616 * @vmi: The vma iterator
617 * @vma: The vma to expand
618 * @start: The start of the vma
619 * @end: The exclusive end of the vma
620 * @pgoff: The page offset of vma
621 * @next: The current of next vma.
622 *
623 * Expand @vma to @start and @end. Can expand off the start and end. Will
624 * expand over @next if it's different from @vma and @end == @next->vm_end.
625 * Checking if the @vma can expand and merge with @next needs to be handled by
626 * the caller.
627 *
628 * Returns: 0 on success
629 */
630int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
631 unsigned long start, unsigned long end, pgoff_t pgoff,
632 struct vm_area_struct *next)
633{
634 struct vm_area_struct *anon_dup = NULL;
635 bool remove_next = false;
636 struct vma_prepare vp;
637
638 vma_start_write(vma);
639 if (next && (vma != next) && (end == next->vm_end)) {
640 int ret;
641
642 remove_next = true;
643 vma_start_write(next);
644 ret = dup_anon_vma(vma, next, &anon_dup);
645 if (ret)
646 return ret;
647 }
648
649 init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL);
650 /* Not merging but overwriting any part of next is not handled. */
651 VM_WARN_ON(next && !vp.remove &&
652 next != vma && end > next->vm_start);
653 /* Only handles expanding */
654 VM_WARN_ON(vma->vm_start < start || vma->vm_end > end);
655
656 /* Note: vma iterator must be pointing to 'start' */
657 vma_iter_config(vmi, start, end);
658 if (vma_iter_prealloc(vmi, vma))
659 goto nomem;
660
661 vma_prepare(&vp);
662 vma_adjust_trans_huge(vma, start, end, 0);
663 vma->vm_start = start;
664 vma->vm_end = end;
665 vma->vm_pgoff = pgoff;
666 vma_iter_store(vmi, vma);
667
668 vma_complete(&vp, vmi, vma->vm_mm);
669 return 0;
670
671nomem:
672 if (anon_dup)
673 unlink_anon_vmas(anon_dup);
674 return -ENOMEM;
675}
676
677/*
678 * vma_shrink() - Reduce an existing VMAs memory area
679 * @vmi: The vma iterator
680 * @vma: The VMA to modify
681 * @start: The new start
682 * @end: The new end
683 *
684 * Returns: 0 on success, -ENOMEM otherwise
685 */
686int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
687 unsigned long start, unsigned long end, pgoff_t pgoff)
688{
689 struct vma_prepare vp;
690
691 WARN_ON((vma->vm_start != start) && (vma->vm_end != end));
692
693 if (vma->vm_start < start)
694 vma_iter_config(vmi, vma->vm_start, start);
695 else
696 vma_iter_config(vmi, end, vma->vm_end);
697
698 if (vma_iter_prealloc(vmi, NULL))
699 return -ENOMEM;
700
701 vma_start_write(vma);
702
703 init_vma_prep(&vp, vma);
704 vma_prepare(&vp);
705 vma_adjust_trans_huge(vma, start, end, 0);
706
707 vma_iter_clear(vmi);
708 vma->vm_start = start;
709 vma->vm_end = end;
710 vma->vm_pgoff = pgoff;
711 vma_complete(&vp, vmi, vma->vm_mm);
712 return 0;
713}
714
715/*
716 * If the vma has a ->close operation then the driver probably needs to release
717 * per-vma resources, so we don't attempt to merge those if the caller indicates
718 * the current vma may be removed as part of the merge.
719 */
720static inline bool is_mergeable_vma(struct vm_area_struct *vma,
721 struct file *file, unsigned long vm_flags,
722 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
723 struct anon_vma_name *anon_name, bool may_remove_vma)
724{
725 /*
726 * VM_SOFTDIRTY should not prevent from VMA merging, if we
727 * match the flags but dirty bit -- the caller should mark
728 * merged VMA as dirty. If dirty bit won't be excluded from
729 * comparison, we increase pressure on the memory system forcing
730 * the kernel to generate new VMAs when old one could be
731 * extended instead.
732 */
733 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
734 return false;
735 if (vma->vm_file != file)
736 return false;
737 if (may_remove_vma && vma->vm_ops && vma->vm_ops->close)
738 return false;
739 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
740 return false;
741 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
742 return false;
743 return true;
744}
745
746static inline bool is_mergeable_anon_vma(struct anon_vma *anon_vma1,
747 struct anon_vma *anon_vma2, struct vm_area_struct *vma)
748{
749 /*
750 * The list_is_singular() test is to avoid merging VMA cloned from
751 * parents. This can improve scalability caused by anon_vma lock.
752 */
753 if ((!anon_vma1 || !anon_vma2) && (!vma ||
754 list_is_singular(&vma->anon_vma_chain)))
755 return true;
756 return anon_vma1 == anon_vma2;
757}
758
759/*
760 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
761 * in front of (at a lower virtual address and file offset than) the vma.
762 *
763 * We cannot merge two vmas if they have differently assigned (non-NULL)
764 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
765 *
766 * We don't check here for the merged mmap wrapping around the end of pagecache
767 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
768 * wrap, nor mmaps which cover the final page at index -1UL.
769 *
770 * We assume the vma may be removed as part of the merge.
771 */
772static bool
773can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
774 struct anon_vma *anon_vma, struct file *file,
775 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
776 struct anon_vma_name *anon_name)
777{
778 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, true) &&
779 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
780 if (vma->vm_pgoff == vm_pgoff)
781 return true;
782 }
783 return false;
784}
785
786/*
787 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
788 * beyond (at a higher virtual address and file offset than) the vma.
789 *
790 * We cannot merge two vmas if they have differently assigned (non-NULL)
791 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
792 *
793 * We assume that vma is not removed as part of the merge.
794 */
795static bool
796can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
797 struct anon_vma *anon_vma, struct file *file,
798 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
799 struct anon_vma_name *anon_name)
800{
801 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, false) &&
802 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
803 pgoff_t vm_pglen;
804 vm_pglen = vma_pages(vma);
805 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
806 return true;
807 }
808 return false;
809}
810
811/*
812 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
813 * figure out whether that can be merged with its predecessor or its
814 * successor. Or both (it neatly fills a hole).
815 *
816 * In most cases - when called for mmap, brk or mremap - [addr,end) is
817 * certain not to be mapped by the time vma_merge is called; but when
818 * called for mprotect, it is certain to be already mapped (either at
819 * an offset within prev, or at the start of next), and the flags of
820 * this area are about to be changed to vm_flags - and the no-change
821 * case has already been eliminated.
822 *
823 * The following mprotect cases have to be considered, where **** is
824 * the area passed down from mprotect_fixup, never extending beyond one
825 * vma, PPPP is the previous vma, CCCC is a concurrent vma that starts
826 * at the same address as **** and is of the same or larger span, and
827 * NNNN the next vma after ****:
828 *
829 * **** **** ****
830 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPCCCCCC
831 * cannot merge might become might become
832 * PPNNNNNNNNNN PPPPPPPPPPCC
833 * mmap, brk or case 4 below case 5 below
834 * mremap move:
835 * **** ****
836 * PPPP NNNN PPPPCCCCNNNN
837 * might become might become
838 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
839 * PPPPPPPPNNNN 2 or PPPPPPPPNNNN 7 or
840 * PPPPNNNNNNNN 3 PPPPNNNNNNNN 8
841 *
842 * It is important for case 8 that the vma CCCC overlapping the
843 * region **** is never going to extended over NNNN. Instead NNNN must
844 * be extended in region **** and CCCC must be removed. This way in
845 * all cases where vma_merge succeeds, the moment vma_merge drops the
846 * rmap_locks, the properties of the merged vma will be already
847 * correct for the whole merged range. Some of those properties like
848 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
849 * be correct for the whole merged range immediately after the
850 * rmap_locks are released. Otherwise if NNNN would be removed and
851 * CCCC would be extended over the NNNN range, remove_migration_ptes
852 * or other rmap walkers (if working on addresses beyond the "end"
853 * parameter) may establish ptes with the wrong permissions of CCCC
854 * instead of the right permissions of NNNN.
855 *
856 * In the code below:
857 * PPPP is represented by *prev
858 * CCCC is represented by *curr or not represented at all (NULL)
859 * NNNN is represented by *next or not represented at all (NULL)
860 * **** is not represented - it will be merged and the vma containing the
861 * area is returned, or the function will return NULL
862 */
863static struct vm_area_struct
864*vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
865 struct vm_area_struct *prev, unsigned long addr, unsigned long end,
866 unsigned long vm_flags, struct anon_vma *anon_vma, struct file *file,
867 pgoff_t pgoff, struct mempolicy *policy,
868 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
869 struct anon_vma_name *anon_name)
870{
871 struct vm_area_struct *curr, *next, *res;
872 struct vm_area_struct *vma, *adjust, *remove, *remove2;
873 struct vm_area_struct *anon_dup = NULL;
874 struct vma_prepare vp;
875 pgoff_t vma_pgoff;
876 int err = 0;
877 bool merge_prev = false;
878 bool merge_next = false;
879 bool vma_expanded = false;
880 unsigned long vma_start = addr;
881 unsigned long vma_end = end;
882 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
883 long adj_start = 0;
884
885 /*
886 * We later require that vma->vm_flags == vm_flags,
887 * so this tests vma->vm_flags & VM_SPECIAL, too.
888 */
889 if (vm_flags & VM_SPECIAL)
890 return NULL;
891
892 /* Does the input range span an existing VMA? (cases 5 - 8) */
893 curr = find_vma_intersection(mm, prev ? prev->vm_end : 0, end);
894
895 if (!curr || /* cases 1 - 4 */
896 end == curr->vm_end) /* cases 6 - 8, adjacent VMA */
897 next = vma_lookup(mm, end);
898 else
899 next = NULL; /* case 5 */
900
901 if (prev) {
902 vma_start = prev->vm_start;
903 vma_pgoff = prev->vm_pgoff;
904
905 /* Can we merge the predecessor? */
906 if (addr == prev->vm_end && mpol_equal(vma_policy(prev), policy)
907 && can_vma_merge_after(prev, vm_flags, anon_vma, file,
908 pgoff, vm_userfaultfd_ctx, anon_name)) {
909 merge_prev = true;
910 vma_prev(vmi);
911 }
912 }
913
914 /* Can we merge the successor? */
915 if (next && mpol_equal(policy, vma_policy(next)) &&
916 can_vma_merge_before(next, vm_flags, anon_vma, file, pgoff+pglen,
917 vm_userfaultfd_ctx, anon_name)) {
918 merge_next = true;
919 }
920
921 /* Verify some invariant that must be enforced by the caller. */
922 VM_WARN_ON(prev && addr <= prev->vm_start);
923 VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
924 VM_WARN_ON(addr >= end);
925
926 if (!merge_prev && !merge_next)
927 return NULL; /* Not mergeable. */
928
929 if (merge_prev)
930 vma_start_write(prev);
931
932 res = vma = prev;
933 remove = remove2 = adjust = NULL;
934
935 /* Can we merge both the predecessor and the successor? */
936 if (merge_prev && merge_next &&
937 is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) {
938 vma_start_write(next);
939 remove = next; /* case 1 */
940 vma_end = next->vm_end;
941 err = dup_anon_vma(prev, next, &anon_dup);
942 if (curr) { /* case 6 */
943 vma_start_write(curr);
944 remove = curr;
945 remove2 = next;
946 /*
947 * Note that the dup_anon_vma below cannot overwrite err
948 * since the first caller would do nothing unless next
949 * has an anon_vma.
950 */
951 if (!next->anon_vma)
952 err = dup_anon_vma(prev, curr, &anon_dup);
953 }
954 } else if (merge_prev) { /* case 2 */
955 if (curr) {
956 vma_start_write(curr);
957 if (end == curr->vm_end) { /* case 7 */
958 /*
959 * can_vma_merge_after() assumed we would not be
960 * removing prev vma, so it skipped the check
961 * for vm_ops->close, but we are removing curr
962 */
963 if (curr->vm_ops && curr->vm_ops->close)
964 err = -EINVAL;
965 remove = curr;
966 } else { /* case 5 */
967 adjust = curr;
968 adj_start = (end - curr->vm_start);
969 }
970 if (!err)
971 err = dup_anon_vma(prev, curr, &anon_dup);
972 }
973 } else { /* merge_next */
974 vma_start_write(next);
975 res = next;
976 if (prev && addr < prev->vm_end) { /* case 4 */
977 vma_start_write(prev);
978 vma_end = addr;
979 adjust = next;
980 adj_start = -(prev->vm_end - addr);
981 err = dup_anon_vma(next, prev, &anon_dup);
982 } else {
983 /*
984 * Note that cases 3 and 8 are the ONLY ones where prev
985 * is permitted to be (but is not necessarily) NULL.
986 */
987 vma = next; /* case 3 */
988 vma_start = addr;
989 vma_end = next->vm_end;
990 vma_pgoff = next->vm_pgoff - pglen;
991 if (curr) { /* case 8 */
992 vma_pgoff = curr->vm_pgoff;
993 vma_start_write(curr);
994 remove = curr;
995 err = dup_anon_vma(next, curr, &anon_dup);
996 }
997 }
998 }
999
1000 /* Error in anon_vma clone. */
1001 if (err)
1002 goto anon_vma_fail;
1003
1004 if (vma_start < vma->vm_start || vma_end > vma->vm_end)
1005 vma_expanded = true;
1006
1007 if (vma_expanded) {
1008 vma_iter_config(vmi, vma_start, vma_end);
1009 } else {
1010 vma_iter_config(vmi, adjust->vm_start + adj_start,
1011 adjust->vm_end);
1012 }
1013
1014 if (vma_iter_prealloc(vmi, vma))
1015 goto prealloc_fail;
1016
1017 init_multi_vma_prep(&vp, vma, adjust, remove, remove2);
1018 VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
1019 vp.anon_vma != adjust->anon_vma);
1020
1021 vma_prepare(&vp);
1022 vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start);
1023
1024 vma->vm_start = vma_start;
1025 vma->vm_end = vma_end;
1026 vma->vm_pgoff = vma_pgoff;
1027
1028 if (vma_expanded)
1029 vma_iter_store(vmi, vma);
1030
1031 if (adj_start) {
1032 adjust->vm_start += adj_start;
1033 adjust->vm_pgoff += adj_start >> PAGE_SHIFT;
1034 if (adj_start < 0) {
1035 WARN_ON(vma_expanded);
1036 vma_iter_store(vmi, next);
1037 }
1038 }
1039
1040 vma_complete(&vp, vmi, mm);
1041 khugepaged_enter_vma(res, vm_flags);
1042 return res;
1043
1044prealloc_fail:
1045 if (anon_dup)
1046 unlink_anon_vmas(anon_dup);
1047
1048anon_vma_fail:
1049 vma_iter_set(vmi, addr);
1050 vma_iter_load(vmi);
1051 return NULL;
1052}
1053
1054/*
1055 * Rough compatibility check to quickly see if it's even worth looking
1056 * at sharing an anon_vma.
1057 *
1058 * They need to have the same vm_file, and the flags can only differ
1059 * in things that mprotect may change.
1060 *
1061 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1062 * we can merge the two vma's. For example, we refuse to merge a vma if
1063 * there is a vm_ops->close() function, because that indicates that the
1064 * driver is doing some kind of reference counting. But that doesn't
1065 * really matter for the anon_vma sharing case.
1066 */
1067static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1068{
1069 return a->vm_end == b->vm_start &&
1070 mpol_equal(vma_policy(a), vma_policy(b)) &&
1071 a->vm_file == b->vm_file &&
1072 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1073 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1074}
1075
1076/*
1077 * Do some basic sanity checking to see if we can re-use the anon_vma
1078 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1079 * the same as 'old', the other will be the new one that is trying
1080 * to share the anon_vma.
1081 *
1082 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1083 * the anon_vma of 'old' is concurrently in the process of being set up
1084 * by another page fault trying to merge _that_. But that's ok: if it
1085 * is being set up, that automatically means that it will be a singleton
1086 * acceptable for merging, so we can do all of this optimistically. But
1087 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1088 *
1089 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1090 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1091 * is to return an anon_vma that is "complex" due to having gone through
1092 * a fork).
1093 *
1094 * We also make sure that the two vma's are compatible (adjacent,
1095 * and with the same memory policies). That's all stable, even with just
1096 * a read lock on the mmap_lock.
1097 */
1098static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1099{
1100 if (anon_vma_compatible(a, b)) {
1101 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1102
1103 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1104 return anon_vma;
1105 }
1106 return NULL;
1107}
1108
1109/*
1110 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1111 * neighbouring vmas for a suitable anon_vma, before it goes off
1112 * to allocate a new anon_vma. It checks because a repetitive
1113 * sequence of mprotects and faults may otherwise lead to distinct
1114 * anon_vmas being allocated, preventing vma merge in subsequent
1115 * mprotect.
1116 */
1117struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1118{
1119 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1120 struct anon_vma *anon_vma = NULL;
1121 struct vm_area_struct *prev, *next;
1122
1123 /* Try next first. */
1124 next = mas_walk(&mas);
1125 if (next) {
1126 anon_vma = reusable_anon_vma(next, vma, next);
1127 if (anon_vma)
1128 return anon_vma;
1129 }
1130
1131 prev = mas_prev(&mas, 0);
1132 VM_BUG_ON_VMA(prev != vma, vma);
1133 prev = mas_prev(&mas, 0);
1134 /* Try prev next. */
1135 if (prev)
1136 anon_vma = reusable_anon_vma(prev, prev, vma);
1137
1138 /*
1139 * We might reach here with anon_vma == NULL if we can't find
1140 * any reusable anon_vma.
1141 * There's no absolute need to look only at touching neighbours:
1142 * we could search further afield for "compatible" anon_vmas.
1143 * But it would probably just be a waste of time searching,
1144 * or lead to too many vmas hanging off the same anon_vma.
1145 * We're trying to allow mprotect remerging later on,
1146 * not trying to minimize memory used for anon_vmas.
1147 */
1148 return anon_vma;
1149}
1150
1151/*
1152 * If a hint addr is less than mmap_min_addr change hint to be as
1153 * low as possible but still greater than mmap_min_addr
1154 */
1155static inline unsigned long round_hint_to_min(unsigned long hint)
1156{
1157 hint &= PAGE_MASK;
1158 if (((void *)hint != NULL) &&
1159 (hint < mmap_min_addr))
1160 return PAGE_ALIGN(mmap_min_addr);
1161 return hint;
1162}
1163
1164bool mlock_future_ok(struct mm_struct *mm, unsigned long flags,
1165 unsigned long bytes)
1166{
1167 unsigned long locked_pages, limit_pages;
1168
1169 if (!(flags & VM_LOCKED) || capable(CAP_IPC_LOCK))
1170 return true;
1171
1172 locked_pages = bytes >> PAGE_SHIFT;
1173 locked_pages += mm->locked_vm;
1174
1175 limit_pages = rlimit(RLIMIT_MEMLOCK);
1176 limit_pages >>= PAGE_SHIFT;
1177
1178 return locked_pages <= limit_pages;
1179}
1180
1181static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1182{
1183 if (S_ISREG(inode->i_mode))
1184 return MAX_LFS_FILESIZE;
1185
1186 if (S_ISBLK(inode->i_mode))
1187 return MAX_LFS_FILESIZE;
1188
1189 if (S_ISSOCK(inode->i_mode))
1190 return MAX_LFS_FILESIZE;
1191
1192 /* Special "we do even unsigned file positions" case */
1193 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1194 return 0;
1195
1196 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1197 return ULONG_MAX;
1198}
1199
1200static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1201 unsigned long pgoff, unsigned long len)
1202{
1203 u64 maxsize = file_mmap_size_max(file, inode);
1204
1205 if (maxsize && len > maxsize)
1206 return false;
1207 maxsize -= len;
1208 if (pgoff > maxsize >> PAGE_SHIFT)
1209 return false;
1210 return true;
1211}
1212
1213/*
1214 * The caller must write-lock current->mm->mmap_lock.
1215 */
1216unsigned long do_mmap(struct file *file, unsigned long addr,
1217 unsigned long len, unsigned long prot,
1218 unsigned long flags, vm_flags_t vm_flags,
1219 unsigned long pgoff, unsigned long *populate,
1220 struct list_head *uf)
1221{
1222 struct mm_struct *mm = current->mm;
1223 int pkey = 0;
1224
1225 *populate = 0;
1226
1227 if (!len)
1228 return -EINVAL;
1229
1230 /*
1231 * Does the application expect PROT_READ to imply PROT_EXEC?
1232 *
1233 * (the exception is when the underlying filesystem is noexec
1234 * mounted, in which case we don't add PROT_EXEC.)
1235 */
1236 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1237 if (!(file && path_noexec(&file->f_path)))
1238 prot |= PROT_EXEC;
1239
1240 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1241 if (flags & MAP_FIXED_NOREPLACE)
1242 flags |= MAP_FIXED;
1243
1244 if (!(flags & MAP_FIXED))
1245 addr = round_hint_to_min(addr);
1246
1247 /* Careful about overflows.. */
1248 len = PAGE_ALIGN(len);
1249 if (!len)
1250 return -ENOMEM;
1251
1252 /* offset overflow? */
1253 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1254 return -EOVERFLOW;
1255
1256 /* Too many mappings? */
1257 if (mm->map_count > sysctl_max_map_count)
1258 return -ENOMEM;
1259
1260 /* Obtain the address to map to. we verify (or select) it and ensure
1261 * that it represents a valid section of the address space.
1262 */
1263 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1264 if (IS_ERR_VALUE(addr))
1265 return addr;
1266
1267 if (flags & MAP_FIXED_NOREPLACE) {
1268 if (find_vma_intersection(mm, addr, addr + len))
1269 return -EEXIST;
1270 }
1271
1272 if (prot == PROT_EXEC) {
1273 pkey = execute_only_pkey(mm);
1274 if (pkey < 0)
1275 pkey = 0;
1276 }
1277
1278 /* Do simple checking here so the lower-level routines won't have
1279 * to. we assume access permissions have been handled by the open
1280 * of the memory object, so we don't do any here.
1281 */
1282 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1283 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1284
1285 if (flags & MAP_LOCKED)
1286 if (!can_do_mlock())
1287 return -EPERM;
1288
1289 if (!mlock_future_ok(mm, vm_flags, len))
1290 return -EAGAIN;
1291
1292 if (file) {
1293 struct inode *inode = file_inode(file);
1294 unsigned long flags_mask;
1295
1296 if (!file_mmap_ok(file, inode, pgoff, len))
1297 return -EOVERFLOW;
1298
1299 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1300
1301 switch (flags & MAP_TYPE) {
1302 case MAP_SHARED:
1303 /*
1304 * Force use of MAP_SHARED_VALIDATE with non-legacy
1305 * flags. E.g. MAP_SYNC is dangerous to use with
1306 * MAP_SHARED as you don't know which consistency model
1307 * you will get. We silently ignore unsupported flags
1308 * with MAP_SHARED to preserve backward compatibility.
1309 */
1310 flags &= LEGACY_MAP_MASK;
1311 fallthrough;
1312 case MAP_SHARED_VALIDATE:
1313 if (flags & ~flags_mask)
1314 return -EOPNOTSUPP;
1315 if (prot & PROT_WRITE) {
1316 if (!(file->f_mode & FMODE_WRITE))
1317 return -EACCES;
1318 if (IS_SWAPFILE(file->f_mapping->host))
1319 return -ETXTBSY;
1320 }
1321
1322 /*
1323 * Make sure we don't allow writing to an append-only
1324 * file..
1325 */
1326 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1327 return -EACCES;
1328
1329 vm_flags |= VM_SHARED | VM_MAYSHARE;
1330 if (!(file->f_mode & FMODE_WRITE))
1331 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1332 fallthrough;
1333 case MAP_PRIVATE:
1334 if (!(file->f_mode & FMODE_READ))
1335 return -EACCES;
1336 if (path_noexec(&file->f_path)) {
1337 if (vm_flags & VM_EXEC)
1338 return -EPERM;
1339 vm_flags &= ~VM_MAYEXEC;
1340 }
1341
1342 if (!file->f_op->mmap)
1343 return -ENODEV;
1344 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1345 return -EINVAL;
1346 break;
1347
1348 default:
1349 return -EINVAL;
1350 }
1351 } else {
1352 switch (flags & MAP_TYPE) {
1353 case MAP_SHARED:
1354 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1355 return -EINVAL;
1356 /*
1357 * Ignore pgoff.
1358 */
1359 pgoff = 0;
1360 vm_flags |= VM_SHARED | VM_MAYSHARE;
1361 break;
1362 case MAP_PRIVATE:
1363 /*
1364 * Set pgoff according to addr for anon_vma.
1365 */
1366 pgoff = addr >> PAGE_SHIFT;
1367 break;
1368 default:
1369 return -EINVAL;
1370 }
1371 }
1372
1373 /*
1374 * Set 'VM_NORESERVE' if we should not account for the
1375 * memory use of this mapping.
1376 */
1377 if (flags & MAP_NORESERVE) {
1378 /* We honor MAP_NORESERVE if allowed to overcommit */
1379 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1380 vm_flags |= VM_NORESERVE;
1381
1382 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1383 if (file && is_file_hugepages(file))
1384 vm_flags |= VM_NORESERVE;
1385 }
1386
1387 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1388 if (!IS_ERR_VALUE(addr) &&
1389 ((vm_flags & VM_LOCKED) ||
1390 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1391 *populate = len;
1392 return addr;
1393}
1394
1395unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1396 unsigned long prot, unsigned long flags,
1397 unsigned long fd, unsigned long pgoff)
1398{
1399 struct file *file = NULL;
1400 unsigned long retval;
1401
1402 if (!(flags & MAP_ANONYMOUS)) {
1403 audit_mmap_fd(fd, flags);
1404 file = fget(fd);
1405 if (!file)
1406 return -EBADF;
1407 if (is_file_hugepages(file)) {
1408 len = ALIGN(len, huge_page_size(hstate_file(file)));
1409 } else if (unlikely(flags & MAP_HUGETLB)) {
1410 retval = -EINVAL;
1411 goto out_fput;
1412 }
1413 } else if (flags & MAP_HUGETLB) {
1414 struct hstate *hs;
1415
1416 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1417 if (!hs)
1418 return -EINVAL;
1419
1420 len = ALIGN(len, huge_page_size(hs));
1421 /*
1422 * VM_NORESERVE is used because the reservations will be
1423 * taken when vm_ops->mmap() is called
1424 */
1425 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1426 VM_NORESERVE,
1427 HUGETLB_ANONHUGE_INODE,
1428 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1429 if (IS_ERR(file))
1430 return PTR_ERR(file);
1431 }
1432
1433 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1434out_fput:
1435 if (file)
1436 fput(file);
1437 return retval;
1438}
1439
1440SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1441 unsigned long, prot, unsigned long, flags,
1442 unsigned long, fd, unsigned long, pgoff)
1443{
1444 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1445}
1446
1447#ifdef __ARCH_WANT_SYS_OLD_MMAP
1448struct mmap_arg_struct {
1449 unsigned long addr;
1450 unsigned long len;
1451 unsigned long prot;
1452 unsigned long flags;
1453 unsigned long fd;
1454 unsigned long offset;
1455};
1456
1457SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1458{
1459 struct mmap_arg_struct a;
1460
1461 if (copy_from_user(&a, arg, sizeof(a)))
1462 return -EFAULT;
1463 if (offset_in_page(a.offset))
1464 return -EINVAL;
1465
1466 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1467 a.offset >> PAGE_SHIFT);
1468}
1469#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1470
1471static bool vm_ops_needs_writenotify(const struct vm_operations_struct *vm_ops)
1472{
1473 return vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite);
1474}
1475
1476static bool vma_is_shared_writable(struct vm_area_struct *vma)
1477{
1478 return (vma->vm_flags & (VM_WRITE | VM_SHARED)) ==
1479 (VM_WRITE | VM_SHARED);
1480}
1481
1482static bool vma_fs_can_writeback(struct vm_area_struct *vma)
1483{
1484 /* No managed pages to writeback. */
1485 if (vma->vm_flags & VM_PFNMAP)
1486 return false;
1487
1488 return vma->vm_file && vma->vm_file->f_mapping &&
1489 mapping_can_writeback(vma->vm_file->f_mapping);
1490}
1491
1492/*
1493 * Does this VMA require the underlying folios to have their dirty state
1494 * tracked?
1495 */
1496bool vma_needs_dirty_tracking(struct vm_area_struct *vma)
1497{
1498 /* Only shared, writable VMAs require dirty tracking. */
1499 if (!vma_is_shared_writable(vma))
1500 return false;
1501
1502 /* Does the filesystem need to be notified? */
1503 if (vm_ops_needs_writenotify(vma->vm_ops))
1504 return true;
1505
1506 /*
1507 * Even if the filesystem doesn't indicate a need for writenotify, if it
1508 * can writeback, dirty tracking is still required.
1509 */
1510 return vma_fs_can_writeback(vma);
1511}
1512
1513/*
1514 * Some shared mappings will want the pages marked read-only
1515 * to track write events. If so, we'll downgrade vm_page_prot
1516 * to the private version (using protection_map[] without the
1517 * VM_SHARED bit).
1518 */
1519int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1520{
1521 /* If it was private or non-writable, the write bit is already clear */
1522 if (!vma_is_shared_writable(vma))
1523 return 0;
1524
1525 /* The backer wishes to know when pages are first written to? */
1526 if (vm_ops_needs_writenotify(vma->vm_ops))
1527 return 1;
1528
1529 /* The open routine did something to the protections that pgprot_modify
1530 * won't preserve? */
1531 if (pgprot_val(vm_page_prot) !=
1532 pgprot_val(vm_pgprot_modify(vm_page_prot, vma->vm_flags)))
1533 return 0;
1534
1535 /*
1536 * Do we need to track softdirty? hugetlb does not support softdirty
1537 * tracking yet.
1538 */
1539 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1540 return 1;
1541
1542 /* Do we need write faults for uffd-wp tracking? */
1543 if (userfaultfd_wp(vma))
1544 return 1;
1545
1546 /* Can the mapping track the dirty pages? */
1547 return vma_fs_can_writeback(vma);
1548}
1549
1550/*
1551 * We account for memory if it's a private writeable mapping,
1552 * not hugepages and VM_NORESERVE wasn't set.
1553 */
1554static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1555{
1556 /*
1557 * hugetlb has its own accounting separate from the core VM
1558 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1559 */
1560 if (file && is_file_hugepages(file))
1561 return 0;
1562
1563 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1564}
1565
1566/**
1567 * unmapped_area() - Find an area between the low_limit and the high_limit with
1568 * the correct alignment and offset, all from @info. Note: current->mm is used
1569 * for the search.
1570 *
1571 * @info: The unmapped area information including the range [low_limit -
1572 * high_limit), the alignment offset and mask.
1573 *
1574 * Return: A memory address or -ENOMEM.
1575 */
1576static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1577{
1578 unsigned long length, gap;
1579 unsigned long low_limit, high_limit;
1580 struct vm_area_struct *tmp;
1581
1582 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1583
1584 /* Adjust search length to account for worst case alignment overhead */
1585 length = info->length + info->align_mask;
1586 if (length < info->length)
1587 return -ENOMEM;
1588
1589 low_limit = info->low_limit;
1590 if (low_limit < mmap_min_addr)
1591 low_limit = mmap_min_addr;
1592 high_limit = info->high_limit;
1593retry:
1594 if (mas_empty_area(&mas, low_limit, high_limit - 1, length))
1595 return -ENOMEM;
1596
1597 gap = mas.index;
1598 gap += (info->align_offset - gap) & info->align_mask;
1599 tmp = mas_next(&mas, ULONG_MAX);
1600 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1601 if (vm_start_gap(tmp) < gap + length - 1) {
1602 low_limit = tmp->vm_end;
1603 mas_reset(&mas);
1604 goto retry;
1605 }
1606 } else {
1607 tmp = mas_prev(&mas, 0);
1608 if (tmp && vm_end_gap(tmp) > gap) {
1609 low_limit = vm_end_gap(tmp);
1610 mas_reset(&mas);
1611 goto retry;
1612 }
1613 }
1614
1615 return gap;
1616}
1617
1618/**
1619 * unmapped_area_topdown() - Find an area between the low_limit and the
1620 * high_limit with the correct alignment and offset at the highest available
1621 * address, all from @info. Note: current->mm is used for the search.
1622 *
1623 * @info: The unmapped area information including the range [low_limit -
1624 * high_limit), the alignment offset and mask.
1625 *
1626 * Return: A memory address or -ENOMEM.
1627 */
1628static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1629{
1630 unsigned long length, gap, gap_end;
1631 unsigned long low_limit, high_limit;
1632 struct vm_area_struct *tmp;
1633
1634 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1635 /* Adjust search length to account for worst case alignment overhead */
1636 length = info->length + info->align_mask;
1637 if (length < info->length)
1638 return -ENOMEM;
1639
1640 low_limit = info->low_limit;
1641 if (low_limit < mmap_min_addr)
1642 low_limit = mmap_min_addr;
1643 high_limit = info->high_limit;
1644retry:
1645 if (mas_empty_area_rev(&mas, low_limit, high_limit - 1, length))
1646 return -ENOMEM;
1647
1648 gap = mas.last + 1 - info->length;
1649 gap -= (gap - info->align_offset) & info->align_mask;
1650 gap_end = mas.last;
1651 tmp = mas_next(&mas, ULONG_MAX);
1652 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1653 if (vm_start_gap(tmp) <= gap_end) {
1654 high_limit = vm_start_gap(tmp);
1655 mas_reset(&mas);
1656 goto retry;
1657 }
1658 } else {
1659 tmp = mas_prev(&mas, 0);
1660 if (tmp && vm_end_gap(tmp) > gap) {
1661 high_limit = tmp->vm_start;
1662 mas_reset(&mas);
1663 goto retry;
1664 }
1665 }
1666
1667 return gap;
1668}
1669
1670/*
1671 * Search for an unmapped address range.
1672 *
1673 * We are looking for a range that:
1674 * - does not intersect with any VMA;
1675 * - is contained within the [low_limit, high_limit) interval;
1676 * - is at least the desired size.
1677 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1678 */
1679unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1680{
1681 unsigned long addr;
1682
1683 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1684 addr = unmapped_area_topdown(info);
1685 else
1686 addr = unmapped_area(info);
1687
1688 trace_vm_unmapped_area(addr, info);
1689 return addr;
1690}
1691
1692/* Get an address range which is currently unmapped.
1693 * For shmat() with addr=0.
1694 *
1695 * Ugly calling convention alert:
1696 * Return value with the low bits set means error value,
1697 * ie
1698 * if (ret & ~PAGE_MASK)
1699 * error = ret;
1700 *
1701 * This function "knows" that -ENOMEM has the bits set.
1702 */
1703unsigned long
1704generic_get_unmapped_area(struct file *filp, unsigned long addr,
1705 unsigned long len, unsigned long pgoff,
1706 unsigned long flags)
1707{
1708 struct mm_struct *mm = current->mm;
1709 struct vm_area_struct *vma, *prev;
1710 struct vm_unmapped_area_info info;
1711 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1712
1713 if (len > mmap_end - mmap_min_addr)
1714 return -ENOMEM;
1715
1716 if (flags & MAP_FIXED)
1717 return addr;
1718
1719 if (addr) {
1720 addr = PAGE_ALIGN(addr);
1721 vma = find_vma_prev(mm, addr, &prev);
1722 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1723 (!vma || addr + len <= vm_start_gap(vma)) &&
1724 (!prev || addr >= vm_end_gap(prev)))
1725 return addr;
1726 }
1727
1728 info.flags = 0;
1729 info.length = len;
1730 info.low_limit = mm->mmap_base;
1731 info.high_limit = mmap_end;
1732 info.align_mask = 0;
1733 info.align_offset = 0;
1734 return vm_unmapped_area(&info);
1735}
1736
1737#ifndef HAVE_ARCH_UNMAPPED_AREA
1738unsigned long
1739arch_get_unmapped_area(struct file *filp, unsigned long addr,
1740 unsigned long len, unsigned long pgoff,
1741 unsigned long flags)
1742{
1743 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1744}
1745#endif
1746
1747/*
1748 * This mmap-allocator allocates new areas top-down from below the
1749 * stack's low limit (the base):
1750 */
1751unsigned long
1752generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1753 unsigned long len, unsigned long pgoff,
1754 unsigned long flags)
1755{
1756 struct vm_area_struct *vma, *prev;
1757 struct mm_struct *mm = current->mm;
1758 struct vm_unmapped_area_info info;
1759 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1760
1761 /* requested length too big for entire address space */
1762 if (len > mmap_end - mmap_min_addr)
1763 return -ENOMEM;
1764
1765 if (flags & MAP_FIXED)
1766 return addr;
1767
1768 /* requesting a specific address */
1769 if (addr) {
1770 addr = PAGE_ALIGN(addr);
1771 vma = find_vma_prev(mm, addr, &prev);
1772 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1773 (!vma || addr + len <= vm_start_gap(vma)) &&
1774 (!prev || addr >= vm_end_gap(prev)))
1775 return addr;
1776 }
1777
1778 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1779 info.length = len;
1780 info.low_limit = PAGE_SIZE;
1781 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1782 info.align_mask = 0;
1783 info.align_offset = 0;
1784 addr = vm_unmapped_area(&info);
1785
1786 /*
1787 * A failed mmap() very likely causes application failure,
1788 * so fall back to the bottom-up function here. This scenario
1789 * can happen with large stack limits and large mmap()
1790 * allocations.
1791 */
1792 if (offset_in_page(addr)) {
1793 VM_BUG_ON(addr != -ENOMEM);
1794 info.flags = 0;
1795 info.low_limit = TASK_UNMAPPED_BASE;
1796 info.high_limit = mmap_end;
1797 addr = vm_unmapped_area(&info);
1798 }
1799
1800 return addr;
1801}
1802
1803#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1804unsigned long
1805arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1806 unsigned long len, unsigned long pgoff,
1807 unsigned long flags)
1808{
1809 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1810}
1811#endif
1812
1813unsigned long
1814get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1815 unsigned long pgoff, unsigned long flags)
1816{
1817 unsigned long (*get_area)(struct file *, unsigned long,
1818 unsigned long, unsigned long, unsigned long);
1819
1820 unsigned long error = arch_mmap_check(addr, len, flags);
1821 if (error)
1822 return error;
1823
1824 /* Careful about overflows.. */
1825 if (len > TASK_SIZE)
1826 return -ENOMEM;
1827
1828 get_area = current->mm->get_unmapped_area;
1829 if (file) {
1830 if (file->f_op->get_unmapped_area)
1831 get_area = file->f_op->get_unmapped_area;
1832 } else if (flags & MAP_SHARED) {
1833 /*
1834 * mmap_region() will call shmem_zero_setup() to create a file,
1835 * so use shmem's get_unmapped_area in case it can be huge.
1836 */
1837 get_area = shmem_get_unmapped_area;
1838 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1839 /* Ensures that larger anonymous mappings are THP aligned. */
1840 get_area = thp_get_unmapped_area;
1841 }
1842
1843 /* Always treat pgoff as zero for anonymous memory. */
1844 if (!file)
1845 pgoff = 0;
1846
1847 addr = get_area(file, addr, len, pgoff, flags);
1848 if (IS_ERR_VALUE(addr))
1849 return addr;
1850
1851 if (addr > TASK_SIZE - len)
1852 return -ENOMEM;
1853 if (offset_in_page(addr))
1854 return -EINVAL;
1855
1856 error = security_mmap_addr(addr);
1857 return error ? error : addr;
1858}
1859
1860EXPORT_SYMBOL(get_unmapped_area);
1861
1862/**
1863 * find_vma_intersection() - Look up the first VMA which intersects the interval
1864 * @mm: The process address space.
1865 * @start_addr: The inclusive start user address.
1866 * @end_addr: The exclusive end user address.
1867 *
1868 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1869 * start_addr < end_addr.
1870 */
1871struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1872 unsigned long start_addr,
1873 unsigned long end_addr)
1874{
1875 unsigned long index = start_addr;
1876
1877 mmap_assert_locked(mm);
1878 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1879}
1880EXPORT_SYMBOL(find_vma_intersection);
1881
1882/**
1883 * find_vma() - Find the VMA for a given address, or the next VMA.
1884 * @mm: The mm_struct to check
1885 * @addr: The address
1886 *
1887 * Returns: The VMA associated with addr, or the next VMA.
1888 * May return %NULL in the case of no VMA at addr or above.
1889 */
1890struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1891{
1892 unsigned long index = addr;
1893
1894 mmap_assert_locked(mm);
1895 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1896}
1897EXPORT_SYMBOL(find_vma);
1898
1899/**
1900 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1901 * set %pprev to the previous VMA, if any.
1902 * @mm: The mm_struct to check
1903 * @addr: The address
1904 * @pprev: The pointer to set to the previous VMA
1905 *
1906 * Note that RCU lock is missing here since the external mmap_lock() is used
1907 * instead.
1908 *
1909 * Returns: The VMA associated with @addr, or the next vma.
1910 * May return %NULL in the case of no vma at addr or above.
1911 */
1912struct vm_area_struct *
1913find_vma_prev(struct mm_struct *mm, unsigned long addr,
1914 struct vm_area_struct **pprev)
1915{
1916 struct vm_area_struct *vma;
1917 MA_STATE(mas, &mm->mm_mt, addr, addr);
1918
1919 vma = mas_walk(&mas);
1920 *pprev = mas_prev(&mas, 0);
1921 if (!vma)
1922 vma = mas_next(&mas, ULONG_MAX);
1923 return vma;
1924}
1925
1926/*
1927 * Verify that the stack growth is acceptable and
1928 * update accounting. This is shared with both the
1929 * grow-up and grow-down cases.
1930 */
1931static int acct_stack_growth(struct vm_area_struct *vma,
1932 unsigned long size, unsigned long grow)
1933{
1934 struct mm_struct *mm = vma->vm_mm;
1935 unsigned long new_start;
1936
1937 /* address space limit tests */
1938 if (!may_expand_vm(mm, vma->vm_flags, grow))
1939 return -ENOMEM;
1940
1941 /* Stack limit test */
1942 if (size > rlimit(RLIMIT_STACK))
1943 return -ENOMEM;
1944
1945 /* mlock limit tests */
1946 if (!mlock_future_ok(mm, vma->vm_flags, grow << PAGE_SHIFT))
1947 return -ENOMEM;
1948
1949 /* Check to ensure the stack will not grow into a hugetlb-only region */
1950 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1951 vma->vm_end - size;
1952 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1953 return -EFAULT;
1954
1955 /*
1956 * Overcommit.. This must be the final test, as it will
1957 * update security statistics.
1958 */
1959 if (security_vm_enough_memory_mm(mm, grow))
1960 return -ENOMEM;
1961
1962 return 0;
1963}
1964
1965#if defined(CONFIG_STACK_GROWSUP)
1966/*
1967 * PA-RISC uses this for its stack.
1968 * vma is the last one with address > vma->vm_end. Have to extend vma.
1969 */
1970static int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1971{
1972 struct mm_struct *mm = vma->vm_mm;
1973 struct vm_area_struct *next;
1974 unsigned long gap_addr;
1975 int error = 0;
1976 MA_STATE(mas, &mm->mm_mt, vma->vm_start, address);
1977
1978 if (!(vma->vm_flags & VM_GROWSUP))
1979 return -EFAULT;
1980
1981 /* Guard against exceeding limits of the address space. */
1982 address &= PAGE_MASK;
1983 if (address >= (TASK_SIZE & PAGE_MASK))
1984 return -ENOMEM;
1985 address += PAGE_SIZE;
1986
1987 /* Enforce stack_guard_gap */
1988 gap_addr = address + stack_guard_gap;
1989
1990 /* Guard against overflow */
1991 if (gap_addr < address || gap_addr > TASK_SIZE)
1992 gap_addr = TASK_SIZE;
1993
1994 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1995 if (next && vma_is_accessible(next)) {
1996 if (!(next->vm_flags & VM_GROWSUP))
1997 return -ENOMEM;
1998 /* Check that both stack segments have the same anon_vma? */
1999 }
2000
2001 if (next)
2002 mas_prev_range(&mas, address);
2003
2004 __mas_set_range(&mas, vma->vm_start, address - 1);
2005 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2006 return -ENOMEM;
2007
2008 /* We must make sure the anon_vma is allocated. */
2009 if (unlikely(anon_vma_prepare(vma))) {
2010 mas_destroy(&mas);
2011 return -ENOMEM;
2012 }
2013
2014 /* Lock the VMA before expanding to prevent concurrent page faults */
2015 vma_start_write(vma);
2016 /*
2017 * vma->vm_start/vm_end cannot change under us because the caller
2018 * is required to hold the mmap_lock in read mode. We need the
2019 * anon_vma lock to serialize against concurrent expand_stacks.
2020 */
2021 anon_vma_lock_write(vma->anon_vma);
2022
2023 /* Somebody else might have raced and expanded it already */
2024 if (address > vma->vm_end) {
2025 unsigned long size, grow;
2026
2027 size = address - vma->vm_start;
2028 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2029
2030 error = -ENOMEM;
2031 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2032 error = acct_stack_growth(vma, size, grow);
2033 if (!error) {
2034 /*
2035 * We only hold a shared mmap_lock lock here, so
2036 * we need to protect against concurrent vma
2037 * expansions. anon_vma_lock_write() doesn't
2038 * help here, as we don't guarantee that all
2039 * growable vmas in a mm share the same root
2040 * anon vma. So, we reuse mm->page_table_lock
2041 * to guard against concurrent vma expansions.
2042 */
2043 spin_lock(&mm->page_table_lock);
2044 if (vma->vm_flags & VM_LOCKED)
2045 mm->locked_vm += grow;
2046 vm_stat_account(mm, vma->vm_flags, grow);
2047 anon_vma_interval_tree_pre_update_vma(vma);
2048 vma->vm_end = address;
2049 /* Overwrite old entry in mtree. */
2050 mas_store_prealloc(&mas, vma);
2051 anon_vma_interval_tree_post_update_vma(vma);
2052 spin_unlock(&mm->page_table_lock);
2053
2054 perf_event_mmap(vma);
2055 }
2056 }
2057 }
2058 anon_vma_unlock_write(vma->anon_vma);
2059 khugepaged_enter_vma(vma, vma->vm_flags);
2060 mas_destroy(&mas);
2061 validate_mm(mm);
2062 return error;
2063}
2064#endif /* CONFIG_STACK_GROWSUP */
2065
2066/*
2067 * vma is the first one with address < vma->vm_start. Have to extend vma.
2068 * mmap_lock held for writing.
2069 */
2070int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2071{
2072 struct mm_struct *mm = vma->vm_mm;
2073 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2074 struct vm_area_struct *prev;
2075 int error = 0;
2076
2077 if (!(vma->vm_flags & VM_GROWSDOWN))
2078 return -EFAULT;
2079
2080 address &= PAGE_MASK;
2081 if (address < mmap_min_addr || address < FIRST_USER_ADDRESS)
2082 return -EPERM;
2083
2084 /* Enforce stack_guard_gap */
2085 prev = mas_prev(&mas, 0);
2086 /* Check that both stack segments have the same anon_vma? */
2087 if (prev) {
2088 if (!(prev->vm_flags & VM_GROWSDOWN) &&
2089 vma_is_accessible(prev) &&
2090 (address - prev->vm_end < stack_guard_gap))
2091 return -ENOMEM;
2092 }
2093
2094 if (prev)
2095 mas_next_range(&mas, vma->vm_start);
2096
2097 __mas_set_range(&mas, address, vma->vm_end - 1);
2098 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2099 return -ENOMEM;
2100
2101 /* We must make sure the anon_vma is allocated. */
2102 if (unlikely(anon_vma_prepare(vma))) {
2103 mas_destroy(&mas);
2104 return -ENOMEM;
2105 }
2106
2107 /* Lock the VMA before expanding to prevent concurrent page faults */
2108 vma_start_write(vma);
2109 /*
2110 * vma->vm_start/vm_end cannot change under us because the caller
2111 * is required to hold the mmap_lock in read mode. We need the
2112 * anon_vma lock to serialize against concurrent expand_stacks.
2113 */
2114 anon_vma_lock_write(vma->anon_vma);
2115
2116 /* Somebody else might have raced and expanded it already */
2117 if (address < vma->vm_start) {
2118 unsigned long size, grow;
2119
2120 size = vma->vm_end - address;
2121 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2122
2123 error = -ENOMEM;
2124 if (grow <= vma->vm_pgoff) {
2125 error = acct_stack_growth(vma, size, grow);
2126 if (!error) {
2127 /*
2128 * We only hold a shared mmap_lock lock here, so
2129 * we need to protect against concurrent vma
2130 * expansions. anon_vma_lock_write() doesn't
2131 * help here, as we don't guarantee that all
2132 * growable vmas in a mm share the same root
2133 * anon vma. So, we reuse mm->page_table_lock
2134 * to guard against concurrent vma expansions.
2135 */
2136 spin_lock(&mm->page_table_lock);
2137 if (vma->vm_flags & VM_LOCKED)
2138 mm->locked_vm += grow;
2139 vm_stat_account(mm, vma->vm_flags, grow);
2140 anon_vma_interval_tree_pre_update_vma(vma);
2141 vma->vm_start = address;
2142 vma->vm_pgoff -= grow;
2143 /* Overwrite old entry in mtree. */
2144 mas_store_prealloc(&mas, vma);
2145 anon_vma_interval_tree_post_update_vma(vma);
2146 spin_unlock(&mm->page_table_lock);
2147
2148 perf_event_mmap(vma);
2149 }
2150 }
2151 }
2152 anon_vma_unlock_write(vma->anon_vma);
2153 khugepaged_enter_vma(vma, vma->vm_flags);
2154 mas_destroy(&mas);
2155 validate_mm(mm);
2156 return error;
2157}
2158
2159/* enforced gap between the expanding stack and other mappings. */
2160unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2161
2162static int __init cmdline_parse_stack_guard_gap(char *p)
2163{
2164 unsigned long val;
2165 char *endptr;
2166
2167 val = simple_strtoul(p, &endptr, 10);
2168 if (!*endptr)
2169 stack_guard_gap = val << PAGE_SHIFT;
2170
2171 return 1;
2172}
2173__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2174
2175#ifdef CONFIG_STACK_GROWSUP
2176int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2177{
2178 return expand_upwards(vma, address);
2179}
2180
2181struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2182{
2183 struct vm_area_struct *vma, *prev;
2184
2185 addr &= PAGE_MASK;
2186 vma = find_vma_prev(mm, addr, &prev);
2187 if (vma && (vma->vm_start <= addr))
2188 return vma;
2189 if (!prev)
2190 return NULL;
2191 if (expand_stack_locked(prev, addr))
2192 return NULL;
2193 if (prev->vm_flags & VM_LOCKED)
2194 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2195 return prev;
2196}
2197#else
2198int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2199{
2200 return expand_downwards(vma, address);
2201}
2202
2203struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2204{
2205 struct vm_area_struct *vma;
2206 unsigned long start;
2207
2208 addr &= PAGE_MASK;
2209 vma = find_vma(mm, addr);
2210 if (!vma)
2211 return NULL;
2212 if (vma->vm_start <= addr)
2213 return vma;
2214 start = vma->vm_start;
2215 if (expand_stack_locked(vma, addr))
2216 return NULL;
2217 if (vma->vm_flags & VM_LOCKED)
2218 populate_vma_page_range(vma, addr, start, NULL);
2219 return vma;
2220}
2221#endif
2222
2223#if defined(CONFIG_STACK_GROWSUP)
2224
2225#define vma_expand_up(vma,addr) expand_upwards(vma, addr)
2226#define vma_expand_down(vma, addr) (-EFAULT)
2227
2228#else
2229
2230#define vma_expand_up(vma,addr) (-EFAULT)
2231#define vma_expand_down(vma, addr) expand_downwards(vma, addr)
2232
2233#endif
2234
2235/*
2236 * expand_stack(): legacy interface for page faulting. Don't use unless
2237 * you have to.
2238 *
2239 * This is called with the mm locked for reading, drops the lock, takes
2240 * the lock for writing, tries to look up a vma again, expands it if
2241 * necessary, and downgrades the lock to reading again.
2242 *
2243 * If no vma is found or it can't be expanded, it returns NULL and has
2244 * dropped the lock.
2245 */
2246struct vm_area_struct *expand_stack(struct mm_struct *mm, unsigned long addr)
2247{
2248 struct vm_area_struct *vma, *prev;
2249
2250 mmap_read_unlock(mm);
2251 if (mmap_write_lock_killable(mm))
2252 return NULL;
2253
2254 vma = find_vma_prev(mm, addr, &prev);
2255 if (vma && vma->vm_start <= addr)
2256 goto success;
2257
2258 if (prev && !vma_expand_up(prev, addr)) {
2259 vma = prev;
2260 goto success;
2261 }
2262
2263 if (vma && !vma_expand_down(vma, addr))
2264 goto success;
2265
2266 mmap_write_unlock(mm);
2267 return NULL;
2268
2269success:
2270 mmap_write_downgrade(mm);
2271 return vma;
2272}
2273
2274/*
2275 * Ok - we have the memory areas we should free on a maple tree so release them,
2276 * and do the vma updates.
2277 *
2278 * Called with the mm semaphore held.
2279 */
2280static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2281{
2282 unsigned long nr_accounted = 0;
2283 struct vm_area_struct *vma;
2284
2285 /* Update high watermark before we lower total_vm */
2286 update_hiwater_vm(mm);
2287 mas_for_each(mas, vma, ULONG_MAX) {
2288 long nrpages = vma_pages(vma);
2289
2290 if (vma->vm_flags & VM_ACCOUNT)
2291 nr_accounted += nrpages;
2292 vm_stat_account(mm, vma->vm_flags, -nrpages);
2293 remove_vma(vma, false);
2294 }
2295 vm_unacct_memory(nr_accounted);
2296}
2297
2298/*
2299 * Get rid of page table information in the indicated region.
2300 *
2301 * Called with the mm semaphore held.
2302 */
2303static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
2304 struct vm_area_struct *vma, struct vm_area_struct *prev,
2305 struct vm_area_struct *next, unsigned long start,
2306 unsigned long end, unsigned long tree_end, bool mm_wr_locked)
2307{
2308 struct mmu_gather tlb;
2309 unsigned long mt_start = mas->index;
2310
2311 lru_add_drain();
2312 tlb_gather_mmu(&tlb, mm);
2313 update_hiwater_rss(mm);
2314 unmap_vmas(&tlb, mas, vma, start, end, tree_end, mm_wr_locked);
2315 mas_set(mas, mt_start);
2316 free_pgtables(&tlb, mas, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2317 next ? next->vm_start : USER_PGTABLES_CEILING,
2318 mm_wr_locked);
2319 tlb_finish_mmu(&tlb);
2320}
2321
2322/*
2323 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2324 * has already been checked or doesn't make sense to fail.
2325 * VMA Iterator will point to the end VMA.
2326 */
2327static int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2328 unsigned long addr, int new_below)
2329{
2330 struct vma_prepare vp;
2331 struct vm_area_struct *new;
2332 int err;
2333
2334 WARN_ON(vma->vm_start >= addr);
2335 WARN_ON(vma->vm_end <= addr);
2336
2337 if (vma->vm_ops && vma->vm_ops->may_split) {
2338 err = vma->vm_ops->may_split(vma, addr);
2339 if (err)
2340 return err;
2341 }
2342
2343 new = vm_area_dup(vma);
2344 if (!new)
2345 return -ENOMEM;
2346
2347 if (new_below) {
2348 new->vm_end = addr;
2349 } else {
2350 new->vm_start = addr;
2351 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2352 }
2353
2354 err = -ENOMEM;
2355 vma_iter_config(vmi, new->vm_start, new->vm_end);
2356 if (vma_iter_prealloc(vmi, new))
2357 goto out_free_vma;
2358
2359 err = vma_dup_policy(vma, new);
2360 if (err)
2361 goto out_free_vmi;
2362
2363 err = anon_vma_clone(new, vma);
2364 if (err)
2365 goto out_free_mpol;
2366
2367 if (new->vm_file)
2368 get_file(new->vm_file);
2369
2370 if (new->vm_ops && new->vm_ops->open)
2371 new->vm_ops->open(new);
2372
2373 vma_start_write(vma);
2374 vma_start_write(new);
2375
2376 init_vma_prep(&vp, vma);
2377 vp.insert = new;
2378 vma_prepare(&vp);
2379 vma_adjust_trans_huge(vma, vma->vm_start, addr, 0);
2380
2381 if (new_below) {
2382 vma->vm_start = addr;
2383 vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT;
2384 } else {
2385 vma->vm_end = addr;
2386 }
2387
2388 /* vma_complete stores the new vma */
2389 vma_complete(&vp, vmi, vma->vm_mm);
2390
2391 /* Success. */
2392 if (new_below)
2393 vma_next(vmi);
2394 return 0;
2395
2396out_free_mpol:
2397 mpol_put(vma_policy(new));
2398out_free_vmi:
2399 vma_iter_free(vmi);
2400out_free_vma:
2401 vm_area_free(new);
2402 return err;
2403}
2404
2405/*
2406 * Split a vma into two pieces at address 'addr', a new vma is allocated
2407 * either for the first part or the tail.
2408 */
2409static int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2410 unsigned long addr, int new_below)
2411{
2412 if (vma->vm_mm->map_count >= sysctl_max_map_count)
2413 return -ENOMEM;
2414
2415 return __split_vma(vmi, vma, addr, new_below);
2416}
2417
2418/*
2419 * We are about to modify one or multiple of a VMA's flags, policy, userfaultfd
2420 * context and anonymous VMA name within the range [start, end).
2421 *
2422 * As a result, we might be able to merge the newly modified VMA range with an
2423 * adjacent VMA with identical properties.
2424 *
2425 * If no merge is possible and the range does not span the entirety of the VMA,
2426 * we then need to split the VMA to accommodate the change.
2427 *
2428 * The function returns either the merged VMA, the original VMA if a split was
2429 * required instead, or an error if the split failed.
2430 */
2431struct vm_area_struct *vma_modify(struct vma_iterator *vmi,
2432 struct vm_area_struct *prev,
2433 struct vm_area_struct *vma,
2434 unsigned long start, unsigned long end,
2435 unsigned long vm_flags,
2436 struct mempolicy *policy,
2437 struct vm_userfaultfd_ctx uffd_ctx,
2438 struct anon_vma_name *anon_name)
2439{
2440 pgoff_t pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
2441 struct vm_area_struct *merged;
2442
2443 merged = vma_merge(vmi, vma->vm_mm, prev, start, end, vm_flags,
2444 vma->anon_vma, vma->vm_file, pgoff, policy,
2445 uffd_ctx, anon_name);
2446 if (merged)
2447 return merged;
2448
2449 if (vma->vm_start < start) {
2450 int err = split_vma(vmi, vma, start, 1);
2451
2452 if (err)
2453 return ERR_PTR(err);
2454 }
2455
2456 if (vma->vm_end > end) {
2457 int err = split_vma(vmi, vma, end, 0);
2458
2459 if (err)
2460 return ERR_PTR(err);
2461 }
2462
2463 return vma;
2464}
2465
2466/*
2467 * Attempt to merge a newly mapped VMA with those adjacent to it. The caller
2468 * must ensure that [start, end) does not overlap any existing VMA.
2469 */
2470static struct vm_area_struct
2471*vma_merge_new_vma(struct vma_iterator *vmi, struct vm_area_struct *prev,
2472 struct vm_area_struct *vma, unsigned long start,
2473 unsigned long end, pgoff_t pgoff)
2474{
2475 return vma_merge(vmi, vma->vm_mm, prev, start, end, vma->vm_flags,
2476 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
2477 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
2478}
2479
2480/*
2481 * Expand vma by delta bytes, potentially merging with an immediately adjacent
2482 * VMA with identical properties.
2483 */
2484struct vm_area_struct *vma_merge_extend(struct vma_iterator *vmi,
2485 struct vm_area_struct *vma,
2486 unsigned long delta)
2487{
2488 pgoff_t pgoff = vma->vm_pgoff + vma_pages(vma);
2489
2490 /* vma is specified as prev, so case 1 or 2 will apply. */
2491 return vma_merge(vmi, vma->vm_mm, vma, vma->vm_end, vma->vm_end + delta,
2492 vma->vm_flags, vma->anon_vma, vma->vm_file, pgoff,
2493 vma_policy(vma), vma->vm_userfaultfd_ctx,
2494 anon_vma_name(vma));
2495}
2496
2497/*
2498 * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
2499 * @vmi: The vma iterator
2500 * @vma: The starting vm_area_struct
2501 * @mm: The mm_struct
2502 * @start: The aligned start address to munmap.
2503 * @end: The aligned end address to munmap.
2504 * @uf: The userfaultfd list_head
2505 * @unlock: Set to true to drop the mmap_lock. unlocking only happens on
2506 * success.
2507 *
2508 * Return: 0 on success and drops the lock if so directed, error and leaves the
2509 * lock held otherwise.
2510 */
2511static int
2512do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
2513 struct mm_struct *mm, unsigned long start,
2514 unsigned long end, struct list_head *uf, bool unlock)
2515{
2516 struct vm_area_struct *prev, *next = NULL;
2517 struct maple_tree mt_detach;
2518 int count = 0;
2519 int error = -ENOMEM;
2520 unsigned long locked_vm = 0;
2521 MA_STATE(mas_detach, &mt_detach, 0, 0);
2522 mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK);
2523 mt_on_stack(mt_detach);
2524
2525 /*
2526 * If we need to split any vma, do it now to save pain later.
2527 *
2528 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2529 * unmapped vm_area_struct will remain in use: so lower split_vma
2530 * places tmp vma above, and higher split_vma places tmp vma below.
2531 */
2532
2533 /* Does it split the first one? */
2534 if (start > vma->vm_start) {
2535
2536 /*
2537 * Make sure that map_count on return from munmap() will
2538 * not exceed its limit; but let map_count go just above
2539 * its limit temporarily, to help free resources as expected.
2540 */
2541 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2542 goto map_count_exceeded;
2543
2544 error = __split_vma(vmi, vma, start, 1);
2545 if (error)
2546 goto start_split_failed;
2547 }
2548
2549 /*
2550 * Detach a range of VMAs from the mm. Using next as a temp variable as
2551 * it is always overwritten.
2552 */
2553 next = vma;
2554 do {
2555 /* Does it split the end? */
2556 if (next->vm_end > end) {
2557 error = __split_vma(vmi, next, end, 0);
2558 if (error)
2559 goto end_split_failed;
2560 }
2561 vma_start_write(next);
2562 mas_set(&mas_detach, count);
2563 error = mas_store_gfp(&mas_detach, next, GFP_KERNEL);
2564 if (error)
2565 goto munmap_gather_failed;
2566 vma_mark_detached(next, true);
2567 if (next->vm_flags & VM_LOCKED)
2568 locked_vm += vma_pages(next);
2569
2570 count++;
2571 if (unlikely(uf)) {
2572 /*
2573 * If userfaultfd_unmap_prep returns an error the vmas
2574 * will remain split, but userland will get a
2575 * highly unexpected error anyway. This is no
2576 * different than the case where the first of the two
2577 * __split_vma fails, but we don't undo the first
2578 * split, despite we could. This is unlikely enough
2579 * failure that it's not worth optimizing it for.
2580 */
2581 error = userfaultfd_unmap_prep(next, start, end, uf);
2582
2583 if (error)
2584 goto userfaultfd_error;
2585 }
2586#ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2587 BUG_ON(next->vm_start < start);
2588 BUG_ON(next->vm_start > end);
2589#endif
2590 } for_each_vma_range(*vmi, next, end);
2591
2592#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2593 /* Make sure no VMAs are about to be lost. */
2594 {
2595 MA_STATE(test, &mt_detach, 0, 0);
2596 struct vm_area_struct *vma_mas, *vma_test;
2597 int test_count = 0;
2598
2599 vma_iter_set(vmi, start);
2600 rcu_read_lock();
2601 vma_test = mas_find(&test, count - 1);
2602 for_each_vma_range(*vmi, vma_mas, end) {
2603 BUG_ON(vma_mas != vma_test);
2604 test_count++;
2605 vma_test = mas_next(&test, count - 1);
2606 }
2607 rcu_read_unlock();
2608 BUG_ON(count != test_count);
2609 }
2610#endif
2611
2612 while (vma_iter_addr(vmi) > start)
2613 vma_iter_prev_range(vmi);
2614
2615 error = vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL);
2616 if (error)
2617 goto clear_tree_failed;
2618
2619 /* Point of no return */
2620 mm->locked_vm -= locked_vm;
2621 mm->map_count -= count;
2622 if (unlock)
2623 mmap_write_downgrade(mm);
2624
2625 prev = vma_iter_prev_range(vmi);
2626 next = vma_next(vmi);
2627 if (next)
2628 vma_iter_prev_range(vmi);
2629
2630 /*
2631 * We can free page tables without write-locking mmap_lock because VMAs
2632 * were isolated before we downgraded mmap_lock.
2633 */
2634 mas_set(&mas_detach, 1);
2635 unmap_region(mm, &mas_detach, vma, prev, next, start, end, count,
2636 !unlock);
2637 /* Statistics and freeing VMAs */
2638 mas_set(&mas_detach, 0);
2639 remove_mt(mm, &mas_detach);
2640 validate_mm(mm);
2641 if (unlock)
2642 mmap_read_unlock(mm);
2643
2644 __mt_destroy(&mt_detach);
2645 return 0;
2646
2647clear_tree_failed:
2648userfaultfd_error:
2649munmap_gather_failed:
2650end_split_failed:
2651 mas_set(&mas_detach, 0);
2652 mas_for_each(&mas_detach, next, end)
2653 vma_mark_detached(next, false);
2654
2655 __mt_destroy(&mt_detach);
2656start_split_failed:
2657map_count_exceeded:
2658 validate_mm(mm);
2659 return error;
2660}
2661
2662/*
2663 * do_vmi_munmap() - munmap a given range.
2664 * @vmi: The vma iterator
2665 * @mm: The mm_struct
2666 * @start: The start address to munmap
2667 * @len: The length of the range to munmap
2668 * @uf: The userfaultfd list_head
2669 * @unlock: set to true if the user wants to drop the mmap_lock on success
2670 *
2671 * This function takes a @mas that is either pointing to the previous VMA or set
2672 * to MA_START and sets it up to remove the mapping(s). The @len will be
2673 * aligned and any arch_unmap work will be preformed.
2674 *
2675 * Return: 0 on success and drops the lock if so directed, error and leaves the
2676 * lock held otherwise.
2677 */
2678int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm,
2679 unsigned long start, size_t len, struct list_head *uf,
2680 bool unlock)
2681{
2682 unsigned long end;
2683 struct vm_area_struct *vma;
2684
2685 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2686 return -EINVAL;
2687
2688 end = start + PAGE_ALIGN(len);
2689 if (end == start)
2690 return -EINVAL;
2691
2692 /* arch_unmap() might do unmaps itself. */
2693 arch_unmap(mm, start, end);
2694
2695 /* Find the first overlapping VMA */
2696 vma = vma_find(vmi, end);
2697 if (!vma) {
2698 if (unlock)
2699 mmap_write_unlock(mm);
2700 return 0;
2701 }
2702
2703 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
2704}
2705
2706/* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2707 * @mm: The mm_struct
2708 * @start: The start address to munmap
2709 * @len: The length to be munmapped.
2710 * @uf: The userfaultfd list_head
2711 *
2712 * Return: 0 on success, error otherwise.
2713 */
2714int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2715 struct list_head *uf)
2716{
2717 VMA_ITERATOR(vmi, mm, start);
2718
2719 return do_vmi_munmap(&vmi, mm, start, len, uf, false);
2720}
2721
2722unsigned long mmap_region(struct file *file, unsigned long addr,
2723 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2724 struct list_head *uf)
2725{
2726 struct mm_struct *mm = current->mm;
2727 struct vm_area_struct *vma = NULL;
2728 struct vm_area_struct *next, *prev, *merge;
2729 pgoff_t pglen = len >> PAGE_SHIFT;
2730 unsigned long charged = 0;
2731 unsigned long end = addr + len;
2732 unsigned long merge_start = addr, merge_end = end;
2733 bool writable_file_mapping = false;
2734 pgoff_t vm_pgoff;
2735 int error;
2736 VMA_ITERATOR(vmi, mm, addr);
2737
2738 /* Check against address space limit. */
2739 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2740 unsigned long nr_pages;
2741
2742 /*
2743 * MAP_FIXED may remove pages of mappings that intersects with
2744 * requested mapping. Account for the pages it would unmap.
2745 */
2746 nr_pages = count_vma_pages_range(mm, addr, end);
2747
2748 if (!may_expand_vm(mm, vm_flags,
2749 (len >> PAGE_SHIFT) - nr_pages))
2750 return -ENOMEM;
2751 }
2752
2753 /* Unmap any existing mapping in the area */
2754 if (do_vmi_munmap(&vmi, mm, addr, len, uf, false))
2755 return -ENOMEM;
2756
2757 /*
2758 * Private writable mapping: check memory availability
2759 */
2760 if (accountable_mapping(file, vm_flags)) {
2761 charged = len >> PAGE_SHIFT;
2762 if (security_vm_enough_memory_mm(mm, charged))
2763 return -ENOMEM;
2764 vm_flags |= VM_ACCOUNT;
2765 }
2766
2767 next = vma_next(&vmi);
2768 prev = vma_prev(&vmi);
2769 if (vm_flags & VM_SPECIAL) {
2770 if (prev)
2771 vma_iter_next_range(&vmi);
2772 goto cannot_expand;
2773 }
2774
2775 /* Attempt to expand an old mapping */
2776 /* Check next */
2777 if (next && next->vm_start == end && !vma_policy(next) &&
2778 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2779 NULL_VM_UFFD_CTX, NULL)) {
2780 merge_end = next->vm_end;
2781 vma = next;
2782 vm_pgoff = next->vm_pgoff - pglen;
2783 }
2784
2785 /* Check prev */
2786 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2787 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2788 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2789 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2790 NULL_VM_UFFD_CTX, NULL))) {
2791 merge_start = prev->vm_start;
2792 vma = prev;
2793 vm_pgoff = prev->vm_pgoff;
2794 } else if (prev) {
2795 vma_iter_next_range(&vmi);
2796 }
2797
2798 /* Actually expand, if possible */
2799 if (vma &&
2800 !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) {
2801 khugepaged_enter_vma(vma, vm_flags);
2802 goto expanded;
2803 }
2804
2805 if (vma == prev)
2806 vma_iter_set(&vmi, addr);
2807cannot_expand:
2808
2809 /*
2810 * Determine the object being mapped and call the appropriate
2811 * specific mapper. the address has already been validated, but
2812 * not unmapped, but the maps are removed from the list.
2813 */
2814 vma = vm_area_alloc(mm);
2815 if (!vma) {
2816 error = -ENOMEM;
2817 goto unacct_error;
2818 }
2819
2820 vma_iter_config(&vmi, addr, end);
2821 vma->vm_start = addr;
2822 vma->vm_end = end;
2823 vm_flags_init(vma, vm_flags);
2824 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2825 vma->vm_pgoff = pgoff;
2826
2827 if (file) {
2828 vma->vm_file = get_file(file);
2829 error = call_mmap(file, vma);
2830 if (error)
2831 goto unmap_and_free_vma;
2832
2833 if (vma_is_shared_maywrite(vma)) {
2834 error = mapping_map_writable(file->f_mapping);
2835 if (error)
2836 goto close_and_free_vma;
2837
2838 writable_file_mapping = true;
2839 }
2840
2841 /*
2842 * Expansion is handled above, merging is handled below.
2843 * Drivers should not alter the address of the VMA.
2844 */
2845 error = -EINVAL;
2846 if (WARN_ON((addr != vma->vm_start)))
2847 goto close_and_free_vma;
2848
2849 vma_iter_config(&vmi, addr, end);
2850 /*
2851 * If vm_flags changed after call_mmap(), we should try merge
2852 * vma again as we may succeed this time.
2853 */
2854 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2855 merge = vma_merge_new_vma(&vmi, prev, vma,
2856 vma->vm_start, vma->vm_end,
2857 vma->vm_pgoff);
2858 if (merge) {
2859 /*
2860 * ->mmap() can change vma->vm_file and fput
2861 * the original file. So fput the vma->vm_file
2862 * here or we would add an extra fput for file
2863 * and cause general protection fault
2864 * ultimately.
2865 */
2866 fput(vma->vm_file);
2867 vm_area_free(vma);
2868 vma = merge;
2869 /* Update vm_flags to pick up the change. */
2870 vm_flags = vma->vm_flags;
2871 goto unmap_writable;
2872 }
2873 }
2874
2875 vm_flags = vma->vm_flags;
2876 } else if (vm_flags & VM_SHARED) {
2877 error = shmem_zero_setup(vma);
2878 if (error)
2879 goto free_vma;
2880 } else {
2881 vma_set_anonymous(vma);
2882 }
2883
2884 if (map_deny_write_exec(vma, vma->vm_flags)) {
2885 error = -EACCES;
2886 goto close_and_free_vma;
2887 }
2888
2889 /* Allow architectures to sanity-check the vm_flags */
2890 error = -EINVAL;
2891 if (!arch_validate_flags(vma->vm_flags))
2892 goto close_and_free_vma;
2893
2894 error = -ENOMEM;
2895 if (vma_iter_prealloc(&vmi, vma))
2896 goto close_and_free_vma;
2897
2898 /* Lock the VMA since it is modified after insertion into VMA tree */
2899 vma_start_write(vma);
2900 vma_iter_store(&vmi, vma);
2901 mm->map_count++;
2902 if (vma->vm_file) {
2903 i_mmap_lock_write(vma->vm_file->f_mapping);
2904 if (vma_is_shared_maywrite(vma))
2905 mapping_allow_writable(vma->vm_file->f_mapping);
2906
2907 flush_dcache_mmap_lock(vma->vm_file->f_mapping);
2908 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
2909 flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
2910 i_mmap_unlock_write(vma->vm_file->f_mapping);
2911 }
2912
2913 /*
2914 * vma_merge() calls khugepaged_enter_vma() either, the below
2915 * call covers the non-merge case.
2916 */
2917 khugepaged_enter_vma(vma, vma->vm_flags);
2918
2919 /* Once vma denies write, undo our temporary denial count */
2920unmap_writable:
2921 if (writable_file_mapping)
2922 mapping_unmap_writable(file->f_mapping);
2923 file = vma->vm_file;
2924 ksm_add_vma(vma);
2925expanded:
2926 perf_event_mmap(vma);
2927
2928 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2929 if (vm_flags & VM_LOCKED) {
2930 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2931 is_vm_hugetlb_page(vma) ||
2932 vma == get_gate_vma(current->mm))
2933 vm_flags_clear(vma, VM_LOCKED_MASK);
2934 else
2935 mm->locked_vm += (len >> PAGE_SHIFT);
2936 }
2937
2938 if (file)
2939 uprobe_mmap(vma);
2940
2941 /*
2942 * New (or expanded) vma always get soft dirty status.
2943 * Otherwise user-space soft-dirty page tracker won't
2944 * be able to distinguish situation when vma area unmapped,
2945 * then new mapped in-place (which must be aimed as
2946 * a completely new data area).
2947 */
2948 vm_flags_set(vma, VM_SOFTDIRTY);
2949
2950 vma_set_page_prot(vma);
2951
2952 validate_mm(mm);
2953 return addr;
2954
2955close_and_free_vma:
2956 if (file && vma->vm_ops && vma->vm_ops->close)
2957 vma->vm_ops->close(vma);
2958
2959 if (file || vma->vm_file) {
2960unmap_and_free_vma:
2961 fput(vma->vm_file);
2962 vma->vm_file = NULL;
2963
2964 vma_iter_set(&vmi, vma->vm_end);
2965 /* Undo any partial mapping done by a device driver. */
2966 unmap_region(mm, &vmi.mas, vma, prev, next, vma->vm_start,
2967 vma->vm_end, vma->vm_end, true);
2968 }
2969 if (writable_file_mapping)
2970 mapping_unmap_writable(file->f_mapping);
2971free_vma:
2972 vm_area_free(vma);
2973unacct_error:
2974 if (charged)
2975 vm_unacct_memory(charged);
2976 validate_mm(mm);
2977 return error;
2978}
2979
2980static int __vm_munmap(unsigned long start, size_t len, bool unlock)
2981{
2982 int ret;
2983 struct mm_struct *mm = current->mm;
2984 LIST_HEAD(uf);
2985 VMA_ITERATOR(vmi, mm, start);
2986
2987 if (mmap_write_lock_killable(mm))
2988 return -EINTR;
2989
2990 ret = do_vmi_munmap(&vmi, mm, start, len, &uf, unlock);
2991 if (ret || !unlock)
2992 mmap_write_unlock(mm);
2993
2994 userfaultfd_unmap_complete(mm, &uf);
2995 return ret;
2996}
2997
2998int vm_munmap(unsigned long start, size_t len)
2999{
3000 return __vm_munmap(start, len, false);
3001}
3002EXPORT_SYMBOL(vm_munmap);
3003
3004SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
3005{
3006 addr = untagged_addr(addr);
3007 return __vm_munmap(addr, len, true);
3008}
3009
3010
3011/*
3012 * Emulation of deprecated remap_file_pages() syscall.
3013 */
3014SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
3015 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
3016{
3017
3018 struct mm_struct *mm = current->mm;
3019 struct vm_area_struct *vma;
3020 unsigned long populate = 0;
3021 unsigned long ret = -EINVAL;
3022 struct file *file;
3023
3024 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
3025 current->comm, current->pid);
3026
3027 if (prot)
3028 return ret;
3029 start = start & PAGE_MASK;
3030 size = size & PAGE_MASK;
3031
3032 if (start + size <= start)
3033 return ret;
3034
3035 /* Does pgoff wrap? */
3036 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
3037 return ret;
3038
3039 if (mmap_write_lock_killable(mm))
3040 return -EINTR;
3041
3042 vma = vma_lookup(mm, start);
3043
3044 if (!vma || !(vma->vm_flags & VM_SHARED))
3045 goto out;
3046
3047 if (start + size > vma->vm_end) {
3048 VMA_ITERATOR(vmi, mm, vma->vm_end);
3049 struct vm_area_struct *next, *prev = vma;
3050
3051 for_each_vma_range(vmi, next, start + size) {
3052 /* hole between vmas ? */
3053 if (next->vm_start != prev->vm_end)
3054 goto out;
3055
3056 if (next->vm_file != vma->vm_file)
3057 goto out;
3058
3059 if (next->vm_flags != vma->vm_flags)
3060 goto out;
3061
3062 if (start + size <= next->vm_end)
3063 break;
3064
3065 prev = next;
3066 }
3067
3068 if (!next)
3069 goto out;
3070 }
3071
3072 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3073 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3074 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3075
3076 flags &= MAP_NONBLOCK;
3077 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3078 if (vma->vm_flags & VM_LOCKED)
3079 flags |= MAP_LOCKED;
3080
3081 file = get_file(vma->vm_file);
3082 ret = do_mmap(vma->vm_file, start, size,
3083 prot, flags, 0, pgoff, &populate, NULL);
3084 fput(file);
3085out:
3086 mmap_write_unlock(mm);
3087 if (populate)
3088 mm_populate(ret, populate);
3089 if (!IS_ERR_VALUE(ret))
3090 ret = 0;
3091 return ret;
3092}
3093
3094/*
3095 * do_vma_munmap() - Unmap a full or partial vma.
3096 * @vmi: The vma iterator pointing at the vma
3097 * @vma: The first vma to be munmapped
3098 * @start: the start of the address to unmap
3099 * @end: The end of the address to unmap
3100 * @uf: The userfaultfd list_head
3101 * @unlock: Drop the lock on success
3102 *
3103 * unmaps a VMA mapping when the vma iterator is already in position.
3104 * Does not handle alignment.
3105 *
3106 * Return: 0 on success drops the lock of so directed, error on failure and will
3107 * still hold the lock.
3108 */
3109int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
3110 unsigned long start, unsigned long end, struct list_head *uf,
3111 bool unlock)
3112{
3113 struct mm_struct *mm = vma->vm_mm;
3114
3115 arch_unmap(mm, start, end);
3116 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
3117}
3118
3119/*
3120 * do_brk_flags() - Increase the brk vma if the flags match.
3121 * @vmi: The vma iterator
3122 * @addr: The start address
3123 * @len: The length of the increase
3124 * @vma: The vma,
3125 * @flags: The VMA Flags
3126 *
3127 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
3128 * do not match then create a new anonymous VMA. Eventually we may be able to
3129 * do some brk-specific accounting here.
3130 */
3131static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
3132 unsigned long addr, unsigned long len, unsigned long flags)
3133{
3134 struct mm_struct *mm = current->mm;
3135 struct vma_prepare vp;
3136
3137 /*
3138 * Check against address space limits by the changed size
3139 * Note: This happens *after* clearing old mappings in some code paths.
3140 */
3141 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3142 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3143 return -ENOMEM;
3144
3145 if (mm->map_count > sysctl_max_map_count)
3146 return -ENOMEM;
3147
3148 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3149 return -ENOMEM;
3150
3151 /*
3152 * Expand the existing vma if possible; Note that singular lists do not
3153 * occur after forking, so the expand will only happen on new VMAs.
3154 */
3155 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
3156 can_vma_merge_after(vma, flags, NULL, NULL,
3157 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
3158 vma_iter_config(vmi, vma->vm_start, addr + len);
3159 if (vma_iter_prealloc(vmi, vma))
3160 goto unacct_fail;
3161
3162 vma_start_write(vma);
3163
3164 init_vma_prep(&vp, vma);
3165 vma_prepare(&vp);
3166 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
3167 vma->vm_end = addr + len;
3168 vm_flags_set(vma, VM_SOFTDIRTY);
3169 vma_iter_store(vmi, vma);
3170
3171 vma_complete(&vp, vmi, mm);
3172 khugepaged_enter_vma(vma, flags);
3173 goto out;
3174 }
3175
3176 if (vma)
3177 vma_iter_next_range(vmi);
3178 /* create a vma struct for an anonymous mapping */
3179 vma = vm_area_alloc(mm);
3180 if (!vma)
3181 goto unacct_fail;
3182
3183 vma_set_anonymous(vma);
3184 vma->vm_start = addr;
3185 vma->vm_end = addr + len;
3186 vma->vm_pgoff = addr >> PAGE_SHIFT;
3187 vm_flags_init(vma, flags);
3188 vma->vm_page_prot = vm_get_page_prot(flags);
3189 vma_start_write(vma);
3190 if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
3191 goto mas_store_fail;
3192
3193 mm->map_count++;
3194 validate_mm(mm);
3195 ksm_add_vma(vma);
3196out:
3197 perf_event_mmap(vma);
3198 mm->total_vm += len >> PAGE_SHIFT;
3199 mm->data_vm += len >> PAGE_SHIFT;
3200 if (flags & VM_LOCKED)
3201 mm->locked_vm += (len >> PAGE_SHIFT);
3202 vm_flags_set(vma, VM_SOFTDIRTY);
3203 return 0;
3204
3205mas_store_fail:
3206 vm_area_free(vma);
3207unacct_fail:
3208 vm_unacct_memory(len >> PAGE_SHIFT);
3209 return -ENOMEM;
3210}
3211
3212int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3213{
3214 struct mm_struct *mm = current->mm;
3215 struct vm_area_struct *vma = NULL;
3216 unsigned long len;
3217 int ret;
3218 bool populate;
3219 LIST_HEAD(uf);
3220 VMA_ITERATOR(vmi, mm, addr);
3221
3222 len = PAGE_ALIGN(request);
3223 if (len < request)
3224 return -ENOMEM;
3225 if (!len)
3226 return 0;
3227
3228 /* Until we need other flags, refuse anything except VM_EXEC. */
3229 if ((flags & (~VM_EXEC)) != 0)
3230 return -EINVAL;
3231
3232 if (mmap_write_lock_killable(mm))
3233 return -EINTR;
3234
3235 ret = check_brk_limits(addr, len);
3236 if (ret)
3237 goto limits_failed;
3238
3239 ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0);
3240 if (ret)
3241 goto munmap_failed;
3242
3243 vma = vma_prev(&vmi);
3244 ret = do_brk_flags(&vmi, vma, addr, len, flags);
3245 populate = ((mm->def_flags & VM_LOCKED) != 0);
3246 mmap_write_unlock(mm);
3247 userfaultfd_unmap_complete(mm, &uf);
3248 if (populate && !ret)
3249 mm_populate(addr, len);
3250 return ret;
3251
3252munmap_failed:
3253limits_failed:
3254 mmap_write_unlock(mm);
3255 return ret;
3256}
3257EXPORT_SYMBOL(vm_brk_flags);
3258
3259/* Release all mmaps. */
3260void exit_mmap(struct mm_struct *mm)
3261{
3262 struct mmu_gather tlb;
3263 struct vm_area_struct *vma;
3264 unsigned long nr_accounted = 0;
3265 MA_STATE(mas, &mm->mm_mt, 0, 0);
3266 int count = 0;
3267
3268 /* mm's last user has gone, and its about to be pulled down */
3269 mmu_notifier_release(mm);
3270
3271 mmap_read_lock(mm);
3272 arch_exit_mmap(mm);
3273
3274 vma = mas_find(&mas, ULONG_MAX);
3275 if (!vma || unlikely(xa_is_zero(vma))) {
3276 /* Can happen if dup_mmap() received an OOM */
3277 mmap_read_unlock(mm);
3278 mmap_write_lock(mm);
3279 goto destroy;
3280 }
3281
3282 lru_add_drain();
3283 flush_cache_mm(mm);
3284 tlb_gather_mmu_fullmm(&tlb, mm);
3285 /* update_hiwater_rss(mm) here? but nobody should be looking */
3286 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3287 unmap_vmas(&tlb, &mas, vma, 0, ULONG_MAX, ULONG_MAX, false);
3288 mmap_read_unlock(mm);
3289
3290 /*
3291 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3292 * because the memory has been already freed.
3293 */
3294 set_bit(MMF_OOM_SKIP, &mm->flags);
3295 mmap_write_lock(mm);
3296 mt_clear_in_rcu(&mm->mm_mt);
3297 mas_set(&mas, vma->vm_end);
3298 free_pgtables(&tlb, &mas, vma, FIRST_USER_ADDRESS,
3299 USER_PGTABLES_CEILING, true);
3300 tlb_finish_mmu(&tlb);
3301
3302 /*
3303 * Walk the list again, actually closing and freeing it, with preemption
3304 * enabled, without holding any MM locks besides the unreachable
3305 * mmap_write_lock.
3306 */
3307 mas_set(&mas, vma->vm_end);
3308 do {
3309 if (vma->vm_flags & VM_ACCOUNT)
3310 nr_accounted += vma_pages(vma);
3311 remove_vma(vma, true);
3312 count++;
3313 cond_resched();
3314 vma = mas_find(&mas, ULONG_MAX);
3315 } while (vma && likely(!xa_is_zero(vma)));
3316
3317 BUG_ON(count != mm->map_count);
3318
3319 trace_exit_mmap(mm);
3320destroy:
3321 __mt_destroy(&mm->mm_mt);
3322 mmap_write_unlock(mm);
3323 vm_unacct_memory(nr_accounted);
3324}
3325
3326/* Insert vm structure into process list sorted by address
3327 * and into the inode's i_mmap tree. If vm_file is non-NULL
3328 * then i_mmap_rwsem is taken here.
3329 */
3330int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3331{
3332 unsigned long charged = vma_pages(vma);
3333
3334
3335 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3336 return -ENOMEM;
3337
3338 if ((vma->vm_flags & VM_ACCOUNT) &&
3339 security_vm_enough_memory_mm(mm, charged))
3340 return -ENOMEM;
3341
3342 /*
3343 * The vm_pgoff of a purely anonymous vma should be irrelevant
3344 * until its first write fault, when page's anon_vma and index
3345 * are set. But now set the vm_pgoff it will almost certainly
3346 * end up with (unless mremap moves it elsewhere before that
3347 * first wfault), so /proc/pid/maps tells a consistent story.
3348 *
3349 * By setting it to reflect the virtual start address of the
3350 * vma, merges and splits can happen in a seamless way, just
3351 * using the existing file pgoff checks and manipulations.
3352 * Similarly in do_mmap and in do_brk_flags.
3353 */
3354 if (vma_is_anonymous(vma)) {
3355 BUG_ON(vma->anon_vma);
3356 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3357 }
3358
3359 if (vma_link(mm, vma)) {
3360 if (vma->vm_flags & VM_ACCOUNT)
3361 vm_unacct_memory(charged);
3362 return -ENOMEM;
3363 }
3364
3365 return 0;
3366}
3367
3368/*
3369 * Copy the vma structure to a new location in the same mm,
3370 * prior to moving page table entries, to effect an mremap move.
3371 */
3372struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3373 unsigned long addr, unsigned long len, pgoff_t pgoff,
3374 bool *need_rmap_locks)
3375{
3376 struct vm_area_struct *vma = *vmap;
3377 unsigned long vma_start = vma->vm_start;
3378 struct mm_struct *mm = vma->vm_mm;
3379 struct vm_area_struct *new_vma, *prev;
3380 bool faulted_in_anon_vma = true;
3381 VMA_ITERATOR(vmi, mm, addr);
3382
3383 /*
3384 * If anonymous vma has not yet been faulted, update new pgoff
3385 * to match new location, to increase its chance of merging.
3386 */
3387 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3388 pgoff = addr >> PAGE_SHIFT;
3389 faulted_in_anon_vma = false;
3390 }
3391
3392 new_vma = find_vma_prev(mm, addr, &prev);
3393 if (new_vma && new_vma->vm_start < addr + len)
3394 return NULL; /* should never get here */
3395
3396 new_vma = vma_merge_new_vma(&vmi, prev, vma, addr, addr + len, pgoff);
3397 if (new_vma) {
3398 /*
3399 * Source vma may have been merged into new_vma
3400 */
3401 if (unlikely(vma_start >= new_vma->vm_start &&
3402 vma_start < new_vma->vm_end)) {
3403 /*
3404 * The only way we can get a vma_merge with
3405 * self during an mremap is if the vma hasn't
3406 * been faulted in yet and we were allowed to
3407 * reset the dst vma->vm_pgoff to the
3408 * destination address of the mremap to allow
3409 * the merge to happen. mremap must change the
3410 * vm_pgoff linearity between src and dst vmas
3411 * (in turn preventing a vma_merge) to be
3412 * safe. It is only safe to keep the vm_pgoff
3413 * linear if there are no pages mapped yet.
3414 */
3415 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3416 *vmap = vma = new_vma;
3417 }
3418 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3419 } else {
3420 new_vma = vm_area_dup(vma);
3421 if (!new_vma)
3422 goto out;
3423 new_vma->vm_start = addr;
3424 new_vma->vm_end = addr + len;
3425 new_vma->vm_pgoff = pgoff;
3426 if (vma_dup_policy(vma, new_vma))
3427 goto out_free_vma;
3428 if (anon_vma_clone(new_vma, vma))
3429 goto out_free_mempol;
3430 if (new_vma->vm_file)
3431 get_file(new_vma->vm_file);
3432 if (new_vma->vm_ops && new_vma->vm_ops->open)
3433 new_vma->vm_ops->open(new_vma);
3434 if (vma_link(mm, new_vma))
3435 goto out_vma_link;
3436 *need_rmap_locks = false;
3437 }
3438 return new_vma;
3439
3440out_vma_link:
3441 if (new_vma->vm_ops && new_vma->vm_ops->close)
3442 new_vma->vm_ops->close(new_vma);
3443
3444 if (new_vma->vm_file)
3445 fput(new_vma->vm_file);
3446
3447 unlink_anon_vmas(new_vma);
3448out_free_mempol:
3449 mpol_put(vma_policy(new_vma));
3450out_free_vma:
3451 vm_area_free(new_vma);
3452out:
3453 return NULL;
3454}
3455
3456/*
3457 * Return true if the calling process may expand its vm space by the passed
3458 * number of pages
3459 */
3460bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3461{
3462 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3463 return false;
3464
3465 if (is_data_mapping(flags) &&
3466 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3467 /* Workaround for Valgrind */
3468 if (rlimit(RLIMIT_DATA) == 0 &&
3469 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3470 return true;
3471
3472 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3473 current->comm, current->pid,
3474 (mm->data_vm + npages) << PAGE_SHIFT,
3475 rlimit(RLIMIT_DATA),
3476 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3477
3478 if (!ignore_rlimit_data)
3479 return false;
3480 }
3481
3482 return true;
3483}
3484
3485void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3486{
3487 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3488
3489 if (is_exec_mapping(flags))
3490 mm->exec_vm += npages;
3491 else if (is_stack_mapping(flags))
3492 mm->stack_vm += npages;
3493 else if (is_data_mapping(flags))
3494 mm->data_vm += npages;
3495}
3496
3497static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3498
3499/*
3500 * Having a close hook prevents vma merging regardless of flags.
3501 */
3502static void special_mapping_close(struct vm_area_struct *vma)
3503{
3504}
3505
3506static const char *special_mapping_name(struct vm_area_struct *vma)
3507{
3508 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3509}
3510
3511static int special_mapping_mremap(struct vm_area_struct *new_vma)
3512{
3513 struct vm_special_mapping *sm = new_vma->vm_private_data;
3514
3515 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3516 return -EFAULT;
3517
3518 if (sm->mremap)
3519 return sm->mremap(sm, new_vma);
3520
3521 return 0;
3522}
3523
3524static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3525{
3526 /*
3527 * Forbid splitting special mappings - kernel has expectations over
3528 * the number of pages in mapping. Together with VM_DONTEXPAND
3529 * the size of vma should stay the same over the special mapping's
3530 * lifetime.
3531 */
3532 return -EINVAL;
3533}
3534
3535static const struct vm_operations_struct special_mapping_vmops = {
3536 .close = special_mapping_close,
3537 .fault = special_mapping_fault,
3538 .mremap = special_mapping_mremap,
3539 .name = special_mapping_name,
3540 /* vDSO code relies that VVAR can't be accessed remotely */
3541 .access = NULL,
3542 .may_split = special_mapping_split,
3543};
3544
3545static const struct vm_operations_struct legacy_special_mapping_vmops = {
3546 .close = special_mapping_close,
3547 .fault = special_mapping_fault,
3548};
3549
3550static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3551{
3552 struct vm_area_struct *vma = vmf->vma;
3553 pgoff_t pgoff;
3554 struct page **pages;
3555
3556 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3557 pages = vma->vm_private_data;
3558 } else {
3559 struct vm_special_mapping *sm = vma->vm_private_data;
3560
3561 if (sm->fault)
3562 return sm->fault(sm, vmf->vma, vmf);
3563
3564 pages = sm->pages;
3565 }
3566
3567 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3568 pgoff--;
3569
3570 if (*pages) {
3571 struct page *page = *pages;
3572 get_page(page);
3573 vmf->page = page;
3574 return 0;
3575 }
3576
3577 return VM_FAULT_SIGBUS;
3578}
3579
3580static struct vm_area_struct *__install_special_mapping(
3581 struct mm_struct *mm,
3582 unsigned long addr, unsigned long len,
3583 unsigned long vm_flags, void *priv,
3584 const struct vm_operations_struct *ops)
3585{
3586 int ret;
3587 struct vm_area_struct *vma;
3588
3589 vma = vm_area_alloc(mm);
3590 if (unlikely(vma == NULL))
3591 return ERR_PTR(-ENOMEM);
3592
3593 vma->vm_start = addr;
3594 vma->vm_end = addr + len;
3595
3596 vm_flags_init(vma, (vm_flags | mm->def_flags |
3597 VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK);
3598 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3599
3600 vma->vm_ops = ops;
3601 vma->vm_private_data = priv;
3602
3603 ret = insert_vm_struct(mm, vma);
3604 if (ret)
3605 goto out;
3606
3607 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3608
3609 perf_event_mmap(vma);
3610
3611 return vma;
3612
3613out:
3614 vm_area_free(vma);
3615 return ERR_PTR(ret);
3616}
3617
3618bool vma_is_special_mapping(const struct vm_area_struct *vma,
3619 const struct vm_special_mapping *sm)
3620{
3621 return vma->vm_private_data == sm &&
3622 (vma->vm_ops == &special_mapping_vmops ||
3623 vma->vm_ops == &legacy_special_mapping_vmops);
3624}
3625
3626/*
3627 * Called with mm->mmap_lock held for writing.
3628 * Insert a new vma covering the given region, with the given flags.
3629 * Its pages are supplied by the given array of struct page *.
3630 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3631 * The region past the last page supplied will always produce SIGBUS.
3632 * The array pointer and the pages it points to are assumed to stay alive
3633 * for as long as this mapping might exist.
3634 */
3635struct vm_area_struct *_install_special_mapping(
3636 struct mm_struct *mm,
3637 unsigned long addr, unsigned long len,
3638 unsigned long vm_flags, const struct vm_special_mapping *spec)
3639{
3640 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3641 &special_mapping_vmops);
3642}
3643
3644int install_special_mapping(struct mm_struct *mm,
3645 unsigned long addr, unsigned long len,
3646 unsigned long vm_flags, struct page **pages)
3647{
3648 struct vm_area_struct *vma = __install_special_mapping(
3649 mm, addr, len, vm_flags, (void *)pages,
3650 &legacy_special_mapping_vmops);
3651
3652 return PTR_ERR_OR_ZERO(vma);
3653}
3654
3655static DEFINE_MUTEX(mm_all_locks_mutex);
3656
3657static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3658{
3659 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3660 /*
3661 * The LSB of head.next can't change from under us
3662 * because we hold the mm_all_locks_mutex.
3663 */
3664 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3665 /*
3666 * We can safely modify head.next after taking the
3667 * anon_vma->root->rwsem. If some other vma in this mm shares
3668 * the same anon_vma we won't take it again.
3669 *
3670 * No need of atomic instructions here, head.next
3671 * can't change from under us thanks to the
3672 * anon_vma->root->rwsem.
3673 */
3674 if (__test_and_set_bit(0, (unsigned long *)
3675 &anon_vma->root->rb_root.rb_root.rb_node))
3676 BUG();
3677 }
3678}
3679
3680static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3681{
3682 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3683 /*
3684 * AS_MM_ALL_LOCKS can't change from under us because
3685 * we hold the mm_all_locks_mutex.
3686 *
3687 * Operations on ->flags have to be atomic because
3688 * even if AS_MM_ALL_LOCKS is stable thanks to the
3689 * mm_all_locks_mutex, there may be other cpus
3690 * changing other bitflags in parallel to us.
3691 */
3692 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3693 BUG();
3694 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3695 }
3696}
3697
3698/*
3699 * This operation locks against the VM for all pte/vma/mm related
3700 * operations that could ever happen on a certain mm. This includes
3701 * vmtruncate, try_to_unmap, and all page faults.
3702 *
3703 * The caller must take the mmap_lock in write mode before calling
3704 * mm_take_all_locks(). The caller isn't allowed to release the
3705 * mmap_lock until mm_drop_all_locks() returns.
3706 *
3707 * mmap_lock in write mode is required in order to block all operations
3708 * that could modify pagetables and free pages without need of
3709 * altering the vma layout. It's also needed in write mode to avoid new
3710 * anon_vmas to be associated with existing vmas.
3711 *
3712 * A single task can't take more than one mm_take_all_locks() in a row
3713 * or it would deadlock.
3714 *
3715 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3716 * mapping->flags avoid to take the same lock twice, if more than one
3717 * vma in this mm is backed by the same anon_vma or address_space.
3718 *
3719 * We take locks in following order, accordingly to comment at beginning
3720 * of mm/rmap.c:
3721 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3722 * hugetlb mapping);
3723 * - all vmas marked locked
3724 * - all i_mmap_rwsem locks;
3725 * - all anon_vma->rwseml
3726 *
3727 * We can take all locks within these types randomly because the VM code
3728 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3729 * mm_all_locks_mutex.
3730 *
3731 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3732 * that may have to take thousand of locks.
3733 *
3734 * mm_take_all_locks() can fail if it's interrupted by signals.
3735 */
3736int mm_take_all_locks(struct mm_struct *mm)
3737{
3738 struct vm_area_struct *vma;
3739 struct anon_vma_chain *avc;
3740 MA_STATE(mas, &mm->mm_mt, 0, 0);
3741
3742 mmap_assert_write_locked(mm);
3743
3744 mutex_lock(&mm_all_locks_mutex);
3745
3746 /*
3747 * vma_start_write() does not have a complement in mm_drop_all_locks()
3748 * because vma_start_write() is always asymmetrical; it marks a VMA as
3749 * being written to until mmap_write_unlock() or mmap_write_downgrade()
3750 * is reached.
3751 */
3752 mas_for_each(&mas, vma, ULONG_MAX) {
3753 if (signal_pending(current))
3754 goto out_unlock;
3755 vma_start_write(vma);
3756 }
3757
3758 mas_set(&mas, 0);
3759 mas_for_each(&mas, vma, ULONG_MAX) {
3760 if (signal_pending(current))
3761 goto out_unlock;
3762 if (vma->vm_file && vma->vm_file->f_mapping &&
3763 is_vm_hugetlb_page(vma))
3764 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3765 }
3766
3767 mas_set(&mas, 0);
3768 mas_for_each(&mas, vma, ULONG_MAX) {
3769 if (signal_pending(current))
3770 goto out_unlock;
3771 if (vma->vm_file && vma->vm_file->f_mapping &&
3772 !is_vm_hugetlb_page(vma))
3773 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3774 }
3775
3776 mas_set(&mas, 0);
3777 mas_for_each(&mas, vma, ULONG_MAX) {
3778 if (signal_pending(current))
3779 goto out_unlock;
3780 if (vma->anon_vma)
3781 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3782 vm_lock_anon_vma(mm, avc->anon_vma);
3783 }
3784
3785 return 0;
3786
3787out_unlock:
3788 mm_drop_all_locks(mm);
3789 return -EINTR;
3790}
3791
3792static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3793{
3794 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3795 /*
3796 * The LSB of head.next can't change to 0 from under
3797 * us because we hold the mm_all_locks_mutex.
3798 *
3799 * We must however clear the bitflag before unlocking
3800 * the vma so the users using the anon_vma->rb_root will
3801 * never see our bitflag.
3802 *
3803 * No need of atomic instructions here, head.next
3804 * can't change from under us until we release the
3805 * anon_vma->root->rwsem.
3806 */
3807 if (!__test_and_clear_bit(0, (unsigned long *)
3808 &anon_vma->root->rb_root.rb_root.rb_node))
3809 BUG();
3810 anon_vma_unlock_write(anon_vma);
3811 }
3812}
3813
3814static void vm_unlock_mapping(struct address_space *mapping)
3815{
3816 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3817 /*
3818 * AS_MM_ALL_LOCKS can't change to 0 from under us
3819 * because we hold the mm_all_locks_mutex.
3820 */
3821 i_mmap_unlock_write(mapping);
3822 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3823 &mapping->flags))
3824 BUG();
3825 }
3826}
3827
3828/*
3829 * The mmap_lock cannot be released by the caller until
3830 * mm_drop_all_locks() returns.
3831 */
3832void mm_drop_all_locks(struct mm_struct *mm)
3833{
3834 struct vm_area_struct *vma;
3835 struct anon_vma_chain *avc;
3836 MA_STATE(mas, &mm->mm_mt, 0, 0);
3837
3838 mmap_assert_write_locked(mm);
3839 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3840
3841 mas_for_each(&mas, vma, ULONG_MAX) {
3842 if (vma->anon_vma)
3843 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3844 vm_unlock_anon_vma(avc->anon_vma);
3845 if (vma->vm_file && vma->vm_file->f_mapping)
3846 vm_unlock_mapping(vma->vm_file->f_mapping);
3847 }
3848
3849 mutex_unlock(&mm_all_locks_mutex);
3850}
3851
3852/*
3853 * initialise the percpu counter for VM
3854 */
3855void __init mmap_init(void)
3856{
3857 int ret;
3858
3859 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3860 VM_BUG_ON(ret);
3861}
3862
3863/*
3864 * Initialise sysctl_user_reserve_kbytes.
3865 *
3866 * This is intended to prevent a user from starting a single memory hogging
3867 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3868 * mode.
3869 *
3870 * The default value is min(3% of free memory, 128MB)
3871 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3872 */
3873static int init_user_reserve(void)
3874{
3875 unsigned long free_kbytes;
3876
3877 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3878
3879 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3880 return 0;
3881}
3882subsys_initcall(init_user_reserve);
3883
3884/*
3885 * Initialise sysctl_admin_reserve_kbytes.
3886 *
3887 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3888 * to log in and kill a memory hogging process.
3889 *
3890 * Systems with more than 256MB will reserve 8MB, enough to recover
3891 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3892 * only reserve 3% of free pages by default.
3893 */
3894static int init_admin_reserve(void)
3895{
3896 unsigned long free_kbytes;
3897
3898 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3899
3900 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3901 return 0;
3902}
3903subsys_initcall(init_admin_reserve);
3904
3905/*
3906 * Reinititalise user and admin reserves if memory is added or removed.
3907 *
3908 * The default user reserve max is 128MB, and the default max for the
3909 * admin reserve is 8MB. These are usually, but not always, enough to
3910 * enable recovery from a memory hogging process using login/sshd, a shell,
3911 * and tools like top. It may make sense to increase or even disable the
3912 * reserve depending on the existence of swap or variations in the recovery
3913 * tools. So, the admin may have changed them.
3914 *
3915 * If memory is added and the reserves have been eliminated or increased above
3916 * the default max, then we'll trust the admin.
3917 *
3918 * If memory is removed and there isn't enough free memory, then we
3919 * need to reset the reserves.
3920 *
3921 * Otherwise keep the reserve set by the admin.
3922 */
3923static int reserve_mem_notifier(struct notifier_block *nb,
3924 unsigned long action, void *data)
3925{
3926 unsigned long tmp, free_kbytes;
3927
3928 switch (action) {
3929 case MEM_ONLINE:
3930 /* Default max is 128MB. Leave alone if modified by operator. */
3931 tmp = sysctl_user_reserve_kbytes;
3932 if (0 < tmp && tmp < (1UL << 17))
3933 init_user_reserve();
3934
3935 /* Default max is 8MB. Leave alone if modified by operator. */
3936 tmp = sysctl_admin_reserve_kbytes;
3937 if (0 < tmp && tmp < (1UL << 13))
3938 init_admin_reserve();
3939
3940 break;
3941 case MEM_OFFLINE:
3942 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3943
3944 if (sysctl_user_reserve_kbytes > free_kbytes) {
3945 init_user_reserve();
3946 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3947 sysctl_user_reserve_kbytes);
3948 }
3949
3950 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3951 init_admin_reserve();
3952 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3953 sysctl_admin_reserve_kbytes);
3954 }
3955 break;
3956 default:
3957 break;
3958 }
3959 return NOTIFY_OK;
3960}
3961
3962static int __meminit init_reserve_notifier(void)
3963{
3964 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3965 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3966
3967 return 0;
3968}
3969subsys_initcall(init_reserve_notifier);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/mmap.c
4 *
5 * Written by obz.
6 *
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
8 */
9
10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12#include <linux/kernel.h>
13#include <linux/slab.h>
14#include <linux/backing-dev.h>
15#include <linux/mm.h>
16#include <linux/mm_inline.h>
17#include <linux/shm.h>
18#include <linux/mman.h>
19#include <linux/pagemap.h>
20#include <linux/swap.h>
21#include <linux/syscalls.h>
22#include <linux/capability.h>
23#include <linux/init.h>
24#include <linux/file.h>
25#include <linux/fs.h>
26#include <linux/personality.h>
27#include <linux/security.h>
28#include <linux/hugetlb.h>
29#include <linux/shmem_fs.h>
30#include <linux/profile.h>
31#include <linux/export.h>
32#include <linux/mount.h>
33#include <linux/mempolicy.h>
34#include <linux/rmap.h>
35#include <linux/mmu_notifier.h>
36#include <linux/mmdebug.h>
37#include <linux/perf_event.h>
38#include <linux/audit.h>
39#include <linux/khugepaged.h>
40#include <linux/uprobes.h>
41#include <linux/notifier.h>
42#include <linux/memory.h>
43#include <linux/printk.h>
44#include <linux/userfaultfd_k.h>
45#include <linux/moduleparam.h>
46#include <linux/pkeys.h>
47#include <linux/oom.h>
48#include <linux/sched/mm.h>
49
50#include <linux/uaccess.h>
51#include <asm/cacheflush.h>
52#include <asm/tlb.h>
53#include <asm/mmu_context.h>
54
55#define CREATE_TRACE_POINTS
56#include <trace/events/mmap.h>
57
58#include "internal.h"
59
60#ifndef arch_mmap_check
61#define arch_mmap_check(addr, len, flags) (0)
62#endif
63
64#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
65const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
66const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
67int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
68#endif
69#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
70const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
71const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
72int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
73#endif
74
75static bool ignore_rlimit_data;
76core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
77
78static void unmap_region(struct mm_struct *mm, struct maple_tree *mt,
79 struct vm_area_struct *vma, struct vm_area_struct *prev,
80 struct vm_area_struct *next, unsigned long start,
81 unsigned long end);
82
83static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
84{
85 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
86}
87
88/* Update vma->vm_page_prot to reflect vma->vm_flags. */
89void vma_set_page_prot(struct vm_area_struct *vma)
90{
91 unsigned long vm_flags = vma->vm_flags;
92 pgprot_t vm_page_prot;
93
94 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
95 if (vma_wants_writenotify(vma, vm_page_prot)) {
96 vm_flags &= ~VM_SHARED;
97 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
98 }
99 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
100 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
101}
102
103/*
104 * Requires inode->i_mapping->i_mmap_rwsem
105 */
106static void __remove_shared_vm_struct(struct vm_area_struct *vma,
107 struct file *file, struct address_space *mapping)
108{
109 if (vma->vm_flags & VM_SHARED)
110 mapping_unmap_writable(mapping);
111
112 flush_dcache_mmap_lock(mapping);
113 vma_interval_tree_remove(vma, &mapping->i_mmap);
114 flush_dcache_mmap_unlock(mapping);
115}
116
117/*
118 * Unlink a file-based vm structure from its interval tree, to hide
119 * vma from rmap and vmtruncate before freeing its page tables.
120 */
121void unlink_file_vma(struct vm_area_struct *vma)
122{
123 struct file *file = vma->vm_file;
124
125 if (file) {
126 struct address_space *mapping = file->f_mapping;
127 i_mmap_lock_write(mapping);
128 __remove_shared_vm_struct(vma, file, mapping);
129 i_mmap_unlock_write(mapping);
130 }
131}
132
133/*
134 * Close a vm structure and free it.
135 */
136static void remove_vma(struct vm_area_struct *vma)
137{
138 might_sleep();
139 if (vma->vm_ops && vma->vm_ops->close)
140 vma->vm_ops->close(vma);
141 if (vma->vm_file)
142 fput(vma->vm_file);
143 mpol_put(vma_policy(vma));
144 vm_area_free(vma);
145}
146
147/*
148 * check_brk_limits() - Use platform specific check of range & verify mlock
149 * limits.
150 * @addr: The address to check
151 * @len: The size of increase.
152 *
153 * Return: 0 on success.
154 */
155static int check_brk_limits(unsigned long addr, unsigned long len)
156{
157 unsigned long mapped_addr;
158
159 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
160 if (IS_ERR_VALUE(mapped_addr))
161 return mapped_addr;
162
163 return mlock_future_check(current->mm, current->mm->def_flags, len);
164}
165static int do_brk_munmap(struct ma_state *mas, struct vm_area_struct *vma,
166 unsigned long newbrk, unsigned long oldbrk,
167 struct list_head *uf);
168static int do_brk_flags(struct ma_state *mas, struct vm_area_struct *brkvma,
169 unsigned long addr, unsigned long request, unsigned long flags);
170SYSCALL_DEFINE1(brk, unsigned long, brk)
171{
172 unsigned long newbrk, oldbrk, origbrk;
173 struct mm_struct *mm = current->mm;
174 struct vm_area_struct *brkvma, *next = NULL;
175 unsigned long min_brk;
176 bool populate;
177 bool downgraded = false;
178 LIST_HEAD(uf);
179 MA_STATE(mas, &mm->mm_mt, 0, 0);
180
181 if (mmap_write_lock_killable(mm))
182 return -EINTR;
183
184 origbrk = mm->brk;
185
186#ifdef CONFIG_COMPAT_BRK
187 /*
188 * CONFIG_COMPAT_BRK can still be overridden by setting
189 * randomize_va_space to 2, which will still cause mm->start_brk
190 * to be arbitrarily shifted
191 */
192 if (current->brk_randomized)
193 min_brk = mm->start_brk;
194 else
195 min_brk = mm->end_data;
196#else
197 min_brk = mm->start_brk;
198#endif
199 if (brk < min_brk)
200 goto out;
201
202 /*
203 * Check against rlimit here. If this check is done later after the test
204 * of oldbrk with newbrk then it can escape the test and let the data
205 * segment grow beyond its set limit the in case where the limit is
206 * not page aligned -Ram Gupta
207 */
208 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
209 mm->end_data, mm->start_data))
210 goto out;
211
212 newbrk = PAGE_ALIGN(brk);
213 oldbrk = PAGE_ALIGN(mm->brk);
214 if (oldbrk == newbrk) {
215 mm->brk = brk;
216 goto success;
217 }
218
219 /*
220 * Always allow shrinking brk.
221 * do_brk_munmap() may downgrade mmap_lock to read.
222 */
223 if (brk <= mm->brk) {
224 int ret;
225
226 /* Search one past newbrk */
227 mas_set(&mas, newbrk);
228 brkvma = mas_find(&mas, oldbrk);
229 if (!brkvma || brkvma->vm_start >= oldbrk)
230 goto out; /* mapping intersects with an existing non-brk vma. */
231 /*
232 * mm->brk must be protected by write mmap_lock.
233 * do_brk_munmap() may downgrade the lock, so update it
234 * before calling do_brk_munmap().
235 */
236 mm->brk = brk;
237 ret = do_brk_munmap(&mas, brkvma, newbrk, oldbrk, &uf);
238 if (ret == 1) {
239 downgraded = true;
240 goto success;
241 } else if (!ret)
242 goto success;
243
244 mm->brk = origbrk;
245 goto out;
246 }
247
248 if (check_brk_limits(oldbrk, newbrk - oldbrk))
249 goto out;
250
251 /*
252 * Only check if the next VMA is within the stack_guard_gap of the
253 * expansion area
254 */
255 mas_set(&mas, oldbrk);
256 next = mas_find(&mas, newbrk - 1 + PAGE_SIZE + stack_guard_gap);
257 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
258 goto out;
259
260 brkvma = mas_prev(&mas, mm->start_brk);
261 /* Ok, looks good - let it rip. */
262 if (do_brk_flags(&mas, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
263 goto out;
264
265 mm->brk = brk;
266
267success:
268 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
269 if (downgraded)
270 mmap_read_unlock(mm);
271 else
272 mmap_write_unlock(mm);
273 userfaultfd_unmap_complete(mm, &uf);
274 if (populate)
275 mm_populate(oldbrk, newbrk - oldbrk);
276 return brk;
277
278out:
279 mmap_write_unlock(mm);
280 return origbrk;
281}
282
283#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
284extern void mt_validate(struct maple_tree *mt);
285extern void mt_dump(const struct maple_tree *mt);
286
287/* Validate the maple tree */
288static void validate_mm_mt(struct mm_struct *mm)
289{
290 struct maple_tree *mt = &mm->mm_mt;
291 struct vm_area_struct *vma_mt;
292
293 MA_STATE(mas, mt, 0, 0);
294
295 mt_validate(&mm->mm_mt);
296 mas_for_each(&mas, vma_mt, ULONG_MAX) {
297 if ((vma_mt->vm_start != mas.index) ||
298 (vma_mt->vm_end - 1 != mas.last)) {
299 pr_emerg("issue in %s\n", current->comm);
300 dump_stack();
301 dump_vma(vma_mt);
302 pr_emerg("mt piv: %p %lu - %lu\n", vma_mt,
303 mas.index, mas.last);
304 pr_emerg("mt vma: %p %lu - %lu\n", vma_mt,
305 vma_mt->vm_start, vma_mt->vm_end);
306
307 mt_dump(mas.tree);
308 if (vma_mt->vm_end != mas.last + 1) {
309 pr_err("vma: %p vma_mt %lu-%lu\tmt %lu-%lu\n",
310 mm, vma_mt->vm_start, vma_mt->vm_end,
311 mas.index, mas.last);
312 mt_dump(mas.tree);
313 }
314 VM_BUG_ON_MM(vma_mt->vm_end != mas.last + 1, mm);
315 if (vma_mt->vm_start != mas.index) {
316 pr_err("vma: %p vma_mt %p %lu - %lu doesn't match\n",
317 mm, vma_mt, vma_mt->vm_start, vma_mt->vm_end);
318 mt_dump(mas.tree);
319 }
320 VM_BUG_ON_MM(vma_mt->vm_start != mas.index, mm);
321 }
322 }
323}
324
325static void validate_mm(struct mm_struct *mm)
326{
327 int bug = 0;
328 int i = 0;
329 struct vm_area_struct *vma;
330 MA_STATE(mas, &mm->mm_mt, 0, 0);
331
332 validate_mm_mt(mm);
333
334 mas_for_each(&mas, vma, ULONG_MAX) {
335#ifdef CONFIG_DEBUG_VM_RB
336 struct anon_vma *anon_vma = vma->anon_vma;
337 struct anon_vma_chain *avc;
338
339 if (anon_vma) {
340 anon_vma_lock_read(anon_vma);
341 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
342 anon_vma_interval_tree_verify(avc);
343 anon_vma_unlock_read(anon_vma);
344 }
345#endif
346 i++;
347 }
348 if (i != mm->map_count) {
349 pr_emerg("map_count %d mas_for_each %d\n", mm->map_count, i);
350 bug = 1;
351 }
352 VM_BUG_ON_MM(bug, mm);
353}
354
355#else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
356#define validate_mm_mt(root) do { } while (0)
357#define validate_mm(mm) do { } while (0)
358#endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
359
360/*
361 * vma has some anon_vma assigned, and is already inserted on that
362 * anon_vma's interval trees.
363 *
364 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
365 * vma must be removed from the anon_vma's interval trees using
366 * anon_vma_interval_tree_pre_update_vma().
367 *
368 * After the update, the vma will be reinserted using
369 * anon_vma_interval_tree_post_update_vma().
370 *
371 * The entire update must be protected by exclusive mmap_lock and by
372 * the root anon_vma's mutex.
373 */
374static inline void
375anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
376{
377 struct anon_vma_chain *avc;
378
379 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
380 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
381}
382
383static inline void
384anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
385{
386 struct anon_vma_chain *avc;
387
388 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
389 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
390}
391
392static unsigned long count_vma_pages_range(struct mm_struct *mm,
393 unsigned long addr, unsigned long end)
394{
395 VMA_ITERATOR(vmi, mm, addr);
396 struct vm_area_struct *vma;
397 unsigned long nr_pages = 0;
398
399 for_each_vma_range(vmi, vma, end) {
400 unsigned long vm_start = max(addr, vma->vm_start);
401 unsigned long vm_end = min(end, vma->vm_end);
402
403 nr_pages += PHYS_PFN(vm_end - vm_start);
404 }
405
406 return nr_pages;
407}
408
409static void __vma_link_file(struct vm_area_struct *vma,
410 struct address_space *mapping)
411{
412 if (vma->vm_flags & VM_SHARED)
413 mapping_allow_writable(mapping);
414
415 flush_dcache_mmap_lock(mapping);
416 vma_interval_tree_insert(vma, &mapping->i_mmap);
417 flush_dcache_mmap_unlock(mapping);
418}
419
420/*
421 * vma_mas_store() - Store a VMA in the maple tree.
422 * @vma: The vm_area_struct
423 * @mas: The maple state
424 *
425 * Efficient way to store a VMA in the maple tree when the @mas has already
426 * walked to the correct location.
427 *
428 * Note: the end address is inclusive in the maple tree.
429 */
430void vma_mas_store(struct vm_area_struct *vma, struct ma_state *mas)
431{
432 trace_vma_store(mas->tree, vma);
433 mas_set_range(mas, vma->vm_start, vma->vm_end - 1);
434 mas_store_prealloc(mas, vma);
435}
436
437/*
438 * vma_mas_remove() - Remove a VMA from the maple tree.
439 * @vma: The vm_area_struct
440 * @mas: The maple state
441 *
442 * Efficient way to remove a VMA from the maple tree when the @mas has already
443 * been established and points to the correct location.
444 * Note: the end address is inclusive in the maple tree.
445 */
446void vma_mas_remove(struct vm_area_struct *vma, struct ma_state *mas)
447{
448 trace_vma_mas_szero(mas->tree, vma->vm_start, vma->vm_end - 1);
449 mas->index = vma->vm_start;
450 mas->last = vma->vm_end - 1;
451 mas_store_prealloc(mas, NULL);
452}
453
454/*
455 * vma_mas_szero() - Set a given range to zero. Used when modifying a
456 * vm_area_struct start or end.
457 *
458 * @mas: The maple tree ma_state
459 * @start: The start address to zero
460 * @end: The end address to zero.
461 */
462static inline void vma_mas_szero(struct ma_state *mas, unsigned long start,
463 unsigned long end)
464{
465 trace_vma_mas_szero(mas->tree, start, end - 1);
466 mas_set_range(mas, start, end - 1);
467 mas_store_prealloc(mas, NULL);
468}
469
470static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
471{
472 MA_STATE(mas, &mm->mm_mt, 0, 0);
473 struct address_space *mapping = NULL;
474
475 if (mas_preallocate(&mas, vma, GFP_KERNEL))
476 return -ENOMEM;
477
478 if (vma->vm_file) {
479 mapping = vma->vm_file->f_mapping;
480 i_mmap_lock_write(mapping);
481 }
482
483 vma_mas_store(vma, &mas);
484
485 if (mapping) {
486 __vma_link_file(vma, mapping);
487 i_mmap_unlock_write(mapping);
488 }
489
490 mm->map_count++;
491 validate_mm(mm);
492 return 0;
493}
494
495/*
496 * vma_expand - Expand an existing VMA
497 *
498 * @mas: The maple state
499 * @vma: The vma to expand
500 * @start: The start of the vma
501 * @end: The exclusive end of the vma
502 * @pgoff: The page offset of vma
503 * @next: The current of next vma.
504 *
505 * Expand @vma to @start and @end. Can expand off the start and end. Will
506 * expand over @next if it's different from @vma and @end == @next->vm_end.
507 * Checking if the @vma can expand and merge with @next needs to be handled by
508 * the caller.
509 *
510 * Returns: 0 on success
511 */
512inline int vma_expand(struct ma_state *mas, struct vm_area_struct *vma,
513 unsigned long start, unsigned long end, pgoff_t pgoff,
514 struct vm_area_struct *next)
515{
516 struct mm_struct *mm = vma->vm_mm;
517 struct address_space *mapping = NULL;
518 struct rb_root_cached *root = NULL;
519 struct anon_vma *anon_vma = vma->anon_vma;
520 struct file *file = vma->vm_file;
521 bool remove_next = false;
522
523 if (next && (vma != next) && (end == next->vm_end)) {
524 remove_next = true;
525 if (next->anon_vma && !vma->anon_vma) {
526 int error;
527
528 anon_vma = next->anon_vma;
529 vma->anon_vma = anon_vma;
530 error = anon_vma_clone(vma, next);
531 if (error)
532 return error;
533 }
534 }
535
536 /* Not merging but overwriting any part of next is not handled. */
537 VM_BUG_ON(next && !remove_next && next != vma && end > next->vm_start);
538 /* Only handles expanding */
539 VM_BUG_ON(vma->vm_start < start || vma->vm_end > end);
540
541 if (mas_preallocate(mas, vma, GFP_KERNEL))
542 goto nomem;
543
544 vma_adjust_trans_huge(vma, start, end, 0);
545
546 if (file) {
547 mapping = file->f_mapping;
548 root = &mapping->i_mmap;
549 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
550 i_mmap_lock_write(mapping);
551 }
552
553 if (anon_vma) {
554 anon_vma_lock_write(anon_vma);
555 anon_vma_interval_tree_pre_update_vma(vma);
556 }
557
558 if (file) {
559 flush_dcache_mmap_lock(mapping);
560 vma_interval_tree_remove(vma, root);
561 }
562
563 vma->vm_start = start;
564 vma->vm_end = end;
565 vma->vm_pgoff = pgoff;
566 /* Note: mas must be pointing to the expanding VMA */
567 vma_mas_store(vma, mas);
568
569 if (file) {
570 vma_interval_tree_insert(vma, root);
571 flush_dcache_mmap_unlock(mapping);
572 }
573
574 /* Expanding over the next vma */
575 if (remove_next && file) {
576 __remove_shared_vm_struct(next, file, mapping);
577 }
578
579 if (anon_vma) {
580 anon_vma_interval_tree_post_update_vma(vma);
581 anon_vma_unlock_write(anon_vma);
582 }
583
584 if (file) {
585 i_mmap_unlock_write(mapping);
586 uprobe_mmap(vma);
587 }
588
589 if (remove_next) {
590 if (file) {
591 uprobe_munmap(next, next->vm_start, next->vm_end);
592 fput(file);
593 }
594 if (next->anon_vma)
595 anon_vma_merge(vma, next);
596 mm->map_count--;
597 mpol_put(vma_policy(next));
598 vm_area_free(next);
599 }
600
601 validate_mm(mm);
602 return 0;
603
604nomem:
605 return -ENOMEM;
606}
607
608/*
609 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
610 * is already present in an i_mmap tree without adjusting the tree.
611 * The following helper function should be used when such adjustments
612 * are necessary. The "insert" vma (if any) is to be inserted
613 * before we drop the necessary locks.
614 */
615int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
616 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
617 struct vm_area_struct *expand)
618{
619 struct mm_struct *mm = vma->vm_mm;
620 struct vm_area_struct *next_next = NULL; /* uninit var warning */
621 struct vm_area_struct *next = find_vma(mm, vma->vm_end);
622 struct vm_area_struct *orig_vma = vma;
623 struct address_space *mapping = NULL;
624 struct rb_root_cached *root = NULL;
625 struct anon_vma *anon_vma = NULL;
626 struct file *file = vma->vm_file;
627 bool vma_changed = false;
628 long adjust_next = 0;
629 int remove_next = 0;
630 MA_STATE(mas, &mm->mm_mt, 0, 0);
631 struct vm_area_struct *exporter = NULL, *importer = NULL;
632
633 if (next && !insert) {
634 if (end >= next->vm_end) {
635 /*
636 * vma expands, overlapping all the next, and
637 * perhaps the one after too (mprotect case 6).
638 * The only other cases that gets here are
639 * case 1, case 7 and case 8.
640 */
641 if (next == expand) {
642 /*
643 * The only case where we don't expand "vma"
644 * and we expand "next" instead is case 8.
645 */
646 VM_WARN_ON(end != next->vm_end);
647 /*
648 * remove_next == 3 means we're
649 * removing "vma" and that to do so we
650 * swapped "vma" and "next".
651 */
652 remove_next = 3;
653 VM_WARN_ON(file != next->vm_file);
654 swap(vma, next);
655 } else {
656 VM_WARN_ON(expand != vma);
657 /*
658 * case 1, 6, 7, remove_next == 2 is case 6,
659 * remove_next == 1 is case 1 or 7.
660 */
661 remove_next = 1 + (end > next->vm_end);
662 if (remove_next == 2)
663 next_next = find_vma(mm, next->vm_end);
664
665 VM_WARN_ON(remove_next == 2 &&
666 end != next_next->vm_end);
667 }
668
669 exporter = next;
670 importer = vma;
671
672 /*
673 * If next doesn't have anon_vma, import from vma after
674 * next, if the vma overlaps with it.
675 */
676 if (remove_next == 2 && !next->anon_vma)
677 exporter = next_next;
678
679 } else if (end > next->vm_start) {
680 /*
681 * vma expands, overlapping part of the next:
682 * mprotect case 5 shifting the boundary up.
683 */
684 adjust_next = (end - next->vm_start);
685 exporter = next;
686 importer = vma;
687 VM_WARN_ON(expand != importer);
688 } else if (end < vma->vm_end) {
689 /*
690 * vma shrinks, and !insert tells it's not
691 * split_vma inserting another: so it must be
692 * mprotect case 4 shifting the boundary down.
693 */
694 adjust_next = -(vma->vm_end - end);
695 exporter = vma;
696 importer = next;
697 VM_WARN_ON(expand != importer);
698 }
699
700 /*
701 * Easily overlooked: when mprotect shifts the boundary,
702 * make sure the expanding vma has anon_vma set if the
703 * shrinking vma had, to cover any anon pages imported.
704 */
705 if (exporter && exporter->anon_vma && !importer->anon_vma) {
706 int error;
707
708 importer->anon_vma = exporter->anon_vma;
709 error = anon_vma_clone(importer, exporter);
710 if (error)
711 return error;
712 }
713 }
714
715 if (mas_preallocate(&mas, vma, GFP_KERNEL))
716 return -ENOMEM;
717
718 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
719 if (file) {
720 mapping = file->f_mapping;
721 root = &mapping->i_mmap;
722 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
723
724 if (adjust_next)
725 uprobe_munmap(next, next->vm_start, next->vm_end);
726
727 i_mmap_lock_write(mapping);
728 if (insert && insert->vm_file) {
729 /*
730 * Put into interval tree now, so instantiated pages
731 * are visible to arm/parisc __flush_dcache_page
732 * throughout; but we cannot insert into address
733 * space until vma start or end is updated.
734 */
735 __vma_link_file(insert, insert->vm_file->f_mapping);
736 }
737 }
738
739 anon_vma = vma->anon_vma;
740 if (!anon_vma && adjust_next)
741 anon_vma = next->anon_vma;
742 if (anon_vma) {
743 VM_WARN_ON(adjust_next && next->anon_vma &&
744 anon_vma != next->anon_vma);
745 anon_vma_lock_write(anon_vma);
746 anon_vma_interval_tree_pre_update_vma(vma);
747 if (adjust_next)
748 anon_vma_interval_tree_pre_update_vma(next);
749 }
750
751 if (file) {
752 flush_dcache_mmap_lock(mapping);
753 vma_interval_tree_remove(vma, root);
754 if (adjust_next)
755 vma_interval_tree_remove(next, root);
756 }
757
758 if (start != vma->vm_start) {
759 if ((vma->vm_start < start) &&
760 (!insert || (insert->vm_end != start))) {
761 vma_mas_szero(&mas, vma->vm_start, start);
762 VM_WARN_ON(insert && insert->vm_start > vma->vm_start);
763 } else {
764 vma_changed = true;
765 }
766 vma->vm_start = start;
767 }
768 if (end != vma->vm_end) {
769 if (vma->vm_end > end) {
770 if (!insert || (insert->vm_start != end)) {
771 vma_mas_szero(&mas, end, vma->vm_end);
772 mas_reset(&mas);
773 VM_WARN_ON(insert &&
774 insert->vm_end < vma->vm_end);
775 }
776 } else {
777 vma_changed = true;
778 }
779 vma->vm_end = end;
780 }
781
782 if (vma_changed)
783 vma_mas_store(vma, &mas);
784
785 vma->vm_pgoff = pgoff;
786 if (adjust_next) {
787 next->vm_start += adjust_next;
788 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
789 vma_mas_store(next, &mas);
790 }
791
792 if (file) {
793 if (adjust_next)
794 vma_interval_tree_insert(next, root);
795 vma_interval_tree_insert(vma, root);
796 flush_dcache_mmap_unlock(mapping);
797 }
798
799 if (remove_next && file) {
800 __remove_shared_vm_struct(next, file, mapping);
801 if (remove_next == 2)
802 __remove_shared_vm_struct(next_next, file, mapping);
803 } else if (insert) {
804 /*
805 * split_vma has split insert from vma, and needs
806 * us to insert it before dropping the locks
807 * (it may either follow vma or precede it).
808 */
809 mas_reset(&mas);
810 vma_mas_store(insert, &mas);
811 mm->map_count++;
812 }
813
814 if (anon_vma) {
815 anon_vma_interval_tree_post_update_vma(vma);
816 if (adjust_next)
817 anon_vma_interval_tree_post_update_vma(next);
818 anon_vma_unlock_write(anon_vma);
819 }
820
821 if (file) {
822 i_mmap_unlock_write(mapping);
823 uprobe_mmap(vma);
824
825 if (adjust_next)
826 uprobe_mmap(next);
827 }
828
829 if (remove_next) {
830again:
831 if (file) {
832 uprobe_munmap(next, next->vm_start, next->vm_end);
833 fput(file);
834 }
835 if (next->anon_vma)
836 anon_vma_merge(vma, next);
837 mm->map_count--;
838 mpol_put(vma_policy(next));
839 if (remove_next != 2)
840 BUG_ON(vma->vm_end < next->vm_end);
841 vm_area_free(next);
842
843 /*
844 * In mprotect's case 6 (see comments on vma_merge),
845 * we must remove next_next too.
846 */
847 if (remove_next == 2) {
848 remove_next = 1;
849 next = next_next;
850 goto again;
851 }
852 }
853 if (insert && file)
854 uprobe_mmap(insert);
855
856 mas_destroy(&mas);
857 validate_mm(mm);
858
859 return 0;
860}
861
862/*
863 * If the vma has a ->close operation then the driver probably needs to release
864 * per-vma resources, so we don't attempt to merge those.
865 */
866static inline int is_mergeable_vma(struct vm_area_struct *vma,
867 struct file *file, unsigned long vm_flags,
868 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
869 struct anon_vma_name *anon_name)
870{
871 /*
872 * VM_SOFTDIRTY should not prevent from VMA merging, if we
873 * match the flags but dirty bit -- the caller should mark
874 * merged VMA as dirty. If dirty bit won't be excluded from
875 * comparison, we increase pressure on the memory system forcing
876 * the kernel to generate new VMAs when old one could be
877 * extended instead.
878 */
879 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
880 return 0;
881 if (vma->vm_file != file)
882 return 0;
883 if (vma->vm_ops && vma->vm_ops->close)
884 return 0;
885 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
886 return 0;
887 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
888 return 0;
889 return 1;
890}
891
892static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
893 struct anon_vma *anon_vma2,
894 struct vm_area_struct *vma)
895{
896 /*
897 * The list_is_singular() test is to avoid merging VMA cloned from
898 * parents. This can improve scalability caused by anon_vma lock.
899 */
900 if ((!anon_vma1 || !anon_vma2) && (!vma ||
901 list_is_singular(&vma->anon_vma_chain)))
902 return 1;
903 return anon_vma1 == anon_vma2;
904}
905
906/*
907 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
908 * in front of (at a lower virtual address and file offset than) the vma.
909 *
910 * We cannot merge two vmas if they have differently assigned (non-NULL)
911 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
912 *
913 * We don't check here for the merged mmap wrapping around the end of pagecache
914 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
915 * wrap, nor mmaps which cover the final page at index -1UL.
916 */
917static int
918can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
919 struct anon_vma *anon_vma, struct file *file,
920 pgoff_t vm_pgoff,
921 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
922 struct anon_vma_name *anon_name)
923{
924 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
925 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
926 if (vma->vm_pgoff == vm_pgoff)
927 return 1;
928 }
929 return 0;
930}
931
932/*
933 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
934 * beyond (at a higher virtual address and file offset than) the vma.
935 *
936 * We cannot merge two vmas if they have differently assigned (non-NULL)
937 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
938 */
939static int
940can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
941 struct anon_vma *anon_vma, struct file *file,
942 pgoff_t vm_pgoff,
943 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
944 struct anon_vma_name *anon_name)
945{
946 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
947 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
948 pgoff_t vm_pglen;
949 vm_pglen = vma_pages(vma);
950 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
951 return 1;
952 }
953 return 0;
954}
955
956/*
957 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
958 * figure out whether that can be merged with its predecessor or its
959 * successor. Or both (it neatly fills a hole).
960 *
961 * In most cases - when called for mmap, brk or mremap - [addr,end) is
962 * certain not to be mapped by the time vma_merge is called; but when
963 * called for mprotect, it is certain to be already mapped (either at
964 * an offset within prev, or at the start of next), and the flags of
965 * this area are about to be changed to vm_flags - and the no-change
966 * case has already been eliminated.
967 *
968 * The following mprotect cases have to be considered, where AAAA is
969 * the area passed down from mprotect_fixup, never extending beyond one
970 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
971 *
972 * AAAA AAAA AAAA
973 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
974 * cannot merge might become might become
975 * PPNNNNNNNNNN PPPPPPPPPPNN
976 * mmap, brk or case 4 below case 5 below
977 * mremap move:
978 * AAAA AAAA
979 * PPPP NNNN PPPPNNNNXXXX
980 * might become might become
981 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
982 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
983 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
984 *
985 * It is important for case 8 that the vma NNNN overlapping the
986 * region AAAA is never going to extended over XXXX. Instead XXXX must
987 * be extended in region AAAA and NNNN must be removed. This way in
988 * all cases where vma_merge succeeds, the moment vma_adjust drops the
989 * rmap_locks, the properties of the merged vma will be already
990 * correct for the whole merged range. Some of those properties like
991 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
992 * be correct for the whole merged range immediately after the
993 * rmap_locks are released. Otherwise if XXXX would be removed and
994 * NNNN would be extended over the XXXX range, remove_migration_ptes
995 * or other rmap walkers (if working on addresses beyond the "end"
996 * parameter) may establish ptes with the wrong permissions of NNNN
997 * instead of the right permissions of XXXX.
998 */
999struct vm_area_struct *vma_merge(struct mm_struct *mm,
1000 struct vm_area_struct *prev, unsigned long addr,
1001 unsigned long end, unsigned long vm_flags,
1002 struct anon_vma *anon_vma, struct file *file,
1003 pgoff_t pgoff, struct mempolicy *policy,
1004 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1005 struct anon_vma_name *anon_name)
1006{
1007 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1008 struct vm_area_struct *mid, *next, *res;
1009 int err = -1;
1010 bool merge_prev = false;
1011 bool merge_next = false;
1012
1013 /*
1014 * We later require that vma->vm_flags == vm_flags,
1015 * so this tests vma->vm_flags & VM_SPECIAL, too.
1016 */
1017 if (vm_flags & VM_SPECIAL)
1018 return NULL;
1019
1020 next = find_vma(mm, prev ? prev->vm_end : 0);
1021 mid = next;
1022 if (next && next->vm_end == end) /* cases 6, 7, 8 */
1023 next = find_vma(mm, next->vm_end);
1024
1025 /* verify some invariant that must be enforced by the caller */
1026 VM_WARN_ON(prev && addr <= prev->vm_start);
1027 VM_WARN_ON(mid && end > mid->vm_end);
1028 VM_WARN_ON(addr >= end);
1029
1030 /* Can we merge the predecessor? */
1031 if (prev && prev->vm_end == addr &&
1032 mpol_equal(vma_policy(prev), policy) &&
1033 can_vma_merge_after(prev, vm_flags,
1034 anon_vma, file, pgoff,
1035 vm_userfaultfd_ctx, anon_name)) {
1036 merge_prev = true;
1037 }
1038 /* Can we merge the successor? */
1039 if (next && end == next->vm_start &&
1040 mpol_equal(policy, vma_policy(next)) &&
1041 can_vma_merge_before(next, vm_flags,
1042 anon_vma, file, pgoff+pglen,
1043 vm_userfaultfd_ctx, anon_name)) {
1044 merge_next = true;
1045 }
1046 /* Can we merge both the predecessor and the successor? */
1047 if (merge_prev && merge_next &&
1048 is_mergeable_anon_vma(prev->anon_vma,
1049 next->anon_vma, NULL)) { /* cases 1, 6 */
1050 err = __vma_adjust(prev, prev->vm_start,
1051 next->vm_end, prev->vm_pgoff, NULL,
1052 prev);
1053 res = prev;
1054 } else if (merge_prev) { /* cases 2, 5, 7 */
1055 err = __vma_adjust(prev, prev->vm_start,
1056 end, prev->vm_pgoff, NULL, prev);
1057 res = prev;
1058 } else if (merge_next) {
1059 if (prev && addr < prev->vm_end) /* case 4 */
1060 err = __vma_adjust(prev, prev->vm_start,
1061 addr, prev->vm_pgoff, NULL, next);
1062 else /* cases 3, 8 */
1063 err = __vma_adjust(mid, addr, next->vm_end,
1064 next->vm_pgoff - pglen, NULL, next);
1065 res = next;
1066 }
1067
1068 /*
1069 * Cannot merge with predecessor or successor or error in __vma_adjust?
1070 */
1071 if (err)
1072 return NULL;
1073 khugepaged_enter_vma(res, vm_flags);
1074 return res;
1075}
1076
1077/*
1078 * Rough compatibility check to quickly see if it's even worth looking
1079 * at sharing an anon_vma.
1080 *
1081 * They need to have the same vm_file, and the flags can only differ
1082 * in things that mprotect may change.
1083 *
1084 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1085 * we can merge the two vma's. For example, we refuse to merge a vma if
1086 * there is a vm_ops->close() function, because that indicates that the
1087 * driver is doing some kind of reference counting. But that doesn't
1088 * really matter for the anon_vma sharing case.
1089 */
1090static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1091{
1092 return a->vm_end == b->vm_start &&
1093 mpol_equal(vma_policy(a), vma_policy(b)) &&
1094 a->vm_file == b->vm_file &&
1095 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1096 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1097}
1098
1099/*
1100 * Do some basic sanity checking to see if we can re-use the anon_vma
1101 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1102 * the same as 'old', the other will be the new one that is trying
1103 * to share the anon_vma.
1104 *
1105 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1106 * the anon_vma of 'old' is concurrently in the process of being set up
1107 * by another page fault trying to merge _that_. But that's ok: if it
1108 * is being set up, that automatically means that it will be a singleton
1109 * acceptable for merging, so we can do all of this optimistically. But
1110 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1111 *
1112 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1113 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1114 * is to return an anon_vma that is "complex" due to having gone through
1115 * a fork).
1116 *
1117 * We also make sure that the two vma's are compatible (adjacent,
1118 * and with the same memory policies). That's all stable, even with just
1119 * a read lock on the mmap_lock.
1120 */
1121static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1122{
1123 if (anon_vma_compatible(a, b)) {
1124 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1125
1126 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1127 return anon_vma;
1128 }
1129 return NULL;
1130}
1131
1132/*
1133 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1134 * neighbouring vmas for a suitable anon_vma, before it goes off
1135 * to allocate a new anon_vma. It checks because a repetitive
1136 * sequence of mprotects and faults may otherwise lead to distinct
1137 * anon_vmas being allocated, preventing vma merge in subsequent
1138 * mprotect.
1139 */
1140struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1141{
1142 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1143 struct anon_vma *anon_vma = NULL;
1144 struct vm_area_struct *prev, *next;
1145
1146 /* Try next first. */
1147 next = mas_walk(&mas);
1148 if (next) {
1149 anon_vma = reusable_anon_vma(next, vma, next);
1150 if (anon_vma)
1151 return anon_vma;
1152 }
1153
1154 prev = mas_prev(&mas, 0);
1155 VM_BUG_ON_VMA(prev != vma, vma);
1156 prev = mas_prev(&mas, 0);
1157 /* Try prev next. */
1158 if (prev)
1159 anon_vma = reusable_anon_vma(prev, prev, vma);
1160
1161 /*
1162 * We might reach here with anon_vma == NULL if we can't find
1163 * any reusable anon_vma.
1164 * There's no absolute need to look only at touching neighbours:
1165 * we could search further afield for "compatible" anon_vmas.
1166 * But it would probably just be a waste of time searching,
1167 * or lead to too many vmas hanging off the same anon_vma.
1168 * We're trying to allow mprotect remerging later on,
1169 * not trying to minimize memory used for anon_vmas.
1170 */
1171 return anon_vma;
1172}
1173
1174/*
1175 * If a hint addr is less than mmap_min_addr change hint to be as
1176 * low as possible but still greater than mmap_min_addr
1177 */
1178static inline unsigned long round_hint_to_min(unsigned long hint)
1179{
1180 hint &= PAGE_MASK;
1181 if (((void *)hint != NULL) &&
1182 (hint < mmap_min_addr))
1183 return PAGE_ALIGN(mmap_min_addr);
1184 return hint;
1185}
1186
1187int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1188 unsigned long len)
1189{
1190 unsigned long locked, lock_limit;
1191
1192 /* mlock MCL_FUTURE? */
1193 if (flags & VM_LOCKED) {
1194 locked = len >> PAGE_SHIFT;
1195 locked += mm->locked_vm;
1196 lock_limit = rlimit(RLIMIT_MEMLOCK);
1197 lock_limit >>= PAGE_SHIFT;
1198 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1199 return -EAGAIN;
1200 }
1201 return 0;
1202}
1203
1204static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1205{
1206 if (S_ISREG(inode->i_mode))
1207 return MAX_LFS_FILESIZE;
1208
1209 if (S_ISBLK(inode->i_mode))
1210 return MAX_LFS_FILESIZE;
1211
1212 if (S_ISSOCK(inode->i_mode))
1213 return MAX_LFS_FILESIZE;
1214
1215 /* Special "we do even unsigned file positions" case */
1216 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1217 return 0;
1218
1219 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1220 return ULONG_MAX;
1221}
1222
1223static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1224 unsigned long pgoff, unsigned long len)
1225{
1226 u64 maxsize = file_mmap_size_max(file, inode);
1227
1228 if (maxsize && len > maxsize)
1229 return false;
1230 maxsize -= len;
1231 if (pgoff > maxsize >> PAGE_SHIFT)
1232 return false;
1233 return true;
1234}
1235
1236/*
1237 * The caller must write-lock current->mm->mmap_lock.
1238 */
1239unsigned long do_mmap(struct file *file, unsigned long addr,
1240 unsigned long len, unsigned long prot,
1241 unsigned long flags, unsigned long pgoff,
1242 unsigned long *populate, struct list_head *uf)
1243{
1244 struct mm_struct *mm = current->mm;
1245 vm_flags_t vm_flags;
1246 int pkey = 0;
1247
1248 validate_mm(mm);
1249 *populate = 0;
1250
1251 if (!len)
1252 return -EINVAL;
1253
1254 /*
1255 * Does the application expect PROT_READ to imply PROT_EXEC?
1256 *
1257 * (the exception is when the underlying filesystem is noexec
1258 * mounted, in which case we dont add PROT_EXEC.)
1259 */
1260 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1261 if (!(file && path_noexec(&file->f_path)))
1262 prot |= PROT_EXEC;
1263
1264 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1265 if (flags & MAP_FIXED_NOREPLACE)
1266 flags |= MAP_FIXED;
1267
1268 if (!(flags & MAP_FIXED))
1269 addr = round_hint_to_min(addr);
1270
1271 /* Careful about overflows.. */
1272 len = PAGE_ALIGN(len);
1273 if (!len)
1274 return -ENOMEM;
1275
1276 /* offset overflow? */
1277 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1278 return -EOVERFLOW;
1279
1280 /* Too many mappings? */
1281 if (mm->map_count > sysctl_max_map_count)
1282 return -ENOMEM;
1283
1284 /* Obtain the address to map to. we verify (or select) it and ensure
1285 * that it represents a valid section of the address space.
1286 */
1287 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1288 if (IS_ERR_VALUE(addr))
1289 return addr;
1290
1291 if (flags & MAP_FIXED_NOREPLACE) {
1292 if (find_vma_intersection(mm, addr, addr + len))
1293 return -EEXIST;
1294 }
1295
1296 if (prot == PROT_EXEC) {
1297 pkey = execute_only_pkey(mm);
1298 if (pkey < 0)
1299 pkey = 0;
1300 }
1301
1302 /* Do simple checking here so the lower-level routines won't have
1303 * to. we assume access permissions have been handled by the open
1304 * of the memory object, so we don't do any here.
1305 */
1306 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1307 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1308
1309 if (flags & MAP_LOCKED)
1310 if (!can_do_mlock())
1311 return -EPERM;
1312
1313 if (mlock_future_check(mm, vm_flags, len))
1314 return -EAGAIN;
1315
1316 if (file) {
1317 struct inode *inode = file_inode(file);
1318 unsigned long flags_mask;
1319
1320 if (!file_mmap_ok(file, inode, pgoff, len))
1321 return -EOVERFLOW;
1322
1323 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1324
1325 switch (flags & MAP_TYPE) {
1326 case MAP_SHARED:
1327 /*
1328 * Force use of MAP_SHARED_VALIDATE with non-legacy
1329 * flags. E.g. MAP_SYNC is dangerous to use with
1330 * MAP_SHARED as you don't know which consistency model
1331 * you will get. We silently ignore unsupported flags
1332 * with MAP_SHARED to preserve backward compatibility.
1333 */
1334 flags &= LEGACY_MAP_MASK;
1335 fallthrough;
1336 case MAP_SHARED_VALIDATE:
1337 if (flags & ~flags_mask)
1338 return -EOPNOTSUPP;
1339 if (prot & PROT_WRITE) {
1340 if (!(file->f_mode & FMODE_WRITE))
1341 return -EACCES;
1342 if (IS_SWAPFILE(file->f_mapping->host))
1343 return -ETXTBSY;
1344 }
1345
1346 /*
1347 * Make sure we don't allow writing to an append-only
1348 * file..
1349 */
1350 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1351 return -EACCES;
1352
1353 vm_flags |= VM_SHARED | VM_MAYSHARE;
1354 if (!(file->f_mode & FMODE_WRITE))
1355 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1356 fallthrough;
1357 case MAP_PRIVATE:
1358 if (!(file->f_mode & FMODE_READ))
1359 return -EACCES;
1360 if (path_noexec(&file->f_path)) {
1361 if (vm_flags & VM_EXEC)
1362 return -EPERM;
1363 vm_flags &= ~VM_MAYEXEC;
1364 }
1365
1366 if (!file->f_op->mmap)
1367 return -ENODEV;
1368 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1369 return -EINVAL;
1370 break;
1371
1372 default:
1373 return -EINVAL;
1374 }
1375 } else {
1376 switch (flags & MAP_TYPE) {
1377 case MAP_SHARED:
1378 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1379 return -EINVAL;
1380 /*
1381 * Ignore pgoff.
1382 */
1383 pgoff = 0;
1384 vm_flags |= VM_SHARED | VM_MAYSHARE;
1385 break;
1386 case MAP_PRIVATE:
1387 /*
1388 * Set pgoff according to addr for anon_vma.
1389 */
1390 pgoff = addr >> PAGE_SHIFT;
1391 break;
1392 default:
1393 return -EINVAL;
1394 }
1395 }
1396
1397 /*
1398 * Set 'VM_NORESERVE' if we should not account for the
1399 * memory use of this mapping.
1400 */
1401 if (flags & MAP_NORESERVE) {
1402 /* We honor MAP_NORESERVE if allowed to overcommit */
1403 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1404 vm_flags |= VM_NORESERVE;
1405
1406 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1407 if (file && is_file_hugepages(file))
1408 vm_flags |= VM_NORESERVE;
1409 }
1410
1411 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1412 if (!IS_ERR_VALUE(addr) &&
1413 ((vm_flags & VM_LOCKED) ||
1414 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1415 *populate = len;
1416 return addr;
1417}
1418
1419unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1420 unsigned long prot, unsigned long flags,
1421 unsigned long fd, unsigned long pgoff)
1422{
1423 struct file *file = NULL;
1424 unsigned long retval;
1425
1426 if (!(flags & MAP_ANONYMOUS)) {
1427 audit_mmap_fd(fd, flags);
1428 file = fget(fd);
1429 if (!file)
1430 return -EBADF;
1431 if (is_file_hugepages(file)) {
1432 len = ALIGN(len, huge_page_size(hstate_file(file)));
1433 } else if (unlikely(flags & MAP_HUGETLB)) {
1434 retval = -EINVAL;
1435 goto out_fput;
1436 }
1437 } else if (flags & MAP_HUGETLB) {
1438 struct hstate *hs;
1439
1440 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1441 if (!hs)
1442 return -EINVAL;
1443
1444 len = ALIGN(len, huge_page_size(hs));
1445 /*
1446 * VM_NORESERVE is used because the reservations will be
1447 * taken when vm_ops->mmap() is called
1448 */
1449 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1450 VM_NORESERVE,
1451 HUGETLB_ANONHUGE_INODE,
1452 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1453 if (IS_ERR(file))
1454 return PTR_ERR(file);
1455 }
1456
1457 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1458out_fput:
1459 if (file)
1460 fput(file);
1461 return retval;
1462}
1463
1464SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1465 unsigned long, prot, unsigned long, flags,
1466 unsigned long, fd, unsigned long, pgoff)
1467{
1468 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1469}
1470
1471#ifdef __ARCH_WANT_SYS_OLD_MMAP
1472struct mmap_arg_struct {
1473 unsigned long addr;
1474 unsigned long len;
1475 unsigned long prot;
1476 unsigned long flags;
1477 unsigned long fd;
1478 unsigned long offset;
1479};
1480
1481SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1482{
1483 struct mmap_arg_struct a;
1484
1485 if (copy_from_user(&a, arg, sizeof(a)))
1486 return -EFAULT;
1487 if (offset_in_page(a.offset))
1488 return -EINVAL;
1489
1490 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1491 a.offset >> PAGE_SHIFT);
1492}
1493#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1494
1495/*
1496 * Some shared mappings will want the pages marked read-only
1497 * to track write events. If so, we'll downgrade vm_page_prot
1498 * to the private version (using protection_map[] without the
1499 * VM_SHARED bit).
1500 */
1501int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1502{
1503 vm_flags_t vm_flags = vma->vm_flags;
1504 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1505
1506 /* If it was private or non-writable, the write bit is already clear */
1507 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1508 return 0;
1509
1510 /* The backer wishes to know when pages are first written to? */
1511 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1512 return 1;
1513
1514 /* The open routine did something to the protections that pgprot_modify
1515 * won't preserve? */
1516 if (pgprot_val(vm_page_prot) !=
1517 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1518 return 0;
1519
1520 /*
1521 * Do we need to track softdirty? hugetlb does not support softdirty
1522 * tracking yet.
1523 */
1524 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1525 return 1;
1526
1527 /* Do we need write faults for uffd-wp tracking? */
1528 if (userfaultfd_wp(vma))
1529 return 1;
1530
1531 /* Specialty mapping? */
1532 if (vm_flags & VM_PFNMAP)
1533 return 0;
1534
1535 /* Can the mapping track the dirty pages? */
1536 return vma->vm_file && vma->vm_file->f_mapping &&
1537 mapping_can_writeback(vma->vm_file->f_mapping);
1538}
1539
1540/*
1541 * We account for memory if it's a private writeable mapping,
1542 * not hugepages and VM_NORESERVE wasn't set.
1543 */
1544static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1545{
1546 /*
1547 * hugetlb has its own accounting separate from the core VM
1548 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1549 */
1550 if (file && is_file_hugepages(file))
1551 return 0;
1552
1553 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1554}
1555
1556/**
1557 * unmapped_area() - Find an area between the low_limit and the high_limit with
1558 * the correct alignment and offset, all from @info. Note: current->mm is used
1559 * for the search.
1560 *
1561 * @info: The unmapped area information including the range (low_limit -
1562 * hight_limit), the alignment offset and mask.
1563 *
1564 * Return: A memory address or -ENOMEM.
1565 */
1566static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1567{
1568 unsigned long length, gap;
1569
1570 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1571
1572 /* Adjust search length to account for worst case alignment overhead */
1573 length = info->length + info->align_mask;
1574 if (length < info->length)
1575 return -ENOMEM;
1576
1577 if (mas_empty_area(&mas, info->low_limit, info->high_limit - 1,
1578 length))
1579 return -ENOMEM;
1580
1581 gap = mas.index;
1582 gap += (info->align_offset - gap) & info->align_mask;
1583 return gap;
1584}
1585
1586/**
1587 * unmapped_area_topdown() - Find an area between the low_limit and the
1588 * high_limit with * the correct alignment and offset at the highest available
1589 * address, all from @info. Note: current->mm is used for the search.
1590 *
1591 * @info: The unmapped area information including the range (low_limit -
1592 * hight_limit), the alignment offset and mask.
1593 *
1594 * Return: A memory address or -ENOMEM.
1595 */
1596static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1597{
1598 unsigned long length, gap;
1599
1600 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1601 /* Adjust search length to account for worst case alignment overhead */
1602 length = info->length + info->align_mask;
1603 if (length < info->length)
1604 return -ENOMEM;
1605
1606 if (mas_empty_area_rev(&mas, info->low_limit, info->high_limit - 1,
1607 length))
1608 return -ENOMEM;
1609
1610 gap = mas.last + 1 - info->length;
1611 gap -= (gap - info->align_offset) & info->align_mask;
1612 return gap;
1613}
1614
1615/*
1616 * Search for an unmapped address range.
1617 *
1618 * We are looking for a range that:
1619 * - does not intersect with any VMA;
1620 * - is contained within the [low_limit, high_limit) interval;
1621 * - is at least the desired size.
1622 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1623 */
1624unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1625{
1626 unsigned long addr;
1627
1628 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1629 addr = unmapped_area_topdown(info);
1630 else
1631 addr = unmapped_area(info);
1632
1633 trace_vm_unmapped_area(addr, info);
1634 return addr;
1635}
1636
1637/* Get an address range which is currently unmapped.
1638 * For shmat() with addr=0.
1639 *
1640 * Ugly calling convention alert:
1641 * Return value with the low bits set means error value,
1642 * ie
1643 * if (ret & ~PAGE_MASK)
1644 * error = ret;
1645 *
1646 * This function "knows" that -ENOMEM has the bits set.
1647 */
1648unsigned long
1649generic_get_unmapped_area(struct file *filp, unsigned long addr,
1650 unsigned long len, unsigned long pgoff,
1651 unsigned long flags)
1652{
1653 struct mm_struct *mm = current->mm;
1654 struct vm_area_struct *vma, *prev;
1655 struct vm_unmapped_area_info info;
1656 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1657
1658 if (len > mmap_end - mmap_min_addr)
1659 return -ENOMEM;
1660
1661 if (flags & MAP_FIXED)
1662 return addr;
1663
1664 if (addr) {
1665 addr = PAGE_ALIGN(addr);
1666 vma = find_vma_prev(mm, addr, &prev);
1667 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1668 (!vma || addr + len <= vm_start_gap(vma)) &&
1669 (!prev || addr >= vm_end_gap(prev)))
1670 return addr;
1671 }
1672
1673 info.flags = 0;
1674 info.length = len;
1675 info.low_limit = mm->mmap_base;
1676 info.high_limit = mmap_end;
1677 info.align_mask = 0;
1678 info.align_offset = 0;
1679 return vm_unmapped_area(&info);
1680}
1681
1682#ifndef HAVE_ARCH_UNMAPPED_AREA
1683unsigned long
1684arch_get_unmapped_area(struct file *filp, unsigned long addr,
1685 unsigned long len, unsigned long pgoff,
1686 unsigned long flags)
1687{
1688 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1689}
1690#endif
1691
1692/*
1693 * This mmap-allocator allocates new areas top-down from below the
1694 * stack's low limit (the base):
1695 */
1696unsigned long
1697generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1698 unsigned long len, unsigned long pgoff,
1699 unsigned long flags)
1700{
1701 struct vm_area_struct *vma, *prev;
1702 struct mm_struct *mm = current->mm;
1703 struct vm_unmapped_area_info info;
1704 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1705
1706 /* requested length too big for entire address space */
1707 if (len > mmap_end - mmap_min_addr)
1708 return -ENOMEM;
1709
1710 if (flags & MAP_FIXED)
1711 return addr;
1712
1713 /* requesting a specific address */
1714 if (addr) {
1715 addr = PAGE_ALIGN(addr);
1716 vma = find_vma_prev(mm, addr, &prev);
1717 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1718 (!vma || addr + len <= vm_start_gap(vma)) &&
1719 (!prev || addr >= vm_end_gap(prev)))
1720 return addr;
1721 }
1722
1723 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1724 info.length = len;
1725 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1726 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1727 info.align_mask = 0;
1728 info.align_offset = 0;
1729 addr = vm_unmapped_area(&info);
1730
1731 /*
1732 * A failed mmap() very likely causes application failure,
1733 * so fall back to the bottom-up function here. This scenario
1734 * can happen with large stack limits and large mmap()
1735 * allocations.
1736 */
1737 if (offset_in_page(addr)) {
1738 VM_BUG_ON(addr != -ENOMEM);
1739 info.flags = 0;
1740 info.low_limit = TASK_UNMAPPED_BASE;
1741 info.high_limit = mmap_end;
1742 addr = vm_unmapped_area(&info);
1743 }
1744
1745 return addr;
1746}
1747
1748#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1749unsigned long
1750arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1751 unsigned long len, unsigned long pgoff,
1752 unsigned long flags)
1753{
1754 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1755}
1756#endif
1757
1758unsigned long
1759get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1760 unsigned long pgoff, unsigned long flags)
1761{
1762 unsigned long (*get_area)(struct file *, unsigned long,
1763 unsigned long, unsigned long, unsigned long);
1764
1765 unsigned long error = arch_mmap_check(addr, len, flags);
1766 if (error)
1767 return error;
1768
1769 /* Careful about overflows.. */
1770 if (len > TASK_SIZE)
1771 return -ENOMEM;
1772
1773 get_area = current->mm->get_unmapped_area;
1774 if (file) {
1775 if (file->f_op->get_unmapped_area)
1776 get_area = file->f_op->get_unmapped_area;
1777 } else if (flags & MAP_SHARED) {
1778 /*
1779 * mmap_region() will call shmem_zero_setup() to create a file,
1780 * so use shmem's get_unmapped_area in case it can be huge.
1781 * do_mmap() will clear pgoff, so match alignment.
1782 */
1783 pgoff = 0;
1784 get_area = shmem_get_unmapped_area;
1785 }
1786
1787 addr = get_area(file, addr, len, pgoff, flags);
1788 if (IS_ERR_VALUE(addr))
1789 return addr;
1790
1791 if (addr > TASK_SIZE - len)
1792 return -ENOMEM;
1793 if (offset_in_page(addr))
1794 return -EINVAL;
1795
1796 error = security_mmap_addr(addr);
1797 return error ? error : addr;
1798}
1799
1800EXPORT_SYMBOL(get_unmapped_area);
1801
1802/**
1803 * find_vma_intersection() - Look up the first VMA which intersects the interval
1804 * @mm: The process address space.
1805 * @start_addr: The inclusive start user address.
1806 * @end_addr: The exclusive end user address.
1807 *
1808 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1809 * start_addr < end_addr.
1810 */
1811struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1812 unsigned long start_addr,
1813 unsigned long end_addr)
1814{
1815 unsigned long index = start_addr;
1816
1817 mmap_assert_locked(mm);
1818 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1819}
1820EXPORT_SYMBOL(find_vma_intersection);
1821
1822/**
1823 * find_vma() - Find the VMA for a given address, or the next VMA.
1824 * @mm: The mm_struct to check
1825 * @addr: The address
1826 *
1827 * Returns: The VMA associated with addr, or the next VMA.
1828 * May return %NULL in the case of no VMA at addr or above.
1829 */
1830struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1831{
1832 unsigned long index = addr;
1833
1834 mmap_assert_locked(mm);
1835 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1836}
1837EXPORT_SYMBOL(find_vma);
1838
1839/**
1840 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1841 * set %pprev to the previous VMA, if any.
1842 * @mm: The mm_struct to check
1843 * @addr: The address
1844 * @pprev: The pointer to set to the previous VMA
1845 *
1846 * Note that RCU lock is missing here since the external mmap_lock() is used
1847 * instead.
1848 *
1849 * Returns: The VMA associated with @addr, or the next vma.
1850 * May return %NULL in the case of no vma at addr or above.
1851 */
1852struct vm_area_struct *
1853find_vma_prev(struct mm_struct *mm, unsigned long addr,
1854 struct vm_area_struct **pprev)
1855{
1856 struct vm_area_struct *vma;
1857 MA_STATE(mas, &mm->mm_mt, addr, addr);
1858
1859 vma = mas_walk(&mas);
1860 *pprev = mas_prev(&mas, 0);
1861 if (!vma)
1862 vma = mas_next(&mas, ULONG_MAX);
1863 return vma;
1864}
1865
1866/*
1867 * Verify that the stack growth is acceptable and
1868 * update accounting. This is shared with both the
1869 * grow-up and grow-down cases.
1870 */
1871static int acct_stack_growth(struct vm_area_struct *vma,
1872 unsigned long size, unsigned long grow)
1873{
1874 struct mm_struct *mm = vma->vm_mm;
1875 unsigned long new_start;
1876
1877 /* address space limit tests */
1878 if (!may_expand_vm(mm, vma->vm_flags, grow))
1879 return -ENOMEM;
1880
1881 /* Stack limit test */
1882 if (size > rlimit(RLIMIT_STACK))
1883 return -ENOMEM;
1884
1885 /* mlock limit tests */
1886 if (mlock_future_check(mm, vma->vm_flags, grow << PAGE_SHIFT))
1887 return -ENOMEM;
1888
1889 /* Check to ensure the stack will not grow into a hugetlb-only region */
1890 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1891 vma->vm_end - size;
1892 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1893 return -EFAULT;
1894
1895 /*
1896 * Overcommit.. This must be the final test, as it will
1897 * update security statistics.
1898 */
1899 if (security_vm_enough_memory_mm(mm, grow))
1900 return -ENOMEM;
1901
1902 return 0;
1903}
1904
1905#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1906/*
1907 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1908 * vma is the last one with address > vma->vm_end. Have to extend vma.
1909 */
1910int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1911{
1912 struct mm_struct *mm = vma->vm_mm;
1913 struct vm_area_struct *next;
1914 unsigned long gap_addr;
1915 int error = 0;
1916 MA_STATE(mas, &mm->mm_mt, 0, 0);
1917
1918 if (!(vma->vm_flags & VM_GROWSUP))
1919 return -EFAULT;
1920
1921 /* Guard against exceeding limits of the address space. */
1922 address &= PAGE_MASK;
1923 if (address >= (TASK_SIZE & PAGE_MASK))
1924 return -ENOMEM;
1925 address += PAGE_SIZE;
1926
1927 /* Enforce stack_guard_gap */
1928 gap_addr = address + stack_guard_gap;
1929
1930 /* Guard against overflow */
1931 if (gap_addr < address || gap_addr > TASK_SIZE)
1932 gap_addr = TASK_SIZE;
1933
1934 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1935 if (next && vma_is_accessible(next)) {
1936 if (!(next->vm_flags & VM_GROWSUP))
1937 return -ENOMEM;
1938 /* Check that both stack segments have the same anon_vma? */
1939 }
1940
1941 if (mas_preallocate(&mas, vma, GFP_KERNEL))
1942 return -ENOMEM;
1943
1944 /* We must make sure the anon_vma is allocated. */
1945 if (unlikely(anon_vma_prepare(vma))) {
1946 mas_destroy(&mas);
1947 return -ENOMEM;
1948 }
1949
1950 /*
1951 * vma->vm_start/vm_end cannot change under us because the caller
1952 * is required to hold the mmap_lock in read mode. We need the
1953 * anon_vma lock to serialize against concurrent expand_stacks.
1954 */
1955 anon_vma_lock_write(vma->anon_vma);
1956
1957 /* Somebody else might have raced and expanded it already */
1958 if (address > vma->vm_end) {
1959 unsigned long size, grow;
1960
1961 size = address - vma->vm_start;
1962 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1963
1964 error = -ENOMEM;
1965 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1966 error = acct_stack_growth(vma, size, grow);
1967 if (!error) {
1968 /*
1969 * We only hold a shared mmap_lock lock here, so
1970 * we need to protect against concurrent vma
1971 * expansions. anon_vma_lock_write() doesn't
1972 * help here, as we don't guarantee that all
1973 * growable vmas in a mm share the same root
1974 * anon vma. So, we reuse mm->page_table_lock
1975 * to guard against concurrent vma expansions.
1976 */
1977 spin_lock(&mm->page_table_lock);
1978 if (vma->vm_flags & VM_LOCKED)
1979 mm->locked_vm += grow;
1980 vm_stat_account(mm, vma->vm_flags, grow);
1981 anon_vma_interval_tree_pre_update_vma(vma);
1982 vma->vm_end = address;
1983 /* Overwrite old entry in mtree. */
1984 vma_mas_store(vma, &mas);
1985 anon_vma_interval_tree_post_update_vma(vma);
1986 spin_unlock(&mm->page_table_lock);
1987
1988 perf_event_mmap(vma);
1989 }
1990 }
1991 }
1992 anon_vma_unlock_write(vma->anon_vma);
1993 khugepaged_enter_vma(vma, vma->vm_flags);
1994 mas_destroy(&mas);
1995 return error;
1996}
1997#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1998
1999/*
2000 * vma is the first one with address < vma->vm_start. Have to extend vma.
2001 */
2002int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2003{
2004 struct mm_struct *mm = vma->vm_mm;
2005 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2006 struct vm_area_struct *prev;
2007 int error = 0;
2008
2009 address &= PAGE_MASK;
2010 if (address < mmap_min_addr)
2011 return -EPERM;
2012
2013 /* Enforce stack_guard_gap */
2014 prev = mas_prev(&mas, 0);
2015 /* Check that both stack segments have the same anon_vma? */
2016 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2017 vma_is_accessible(prev)) {
2018 if (address - prev->vm_end < stack_guard_gap)
2019 return -ENOMEM;
2020 }
2021
2022 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2023 return -ENOMEM;
2024
2025 /* We must make sure the anon_vma is allocated. */
2026 if (unlikely(anon_vma_prepare(vma))) {
2027 mas_destroy(&mas);
2028 return -ENOMEM;
2029 }
2030
2031 /*
2032 * vma->vm_start/vm_end cannot change under us because the caller
2033 * is required to hold the mmap_lock in read mode. We need the
2034 * anon_vma lock to serialize against concurrent expand_stacks.
2035 */
2036 anon_vma_lock_write(vma->anon_vma);
2037
2038 /* Somebody else might have raced and expanded it already */
2039 if (address < vma->vm_start) {
2040 unsigned long size, grow;
2041
2042 size = vma->vm_end - address;
2043 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2044
2045 error = -ENOMEM;
2046 if (grow <= vma->vm_pgoff) {
2047 error = acct_stack_growth(vma, size, grow);
2048 if (!error) {
2049 /*
2050 * We only hold a shared mmap_lock lock here, so
2051 * we need to protect against concurrent vma
2052 * expansions. anon_vma_lock_write() doesn't
2053 * help here, as we don't guarantee that all
2054 * growable vmas in a mm share the same root
2055 * anon vma. So, we reuse mm->page_table_lock
2056 * to guard against concurrent vma expansions.
2057 */
2058 spin_lock(&mm->page_table_lock);
2059 if (vma->vm_flags & VM_LOCKED)
2060 mm->locked_vm += grow;
2061 vm_stat_account(mm, vma->vm_flags, grow);
2062 anon_vma_interval_tree_pre_update_vma(vma);
2063 vma->vm_start = address;
2064 vma->vm_pgoff -= grow;
2065 /* Overwrite old entry in mtree. */
2066 vma_mas_store(vma, &mas);
2067 anon_vma_interval_tree_post_update_vma(vma);
2068 spin_unlock(&mm->page_table_lock);
2069
2070 perf_event_mmap(vma);
2071 }
2072 }
2073 }
2074 anon_vma_unlock_write(vma->anon_vma);
2075 khugepaged_enter_vma(vma, vma->vm_flags);
2076 mas_destroy(&mas);
2077 return error;
2078}
2079
2080/* enforced gap between the expanding stack and other mappings. */
2081unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2082
2083static int __init cmdline_parse_stack_guard_gap(char *p)
2084{
2085 unsigned long val;
2086 char *endptr;
2087
2088 val = simple_strtoul(p, &endptr, 10);
2089 if (!*endptr)
2090 stack_guard_gap = val << PAGE_SHIFT;
2091
2092 return 1;
2093}
2094__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2095
2096#ifdef CONFIG_STACK_GROWSUP
2097int expand_stack(struct vm_area_struct *vma, unsigned long address)
2098{
2099 return expand_upwards(vma, address);
2100}
2101
2102struct vm_area_struct *
2103find_extend_vma(struct mm_struct *mm, unsigned long addr)
2104{
2105 struct vm_area_struct *vma, *prev;
2106
2107 addr &= PAGE_MASK;
2108 vma = find_vma_prev(mm, addr, &prev);
2109 if (vma && (vma->vm_start <= addr))
2110 return vma;
2111 if (!prev || expand_stack(prev, addr))
2112 return NULL;
2113 if (prev->vm_flags & VM_LOCKED)
2114 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2115 return prev;
2116}
2117#else
2118int expand_stack(struct vm_area_struct *vma, unsigned long address)
2119{
2120 return expand_downwards(vma, address);
2121}
2122
2123struct vm_area_struct *
2124find_extend_vma(struct mm_struct *mm, unsigned long addr)
2125{
2126 struct vm_area_struct *vma;
2127 unsigned long start;
2128
2129 addr &= PAGE_MASK;
2130 vma = find_vma(mm, addr);
2131 if (!vma)
2132 return NULL;
2133 if (vma->vm_start <= addr)
2134 return vma;
2135 if (!(vma->vm_flags & VM_GROWSDOWN))
2136 return NULL;
2137 start = vma->vm_start;
2138 if (expand_stack(vma, addr))
2139 return NULL;
2140 if (vma->vm_flags & VM_LOCKED)
2141 populate_vma_page_range(vma, addr, start, NULL);
2142 return vma;
2143}
2144#endif
2145
2146EXPORT_SYMBOL_GPL(find_extend_vma);
2147
2148/*
2149 * Ok - we have the memory areas we should free on a maple tree so release them,
2150 * and do the vma updates.
2151 *
2152 * Called with the mm semaphore held.
2153 */
2154static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2155{
2156 unsigned long nr_accounted = 0;
2157 struct vm_area_struct *vma;
2158
2159 /* Update high watermark before we lower total_vm */
2160 update_hiwater_vm(mm);
2161 mas_for_each(mas, vma, ULONG_MAX) {
2162 long nrpages = vma_pages(vma);
2163
2164 if (vma->vm_flags & VM_ACCOUNT)
2165 nr_accounted += nrpages;
2166 vm_stat_account(mm, vma->vm_flags, -nrpages);
2167 remove_vma(vma);
2168 }
2169 vm_unacct_memory(nr_accounted);
2170 validate_mm(mm);
2171}
2172
2173/*
2174 * Get rid of page table information in the indicated region.
2175 *
2176 * Called with the mm semaphore held.
2177 */
2178static void unmap_region(struct mm_struct *mm, struct maple_tree *mt,
2179 struct vm_area_struct *vma, struct vm_area_struct *prev,
2180 struct vm_area_struct *next,
2181 unsigned long start, unsigned long end)
2182{
2183 struct mmu_gather tlb;
2184
2185 lru_add_drain();
2186 tlb_gather_mmu(&tlb, mm);
2187 update_hiwater_rss(mm);
2188 unmap_vmas(&tlb, mt, vma, start, end);
2189 free_pgtables(&tlb, mt, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2190 next ? next->vm_start : USER_PGTABLES_CEILING);
2191 tlb_finish_mmu(&tlb);
2192}
2193
2194/*
2195 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2196 * has already been checked or doesn't make sense to fail.
2197 */
2198int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2199 unsigned long addr, int new_below)
2200{
2201 struct vm_area_struct *new;
2202 int err;
2203 validate_mm_mt(mm);
2204
2205 if (vma->vm_ops && vma->vm_ops->may_split) {
2206 err = vma->vm_ops->may_split(vma, addr);
2207 if (err)
2208 return err;
2209 }
2210
2211 new = vm_area_dup(vma);
2212 if (!new)
2213 return -ENOMEM;
2214
2215 if (new_below)
2216 new->vm_end = addr;
2217 else {
2218 new->vm_start = addr;
2219 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2220 }
2221
2222 err = vma_dup_policy(vma, new);
2223 if (err)
2224 goto out_free_vma;
2225
2226 err = anon_vma_clone(new, vma);
2227 if (err)
2228 goto out_free_mpol;
2229
2230 if (new->vm_file)
2231 get_file(new->vm_file);
2232
2233 if (new->vm_ops && new->vm_ops->open)
2234 new->vm_ops->open(new);
2235
2236 if (new_below)
2237 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2238 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2239 else
2240 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2241
2242 /* Success. */
2243 if (!err)
2244 return 0;
2245
2246 /* Avoid vm accounting in close() operation */
2247 new->vm_start = new->vm_end;
2248 new->vm_pgoff = 0;
2249 /* Clean everything up if vma_adjust failed. */
2250 if (new->vm_ops && new->vm_ops->close)
2251 new->vm_ops->close(new);
2252 if (new->vm_file)
2253 fput(new->vm_file);
2254 unlink_anon_vmas(new);
2255 out_free_mpol:
2256 mpol_put(vma_policy(new));
2257 out_free_vma:
2258 vm_area_free(new);
2259 validate_mm_mt(mm);
2260 return err;
2261}
2262
2263/*
2264 * Split a vma into two pieces at address 'addr', a new vma is allocated
2265 * either for the first part or the tail.
2266 */
2267int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2268 unsigned long addr, int new_below)
2269{
2270 if (mm->map_count >= sysctl_max_map_count)
2271 return -ENOMEM;
2272
2273 return __split_vma(mm, vma, addr, new_below);
2274}
2275
2276static inline int munmap_sidetree(struct vm_area_struct *vma,
2277 struct ma_state *mas_detach)
2278{
2279 mas_set_range(mas_detach, vma->vm_start, vma->vm_end - 1);
2280 if (mas_store_gfp(mas_detach, vma, GFP_KERNEL))
2281 return -ENOMEM;
2282
2283 if (vma->vm_flags & VM_LOCKED)
2284 vma->vm_mm->locked_vm -= vma_pages(vma);
2285
2286 return 0;
2287}
2288
2289/*
2290 * do_mas_align_munmap() - munmap the aligned region from @start to @end.
2291 * @mas: The maple_state, ideally set up to alter the correct tree location.
2292 * @vma: The starting vm_area_struct
2293 * @mm: The mm_struct
2294 * @start: The aligned start address to munmap.
2295 * @end: The aligned end address to munmap.
2296 * @uf: The userfaultfd list_head
2297 * @downgrade: Set to true to attempt a write downgrade of the mmap_lock
2298 *
2299 * If @downgrade is true, check return code for potential release of the lock.
2300 */
2301static int
2302do_mas_align_munmap(struct ma_state *mas, struct vm_area_struct *vma,
2303 struct mm_struct *mm, unsigned long start,
2304 unsigned long end, struct list_head *uf, bool downgrade)
2305{
2306 struct vm_area_struct *prev, *next = NULL;
2307 struct maple_tree mt_detach;
2308 int count = 0;
2309 int error = -ENOMEM;
2310 MA_STATE(mas_detach, &mt_detach, 0, 0);
2311 mt_init_flags(&mt_detach, MT_FLAGS_LOCK_EXTERN);
2312 mt_set_external_lock(&mt_detach, &mm->mmap_lock);
2313
2314 if (mas_preallocate(mas, vma, GFP_KERNEL))
2315 return -ENOMEM;
2316
2317 mas->last = end - 1;
2318 /*
2319 * If we need to split any vma, do it now to save pain later.
2320 *
2321 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2322 * unmapped vm_area_struct will remain in use: so lower split_vma
2323 * places tmp vma above, and higher split_vma places tmp vma below.
2324 */
2325
2326 /* Does it split the first one? */
2327 if (start > vma->vm_start) {
2328
2329 /*
2330 * Make sure that map_count on return from munmap() will
2331 * not exceed its limit; but let map_count go just above
2332 * its limit temporarily, to help free resources as expected.
2333 */
2334 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2335 goto map_count_exceeded;
2336
2337 /*
2338 * mas_pause() is not needed since mas->index needs to be set
2339 * differently than vma->vm_end anyways.
2340 */
2341 error = __split_vma(mm, vma, start, 0);
2342 if (error)
2343 goto start_split_failed;
2344
2345 mas_set(mas, start);
2346 vma = mas_walk(mas);
2347 }
2348
2349 prev = mas_prev(mas, 0);
2350 if (unlikely((!prev)))
2351 mas_set(mas, start);
2352
2353 /*
2354 * Detach a range of VMAs from the mm. Using next as a temp variable as
2355 * it is always overwritten.
2356 */
2357 mas_for_each(mas, next, end - 1) {
2358 /* Does it split the end? */
2359 if (next->vm_end > end) {
2360 struct vm_area_struct *split;
2361
2362 error = __split_vma(mm, next, end, 1);
2363 if (error)
2364 goto end_split_failed;
2365
2366 mas_set(mas, end);
2367 split = mas_prev(mas, 0);
2368 error = munmap_sidetree(split, &mas_detach);
2369 if (error)
2370 goto munmap_sidetree_failed;
2371
2372 count++;
2373 if (vma == next)
2374 vma = split;
2375 break;
2376 }
2377 error = munmap_sidetree(next, &mas_detach);
2378 if (error)
2379 goto munmap_sidetree_failed;
2380
2381 count++;
2382#ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2383 BUG_ON(next->vm_start < start);
2384 BUG_ON(next->vm_start > end);
2385#endif
2386 }
2387
2388 if (!next)
2389 next = mas_next(mas, ULONG_MAX);
2390
2391 if (unlikely(uf)) {
2392 /*
2393 * If userfaultfd_unmap_prep returns an error the vmas
2394 * will remain split, but userland will get a
2395 * highly unexpected error anyway. This is no
2396 * different than the case where the first of the two
2397 * __split_vma fails, but we don't undo the first
2398 * split, despite we could. This is unlikely enough
2399 * failure that it's not worth optimizing it for.
2400 */
2401 error = userfaultfd_unmap_prep(mm, start, end, uf);
2402
2403 if (error)
2404 goto userfaultfd_error;
2405 }
2406
2407 /* Point of no return */
2408 mas_set_range(mas, start, end - 1);
2409#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2410 /* Make sure no VMAs are about to be lost. */
2411 {
2412 MA_STATE(test, &mt_detach, start, end - 1);
2413 struct vm_area_struct *vma_mas, *vma_test;
2414 int test_count = 0;
2415
2416 rcu_read_lock();
2417 vma_test = mas_find(&test, end - 1);
2418 mas_for_each(mas, vma_mas, end - 1) {
2419 BUG_ON(vma_mas != vma_test);
2420 test_count++;
2421 vma_test = mas_next(&test, end - 1);
2422 }
2423 rcu_read_unlock();
2424 BUG_ON(count != test_count);
2425 mas_set_range(mas, start, end - 1);
2426 }
2427#endif
2428 mas_store_prealloc(mas, NULL);
2429 mm->map_count -= count;
2430 /*
2431 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2432 * VM_GROWSUP VMA. Such VMAs can change their size under
2433 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2434 */
2435 if (downgrade) {
2436 if (next && (next->vm_flags & VM_GROWSDOWN))
2437 downgrade = false;
2438 else if (prev && (prev->vm_flags & VM_GROWSUP))
2439 downgrade = false;
2440 else
2441 mmap_write_downgrade(mm);
2442 }
2443
2444 unmap_region(mm, &mt_detach, vma, prev, next, start, end);
2445 /* Statistics and freeing VMAs */
2446 mas_set(&mas_detach, start);
2447 remove_mt(mm, &mas_detach);
2448 __mt_destroy(&mt_detach);
2449
2450
2451 validate_mm(mm);
2452 return downgrade ? 1 : 0;
2453
2454userfaultfd_error:
2455munmap_sidetree_failed:
2456end_split_failed:
2457 __mt_destroy(&mt_detach);
2458start_split_failed:
2459map_count_exceeded:
2460 mas_destroy(mas);
2461 return error;
2462}
2463
2464/*
2465 * do_mas_munmap() - munmap a given range.
2466 * @mas: The maple state
2467 * @mm: The mm_struct
2468 * @start: The start address to munmap
2469 * @len: The length of the range to munmap
2470 * @uf: The userfaultfd list_head
2471 * @downgrade: set to true if the user wants to attempt to write_downgrade the
2472 * mmap_lock
2473 *
2474 * This function takes a @mas that is either pointing to the previous VMA or set
2475 * to MA_START and sets it up to remove the mapping(s). The @len will be
2476 * aligned and any arch_unmap work will be preformed.
2477 *
2478 * Returns: -EINVAL on failure, 1 on success and unlock, 0 otherwise.
2479 */
2480int do_mas_munmap(struct ma_state *mas, struct mm_struct *mm,
2481 unsigned long start, size_t len, struct list_head *uf,
2482 bool downgrade)
2483{
2484 unsigned long end;
2485 struct vm_area_struct *vma;
2486
2487 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2488 return -EINVAL;
2489
2490 end = start + PAGE_ALIGN(len);
2491 if (end == start)
2492 return -EINVAL;
2493
2494 /* arch_unmap() might do unmaps itself. */
2495 arch_unmap(mm, start, end);
2496
2497 /* Find the first overlapping VMA */
2498 vma = mas_find(mas, end - 1);
2499 if (!vma)
2500 return 0;
2501
2502 return do_mas_align_munmap(mas, vma, mm, start, end, uf, downgrade);
2503}
2504
2505/* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2506 * @mm: The mm_struct
2507 * @start: The start address to munmap
2508 * @len: The length to be munmapped.
2509 * @uf: The userfaultfd list_head
2510 */
2511int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2512 struct list_head *uf)
2513{
2514 MA_STATE(mas, &mm->mm_mt, start, start);
2515
2516 return do_mas_munmap(&mas, mm, start, len, uf, false);
2517}
2518
2519unsigned long mmap_region(struct file *file, unsigned long addr,
2520 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2521 struct list_head *uf)
2522{
2523 struct mm_struct *mm = current->mm;
2524 struct vm_area_struct *vma = NULL;
2525 struct vm_area_struct *next, *prev, *merge;
2526 pgoff_t pglen = len >> PAGE_SHIFT;
2527 unsigned long charged = 0;
2528 unsigned long end = addr + len;
2529 unsigned long merge_start = addr, merge_end = end;
2530 pgoff_t vm_pgoff;
2531 int error;
2532 MA_STATE(mas, &mm->mm_mt, addr, end - 1);
2533
2534 /* Check against address space limit. */
2535 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2536 unsigned long nr_pages;
2537
2538 /*
2539 * MAP_FIXED may remove pages of mappings that intersects with
2540 * requested mapping. Account for the pages it would unmap.
2541 */
2542 nr_pages = count_vma_pages_range(mm, addr, end);
2543
2544 if (!may_expand_vm(mm, vm_flags,
2545 (len >> PAGE_SHIFT) - nr_pages))
2546 return -ENOMEM;
2547 }
2548
2549 /* Unmap any existing mapping in the area */
2550 if (do_mas_munmap(&mas, mm, addr, len, uf, false))
2551 return -ENOMEM;
2552
2553 /*
2554 * Private writable mapping: check memory availability
2555 */
2556 if (accountable_mapping(file, vm_flags)) {
2557 charged = len >> PAGE_SHIFT;
2558 if (security_vm_enough_memory_mm(mm, charged))
2559 return -ENOMEM;
2560 vm_flags |= VM_ACCOUNT;
2561 }
2562
2563 next = mas_next(&mas, ULONG_MAX);
2564 prev = mas_prev(&mas, 0);
2565 if (vm_flags & VM_SPECIAL)
2566 goto cannot_expand;
2567
2568 /* Attempt to expand an old mapping */
2569 /* Check next */
2570 if (next && next->vm_start == end && !vma_policy(next) &&
2571 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2572 NULL_VM_UFFD_CTX, NULL)) {
2573 merge_end = next->vm_end;
2574 vma = next;
2575 vm_pgoff = next->vm_pgoff - pglen;
2576 }
2577
2578 /* Check prev */
2579 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2580 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2581 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2582 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2583 NULL_VM_UFFD_CTX, NULL))) {
2584 merge_start = prev->vm_start;
2585 vma = prev;
2586 vm_pgoff = prev->vm_pgoff;
2587 }
2588
2589
2590 /* Actually expand, if possible */
2591 if (vma &&
2592 !vma_expand(&mas, vma, merge_start, merge_end, vm_pgoff, next)) {
2593 khugepaged_enter_vma(vma, vm_flags);
2594 goto expanded;
2595 }
2596
2597 mas.index = addr;
2598 mas.last = end - 1;
2599cannot_expand:
2600 /*
2601 * Determine the object being mapped and call the appropriate
2602 * specific mapper. the address has already been validated, but
2603 * not unmapped, but the maps are removed from the list.
2604 */
2605 vma = vm_area_alloc(mm);
2606 if (!vma) {
2607 error = -ENOMEM;
2608 goto unacct_error;
2609 }
2610
2611 vma->vm_start = addr;
2612 vma->vm_end = end;
2613 vma->vm_flags = vm_flags;
2614 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2615 vma->vm_pgoff = pgoff;
2616
2617 if (file) {
2618 if (vm_flags & VM_SHARED) {
2619 error = mapping_map_writable(file->f_mapping);
2620 if (error)
2621 goto free_vma;
2622 }
2623
2624 vma->vm_file = get_file(file);
2625 error = call_mmap(file, vma);
2626 if (error)
2627 goto unmap_and_free_vma;
2628
2629 /*
2630 * Expansion is handled above, merging is handled below.
2631 * Drivers should not alter the address of the VMA.
2632 */
2633 if (WARN_ON((addr != vma->vm_start))) {
2634 error = -EINVAL;
2635 goto close_and_free_vma;
2636 }
2637 mas_reset(&mas);
2638
2639 /*
2640 * If vm_flags changed after call_mmap(), we should try merge
2641 * vma again as we may succeed this time.
2642 */
2643 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2644 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
2645 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
2646 if (merge) {
2647 /*
2648 * ->mmap() can change vma->vm_file and fput
2649 * the original file. So fput the vma->vm_file
2650 * here or we would add an extra fput for file
2651 * and cause general protection fault
2652 * ultimately.
2653 */
2654 fput(vma->vm_file);
2655 vm_area_free(vma);
2656 vma = merge;
2657 /* Update vm_flags to pick up the change. */
2658 vm_flags = vma->vm_flags;
2659 goto unmap_writable;
2660 }
2661 }
2662
2663 vm_flags = vma->vm_flags;
2664 } else if (vm_flags & VM_SHARED) {
2665 error = shmem_zero_setup(vma);
2666 if (error)
2667 goto free_vma;
2668 } else {
2669 vma_set_anonymous(vma);
2670 }
2671
2672 /* Allow architectures to sanity-check the vm_flags */
2673 if (!arch_validate_flags(vma->vm_flags)) {
2674 error = -EINVAL;
2675 if (file)
2676 goto close_and_free_vma;
2677 else if (vma->vm_file)
2678 goto unmap_and_free_vma;
2679 else
2680 goto free_vma;
2681 }
2682
2683 if (mas_preallocate(&mas, vma, GFP_KERNEL)) {
2684 error = -ENOMEM;
2685 if (file)
2686 goto close_and_free_vma;
2687 else if (vma->vm_file)
2688 goto unmap_and_free_vma;
2689 else
2690 goto free_vma;
2691 }
2692
2693 if (vma->vm_file)
2694 i_mmap_lock_write(vma->vm_file->f_mapping);
2695
2696 vma_mas_store(vma, &mas);
2697 mm->map_count++;
2698 if (vma->vm_file) {
2699 if (vma->vm_flags & VM_SHARED)
2700 mapping_allow_writable(vma->vm_file->f_mapping);
2701
2702 flush_dcache_mmap_lock(vma->vm_file->f_mapping);
2703 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
2704 flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
2705 i_mmap_unlock_write(vma->vm_file->f_mapping);
2706 }
2707
2708 /*
2709 * vma_merge() calls khugepaged_enter_vma() either, the below
2710 * call covers the non-merge case.
2711 */
2712 khugepaged_enter_vma(vma, vma->vm_flags);
2713
2714 /* Once vma denies write, undo our temporary denial count */
2715unmap_writable:
2716 if (file && vm_flags & VM_SHARED)
2717 mapping_unmap_writable(file->f_mapping);
2718 file = vma->vm_file;
2719expanded:
2720 perf_event_mmap(vma);
2721
2722 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2723 if (vm_flags & VM_LOCKED) {
2724 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2725 is_vm_hugetlb_page(vma) ||
2726 vma == get_gate_vma(current->mm))
2727 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
2728 else
2729 mm->locked_vm += (len >> PAGE_SHIFT);
2730 }
2731
2732 if (file)
2733 uprobe_mmap(vma);
2734
2735 /*
2736 * New (or expanded) vma always get soft dirty status.
2737 * Otherwise user-space soft-dirty page tracker won't
2738 * be able to distinguish situation when vma area unmapped,
2739 * then new mapped in-place (which must be aimed as
2740 * a completely new data area).
2741 */
2742 vma->vm_flags |= VM_SOFTDIRTY;
2743
2744 vma_set_page_prot(vma);
2745
2746 validate_mm(mm);
2747 return addr;
2748
2749close_and_free_vma:
2750 if (vma->vm_ops && vma->vm_ops->close)
2751 vma->vm_ops->close(vma);
2752unmap_and_free_vma:
2753 fput(vma->vm_file);
2754 vma->vm_file = NULL;
2755
2756 /* Undo any partial mapping done by a device driver. */
2757 unmap_region(mm, mas.tree, vma, prev, next, vma->vm_start, vma->vm_end);
2758 if (file && (vm_flags & VM_SHARED))
2759 mapping_unmap_writable(file->f_mapping);
2760free_vma:
2761 vm_area_free(vma);
2762unacct_error:
2763 if (charged)
2764 vm_unacct_memory(charged);
2765 validate_mm(mm);
2766 return error;
2767}
2768
2769static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2770{
2771 int ret;
2772 struct mm_struct *mm = current->mm;
2773 LIST_HEAD(uf);
2774 MA_STATE(mas, &mm->mm_mt, start, start);
2775
2776 if (mmap_write_lock_killable(mm))
2777 return -EINTR;
2778
2779 ret = do_mas_munmap(&mas, mm, start, len, &uf, downgrade);
2780 /*
2781 * Returning 1 indicates mmap_lock is downgraded.
2782 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2783 * it to 0 before return.
2784 */
2785 if (ret == 1) {
2786 mmap_read_unlock(mm);
2787 ret = 0;
2788 } else
2789 mmap_write_unlock(mm);
2790
2791 userfaultfd_unmap_complete(mm, &uf);
2792 return ret;
2793}
2794
2795int vm_munmap(unsigned long start, size_t len)
2796{
2797 return __vm_munmap(start, len, false);
2798}
2799EXPORT_SYMBOL(vm_munmap);
2800
2801SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2802{
2803 addr = untagged_addr(addr);
2804 return __vm_munmap(addr, len, true);
2805}
2806
2807
2808/*
2809 * Emulation of deprecated remap_file_pages() syscall.
2810 */
2811SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2812 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2813{
2814
2815 struct mm_struct *mm = current->mm;
2816 struct vm_area_struct *vma;
2817 unsigned long populate = 0;
2818 unsigned long ret = -EINVAL;
2819 struct file *file;
2820
2821 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
2822 current->comm, current->pid);
2823
2824 if (prot)
2825 return ret;
2826 start = start & PAGE_MASK;
2827 size = size & PAGE_MASK;
2828
2829 if (start + size <= start)
2830 return ret;
2831
2832 /* Does pgoff wrap? */
2833 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2834 return ret;
2835
2836 if (mmap_write_lock_killable(mm))
2837 return -EINTR;
2838
2839 vma = vma_lookup(mm, start);
2840
2841 if (!vma || !(vma->vm_flags & VM_SHARED))
2842 goto out;
2843
2844 if (start + size > vma->vm_end) {
2845 VMA_ITERATOR(vmi, mm, vma->vm_end);
2846 struct vm_area_struct *next, *prev = vma;
2847
2848 for_each_vma_range(vmi, next, start + size) {
2849 /* hole between vmas ? */
2850 if (next->vm_start != prev->vm_end)
2851 goto out;
2852
2853 if (next->vm_file != vma->vm_file)
2854 goto out;
2855
2856 if (next->vm_flags != vma->vm_flags)
2857 goto out;
2858
2859 if (start + size <= next->vm_end)
2860 break;
2861
2862 prev = next;
2863 }
2864
2865 if (!next)
2866 goto out;
2867 }
2868
2869 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2870 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2871 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2872
2873 flags &= MAP_NONBLOCK;
2874 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2875 if (vma->vm_flags & VM_LOCKED)
2876 flags |= MAP_LOCKED;
2877
2878 file = get_file(vma->vm_file);
2879 ret = do_mmap(vma->vm_file, start, size,
2880 prot, flags, pgoff, &populate, NULL);
2881 fput(file);
2882out:
2883 mmap_write_unlock(mm);
2884 if (populate)
2885 mm_populate(ret, populate);
2886 if (!IS_ERR_VALUE(ret))
2887 ret = 0;
2888 return ret;
2889}
2890
2891/*
2892 * brk_munmap() - Unmap a parital vma.
2893 * @mas: The maple tree state.
2894 * @vma: The vma to be modified
2895 * @newbrk: the start of the address to unmap
2896 * @oldbrk: The end of the address to unmap
2897 * @uf: The userfaultfd list_head
2898 *
2899 * Returns: 1 on success.
2900 * unmaps a partial VMA mapping. Does not handle alignment, downgrades lock if
2901 * possible.
2902 */
2903static int do_brk_munmap(struct ma_state *mas, struct vm_area_struct *vma,
2904 unsigned long newbrk, unsigned long oldbrk,
2905 struct list_head *uf)
2906{
2907 struct mm_struct *mm = vma->vm_mm;
2908 int ret;
2909
2910 arch_unmap(mm, newbrk, oldbrk);
2911 ret = do_mas_align_munmap(mas, vma, mm, newbrk, oldbrk, uf, true);
2912 validate_mm_mt(mm);
2913 return ret;
2914}
2915
2916/*
2917 * do_brk_flags() - Increase the brk vma if the flags match.
2918 * @mas: The maple tree state.
2919 * @addr: The start address
2920 * @len: The length of the increase
2921 * @vma: The vma,
2922 * @flags: The VMA Flags
2923 *
2924 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
2925 * do not match then create a new anonymous VMA. Eventually we may be able to
2926 * do some brk-specific accounting here.
2927 */
2928static int do_brk_flags(struct ma_state *mas, struct vm_area_struct *vma,
2929 unsigned long addr, unsigned long len, unsigned long flags)
2930{
2931 struct mm_struct *mm = current->mm;
2932
2933 validate_mm_mt(mm);
2934 /*
2935 * Check against address space limits by the changed size
2936 * Note: This happens *after* clearing old mappings in some code paths.
2937 */
2938 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2939 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2940 return -ENOMEM;
2941
2942 if (mm->map_count > sysctl_max_map_count)
2943 return -ENOMEM;
2944
2945 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2946 return -ENOMEM;
2947
2948 /*
2949 * Expand the existing vma if possible; Note that singular lists do not
2950 * occur after forking, so the expand will only happen on new VMAs.
2951 */
2952 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
2953 can_vma_merge_after(vma, flags, NULL, NULL,
2954 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
2955 mas_set_range(mas, vma->vm_start, addr + len - 1);
2956 if (mas_preallocate(mas, vma, GFP_KERNEL))
2957 goto unacct_fail;
2958
2959 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
2960 if (vma->anon_vma) {
2961 anon_vma_lock_write(vma->anon_vma);
2962 anon_vma_interval_tree_pre_update_vma(vma);
2963 }
2964 vma->vm_end = addr + len;
2965 vma->vm_flags |= VM_SOFTDIRTY;
2966 mas_store_prealloc(mas, vma);
2967
2968 if (vma->anon_vma) {
2969 anon_vma_interval_tree_post_update_vma(vma);
2970 anon_vma_unlock_write(vma->anon_vma);
2971 }
2972 khugepaged_enter_vma(vma, flags);
2973 goto out;
2974 }
2975
2976 /* create a vma struct for an anonymous mapping */
2977 vma = vm_area_alloc(mm);
2978 if (!vma)
2979 goto unacct_fail;
2980
2981 vma_set_anonymous(vma);
2982 vma->vm_start = addr;
2983 vma->vm_end = addr + len;
2984 vma->vm_pgoff = addr >> PAGE_SHIFT;
2985 vma->vm_flags = flags;
2986 vma->vm_page_prot = vm_get_page_prot(flags);
2987 mas_set_range(mas, vma->vm_start, addr + len - 1);
2988 if (mas_store_gfp(mas, vma, GFP_KERNEL))
2989 goto mas_store_fail;
2990
2991 mm->map_count++;
2992out:
2993 perf_event_mmap(vma);
2994 mm->total_vm += len >> PAGE_SHIFT;
2995 mm->data_vm += len >> PAGE_SHIFT;
2996 if (flags & VM_LOCKED)
2997 mm->locked_vm += (len >> PAGE_SHIFT);
2998 vma->vm_flags |= VM_SOFTDIRTY;
2999 validate_mm(mm);
3000 return 0;
3001
3002mas_store_fail:
3003 vm_area_free(vma);
3004unacct_fail:
3005 vm_unacct_memory(len >> PAGE_SHIFT);
3006 return -ENOMEM;
3007}
3008
3009int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3010{
3011 struct mm_struct *mm = current->mm;
3012 struct vm_area_struct *vma = NULL;
3013 unsigned long len;
3014 int ret;
3015 bool populate;
3016 LIST_HEAD(uf);
3017 MA_STATE(mas, &mm->mm_mt, addr, addr);
3018
3019 len = PAGE_ALIGN(request);
3020 if (len < request)
3021 return -ENOMEM;
3022 if (!len)
3023 return 0;
3024
3025 if (mmap_write_lock_killable(mm))
3026 return -EINTR;
3027
3028 /* Until we need other flags, refuse anything except VM_EXEC. */
3029 if ((flags & (~VM_EXEC)) != 0)
3030 return -EINVAL;
3031
3032 ret = check_brk_limits(addr, len);
3033 if (ret)
3034 goto limits_failed;
3035
3036 ret = do_mas_munmap(&mas, mm, addr, len, &uf, 0);
3037 if (ret)
3038 goto munmap_failed;
3039
3040 vma = mas_prev(&mas, 0);
3041 ret = do_brk_flags(&mas, vma, addr, len, flags);
3042 populate = ((mm->def_flags & VM_LOCKED) != 0);
3043 mmap_write_unlock(mm);
3044 userfaultfd_unmap_complete(mm, &uf);
3045 if (populate && !ret)
3046 mm_populate(addr, len);
3047 return ret;
3048
3049munmap_failed:
3050limits_failed:
3051 mmap_write_unlock(mm);
3052 return ret;
3053}
3054EXPORT_SYMBOL(vm_brk_flags);
3055
3056int vm_brk(unsigned long addr, unsigned long len)
3057{
3058 return vm_brk_flags(addr, len, 0);
3059}
3060EXPORT_SYMBOL(vm_brk);
3061
3062/* Release all mmaps. */
3063void exit_mmap(struct mm_struct *mm)
3064{
3065 struct mmu_gather tlb;
3066 struct vm_area_struct *vma;
3067 unsigned long nr_accounted = 0;
3068 MA_STATE(mas, &mm->mm_mt, 0, 0);
3069 int count = 0;
3070
3071 /* mm's last user has gone, and its about to be pulled down */
3072 mmu_notifier_release(mm);
3073
3074 mmap_read_lock(mm);
3075 arch_exit_mmap(mm);
3076
3077 vma = mas_find(&mas, ULONG_MAX);
3078 if (!vma) {
3079 /* Can happen if dup_mmap() received an OOM */
3080 mmap_read_unlock(mm);
3081 return;
3082 }
3083
3084 lru_add_drain();
3085 flush_cache_mm(mm);
3086 tlb_gather_mmu_fullmm(&tlb, mm);
3087 /* update_hiwater_rss(mm) here? but nobody should be looking */
3088 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3089 unmap_vmas(&tlb, &mm->mm_mt, vma, 0, ULONG_MAX);
3090 mmap_read_unlock(mm);
3091
3092 /*
3093 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3094 * because the memory has been already freed.
3095 */
3096 set_bit(MMF_OOM_SKIP, &mm->flags);
3097 mmap_write_lock(mm);
3098 free_pgtables(&tlb, &mm->mm_mt, vma, FIRST_USER_ADDRESS,
3099 USER_PGTABLES_CEILING);
3100 tlb_finish_mmu(&tlb);
3101
3102 /*
3103 * Walk the list again, actually closing and freeing it, with preemption
3104 * enabled, without holding any MM locks besides the unreachable
3105 * mmap_write_lock.
3106 */
3107 do {
3108 if (vma->vm_flags & VM_ACCOUNT)
3109 nr_accounted += vma_pages(vma);
3110 remove_vma(vma);
3111 count++;
3112 cond_resched();
3113 } while ((vma = mas_find(&mas, ULONG_MAX)) != NULL);
3114
3115 BUG_ON(count != mm->map_count);
3116
3117 trace_exit_mmap(mm);
3118 __mt_destroy(&mm->mm_mt);
3119 mmap_write_unlock(mm);
3120 vm_unacct_memory(nr_accounted);
3121}
3122
3123/* Insert vm structure into process list sorted by address
3124 * and into the inode's i_mmap tree. If vm_file is non-NULL
3125 * then i_mmap_rwsem is taken here.
3126 */
3127int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3128{
3129 unsigned long charged = vma_pages(vma);
3130
3131
3132 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3133 return -ENOMEM;
3134
3135 if ((vma->vm_flags & VM_ACCOUNT) &&
3136 security_vm_enough_memory_mm(mm, charged))
3137 return -ENOMEM;
3138
3139 /*
3140 * The vm_pgoff of a purely anonymous vma should be irrelevant
3141 * until its first write fault, when page's anon_vma and index
3142 * are set. But now set the vm_pgoff it will almost certainly
3143 * end up with (unless mremap moves it elsewhere before that
3144 * first wfault), so /proc/pid/maps tells a consistent story.
3145 *
3146 * By setting it to reflect the virtual start address of the
3147 * vma, merges and splits can happen in a seamless way, just
3148 * using the existing file pgoff checks and manipulations.
3149 * Similarly in do_mmap and in do_brk_flags.
3150 */
3151 if (vma_is_anonymous(vma)) {
3152 BUG_ON(vma->anon_vma);
3153 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3154 }
3155
3156 if (vma_link(mm, vma)) {
3157 vm_unacct_memory(charged);
3158 return -ENOMEM;
3159 }
3160
3161 return 0;
3162}
3163
3164/*
3165 * Copy the vma structure to a new location in the same mm,
3166 * prior to moving page table entries, to effect an mremap move.
3167 */
3168struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3169 unsigned long addr, unsigned long len, pgoff_t pgoff,
3170 bool *need_rmap_locks)
3171{
3172 struct vm_area_struct *vma = *vmap;
3173 unsigned long vma_start = vma->vm_start;
3174 struct mm_struct *mm = vma->vm_mm;
3175 struct vm_area_struct *new_vma, *prev;
3176 bool faulted_in_anon_vma = true;
3177
3178 validate_mm_mt(mm);
3179 /*
3180 * If anonymous vma has not yet been faulted, update new pgoff
3181 * to match new location, to increase its chance of merging.
3182 */
3183 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3184 pgoff = addr >> PAGE_SHIFT;
3185 faulted_in_anon_vma = false;
3186 }
3187
3188 new_vma = find_vma_prev(mm, addr, &prev);
3189 if (new_vma && new_vma->vm_start < addr + len)
3190 return NULL; /* should never get here */
3191
3192 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3193 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3194 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
3195 if (new_vma) {
3196 /*
3197 * Source vma may have been merged into new_vma
3198 */
3199 if (unlikely(vma_start >= new_vma->vm_start &&
3200 vma_start < new_vma->vm_end)) {
3201 /*
3202 * The only way we can get a vma_merge with
3203 * self during an mremap is if the vma hasn't
3204 * been faulted in yet and we were allowed to
3205 * reset the dst vma->vm_pgoff to the
3206 * destination address of the mremap to allow
3207 * the merge to happen. mremap must change the
3208 * vm_pgoff linearity between src and dst vmas
3209 * (in turn preventing a vma_merge) to be
3210 * safe. It is only safe to keep the vm_pgoff
3211 * linear if there are no pages mapped yet.
3212 */
3213 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3214 *vmap = vma = new_vma;
3215 }
3216 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3217 } else {
3218 new_vma = vm_area_dup(vma);
3219 if (!new_vma)
3220 goto out;
3221 new_vma->vm_start = addr;
3222 new_vma->vm_end = addr + len;
3223 new_vma->vm_pgoff = pgoff;
3224 if (vma_dup_policy(vma, new_vma))
3225 goto out_free_vma;
3226 if (anon_vma_clone(new_vma, vma))
3227 goto out_free_mempol;
3228 if (new_vma->vm_file)
3229 get_file(new_vma->vm_file);
3230 if (new_vma->vm_ops && new_vma->vm_ops->open)
3231 new_vma->vm_ops->open(new_vma);
3232 if (vma_link(mm, new_vma))
3233 goto out_vma_link;
3234 *need_rmap_locks = false;
3235 }
3236 validate_mm_mt(mm);
3237 return new_vma;
3238
3239out_vma_link:
3240 if (new_vma->vm_ops && new_vma->vm_ops->close)
3241 new_vma->vm_ops->close(new_vma);
3242
3243 if (new_vma->vm_file)
3244 fput(new_vma->vm_file);
3245
3246 unlink_anon_vmas(new_vma);
3247out_free_mempol:
3248 mpol_put(vma_policy(new_vma));
3249out_free_vma:
3250 vm_area_free(new_vma);
3251out:
3252 validate_mm_mt(mm);
3253 return NULL;
3254}
3255
3256/*
3257 * Return true if the calling process may expand its vm space by the passed
3258 * number of pages
3259 */
3260bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3261{
3262 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3263 return false;
3264
3265 if (is_data_mapping(flags) &&
3266 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3267 /* Workaround for Valgrind */
3268 if (rlimit(RLIMIT_DATA) == 0 &&
3269 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3270 return true;
3271
3272 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3273 current->comm, current->pid,
3274 (mm->data_vm + npages) << PAGE_SHIFT,
3275 rlimit(RLIMIT_DATA),
3276 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3277
3278 if (!ignore_rlimit_data)
3279 return false;
3280 }
3281
3282 return true;
3283}
3284
3285void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3286{
3287 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3288
3289 if (is_exec_mapping(flags))
3290 mm->exec_vm += npages;
3291 else if (is_stack_mapping(flags))
3292 mm->stack_vm += npages;
3293 else if (is_data_mapping(flags))
3294 mm->data_vm += npages;
3295}
3296
3297static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3298
3299/*
3300 * Having a close hook prevents vma merging regardless of flags.
3301 */
3302static void special_mapping_close(struct vm_area_struct *vma)
3303{
3304}
3305
3306static const char *special_mapping_name(struct vm_area_struct *vma)
3307{
3308 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3309}
3310
3311static int special_mapping_mremap(struct vm_area_struct *new_vma)
3312{
3313 struct vm_special_mapping *sm = new_vma->vm_private_data;
3314
3315 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3316 return -EFAULT;
3317
3318 if (sm->mremap)
3319 return sm->mremap(sm, new_vma);
3320
3321 return 0;
3322}
3323
3324static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3325{
3326 /*
3327 * Forbid splitting special mappings - kernel has expectations over
3328 * the number of pages in mapping. Together with VM_DONTEXPAND
3329 * the size of vma should stay the same over the special mapping's
3330 * lifetime.
3331 */
3332 return -EINVAL;
3333}
3334
3335static const struct vm_operations_struct special_mapping_vmops = {
3336 .close = special_mapping_close,
3337 .fault = special_mapping_fault,
3338 .mremap = special_mapping_mremap,
3339 .name = special_mapping_name,
3340 /* vDSO code relies that VVAR can't be accessed remotely */
3341 .access = NULL,
3342 .may_split = special_mapping_split,
3343};
3344
3345static const struct vm_operations_struct legacy_special_mapping_vmops = {
3346 .close = special_mapping_close,
3347 .fault = special_mapping_fault,
3348};
3349
3350static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3351{
3352 struct vm_area_struct *vma = vmf->vma;
3353 pgoff_t pgoff;
3354 struct page **pages;
3355
3356 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3357 pages = vma->vm_private_data;
3358 } else {
3359 struct vm_special_mapping *sm = vma->vm_private_data;
3360
3361 if (sm->fault)
3362 return sm->fault(sm, vmf->vma, vmf);
3363
3364 pages = sm->pages;
3365 }
3366
3367 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3368 pgoff--;
3369
3370 if (*pages) {
3371 struct page *page = *pages;
3372 get_page(page);
3373 vmf->page = page;
3374 return 0;
3375 }
3376
3377 return VM_FAULT_SIGBUS;
3378}
3379
3380static struct vm_area_struct *__install_special_mapping(
3381 struct mm_struct *mm,
3382 unsigned long addr, unsigned long len,
3383 unsigned long vm_flags, void *priv,
3384 const struct vm_operations_struct *ops)
3385{
3386 int ret;
3387 struct vm_area_struct *vma;
3388
3389 validate_mm_mt(mm);
3390 vma = vm_area_alloc(mm);
3391 if (unlikely(vma == NULL))
3392 return ERR_PTR(-ENOMEM);
3393
3394 vma->vm_start = addr;
3395 vma->vm_end = addr + len;
3396
3397 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3398 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
3399 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3400
3401 vma->vm_ops = ops;
3402 vma->vm_private_data = priv;
3403
3404 ret = insert_vm_struct(mm, vma);
3405 if (ret)
3406 goto out;
3407
3408 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3409
3410 perf_event_mmap(vma);
3411
3412 validate_mm_mt(mm);
3413 return vma;
3414
3415out:
3416 vm_area_free(vma);
3417 validate_mm_mt(mm);
3418 return ERR_PTR(ret);
3419}
3420
3421bool vma_is_special_mapping(const struct vm_area_struct *vma,
3422 const struct vm_special_mapping *sm)
3423{
3424 return vma->vm_private_data == sm &&
3425 (vma->vm_ops == &special_mapping_vmops ||
3426 vma->vm_ops == &legacy_special_mapping_vmops);
3427}
3428
3429/*
3430 * Called with mm->mmap_lock held for writing.
3431 * Insert a new vma covering the given region, with the given flags.
3432 * Its pages are supplied by the given array of struct page *.
3433 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3434 * The region past the last page supplied will always produce SIGBUS.
3435 * The array pointer and the pages it points to are assumed to stay alive
3436 * for as long as this mapping might exist.
3437 */
3438struct vm_area_struct *_install_special_mapping(
3439 struct mm_struct *mm,
3440 unsigned long addr, unsigned long len,
3441 unsigned long vm_flags, const struct vm_special_mapping *spec)
3442{
3443 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3444 &special_mapping_vmops);
3445}
3446
3447int install_special_mapping(struct mm_struct *mm,
3448 unsigned long addr, unsigned long len,
3449 unsigned long vm_flags, struct page **pages)
3450{
3451 struct vm_area_struct *vma = __install_special_mapping(
3452 mm, addr, len, vm_flags, (void *)pages,
3453 &legacy_special_mapping_vmops);
3454
3455 return PTR_ERR_OR_ZERO(vma);
3456}
3457
3458static DEFINE_MUTEX(mm_all_locks_mutex);
3459
3460static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3461{
3462 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3463 /*
3464 * The LSB of head.next can't change from under us
3465 * because we hold the mm_all_locks_mutex.
3466 */
3467 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3468 /*
3469 * We can safely modify head.next after taking the
3470 * anon_vma->root->rwsem. If some other vma in this mm shares
3471 * the same anon_vma we won't take it again.
3472 *
3473 * No need of atomic instructions here, head.next
3474 * can't change from under us thanks to the
3475 * anon_vma->root->rwsem.
3476 */
3477 if (__test_and_set_bit(0, (unsigned long *)
3478 &anon_vma->root->rb_root.rb_root.rb_node))
3479 BUG();
3480 }
3481}
3482
3483static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3484{
3485 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3486 /*
3487 * AS_MM_ALL_LOCKS can't change from under us because
3488 * we hold the mm_all_locks_mutex.
3489 *
3490 * Operations on ->flags have to be atomic because
3491 * even if AS_MM_ALL_LOCKS is stable thanks to the
3492 * mm_all_locks_mutex, there may be other cpus
3493 * changing other bitflags in parallel to us.
3494 */
3495 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3496 BUG();
3497 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3498 }
3499}
3500
3501/*
3502 * This operation locks against the VM for all pte/vma/mm related
3503 * operations that could ever happen on a certain mm. This includes
3504 * vmtruncate, try_to_unmap, and all page faults.
3505 *
3506 * The caller must take the mmap_lock in write mode before calling
3507 * mm_take_all_locks(). The caller isn't allowed to release the
3508 * mmap_lock until mm_drop_all_locks() returns.
3509 *
3510 * mmap_lock in write mode is required in order to block all operations
3511 * that could modify pagetables and free pages without need of
3512 * altering the vma layout. It's also needed in write mode to avoid new
3513 * anon_vmas to be associated with existing vmas.
3514 *
3515 * A single task can't take more than one mm_take_all_locks() in a row
3516 * or it would deadlock.
3517 *
3518 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3519 * mapping->flags avoid to take the same lock twice, if more than one
3520 * vma in this mm is backed by the same anon_vma or address_space.
3521 *
3522 * We take locks in following order, accordingly to comment at beginning
3523 * of mm/rmap.c:
3524 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3525 * hugetlb mapping);
3526 * - all i_mmap_rwsem locks;
3527 * - all anon_vma->rwseml
3528 *
3529 * We can take all locks within these types randomly because the VM code
3530 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3531 * mm_all_locks_mutex.
3532 *
3533 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3534 * that may have to take thousand of locks.
3535 *
3536 * mm_take_all_locks() can fail if it's interrupted by signals.
3537 */
3538int mm_take_all_locks(struct mm_struct *mm)
3539{
3540 struct vm_area_struct *vma;
3541 struct anon_vma_chain *avc;
3542 MA_STATE(mas, &mm->mm_mt, 0, 0);
3543
3544 mmap_assert_write_locked(mm);
3545
3546 mutex_lock(&mm_all_locks_mutex);
3547
3548 mas_for_each(&mas, vma, ULONG_MAX) {
3549 if (signal_pending(current))
3550 goto out_unlock;
3551 if (vma->vm_file && vma->vm_file->f_mapping &&
3552 is_vm_hugetlb_page(vma))
3553 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3554 }
3555
3556 mas_set(&mas, 0);
3557 mas_for_each(&mas, vma, ULONG_MAX) {
3558 if (signal_pending(current))
3559 goto out_unlock;
3560 if (vma->vm_file && vma->vm_file->f_mapping &&
3561 !is_vm_hugetlb_page(vma))
3562 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3563 }
3564
3565 mas_set(&mas, 0);
3566 mas_for_each(&mas, vma, ULONG_MAX) {
3567 if (signal_pending(current))
3568 goto out_unlock;
3569 if (vma->anon_vma)
3570 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3571 vm_lock_anon_vma(mm, avc->anon_vma);
3572 }
3573
3574 return 0;
3575
3576out_unlock:
3577 mm_drop_all_locks(mm);
3578 return -EINTR;
3579}
3580
3581static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3582{
3583 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3584 /*
3585 * The LSB of head.next can't change to 0 from under
3586 * us because we hold the mm_all_locks_mutex.
3587 *
3588 * We must however clear the bitflag before unlocking
3589 * the vma so the users using the anon_vma->rb_root will
3590 * never see our bitflag.
3591 *
3592 * No need of atomic instructions here, head.next
3593 * can't change from under us until we release the
3594 * anon_vma->root->rwsem.
3595 */
3596 if (!__test_and_clear_bit(0, (unsigned long *)
3597 &anon_vma->root->rb_root.rb_root.rb_node))
3598 BUG();
3599 anon_vma_unlock_write(anon_vma);
3600 }
3601}
3602
3603static void vm_unlock_mapping(struct address_space *mapping)
3604{
3605 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3606 /*
3607 * AS_MM_ALL_LOCKS can't change to 0 from under us
3608 * because we hold the mm_all_locks_mutex.
3609 */
3610 i_mmap_unlock_write(mapping);
3611 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3612 &mapping->flags))
3613 BUG();
3614 }
3615}
3616
3617/*
3618 * The mmap_lock cannot be released by the caller until
3619 * mm_drop_all_locks() returns.
3620 */
3621void mm_drop_all_locks(struct mm_struct *mm)
3622{
3623 struct vm_area_struct *vma;
3624 struct anon_vma_chain *avc;
3625 MA_STATE(mas, &mm->mm_mt, 0, 0);
3626
3627 mmap_assert_write_locked(mm);
3628 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3629
3630 mas_for_each(&mas, vma, ULONG_MAX) {
3631 if (vma->anon_vma)
3632 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3633 vm_unlock_anon_vma(avc->anon_vma);
3634 if (vma->vm_file && vma->vm_file->f_mapping)
3635 vm_unlock_mapping(vma->vm_file->f_mapping);
3636 }
3637
3638 mutex_unlock(&mm_all_locks_mutex);
3639}
3640
3641/*
3642 * initialise the percpu counter for VM
3643 */
3644void __init mmap_init(void)
3645{
3646 int ret;
3647
3648 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3649 VM_BUG_ON(ret);
3650}
3651
3652/*
3653 * Initialise sysctl_user_reserve_kbytes.
3654 *
3655 * This is intended to prevent a user from starting a single memory hogging
3656 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3657 * mode.
3658 *
3659 * The default value is min(3% of free memory, 128MB)
3660 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3661 */
3662static int init_user_reserve(void)
3663{
3664 unsigned long free_kbytes;
3665
3666 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3667
3668 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3669 return 0;
3670}
3671subsys_initcall(init_user_reserve);
3672
3673/*
3674 * Initialise sysctl_admin_reserve_kbytes.
3675 *
3676 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3677 * to log in and kill a memory hogging process.
3678 *
3679 * Systems with more than 256MB will reserve 8MB, enough to recover
3680 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3681 * only reserve 3% of free pages by default.
3682 */
3683static int init_admin_reserve(void)
3684{
3685 unsigned long free_kbytes;
3686
3687 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3688
3689 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3690 return 0;
3691}
3692subsys_initcall(init_admin_reserve);
3693
3694/*
3695 * Reinititalise user and admin reserves if memory is added or removed.
3696 *
3697 * The default user reserve max is 128MB, and the default max for the
3698 * admin reserve is 8MB. These are usually, but not always, enough to
3699 * enable recovery from a memory hogging process using login/sshd, a shell,
3700 * and tools like top. It may make sense to increase or even disable the
3701 * reserve depending on the existence of swap or variations in the recovery
3702 * tools. So, the admin may have changed them.
3703 *
3704 * If memory is added and the reserves have been eliminated or increased above
3705 * the default max, then we'll trust the admin.
3706 *
3707 * If memory is removed and there isn't enough free memory, then we
3708 * need to reset the reserves.
3709 *
3710 * Otherwise keep the reserve set by the admin.
3711 */
3712static int reserve_mem_notifier(struct notifier_block *nb,
3713 unsigned long action, void *data)
3714{
3715 unsigned long tmp, free_kbytes;
3716
3717 switch (action) {
3718 case MEM_ONLINE:
3719 /* Default max is 128MB. Leave alone if modified by operator. */
3720 tmp = sysctl_user_reserve_kbytes;
3721 if (0 < tmp && tmp < (1UL << 17))
3722 init_user_reserve();
3723
3724 /* Default max is 8MB. Leave alone if modified by operator. */
3725 tmp = sysctl_admin_reserve_kbytes;
3726 if (0 < tmp && tmp < (1UL << 13))
3727 init_admin_reserve();
3728
3729 break;
3730 case MEM_OFFLINE:
3731 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3732
3733 if (sysctl_user_reserve_kbytes > free_kbytes) {
3734 init_user_reserve();
3735 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3736 sysctl_user_reserve_kbytes);
3737 }
3738
3739 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3740 init_admin_reserve();
3741 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3742 sysctl_admin_reserve_kbytes);
3743 }
3744 break;
3745 default:
3746 break;
3747 }
3748 return NOTIFY_OK;
3749}
3750
3751static int __meminit init_reserve_notifier(void)
3752{
3753 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3754 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3755
3756 return 0;
3757}
3758subsys_initcall(init_reserve_notifier);