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/*
2 * mm/mmap.c
3 *
4 * Written by obz.
5 *
6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
7 */
8
9#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10
11#include <linux/kernel.h>
12#include <linux/slab.h>
13#include <linux/backing-dev.h>
14#include <linux/mm.h>
15#include <linux/vmacache.h>
16#include <linux/shm.h>
17#include <linux/mman.h>
18#include <linux/pagemap.h>
19#include <linux/swap.h>
20#include <linux/syscalls.h>
21#include <linux/capability.h>
22#include <linux/init.h>
23#include <linux/file.h>
24#include <linux/fs.h>
25#include <linux/personality.h>
26#include <linux/security.h>
27#include <linux/hugetlb.h>
28#include <linux/shmem_fs.h>
29#include <linux/profile.h>
30#include <linux/export.h>
31#include <linux/mount.h>
32#include <linux/mempolicy.h>
33#include <linux/rmap.h>
34#include <linux/mmu_notifier.h>
35#include <linux/mmdebug.h>
36#include <linux/perf_event.h>
37#include <linux/audit.h>
38#include <linux/khugepaged.h>
39#include <linux/uprobes.h>
40#include <linux/rbtree_augmented.h>
41#include <linux/notifier.h>
42#include <linux/memory.h>
43#include <linux/printk.h>
44#include <linux/userfaultfd_k.h>
45#include <linux/moduleparam.h>
46#include <linux/pkeys.h>
47#include <linux/oom.h>
48
49#include <linux/uaccess.h>
50#include <asm/cacheflush.h>
51#include <asm/tlb.h>
52#include <asm/mmu_context.h>
53
54#include "internal.h"
55
56#ifndef arch_mmap_check
57#define arch_mmap_check(addr, len, flags) (0)
58#endif
59
60#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
61const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
62const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
63int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
64#endif
65#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
66const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
67const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
68int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
69#endif
70
71static bool ignore_rlimit_data;
72core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
73
74static void unmap_region(struct mm_struct *mm,
75 struct vm_area_struct *vma, struct vm_area_struct *prev,
76 unsigned long start, unsigned long end);
77
78/* description of effects of mapping type and prot in current implementation.
79 * this is due to the limited x86 page protection hardware. The expected
80 * behavior is in parens:
81 *
82 * map_type prot
83 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
84 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
85 * w: (no) no w: (no) no w: (yes) yes w: (no) no
86 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
87 *
88 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
89 * w: (no) no w: (no) no w: (copy) copy w: (no) no
90 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
91 *
92 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
93 * MAP_PRIVATE:
94 * r: (no) no
95 * w: (no) no
96 * x: (yes) yes
97 */
98pgprot_t protection_map[16] __ro_after_init = {
99 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
100 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
101};
102
103#ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
104static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
105{
106 return prot;
107}
108#endif
109
110pgprot_t vm_get_page_prot(unsigned long vm_flags)
111{
112 pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
113 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
114 pgprot_val(arch_vm_get_page_prot(vm_flags)));
115
116 return arch_filter_pgprot(ret);
117}
118EXPORT_SYMBOL(vm_get_page_prot);
119
120static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
121{
122 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
123}
124
125/* Update vma->vm_page_prot to reflect vma->vm_flags. */
126void vma_set_page_prot(struct vm_area_struct *vma)
127{
128 unsigned long vm_flags = vma->vm_flags;
129 pgprot_t vm_page_prot;
130
131 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
132 if (vma_wants_writenotify(vma, vm_page_prot)) {
133 vm_flags &= ~VM_SHARED;
134 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
135 }
136 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
137 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
138}
139
140/*
141 * Requires inode->i_mapping->i_mmap_rwsem
142 */
143static void __remove_shared_vm_struct(struct vm_area_struct *vma,
144 struct file *file, struct address_space *mapping)
145{
146 if (vma->vm_flags & VM_DENYWRITE)
147 atomic_inc(&file_inode(file)->i_writecount);
148 if (vma->vm_flags & VM_SHARED)
149 mapping_unmap_writable(mapping);
150
151 flush_dcache_mmap_lock(mapping);
152 vma_interval_tree_remove(vma, &mapping->i_mmap);
153 flush_dcache_mmap_unlock(mapping);
154}
155
156/*
157 * Unlink a file-based vm structure from its interval tree, to hide
158 * vma from rmap and vmtruncate before freeing its page tables.
159 */
160void unlink_file_vma(struct vm_area_struct *vma)
161{
162 struct file *file = vma->vm_file;
163
164 if (file) {
165 struct address_space *mapping = file->f_mapping;
166 i_mmap_lock_write(mapping);
167 __remove_shared_vm_struct(vma, file, mapping);
168 i_mmap_unlock_write(mapping);
169 }
170}
171
172/*
173 * Close a vm structure and free it, returning the next.
174 */
175static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
176{
177 struct vm_area_struct *next = vma->vm_next;
178
179 might_sleep();
180 if (vma->vm_ops && vma->vm_ops->close)
181 vma->vm_ops->close(vma);
182 if (vma->vm_file)
183 fput(vma->vm_file);
184 mpol_put(vma_policy(vma));
185 kmem_cache_free(vm_area_cachep, vma);
186 return next;
187}
188
189static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf);
190
191SYSCALL_DEFINE1(brk, unsigned long, brk)
192{
193 unsigned long retval;
194 unsigned long newbrk, oldbrk;
195 struct mm_struct *mm = current->mm;
196 struct vm_area_struct *next;
197 unsigned long min_brk;
198 bool populate;
199 LIST_HEAD(uf);
200
201 if (down_write_killable(&mm->mmap_sem))
202 return -EINTR;
203
204#ifdef CONFIG_COMPAT_BRK
205 /*
206 * CONFIG_COMPAT_BRK can still be overridden by setting
207 * randomize_va_space to 2, which will still cause mm->start_brk
208 * to be arbitrarily shifted
209 */
210 if (current->brk_randomized)
211 min_brk = mm->start_brk;
212 else
213 min_brk = mm->end_data;
214#else
215 min_brk = mm->start_brk;
216#endif
217 if (brk < min_brk)
218 goto out;
219
220 /*
221 * Check against rlimit here. If this check is done later after the test
222 * of oldbrk with newbrk then it can escape the test and let the data
223 * segment grow beyond its set limit the in case where the limit is
224 * not page aligned -Ram Gupta
225 */
226 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
227 mm->end_data, mm->start_data))
228 goto out;
229
230 newbrk = PAGE_ALIGN(brk);
231 oldbrk = PAGE_ALIGN(mm->brk);
232 if (oldbrk == newbrk)
233 goto set_brk;
234
235 /* Always allow shrinking brk. */
236 if (brk <= mm->brk) {
237 if (!do_munmap(mm, newbrk, oldbrk-newbrk, &uf))
238 goto set_brk;
239 goto out;
240 }
241
242 /* Check against existing mmap mappings. */
243 next = find_vma(mm, oldbrk);
244 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
245 goto out;
246
247 /* Ok, looks good - let it rip. */
248 if (do_brk(oldbrk, newbrk-oldbrk, &uf) < 0)
249 goto out;
250
251set_brk:
252 mm->brk = brk;
253 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
254 up_write(&mm->mmap_sem);
255 userfaultfd_unmap_complete(mm, &uf);
256 if (populate)
257 mm_populate(oldbrk, newbrk - oldbrk);
258 return brk;
259
260out:
261 retval = mm->brk;
262 up_write(&mm->mmap_sem);
263 return retval;
264}
265
266static long vma_compute_subtree_gap(struct vm_area_struct *vma)
267{
268 unsigned long max, prev_end, subtree_gap;
269
270 /*
271 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
272 * allow two stack_guard_gaps between them here, and when choosing
273 * an unmapped area; whereas when expanding we only require one.
274 * That's a little inconsistent, but keeps the code here simpler.
275 */
276 max = vm_start_gap(vma);
277 if (vma->vm_prev) {
278 prev_end = vm_end_gap(vma->vm_prev);
279 if (max > prev_end)
280 max -= prev_end;
281 else
282 max = 0;
283 }
284 if (vma->vm_rb.rb_left) {
285 subtree_gap = rb_entry(vma->vm_rb.rb_left,
286 struct vm_area_struct, vm_rb)->rb_subtree_gap;
287 if (subtree_gap > max)
288 max = subtree_gap;
289 }
290 if (vma->vm_rb.rb_right) {
291 subtree_gap = rb_entry(vma->vm_rb.rb_right,
292 struct vm_area_struct, vm_rb)->rb_subtree_gap;
293 if (subtree_gap > max)
294 max = subtree_gap;
295 }
296 return max;
297}
298
299#ifdef CONFIG_DEBUG_VM_RB
300static int browse_rb(struct mm_struct *mm)
301{
302 struct rb_root *root = &mm->mm_rb;
303 int i = 0, j, bug = 0;
304 struct rb_node *nd, *pn = NULL;
305 unsigned long prev = 0, pend = 0;
306
307 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
308 struct vm_area_struct *vma;
309 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
310 if (vma->vm_start < prev) {
311 pr_emerg("vm_start %lx < prev %lx\n",
312 vma->vm_start, prev);
313 bug = 1;
314 }
315 if (vma->vm_start < pend) {
316 pr_emerg("vm_start %lx < pend %lx\n",
317 vma->vm_start, pend);
318 bug = 1;
319 }
320 if (vma->vm_start > vma->vm_end) {
321 pr_emerg("vm_start %lx > vm_end %lx\n",
322 vma->vm_start, vma->vm_end);
323 bug = 1;
324 }
325 spin_lock(&mm->page_table_lock);
326 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
327 pr_emerg("free gap %lx, correct %lx\n",
328 vma->rb_subtree_gap,
329 vma_compute_subtree_gap(vma));
330 bug = 1;
331 }
332 spin_unlock(&mm->page_table_lock);
333 i++;
334 pn = nd;
335 prev = vma->vm_start;
336 pend = vma->vm_end;
337 }
338 j = 0;
339 for (nd = pn; nd; nd = rb_prev(nd))
340 j++;
341 if (i != j) {
342 pr_emerg("backwards %d, forwards %d\n", j, i);
343 bug = 1;
344 }
345 return bug ? -1 : i;
346}
347
348static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
349{
350 struct rb_node *nd;
351
352 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
353 struct vm_area_struct *vma;
354 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
355 VM_BUG_ON_VMA(vma != ignore &&
356 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
357 vma);
358 }
359}
360
361static void validate_mm(struct mm_struct *mm)
362{
363 int bug = 0;
364 int i = 0;
365 unsigned long highest_address = 0;
366 struct vm_area_struct *vma = mm->mmap;
367
368 while (vma) {
369 struct anon_vma *anon_vma = vma->anon_vma;
370 struct anon_vma_chain *avc;
371
372 if (anon_vma) {
373 anon_vma_lock_read(anon_vma);
374 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
375 anon_vma_interval_tree_verify(avc);
376 anon_vma_unlock_read(anon_vma);
377 }
378
379 highest_address = vm_end_gap(vma);
380 vma = vma->vm_next;
381 i++;
382 }
383 if (i != mm->map_count) {
384 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
385 bug = 1;
386 }
387 if (highest_address != mm->highest_vm_end) {
388 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
389 mm->highest_vm_end, highest_address);
390 bug = 1;
391 }
392 i = browse_rb(mm);
393 if (i != mm->map_count) {
394 if (i != -1)
395 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
396 bug = 1;
397 }
398 VM_BUG_ON_MM(bug, mm);
399}
400#else
401#define validate_mm_rb(root, ignore) do { } while (0)
402#define validate_mm(mm) do { } while (0)
403#endif
404
405RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
406 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
407
408/*
409 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
410 * vma->vm_prev->vm_end values changed, without modifying the vma's position
411 * in the rbtree.
412 */
413static void vma_gap_update(struct vm_area_struct *vma)
414{
415 /*
416 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
417 * function that does exacltly what we want.
418 */
419 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
420}
421
422static inline void vma_rb_insert(struct vm_area_struct *vma,
423 struct rb_root *root)
424{
425 /* All rb_subtree_gap values must be consistent prior to insertion */
426 validate_mm_rb(root, NULL);
427
428 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
429}
430
431static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
432{
433 /*
434 * Note rb_erase_augmented is a fairly large inline function,
435 * so make sure we instantiate it only once with our desired
436 * augmented rbtree callbacks.
437 */
438 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
439}
440
441static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
442 struct rb_root *root,
443 struct vm_area_struct *ignore)
444{
445 /*
446 * All rb_subtree_gap values must be consistent prior to erase,
447 * with the possible exception of the "next" vma being erased if
448 * next->vm_start was reduced.
449 */
450 validate_mm_rb(root, ignore);
451
452 __vma_rb_erase(vma, root);
453}
454
455static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
456 struct rb_root *root)
457{
458 /*
459 * All rb_subtree_gap values must be consistent prior to erase,
460 * with the possible exception of the vma being erased.
461 */
462 validate_mm_rb(root, vma);
463
464 __vma_rb_erase(vma, root);
465}
466
467/*
468 * vma has some anon_vma assigned, and is already inserted on that
469 * anon_vma's interval trees.
470 *
471 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
472 * vma must be removed from the anon_vma's interval trees using
473 * anon_vma_interval_tree_pre_update_vma().
474 *
475 * After the update, the vma will be reinserted using
476 * anon_vma_interval_tree_post_update_vma().
477 *
478 * The entire update must be protected by exclusive mmap_sem and by
479 * the root anon_vma's mutex.
480 */
481static inline void
482anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
483{
484 struct anon_vma_chain *avc;
485
486 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
487 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
488}
489
490static inline void
491anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
492{
493 struct anon_vma_chain *avc;
494
495 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
496 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
497}
498
499static int find_vma_links(struct mm_struct *mm, unsigned long addr,
500 unsigned long end, struct vm_area_struct **pprev,
501 struct rb_node ***rb_link, struct rb_node **rb_parent)
502{
503 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
504
505 __rb_link = &mm->mm_rb.rb_node;
506 rb_prev = __rb_parent = NULL;
507
508 while (*__rb_link) {
509 struct vm_area_struct *vma_tmp;
510
511 __rb_parent = *__rb_link;
512 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
513
514 if (vma_tmp->vm_end > addr) {
515 /* Fail if an existing vma overlaps the area */
516 if (vma_tmp->vm_start < end)
517 return -ENOMEM;
518 __rb_link = &__rb_parent->rb_left;
519 } else {
520 rb_prev = __rb_parent;
521 __rb_link = &__rb_parent->rb_right;
522 }
523 }
524
525 *pprev = NULL;
526 if (rb_prev)
527 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
528 *rb_link = __rb_link;
529 *rb_parent = __rb_parent;
530 return 0;
531}
532
533static unsigned long count_vma_pages_range(struct mm_struct *mm,
534 unsigned long addr, unsigned long end)
535{
536 unsigned long nr_pages = 0;
537 struct vm_area_struct *vma;
538
539 /* Find first overlaping mapping */
540 vma = find_vma_intersection(mm, addr, end);
541 if (!vma)
542 return 0;
543
544 nr_pages = (min(end, vma->vm_end) -
545 max(addr, vma->vm_start)) >> PAGE_SHIFT;
546
547 /* Iterate over the rest of the overlaps */
548 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
549 unsigned long overlap_len;
550
551 if (vma->vm_start > end)
552 break;
553
554 overlap_len = min(end, vma->vm_end) - vma->vm_start;
555 nr_pages += overlap_len >> PAGE_SHIFT;
556 }
557
558 return nr_pages;
559}
560
561void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
562 struct rb_node **rb_link, struct rb_node *rb_parent)
563{
564 /* Update tracking information for the gap following the new vma. */
565 if (vma->vm_next)
566 vma_gap_update(vma->vm_next);
567 else
568 mm->highest_vm_end = vm_end_gap(vma);
569
570 /*
571 * vma->vm_prev wasn't known when we followed the rbtree to find the
572 * correct insertion point for that vma. As a result, we could not
573 * update the vma vm_rb parents rb_subtree_gap values on the way down.
574 * So, we first insert the vma with a zero rb_subtree_gap value
575 * (to be consistent with what we did on the way down), and then
576 * immediately update the gap to the correct value. Finally we
577 * rebalance the rbtree after all augmented values have been set.
578 */
579 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
580 vma->rb_subtree_gap = 0;
581 vma_gap_update(vma);
582 vma_rb_insert(vma, &mm->mm_rb);
583}
584
585static void __vma_link_file(struct vm_area_struct *vma)
586{
587 struct file *file;
588
589 file = vma->vm_file;
590 if (file) {
591 struct address_space *mapping = file->f_mapping;
592
593 if (vma->vm_flags & VM_DENYWRITE)
594 atomic_dec(&file_inode(file)->i_writecount);
595 if (vma->vm_flags & VM_SHARED)
596 atomic_inc(&mapping->i_mmap_writable);
597
598 flush_dcache_mmap_lock(mapping);
599 vma_interval_tree_insert(vma, &mapping->i_mmap);
600 flush_dcache_mmap_unlock(mapping);
601 }
602}
603
604static void
605__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
606 struct vm_area_struct *prev, struct rb_node **rb_link,
607 struct rb_node *rb_parent)
608{
609 __vma_link_list(mm, vma, prev, rb_parent);
610 __vma_link_rb(mm, vma, rb_link, rb_parent);
611}
612
613static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
614 struct vm_area_struct *prev, struct rb_node **rb_link,
615 struct rb_node *rb_parent)
616{
617 struct address_space *mapping = NULL;
618
619 if (vma->vm_file) {
620 mapping = vma->vm_file->f_mapping;
621 i_mmap_lock_write(mapping);
622 }
623
624 __vma_link(mm, vma, prev, rb_link, rb_parent);
625 __vma_link_file(vma);
626
627 if (mapping)
628 i_mmap_unlock_write(mapping);
629
630 mm->map_count++;
631 validate_mm(mm);
632}
633
634/*
635 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
636 * mm's list and rbtree. It has already been inserted into the interval tree.
637 */
638static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
639{
640 struct vm_area_struct *prev;
641 struct rb_node **rb_link, *rb_parent;
642
643 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
644 &prev, &rb_link, &rb_parent))
645 BUG();
646 __vma_link(mm, vma, prev, rb_link, rb_parent);
647 mm->map_count++;
648}
649
650static __always_inline void __vma_unlink_common(struct mm_struct *mm,
651 struct vm_area_struct *vma,
652 struct vm_area_struct *prev,
653 bool has_prev,
654 struct vm_area_struct *ignore)
655{
656 struct vm_area_struct *next;
657
658 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
659 next = vma->vm_next;
660 if (has_prev)
661 prev->vm_next = next;
662 else {
663 prev = vma->vm_prev;
664 if (prev)
665 prev->vm_next = next;
666 else
667 mm->mmap = next;
668 }
669 if (next)
670 next->vm_prev = prev;
671
672 /* Kill the cache */
673 vmacache_invalidate(mm);
674}
675
676static inline void __vma_unlink_prev(struct mm_struct *mm,
677 struct vm_area_struct *vma,
678 struct vm_area_struct *prev)
679{
680 __vma_unlink_common(mm, vma, prev, true, vma);
681}
682
683/*
684 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
685 * is already present in an i_mmap tree without adjusting the tree.
686 * The following helper function should be used when such adjustments
687 * are necessary. The "insert" vma (if any) is to be inserted
688 * before we drop the necessary locks.
689 */
690int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
691 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
692 struct vm_area_struct *expand)
693{
694 struct mm_struct *mm = vma->vm_mm;
695 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
696 struct address_space *mapping = NULL;
697 struct rb_root_cached *root = NULL;
698 struct anon_vma *anon_vma = NULL;
699 struct file *file = vma->vm_file;
700 bool start_changed = false, end_changed = false;
701 long adjust_next = 0;
702 int remove_next = 0;
703
704 if (next && !insert) {
705 struct vm_area_struct *exporter = NULL, *importer = NULL;
706
707 if (end >= next->vm_end) {
708 /*
709 * vma expands, overlapping all the next, and
710 * perhaps the one after too (mprotect case 6).
711 * The only other cases that gets here are
712 * case 1, case 7 and case 8.
713 */
714 if (next == expand) {
715 /*
716 * The only case where we don't expand "vma"
717 * and we expand "next" instead is case 8.
718 */
719 VM_WARN_ON(end != next->vm_end);
720 /*
721 * remove_next == 3 means we're
722 * removing "vma" and that to do so we
723 * swapped "vma" and "next".
724 */
725 remove_next = 3;
726 VM_WARN_ON(file != next->vm_file);
727 swap(vma, next);
728 } else {
729 VM_WARN_ON(expand != vma);
730 /*
731 * case 1, 6, 7, remove_next == 2 is case 6,
732 * remove_next == 1 is case 1 or 7.
733 */
734 remove_next = 1 + (end > next->vm_end);
735 VM_WARN_ON(remove_next == 2 &&
736 end != next->vm_next->vm_end);
737 VM_WARN_ON(remove_next == 1 &&
738 end != next->vm_end);
739 /* trim end to next, for case 6 first pass */
740 end = next->vm_end;
741 }
742
743 exporter = next;
744 importer = vma;
745
746 /*
747 * If next doesn't have anon_vma, import from vma after
748 * next, if the vma overlaps with it.
749 */
750 if (remove_next == 2 && !next->anon_vma)
751 exporter = next->vm_next;
752
753 } else if (end > next->vm_start) {
754 /*
755 * vma expands, overlapping part of the next:
756 * mprotect case 5 shifting the boundary up.
757 */
758 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
759 exporter = next;
760 importer = vma;
761 VM_WARN_ON(expand != importer);
762 } else if (end < vma->vm_end) {
763 /*
764 * vma shrinks, and !insert tells it's not
765 * split_vma inserting another: so it must be
766 * mprotect case 4 shifting the boundary down.
767 */
768 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
769 exporter = vma;
770 importer = next;
771 VM_WARN_ON(expand != importer);
772 }
773
774 /*
775 * Easily overlooked: when mprotect shifts the boundary,
776 * make sure the expanding vma has anon_vma set if the
777 * shrinking vma had, to cover any anon pages imported.
778 */
779 if (exporter && exporter->anon_vma && !importer->anon_vma) {
780 int error;
781
782 importer->anon_vma = exporter->anon_vma;
783 error = anon_vma_clone(importer, exporter);
784 if (error)
785 return error;
786 }
787 }
788again:
789 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
790
791 if (file) {
792 mapping = file->f_mapping;
793 root = &mapping->i_mmap;
794 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
795
796 if (adjust_next)
797 uprobe_munmap(next, next->vm_start, next->vm_end);
798
799 i_mmap_lock_write(mapping);
800 if (insert) {
801 /*
802 * Put into interval tree now, so instantiated pages
803 * are visible to arm/parisc __flush_dcache_page
804 * throughout; but we cannot insert into address
805 * space until vma start or end is updated.
806 */
807 __vma_link_file(insert);
808 }
809 }
810
811 anon_vma = vma->anon_vma;
812 if (!anon_vma && adjust_next)
813 anon_vma = next->anon_vma;
814 if (anon_vma) {
815 VM_WARN_ON(adjust_next && next->anon_vma &&
816 anon_vma != next->anon_vma);
817 anon_vma_lock_write(anon_vma);
818 anon_vma_interval_tree_pre_update_vma(vma);
819 if (adjust_next)
820 anon_vma_interval_tree_pre_update_vma(next);
821 }
822
823 if (root) {
824 flush_dcache_mmap_lock(mapping);
825 vma_interval_tree_remove(vma, root);
826 if (adjust_next)
827 vma_interval_tree_remove(next, root);
828 }
829
830 if (start != vma->vm_start) {
831 vma->vm_start = start;
832 start_changed = true;
833 }
834 if (end != vma->vm_end) {
835 vma->vm_end = end;
836 end_changed = true;
837 }
838 vma->vm_pgoff = pgoff;
839 if (adjust_next) {
840 next->vm_start += adjust_next << PAGE_SHIFT;
841 next->vm_pgoff += adjust_next;
842 }
843
844 if (root) {
845 if (adjust_next)
846 vma_interval_tree_insert(next, root);
847 vma_interval_tree_insert(vma, root);
848 flush_dcache_mmap_unlock(mapping);
849 }
850
851 if (remove_next) {
852 /*
853 * vma_merge has merged next into vma, and needs
854 * us to remove next before dropping the locks.
855 */
856 if (remove_next != 3)
857 __vma_unlink_prev(mm, next, vma);
858 else
859 /*
860 * vma is not before next if they've been
861 * swapped.
862 *
863 * pre-swap() next->vm_start was reduced so
864 * tell validate_mm_rb to ignore pre-swap()
865 * "next" (which is stored in post-swap()
866 * "vma").
867 */
868 __vma_unlink_common(mm, next, NULL, false, vma);
869 if (file)
870 __remove_shared_vm_struct(next, file, mapping);
871 } else if (insert) {
872 /*
873 * split_vma has split insert from vma, and needs
874 * us to insert it before dropping the locks
875 * (it may either follow vma or precede it).
876 */
877 __insert_vm_struct(mm, insert);
878 } else {
879 if (start_changed)
880 vma_gap_update(vma);
881 if (end_changed) {
882 if (!next)
883 mm->highest_vm_end = vm_end_gap(vma);
884 else if (!adjust_next)
885 vma_gap_update(next);
886 }
887 }
888
889 if (anon_vma) {
890 anon_vma_interval_tree_post_update_vma(vma);
891 if (adjust_next)
892 anon_vma_interval_tree_post_update_vma(next);
893 anon_vma_unlock_write(anon_vma);
894 }
895 if (mapping)
896 i_mmap_unlock_write(mapping);
897
898 if (root) {
899 uprobe_mmap(vma);
900
901 if (adjust_next)
902 uprobe_mmap(next);
903 }
904
905 if (remove_next) {
906 if (file) {
907 uprobe_munmap(next, next->vm_start, next->vm_end);
908 fput(file);
909 }
910 if (next->anon_vma)
911 anon_vma_merge(vma, next);
912 mm->map_count--;
913 mpol_put(vma_policy(next));
914 kmem_cache_free(vm_area_cachep, next);
915 /*
916 * In mprotect's case 6 (see comments on vma_merge),
917 * we must remove another next too. It would clutter
918 * up the code too much to do both in one go.
919 */
920 if (remove_next != 3) {
921 /*
922 * If "next" was removed and vma->vm_end was
923 * expanded (up) over it, in turn
924 * "next->vm_prev->vm_end" changed and the
925 * "vma->vm_next" gap must be updated.
926 */
927 next = vma->vm_next;
928 } else {
929 /*
930 * For the scope of the comment "next" and
931 * "vma" considered pre-swap(): if "vma" was
932 * removed, next->vm_start was expanded (down)
933 * over it and the "next" gap must be updated.
934 * Because of the swap() the post-swap() "vma"
935 * actually points to pre-swap() "next"
936 * (post-swap() "next" as opposed is now a
937 * dangling pointer).
938 */
939 next = vma;
940 }
941 if (remove_next == 2) {
942 remove_next = 1;
943 end = next->vm_end;
944 goto again;
945 }
946 else if (next)
947 vma_gap_update(next);
948 else {
949 /*
950 * If remove_next == 2 we obviously can't
951 * reach this path.
952 *
953 * If remove_next == 3 we can't reach this
954 * path because pre-swap() next is always not
955 * NULL. pre-swap() "next" is not being
956 * removed and its next->vm_end is not altered
957 * (and furthermore "end" already matches
958 * next->vm_end in remove_next == 3).
959 *
960 * We reach this only in the remove_next == 1
961 * case if the "next" vma that was removed was
962 * the highest vma of the mm. However in such
963 * case next->vm_end == "end" and the extended
964 * "vma" has vma->vm_end == next->vm_end so
965 * mm->highest_vm_end doesn't need any update
966 * in remove_next == 1 case.
967 */
968 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
969 }
970 }
971 if (insert && file)
972 uprobe_mmap(insert);
973
974 validate_mm(mm);
975
976 return 0;
977}
978
979/*
980 * If the vma has a ->close operation then the driver probably needs to release
981 * per-vma resources, so we don't attempt to merge those.
982 */
983static inline int is_mergeable_vma(struct vm_area_struct *vma,
984 struct file *file, unsigned long vm_flags,
985 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
986{
987 /*
988 * VM_SOFTDIRTY should not prevent from VMA merging, if we
989 * match the flags but dirty bit -- the caller should mark
990 * merged VMA as dirty. If dirty bit won't be excluded from
991 * comparison, we increase pressue on the memory system forcing
992 * the kernel to generate new VMAs when old one could be
993 * extended instead.
994 */
995 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
996 return 0;
997 if (vma->vm_file != file)
998 return 0;
999 if (vma->vm_ops && vma->vm_ops->close)
1000 return 0;
1001 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1002 return 0;
1003 return 1;
1004}
1005
1006static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1007 struct anon_vma *anon_vma2,
1008 struct vm_area_struct *vma)
1009{
1010 /*
1011 * The list_is_singular() test is to avoid merging VMA cloned from
1012 * parents. This can improve scalability caused by anon_vma lock.
1013 */
1014 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1015 list_is_singular(&vma->anon_vma_chain)))
1016 return 1;
1017 return anon_vma1 == anon_vma2;
1018}
1019
1020/*
1021 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1022 * in front of (at a lower virtual address and file offset than) the vma.
1023 *
1024 * We cannot merge two vmas if they have differently assigned (non-NULL)
1025 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1026 *
1027 * We don't check here for the merged mmap wrapping around the end of pagecache
1028 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1029 * wrap, nor mmaps which cover the final page at index -1UL.
1030 */
1031static int
1032can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1033 struct anon_vma *anon_vma, struct file *file,
1034 pgoff_t vm_pgoff,
1035 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1036{
1037 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1038 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1039 if (vma->vm_pgoff == vm_pgoff)
1040 return 1;
1041 }
1042 return 0;
1043}
1044
1045/*
1046 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1047 * beyond (at a higher virtual address and file offset than) the vma.
1048 *
1049 * We cannot merge two vmas if they have differently assigned (non-NULL)
1050 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1051 */
1052static int
1053can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1054 struct anon_vma *anon_vma, struct file *file,
1055 pgoff_t vm_pgoff,
1056 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1057{
1058 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1059 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1060 pgoff_t vm_pglen;
1061 vm_pglen = vma_pages(vma);
1062 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1063 return 1;
1064 }
1065 return 0;
1066}
1067
1068/*
1069 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1070 * whether that can be merged with its predecessor or its successor.
1071 * Or both (it neatly fills a hole).
1072 *
1073 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1074 * certain not to be mapped by the time vma_merge is called; but when
1075 * called for mprotect, it is certain to be already mapped (either at
1076 * an offset within prev, or at the start of next), and the flags of
1077 * this area are about to be changed to vm_flags - and the no-change
1078 * case has already been eliminated.
1079 *
1080 * The following mprotect cases have to be considered, where AAAA is
1081 * the area passed down from mprotect_fixup, never extending beyond one
1082 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1083 *
1084 * AAAA AAAA AAAA AAAA
1085 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1086 * cannot merge might become might become might become
1087 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1088 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1089 * mremap move: PPPPXXXXXXXX 8
1090 * AAAA
1091 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1092 * might become case 1 below case 2 below case 3 below
1093 *
1094 * It is important for case 8 that the the vma NNNN overlapping the
1095 * region AAAA is never going to extended over XXXX. Instead XXXX must
1096 * be extended in region AAAA and NNNN must be removed. This way in
1097 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1098 * rmap_locks, the properties of the merged vma will be already
1099 * correct for the whole merged range. Some of those properties like
1100 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1101 * be correct for the whole merged range immediately after the
1102 * rmap_locks are released. Otherwise if XXXX would be removed and
1103 * NNNN would be extended over the XXXX range, remove_migration_ptes
1104 * or other rmap walkers (if working on addresses beyond the "end"
1105 * parameter) may establish ptes with the wrong permissions of NNNN
1106 * instead of the right permissions of XXXX.
1107 */
1108struct vm_area_struct *vma_merge(struct mm_struct *mm,
1109 struct vm_area_struct *prev, unsigned long addr,
1110 unsigned long end, unsigned long vm_flags,
1111 struct anon_vma *anon_vma, struct file *file,
1112 pgoff_t pgoff, struct mempolicy *policy,
1113 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1114{
1115 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1116 struct vm_area_struct *area, *next;
1117 int err;
1118
1119 /*
1120 * We later require that vma->vm_flags == vm_flags,
1121 * so this tests vma->vm_flags & VM_SPECIAL, too.
1122 */
1123 if (vm_flags & VM_SPECIAL)
1124 return NULL;
1125
1126 if (prev)
1127 next = prev->vm_next;
1128 else
1129 next = mm->mmap;
1130 area = next;
1131 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1132 next = next->vm_next;
1133
1134 /* verify some invariant that must be enforced by the caller */
1135 VM_WARN_ON(prev && addr <= prev->vm_start);
1136 VM_WARN_ON(area && end > area->vm_end);
1137 VM_WARN_ON(addr >= end);
1138
1139 /*
1140 * Can it merge with the predecessor?
1141 */
1142 if (prev && prev->vm_end == addr &&
1143 mpol_equal(vma_policy(prev), policy) &&
1144 can_vma_merge_after(prev, vm_flags,
1145 anon_vma, file, pgoff,
1146 vm_userfaultfd_ctx)) {
1147 /*
1148 * OK, it can. Can we now merge in the successor as well?
1149 */
1150 if (next && end == next->vm_start &&
1151 mpol_equal(policy, vma_policy(next)) &&
1152 can_vma_merge_before(next, vm_flags,
1153 anon_vma, file,
1154 pgoff+pglen,
1155 vm_userfaultfd_ctx) &&
1156 is_mergeable_anon_vma(prev->anon_vma,
1157 next->anon_vma, NULL)) {
1158 /* cases 1, 6 */
1159 err = __vma_adjust(prev, prev->vm_start,
1160 next->vm_end, prev->vm_pgoff, NULL,
1161 prev);
1162 } else /* cases 2, 5, 7 */
1163 err = __vma_adjust(prev, prev->vm_start,
1164 end, prev->vm_pgoff, NULL, prev);
1165 if (err)
1166 return NULL;
1167 khugepaged_enter_vma_merge(prev, vm_flags);
1168 return prev;
1169 }
1170
1171 /*
1172 * Can this new request be merged in front of next?
1173 */
1174 if (next && end == next->vm_start &&
1175 mpol_equal(policy, vma_policy(next)) &&
1176 can_vma_merge_before(next, vm_flags,
1177 anon_vma, file, pgoff+pglen,
1178 vm_userfaultfd_ctx)) {
1179 if (prev && addr < prev->vm_end) /* case 4 */
1180 err = __vma_adjust(prev, prev->vm_start,
1181 addr, prev->vm_pgoff, NULL, next);
1182 else { /* cases 3, 8 */
1183 err = __vma_adjust(area, addr, next->vm_end,
1184 next->vm_pgoff - pglen, NULL, next);
1185 /*
1186 * In case 3 area is already equal to next and
1187 * this is a noop, but in case 8 "area" has
1188 * been removed and next was expanded over it.
1189 */
1190 area = next;
1191 }
1192 if (err)
1193 return NULL;
1194 khugepaged_enter_vma_merge(area, vm_flags);
1195 return area;
1196 }
1197
1198 return NULL;
1199}
1200
1201/*
1202 * Rough compatbility check to quickly see if it's even worth looking
1203 * at sharing an anon_vma.
1204 *
1205 * They need to have the same vm_file, and the flags can only differ
1206 * in things that mprotect may change.
1207 *
1208 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1209 * we can merge the two vma's. For example, we refuse to merge a vma if
1210 * there is a vm_ops->close() function, because that indicates that the
1211 * driver is doing some kind of reference counting. But that doesn't
1212 * really matter for the anon_vma sharing case.
1213 */
1214static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1215{
1216 return a->vm_end == b->vm_start &&
1217 mpol_equal(vma_policy(a), vma_policy(b)) &&
1218 a->vm_file == b->vm_file &&
1219 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1220 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1221}
1222
1223/*
1224 * Do some basic sanity checking to see if we can re-use the anon_vma
1225 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1226 * the same as 'old', the other will be the new one that is trying
1227 * to share the anon_vma.
1228 *
1229 * NOTE! This runs with mm_sem held for reading, so it is possible that
1230 * the anon_vma of 'old' is concurrently in the process of being set up
1231 * by another page fault trying to merge _that_. But that's ok: if it
1232 * is being set up, that automatically means that it will be a singleton
1233 * acceptable for merging, so we can do all of this optimistically. But
1234 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1235 *
1236 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1237 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1238 * is to return an anon_vma that is "complex" due to having gone through
1239 * a fork).
1240 *
1241 * We also make sure that the two vma's are compatible (adjacent,
1242 * and with the same memory policies). That's all stable, even with just
1243 * a read lock on the mm_sem.
1244 */
1245static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1246{
1247 if (anon_vma_compatible(a, b)) {
1248 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1249
1250 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1251 return anon_vma;
1252 }
1253 return NULL;
1254}
1255
1256/*
1257 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1258 * neighbouring vmas for a suitable anon_vma, before it goes off
1259 * to allocate a new anon_vma. It checks because a repetitive
1260 * sequence of mprotects and faults may otherwise lead to distinct
1261 * anon_vmas being allocated, preventing vma merge in subsequent
1262 * mprotect.
1263 */
1264struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1265{
1266 struct anon_vma *anon_vma;
1267 struct vm_area_struct *near;
1268
1269 near = vma->vm_next;
1270 if (!near)
1271 goto try_prev;
1272
1273 anon_vma = reusable_anon_vma(near, vma, near);
1274 if (anon_vma)
1275 return anon_vma;
1276try_prev:
1277 near = vma->vm_prev;
1278 if (!near)
1279 goto none;
1280
1281 anon_vma = reusable_anon_vma(near, near, vma);
1282 if (anon_vma)
1283 return anon_vma;
1284none:
1285 /*
1286 * There's no absolute need to look only at touching neighbours:
1287 * we could search further afield for "compatible" anon_vmas.
1288 * But it would probably just be a waste of time searching,
1289 * or lead to too many vmas hanging off the same anon_vma.
1290 * We're trying to allow mprotect remerging later on,
1291 * not trying to minimize memory used for anon_vmas.
1292 */
1293 return NULL;
1294}
1295
1296/*
1297 * If a hint addr is less than mmap_min_addr change hint to be as
1298 * low as possible but still greater than mmap_min_addr
1299 */
1300static inline unsigned long round_hint_to_min(unsigned long hint)
1301{
1302 hint &= PAGE_MASK;
1303 if (((void *)hint != NULL) &&
1304 (hint < mmap_min_addr))
1305 return PAGE_ALIGN(mmap_min_addr);
1306 return hint;
1307}
1308
1309static inline int mlock_future_check(struct mm_struct *mm,
1310 unsigned long flags,
1311 unsigned long len)
1312{
1313 unsigned long locked, lock_limit;
1314
1315 /* mlock MCL_FUTURE? */
1316 if (flags & VM_LOCKED) {
1317 locked = len >> PAGE_SHIFT;
1318 locked += mm->locked_vm;
1319 lock_limit = rlimit(RLIMIT_MEMLOCK);
1320 lock_limit >>= PAGE_SHIFT;
1321 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1322 return -EAGAIN;
1323 }
1324 return 0;
1325}
1326
1327static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1328{
1329 if (S_ISREG(inode->i_mode))
1330 return MAX_LFS_FILESIZE;
1331
1332 if (S_ISBLK(inode->i_mode))
1333 return MAX_LFS_FILESIZE;
1334
1335 /* Special "we do even unsigned file positions" case */
1336 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1337 return 0;
1338
1339 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1340 return ULONG_MAX;
1341}
1342
1343static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1344 unsigned long pgoff, unsigned long len)
1345{
1346 u64 maxsize = file_mmap_size_max(file, inode);
1347
1348 if (maxsize && len > maxsize)
1349 return false;
1350 maxsize -= len;
1351 if (pgoff > maxsize >> PAGE_SHIFT)
1352 return false;
1353 return true;
1354}
1355
1356/*
1357 * The caller must hold down_write(¤t->mm->mmap_sem).
1358 */
1359unsigned long do_mmap(struct file *file, unsigned long addr,
1360 unsigned long len, unsigned long prot,
1361 unsigned long flags, vm_flags_t vm_flags,
1362 unsigned long pgoff, unsigned long *populate,
1363 struct list_head *uf)
1364{
1365 struct mm_struct *mm = current->mm;
1366 int pkey = 0;
1367
1368 *populate = 0;
1369
1370 if (!len)
1371 return -EINVAL;
1372
1373 /*
1374 * Does the application expect PROT_READ to imply PROT_EXEC?
1375 *
1376 * (the exception is when the underlying filesystem is noexec
1377 * mounted, in which case we dont add PROT_EXEC.)
1378 */
1379 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1380 if (!(file && path_noexec(&file->f_path)))
1381 prot |= PROT_EXEC;
1382
1383 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1384 if (flags & MAP_FIXED_NOREPLACE)
1385 flags |= MAP_FIXED;
1386
1387 if (!(flags & MAP_FIXED))
1388 addr = round_hint_to_min(addr);
1389
1390 /* Careful about overflows.. */
1391 len = PAGE_ALIGN(len);
1392 if (!len)
1393 return -ENOMEM;
1394
1395 /* offset overflow? */
1396 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1397 return -EOVERFLOW;
1398
1399 /* Too many mappings? */
1400 if (mm->map_count > sysctl_max_map_count)
1401 return -ENOMEM;
1402
1403 /* Obtain the address to map to. we verify (or select) it and ensure
1404 * that it represents a valid section of the address space.
1405 */
1406 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1407 if (offset_in_page(addr))
1408 return addr;
1409
1410 if (flags & MAP_FIXED_NOREPLACE) {
1411 struct vm_area_struct *vma = find_vma(mm, addr);
1412
1413 if (vma && vma->vm_start <= addr)
1414 return -EEXIST;
1415 }
1416
1417 if (prot == PROT_EXEC) {
1418 pkey = execute_only_pkey(mm);
1419 if (pkey < 0)
1420 pkey = 0;
1421 }
1422
1423 /* Do simple checking here so the lower-level routines won't have
1424 * to. we assume access permissions have been handled by the open
1425 * of the memory object, so we don't do any here.
1426 */
1427 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1428 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1429
1430 if (flags & MAP_LOCKED)
1431 if (!can_do_mlock())
1432 return -EPERM;
1433
1434 if (mlock_future_check(mm, vm_flags, len))
1435 return -EAGAIN;
1436
1437 if (file) {
1438 struct inode *inode = file_inode(file);
1439 unsigned long flags_mask;
1440
1441 if (!file_mmap_ok(file, inode, pgoff, len))
1442 return -EOVERFLOW;
1443
1444 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1445
1446 switch (flags & MAP_TYPE) {
1447 case MAP_SHARED:
1448 /*
1449 * Force use of MAP_SHARED_VALIDATE with non-legacy
1450 * flags. E.g. MAP_SYNC is dangerous to use with
1451 * MAP_SHARED as you don't know which consistency model
1452 * you will get. We silently ignore unsupported flags
1453 * with MAP_SHARED to preserve backward compatibility.
1454 */
1455 flags &= LEGACY_MAP_MASK;
1456 /* fall through */
1457 case MAP_SHARED_VALIDATE:
1458 if (flags & ~flags_mask)
1459 return -EOPNOTSUPP;
1460 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1461 return -EACCES;
1462
1463 /*
1464 * Make sure we don't allow writing to an append-only
1465 * file..
1466 */
1467 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1468 return -EACCES;
1469
1470 /*
1471 * Make sure there are no mandatory locks on the file.
1472 */
1473 if (locks_verify_locked(file))
1474 return -EAGAIN;
1475
1476 vm_flags |= VM_SHARED | VM_MAYSHARE;
1477 if (!(file->f_mode & FMODE_WRITE))
1478 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1479
1480 /* fall through */
1481 case MAP_PRIVATE:
1482 if (!(file->f_mode & FMODE_READ))
1483 return -EACCES;
1484 if (path_noexec(&file->f_path)) {
1485 if (vm_flags & VM_EXEC)
1486 return -EPERM;
1487 vm_flags &= ~VM_MAYEXEC;
1488 }
1489
1490 if (!file->f_op->mmap)
1491 return -ENODEV;
1492 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1493 return -EINVAL;
1494 break;
1495
1496 default:
1497 return -EINVAL;
1498 }
1499 } else {
1500 switch (flags & MAP_TYPE) {
1501 case MAP_SHARED:
1502 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1503 return -EINVAL;
1504 /*
1505 * Ignore pgoff.
1506 */
1507 pgoff = 0;
1508 vm_flags |= VM_SHARED | VM_MAYSHARE;
1509 break;
1510 case MAP_PRIVATE:
1511 /*
1512 * Set pgoff according to addr for anon_vma.
1513 */
1514 pgoff = addr >> PAGE_SHIFT;
1515 break;
1516 default:
1517 return -EINVAL;
1518 }
1519 }
1520
1521 /*
1522 * Set 'VM_NORESERVE' if we should not account for the
1523 * memory use of this mapping.
1524 */
1525 if (flags & MAP_NORESERVE) {
1526 /* We honor MAP_NORESERVE if allowed to overcommit */
1527 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1528 vm_flags |= VM_NORESERVE;
1529
1530 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1531 if (file && is_file_hugepages(file))
1532 vm_flags |= VM_NORESERVE;
1533 }
1534
1535 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1536 if (!IS_ERR_VALUE(addr) &&
1537 ((vm_flags & VM_LOCKED) ||
1538 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1539 *populate = len;
1540 return addr;
1541}
1542
1543unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1544 unsigned long prot, unsigned long flags,
1545 unsigned long fd, unsigned long pgoff)
1546{
1547 struct file *file = NULL;
1548 unsigned long retval;
1549
1550 if (!(flags & MAP_ANONYMOUS)) {
1551 audit_mmap_fd(fd, flags);
1552 file = fget(fd);
1553 if (!file)
1554 return -EBADF;
1555 if (is_file_hugepages(file))
1556 len = ALIGN(len, huge_page_size(hstate_file(file)));
1557 retval = -EINVAL;
1558 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1559 goto out_fput;
1560 } else if (flags & MAP_HUGETLB) {
1561 struct user_struct *user = NULL;
1562 struct hstate *hs;
1563
1564 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1565 if (!hs)
1566 return -EINVAL;
1567
1568 len = ALIGN(len, huge_page_size(hs));
1569 /*
1570 * VM_NORESERVE is used because the reservations will be
1571 * taken when vm_ops->mmap() is called
1572 * A dummy user value is used because we are not locking
1573 * memory so no accounting is necessary
1574 */
1575 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1576 VM_NORESERVE,
1577 &user, HUGETLB_ANONHUGE_INODE,
1578 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1579 if (IS_ERR(file))
1580 return PTR_ERR(file);
1581 }
1582
1583 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1584
1585 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1586out_fput:
1587 if (file)
1588 fput(file);
1589 return retval;
1590}
1591
1592SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1593 unsigned long, prot, unsigned long, flags,
1594 unsigned long, fd, unsigned long, pgoff)
1595{
1596 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1597}
1598
1599#ifdef __ARCH_WANT_SYS_OLD_MMAP
1600struct mmap_arg_struct {
1601 unsigned long addr;
1602 unsigned long len;
1603 unsigned long prot;
1604 unsigned long flags;
1605 unsigned long fd;
1606 unsigned long offset;
1607};
1608
1609SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1610{
1611 struct mmap_arg_struct a;
1612
1613 if (copy_from_user(&a, arg, sizeof(a)))
1614 return -EFAULT;
1615 if (offset_in_page(a.offset))
1616 return -EINVAL;
1617
1618 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1619 a.offset >> PAGE_SHIFT);
1620}
1621#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1622
1623/*
1624 * Some shared mappigns will want the pages marked read-only
1625 * to track write events. If so, we'll downgrade vm_page_prot
1626 * to the private version (using protection_map[] without the
1627 * VM_SHARED bit).
1628 */
1629int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1630{
1631 vm_flags_t vm_flags = vma->vm_flags;
1632 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1633
1634 /* If it was private or non-writable, the write bit is already clear */
1635 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1636 return 0;
1637
1638 /* The backer wishes to know when pages are first written to? */
1639 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1640 return 1;
1641
1642 /* The open routine did something to the protections that pgprot_modify
1643 * won't preserve? */
1644 if (pgprot_val(vm_page_prot) !=
1645 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1646 return 0;
1647
1648 /* Do we need to track softdirty? */
1649 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1650 return 1;
1651
1652 /* Specialty mapping? */
1653 if (vm_flags & VM_PFNMAP)
1654 return 0;
1655
1656 /* Can the mapping track the dirty pages? */
1657 return vma->vm_file && vma->vm_file->f_mapping &&
1658 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1659}
1660
1661/*
1662 * We account for memory if it's a private writeable mapping,
1663 * not hugepages and VM_NORESERVE wasn't set.
1664 */
1665static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1666{
1667 /*
1668 * hugetlb has its own accounting separate from the core VM
1669 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1670 */
1671 if (file && is_file_hugepages(file))
1672 return 0;
1673
1674 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1675}
1676
1677unsigned long mmap_region(struct file *file, unsigned long addr,
1678 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1679 struct list_head *uf)
1680{
1681 struct mm_struct *mm = current->mm;
1682 struct vm_area_struct *vma, *prev;
1683 int error;
1684 struct rb_node **rb_link, *rb_parent;
1685 unsigned long charged = 0;
1686
1687 /* Check against address space limit. */
1688 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1689 unsigned long nr_pages;
1690
1691 /*
1692 * MAP_FIXED may remove pages of mappings that intersects with
1693 * requested mapping. Account for the pages it would unmap.
1694 */
1695 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1696
1697 if (!may_expand_vm(mm, vm_flags,
1698 (len >> PAGE_SHIFT) - nr_pages))
1699 return -ENOMEM;
1700 }
1701
1702 /* Clear old maps */
1703 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1704 &rb_parent)) {
1705 if (do_munmap(mm, addr, len, uf))
1706 return -ENOMEM;
1707 }
1708
1709 /*
1710 * Private writable mapping: check memory availability
1711 */
1712 if (accountable_mapping(file, vm_flags)) {
1713 charged = len >> PAGE_SHIFT;
1714 if (security_vm_enough_memory_mm(mm, charged))
1715 return -ENOMEM;
1716 vm_flags |= VM_ACCOUNT;
1717 }
1718
1719 /*
1720 * Can we just expand an old mapping?
1721 */
1722 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1723 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1724 if (vma)
1725 goto out;
1726
1727 /*
1728 * Determine the object being mapped and call the appropriate
1729 * specific mapper. the address has already been validated, but
1730 * not unmapped, but the maps are removed from the list.
1731 */
1732 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1733 if (!vma) {
1734 error = -ENOMEM;
1735 goto unacct_error;
1736 }
1737
1738 vma->vm_mm = mm;
1739 vma->vm_start = addr;
1740 vma->vm_end = addr + len;
1741 vma->vm_flags = vm_flags;
1742 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1743 vma->vm_pgoff = pgoff;
1744 INIT_LIST_HEAD(&vma->anon_vma_chain);
1745
1746 if (file) {
1747 if (vm_flags & VM_DENYWRITE) {
1748 error = deny_write_access(file);
1749 if (error)
1750 goto free_vma;
1751 }
1752 if (vm_flags & VM_SHARED) {
1753 error = mapping_map_writable(file->f_mapping);
1754 if (error)
1755 goto allow_write_and_free_vma;
1756 }
1757
1758 /* ->mmap() can change vma->vm_file, but must guarantee that
1759 * vma_link() below can deny write-access if VM_DENYWRITE is set
1760 * and map writably if VM_SHARED is set. This usually means the
1761 * new file must not have been exposed to user-space, yet.
1762 */
1763 vma->vm_file = get_file(file);
1764 error = call_mmap(file, vma);
1765 if (error)
1766 goto unmap_and_free_vma;
1767
1768 /* Can addr have changed??
1769 *
1770 * Answer: Yes, several device drivers can do it in their
1771 * f_op->mmap method. -DaveM
1772 * Bug: If addr is changed, prev, rb_link, rb_parent should
1773 * be updated for vma_link()
1774 */
1775 WARN_ON_ONCE(addr != vma->vm_start);
1776
1777 addr = vma->vm_start;
1778 vm_flags = vma->vm_flags;
1779 } else if (vm_flags & VM_SHARED) {
1780 error = shmem_zero_setup(vma);
1781 if (error)
1782 goto free_vma;
1783 }
1784
1785 vma_link(mm, vma, prev, rb_link, rb_parent);
1786 /* Once vma denies write, undo our temporary denial count */
1787 if (file) {
1788 if (vm_flags & VM_SHARED)
1789 mapping_unmap_writable(file->f_mapping);
1790 if (vm_flags & VM_DENYWRITE)
1791 allow_write_access(file);
1792 }
1793 file = vma->vm_file;
1794out:
1795 perf_event_mmap(vma);
1796
1797 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1798 if (vm_flags & VM_LOCKED) {
1799 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1800 vma == get_gate_vma(current->mm)))
1801 mm->locked_vm += (len >> PAGE_SHIFT);
1802 else
1803 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1804 }
1805
1806 if (file)
1807 uprobe_mmap(vma);
1808
1809 /*
1810 * New (or expanded) vma always get soft dirty status.
1811 * Otherwise user-space soft-dirty page tracker won't
1812 * be able to distinguish situation when vma area unmapped,
1813 * then new mapped in-place (which must be aimed as
1814 * a completely new data area).
1815 */
1816 vma->vm_flags |= VM_SOFTDIRTY;
1817
1818 vma_set_page_prot(vma);
1819
1820 return addr;
1821
1822unmap_and_free_vma:
1823 vma->vm_file = NULL;
1824 fput(file);
1825
1826 /* Undo any partial mapping done by a device driver. */
1827 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1828 charged = 0;
1829 if (vm_flags & VM_SHARED)
1830 mapping_unmap_writable(file->f_mapping);
1831allow_write_and_free_vma:
1832 if (vm_flags & VM_DENYWRITE)
1833 allow_write_access(file);
1834free_vma:
1835 kmem_cache_free(vm_area_cachep, vma);
1836unacct_error:
1837 if (charged)
1838 vm_unacct_memory(charged);
1839 return error;
1840}
1841
1842unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1843{
1844 /*
1845 * We implement the search by looking for an rbtree node that
1846 * immediately follows a suitable gap. That is,
1847 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1848 * - gap_end = vma->vm_start >= info->low_limit + length;
1849 * - gap_end - gap_start >= length
1850 */
1851
1852 struct mm_struct *mm = current->mm;
1853 struct vm_area_struct *vma;
1854 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1855
1856 /* Adjust search length to account for worst case alignment overhead */
1857 length = info->length + info->align_mask;
1858 if (length < info->length)
1859 return -ENOMEM;
1860
1861 /* Adjust search limits by the desired length */
1862 if (info->high_limit < length)
1863 return -ENOMEM;
1864 high_limit = info->high_limit - length;
1865
1866 if (info->low_limit > high_limit)
1867 return -ENOMEM;
1868 low_limit = info->low_limit + length;
1869
1870 /* Check if rbtree root looks promising */
1871 if (RB_EMPTY_ROOT(&mm->mm_rb))
1872 goto check_highest;
1873 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1874 if (vma->rb_subtree_gap < length)
1875 goto check_highest;
1876
1877 while (true) {
1878 /* Visit left subtree if it looks promising */
1879 gap_end = vm_start_gap(vma);
1880 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1881 struct vm_area_struct *left =
1882 rb_entry(vma->vm_rb.rb_left,
1883 struct vm_area_struct, vm_rb);
1884 if (left->rb_subtree_gap >= length) {
1885 vma = left;
1886 continue;
1887 }
1888 }
1889
1890 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1891check_current:
1892 /* Check if current node has a suitable gap */
1893 if (gap_start > high_limit)
1894 return -ENOMEM;
1895 if (gap_end >= low_limit &&
1896 gap_end > gap_start && gap_end - gap_start >= length)
1897 goto found;
1898
1899 /* Visit right subtree if it looks promising */
1900 if (vma->vm_rb.rb_right) {
1901 struct vm_area_struct *right =
1902 rb_entry(vma->vm_rb.rb_right,
1903 struct vm_area_struct, vm_rb);
1904 if (right->rb_subtree_gap >= length) {
1905 vma = right;
1906 continue;
1907 }
1908 }
1909
1910 /* Go back up the rbtree to find next candidate node */
1911 while (true) {
1912 struct rb_node *prev = &vma->vm_rb;
1913 if (!rb_parent(prev))
1914 goto check_highest;
1915 vma = rb_entry(rb_parent(prev),
1916 struct vm_area_struct, vm_rb);
1917 if (prev == vma->vm_rb.rb_left) {
1918 gap_start = vm_end_gap(vma->vm_prev);
1919 gap_end = vm_start_gap(vma);
1920 goto check_current;
1921 }
1922 }
1923 }
1924
1925check_highest:
1926 /* Check highest gap, which does not precede any rbtree node */
1927 gap_start = mm->highest_vm_end;
1928 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1929 if (gap_start > high_limit)
1930 return -ENOMEM;
1931
1932found:
1933 /* We found a suitable gap. Clip it with the original low_limit. */
1934 if (gap_start < info->low_limit)
1935 gap_start = info->low_limit;
1936
1937 /* Adjust gap address to the desired alignment */
1938 gap_start += (info->align_offset - gap_start) & info->align_mask;
1939
1940 VM_BUG_ON(gap_start + info->length > info->high_limit);
1941 VM_BUG_ON(gap_start + info->length > gap_end);
1942 return gap_start;
1943}
1944
1945unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1946{
1947 struct mm_struct *mm = current->mm;
1948 struct vm_area_struct *vma;
1949 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1950
1951 /* Adjust search length to account for worst case alignment overhead */
1952 length = info->length + info->align_mask;
1953 if (length < info->length)
1954 return -ENOMEM;
1955
1956 /*
1957 * Adjust search limits by the desired length.
1958 * See implementation comment at top of unmapped_area().
1959 */
1960 gap_end = info->high_limit;
1961 if (gap_end < length)
1962 return -ENOMEM;
1963 high_limit = gap_end - length;
1964
1965 if (info->low_limit > high_limit)
1966 return -ENOMEM;
1967 low_limit = info->low_limit + length;
1968
1969 /* Check highest gap, which does not precede any rbtree node */
1970 gap_start = mm->highest_vm_end;
1971 if (gap_start <= high_limit)
1972 goto found_highest;
1973
1974 /* Check if rbtree root looks promising */
1975 if (RB_EMPTY_ROOT(&mm->mm_rb))
1976 return -ENOMEM;
1977 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1978 if (vma->rb_subtree_gap < length)
1979 return -ENOMEM;
1980
1981 while (true) {
1982 /* Visit right subtree if it looks promising */
1983 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1984 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1985 struct vm_area_struct *right =
1986 rb_entry(vma->vm_rb.rb_right,
1987 struct vm_area_struct, vm_rb);
1988 if (right->rb_subtree_gap >= length) {
1989 vma = right;
1990 continue;
1991 }
1992 }
1993
1994check_current:
1995 /* Check if current node has a suitable gap */
1996 gap_end = vm_start_gap(vma);
1997 if (gap_end < low_limit)
1998 return -ENOMEM;
1999 if (gap_start <= high_limit &&
2000 gap_end > gap_start && gap_end - gap_start >= length)
2001 goto found;
2002
2003 /* Visit left subtree if it looks promising */
2004 if (vma->vm_rb.rb_left) {
2005 struct vm_area_struct *left =
2006 rb_entry(vma->vm_rb.rb_left,
2007 struct vm_area_struct, vm_rb);
2008 if (left->rb_subtree_gap >= length) {
2009 vma = left;
2010 continue;
2011 }
2012 }
2013
2014 /* Go back up the rbtree to find next candidate node */
2015 while (true) {
2016 struct rb_node *prev = &vma->vm_rb;
2017 if (!rb_parent(prev))
2018 return -ENOMEM;
2019 vma = rb_entry(rb_parent(prev),
2020 struct vm_area_struct, vm_rb);
2021 if (prev == vma->vm_rb.rb_right) {
2022 gap_start = vma->vm_prev ?
2023 vm_end_gap(vma->vm_prev) : 0;
2024 goto check_current;
2025 }
2026 }
2027 }
2028
2029found:
2030 /* We found a suitable gap. Clip it with the original high_limit. */
2031 if (gap_end > info->high_limit)
2032 gap_end = info->high_limit;
2033
2034found_highest:
2035 /* Compute highest gap address at the desired alignment */
2036 gap_end -= info->length;
2037 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2038
2039 VM_BUG_ON(gap_end < info->low_limit);
2040 VM_BUG_ON(gap_end < gap_start);
2041 return gap_end;
2042}
2043
2044/* Get an address range which is currently unmapped.
2045 * For shmat() with addr=0.
2046 *
2047 * Ugly calling convention alert:
2048 * Return value with the low bits set means error value,
2049 * ie
2050 * if (ret & ~PAGE_MASK)
2051 * error = ret;
2052 *
2053 * This function "knows" that -ENOMEM has the bits set.
2054 */
2055#ifndef HAVE_ARCH_UNMAPPED_AREA
2056unsigned long
2057arch_get_unmapped_area(struct file *filp, unsigned long addr,
2058 unsigned long len, unsigned long pgoff, unsigned long flags)
2059{
2060 struct mm_struct *mm = current->mm;
2061 struct vm_area_struct *vma, *prev;
2062 struct vm_unmapped_area_info info;
2063
2064 if (len > TASK_SIZE - mmap_min_addr)
2065 return -ENOMEM;
2066
2067 if (flags & MAP_FIXED)
2068 return addr;
2069
2070 if (addr) {
2071 addr = PAGE_ALIGN(addr);
2072 vma = find_vma_prev(mm, addr, &prev);
2073 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2074 (!vma || addr + len <= vm_start_gap(vma)) &&
2075 (!prev || addr >= vm_end_gap(prev)))
2076 return addr;
2077 }
2078
2079 info.flags = 0;
2080 info.length = len;
2081 info.low_limit = mm->mmap_base;
2082 info.high_limit = TASK_SIZE;
2083 info.align_mask = 0;
2084 return vm_unmapped_area(&info);
2085}
2086#endif
2087
2088/*
2089 * This mmap-allocator allocates new areas top-down from below the
2090 * stack's low limit (the base):
2091 */
2092#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2093unsigned long
2094arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2095 const unsigned long len, const unsigned long pgoff,
2096 const unsigned long flags)
2097{
2098 struct vm_area_struct *vma, *prev;
2099 struct mm_struct *mm = current->mm;
2100 unsigned long addr = addr0;
2101 struct vm_unmapped_area_info info;
2102
2103 /* requested length too big for entire address space */
2104 if (len > TASK_SIZE - mmap_min_addr)
2105 return -ENOMEM;
2106
2107 if (flags & MAP_FIXED)
2108 return addr;
2109
2110 /* requesting a specific address */
2111 if (addr) {
2112 addr = PAGE_ALIGN(addr);
2113 vma = find_vma_prev(mm, addr, &prev);
2114 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2115 (!vma || addr + len <= vm_start_gap(vma)) &&
2116 (!prev || addr >= vm_end_gap(prev)))
2117 return addr;
2118 }
2119
2120 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2121 info.length = len;
2122 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2123 info.high_limit = mm->mmap_base;
2124 info.align_mask = 0;
2125 addr = vm_unmapped_area(&info);
2126
2127 /*
2128 * A failed mmap() very likely causes application failure,
2129 * so fall back to the bottom-up function here. This scenario
2130 * can happen with large stack limits and large mmap()
2131 * allocations.
2132 */
2133 if (offset_in_page(addr)) {
2134 VM_BUG_ON(addr != -ENOMEM);
2135 info.flags = 0;
2136 info.low_limit = TASK_UNMAPPED_BASE;
2137 info.high_limit = TASK_SIZE;
2138 addr = vm_unmapped_area(&info);
2139 }
2140
2141 return addr;
2142}
2143#endif
2144
2145unsigned long
2146get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2147 unsigned long pgoff, unsigned long flags)
2148{
2149 unsigned long (*get_area)(struct file *, unsigned long,
2150 unsigned long, unsigned long, unsigned long);
2151
2152 unsigned long error = arch_mmap_check(addr, len, flags);
2153 if (error)
2154 return error;
2155
2156 /* Careful about overflows.. */
2157 if (len > TASK_SIZE)
2158 return -ENOMEM;
2159
2160 get_area = current->mm->get_unmapped_area;
2161 if (file) {
2162 if (file->f_op->get_unmapped_area)
2163 get_area = file->f_op->get_unmapped_area;
2164 } else if (flags & MAP_SHARED) {
2165 /*
2166 * mmap_region() will call shmem_zero_setup() to create a file,
2167 * so use shmem's get_unmapped_area in case it can be huge.
2168 * do_mmap_pgoff() will clear pgoff, so match alignment.
2169 */
2170 pgoff = 0;
2171 get_area = shmem_get_unmapped_area;
2172 }
2173
2174 addr = get_area(file, addr, len, pgoff, flags);
2175 if (IS_ERR_VALUE(addr))
2176 return addr;
2177
2178 if (addr > TASK_SIZE - len)
2179 return -ENOMEM;
2180 if (offset_in_page(addr))
2181 return -EINVAL;
2182
2183 error = security_mmap_addr(addr);
2184 return error ? error : addr;
2185}
2186
2187EXPORT_SYMBOL(get_unmapped_area);
2188
2189/* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2190struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2191{
2192 struct rb_node *rb_node;
2193 struct vm_area_struct *vma;
2194
2195 /* Check the cache first. */
2196 vma = vmacache_find(mm, addr);
2197 if (likely(vma))
2198 return vma;
2199
2200 rb_node = mm->mm_rb.rb_node;
2201
2202 while (rb_node) {
2203 struct vm_area_struct *tmp;
2204
2205 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2206
2207 if (tmp->vm_end > addr) {
2208 vma = tmp;
2209 if (tmp->vm_start <= addr)
2210 break;
2211 rb_node = rb_node->rb_left;
2212 } else
2213 rb_node = rb_node->rb_right;
2214 }
2215
2216 if (vma)
2217 vmacache_update(addr, vma);
2218 return vma;
2219}
2220
2221EXPORT_SYMBOL(find_vma);
2222
2223/*
2224 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2225 */
2226struct vm_area_struct *
2227find_vma_prev(struct mm_struct *mm, unsigned long addr,
2228 struct vm_area_struct **pprev)
2229{
2230 struct vm_area_struct *vma;
2231
2232 vma = find_vma(mm, addr);
2233 if (vma) {
2234 *pprev = vma->vm_prev;
2235 } else {
2236 struct rb_node *rb_node = mm->mm_rb.rb_node;
2237 *pprev = NULL;
2238 while (rb_node) {
2239 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2240 rb_node = rb_node->rb_right;
2241 }
2242 }
2243 return vma;
2244}
2245
2246/*
2247 * Verify that the stack growth is acceptable and
2248 * update accounting. This is shared with both the
2249 * grow-up and grow-down cases.
2250 */
2251static int acct_stack_growth(struct vm_area_struct *vma,
2252 unsigned long size, unsigned long grow)
2253{
2254 struct mm_struct *mm = vma->vm_mm;
2255 unsigned long new_start;
2256
2257 /* address space limit tests */
2258 if (!may_expand_vm(mm, vma->vm_flags, grow))
2259 return -ENOMEM;
2260
2261 /* Stack limit test */
2262 if (size > rlimit(RLIMIT_STACK))
2263 return -ENOMEM;
2264
2265 /* mlock limit tests */
2266 if (vma->vm_flags & VM_LOCKED) {
2267 unsigned long locked;
2268 unsigned long limit;
2269 locked = mm->locked_vm + grow;
2270 limit = rlimit(RLIMIT_MEMLOCK);
2271 limit >>= PAGE_SHIFT;
2272 if (locked > limit && !capable(CAP_IPC_LOCK))
2273 return -ENOMEM;
2274 }
2275
2276 /* Check to ensure the stack will not grow into a hugetlb-only region */
2277 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2278 vma->vm_end - size;
2279 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2280 return -EFAULT;
2281
2282 /*
2283 * Overcommit.. This must be the final test, as it will
2284 * update security statistics.
2285 */
2286 if (security_vm_enough_memory_mm(mm, grow))
2287 return -ENOMEM;
2288
2289 return 0;
2290}
2291
2292#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2293/*
2294 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2295 * vma is the last one with address > vma->vm_end. Have to extend vma.
2296 */
2297int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2298{
2299 struct mm_struct *mm = vma->vm_mm;
2300 struct vm_area_struct *next;
2301 unsigned long gap_addr;
2302 int error = 0;
2303
2304 if (!(vma->vm_flags & VM_GROWSUP))
2305 return -EFAULT;
2306
2307 /* Guard against exceeding limits of the address space. */
2308 address &= PAGE_MASK;
2309 if (address >= (TASK_SIZE & PAGE_MASK))
2310 return -ENOMEM;
2311 address += PAGE_SIZE;
2312
2313 /* Enforce stack_guard_gap */
2314 gap_addr = address + stack_guard_gap;
2315
2316 /* Guard against overflow */
2317 if (gap_addr < address || gap_addr > TASK_SIZE)
2318 gap_addr = TASK_SIZE;
2319
2320 next = vma->vm_next;
2321 if (next && next->vm_start < gap_addr &&
2322 (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2323 if (!(next->vm_flags & VM_GROWSUP))
2324 return -ENOMEM;
2325 /* Check that both stack segments have the same anon_vma? */
2326 }
2327
2328 /* We must make sure the anon_vma is allocated. */
2329 if (unlikely(anon_vma_prepare(vma)))
2330 return -ENOMEM;
2331
2332 /*
2333 * vma->vm_start/vm_end cannot change under us because the caller
2334 * is required to hold the mmap_sem in read mode. We need the
2335 * anon_vma lock to serialize against concurrent expand_stacks.
2336 */
2337 anon_vma_lock_write(vma->anon_vma);
2338
2339 /* Somebody else might have raced and expanded it already */
2340 if (address > vma->vm_end) {
2341 unsigned long size, grow;
2342
2343 size = address - vma->vm_start;
2344 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2345
2346 error = -ENOMEM;
2347 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2348 error = acct_stack_growth(vma, size, grow);
2349 if (!error) {
2350 /*
2351 * vma_gap_update() doesn't support concurrent
2352 * updates, but we only hold a shared mmap_sem
2353 * lock here, so we need to protect against
2354 * concurrent vma expansions.
2355 * anon_vma_lock_write() doesn't help here, as
2356 * we don't guarantee that all growable vmas
2357 * in a mm share the same root anon vma.
2358 * So, we reuse mm->page_table_lock to guard
2359 * against concurrent vma expansions.
2360 */
2361 spin_lock(&mm->page_table_lock);
2362 if (vma->vm_flags & VM_LOCKED)
2363 mm->locked_vm += grow;
2364 vm_stat_account(mm, vma->vm_flags, grow);
2365 anon_vma_interval_tree_pre_update_vma(vma);
2366 vma->vm_end = address;
2367 anon_vma_interval_tree_post_update_vma(vma);
2368 if (vma->vm_next)
2369 vma_gap_update(vma->vm_next);
2370 else
2371 mm->highest_vm_end = vm_end_gap(vma);
2372 spin_unlock(&mm->page_table_lock);
2373
2374 perf_event_mmap(vma);
2375 }
2376 }
2377 }
2378 anon_vma_unlock_write(vma->anon_vma);
2379 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2380 validate_mm(mm);
2381 return error;
2382}
2383#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2384
2385/*
2386 * vma is the first one with address < vma->vm_start. Have to extend vma.
2387 */
2388int expand_downwards(struct vm_area_struct *vma,
2389 unsigned long address)
2390{
2391 struct mm_struct *mm = vma->vm_mm;
2392 struct vm_area_struct *prev;
2393 int error;
2394
2395 address &= PAGE_MASK;
2396 error = security_mmap_addr(address);
2397 if (error)
2398 return error;
2399
2400 /* Enforce stack_guard_gap */
2401 prev = vma->vm_prev;
2402 /* Check that both stack segments have the same anon_vma? */
2403 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2404 (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2405 if (address - prev->vm_end < stack_guard_gap)
2406 return -ENOMEM;
2407 }
2408
2409 /* We must make sure the anon_vma is allocated. */
2410 if (unlikely(anon_vma_prepare(vma)))
2411 return -ENOMEM;
2412
2413 /*
2414 * vma->vm_start/vm_end cannot change under us because the caller
2415 * is required to hold the mmap_sem in read mode. We need the
2416 * anon_vma lock to serialize against concurrent expand_stacks.
2417 */
2418 anon_vma_lock_write(vma->anon_vma);
2419
2420 /* Somebody else might have raced and expanded it already */
2421 if (address < vma->vm_start) {
2422 unsigned long size, grow;
2423
2424 size = vma->vm_end - address;
2425 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2426
2427 error = -ENOMEM;
2428 if (grow <= vma->vm_pgoff) {
2429 error = acct_stack_growth(vma, size, grow);
2430 if (!error) {
2431 /*
2432 * vma_gap_update() doesn't support concurrent
2433 * updates, but we only hold a shared mmap_sem
2434 * lock here, so we need to protect against
2435 * concurrent vma expansions.
2436 * anon_vma_lock_write() doesn't help here, as
2437 * we don't guarantee that all growable vmas
2438 * in a mm share the same root anon vma.
2439 * So, we reuse mm->page_table_lock to guard
2440 * against concurrent vma expansions.
2441 */
2442 spin_lock(&mm->page_table_lock);
2443 if (vma->vm_flags & VM_LOCKED)
2444 mm->locked_vm += grow;
2445 vm_stat_account(mm, vma->vm_flags, grow);
2446 anon_vma_interval_tree_pre_update_vma(vma);
2447 vma->vm_start = address;
2448 vma->vm_pgoff -= grow;
2449 anon_vma_interval_tree_post_update_vma(vma);
2450 vma_gap_update(vma);
2451 spin_unlock(&mm->page_table_lock);
2452
2453 perf_event_mmap(vma);
2454 }
2455 }
2456 }
2457 anon_vma_unlock_write(vma->anon_vma);
2458 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2459 validate_mm(mm);
2460 return error;
2461}
2462
2463/* enforced gap between the expanding stack and other mappings. */
2464unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2465
2466static int __init cmdline_parse_stack_guard_gap(char *p)
2467{
2468 unsigned long val;
2469 char *endptr;
2470
2471 val = simple_strtoul(p, &endptr, 10);
2472 if (!*endptr)
2473 stack_guard_gap = val << PAGE_SHIFT;
2474
2475 return 0;
2476}
2477__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2478
2479#ifdef CONFIG_STACK_GROWSUP
2480int expand_stack(struct vm_area_struct *vma, unsigned long address)
2481{
2482 return expand_upwards(vma, address);
2483}
2484
2485struct vm_area_struct *
2486find_extend_vma(struct mm_struct *mm, unsigned long addr)
2487{
2488 struct vm_area_struct *vma, *prev;
2489
2490 addr &= PAGE_MASK;
2491 vma = find_vma_prev(mm, addr, &prev);
2492 if (vma && (vma->vm_start <= addr))
2493 return vma;
2494 if (!prev || expand_stack(prev, addr))
2495 return NULL;
2496 if (prev->vm_flags & VM_LOCKED)
2497 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2498 return prev;
2499}
2500#else
2501int expand_stack(struct vm_area_struct *vma, unsigned long address)
2502{
2503 return expand_downwards(vma, address);
2504}
2505
2506struct vm_area_struct *
2507find_extend_vma(struct mm_struct *mm, unsigned long addr)
2508{
2509 struct vm_area_struct *vma;
2510 unsigned long start;
2511
2512 addr &= PAGE_MASK;
2513 vma = find_vma(mm, addr);
2514 if (!vma)
2515 return NULL;
2516 if (vma->vm_start <= addr)
2517 return vma;
2518 if (!(vma->vm_flags & VM_GROWSDOWN))
2519 return NULL;
2520 start = vma->vm_start;
2521 if (expand_stack(vma, addr))
2522 return NULL;
2523 if (vma->vm_flags & VM_LOCKED)
2524 populate_vma_page_range(vma, addr, start, NULL);
2525 return vma;
2526}
2527#endif
2528
2529EXPORT_SYMBOL_GPL(find_extend_vma);
2530
2531/*
2532 * Ok - we have the memory areas we should free on the vma list,
2533 * so release them, and do the vma updates.
2534 *
2535 * Called with the mm semaphore held.
2536 */
2537static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2538{
2539 unsigned long nr_accounted = 0;
2540
2541 /* Update high watermark before we lower total_vm */
2542 update_hiwater_vm(mm);
2543 do {
2544 long nrpages = vma_pages(vma);
2545
2546 if (vma->vm_flags & VM_ACCOUNT)
2547 nr_accounted += nrpages;
2548 vm_stat_account(mm, vma->vm_flags, -nrpages);
2549 vma = remove_vma(vma);
2550 } while (vma);
2551 vm_unacct_memory(nr_accounted);
2552 validate_mm(mm);
2553}
2554
2555/*
2556 * Get rid of page table information in the indicated region.
2557 *
2558 * Called with the mm semaphore held.
2559 */
2560static void unmap_region(struct mm_struct *mm,
2561 struct vm_area_struct *vma, struct vm_area_struct *prev,
2562 unsigned long start, unsigned long end)
2563{
2564 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2565 struct mmu_gather tlb;
2566
2567 lru_add_drain();
2568 tlb_gather_mmu(&tlb, mm, start, end);
2569 update_hiwater_rss(mm);
2570 unmap_vmas(&tlb, vma, start, end);
2571 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2572 next ? next->vm_start : USER_PGTABLES_CEILING);
2573 tlb_finish_mmu(&tlb, start, end);
2574}
2575
2576/*
2577 * Create a list of vma's touched by the unmap, removing them from the mm's
2578 * vma list as we go..
2579 */
2580static void
2581detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2582 struct vm_area_struct *prev, unsigned long end)
2583{
2584 struct vm_area_struct **insertion_point;
2585 struct vm_area_struct *tail_vma = NULL;
2586
2587 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2588 vma->vm_prev = NULL;
2589 do {
2590 vma_rb_erase(vma, &mm->mm_rb);
2591 mm->map_count--;
2592 tail_vma = vma;
2593 vma = vma->vm_next;
2594 } while (vma && vma->vm_start < end);
2595 *insertion_point = vma;
2596 if (vma) {
2597 vma->vm_prev = prev;
2598 vma_gap_update(vma);
2599 } else
2600 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2601 tail_vma->vm_next = NULL;
2602
2603 /* Kill the cache */
2604 vmacache_invalidate(mm);
2605}
2606
2607/*
2608 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2609 * has already been checked or doesn't make sense to fail.
2610 */
2611int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2612 unsigned long addr, int new_below)
2613{
2614 struct vm_area_struct *new;
2615 int err;
2616
2617 if (vma->vm_ops && vma->vm_ops->split) {
2618 err = vma->vm_ops->split(vma, addr);
2619 if (err)
2620 return err;
2621 }
2622
2623 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2624 if (!new)
2625 return -ENOMEM;
2626
2627 /* most fields are the same, copy all, and then fixup */
2628 *new = *vma;
2629
2630 INIT_LIST_HEAD(&new->anon_vma_chain);
2631
2632 if (new_below)
2633 new->vm_end = addr;
2634 else {
2635 new->vm_start = addr;
2636 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2637 }
2638
2639 err = vma_dup_policy(vma, new);
2640 if (err)
2641 goto out_free_vma;
2642
2643 err = anon_vma_clone(new, vma);
2644 if (err)
2645 goto out_free_mpol;
2646
2647 if (new->vm_file)
2648 get_file(new->vm_file);
2649
2650 if (new->vm_ops && new->vm_ops->open)
2651 new->vm_ops->open(new);
2652
2653 if (new_below)
2654 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2655 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2656 else
2657 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2658
2659 /* Success. */
2660 if (!err)
2661 return 0;
2662
2663 /* Clean everything up if vma_adjust failed. */
2664 if (new->vm_ops && new->vm_ops->close)
2665 new->vm_ops->close(new);
2666 if (new->vm_file)
2667 fput(new->vm_file);
2668 unlink_anon_vmas(new);
2669 out_free_mpol:
2670 mpol_put(vma_policy(new));
2671 out_free_vma:
2672 kmem_cache_free(vm_area_cachep, new);
2673 return err;
2674}
2675
2676/*
2677 * Split a vma into two pieces at address 'addr', a new vma is allocated
2678 * either for the first part or the tail.
2679 */
2680int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2681 unsigned long addr, int new_below)
2682{
2683 if (mm->map_count >= sysctl_max_map_count)
2684 return -ENOMEM;
2685
2686 return __split_vma(mm, vma, addr, new_below);
2687}
2688
2689/* Munmap is split into 2 main parts -- this part which finds
2690 * what needs doing, and the areas themselves, which do the
2691 * work. This now handles partial unmappings.
2692 * Jeremy Fitzhardinge <jeremy@goop.org>
2693 */
2694int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2695 struct list_head *uf)
2696{
2697 unsigned long end;
2698 struct vm_area_struct *vma, *prev, *last;
2699
2700 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2701 return -EINVAL;
2702
2703 len = PAGE_ALIGN(len);
2704 if (len == 0)
2705 return -EINVAL;
2706
2707 /* Find the first overlapping VMA */
2708 vma = find_vma(mm, start);
2709 if (!vma)
2710 return 0;
2711 prev = vma->vm_prev;
2712 /* we have start < vma->vm_end */
2713
2714 /* if it doesn't overlap, we have nothing.. */
2715 end = start + len;
2716 if (vma->vm_start >= end)
2717 return 0;
2718
2719 /*
2720 * If we need to split any vma, do it now to save pain later.
2721 *
2722 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2723 * unmapped vm_area_struct will remain in use: so lower split_vma
2724 * places tmp vma above, and higher split_vma places tmp vma below.
2725 */
2726 if (start > vma->vm_start) {
2727 int error;
2728
2729 /*
2730 * Make sure that map_count on return from munmap() will
2731 * not exceed its limit; but let map_count go just above
2732 * its limit temporarily, to help free resources as expected.
2733 */
2734 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2735 return -ENOMEM;
2736
2737 error = __split_vma(mm, vma, start, 0);
2738 if (error)
2739 return error;
2740 prev = vma;
2741 }
2742
2743 /* Does it split the last one? */
2744 last = find_vma(mm, end);
2745 if (last && end > last->vm_start) {
2746 int error = __split_vma(mm, last, end, 1);
2747 if (error)
2748 return error;
2749 }
2750 vma = prev ? prev->vm_next : mm->mmap;
2751
2752 if (unlikely(uf)) {
2753 /*
2754 * If userfaultfd_unmap_prep returns an error the vmas
2755 * will remain splitted, but userland will get a
2756 * highly unexpected error anyway. This is no
2757 * different than the case where the first of the two
2758 * __split_vma fails, but we don't undo the first
2759 * split, despite we could. This is unlikely enough
2760 * failure that it's not worth optimizing it for.
2761 */
2762 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2763 if (error)
2764 return error;
2765 }
2766
2767 /*
2768 * unlock any mlock()ed ranges before detaching vmas
2769 */
2770 if (mm->locked_vm) {
2771 struct vm_area_struct *tmp = vma;
2772 while (tmp && tmp->vm_start < end) {
2773 if (tmp->vm_flags & VM_LOCKED) {
2774 mm->locked_vm -= vma_pages(tmp);
2775 munlock_vma_pages_all(tmp);
2776 }
2777 tmp = tmp->vm_next;
2778 }
2779 }
2780
2781 /*
2782 * Remove the vma's, and unmap the actual pages
2783 */
2784 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2785 unmap_region(mm, vma, prev, start, end);
2786
2787 arch_unmap(mm, vma, start, end);
2788
2789 /* Fix up all other VM information */
2790 remove_vma_list(mm, vma);
2791
2792 return 0;
2793}
2794
2795int vm_munmap(unsigned long start, size_t len)
2796{
2797 int ret;
2798 struct mm_struct *mm = current->mm;
2799 LIST_HEAD(uf);
2800
2801 if (down_write_killable(&mm->mmap_sem))
2802 return -EINTR;
2803
2804 ret = do_munmap(mm, start, len, &uf);
2805 up_write(&mm->mmap_sem);
2806 userfaultfd_unmap_complete(mm, &uf);
2807 return ret;
2808}
2809EXPORT_SYMBOL(vm_munmap);
2810
2811SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2812{
2813 profile_munmap(addr);
2814 return vm_munmap(addr, len);
2815}
2816
2817
2818/*
2819 * Emulation of deprecated remap_file_pages() syscall.
2820 */
2821SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2822 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2823{
2824
2825 struct mm_struct *mm = current->mm;
2826 struct vm_area_struct *vma;
2827 unsigned long populate = 0;
2828 unsigned long ret = -EINVAL;
2829 struct file *file;
2830
2831 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2832 current->comm, current->pid);
2833
2834 if (prot)
2835 return ret;
2836 start = start & PAGE_MASK;
2837 size = size & PAGE_MASK;
2838
2839 if (start + size <= start)
2840 return ret;
2841
2842 /* Does pgoff wrap? */
2843 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2844 return ret;
2845
2846 if (down_write_killable(&mm->mmap_sem))
2847 return -EINTR;
2848
2849 vma = find_vma(mm, start);
2850
2851 if (!vma || !(vma->vm_flags & VM_SHARED))
2852 goto out;
2853
2854 if (start < vma->vm_start)
2855 goto out;
2856
2857 if (start + size > vma->vm_end) {
2858 struct vm_area_struct *next;
2859
2860 for (next = vma->vm_next; next; next = next->vm_next) {
2861 /* hole between vmas ? */
2862 if (next->vm_start != next->vm_prev->vm_end)
2863 goto out;
2864
2865 if (next->vm_file != vma->vm_file)
2866 goto out;
2867
2868 if (next->vm_flags != vma->vm_flags)
2869 goto out;
2870
2871 if (start + size <= next->vm_end)
2872 break;
2873 }
2874
2875 if (!next)
2876 goto out;
2877 }
2878
2879 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2880 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2881 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2882
2883 flags &= MAP_NONBLOCK;
2884 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2885 if (vma->vm_flags & VM_LOCKED) {
2886 struct vm_area_struct *tmp;
2887 flags |= MAP_LOCKED;
2888
2889 /* drop PG_Mlocked flag for over-mapped range */
2890 for (tmp = vma; tmp->vm_start >= start + size;
2891 tmp = tmp->vm_next) {
2892 /*
2893 * Split pmd and munlock page on the border
2894 * of the range.
2895 */
2896 vma_adjust_trans_huge(tmp, start, start + size, 0);
2897
2898 munlock_vma_pages_range(tmp,
2899 max(tmp->vm_start, start),
2900 min(tmp->vm_end, start + size));
2901 }
2902 }
2903
2904 file = get_file(vma->vm_file);
2905 ret = do_mmap_pgoff(vma->vm_file, start, size,
2906 prot, flags, pgoff, &populate, NULL);
2907 fput(file);
2908out:
2909 up_write(&mm->mmap_sem);
2910 if (populate)
2911 mm_populate(ret, populate);
2912 if (!IS_ERR_VALUE(ret))
2913 ret = 0;
2914 return ret;
2915}
2916
2917static inline void verify_mm_writelocked(struct mm_struct *mm)
2918{
2919#ifdef CONFIG_DEBUG_VM
2920 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2921 WARN_ON(1);
2922 up_read(&mm->mmap_sem);
2923 }
2924#endif
2925}
2926
2927/*
2928 * this is really a simplified "do_mmap". it only handles
2929 * anonymous maps. eventually we may be able to do some
2930 * brk-specific accounting here.
2931 */
2932static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags, struct list_head *uf)
2933{
2934 struct mm_struct *mm = current->mm;
2935 struct vm_area_struct *vma, *prev;
2936 unsigned long len;
2937 struct rb_node **rb_link, *rb_parent;
2938 pgoff_t pgoff = addr >> PAGE_SHIFT;
2939 int error;
2940
2941 len = PAGE_ALIGN(request);
2942 if (len < request)
2943 return -ENOMEM;
2944 if (!len)
2945 return 0;
2946
2947 /* Until we need other flags, refuse anything except VM_EXEC. */
2948 if ((flags & (~VM_EXEC)) != 0)
2949 return -EINVAL;
2950 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2951
2952 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2953 if (offset_in_page(error))
2954 return error;
2955
2956 error = mlock_future_check(mm, mm->def_flags, len);
2957 if (error)
2958 return error;
2959
2960 /*
2961 * mm->mmap_sem is required to protect against another thread
2962 * changing the mappings in case we sleep.
2963 */
2964 verify_mm_writelocked(mm);
2965
2966 /*
2967 * Clear old maps. this also does some error checking for us
2968 */
2969 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2970 &rb_parent)) {
2971 if (do_munmap(mm, addr, len, uf))
2972 return -ENOMEM;
2973 }
2974
2975 /* Check against address space limits *after* clearing old maps... */
2976 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2977 return -ENOMEM;
2978
2979 if (mm->map_count > sysctl_max_map_count)
2980 return -ENOMEM;
2981
2982 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2983 return -ENOMEM;
2984
2985 /* Can we just expand an old private anonymous mapping? */
2986 vma = vma_merge(mm, prev, addr, addr + len, flags,
2987 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2988 if (vma)
2989 goto out;
2990
2991 /*
2992 * create a vma struct for an anonymous mapping
2993 */
2994 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2995 if (!vma) {
2996 vm_unacct_memory(len >> PAGE_SHIFT);
2997 return -ENOMEM;
2998 }
2999
3000 INIT_LIST_HEAD(&vma->anon_vma_chain);
3001 vma->vm_mm = mm;
3002 vma->vm_start = addr;
3003 vma->vm_end = addr + len;
3004 vma->vm_pgoff = pgoff;
3005 vma->vm_flags = flags;
3006 vma->vm_page_prot = vm_get_page_prot(flags);
3007 vma_link(mm, vma, prev, rb_link, rb_parent);
3008out:
3009 perf_event_mmap(vma);
3010 mm->total_vm += len >> PAGE_SHIFT;
3011 mm->data_vm += len >> PAGE_SHIFT;
3012 if (flags & VM_LOCKED)
3013 mm->locked_vm += (len >> PAGE_SHIFT);
3014 vma->vm_flags |= VM_SOFTDIRTY;
3015 return 0;
3016}
3017
3018static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf)
3019{
3020 return do_brk_flags(addr, len, 0, uf);
3021}
3022
3023int vm_brk_flags(unsigned long addr, unsigned long len, unsigned long flags)
3024{
3025 struct mm_struct *mm = current->mm;
3026 int ret;
3027 bool populate;
3028 LIST_HEAD(uf);
3029
3030 if (down_write_killable(&mm->mmap_sem))
3031 return -EINTR;
3032
3033 ret = do_brk_flags(addr, len, flags, &uf);
3034 populate = ((mm->def_flags & VM_LOCKED) != 0);
3035 up_write(&mm->mmap_sem);
3036 userfaultfd_unmap_complete(mm, &uf);
3037 if (populate && !ret)
3038 mm_populate(addr, len);
3039 return ret;
3040}
3041EXPORT_SYMBOL(vm_brk_flags);
3042
3043int vm_brk(unsigned long addr, unsigned long len)
3044{
3045 return vm_brk_flags(addr, len, 0);
3046}
3047EXPORT_SYMBOL(vm_brk);
3048
3049/* Release all mmaps. */
3050void exit_mmap(struct mm_struct *mm)
3051{
3052 struct mmu_gather tlb;
3053 struct vm_area_struct *vma;
3054 unsigned long nr_accounted = 0;
3055
3056 /* mm's last user has gone, and its about to be pulled down */
3057 mmu_notifier_release(mm);
3058
3059 if (unlikely(mm_is_oom_victim(mm))) {
3060 /*
3061 * Manually reap the mm to free as much memory as possible.
3062 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3063 * this mm from further consideration. Taking mm->mmap_sem for
3064 * write after setting MMF_OOM_SKIP will guarantee that the oom
3065 * reaper will not run on this mm again after mmap_sem is
3066 * dropped.
3067 *
3068 * Nothing can be holding mm->mmap_sem here and the above call
3069 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3070 * __oom_reap_task_mm() will not block.
3071 *
3072 * This needs to be done before calling munlock_vma_pages_all(),
3073 * which clears VM_LOCKED, otherwise the oom reaper cannot
3074 * reliably test it.
3075 */
3076 mutex_lock(&oom_lock);
3077 __oom_reap_task_mm(mm);
3078 mutex_unlock(&oom_lock);
3079
3080 set_bit(MMF_OOM_SKIP, &mm->flags);
3081 down_write(&mm->mmap_sem);
3082 up_write(&mm->mmap_sem);
3083 }
3084
3085 if (mm->locked_vm) {
3086 vma = mm->mmap;
3087 while (vma) {
3088 if (vma->vm_flags & VM_LOCKED)
3089 munlock_vma_pages_all(vma);
3090 vma = vma->vm_next;
3091 }
3092 }
3093
3094 arch_exit_mmap(mm);
3095
3096 vma = mm->mmap;
3097 if (!vma) /* Can happen if dup_mmap() received an OOM */
3098 return;
3099
3100 lru_add_drain();
3101 flush_cache_mm(mm);
3102 tlb_gather_mmu(&tlb, mm, 0, -1);
3103 /* update_hiwater_rss(mm) here? but nobody should be looking */
3104 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3105 unmap_vmas(&tlb, vma, 0, -1);
3106 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3107 tlb_finish_mmu(&tlb, 0, -1);
3108
3109 /*
3110 * Walk the list again, actually closing and freeing it,
3111 * with preemption enabled, without holding any MM locks.
3112 */
3113 while (vma) {
3114 if (vma->vm_flags & VM_ACCOUNT)
3115 nr_accounted += vma_pages(vma);
3116 vma = remove_vma(vma);
3117 }
3118 vm_unacct_memory(nr_accounted);
3119}
3120
3121/* Insert vm structure into process list sorted by address
3122 * and into the inode's i_mmap tree. If vm_file is non-NULL
3123 * then i_mmap_rwsem is taken here.
3124 */
3125int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3126{
3127 struct vm_area_struct *prev;
3128 struct rb_node **rb_link, *rb_parent;
3129
3130 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3131 &prev, &rb_link, &rb_parent))
3132 return -ENOMEM;
3133 if ((vma->vm_flags & VM_ACCOUNT) &&
3134 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3135 return -ENOMEM;
3136
3137 /*
3138 * The vm_pgoff of a purely anonymous vma should be irrelevant
3139 * until its first write fault, when page's anon_vma and index
3140 * are set. But now set the vm_pgoff it will almost certainly
3141 * end up with (unless mremap moves it elsewhere before that
3142 * first wfault), so /proc/pid/maps tells a consistent story.
3143 *
3144 * By setting it to reflect the virtual start address of the
3145 * vma, merges and splits can happen in a seamless way, just
3146 * using the existing file pgoff checks and manipulations.
3147 * Similarly in do_mmap_pgoff and in do_brk.
3148 */
3149 if (vma_is_anonymous(vma)) {
3150 BUG_ON(vma->anon_vma);
3151 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3152 }
3153
3154 vma_link(mm, vma, prev, rb_link, rb_parent);
3155 return 0;
3156}
3157
3158/*
3159 * Copy the vma structure to a new location in the same mm,
3160 * prior to moving page table entries, to effect an mremap move.
3161 */
3162struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3163 unsigned long addr, unsigned long len, pgoff_t pgoff,
3164 bool *need_rmap_locks)
3165{
3166 struct vm_area_struct *vma = *vmap;
3167 unsigned long vma_start = vma->vm_start;
3168 struct mm_struct *mm = vma->vm_mm;
3169 struct vm_area_struct *new_vma, *prev;
3170 struct rb_node **rb_link, *rb_parent;
3171 bool faulted_in_anon_vma = true;
3172
3173 /*
3174 * If anonymous vma has not yet been faulted, update new pgoff
3175 * to match new location, to increase its chance of merging.
3176 */
3177 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3178 pgoff = addr >> PAGE_SHIFT;
3179 faulted_in_anon_vma = false;
3180 }
3181
3182 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3183 return NULL; /* should never get here */
3184 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3185 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3186 vma->vm_userfaultfd_ctx);
3187 if (new_vma) {
3188 /*
3189 * Source vma may have been merged into new_vma
3190 */
3191 if (unlikely(vma_start >= new_vma->vm_start &&
3192 vma_start < new_vma->vm_end)) {
3193 /*
3194 * The only way we can get a vma_merge with
3195 * self during an mremap is if the vma hasn't
3196 * been faulted in yet and we were allowed to
3197 * reset the dst vma->vm_pgoff to the
3198 * destination address of the mremap to allow
3199 * the merge to happen. mremap must change the
3200 * vm_pgoff linearity between src and dst vmas
3201 * (in turn preventing a vma_merge) to be
3202 * safe. It is only safe to keep the vm_pgoff
3203 * linear if there are no pages mapped yet.
3204 */
3205 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3206 *vmap = vma = new_vma;
3207 }
3208 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3209 } else {
3210 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3211 if (!new_vma)
3212 goto out;
3213 *new_vma = *vma;
3214 new_vma->vm_start = addr;
3215 new_vma->vm_end = addr + len;
3216 new_vma->vm_pgoff = pgoff;
3217 if (vma_dup_policy(vma, new_vma))
3218 goto out_free_vma;
3219 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3220 if (anon_vma_clone(new_vma, vma))
3221 goto out_free_mempol;
3222 if (new_vma->vm_file)
3223 get_file(new_vma->vm_file);
3224 if (new_vma->vm_ops && new_vma->vm_ops->open)
3225 new_vma->vm_ops->open(new_vma);
3226 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3227 *need_rmap_locks = false;
3228 }
3229 return new_vma;
3230
3231out_free_mempol:
3232 mpol_put(vma_policy(new_vma));
3233out_free_vma:
3234 kmem_cache_free(vm_area_cachep, new_vma);
3235out:
3236 return NULL;
3237}
3238
3239/*
3240 * Return true if the calling process may expand its vm space by the passed
3241 * number of pages
3242 */
3243bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3244{
3245 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3246 return false;
3247
3248 if (is_data_mapping(flags) &&
3249 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3250 /* Workaround for Valgrind */
3251 if (rlimit(RLIMIT_DATA) == 0 &&
3252 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3253 return true;
3254
3255 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3256 current->comm, current->pid,
3257 (mm->data_vm + npages) << PAGE_SHIFT,
3258 rlimit(RLIMIT_DATA),
3259 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3260
3261 if (!ignore_rlimit_data)
3262 return false;
3263 }
3264
3265 return true;
3266}
3267
3268void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3269{
3270 mm->total_vm += npages;
3271
3272 if (is_exec_mapping(flags))
3273 mm->exec_vm += npages;
3274 else if (is_stack_mapping(flags))
3275 mm->stack_vm += npages;
3276 else if (is_data_mapping(flags))
3277 mm->data_vm += npages;
3278}
3279
3280static int special_mapping_fault(struct vm_fault *vmf);
3281
3282/*
3283 * Having a close hook prevents vma merging regardless of flags.
3284 */
3285static void special_mapping_close(struct vm_area_struct *vma)
3286{
3287}
3288
3289static const char *special_mapping_name(struct vm_area_struct *vma)
3290{
3291 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3292}
3293
3294static int special_mapping_mremap(struct vm_area_struct *new_vma)
3295{
3296 struct vm_special_mapping *sm = new_vma->vm_private_data;
3297
3298 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3299 return -EFAULT;
3300
3301 if (sm->mremap)
3302 return sm->mremap(sm, new_vma);
3303
3304 return 0;
3305}
3306
3307static const struct vm_operations_struct special_mapping_vmops = {
3308 .close = special_mapping_close,
3309 .fault = special_mapping_fault,
3310 .mremap = special_mapping_mremap,
3311 .name = special_mapping_name,
3312};
3313
3314static const struct vm_operations_struct legacy_special_mapping_vmops = {
3315 .close = special_mapping_close,
3316 .fault = special_mapping_fault,
3317};
3318
3319static int special_mapping_fault(struct vm_fault *vmf)
3320{
3321 struct vm_area_struct *vma = vmf->vma;
3322 pgoff_t pgoff;
3323 struct page **pages;
3324
3325 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3326 pages = vma->vm_private_data;
3327 } else {
3328 struct vm_special_mapping *sm = vma->vm_private_data;
3329
3330 if (sm->fault)
3331 return sm->fault(sm, vmf->vma, vmf);
3332
3333 pages = sm->pages;
3334 }
3335
3336 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3337 pgoff--;
3338
3339 if (*pages) {
3340 struct page *page = *pages;
3341 get_page(page);
3342 vmf->page = page;
3343 return 0;
3344 }
3345
3346 return VM_FAULT_SIGBUS;
3347}
3348
3349static struct vm_area_struct *__install_special_mapping(
3350 struct mm_struct *mm,
3351 unsigned long addr, unsigned long len,
3352 unsigned long vm_flags, void *priv,
3353 const struct vm_operations_struct *ops)
3354{
3355 int ret;
3356 struct vm_area_struct *vma;
3357
3358 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3359 if (unlikely(vma == NULL))
3360 return ERR_PTR(-ENOMEM);
3361
3362 INIT_LIST_HEAD(&vma->anon_vma_chain);
3363 vma->vm_mm = mm;
3364 vma->vm_start = addr;
3365 vma->vm_end = addr + len;
3366
3367 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3368 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3369
3370 vma->vm_ops = ops;
3371 vma->vm_private_data = priv;
3372
3373 ret = insert_vm_struct(mm, vma);
3374 if (ret)
3375 goto out;
3376
3377 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3378
3379 perf_event_mmap(vma);
3380
3381 return vma;
3382
3383out:
3384 kmem_cache_free(vm_area_cachep, vma);
3385 return ERR_PTR(ret);
3386}
3387
3388bool vma_is_special_mapping(const struct vm_area_struct *vma,
3389 const struct vm_special_mapping *sm)
3390{
3391 return vma->vm_private_data == sm &&
3392 (vma->vm_ops == &special_mapping_vmops ||
3393 vma->vm_ops == &legacy_special_mapping_vmops);
3394}
3395
3396/*
3397 * Called with mm->mmap_sem held for writing.
3398 * Insert a new vma covering the given region, with the given flags.
3399 * Its pages are supplied by the given array of struct page *.
3400 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3401 * The region past the last page supplied will always produce SIGBUS.
3402 * The array pointer and the pages it points to are assumed to stay alive
3403 * for as long as this mapping might exist.
3404 */
3405struct vm_area_struct *_install_special_mapping(
3406 struct mm_struct *mm,
3407 unsigned long addr, unsigned long len,
3408 unsigned long vm_flags, const struct vm_special_mapping *spec)
3409{
3410 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3411 &special_mapping_vmops);
3412}
3413
3414int install_special_mapping(struct mm_struct *mm,
3415 unsigned long addr, unsigned long len,
3416 unsigned long vm_flags, struct page **pages)
3417{
3418 struct vm_area_struct *vma = __install_special_mapping(
3419 mm, addr, len, vm_flags, (void *)pages,
3420 &legacy_special_mapping_vmops);
3421
3422 return PTR_ERR_OR_ZERO(vma);
3423}
3424
3425static DEFINE_MUTEX(mm_all_locks_mutex);
3426
3427static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3428{
3429 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3430 /*
3431 * The LSB of head.next can't change from under us
3432 * because we hold the mm_all_locks_mutex.
3433 */
3434 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3435 /*
3436 * We can safely modify head.next after taking the
3437 * anon_vma->root->rwsem. If some other vma in this mm shares
3438 * the same anon_vma we won't take it again.
3439 *
3440 * No need of atomic instructions here, head.next
3441 * can't change from under us thanks to the
3442 * anon_vma->root->rwsem.
3443 */
3444 if (__test_and_set_bit(0, (unsigned long *)
3445 &anon_vma->root->rb_root.rb_root.rb_node))
3446 BUG();
3447 }
3448}
3449
3450static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3451{
3452 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3453 /*
3454 * AS_MM_ALL_LOCKS can't change from under us because
3455 * we hold the mm_all_locks_mutex.
3456 *
3457 * Operations on ->flags have to be atomic because
3458 * even if AS_MM_ALL_LOCKS is stable thanks to the
3459 * mm_all_locks_mutex, there may be other cpus
3460 * changing other bitflags in parallel to us.
3461 */
3462 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3463 BUG();
3464 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3465 }
3466}
3467
3468/*
3469 * This operation locks against the VM for all pte/vma/mm related
3470 * operations that could ever happen on a certain mm. This includes
3471 * vmtruncate, try_to_unmap, and all page faults.
3472 *
3473 * The caller must take the mmap_sem in write mode before calling
3474 * mm_take_all_locks(). The caller isn't allowed to release the
3475 * mmap_sem until mm_drop_all_locks() returns.
3476 *
3477 * mmap_sem in write mode is required in order to block all operations
3478 * that could modify pagetables and free pages without need of
3479 * altering the vma layout. It's also needed in write mode to avoid new
3480 * anon_vmas to be associated with existing vmas.
3481 *
3482 * A single task can't take more than one mm_take_all_locks() in a row
3483 * or it would deadlock.
3484 *
3485 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3486 * mapping->flags avoid to take the same lock twice, if more than one
3487 * vma in this mm is backed by the same anon_vma or address_space.
3488 *
3489 * We take locks in following order, accordingly to comment at beginning
3490 * of mm/rmap.c:
3491 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3492 * hugetlb mapping);
3493 * - all i_mmap_rwsem locks;
3494 * - all anon_vma->rwseml
3495 *
3496 * We can take all locks within these types randomly because the VM code
3497 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3498 * mm_all_locks_mutex.
3499 *
3500 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3501 * that may have to take thousand of locks.
3502 *
3503 * mm_take_all_locks() can fail if it's interrupted by signals.
3504 */
3505int mm_take_all_locks(struct mm_struct *mm)
3506{
3507 struct vm_area_struct *vma;
3508 struct anon_vma_chain *avc;
3509
3510 BUG_ON(down_read_trylock(&mm->mmap_sem));
3511
3512 mutex_lock(&mm_all_locks_mutex);
3513
3514 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3515 if (signal_pending(current))
3516 goto out_unlock;
3517 if (vma->vm_file && vma->vm_file->f_mapping &&
3518 is_vm_hugetlb_page(vma))
3519 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3520 }
3521
3522 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3523 if (signal_pending(current))
3524 goto out_unlock;
3525 if (vma->vm_file && vma->vm_file->f_mapping &&
3526 !is_vm_hugetlb_page(vma))
3527 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3528 }
3529
3530 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3531 if (signal_pending(current))
3532 goto out_unlock;
3533 if (vma->anon_vma)
3534 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3535 vm_lock_anon_vma(mm, avc->anon_vma);
3536 }
3537
3538 return 0;
3539
3540out_unlock:
3541 mm_drop_all_locks(mm);
3542 return -EINTR;
3543}
3544
3545static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3546{
3547 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3548 /*
3549 * The LSB of head.next can't change to 0 from under
3550 * us because we hold the mm_all_locks_mutex.
3551 *
3552 * We must however clear the bitflag before unlocking
3553 * the vma so the users using the anon_vma->rb_root will
3554 * never see our bitflag.
3555 *
3556 * No need of atomic instructions here, head.next
3557 * can't change from under us until we release the
3558 * anon_vma->root->rwsem.
3559 */
3560 if (!__test_and_clear_bit(0, (unsigned long *)
3561 &anon_vma->root->rb_root.rb_root.rb_node))
3562 BUG();
3563 anon_vma_unlock_write(anon_vma);
3564 }
3565}
3566
3567static void vm_unlock_mapping(struct address_space *mapping)
3568{
3569 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3570 /*
3571 * AS_MM_ALL_LOCKS can't change to 0 from under us
3572 * because we hold the mm_all_locks_mutex.
3573 */
3574 i_mmap_unlock_write(mapping);
3575 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3576 &mapping->flags))
3577 BUG();
3578 }
3579}
3580
3581/*
3582 * The mmap_sem cannot be released by the caller until
3583 * mm_drop_all_locks() returns.
3584 */
3585void mm_drop_all_locks(struct mm_struct *mm)
3586{
3587 struct vm_area_struct *vma;
3588 struct anon_vma_chain *avc;
3589
3590 BUG_ON(down_read_trylock(&mm->mmap_sem));
3591 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3592
3593 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3594 if (vma->anon_vma)
3595 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3596 vm_unlock_anon_vma(avc->anon_vma);
3597 if (vma->vm_file && vma->vm_file->f_mapping)
3598 vm_unlock_mapping(vma->vm_file->f_mapping);
3599 }
3600
3601 mutex_unlock(&mm_all_locks_mutex);
3602}
3603
3604/*
3605 * initialise the percpu counter for VM
3606 */
3607void __init mmap_init(void)
3608{
3609 int ret;
3610
3611 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3612 VM_BUG_ON(ret);
3613}
3614
3615/*
3616 * Initialise sysctl_user_reserve_kbytes.
3617 *
3618 * This is intended to prevent a user from starting a single memory hogging
3619 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3620 * mode.
3621 *
3622 * The default value is min(3% of free memory, 128MB)
3623 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3624 */
3625static int init_user_reserve(void)
3626{
3627 unsigned long free_kbytes;
3628
3629 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3630
3631 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3632 return 0;
3633}
3634subsys_initcall(init_user_reserve);
3635
3636/*
3637 * Initialise sysctl_admin_reserve_kbytes.
3638 *
3639 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3640 * to log in and kill a memory hogging process.
3641 *
3642 * Systems with more than 256MB will reserve 8MB, enough to recover
3643 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3644 * only reserve 3% of free pages by default.
3645 */
3646static int init_admin_reserve(void)
3647{
3648 unsigned long free_kbytes;
3649
3650 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3651
3652 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3653 return 0;
3654}
3655subsys_initcall(init_admin_reserve);
3656
3657/*
3658 * Reinititalise user and admin reserves if memory is added or removed.
3659 *
3660 * The default user reserve max is 128MB, and the default max for the
3661 * admin reserve is 8MB. These are usually, but not always, enough to
3662 * enable recovery from a memory hogging process using login/sshd, a shell,
3663 * and tools like top. It may make sense to increase or even disable the
3664 * reserve depending on the existence of swap or variations in the recovery
3665 * tools. So, the admin may have changed them.
3666 *
3667 * If memory is added and the reserves have been eliminated or increased above
3668 * the default max, then we'll trust the admin.
3669 *
3670 * If memory is removed and there isn't enough free memory, then we
3671 * need to reset the reserves.
3672 *
3673 * Otherwise keep the reserve set by the admin.
3674 */
3675static int reserve_mem_notifier(struct notifier_block *nb,
3676 unsigned long action, void *data)
3677{
3678 unsigned long tmp, free_kbytes;
3679
3680 switch (action) {
3681 case MEM_ONLINE:
3682 /* Default max is 128MB. Leave alone if modified by operator. */
3683 tmp = sysctl_user_reserve_kbytes;
3684 if (0 < tmp && tmp < (1UL << 17))
3685 init_user_reserve();
3686
3687 /* Default max is 8MB. Leave alone if modified by operator. */
3688 tmp = sysctl_admin_reserve_kbytes;
3689 if (0 < tmp && tmp < (1UL << 13))
3690 init_admin_reserve();
3691
3692 break;
3693 case MEM_OFFLINE:
3694 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3695
3696 if (sysctl_user_reserve_kbytes > free_kbytes) {
3697 init_user_reserve();
3698 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3699 sysctl_user_reserve_kbytes);
3700 }
3701
3702 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3703 init_admin_reserve();
3704 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3705 sysctl_admin_reserve_kbytes);
3706 }
3707 break;
3708 default:
3709 break;
3710 }
3711 return NOTIFY_OK;
3712}
3713
3714static struct notifier_block reserve_mem_nb = {
3715 .notifier_call = reserve_mem_notifier,
3716};
3717
3718static int __meminit init_reserve_notifier(void)
3719{
3720 if (register_hotmemory_notifier(&reserve_mem_nb))
3721 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3722
3723 return 0;
3724}
3725subsys_initcall(init_reserve_notifier);